Forecasting Urban Water Demand

Forecasting
Forecasting Urban Water Demand, Second Edition

Urban
Completely revised and updated, Forecasting Urban Water Demand
includes new materials on selecting a forecasting approach, forecasting
water demand by customer class, seasonal water demand models, short-
term water demand models, climate change, emergency water conservation
Water
or water restriction measures, long-term water conservation, and risk
simulation. There also are expanded treatments of rapidly-growing and
vitally important areas such as consumer rate responses and short-term
water conservation.
Demand
Forecasting Urban Water Demand gives a comprehensive viewpoint for
methods of statistical analysis, criteria for forecasts, and information
about population and employment trends, economics, weather, climate,
and customer responses to water rates and conservation programs.
R. Bruce Billings
Clive V. Jones
Second Edition
AWWA is the authoritative resource for knowledge, information, and advocacy
to improve the quality and supply of water in North America and beyond. AWWA
Advocacy
is the largest organization of water professionals in the world. AWWA advances
Communications
public health, safety and welfare by uniting the efforts of the full spectrum of
Conferences
the entire water community. Through our collective strength we become better
Education and Training
stewards of water for the greatest good of the people and the environment.
c Science and Technology
Sections
1P-7.5C-20410-2/08-SB
The Authoritative Resource on Safe Water ®
20410 cover 1.indd 1 2/6/2008 8:31:10 AM

20410-A Forecasting.book Page i Tuesday, January 29, 2008 3:46 PM
FORECASTING
URBAN
WATER
DEMAND
20410-A Forecasting.book Page ii Tuesday, January 29, 2008 3:46 PM
20410-A Forecasting.book Page iii Tuesday, January 29, 2008 3:46 PM
FORECASTING
URBAN
WATER
DEMAND
Second Edition
R. Bruce Billings
Clive V. Jones
Science and Technology

AWWA unites the drinking water community by developing and distributing authoritative scientific and technological
knowledge. Through its members, AWWA develops industry standards for products and processes that advance
public health and safety. AWWA also provides quality improvement programs for water and wastewater utilities.
20410-A Forecasting.book Page iv Tuesday, January 29, 2008 3:46 PM
Forecasting Urban Water Demand, Second Edition

Copyright ©1996, 2008 American Water Works Association
All rights reserved. No part of this publication may be reproduced or transmitted in any form
or by any means, electronic or mechanical, including photocopy, recording, or any
information or retrieval system, except in the form of brief excerpts or quotations for review
purposes, without the written permission of the publisher.
Disclaimer
This book is provided for informational purposes only, with the understanding that the
publisher, editors, and authors are not thereby engaged in rendering engineering or other
professional services. The authors, editors, and publisher make no claim as to the accuracy of
the book's contents, or their applicability to any particular circumstance. The editors, authors,
and publisher accept no liability to any person for the information or advice provided in this
book or for loss or damages incurred by any person as a result of reliance on its contents. The
reader is urged to consult with an appropriate licensed professional before taking any action or
making any interpretation that is within the realm of a licensed professional practice.
AWWA Publications Manager: Gay Porter De Nileon

Senior Technical Editor/Project Manager: Melissa Valentine
Copy Editor: René S. Howard
Cover Art/Production Editor: Cheryl Armstrong
Library of Congress Cataloging-in-Publication Data
Billings, R. Bruce.
Forecasting Urban Water Demand / R. Bruce Billings, Clive V. Jones. -- 2nd ed.
p. cm.
Includes bibliographical references and index.
ISBN-13: 978-1-58231-537-1
ISBN-10: 1-58231-537-9
1. Municipal water supply--Management. 2. Municipal water supply--
Forecasting--Statistical methods. I. Jones, Clive Vaughan. II. Title.
HD4456.B55 2007
333.91'2--dc22
2007025815
ISBN 1-58321-537-9
9781583215371
6666 West Quincy Avenue

Denver, CO 80235-3098
303.794.7711
www.awwa.org
20410-A Forecasting.book Page v Tuesday, January 29, 2008 3:47 PM
CONTENTS
LIST OF FIGURES ix
LIST OF TABLES xiii
INTRODUCTION xv
ACKNOWLEDGMENTS xvii
CHAPTER 1
Introduction To Water Forecasting 1
What Is Water Demand? 3
Forecast Horizon and Forecast Applications 6
Major Influences on Water Demand 9
Water-Demand Forecasting 10
Organization of This Book 11
References 16
CHAPTER 2
Basic Facts About Urban Water Demand 17
Trends and Patterns in per Capita Water Use 18
Statistical Profiles of Urban Water Use 20
Major Components of Water System Demand 21
Other Basic Facts About Urban Water Use 27
Summary and Conclusions 28
References 30
CHAPTER 3
Choosing A Forecasting Approach 31
Forecast Methods and Models Currently in Use 31
Forecasting Models 33
Selecting a Forecasting Method—A Pragmatic Approach 35
Forecasting Purpose 37
Customer Disaggregation 37
Data Availability 38
References 43
v
20410-A Forecasting.book Page vi Tuesday, January 29, 2008 3:47 PM
CHAPTER 4
Data and Data Structures for Water-Demand Forecasting 45
Collecting Water-Use Data 46
Organizing Data 47
How Much Data Disaggregation? 48
Other Variables 50
Basic Analysis of Water Use 51
Summary 58
CHAPTER 5
Simple Forecasting Methods and Reality Checks 59
Per Capita Water-Demand Forecasts 61
Sectoral Water-Demand Forecasts 65
Reality Checks 79
Summary and Concusions 80
References 82
CHAPTER 6
Forecasting Seasonal and Peak Water Demand 83
Peak-to-Average Day Demand Ratios 86
Averaging for Seasonal Patterns and Interannual Trends 87
Regression Approaches 93
Applications to Forecasting and Risk Evaluation 104
Summary 105
References 107
CHAPTER 7
Population, Employment, and Technology Forecasts 109
Population Forecasts 110
National Population Trends 110
Regional Population Trends 112
Regional and Local Population Forecasts 113
Employment Forecasts 114
Regional Economic Forecasts 116
Technology Forecasts 119
Checklists 121
References 125
CHAPTER 8
Weather and Climate 127
Weather 128
Climate 135
vi
20410-A Forecasting.book Page vii Tuesday, January 29, 2008 3:47 PM

References 144
CHAPTER 9
Price Effects 147
Demand Curves for Water 148
Schedules of Rates and Fees 149
Rate Design and Revenue Stability 154
What Is the “Price” of Water? 158
Price Elasticities of Demand and Revenue 160
Using Price Elasticities in Demand Forecasts 166
Estimating the Impact of Metering and Rate Changes 169
Econometric Studies 170
Single-family Residential Price Elasticity of Water Demand 172
Summary 176
References 178
CHAPTER 10
Long-Term Water Conservation 181
Evaluating Water Conservation Programs 183
Structural Conservation Programs 188
Behavioral Conservation Programs 204
Incorporating Conservation Into the Forecast 211
Management of Conservation Programs 214
Summary 222
References 224
CHAPTER 11
Short-Term Emergency Conservation 227
Effectiveness of Short-Term Conservation Programs 228
Considerations in Developing a Conservation Program 230
Program Elements 233
Drought Planning 239
Summary 251
References 252
CHAPTER 12
Forecasting With Regression 253
Regression Modeling—General Concepts 254
Estimating a Regression Model With Household Data 259
Estimating a Regression Model With Time-Series Data 269
Cross-Section, Time-Series Models 271
The Validity of Regression Models 272
Summary 277
vii
20410-A Forecasting.book Page viii Tuesday, January 29, 2008 3:47 PM
References 279
CHAPTER 13
Advanced Topics and Methods 281
Time-Series Analysis 281
Neural Network Models 285
Statistical Analysis of Consumer Response to Rate Schedules 288
Risk Simulation 293
Summary 297
References 298
CHAPTER 14
Forecast Uses, Evaluation, and Improvement 299
Forecast Horizon and Approach 300
Forecast Interval Estimates 302
Forecast Evaluation 303
Error Measures and Loss Functions 305
Strategies to Improve Accuracy 305
Why Forecast Formally? 306
References 308
APPENDIX A
Survey Methods for Direct Data Collection 309
APPENDIX B
AWWA Water Demand Survey Findings 321
GLOSSARY 329
AUTHOR BIOGRAPHIES 335
INDEX 337
CD-ROM OF SELECTED DATA TABLES (INSIDE BACK COVER)
viii
20410-A Forecasting.book Page ix Tuesday, January 29, 2008 3:47 PM
LIST OF FIGURES
1-1 Schematic of Demand Curve....................................................................... 5
2-1 Per Capita Water Usage in the United States in 2000 ............................18
2-2 Total System Water Use, Annual Data ......................................................20
2-3 Total System Water Use, Monthly Data ....................................................21
2-4 Peak-to-Average Day Pumpage Ratios......................................................22
2-5 Monthly Residential Water Use, City of Los Angeles, 1980–1995.........22
2-6 Distribution of Water Use by Residential Customers...............................23
2-7 Water Losses Reported in the AWWA WDS ..............................................27
4-1 Time-Series of Water Use by Month..........................................................53

4-2 Cross-Section of Water Use per Customer ...............................................53
4-3 Peak-to-Average Water Use Ratio..............................................................57
5-1 Simplified Flow of Work Diagram for Sectoral Water-Demand

Forecast ...............................................................................................67
5-2 Historical Residential and CI Water-Use Rates ........................................76
6-1 Historical Peak-to-Average Day Demand ..................................................84

6-2 Average Monthly Proportions of Annual Water Production .....................85
6-3 Monthly Total System Water Use ..............................................................85
6-4 Indoor and Outdoor Water Use..................................................................90
6-5 Trend Indicated by a Centered Moving Average.......................................93
6-6 Comparison of Summer Water Use and Values Predicted by
Monthly Water Demand ....................................................................101
6-7 Projected and Actual Water Use..............................................................104
7-1 Historical Population From 1960 to 2000 and Projections to 2050 ...111
8-1 Average Daily Temperatures, Atlanta, Ga. ..............................................129

8-2 Daily Precipitation, Atlanta, Ga................................................................129
8-3 Palmer Drought Severity Index ................................................................132
8-4 Climograph for Austin, Texas. ..................................................................136
8-5 Climograph for Boston, Mass. .................................................................136
8-6 Climograph for Chicago, Ill. ......................................................................137
8-7 Climograph for Phoenix, Ariz....................................................................137
8-8 Climograph for Kansas City, Mo. .............................................................138
8-9 Climograph for Los Angeles, Calif............................................................138
8-10 Climograph for San Francisco, Calif. .......................................................139
8-11 Climograph for Tampa, Fla.......................................................................139
9-1 Typical Linear and Exponential Demand Curves....................................149

9-2 Uniform Commodity Rate and Average Price With a
$7.00 Monthly Charge ......................................................................151
ix
20410-A Forecasting.book Page x Tuesday, January 29, 2008 3:47 PM
9-3 Decreasing Block Rates and Average Price With a $7.00

Monthly Charge................................................................................. 151
9-4 Increasing Block Rates and Average Price With a $7.00
Monthly Charge................................................................................. 152
9-5 Typical Increasing Step Rate................................................................... 154
9-6 Demand for Water by Elastic Generation............................................... 176
10-1 Water Use in Cities .................................................................................. 186

10-2 Typical Single-Family Residential Indoor Water Use, by Percentage ... 190
10-3 Toilet Flush Volume ................................................................................. 191
10-4 Water Use by Washroom Faucets .......................................................... 197
10-5 Water Use by Turf and Xeriscape ........................................................... 199
10-6 Area Converted From Turf to Xeriscape, Las Vegas, Nev...................... 199
10-7 Industrial Conservation Potential by Sector........................................... 202
10-8 A Water Bill That Fails to Effectively Show the Marginal
Price of Water................................................................................ 206
10-9 A Water Bill That Effectively Displays the Marginal Cost of Water ....... 207
10-10 Residential Water Conservation Model.................................................. 219
11-1 Flow Diagram of Goal Achievement Process ......................................... 240
12-1 Frequency Distribution of Water Use ..................................................... 263
13-1 Historical Population and Trend Extrapolations .................................... 283

13-2 A Simple Diagram of a Multilayer Neural Network ................................ 286
13-3 Five Simulations of Population Growth Over the Forecast Horizon ..... 295
13-4 Frequency Distribution of Simulated Population................................... 296
13-5 Simulation Results for 2037 System Water Demand........................... 296
x
20410-A Forecasting.book Page xi Tuesday, January 29, 2008 3:47 PM
LIST OF TABLES
1-1 Types of Urban Water-Demand Forecasts Reported in the
AWWA WDS ............................................................................................ 2
1-2 Types of Water-Demand Forecasts and Major Applications...................... 7
2-1 Water Withdrawals for Public Water Supplies and per Capita
Water Use in the United States ..........................................................19
2-2 Indoor Residential End Uses of Water ......................................................25
2-3 Commercial and Industrial Water Use by Major Category.......................26
3-1 Methods for Developing Water-Demand Forecasts Reported

in the AWWA Water-Demand Survey..................................................32
4-1 Time-Series Data Format ...........................................................................52

4-2 Cross-Section Data Format........................................................................52
4-3 Simple Correlations of Monthly Data, Full Years......................................56
4-4 Simple Correlations, Summer Only ...........................................................56
4-5 Peak-to-Average Water Use by Month.......................................................57
5-1 Simple per Capita Forecast Approach ......................................................61

5-2 Database for Analysis of per Capita Water Demand ...............................63
5-3 Data for Sectoral Water-Demand Forecast ..............................................68
5-4 Data for Sectoral Water-Demand Forecast, Reference Forecast............71
5-5 Data for Sectoral Water-Demand Forecast, Adjusted Forecast ..............72
5-6 Water Savings From Mandate and Rebate Conservation Program ........73
5-7 Comparing the Forecasts...........................................................................78
6-1 Data Array for Calculating Monthly Proportions of

Annual Water Usage............................................................................88
6-2 Monthly Proportions of Annual Water Usage............................................88
6-3 A Simple Moving Average...........................................................................91
6-4 Calculating a Centered Moving Average ...................................................92
6-5 Sample of Regression Database...............................................................96
6-6 Regression Output for Estimating Equation 6-3.......................................99
6-7 Daily Demand Regression Model ........................................................... 100
6-8 Daily Demand Regression Model ........................................................... 103
7-1 Most Rapidly Growing Metropolitan Areas, 1990–2000 ..................... 112

7-2 Employment and Output by Industry, 1994, 2004,
and Projected 2014 ......................................................................... 115
9-1 Water Rate Schedules ............................................................................ 152

9-2 Typical Step Rate Schedule, Customer Bill, and Marginal Cost ........... 155
9-3 Effect on Revenue of a 10 Percent Drop in Sales by Rate Structure .. 156
9-4 Effect of a Drop in Use, Marginal Cost, and Rebate Schedule............. 158
9-5 Rule-of-Thumb Elasticity Values ............................................................. 165
xi
20410-A Forecasting.book Page xii Tuesday, January 29, 2008 3:47 PM
9-6 Detailed Computations for a 200 Percent Price Increase.................... 169

9-7 Water Use in Metered and Flat Rate Areas, Western United States.... 170
9-8 Controlled Comparison of Water Usage Rates ...................................... 170
9-9 Selected Residential Price Elasciticy Estimates.................................... 173
9-10 Estimated Price Elasticities by Region for Selected Price Levels......... 174
9-11 Commercial Price Elasticity Estimates ................................................... 174
9-12 Industrial Price Elasticity Estimates ....................................................... 175
10-1 Anticipated Water Savings by Device or Program ................................. 182

10-2 Water Conservation Programs, US Utilities, October 2006 .................. 183
10-3 Annual Indoor and Outdoor Water Use per Household......................... 189
10-4 Metropolitan Water District Toilet Inventory .......................................... 193
10-5 Estimated Savings per Toilet .................................................................. 194
10-6 Daily Shower Use ..................................................................................... 195
10-7 Potential Water Savings and Payback for Santa Clara, Calif................ 203
10-8 Confidence Intervals for Price Effects on Water Use and Revenue ........208
10-9 Shift From Constant to Increasing Block Rates, Uniform Elasticity ..... 210
10-10 Iterative Computation of Large Price Effects on Quantity..................... 211
10-11 Shift From Uniform to Increasing Block Rates:
Differentiated Elasticities................................................................. 212
10-12 San Francisco Wholesale Demand Forecast and
Conservation Savings ....................................................................... 214
11-1 Projected Savings by Program ................................................................ 229

11-2 Survey Concern Level for Demand Hardening....................................... 233
11-3 Distribution of Customers by Water Use Under Uniform
and Increasing Block Rates, Alameda County, California.............. 242
11-4 Conservation Programs, Los Angeles and San Diego, 1990–1991 .... 243
11-5 Service Providers and Programs, San Francisco Area .......................... 245
12-1 Descriptive Statistics............................................................................... 262

12-2 Correlation Matrix for Residential Microdata—Raw and
Log-Transformed Versions ............................................................... 263
12-3 Multiple Regression Results ................................................................... 264
13-1 Distribution of Customers by Water Use Under Uniform

and Increasing Block Rates ............................................................. 292
13-2 Price Elasticity Estimates for Discrete-Continuous
Choice Model .................................................................................... 293
B-1 WDS Participants by US State or Canadian Province............................ 322
xii
20410-A Forecasting.book Page xiii Tuesday, January 29, 2008 3:48 PM
INTRODUCTION
AWWA and the authors are pleased to offer a this second edition of Forecasting Urban
Water Demand. Extensively rewritten and expanded, this new edition reflects many of
the dramatic changes in urban water-demand research and practice since the first
edition came out in 1996.
This book also presents findings from the Water-Demand Survey—a polling of
more than 600 US water utility managers on their forecast practices, conservation
programs, and current and anticipated trends in per capita water use in their systems.
In addition, data used to develop many of the tables in the book are included in the
accompanying data disk, along with other information collected from contributing
utilities.
Water conservation is an important focus for many state and local governments
in the United States. This led to creating two new chapters to deal with these issues;
chapter 10 addresses long-term conservation programs and technology, and chapter
11 focuses on emergency conservation or water restriction programs. Also, the
chapter on long-term water-demand forecasts (chapter 5) details the way per capita
and sectoral or per customer forecasts can be developed with a series of linked
examples.
This book also includes updated and expanded tables on price elasticities with
example calculations. The extensive statistical studies relating to price elasticities are
translated into easy-to-apply rules of thumb for calculating price impacts in the
water-demand forecast.
An expanded discussion in chapter 8 of weather and climate reports on the
growing research on climate change and its potential impact on water-demand
modeling. This new edition also includes new and expanded discussions of seasonal,
monthly, and daily water-demand forecasts.
As with the first edition, Forecasting Urban Water Demand, second edition,
examines the full range of influences on urban water demand—population, weather,
climate, water prices/rates, and short- and long-term conservation programs,
measures, and technology. This book is distinct in the industry for this scope and also
in its attention to the development and interpretation of statistical regression
procedures.
xiii
20410-A Forecasting.book Page xiv Tuesday, January 29, 2008 3:48 PM
This page intentionally blank.

20410-A Forecasting.book Page xv Tuesday, January 29, 2008 3:48 PM
ACKNOWLEDGMENTS
For discussions of forecasting issues and data, thanks go to many persons, including
Alvin Bautista, Claudia Borchert, Kees Corssmit, Paul Fesko, Greg Fischer, Tom Fox,
Chris Goemans, Professor Michael Haneman, Mary Hoddinot, Kurt Keeley, Jack C.
Kiefer, Jan Klein, Martina Kopat-Wright, George Kunkel, Ellen Levin, Warren Liebold,
Teresa Lutes, William Maddaus, Chandler Peter, Jerri Russell, Dan Schaefer, Esther
Siskind, Dave Visintainer, Gene Reetz, Amy Shallcross, and Bronwyn Weygandt.
These and other persons facilitated collection of data and reports from Calgary,
Alta.; Boulder, Colo.; Santa Fe, N.M.; New York City; Portland, Ore.; Los Angeles,
Calif.; Philadelphia, Pa.; and the Washington, D.C. area.
Special thanks for review and comments go to Donald E. Agthe, W. Mark Day,
Andrew G. Graham, Erik Hagen, Peter Mayer, Robert McGinness, G. Hossein
Parandvash, Ph.D., and Lorna Stickel.
The authors are particularly grateful for inspiration and encouragement from
Helmut Frank, Lester Taylor, Charles W. Howe, and Colin Murcray.
We also appreciate the many students and clients who have helped us develop a
deeper understanding of forecasting issues.
xv
20410-A Forecasting.book Page xvi Tuesday, January 29, 2008 3:48 PM

20410-A Forecasting.book Page 1 Tuesday, January 29, 2008 3:14 PM
Forecasting Urban Water Demand
CHAPTER 1
Introduction to Water
Forecasting
Urban water-demand forecasting is the process of making predictions about future

water use based on knowledge of historical water-use patterns. Forecasts help utility
officials answer questions such as the following:
• How much additional water supply and storage will be required to meet
future needs?
• How big must the treatment plant and water supply system be to
accommodate peak usage?
• What is the likelihood that water-use restrictions will be necessary to meet a
short-term supply shortage or emergency demand caused by unusual
weather?
• What are the likely effects of water-conservation programs and policies on
future water use?
• How will water usage and utility revenues change with increases in water
commodity charges and rates?
• What is the potential of, and limitations of, demand management in
responding to increases in water usage?
In this book, these problems are addressed from a comprehensive viewpoint,
suggesting methods of statistical analysis and forecast criteria, and providing
information about population and employment trends, economics, weather, climate,
and customer responses to water rates.
1
2 FORECASTING URBAN WATER DEMAND
Background data for this book1 was obtained from the 662 North American
public water supply systems that participated in the American Water Works
Association (AWWA) Water Demand Survey (WDS). The WDS is the most complete
information so far assembled on water-demand forecasting and key parameters, such
as peak- to average-day demand ratios or percentages of unaccounted-for water. The
responding utilities are geographically distributed across North America and repre-
sent the full range of utility sizes. Together, the participating companies and
organizations serve more than 80 million individuals in North America. Table 1-1 lists
the major types of water-demand forecasts and their prevalence, based on the WDS.
In general, only extremely small utilities did not develop water-demand forecasts.
Table 1–1 Types of Urban Water-Demand Forecasts Reported in the AWWA WDS
Percentage of Utilities
Reporting Forecast Type Forecast Type
65.4 Annual per capita water-demand forecasts

58.0 Annual water-demand forecasts by major customer
class (e.g., residential, commercial, industrial)
73.9 Peak-day forecasts
65.6 Monthly system water-demand forecasts
65.9 Daily water-demand forecasts
57.9 Revenue forecasts linked with water-demand
forecasts
Many water utilities rely on simple forecasting techniques, such as multiplying

water use per person by the projected population. These methods rely on the ability
of analysts to identify reasonable numbers for gallons per capita per day (or liters per
capita per day) and the accuracy of population forecasts that are typically produced by
other agencies.
Today, many larger utilities analyze water-use patterns by customer class
because additional detail yields better forecasts and more information for planning.
These more sophisticated methods focus on the dynamics of residential, commercial,
industrial, and public customers—issues that ultimately relate to the form and growth
of a community or region.
There are many books on forecasting and a growing list of studies of urban
water demand and its major components. For practitioners, Makridakis et al. (1998)
is an especially clearly written exposition. We also recommend several more recent
books, listed in the reference list to this chapter.
1. The WDS was sent to the utilities with e-mail addresses in the AWWA water utility database (the
database does not contain e-mail addresses for some utilities). Because the results reported are from
utilities that chose to respond, it does not represent a random sample of all utilities.
INTRODUCTION 3
Numerous research projects sponsored by the AWWA Research Foundation

(AwwaRF) report on key aspects of urban water demand. For instance, advances in
electronics and telecommunications support the use of data loggers at specific
residences or commercial sites. Data loggers can gather real-time information on water
use, which can then be analyzed and identified as specific water-using events (e.g.,
flushing toilets or running a load of laundry). Such research establishes important
benchmarks on end uses of water in residences and commercial establishments.
State agencies are also increasingly supporting water-demand research. The
Texas Water Development Commission (TWDC), for example, has a grant program
for analyzing community water demand. Recent completion reports describe statisti-
cal regression models based on water use for communities with utilities that serve
more than nine million individuals in Texas (Bell and Griffin 2006). In California, the
largest research database yet compiled with “microdata” on water use—billing records
of more than one million residential customers—is being put to work assessing the
potential impacts of global warming (Dale et al. 2005).
In this book, we provide an overview of findings of water-demand research and
consider how this growing stream of information can be integrated into water-
demand forecasts. The emphasis is on applications instead of theory, and the text is
supported by examples presented on the accompanying data compact disk (CD).
Technical terms are explained when first introduced. Many important statistical
and econometric terms also are defined in the glossary.
Forecasting Urban Water Demand is designed for managers, engineers, and
planners with forecasting responsibilities. It also should be valuable as a supplemental
exposition for water resource analysts; as an addition to community library holdings
on municipal planning issues; and as a window into applied work for civil
engineering, urban planning, and community design students.
This second edition is updated and substantially expanded. New materials cover
selecting a forecasting approach and forecasting water demand by customer class.
Seasonal water-demand models, short-term water-demand models, climate change,
emergency water conservation or water restriction measures, and risk simulation are
also covered. In addition, this book expands the discussion of consumer-rate
responses and short-term water conservation.
The subsequent discussion in this chapter focuses on the connotations and
meanings associated with the concept of water demand, the benefits of forecasting,
the major drivers of urban water usage, forecasting approaches, and the organization
of the book.
WHAT IS WATER DEMAND?

A public water supply delivers volumes of water to residents, businesses, and other
organizations in a community or urban area. These customers use the water,
returning some part of it back, usually to a water treatment facility that discharges its
effluent into rivers, streams, or other environments. The purpose of water-demand
forecasting is to make forward-looking information available to public water suppliers

as they conduct their business.
Accordingly, the basic meaning of water demand, in the context of a public
water supply, is the total volume of water necessary or needed to supply customers within
a certain period of time. In this sense, system water demand and total water production
in a public water supply are ultimately equivalent concepts. Total system water
demand includes an inevitable volume or component of water losses.2
Economics, which has contributed to urban water studies, loads “demand” with
more technical meanings. One of the most widely recognized definitions is “the
schedule or curve describing the amount the consumer is willing to purchase of a good
or service at various prices, other things being equal.” More generally, the “demand
function” is the mathematical relationship between the amount of goods a consumer
buys and the values and levels of various variables that determine this decision.
Figure 1-1 shows a standard picture of the demand curve. This curve traces how
much a consumer (or a group of consumers) is willing to pay for a commodity at
various prices. The curve in Figure 1-1 is downward-sloping, indicating that
consumers are willing to purchase more of a good as its price is reduced. This
relationship is generally known as the Law of Demand.
There is a caveat—namely that “all other things are [held] equal.” In other
words, other influences affect consumer demand beside price, particularly consumer
income, the prices of substitute and complementary goods, and environmental
parameters. Changes in these other demand influences can be considered to shift the
curve in Figure 1-1 in various ways.3
What we are really talking about here is a general functional relationship
between the quantity q of a good the consumer is willing to purchase and some
number, say k, of explanatory variables, x1, x2, …,xk. This relationship can be
symbolized by the mathematical function f(..) and summarized in mathematical
notation as
q = f ( x 1, x 2, …x k ) (Eq 1-1)
The existence of this function for a consumer or group of consumers—sometimes

called a market—is really a hypothesis: the demand hypothesis. Extensive research
2. For example, nearly 10 percent of the public water supply systems in the United States do not
meter customer water usage. These typically smaller systems instead charge a fixed fee per billing period
for connection to the water system. In such “unmetered” water systems, water production is the only
parameter that the utility can measure. In other systems, where each water connection has an associated
water meter that determines the bill based on the quantity of water sold, the purchase and sale of water
are approximately equivalent, as in over-the-counter transactions in a store. Here a utility can measure
“customer water demand” or usage and add up all these customer sales for a period of time to arrive at
total system water demand (net of the inevitable system water loss).
3. The curve in Figure 1-1 shifts to the right with, for example, increases in consumer income. In
other words, if consumers have more money, they tend to be willing to purchase more of a good at any
given price than they had been previously.
INTRODUCTION 5
20
18
16
14
Price ($) per Unit
12
10
0
40
60
80
0
10
12
14
16
18
20
22
Quantity
Figure 1–1 Schematic of Demand Curve
demonstrates its basic truth. Regular quantitative relationships exist between the
amount of a commodity that consumers purchase and a number of underlying
factors—the xi above.
So, in the space of a few sentences, the discussion has traversed from water
volumes to these mathematically worrisome expressions that allow for many factors
driving customer water use.
This conceptual tension cannot be completely resolved. The literature on urban
water demand and its components is extensive, and is contributed by various types of
researchers. Engineers and utility practitioners tend to emphasize total system
volumes and their flows. Social scientists, especially economists, focus on statistical
regressions that seek to capture the underlying functional relationships governing
water usage of some group of customers for a period of time.
The point is to develop a working synthesis of these perspectives—one of the
primary objectives of this book.
Why Forecast? Focusing on demand modeling and forecasting practices can
be useful and beneficial for many reasons, covered in the paragraphs that follow.
Capital investments associated with public water supply systems are extremely
expensive, costing millions—even hundreds of millions—of dollars. It thus behooves
utility management to make continuing comparisons between current conditions and
longer-term forecasts.
Although generally priced inexpensively, water is a prized resource, and its
allocation can be subject to intense conflict. Accurate water-demand forecasts provide
an opportunity to mitigate conflict and establish a credible basis for water use by
specific public water utilities.
Much urban and suburban growth is occurring in relatively arid regions of the
United States, leading to growing interest in urban water conservation. Water-demand
modeling and forecasting makes important contributions to policy analysis in this
regard. We cannot analyze the benefits and costs of water conservation without
considering the water-demand picture. Conservation programs, such as free distribu-
tion of water-saving kits, incentives for low-flow plumbing fixtures, landscape
regulations, and expanded leak detection have benefits and costs that are directly
related to the trajectory of water demand in a system.
More generally, water-demand forecasting contributes to system optimization.
Short- and medium-term forecast models, developed with an eye to accuracy, help
utility managers supply water to customers in a least-cost manner.
A spin-off of developing a water-demand forecast is a better understanding of
the dynamics and underlying factors that affect water use in the system.
A successful forecasting process serves management by providing organized
information about the past and its implications for the future. One benefit of formal
methods is that they encourage managers and analysts to examine the assumptions of
the forecast. The forecaster must collect and analyze historical data, thereby avoiding
errors based on unsupported beliefs about the past.
FORECAST HORIZON AND FORECAST APPLICATIONS

The forecasting literature often makes a distinction between short-, medium-, and
long-term forecasts. Typically, different forecast horizons are associated with different
uses or purposes, different types of forecast models, and different levels of reliability.
Table 1-2 shows major applications associated with water-demand forecasts
with these forecast horizons, including very-short-term forecasts. We examine each of
these four types more closely in following sections.
Long-Term Water-Demand Forecasts. Long-term water-demand forecasts

are developed for forecast horizons of a decade or more, and sometimes up to
30 years. The length of this forecast horizon is related to the time it takes to develop
new water plants and equipment. Storage options and long-distance water diversions
can require a decade or more to plan, permit, and bring to operational capability.
Such investments are capital-intensive. Capital-to-output ratios for potable water
typically are higher than for water used in agricultural or manufacturing operations.
Statistics, as well as research into the ex-post (after the fact)4 performance of
forecasts, suggest that forecast errors typically increase with the length of the forecast
period.
4. Forecasts evaluated ex-post can be compared with actual numbers.

INTRODUCTION 7
Table 1–2 Types of Water-Demand Forecasts and Major Applications
Forecast Type Forecast Horizon Applications
Long-Term Decades, 10–50 years Sizing system capacity, raw

water supply
Medium-Term Years to a decade, 7–10 years Sizing, staging treatment

and distribution system
improvements, investments,
setting water rates
Short-Term Years, 1–2 years Budgeting, program tracking

and evaluation, revenue forecasting
Very-Short-Term Hours, days, weeks, Optimizing, managing system

up to two weeks operations, pumping
As the forecast horizon extends, growing forecast errors pose two-sided risks to
water utilities. If the utility builds facilities that are too large, a smaller-than-planned
number of customers will bear the capital costs for excess capacity. This can be
burdensome, meaning higher fees, charges, and rates. On the other hand, if long-
range capacity issues are inadequately addressed, there can be water shortages. Water
shortages also impose costs in the form of losses in landscaping, convenience of water
use, and constraints on new construction.
In the years after World War II, water-supply risks were usually managed by
opting for the largest facility possible. As long as economic and population growth
was the general rule, this tactic worked well. With geographical shifts in population
and employment growth and expansion of suburbs and “ex-urbs” at the expense of
core metro areas, however, simple extrapolation of population trends from past rates
can be problematic.
In this context, many analysts suggest linking water-demand forecasts and their
potential errors with the types and sizes of projects selected for future development.
Smaller, more modular projects—potentially with higher unit costs, but lower overall
capital costs—can offer flexibility to the water agency.
One focus, therefore, is on assessing potential forecast errors over the planning
horizon. Such information can support financial and other types of risk analysis in
infrastructure planning.
Medium-Term Water-Demand Forecasts. These forecasts (for one to ten
years) are commonly developed for planning improvements to the water distribution
and treatment system and for setting water rates. Medium-term forecasts focus on the
variability of water consumption by a fixed or slowly changing customer base.
Variations in demand over this time period are driven by weather cycles, changes in
the composition of the customer base, and the business cycle.
Short-Term Water-Demand Forecasts. These forecasts support water sys-
tem operations, as well as budgeting and financial management. Forecast horizons are
commonly a year or so. Detail over periods as short as a month or even a day can be
relevant. Forecast errors in the short term arise primarily from the inherent variability
and unpredictability of the weather and human behavior. These effects can be
documented and effectively analyzed by short-term forecasting methods.
Very-Short-Term Water-Demand Forecasts. These forecast models, focus-

ing on horizons of a few days or weeks, are now being developed with relatively
advanced forecasting techniques. These methods often use weather forecasts of a few
days to a few weeks to develop forecasts of water use aimed at optimizing day-to-day
utility operations. The forecast models can also be helpful in planning facility
maintenance schedules.
Related Applications. Special studies, sponsored by utilities or conducted

by the research community, support the accuracy of forecasting over a variety of
forecast horizons. These related applications generally fall into two categories: (1)
studies of key water system parameters, and (2) demand monitoring.
Studies of key water system parameters. The responsiveness of water customers
to water rates is, to a degree, system-specific and a good candidate for special on-site
research. Many public water suppliers also study the effectiveness of water-conserva-
tion programs, comparing end-use estimates and econometric estimates of water
savings. Another example is analyzing data on individual household water usage to
determine whether newer homes in a community use more or less water than older
residences—an issue in assessing future residential-customer use rates. Water savings
from converting from a flat rate to universal metering may also be an issue of
sufficient importance to warrant specialized research with household microdata.
Demand monitoring. Evaluating the accuracy of forecasts is especially challeng-
ing in the water field because of the importance and unpredictability of the weather.
More extensive data-collection methods and technologies in urban water systems, as
well as developments in statistical techniques, make real-time demand monitoring
increasingly feasible.
Answers to a number of questions may offer insight into forecast adequacy. Are
regulatory requirements being met, or are current high (or low) levels of water usage
a fluke of the weather or other factors, such as changes in local business conditions? Is
the water system on course for meeting planning targets for conservation, per capita
usage, or other designated quantitative goals and objectives? What is the probability
that peak-day demands will exceed current treatment capacity this year?
Demand-monitoring models address these types of questions, often drawing on
advanced modeling techniques to generate answers. The performance of forecasting
models over any forecasting horizon is also a statistical issue, suitable for special
research. Such performance evaluations are essential to improving forecasting models.
INTRODUCTION 9
MAJOR INFLUENCES ON WATER DEMAND

A number of factors have a significant impact on water demand, including
population, employment, economic cycles, technology, weather and climate, price,
and conservation programs. We discuss these and other factors in the following
sections.
Population, Employment, and Technology. Population growth is often
the major trend factor in water use. In the United States, the Southwest and Florida
continue to be magnets for migration. Conversely, many of the older urban areas of
the nation are losing population and employment, implying reduced needs for water
in the future.
Business-cycle factors affect water use because fluctuations in industrial and
commercial production translate into commensurate changes in water demand.
Changes in personal income also influence water use because, in the jargon of
economics, water is a “normal” good. Water consumption increases, other things
being equal, when family income rises. Consumption is likely to drop if, for example,
the primary breadwinner in the household is out of work.
Technological change can also affect water use over time. For example,
widespread installation of garbage disposals and automatic dishwashers in homes may
increase domestic water use. Adoption of drip irrigation, on the other hand, may
reduce water use. New industrial production methods may use either far more or far
less water than previous methods, while requirements and opportunities for water
reuse may dramatically cut water requirements.
Weather and Climate. Seasonal components in water use are generated
primarily by the local climate. Summer peaking demand is typical. Higher summer
demand levels are related to water use for outdoor activities, including lawn watering
and gardening, and to the use of evaporative coolers. Some cold climates have winter-
peaking systems, in which water flow is used to prevent the freezing of pipes on the
coldest days. Seasonal demand patterns are important in planning the capacity of
water treatment and distribution systems. Short-term patterns are also critical for
scheduling maintenance times for reservoirs, pumps, and mains.
Price. Price effects are important for short-, medium-, and long-term
forecasts. Both water use and utility revenue are directly affected by water rate
changes. In the short term of a few months, rate hikes can cause consumers to change
their behavior. These changes can include taking shorter showers or reducing car
washing and lawn watering. In the longer term, if a noticeable rate hike keeps pace
with inflation, consumers may adapt through their selection of water-using durable
goods, favoring appliances with lower water-use ratings and possibly innovative
landscaping designed to cut back on water use.
Efficiency and Conservation Programs. Water efficiency and conserva-
tion programs typically couple an appeal to civic virtue with information on how to
use less water. Crisis programs resulting from drought or other supply interruptions
have generated large, albeit temporary, reductions in water use. Programs designed to
permanently change individual behavior are capable of generating long-term reduc-
tions in water usage. Particularly effective are programs that encourage structural
changes, such as the use of low-flow toilets and low-water use vegetation. Conserva-
tion program effects must be thoughtfully included in the water-demand forecasting
model to minimize errors in projected water use and revenue. Regardless of the time
frame of the forecast, many conservation programs will cause a reduction in water use
and a drop in revenue. In contrast, a conservation program that relies heavily on rate
increases will cause revenue to rise while reducing water use.
Other Factors. Physical deterioration in the water distribution system is

often overlooked as a factor influencing water-production requirements. Degradation
of the water distribution infrastructure is likely to result in increased losses from leaky
and broken pipes. The size of these losses in an older system can be hard to estimate,
however, because certain types of water meters may underregister as they age. As a
result, some unaccounted-for water may be delivered to customers instead of being
lost. End-user water losses may also tend to increase with the age of an urban area.
Including estimates of future lost water in a complete water-use model may therefore
be desirable. Projection of water losses helps to provide a rational basis for
replacement and maintenance policies for pipelines, valves, and meters.
WATER-DEMAND FORECASTING
A wide range of methods is used for forecasting. The method(s) a particular utility
chooses depends on the technical sophistication of the analyst, the resources devoted
to the forecasting process, and the available data. Generally, the quality and
sophistication of forecasts improve as the utility gains experience and the commit-
ment of utility management to the process increases. Forecasting is useful even if the
accuracy of long-term forecast models is likely to be low. A conceptually valid water-
demand model aids in considering the impact of changes in such variables as
population, industrial production, climatic conditions, price, and water-conservation
measures.
Many urban water-forecasting approaches use a statistical appraisal of per
capita (or per customer) water-use rates. Projected water use per capita (or per
customer) is multiplied by the population forecast to obtain a forecast of total water
use. The greatest differences among forecasting methods involve how the per capita
forecast is derived and the degree of disaggregation by customer type.
Per Capita and Other Unit Use Coefficient Approaches. The AWWA WDS
infers that 65 percent of North American water utilities develop per capita water-
demand forecasts.
Larger urban water systems tend to develop sectoral demand forecasts on a per-
customer basis, calculating unit water-use coefficients for customers by categories
such as residential, commercial, industrial, and public. A variant of the unit-use
INTRODUCTION 11
coefficient approach calibrates the demand forecast to the land-use plan in the utility
service area. Residential, commercial, and industrial land uses are estimated to
consume certain amounts of water per acre per year. The effectiveness of water-
demand forecasts based on land use is greatest in those areas with strict land-use
regulations, comprehensive land-use planning, and a stable industrial structure.
End-Use Models. End-use model forecasts use extensive, detailed informa-

tion about customer behavior. They build up a forecast from data on the inventory of
water-using appliances and fixtures and typical patterns of behavior. For example, the
number of toilet flushes for a typical family size may be estimated from minute-by-
minute logging of water use in sample households (refer to Table 2-2). The end-use
approach is also applicable to industrial water demand, where engineering techniques
such as linear programming can be applied to investment decisions about water
recycling equipment. IWR-MAIN and the DSS Model, discussed in chapter 3, contain
software templates for water-demand forecasts based on end-use estimates for various
types of water customers.
Time-Series Models. Time-series extrapolation encompasses a variety of

techniques, including simple time trends, exponential smoothing, and Box-Jenkins
(autoregressive integrated moving-average) models to project historical water-use
trends into the future. These models rely on the assertion that future changes in water
use can be predicted based on historical changes in water use (ignoring all other
possible influences). These methods can provide reasonably accurate forecasts as long
as the future is essentially similar to the past. Accuracy is improved by disaggregating
water use by customer class and, if appropriate, through geographic considerations.
Because future trends are unlikely to repeat past trends, extrapolation methods are of
limited value for long-range forecasting.
Regression Models. Since the 1960s, economists and social scientists have
criticized per capita and unit-use coefficient forecasting methods for ignoring
socioeconomic factors. In particular, such fixed coefficients do not reflect the level of
water rates (i.e., price effects), which is problematic because water usage responds to
price changes and real (inflation-adjusted) water rates do increase from time to time.
If trends in water prices, personal income, ownership of water-using appliances,
population, urban density, and other factors are to be used together in a forecasting
model, regression modeling is appropriate.
Other Structural Forecast Models. Many hybrid types of forecast models

incorporate, for example, both time-series and structural regression methods. In
addition, nonparametric forecasting models apply neural networks or fuzzy logic
systems.
ORGANIZATION OF THIS BOOK

The details on all the remaining chapters are in the subsections that follow.
Chapter 2—Basic Facts About Urban Water Demand. This chapter sum-
marizes facts about major water-demand components, such as residential demand
and commercial/industrial demand. This information is presented with data on
national per capita water-use rates, charts illustrating the allocation of water use by
customer grouping, and discussions of regional variations in indoor and outdoor
water use by season.
Recent studies (cited elsewhere in this book) give us a better understanding of
indoor water use by residential customers—for both single-family residences and
apartment dwellers (multiple-family residences). Other, more finely calibrated
studies, conducted under AwwaRF auspices, present new information on water use by
commercial and institutional customers in several categories. The chapter also
contains an analysis of the conservation effectiveness of low-flow toilets, faucets,
showerheads, and conservation water rates. The chapter summarizes major findings
in these areas, directing readers to sources and research for more exhaustive
information.
Chapter 2 also considers urban growth and development, along with system
leakage detection and estimation.
Chapter 3—Choosing a Forecasting Approach. In this chapter, we in-

ventory major water-demand forecasting approaches, highlighting the strengths of the
various modeling approaches in the context of factors, such as utility size, forecasting
application purpose, and data availability. The discussion is couched in terms of
generic features to keep the evaluation pertinent to the changing scene of modeling
products and approaches.
Chapter 4—Data and Data Structures for Water-Demand Forecasting.

This chapter focuses on three major topics: (1) the visual structure of usefully
organized water data in spreadsheets; (2) data-collection and data-quality trouble-
shooting; and (3) sources of data on water-use drivers (socioeconomic, demographic,
climatic).
Times-series on water sales and production can highlight basic trends (annual
data), seasonal variation (monthly data), and peak-to-average water-use relationships
(daily data). Customer billings usually comprise the most readily available cross-
sectional data. Data collection presents a surprising number of issues. Questions arise
about how to treat missing values, how to identify and scrub outliers, and how much
disaggregation to seek in terms of data frequency and/or drill-down to the customer
level. We suggest that data collection should be accompanied by graphs and
descriptive statistics, which often reveal features not apparent in a spreadsheet format.
Chapter 5—Simple Forecasting Methods and Reality Checks. This chap-

ter focuses on two basic types of water-demand models: per capita and sectoral
forecasts. The chapter develops discussions of these two major long-term forecasting
approaches with coordinated data from an example water system.
INTRODUCTION 13
Chapter 6—Forecasting Seasonal and Peak Water Demands.

Seasonal and peak water demands drive treatment capacity and are a major factor in
utility capital costs. This chapter focuses on seasonal or monthly water demand, daily
water-demand models, and estimates of the peak-day to average-day demand ratio.
Topics include averaging methods, such as simple curve extrapolation and modeling
approaches using multiple-regression with seasonal dummy variables.
Chapter 7—Population, Employment, and Technology Forecasts. I n
this chapter, we review facts and guidelines for forecasts relating to population and
employment, in addition to the technology of water consumption. The focus is on
trends and developments, such as the movement of the “center” of the US population
west and south, the impacts of immigration, and the fastest growing states and
metropolitan areas. Methods of forecasting regional and local populations are also
considered.
The chapter concludes with checklists that offer suggestions about how to
evaluate population, employment, and regional economic forecasts.
Chapter 8—Weather and Climate. The weather section of this chapter
discusses weather patterns and their variability, weather variables, sources of weather
data, statistical analysis of weather data, simulation of weather data, and weather
forecasts. We also consider extreme weather events, such as droughts, heat waves, and
weather cycles.
Next we present a major section focusing on climate. The discussion considers
some of the basic science behind the projections of warming, global climate models,
downscaling (localizing) global climate models, and integrating climate-change
scenarios into long-range planning.
Chapter 9—Price Effects. In this chapter, rules of thumb are provided for
including price effects in the demand forecast, as well as an overview of common
types of water rate and fee schedules. The discussion considers (1) impacts of water
metering; (2) controlled comparisons of communities with similar socioeconomic
and weather conditions but different levels of water rates; and (3) econometric studies
of water demand. Alternative features of utility billing schedules include monthly
service fees, constant volumetric fees, increasing and decreasing block rates, lifeline
rates, and sewer-use fees based on water volume. The price elasticity of demand and
price elasticity of revenue are carefully explained, with examples showing the effects of
price changes on utility sales and revenues.
Chapter 10—Long-Term Water Conservation. This chapter discusses a
variety of conservation initiatives, with engineering estimates of their potential for
saving water and analysis of the conservation experience in several utility districts.
The chapter lists methods of conservation by customer type (residential, multifamily
residential, commercial, industrial) and evaluates incentives, such as future cost
savings, civic-mindedness, and higher marginal prices for water and sewer service. We
also discuss revenue effects and requirements, which can be a challenge in terms of
avoiding backlash from conservation-induced rate increases.
The conservation potential of shifting to increasing block rates is presented with

a numerical example. In addition, we cover methods of estimating water-conservation
effects with local studies using aggregate before-and-after analysis, control group
experiments, and statistical modeling.
Chapter 11—Short-Term Water Conservation. In this chapter, useful

short-term conservation programs are reviewed and their track record summarized as
gleaned from case studies. Incentives for conservation, such as publicity, penalty water
rates, and direct regulations, are discussed and evaluated. Major features of this
chapter are a simulation of the effect of various conservation programs on water
demand, and a discussion of incorporating these effects into short- to medium-term
demand forecasts. We also consider methods for evaluating programs and incorporat-
ing evaluation instruments into a conservation program.
Chapter 12—Forecasting With Regression. This chapter provides an in-

depth discussion of forecasting with regression, covering the basics of estimating and
interpreting regression equations and illustrating key points with an example
regression based on household microdata. The material in this chapter is specially
designed to support the use of regression methods for special studies (e.g.,
determining whether new homes in an area consume, on average, more or less water
than older residences).
Chapter 13—Advanced Topics and Methods. This chapter presents con-

ceptual-level discussions of time-series analysis, neural networks, recent developments
in the econometric analysis of consumer responses to water rates, and risk simulation
of water-demand forecasts.
Chapter 14—Forecast Uses, Evaluation, and Improvement. In this chap-

ter, we discuss (1) error measures and loss functions, (2) comparing prior forecasts
with actual data to improve future results, (3) backcasting and withholding data as
evaluation techniques, and (4) other strategies to improve accuracy. We also consider
the motivations for formal forecasts, compare formal statistical forecasts with
judgment forecasts, and outline the major uses of forecasts.
Appendix A—Survey Methods and Direct Data Collection. The appen-

dix introduces and reviews methods of direct data collection, including customer
feedback acquired from ordinary utility interactions with customers, from random
sampling, and from Delphi methods. We include examples of mail and telephone
survey approaches.
Appendix B—AWWA WDS Findings. The appendix reports findings from

the Water Demand Survey. The appendix reports on the types of water-demand
forecasts that are being developed and the general classes of water conservation
programs that are being implemented. The WDS also profiles peak and average water
demands and utility expectations about future growth of their systems.
INTRODUCTION 15
Glossary. The glossary lists key words along with their definitions as applied
to water-demand forecasting.
Data CD. We have included data for the examples covered in the text on a CD
so the reader can perform and experiment with the procedures described in the book.
The data sets and spreadsheets are organized by chapter. The files on the CD are in
two forms: raw data files and Microsoft Excel (Microsoft, Redmond, Wash.)
spreadsheets. To make full use of the examples in the data CD, the Microsoft Excel
Analysis Toolpak of statistical and engineering routines must be installed.5 To
conclude this chapter, we offer poetry fans this ode composed by the late Kenneth
Boulding, perhaps most famous for his essay “The Economics of Spaceship Earth,”
and the author of many books, tracts, and articles in economics.
Ode, On the General Subject of Water

by Kenneth E. Boulding; Feather River Anthology
Water is far from a simple commodity,

Water’s a sociological oddity,
Water’s a pasture for science to forage in,
Water’s a mark of our dubious origin,
Water’s a link with a distant futurity,
Water’s a symbol of ritual purity.
Water is politics, Water’s religion,
Water is just about anyone’s pigeon.
Water is frightening, water’s endearing,
Water’s a lot more than mere engineering.
Water is tragical, water is comical,
Water is far from Pure Economical,
So studies of water, though free from aridity
Are apt to produce a good deal of turbidity.
5. With this add-on (available online and at computer stores), and the ability to write custom rou-
tines called macros in Visual Basic, Excel can act as a powerful statistical tool.
REFERENCES
American Water Works Association. 1993. Evaluating Urban Water Conservation Programs: A Procedures
Manual. Denver, Colo.: AWWA.
Armstrong, J.S. 1985. Long-Range Forecasting: From Crystal Ball to Computer. New York: Wiley
Interscience.
Bails, D.G., and L.C. Peppers. 1993. Business Fluctuations: Forecasting Techniques and Applications. 2nd
ed. Englewood Cliffs, N.J.: Prentice Hall.
Bell, D.R., and R.C. Griffin. 2006. Community Water Demand in Texas as a Century is Turned. Natural
Resource and Environmental Economics Working Group. College Station: Texas A&M University
Department of Agricultural Economics.
Boland, J.J., J.L. Pacey, W.S. Moy, and R.C. Steiner. 1983. Forecasting Municipal and Industrial Water Use:
A Handbook of Methods. Contract Report 83C-01. Alexandria, Va.: US Army Corps of Engineers,
Engineer Institute for Water Resources.
Boulding, K.E. The Feather River Anthology, stanza III. Industrial Water Engineering 3(12): 32–33.
Dale, L., M. Hanemann, C. Dyckman, and C. Nauges. 2005. Economic Impacts of Climate Change on
Urban Water Use in California. PIER Energy-Related Environmental Research. CEC-500-2005-124.
Sacramento: California Energy Commission.
Grima, A.P. 1972. Residential Water Demand: Alternative Choices for Management. Toronto, Ont.:
University of Toronto Press.
Haneman, W.M. 1998. Determinants of Urban Water Use. In Urban Water Demand Management and
Planning. Edited by D.D. Bauman, J.J. Boland, and W.M. Hanneman. New York: McGraw-Hill.
Jones, C.V., J.J. Boland, J.E. Crews, C.F. DeKay, and J.R. Morris. 1984. Municipal Water Demand:
Statistical and Management Issues, Studies in Water Policy and Management No. 4. Boulder,
Colo.: Westview Press.
Kennedy, P. 2003. A Guide to Econometrics. 5th ed. Cambridge, Mass.: MIT Press.
Kindler, J., and C.S. Russell in collaboration with B.T. Bower, I. Gouevsky, D.R. Maidment, and W.R.D.
Sewell. 1984. Modeling Water Demands. London: Academic Press.
Linaweaver, F.P. Jr., J.C. Geyer, and J.B. Wolff. 1966. Residential Water Use Project–Final and Summary
Report, Report V, Phase Two. Baltimore, Md.: Johns Hopkins University Press.
Makridakis, S., S.C. Wheelwright, and R.J. Hyndman. 1998. Forecasting: Methods and Applications. New
York: John Wiley & Sons.
Pindyck, R.S., and D.L. Rubinfeld. 2000. Econometric Models and Economic Forecasts. 4th ed. New York:
McGraw-Hill.
US Army Corps of Engineers. 1981. An Assessment of Municipal and Industrial Water Use Forecasting
Approaches: Water Conservation and Supply Information Transfer and Analysis Program.
Contract Report 81-005. Alexandria, Va.: US Army Corps of Engineers, Engineer Institute for
Water Resources.
CHAPTER 2
Basic Facts About

Urban Water Demand
Public water supplies1 served an estimated 242 million people (approximately

85 percent of the population) in the United States in 2000. These public water systems
utilized about 11 percent of the 408,000 million gallons per day (mgd) of total US
water withdrawals, or an estimated 43,300 mgd. By comparison, the other 15 percent
of the population relies on self-supplied domestic use—largely wells—which are
estimated to account for less than 1 percent of total water withdrawals in 2000. These
public water supplies include hundreds of larger water systems, several multicounty
metropolitan area utilities, and thousands of smaller water utilities.
In this chapter, we consider trends and geographic patterns in per capita water
use and present basic facts about the water use of major customer groups in urban
water systems. The next section discusses how water-use data can be manipulated to
profile important features of the water system, such as longer-term trends, seasonal
patterns, and the typical distribution of large and small users in a given period.
Subsequently, we also report on the prevalence of metering in public water supplies
and reported levels of unaccounted-for water.
1. A public water supply system refers to public and private water suppliers serving at least 25 peo-
ple or with a minimum of 15 connections.
17
TRENDS AND PATTERNS IN PER CAPITA WATER USE

The US map in Figure 2-1 illustrates per capita water use in public water supplies in
2000. Texas, Arizona, and California are among the higher per capita water users.
Many states east of the Mississippi have lower per capita water usage. Figure 2-1 shows
some surprises, such as states with higher use codes where urban water use had not
been commonly perceived as intensive.
WA
MT ME
ND
OR MN VT
WI NY NH
ID SD
WY MI MA
PA RI
NE IA NJ CT
NV
UT IL IN OH DE
CO WV VA MD
CA KS MO DC
KY
NC
TN
AZ NM OK SC
AR
PR
MS AL GA
TX LA
HI VI
FL
AK
Consumption Intensity,
(gpcd)
Less than 130–149 150–169 170–189 190–219 More than

130 220
Source: Hutson et al. 2004.
Figure 2–1 Per Capita Water Usage in the United States in 2000
Based on US Geological Survey (USGS) research, per capita water use in US

public water supplies stabilized at around 179 gallons per capita per day (gpcd) in
1995 and 2000, after dropping from around 183 gpcd in 1985 and 1990. The net
decrease in per capita water use from 1985 to the present follows earlier increases
BASIC FACTS ABOUT URBAN WATER DEMAND 19
recorded in the five-year surveys of water use that the USGS has conducted since
1950.
Table 2-1 outlines the underlying data. Note that the population served by
public water supply systems rose from approximately 82 to 85 percent of the total US
population over the 15 years from 1985 to 2000. In 1950, the percentage of the
population served by public water supply systems was only around 62 percent.
Table 2–1 Water Withdrawals for Public Water Supplies and per Capita Water Use
in the United States
Year
1985 1990 1995 2000
Population (millions) 242.4 252.3 267.1 281.4
Population Served
by Public Water
Supplies (millions) 199.5 210 225 242
Total Withdrawals (mgd) 399,000 408,000 402,000 408,000
Public Supply (mgd) 36,500 38,500 40,200 43,300
Per Capita Water Use

in Public Water
Supplies (gpcd) 183.0 183.3 178.7 178.9
The AWWA Water Demand Survey (WDS; see appendix B) confirms the
stabilizing trend in per capita water use.2 Some 34.6 percent of the utilities responding
to the WDS reported increasing per capita water demand in recent years. On the other
hand, 23.9 percent reported recent decreases in per capita demand. Many systems
reporting lower per capita water demand are core metropolitan areas located in the
Northeast and the Midwest. The remaining 9.3 percent of the 662 utilities indicated
that recent per capita demands were level or showed no clear trend.
Many factors contribute to changes in per capita water use, including
geographic shifts in population, migration from central metropolitan areas to the
suburbs, and technological change. Historically, lavish water use has been associated
with wealth. As living standards rose after World War II, owning a house with a
variety of water-using appliances and a lawn became a middle-class goal, causing per
capita water use to surge (Solley et al. 1998). Water withdrawals follow a similar
pattern. Based on the steady increase from 1950 to 1980:
The expectation was that as population increased, so would water use.
Contrary to expectation, reported water withdrawals declined in 1985 and
2. For details on ways to design and conduct a survey, see appendix A. Appendix B contains details
about and results of the AWWA WDS.
have remained relatively stable since then. Changes in technology, in state

and federal laws, and in economic factors, along with increased awareness of
the need for water conservation, have resulted in more efficient use of the
water from the nation’s rivers, lakes, reservoirs, and aquifers.
(Statement of Robert M. Hirsch, USGS associate director for water, in
Hutson et al. 2004, iii)
STATISTICAL PROFILES OF URBAN WATER USE

Water-use data can be presented in several different ways, providing statistical profiles
of an urban water system and its major components. There is, for example, the long
view—total system water use over several years. In many water systems, graphing a
20- to 30-year series shows an increasing trend in total use as population increases
and per capita use expands. In other systems, recent declines in per capita use,
perhaps resulting from plumbing codes, may offset increasing population, resulting in
a flattening or even a declining trend in more recent years.
As an example, Figure 2-2 shows total monthly system water production for a
major North American city (Calgary, Alberta) with data dating back to the late 1960s.
This chart shows water production stabilizing in recent years, despite continuing and
sometimes rapid increases in population of this urban area.
200
180
160
140
Billion Liters
120
100
80
60
40
20
69
71
73
75
77
79
81
83
85
87
89
91
93
95
97
99
01
03
19
19
19
19
19
19
19
19
19
19
19
19
19
19
19
19
20
20
Year
Source: Fesko 2006.
Figure 2–2 Total System Water Use, Annual Data

800
700
600
Million Liters
500
400
300
200
100
0
19 0:1
19 :9
19 :7
19 :5
19 :3
19 1:1
19 :9
19 :7
19 :5
19 :3
19 :9
19 :7
19 2:1
20 :5
20 :3
20 3:1
19 1
19 1
19 1
1
:1
:1
:1
:1
73
75
77
79
84
86
88
90
95
97
99
01
7
0
71
82
93
04
19
Year and Month (January [1], October [10], July [7], and April [4])
Source: Fesko 2006.
Figure 2–3 Total System Water Use, Average Daily Water Use by Month
The seasonal pattern is another important perspective. This, again, can be

expressed with total system water use data, but in a monthly or higher-frequency
format. Typically, water use in public water supply systems exhibits seasonality,
usually peaking during the summer because of demand for lawns, gardens, and other
outdoor uses. Figure 2-3 illustrates the monthly detail for Calgary’s total system water
production dating back to the late 1960s. Note the volatility in peak month water use
in this system.
Peak demands are another important feature of water system demand, usually
measured as a ratio between peak- and average-day demand. Figure 2-4 shows the
ratio between peak- and average-day demand for a Midwestern community of
approximately 300,000 residents between 1960 and 2003. In this system, an upper and
lower bound effectively contain the peak- to average-day demand ratio over this
period of more than 40 years.
MAJOR COMPONENTS OF WATER SYSTEM DEMAND

Most public water supply systems serve multiple types of customers including single-
family residences, multifamily residences, industrial establishments, commercial
businesses, institutional customers, and public parks and recreational facilities.
Inevitably, there also are “water losses” and unaccounted-for water or differences
between water production and billed usage.
1.9
1.8
Peak-to-Average-Day Demand Ratio
1.7
1.6
1.5
1.4
1.3
1.2
1.1
1
1960
1962
1964
1966
1968
1970
1972
1974
1976
1978
1980
1982
1984
1986
1988
1990
1992
1994
1996
1998
2000
2002
Year
Source: Confidential.
Figure 2–4 Peak-to-Average Day Pumpage Ratios
10,000
9,000
8,000
7,000
Million Gallons
6,000
5,000
4,000
3,000
2,000
1,000
0
1980:1 1981:7 1983:1 1984:7 1986:1 1987:7 1989:1 1990:7 1992:1 1993:7 1995:1
Year and Month (January [1] or July [7])
Source: Michelson et al. 1998.
Figure 2–5 Monthly Residential Water Use, City of Los Angeles, 1980–1995
Residential Water Demand. In many communities across North America,

residential customers use 50 to 60 percent of total water production or sales. The
proportion varies by the size of the water system. In larger urban areas, greater
proportions of total water are devoted to commercial, industrial, and public sector
uses.
Usually, single-family residential customers are major contributors to water
system peaking patterns. Figure 2-5 illustrates summer-peaking seasonal variation
with monthly data on total residential water use for Los Angeles, Calif., from 1980 to
1995. Notice the strong seasonal pattern. This data series is also notable because it
encompasses a severe drought, which saw reductions in total use as well as
suppression of peak demands.
Based on an AWWA database system from the 1990s (AWWA 1991), annual
residential water purchases ranged from a low of 70,500 gal per household per year in
a region encompassing the District of Columbia, Delaware, Maryland, Pennsylvania,
Virginia, and West Virginia to a high of 169,800 gal per household per year in a region
comprising Arizona, California, Hawaii, and Nevada.
Within a public water supply, residential water use can vary widely by customer.
Figure 2-6 shows, for example, residential customer water use for the January 1996
billing period for a sample of residential customers from Boulder, Colo.
100
90
80
70
Customer Count
60
50
40
30
20
10
0
0 25 50 75 100 125 150 175 200 225 250 275 300 325 350 375 400
Gallons per Customer per Day
Source: Personal communication from Peter Mayer and William DeOreo,
Aquacraft, Boulder, Colo., July 2006.
Figure 2–6 Distribution of Water Use by Residential Customers

Figure 2-6 illustrates a pattern we see in many contexts when we examine a

cross section of billing data for a particular customer group. The figure suggests an
underlying lognormal distribution.3 The average January water use for this group is
approximately 134 gallons per customer per day. The distribution of water use for
this sample of customers, then, is not symmetric but skewed with a long tail to the
right. Another way of saying this is that the average quantity is greater than the
median quantity in this group of customers. There are a few large users of water and
many more customers consuming smaller quantities of water.
Indoor and Outdoor Use. Many temperate zones in North America have
distinct winter-season months in which maximum temperatures are close to or
below freezing. In these areas, indoor water use can be identified with water-use
levels during the months of lowest water use. When we subtract these minimum
levels from monthly water sales through the year, we arrive at estimates of water use
for outdoor purposes (see example in chapter 6).
In many water systems, outdoor water use is largely from residential
customers. This can have implications for utility rates and revenues. Residential
water use typically is more responsive to changes in water commodity charges, or
rates, than other customer categories. Furthermore, outdoor residential water use is
more responsive to rate changes than indoor water usage.
Studies conducted under the auspices of the AWWA Research Foundation
(AwwaRF) offer important insights into residential indoor water use and are widely
cited and utilized in water utility planning.
The breakdown of indoor water use by end use in the study cited in Table 2-
2 is especially significant. The table summarizes the percentage of indoor water use
going to the major indoor residential end uses of water.
The AwwaRF Residential End Uses of Water study (Mayer 1999) gathered data
on approximately 1,000 single-family detached residential customers in 12 public
water systems. With electronic data loggers, minute-by-minute water flow data can
be analyzed into discrete end-use events, such as a resident showering or flushing a
toilet.
Average annual water use in the sample varied from 69,900 to 301,100 gal.
Indoor water use averages 69.3 gpcd, distributed across the end uses shown in Table
2-2. Residential indoor water use per customer is much less variable than residential
outdoor water use per customer.
The Mayer (1999) analysis shows that household size is a key factor for many
indoor end uses of water. Household size materially influences the amount of water
for shower and bath use, toilet flushing, dishwasher use, and clothes washer use.
3. The lognormal distribution is transformed into the bell-shaped curve or normal distribution by
a logarithmic transformation. As a result, if the quantities qi are lognormally distributed, the quantities
log(qi) are distributed according to a normal distribution.
Table 2–2 Indoor Residential End Uses of Water
End Use Percentage of Indoor Water Use
Toilets 27
Clothes washers 22
Showers 17
Faucets 16
Leakage 14
Other domestic 2
Baths 2
Dishwashers 1
Total indoor* 100
Source: Mayer et al. 1999.

*Adds to 101% because of rounding error.
Interestingly, economies of scale emerge with respect to household size and

indoor water use. Water use goes up less than proportionally with the size of the
household.
Daily Pattern of Residential Water Use. The Mayer (1999) study also
casts light on the diurnal or daily pattern of water use over a 24-hour cycle of day
and night. Most households saw peaks in water use in the morning and in the late
afternoon or evening. The daily peak was usually in the morning, with nighttime
water use characteristically being the lowest.
Commercial and Industrial (CI) Water Demand. CI customers use ap-
proximately one-fourth the total quantity of water produced by public supply
systems (USGS 1995). Studies show that CI water customers are heterogeneous in
their water use among categories such as irrigation-intensive recreation (urban
irrigation), office buildings, schools and colleges, restaurants, hotels and motels,
laundries and laundromats, hospitals and medical offices, food stores, auto shops,
membership organizations, and car washes. Water is used for cooling, cleaning,
sanitation, and landscaping.
Other research has developed detailed comparisons of water use of establish-
ments in these categories, based on data from participating water providers.4 For
instance, using data loggers along with billing data from cooperating utilities,
Dziegielewski and colleagues (2000) present important information for water-
demand analysis and planning in their Commercial and Institutional End Uses of
Water (CIEW) study. Table 2-3 shows some key characteristics of these CI categories
4. The Los Angeles Department of Water and Power, the Irvine Ranch Water District, the City of
San Diego Water Utilities Department, and the City of Santa Monica in California; and the City of
Phoenix in Arizona.
in the CIEW sample data. The categories are ranked in terms of their percentage of
total CI usage in the sample cities.
Table 2–3 Commercial and Industrial Water Use by Major Category

Average Annual Percentage Percentage
Daily Use (gpcd*) of Total CI Use of CI Customers
Urban irrigation 2,596 28 30
Office buildings 1,204 10 12
Schools and colleges 2,117 9 11
Restaurants 906 9 11
Hotels and motels 7,113 6 2
Laundries/laundromats 3,290 4 1
Hospitals/medical offices 1,236 4 4
Food stores 729 3 5
Auto shops 687 2 7
Membership organizations 629 2 6
Car washes 3,031 1 0
Totals 78 83
* gallons per day per customer

Source: Commercial and Institutional End Uses of Water, 2000, p. 110.
Urban irrigation, which contributes to peak demands, is the highest single use
of CI water at the sample utilities. The CIEW study underlines the variability of CI
water use across and within types of establishments. One measure of this variability
is provided by the standard deviations of the average annual daily water use by
customer type. The standard deviations of water use of each of the 11 customer
types in Table 2-3 exceed the listed averages. This is possible only for a skewed,
nonsymmetric distribution of customer water use. The CIEW sample encompasses
a few extremely large water users, along with a larger number of smaller water users
in each CI category.
The CIEW develops predictive regression equations for categories of CI
customers. The number of employees per establishment and the building area figure
as important explanatory variables in several of these statistical models. In some
cases, water-demand projections are best served by separating out the largest CI
water users in a system and conducting specific analysis and projections for these
large customers.
Many industrial water users are self-supplied. Industrial water-use withdraw-
als in the United States are an order of magnitude larger than treated water use by
industrial concerns.
Water Loss. Figure 2-7 shows the distribution of unaccounted-for water or

water losses reported in the AWWA WDS. On average, 12.33 percent of annual
70
60
50
Number of Utilities
40
30
20
10
0
1 3 5 7 9 11 13 15 17 19 21 23 25
Water Losses as a Percentage of Total Water Production
Figure 2–7 Water Losses Reported in the AWWA WDS
water production in 2005 is reported as water loss.5 This percentage is somewhat

lower than other, earlier estimates, such as those found in AWWA (2003) and
Hutson et al. (2004).
For example, Solley and colleagues (1998) estimate that only 34 billion
gallons per day (bgd; 12,873,000 m3/d) of the 40 bgd (15,145,000 m3/d) withdrawn
in 1995 by water utilities is documented as consumed by end users. This suggests
water losses of approximately 15 percent, or “enough to meet the needs of the
10 largest US cities.”6
Water losses typically include both water leakage from mains and the
distribution system and water usage by customers that is not properly measured, paid
for, or credited. Water-loss control can pay dividends, especially for utilities
experiencing increasing variable costs for pumping and looking at higher prices for
additional supplies of raw or wholesale water.
5. Reporting of simple percentages as a measure of water losses has been criticized on the grounds
that better volumetric measures, taking into account the specific infrastructure setup of a water system,
can now be developed (see AWWA Water Loss Control Committee 2003). The idea is that unavoidable
water losses are related to the length and type of water transmission lines, connections, and other fea-
tures of the water system infrastructure, which may not be comparable across topographies and geo-
graphic extent of different urban areas.
6. AWWA Water Loss Control Committee (2003), p. 69.
OTHER BASIC FACTS ABOUT URBAN WATER USE

Water customers respond to how water is billed, and metering has a definite impact.
We can see dramatic evidence of this in water systems that convert from flat rates to
metered billing arrangements. Studies of the conversion to “universal metering”7
show big water savings when residential customers are billed according to how much
water they use, instead of paying a fixed fee (flat rate) per billing period for any
quantity of water use. Differences in water use among customers with broadly the
same characteristics (e.g., family size, lot size, and income, among others) are
consistently on the order of 30 percent, ranging from 10 to as much as 50 percent.
Metering also pays dividends in management and engineering terms. When
water customers are metered, utilities can track water system performance and water
losses much more readily.
Water Conservation. Water conservation is a special concern in the western
part of the United States because of rapid population growth and arid conditions.
When water systems face constraints resulting from litigation over surface sources,
water-quality problems, and depletion of groundwater sources, among other reasons,
water conservation becomes a priority. Many state and regional authorities require
municipalities and public supply systems to develop and submit water-conservation
plans.
Against the backdrop of this increasing level of interest in water-conservation
planning, the Energy Policy Act of 1992 (EPAct) resulted in ongoing reductions in
indoor water usage. The act requires standard flow rates and flush volumes for
plumbing manufacturers supplying both new construction and the replacement
market.
Research sponsored by AWWA and the California Urban Water Conservation
Council estimates that this federal legislation will reduce public water demands by
5 percent by 2010, compared to what overall urban water demand would be otherwise
(Dickinson et al. 2003).
A range of studies (see chapter 10) conclusively demonstrate the conservation
effectiveness of low-flow toilets, faucets, and conservation water rates. Accordingly,
the potential for achieving water savings in urban water systems may be on the order
of 15 to 25 percent over a planning horizon of one to two decades. One issue is the
cost of implementing conservation measures in comparison with the benefits of
conservation to customers, utilities, and other water-supply and environmental
stakeholders.
SUMMARY AND CONCLUSIONS

Public water supplies served an estimated 242 million people in the United States in
2000, utilizing about 11 percent of total US water withdrawals. Total per capita water
7. See, for example, Maddaus (2001).

use in public water supplies is estimated to have dropped from around 183 gpcd in
1985 and 1990 to around 179 gpcd in 1995 and 2000. Overall, per capita urban water
demands are highest in the US Rocky Mountain West and Southwest and on the West
Coast.
From the standpoint of water demand, most urban water systems share a
number of common features. There is, for example, the long view—total system water
use spanning a decade or more. In larger water systems, graphing several decades of
water use often suggests an increasing total system and per capita water use trend in
the early years, followed by a flattening, or even a decline, in later years.
The seasonal pattern of water use is another important perspective. Most public
water supply systems see peak demand in the summer because of outdoor water
demand for landscaping, lawns, gardens, and other similar uses.
Other facts relate to major components of urban water use. Usually, billing
arrangements in urban water systems distinguish among single-family residences,
multifamily residences, and CI and institutional customers. In addition, there can be
public uses of water for parks and recreational facilities and, inevitably, there are water
losses, which are calculated as the difference between water production and sales.
In many communities across North America, residential users account for 50 to
60 percent of total water production or sales, with the proportion varying by the size
of the water system. Single-family residences usually make major contributions to
system peak demand, generally because these residences account for a large
percentage of the total outdoor water use in water systems.
Recent studies give us a clearer picture of the indoor water use of residential
customers, as well as new information on the water use of several categories of CI and
institutional customers. We now also have a better understanding of the conservation
effectiveness of low-flow toilets, faucets, showerheads, and conservation water rates.
REFERENCES
AWWA. 1991. Water Industry Data Base: Utility Profiles. Denver, Colo.: AWWA.
AWWA. 2003. Water Stats 2002 Distribution Survey CD-ROM. Denver, Colo.: AWWA.
AWWA Water Loss Control Committee. 2003. Committee Report: Applying Worldwide BMPs in Water
Loss Control. Journal AWWA 95:8.
Dickinson, M.A., L.A. Maddaus, and W.O. Maddaus. 2003. Benefits of the United States Nationwide
Plumbing Efficiency Standards. Water Science and Technology: Water Supply 3(3): 231–237.
Dziegielewski, B., J.C. Kiefer, E. Opitz, G.A. Porter, L.L. Lantz, W.B. DeOreo, P.W. Mayer, and J.O. Nelson.
2000. Commercial and Institutional End Uses of Water. Denver, Colo.: AwwaRF.
Fesko, P., Manager, Office of Strategic Services, City of Calgary Waterworks. Personal communication,
November 2006. Data collated by Martina Kopat-Wright.
Hutson, S.S., N.L. Barber, J.F. Kenny, K.S. Lindsey, D.S. Lumia, and M.A. Maupin. 2004. Estimated Use of
Water in the United States in 2000. USGS Circular 1268. Reston, Va.: USGS. Available online at
pubs.usgs.gov/circ/2004/circ1268/.
Maddaus, L. 2001. Effects of Metering on Residential Water Demand for Davis, California. Prepared for
Masters of Science Project Requirement, Civil & Environmental Engineering Department,
University of California, Davis.
Mayer, P.W., W.B. DeOreo, E.M. Opitz, J.C. Kiefer, W.Y. Davis, B. Dziegielewski, and J.O. Nelson. 1999.
Residential End Uses of Water. Denver, Colo.: AwwaRF.
Miichelsen, A.M., J. Thomas, D.M. McGuckin, J. Reid, L. Thelen, and P. Manning. 1998. Residential
Water Use, Rate, Revenue and Nonprice Conservation Program Database. Denver, Colo.: AwwaRF
and AWWA.
Solley, W.B., R.R. Pierce, and H.A. Perlman. 1998. Estimated Use of Water in the United States in 1995.
USGS Circular 1200. Reston, Va.: USGS.
CHAPTER 3
Choosing a
Forecasting Approach
The choice of a forecasting approach depends on the expected uses for the forecast
results, as well as the size and other characteristics of the utility and its service area.
Generally, the amount of effort invested in forecasting increases with the size of the
utility and with the importance of the decisions that will be influenced by the forecast
results. Simple forecast methods are appropriate for systems with slowly changing
customer bases that consist mostly of residential and small commercial customers.
The simple methods are also appropriately used to support decisions that are easily
revised or incur only small resource costs. More complex models are well suited to
larger utilities with diverse customers and significant growth and to support decisions
that involve a large commitment of resources and are not easily rescinded.
FORECAST METHODS AND MODELS CURRENTLY IN USE

Approximately two-thirds of the water utilities responding to the AWWA Water
Demand Survey (WDS) develop and apply at least one of the following basic types of
water-demand models:
• Annual per capita water-demand forecasts
• Annual water-demand forecasts by major customer class
• Peak-day forecasts
• Monthly system water-demand forecasts
31
• Daily water-demand forecasts
• Revenue forecasts linked with water-demand forecasts
The AWWA WDS also collected data about methods used to develop water-
demand forecasts, as shown in Table 3-1.1 The majority of utilities develop demand
forecasts in house, although nearly 43 percent hire consultants. About 11 percent of the
utilities surveyed develop water-demand forecasts with statistical regression analysis.
Table 3–1 Methods for Developing Water-Demand Forecasts Reported in the AWWA WDS
Response* Count Percentage
Develop demand forecast in house 354 53.6
Rely on consultant to develop

demand forecast 281 42.6
Forecast provided by larger area

planning entity 69 10.5
Develop and implement forecasting

models with spreadsheets 150 22.7
Develop water-demand forecasts

with statistical regression analysis
of water-use data 74 11.2
Use proprietary software 9 1.4
Use other off-the-shelf software† 13 2.0
*Questions allowed for multiple positive responses.

†If this option was checked, the respondent was asked to please list the name of the software in the
associated “Additional Comments” box.
Off-the-shelf software is used by 2 percent of the responding utilities, mostly in

large cities or regional water systems. An example is Forecast Pro (Business Forecast
Systems, Inc., Belmont, Mass.), which is an “automatic forecasting program” used by
many business firms. These programs purport to select the best forecast method,
given input data.
Only nine very large utilities reported using specialized, proprietary, water-demand
forecasting software. These specialized forecasting packages offer a framework for
organizing information about the water use of various customer groups and conservation
and pricing impacts. IWR-MAIN and the DSS model are two examples. The focus of both
packages is on long-term water-demand forecasting for capacity planning.
1. The survey question was “What methods, resources, and software do you use to forecast your
water demand? Please mark all that apply.”
CHOOSING A FORECASTING APPROACH 33
IWR-MAIN dates back to research on national patterns of residential water

demand conducted in the 1960s. IWR-MAIN is structured as a sectoral forecast
model. Earlier versions had built-in default parameters that allowed users to make
adjustments. The software can be calibrated to forecast commercial and industrial
(CI) water use at a high degree of disaggregation (i.e., the two- or three-digit Standard
Industrial Classification [SIC]) level. The basic water-demand model in IWR-MAIN
is in double logarithmic form so that estimated coefficients are elasticities of the
explanatory variables, such as water price or consumer income. IWR-MAIN migrated
to the Microsoft Windows platform in the 1990s, and has been or is being applied in
larger metropolitan areas, such as the Metropolitan Water District of Southern
California. At this time, the software is probably best considered as providing a
framework for demand modeling by customer grouping, because many of the key
inputs, elasticities, and so forth are supplied by intensive analysis of on-site data.
The DSS model—short for “Least Cost Planning Demand Management
Decision Support System”—has been used by a broad range of water utilities in recent
years. Developed by the California water-consulting firm Maddaus Water Manage-
ment, the DSS model has recently been offered as a planning tool by the California
Urban Water Conservation Council. The DSS model is a Microsoft Excel–based
application that provides a framework for the development of an end-use model,
water-demand and sewer-flow forecasts, and benefit-cost analysis of water conserva-
tion measures.
The DSS employs top-down and bottom-up methods to calibrate end-use
models of water use for major customer segments, such as single-family residential,
multifamily residential, CI, institutional (schools), and other (e.g., fire hydrant and
construction) accounts. These customer class end-use models are linked with
estimates of conservation savings for which benefit-cost analysis is used to determine
the most attractive package of water-conservation policies and measures.
Long-term water-demand forecasts for periods of 10 to 30 years are developed
in relation to population and employment forecasts, along with the portfolio of
conservation measures adopted by the water utility. The DSS model recently saw
application in regional water-demand and conservation forecasts in the San Francisco
Bay area (Levin et al. 2006).
FORECASTING MODELS
Water utility analysts can choose from a wide range of available forecasting models.
Generally, utility representatives will seek a new forecasting approach when the utility
has a problem that is not being adequately addressed by existing information. As a
starting point, analysts determine the forecasting method(s) currently in use. These
methods can range from simple informal forecasts, in which decision makers judge
that the future will be just like their recollections of the past, to complex formal
models requiring many variables, large amounts of data, and a significant commit-
ment of resources. The following list is helpful in considering the range of alternatives.
We discuss many of these extensively in subsequent chapters of this book.
• Subjective judgment-based methods include (1) expert opinion, (2) Delphi

methods, (3) alternative scenario building, and (4) surveys.
• Extrapolation methods include (1) averaging, (2) trend analysis, (3) exponen-
tial smoothing, and (4) Box-Jenkins/ARIMA (autoregressive integrated mov-
ing average).
• Multivariate regression methods.
• Nonparametric methods include neural networks and fuzzy logic systems.
Accurate information on current water sales and revenues is essential as a starting
point for any forecast. Current population and trends in population of the utility service
area, which we discuss in chapter 7, are also critical to the forecasting effort.
Subjective Methods. Judgment-based forecasting methods vary widely,

ranging from the informed opinion of utility management to highly structured
Delphi and scenario-building methods. Presumably, utilities that do not use a formal
forecasting method rely on the informed opinion of management to make decisions.
For many small utilities that are experiencing slowly changing conditions, this may be
sufficient. Larger utilities and those facing more rapid changes in their service areas
would most likely benefit from more elaborate methods.
Scenario building involves brainstorming among an informed group of
stakeholders about future conditions. Generally, the participants develop several
alternatives, which might include, for example, high growth, low growth, and no
growth of the service area. The implications of each are worked out, based on various
assumptions about the future, as discussed in chapter 13. Such processes can help
identify long-term trends and inform long-term planning.
Extrapolation. With these methods, analysts rely on historical data on water

use (or revenue) to forecast future values. Several of these methods are addressed in
chapter 5. They can range from simple averages of historical data to relatively complex
Box-Jenkins/ARIMA models that make use of past variations in the data series to project
its future values. The strength of extrapolation models is that the only data required are
the historical data on the variable being forecasted. This means that if a utility’s analysts
have 30 years of annual data on aggregate system water deliveries, they can prepare a
multiyear forecast of future water deliveries with any of these methods.
Five years of monthly water-use data could be projected for up to five years into
the future using moving averages, exponential smoothing, or ARIMA methods. Daily
or hourly data could also be forecasted with these methods. All of these single-
variable methods share a major limitation—they do not account for changes such as
population shifts, conservation programs, or price increases. The methods assume
that the future will be essentially like the past.
These methods are often used, however, to forecast per capita or per-customer
water use. The result is multiplied by a population forecast to obtain a forecast of total
system water use. Forecasters often make adjustments for predictable changes such as
new conservation programs, real water prices, the continuing impact of plumbing
codes, and changing area incomes. This process is fully demonstrated in chapter 5.
The utility might also decide to disaggregate its customers into categories such
as residential and industrial, and forecast use within each category separately, using
forecasts of the number of customers by class. This is often referred to as a “sectoral
forecast.” For larger utilities with diverse customer groups, sectoral forecasts are
usually more accurate than the simpler per capita method.
Multivariate Regression. The essential feature of these models is the use
of a set of driver or explanatory variables to show why water use has changed
historically and to forecast its future values. The models directly incorporate
anticipated changes in driver variables such as, for example, customer income
levels, water rates, conservation programs, weather factors, and technology ad-
vancements. The challenge arises because these driver variables must be forecasted
first, before water-use forecasts can be developed. This makes the entire effort far
more complex. We explain regression methods in Chapter 12 and use examples
throughout this work.
During the estimation phase of regression model building, explicit mathemati-
cal relationships, which relate changes in each explanatory variable to the target
variable (water use or revenue), are developed and tested. These relationships are
assumed to remain fixed during the forecasting period so that they can be used to
compute the forecast.
Nonparametric Methods. Methods such as neural networks and fuzzy logic
present intriguing alternatives to the usual regression model, in that the estimating
program “learns” about the relationships among the variables in the model, and uses
that knowledge to produce forecasts. These methods are more difficult to use than the
other models, and much harder to explain to potential users. Although there may be
some potential gains in forecast accuracy from using these models, their obscurity
makes them all but unavailable to most water analysts.
SELECTING A FORECASTING METHOD—

A PRAGMATIC APPROACH
Faced with such a welter of possibilities, how is the water utility professional to select
a water-demand model and modeling approach? We offer some guidelines in this
section.
From a practical standpoint, the size of the utility and the motivation for the
water-demand model are of commanding importance. In general, the development of
sophisticated water-demand forecast models, often employing a variety of forecasting
methods, is confined to larger water systems and, often, to utilities where there is a
perceived problem surrounding future water availability or cost.
The nature of these perceived problems can be broad. In some cases, there are
constraints on water availability. Perhaps current supplies depend on depleting
sources, such as systems that mine groundwater. Surface water sources fed by melting
glaciers may also be seen as potentially compromised in future years. Generally,
however, water supply limitations are not exclusively hydrological, but are mediated
by administrative compacts, water courts, and allocation arrangements. Several major
urban water systems, for example, are obligated to justify future water withdrawals
based on the appropriateness of their customers’ demands and the credibility of their
efforts to conserve water.
Water quality, both of intake water and requirements for sewerage treatment
discharges, is another problem that can stimulate efforts to come to grips with the
likely path of future water demand. Given motivation in terms of an underlying
problem, water-demand modeling is a process that often expands as data streams and
general understanding improve.
Choosing the best forecasting approach involves several decisions. First, utilities
consider the motivation. Is there a problem that justifies expending special effort on
water-demand forecasts? Increasingly, the answer is yes, leading to additional
questions: Should we develop water-demand forecasts in house or hire consultants?
Should we develop water-demand forecasts from scratch or apply specialized water-
demand forecasting software?
There are few shortcuts in statistics. For example, specialized water-demand
forecasting packages typically require considerable data on local water use and other
factors. National studies supply reasonable parameter estimates, such as summer
price elasticities or the percentage of residential indoors water used to flush toilets.
These numbers, however, vary in different communities and places. Accordingly,
when developing water-demand forecasts from scratch, utilities must take several
steps to calibrate specialized water-demand forecasting software and to monitor the
accuracy and reliability of the resulting forecasts.
Once a utility decides to develop a new forecast, management can decide to
develop the needed expertise in house, to hire a consultant, or to select elements of
each approach. Either way, forecasts can be developed from scratch using any of the
methods we discussed previously, or specialized water-demand forecasting software
can be procured and used.
Relying on engineering and other consultants to produce long-range and other
water-demand forecasts helps utilities manage the costs of staff training and obtain
access to current methods and techniques. Nevertheless, if the water utility is planning
on doing anything more than accepting the forecasts “on faith,” some knowledge must reside
in house.
Running through key criteria for selecting and evaluating a water-demand
forecasting method, then, is useful. These criteria can include (1) purpose of the
forecast, (2) data availability, (3) requirements for accuracy of the forecast, and (4)
performance and diagnostics of the forecasting model. In addition, factors such as
how well the forecasting model can be explained to stakeholders in the water-
planning process and the ease of updating the model come into play.2
FORECASTING PURPOSE
The choice of methodology, including the forecast horizon, is directly linked with the
intended purpose for the forecast results. The basic application areas for water-
demand forecasts include:
• Sizing system capacity and raw water supply
• Sizing and staging treatment and distribution system improvements
• Water-rate setting, revenue forecasting, and budgeting
• Program tracking and evaluation
• System operations management and optimization
Capacity issues and raw water supply usually relate to long-term forecast
horizons that range from one to several decades. Sizing and staging treatment and
distribution system improvements in a water system usually involve a medium-term
forecast horizon of several years to a decade. In the short term, a few months to a few
years, the forecast focus is on budgeting, program tracking and evaluation, revenue
forecasting, and rate setting. Finally, managing and optimizing system operations,
such as pumping and maintenance schedules, involve very-short-term forecasts—
periods of hours, days, or weeks.
CUSTOMER DISAGGREGATION
Forecast accuracy can often be improved, regardless of the choice of method, by
segmenting utility customers into relatively homogeneous groups. Single-family
residential, multifamily residential, CI, and governmental customer groups can be
individually analyzed. The choice of segments depends on the characteristics of the
utility service area and may include additional categories such as high- and low-
valued housing areas.
For the thousands of smaller public water supply systems, relatively simple
forecast methods suffice, and not just because of costs. With smaller numbers of
customers, disaggregating water use by categories is more likely to result in excessive
volatility within each category.3 Simpler forecasting methods, such as the tried-and-
2. Baumann and colleagues (1998) suggest similar criteria for evaluating water-demand forecasting
methods, calling the following “correlates of accuracy”:
• Is the chosen scope appropriate?
• Is there adequate disaggregation of the data?
• Do the model and assumptions (elasticities) reflect expectations?
• Does the model make use of the resources available?
• Is the model simple yet effective?
• Are the model and its assumptions robust?
3. Indeed, we can argue that this volatility ought to be a primary focus of statistical research in
smaller water systems, to help in planning for peak usage and shifts in the pattern of demand over time.
true per capita water-demand forecasting approach with no disaggregation, are

appropriate. For unmetered systems, the only available data may be system-wide,
making a segmentation strategy unavailable. At the same time, small utilities that
meter individual customer use should be alert to new or changing customer demands
that might make sectoral models worth the extra effort.
As the water system grows in size, customer water-use aggregates become more
predictable. Developing a sectoral water-demand forecast, which focuses on move-
ments of water use by major customer class, can result in gains in accuracy and
explicability. In chapter 5, we present an in-depth discussion of the development of
sectoral or per-customer forecasting models, coordinated with estimates of the
impacts of a generic water conservation program focused on water-efficient fixtures
for residences.
In general, sectoral water-demand models are more complex, but offer a better
framework for considering “what-ifs.” Evaluation of utility programs and policies
usually requires more detail than is available with per capita water-demand
forecasting. Sectoral water-demand forecasts also provide better benchmarks for
tracking water demand in the near term. Maintaining a sectoral water-demand
forecast is good business practice when data availability and system financial resources
allow it and heterogeneous groups of customers make it worthwhile.
Special studies complement sectoral water-demand forecasting models by
providing detailed information on key parameters. Consequently, utility analysts
might wish to develop a random sample of single-family residences to clearly
determine whether new homes use more or less water than older residences. Impacts
of innovative rate design, such as water budgets, are also good candidates for detailed
statistical studies.
DATA AVAILABILITY
The availability of data is often a primary constraint on developing forecasting
models. Some models such as ARIMA, exponential smoothing, and regression require
many years of monthly data to properly configure. If these data are not available, such
models cannot be used. On the other hand, if data resources are abundant and the
time depth of historical series is on the order of a decade or more, demand modeling
options open up. Key questions include:
1. What data are needed to estimate the forecasting model?
2. Are these data available in the water system and the community?
3. How much does collecting these data cost?
In general, several years of data are needed to develop medium- to long-term
water-demand forecasting models. This requirement for time or historical depth of
the data is closely related to the importance and unpredictability of weather on urban
water use. The historical data must be of sufficient length or time depth to allow
unusual weather effects—such as droughts or exceptionally wet, rainy, and cool

periods—to be averaged out in the historical record.
At the same time, new forces can emerge in the community, causing changes in
water use patterns. Examples can include “densification” of settlement patterns, or
construction of substantially larger houses with more bathrooms and water-using
appliances on larger landscaped lots. Conversely, new developments, especially of
townhouses and condominiums may be built around natural areas with no cultivated
landscape. Carefully examining community trends helps analysts determine how
many years of data are really relevant in developing a forecast.
Another hurdle to overcome is getting data with the proper frequency. For
example, annual data may be available for an extensive period, but monthly data may
not be available. Or perhaps monthly data can only be constructed from more
complex and staggered data from billing cycles in a water system.
Because new accounting and data systems may have come online in recent
years, previous water-use data may be available only in paper archives (old reports)
and may have to be manually extracted and put into the current format. Changes
from one billing system to another and, perhaps, changes in the designation of major
customer classes, among other alterations, can cause discontinuities in this record.
Total water production data, sometimes available in a daily series, is another good
series for water-demand modeling.
Once water-use data are identified as collectible, more questions arise: (1) What
will the scaling factor be? and (2) What type of forecasting model will be developed
with water use or revenue as the dependent variable?
For regression models, the frequency and disaggregation of the selected driver
variables become important. Population counts, for example, are most accurate in
census years. For intervening years, data on births, deaths, residential occupancy, and
customer counts can be assembled to estimate year-by-year changes in population.
Generally, monthly population data must be interpolated from such annual estimates.
Presumably the population changes slowly enough that these values are sufficiently
accurate for forecasting purposes. Customer counts for a metered water system should
be available from billing records.
Some medium- and long-term water-demand forecasting models have been
developed using regression. These models can include explanatory variables such as
consumer income. Again, though, data frequency and data disaggregation can be
issues. Consumer income levels are estimated in connection with the US Decennial
Census, which is conducted every 10 years. The US Internal Revenue Service, state tax
agencies, research universities, and others can also provide estimates of consumer
income, which is often categorized by year and for entire states. Using these data for
individual utility districts is problematic, leading to the frequent use of assessed
housing values as a proxy variable.
Model Estimation and Accuracy. Accuracy in urban water-demand fore-

casting may have been lax in the past, but there are reasons to expect that higher
standards will be required in the future. In the past, the idea was to “build it and they
will come.” Utility professionals thought that underestimating demand was much
worse than overestimating it. But the future looks different. For one thing, water
supplies are constrained in many locations. And some utilities are grappling with
revenue issues as their demand stabilizes or expenses escalate. Finally, there are new
possibilities for short-term demand modeling for system optimization, and water-
demand monitoring also places a premium on accuracy.
Two questions are relevant in this context: (1) What are the costs of forecasting
errors? and (2) How much accuracy is feasible in water-demand forecasts?
Water-demand forecasting errors can be quantified with careful analysis, not
unlike that for calculating benefits and costs of programs in an integrated supply plan.
If, for example, the utility builds facilities that are too large, based on a “high-ball”
forecast, an extra cost burden can be passed on to customers, meaning that a smaller-
than-planned number of customers must bear the capital costs for excess capacity.
Excessive capacity can result in higher water rates or taxes, along with political
problems that can waylay utility management. On the other hand, if long-range
capacity issues are inadequately addressed, water shortages can result, causing even
worse problems. Water shortages may impose costs in the form of losses in
landscaping, convenience of water use, and constraints on new construction. In either
case, costs can be evaluated within the framework of discounted present values and
benefit-cost models.
The question of how much accuracy is feasible, on the other hand, can be
assessed with models such as those illustrated in chapter 6. These models break out
the effect of unusual weather on monthly or seasonal demand. Month-to-month
deviations from long-term norms or averages are essentially impossible to predict,
except a few days ahead. Accordingly, the ability to forecast water demand in a system
depends largely, in the short to medium term, on the volatility of weather patterns
compared to long-term climatic patterns by season.
In the longer run, water analysts must evaluate and quantify risks associated
with population or employment forecasts. Here, expert opinion is probably the best
source. History shows that demographic and employment forecasts in many locations
can deviate substantially from what actually materializes.
Broadly, there are four basic steps in estimating a water-demand model: (1)
collect the data, (2) estimate the model, (3) test the model, and (4) use the model for
forecasting.
These steps resolve into several related tasks, tests, and procedures:
• Collect data on water use and drivers of water demand
• Analyze key drivers of water use in the database
• Identify the preferred water-demand model
• Augment the database if necessary
• Make a first-cut estimation of the water-demand model
• Perform diagnostic testing of the water-demand model
• Carry out final estimation or calibration of the water-demand model

• Develop forecasts of the driver (independent) variables, if any
• Develop the water-demand forecast
Estimating a water-demand model can be as simple as picking a per capita
water-use figure from the historical record for recent years—a figure that reflects a
year of normal weather and typical business conditions.
Estimating a statistical regression model or a time-series model is more
involved. as we discuss in chapter 12, various “diagnostic” statistics usually accompa-
ny the estimates of model parameters, often leading to model refinements. Similarly,
with other more advanced forecasting approaches, there are diagnostics relating to
how well the model fits the data, whether the estimated parameters can be taken to be
reliable, and so forth.
Reference regressions have been widely developed, often by university or other
researchers, to explore issues such as the responsiveness of various customers or the
entire water system to changes in water rates, the effect of water conservation
programs, demographic shifts, and weather effects.
Ideally, the performance of the forecasting model can be measured by how well
the model forecasts. With water-demand forecasts, however, applying this criterion is
complicated by the importance and inherent unpredictability of transitory weather
events. As a result, we strongly recommend using tracking or monitoring models of
total system water use. This essentially means developing a comprehensive water-
demand regression (probably on a monthly or seasonal basis), and studying the
performance of predicted parameters (such as water-use rates of residential house-
holds) when making allowance for specific weather conditions.
The ultimate test of any forecast is how close it came to predicting what actually
happened. This suggests that utilities undertake periodic comparisons between
previous forecasts and realized values for water demand and utility revenues.

In manufacturing or sales of durable goods, forecasting pays off in terms of matching
inventory to customer needs. When demand is overestimated, there are definite costs
for holding excess inventory. On the other hand, a company can lose market share if
deliveries cannot be made in a timely fashion and customers go to a competitor to buy
product. The issues are similar, but perhaps more complicated, for urban water
markets.
There is, for example, an issue of fitting the tool to the task. First, the costs and
resources necessary for applying these forecasting methods vary widely. Second, it is
usually effective for water-utility staff to do the preliminary groundwork, including
some basic model development, even if the utility plans to contract out its major
forecasting projects.
The spread of automatic meter-reading systems, which use telecommunications

links with the utility billing system, should yield more and more real-time data.
Although this may be another example of “build it, and they will come,” the new spin
will be “We have all this information, so why can’t we do something with it?”
On a more serious note, we find it useful to keep in mind that knowledge about
urban water use and skills in forecasting are acquired only by grappling with actual
water-use data and the facts of particular urban water systems. Although budget
constraints usually mean that one forecasting approach must be favored over others,
there is an important sense in which the several approaches to forecasting urban water
demand complement one another. As a consequence, it is probably always useful, and
not very costly, to develop per capita extrapolations of water demand for a system—if
only to establish a benchmark or context for evaluating other approaches. Similarly, if
there were an automatic method for forecasting urban water use, seeing its output
would be helpful (assuming costs were within budget). To deepen our insight,
however, there are few substitutes for grinding through the development of a sectoral
forecasting model, based on local data and, ideally, with reference to local statistical
samples and studies to identify the values of key parameters.
REFERENCES
Army Institute for Water Resources. 1987. IWR-MAIN Water Use Forecasting System, Version 5.1, Users
Manual and System Description. Fort Belvoir, Va.: Army Institute for Water Resources.
Baumann, D.D., J.J. Boland, and W.M. Hanemann. 1998. Urban Water Demand Management and
Planning. New York: McGraw-Hill.
Levin, E.R., W.O. Maddaus, N.M. Sandkulla, and H. Pohl. 2006. Forecasting Wholesale Demand and
Conservation Savings. Journal AWWA 98(2): 102–111.
Maddaus, W.O., and M.L. Maddaus. 2004. Evaluating Water Conservation Cost-Effectiveness with an
End Use Model. Proceedings AWWA 2004 Water Sources Conference, Austin, Tex., January 12–14.

CHAPTER 4
Data and Data Structures for

Water-Demand Forecasting
Effective forecasting requires extensive amounts of accurate historical data. Without

such data, the only forecast possible is that tomorrow will be just like today, and the
same day a year from now will also be identical. Seasonality could be included in such
a simplistic forecast without historical data beyond one year—each month next year
will be the same as the corresponding month this year.
This chapter discusses issues surrounding the acquisition, organization, storage,
and preliminary examination of data that have proven useful in forecasting urban
water demand. Time series on water sales and production can highlight basic trends
(annual data), seasonal variation (monthly data), and peak-to-average water-use
relationships (daily data). Customer billing records, which document water-sales data
by individual customer, usually comprise the most available cross-section data. For
moderate to large water systems, these data, viewed by customer class, typically have
a skewed distribution with a long tail toward large quantities (lognormal).
Data collection presents a surprising number of issues. Questions arise about
how to treat missing values, how to identify and scrub outliers, and how much
disaggregation to use in terms of data frequency and drill-down to the customer level.
Also, recommendations and illustrations of exploratory data analysis are described in
this chapter. It is suggested to augment data collection with graphs and descriptive
statistics, which often reveal water-use characteristics that are not apparent in a
spreadsheet dense with numbers.
45
COLLECTING WATER-USE DATA

Total water-production and/or water-delivery records, along with individual usage
and billing records, are primary sources of information about system water use for a
given utility.
Aggregate System Deliveries. Every water utility should have a reasonably
accurate record of total water delivered to its distribution system by year, month, day,
and perhaps even hour. Utilities should keep records of annual and monthly deliveries
in an available form for as long as 30 to 40 years. These records are an essential
starting point for planning changes in total system capacity. Monthly or other
periodic data, perhaps corresponding to the utility’s customer billing periods, can be
used to display and analyze seasonal patterns of water use. And daily and even hourly
water-delivery rates are necessary for computing peak system capacity requirements.
Utilities should carefully record and preserve the timing and extent of unusual
water-delivery events. For example, a large fire or a water main break puts unusual
stress on the capacity to deliver water. A record of the additional amounts of water
delivered not only helps to account for otherwise unexplainable spikes in deliveries,
but also yields valuable insights into capacity requirements for firefighting and the
cost in lost water from insufficient maintenance work.
Records of the number of customers served in each billing period should also be
maintained for 30 to 40 years so that usage rates per customer can be incorporated
into long-term forecasts. Utilities should also retain breakdowns of customers by
meter size or by category (e.g., residential and commercial and industrial [CI]). Long-
term graphical analysis of the number of connections and each major category of
connection generates a useful snapshot of trends in the utility service area.
Rate schedules that apply to customers within the utility service area should be
compiled and kept indefinitely, allowing utilities to identify and analyze the various
price/rate proxies as discussed in chapter 9. A detailed history of the implementation
of water-conservation and water-restriction measures is also essential. Without these
records, analysts cannot determine the effect and effectiveness of any water-
conservation programs that might have been implemented, with a resulting loss of
precision for water-demand forecasts.
Individual User Records. Historical data showing water use by billing
period for each customer are essential for more precise forecasts. Analyzing individual
customer responses to changes in water rates, to conservation programs, and to
changes in the economic and social environment that are beyond the control of the
utility is often useful. In the past, utilities typically purged customer billing records at
relatively frequent intervals, such as every two years, to make room for new data.
Today, with computer hard drive sizes reaching one terabyte and inexpensive flash
memory offering almost limitless data storage capabilities, there is no reason to delete
billing information.
For this reason, recording all information on water purchases in a standard
format that allows convenient retrieval at any time is a wise choice for a utility.
DATA AND DATA STRUCTURES 47
Maintaining records for each billing period by individual customer allows future
analysts to aggregate the water-use data however they wish, providing the analyst with
a rich source of information.
Individual billing records also present a challenge to forecast accuracy, because
not all meters are read on the same day of the month. For some customers, a July bill
may include water use from a time period as early as May 26–June 25 or as late as
June 22–July 23. As a result, utilities should keep meter-reading dates as part of the
customer record. In addition, meters can malfunction, leading to erroneous data. And
sometimes, for various reasons, a meter is not actually read, with the utility supplying
an estimate of water use for that billing period. Random events can also affect
individual customers and cause their actual water use to differ markedly from their
intended water use for an individual billing period. For example, a hose bib that is
inadvertently left on or a toilet or irrigation system valve that leaks occasionally can
distort a customer’s actual use. Analysts usually assume that the “law of large
numbers” washes out the impact of random events by individual customer through
aggregation of many customers. As an example, while a leak at one residence may
increase its monthly use from, say, 3,000 gallons to 12,000 gallons, when this residence
is averaged in with 100 others, the impact is negligible.
At some point, a question arises: “How many years of data are needed for
analysis?” In general, more information is better than less information. Carefully
documenting the data formats and definitions ensures that stored data will be usable
in the future. In purely statistical terms, at a minimum, 20 years of annual data and 4
to 5 years of monthly data are necessary to support the development of forecasting
models. Analysts can obtain more reliable results from these models with 30 to
40 years of annual data and 10 to 15 years of monthly data.
ORGANIZING DATA
Depending on the intended end use, data can be organized in various ways. This
section discusses the organization of data by time series and cross section, followed by
pooling of the two.
Time-Series Data. A variable record by day, month, or year is often referred
to as a time series, which is simply a collection of observations on a particular
variable. The variable can represent either a flow or a changing status over time. A
flow variable, such as water sales, is often measured by the hour, day, month, or year.
The time period of measurement always accompanies descriptions of the variable,
such as water sales per year. In contrast, a status variable, such as the number of water
customers, is recorded at a specific point in time, such as December 31 of each year.
The number of customers could also be presented as an annual average of the number
at the end of each month. Ordinarily the number of customers is changing slowly
enough that the exact form of this variable makes no difference.
Level of aggregation is another dimension for classifying data. An extreme level
of disaggregation is a time series containing water used by each customer recorded at
close intervals, perhaps hourly, by meters linked by telecommunications with a central

computer. Utilities have collected short-term data from samples of water users for
limited time periods to assess the effectiveness of conservation programs. More
common are data aggregated by customer class; by geographic region; and by month,
quarter, or year.
Cross-Section Data. This type of data consists of a “slice” of water-use

patterns for a specific time interval (e.g., recorded water use for each individual
customer in a district for a single billing period or year). Cross-section data for
individual households are frequently the basis of research and analysis published in
professional journals. Many of these studies combine cross-section data on water use
with survey information on customer characteristics.1 These data support the
statistical estimation of demand functions for residential or other customers, showing
the response of water sales to influences such as differences in water rates, consumer
income, housing characteristics, and family size.2
Obtaining cross-section data to support multivariate or structural water-
demand forecasting models requires random samples of water purchases and data
collection. The samples focus on characteristics of the sampled customers. Within an
individual utility, a cross section might consist of a random sample of individual
customers by customer category. Utilities can also gather water-use information for
groups of customers broken out by distribution areas, census tracts, or other
divisions. This is often useful when demographic data is available for these
breakdowns.
Pooled Cross-Section and Time-Series Data. Some research uses com-

bine (“pool”) time-series and cross-section data, which consist of a number of
individual customers or customer groups with water use measured over many days,
months, or years. Pooled data are especially valuable when analysts are investigating
issues such as how water customers in different income groups respond and adapt to
various conservation programs or to increases in water rates over time.
HOW MUCH DATA DISAGGREGATION?

For forecasting purposes, how much data disaggregation should be undertaken? An
example is given in chapter 5 that shows how forecast accuracy can be improved by
separately analyzing residential, CI, and public water customers. How useful is
disaggregating water-use data to narrowly defined categories such as residential
customers of certain types in certain areas of the water district? Although no hard and
fast rule covers all cases, the general principles of composition and noise can be
applied.
1. Appendix A describes methods of collecting survey data.

2. Such regression models are discussed in chapter 9.
The principle of composition provides the impetus to disaggregate demand by

category. The idea is to separate system demand into categories and develop
individual explanations for the typical water use of each customer grouping. Total
water use for each group is obtained by multiplying these per customer results by the
forecasted size of each group. Overall totals are then obtained by adding all the groups
together. Because each group may respond differently to changes in rates or other
programs, this can add precision to forecasts and individual models can capture these
effects. The technique also facilitates the use of the most appropriate forecast method
for each group.
The principle of noise is based on the theory that, because of random
(capricious) components in behavior, as well as inevitable errors in observation, it is
easier to predict behavior in the aggregate than for individuals or smaller groups.
Ideally, efforts to predict how much a consumer will buy are just as likely to
undershoot as overshoot the correct amount, unless an important driver variable such
as temperature or rainfall has been ignored. Forecasting errors will tend to “net out”
to zero if enough customer-use patterns are combined. The water use of several
hundred or several thousand customers can usually be predicted with a smaller
percentage error than the water use of a single or a few score customers. Predicting the
water use of 100 groups of 200 customers each and adding the results may yield more
accuracy than a single forecast for all 20,000 combined, especially if there are marked
dissimilarities among the groups. One challenge for water-demand forecasters is to
identify the appropriate level of disaggregation, striking a balance between the
principles of composition and noise.
Outliers and Missing Values. Two of the problems often encountered in

data analysis are outliers and missing values. Outliers are values imbedded in the data
series that do not seem to make sense. They often result from data-entry errors or
unusual events. When an unexpectedly large or small value of a variable is
encountered, the analyst should look for an explanation. A broken meter may have
recorded far less water than a customer actually used. Total system deliveries may have
been extraordinarily large in a particular month because of a fire or a water-main
break. Disabled pumps or supply shortages may have resulted in unusually small
deliveries. If the analyst can identify such events, they can be included in the
forecasting model, or the data can be adjusted to remove their effects. If no cause can
be discovered for unusual values, the analyst must choose among adjusting these
values to “normal” levels, omitting the offending cases from the file, or ignoring them.
Analysts should give careful consideration, including a well-thought-out rationale, to
any data adjustments.
Similar considerations apply to missing values. In some cases, analysts can fill in
a missing data point by interpolating from neighboring values. This is usually
appropriate for gaps in a data series such as personal income or summer tempera-
tures. The other alternative is to drop the entire case. If there are sufficient data and
the “proper” value for the missing information is not obvious, this is often the
preferred approach.
OTHER VARIABLES
Water-demand forecasters use a number of weather, climate, demographic, structural,
and economic variables. The following section provides a partial list of these variables,
together with likely data sources. Each source is examined in greater detail in
subsequent chapters.
Weather and Climate. Climate refers to the typical characteristics of a

location, such as hot and dry or cold and wet. In contrast, weather refers to short-
term variations from the long-term averages. Weather and climate data are available
from local weather stations and published reports of the National Weather Service.
Additional data may be available from commercial reporting services, research
universities, and land grant colleges. The basic variables are temperature and rainfall,
with these and other variables available on an hourly, daily, or monthly basis. Chapter
8 provides a detailed treatment of how to most effectively incorporate weather and
climate variables into a forecast. Including deviations of actual temperature and
rainfall from long-term average values is often an effective strategy. Winter precipita-
tion has little impact on water use, but rain during the growing season has substantial
effects. Researchers often use evapotranspiration minus rainfall as a weather variable.
Evapotranspiration can be computed using readily available formulas and data, as
discussed in chapter 8.
Population. The US Census Bureau typically publishes historical population

data and forecasts, which are then promulgated by state and local economic
development agencies. Intercensus population figures are often projected using the
number of utility customers, school attendance records, and similar indicators. Other
demographic characteristics, such as household size, are available in the US Decennial
Census, which is taken every 10 years. These variables may also be available from
some of the sources mentioned previously on a more frequent basis. The availability
and use of population forecasts is discussed in detail in chapter 7.
Personal Income, Employment, and Production. Income and employ-

ment data by community are available from most state economic security agencies,
university economic research departments, the US Census Bureau (on an annual
basis), Sales and Marketing Management Magazine (available by subscription at
www.salesandmarketing.com/msg/publications/smm.jsp), the US Internal Revenue
Service, and state tax departments. Some banks and state or local government
agencies periodically publish economic data. Personal income has been persuasively
shown to have an impact on residential water use, with higher income households
using more water.
Measures of CI activity, such as employment, are essential for forecasting water
use by businesses. Data on CI activity for each state and county in the United States
are reported annually by the US Department of Commerce’s US Census Bureau in
County Business Patterns, the annual report on business activity by industry (available
online at www.census.gov/epcd/cbp/view/cbpview.html). More detailed information
is available in the Economic Census, also published by the US Census Bureau every
five years (available online at www.census.gov/econ/www/). Finally, additional data
may be available from economic development agencies, university research bureaus,
and other sources.
BASIC ANALYSIS OF WATER USE

This section provides an overview of the basic tools and techniques used in water-use
analysis.
Data Formats for Spreadsheet Analysis. Most analysts organize the data
they wish to analyze in a spreadsheet format, using a program such as Microsoft Excel
or Lotus (IBM Corporation, Armonk, New York; formerly the Lotus Development
Corporation). Table 4-1 gives an example of a useful way to organize time-series data,
in which each row starts with the date (by month in this example) and each column
starts with the name of the variable placed in that column. In this example, Quantity
is the average monthly water use of residential customers, Price is the marginal price
paid per unit of water, Temperature is the average temperature for the month, and
Rain is precipitation during the month. An actual analysis would likely include
additional variables such as personal income per household and four or more years of
monthly data. Cross-section data, such as those shown in Table 4-2, are organized in
much the same way, except that each row represents an individual customer or group
of customers instead of a time period. The variables given in each column are similar,
starting with a customer number. The CD that accompanies this book contains
complete databases.
Simple Data Analysis.

Graphs. An analyst’s first step may be to graph the data. The data presented in
Table 4-1 along with additional years of data are graphed in Figure 4-1 for illustration
purposes. In this example, the analyst used a spreadsheet program to develop a scatter
plot, with each quantity of water plotted on the vertical (y-) axis and each month
plotted on the horizontal (x-) axis. This graph clearly shows the seasonal pattern of
water use for this utility for each of the 12 months.
Figure 4-2 shows an equivalent cross-sectional graph, in which the number of
customers using each amount of water is plotted on the y-axis and increasing
quantities of water are plotted on the x-axis. Figures such as this are often referred to
as frequency distributions, which show the distribution of responses to a specific
question. For a variable that can take on a large number of different values, an analyst
must sort the data observations in categories to construct a frequency distribution.
For example, a frequency distribution of water use might summarize individual
customer use into categories, such as 0–5 units, 6–10 units, 11–15 units, and so forth.
Figure 4-2 shows the results for all customers of a large utility for a single
month. Note that the most common volume interval for these consumers is between
Table 4–1 Time-Series Data Format
Month Quantity Price Temp Rain
1 8.0 80.1 51.1 1.0

2 7.5 79.2 52.2 0.0
3 9.6 97.7 60.6 0.6
4 13.2 96.7 65.3 0.0
5 16.2 96.7 74.5 0.0
6 22.0 96.5 86.7 0.0
7 21.1 95.4 85.0 2.4
8 15.6 95.0 84.0 1.6
9 17.0 82.5 79.4 2.5
10 11.0 81.7 70.6 2.3
11 8.7 81.1 57.6 0.8
12 7.6 80.5 47.3 0.3
Table 4–2 Cross-Section Data Format
Account Year Month Consumption
28.02 5 7 0
30.04 5 7 15
33.05 5 7 13
34.09 5 7 2
35 5 7 0
36.02 5 7 7
37 5 7 9
38 5 7 17
40 5 7 0
11 and 15 units of water, and that the distribution tapers off to a long tail toward
larger quantities. Water use is clearly not normally distributed.
Maximum and minimum values. Obtaining the maximum and minimum values
for the data series is another good starting point. Examining these values often reveals
outliers or obvious data entry errors that will need the analyst’s attention. The
minimum value in the data series used to obtain Figure 4-2 is 0 and the maximum is
98,900. Because this one high value was several magnitudes larger than any other
observation, it was dropped from the analysis. It does warrant further investigation,
25
20
15
Water Use
10
0
1 2 3 4 5 6 7 8 9 10 11 12
Month
Figure 4–1 Time-Series of Water Use by Month
1000
900
800
700
Frequency
600
500
400
300
200
100
0
More
10
15
20
25
30
35
40
45
50
5
Consumption Range
Figure 4–2 Cross-Section of Water Use per Customer
however, because it may represent a single industrial customer who uses very large
amounts of water. In this case, the analysts would want to forecast water use for this
customer separately from the other customers. This is likely to involve direct
discussions between utility and customer management to assess future water
demands.
1
Average (mean). The arithmetic mean = ---- ∑ Q , where Σ is the summation sign
N
and Q is water use from the quantity column in Table 4-2 or Figure 4-2 and N is the
number of rows of data. The arithmetic mean is widely used as a convenient way of
describing a set of data. Common uses would be to say, for example, that the average
residential customer used 21 units of water per month during the past four years, or
that average June use by residential customers was 43 units.
Median. Median water use is that of the customer in the middle—half the
customers use less water and the other half use more. The median is often considered
to represent the typical customer, which is especially useful when water use is not
normally distributed. Because the distribution of actual water use often appears as in
Figure 4-2 (with a long tail in the distribution toward higher water use), the average
is not as good a measure of the typical consumer as the median.
Mode. The mode or modal value describes the water-use volume category with
the highest frequency. For the data represented in Figure 4-2, the modal value is 14.
Compare this value with the average water use of 18.5 and median water use of 17.
The maximum water use of any customer in this particular distribution is 206. This
customer—along with other high-use customers evident in the skewed distribution—
causes the average to be significantly higher than either the mode or the median.
Variance and standard deviation. These are measures of dispersion around the
central tendency or mean value of a variable.
∑ ( Qi – Q )
1 2
Variance = ---- (Eq 4-1)
N
i=1
where:
Qi = quantity of water for to the ith customer, i =1 to N
Q = mean water use
The standard deviation, which is the square root of the variance, can be
interpreted as the typical difference between an individual observation and the
average of the data series. The standard deviation is particularly useful when the data
are normally distributed. The normal distribution is often referred to as the bell-
shaped curve because it is distributed symmetrically around its average value in
roughly the shape of a bell. The characteristics of a normal distribution are entirely
summarized or described by its mean and standard deviation. When a variable is
normally distributed, for example, an additional observation has a 99 percent chance
of being within ± 2.57 standard deviations from its average value and a 95 percent
chance of being within ± 1.96 standard deviations.
Confidence interval. This is another way of showing the statistical accuracy of an
estimated parameter, such as a mean. A 95 percent confidence interval shows the
range within which the mean value of a sample drawn from some population is
expected to fall 95 percent of the time. If 100 individual samples were drawn from
some population, the average value from 95 of them could be expected to be within
the 95 percent confidence interval of the true, but unknown, population mean.
The confidence interval for an average is
S
Q ± Z --------- (Eq 4-2)
N
where:
Q = average (mean) water use
Z = the number of standard deviations corresponding to the
chosen significance level from a normal distribution table.
For small samples, Z is replaced by the t-statistic from a
table of the t distribution (found in the back of most
statistics books).
S = standard deviation computed from a random sample of the
population
N = sample size
The Z for a 95 percent confidence interval is 1.96. Thus, the 95 percent

confidence interval for the water-use distribution shown in Figure 4-2, which is a
sample of N = 4087 with estimated Q = 18.46 and S = 11.43, is
18.46 ± 1.96 × 11.43 ÷ N , which is 18.46 ± 0.35.
Therefore, the confidence interval for this mean value is
18.11 ≤ true Q ≤ 18.81.
This is a very tight confidence interval because of the large N. If the N had been
only 250, the confidence interval would be
17.29 ≤ true Q ≤ 19.88, which is a much wider interval.
Simple correlations. The simple correlations between pairs of variables in a data
array show the strength of their relationship. The correlations shown in Table 4-3
were computed using four years of data similar to those used in Table 4-1. The highest
correlation, 0.88, is between quantity of water and monthly temperature, which
reflects increased summer irrigation of lawns and other landscaping. Temperature
would play a role in any forecasting model using these data. The remaining
correlation coefficients are not encouraging for an analysis of this data set. Quantity
and price have a positive simple correlation, which undoubtedly resulted from the
imposition of summer surcharges in this water system, in addition to the lack of any
large rate increases during this time period. The positive correlation between price
and temperature results from this same relationship. Income does not appear to be
strongly correlated with any of the other variables. Because correlation does not, by
itself, imply causation, the analyst must have a sound theoretical reason for
interpreting a high correlation as a measure of causation between two variables.
Table 4–3 Simple Correlations of Monthly Data, Full Years
Quantity Price Temp Rain Income
Quantity 1
Price 0.428597 1
Temp 0.876109 0.510692 1
Rain 0.178837 –0.13247 0.243906 1
Income –0.10463 0.065906 –0.00437 0.1058 1
When only the summer months are included in the correlation matrix (Table 4-4),
the relationship between quantity and temperature is weaker because all the months
are relatively hot. But now a negative correlation appears between price and quantity,
suggesting that when rates for summer use are higher, usage is lower.
Table 4–4 Simple Correlations, Summer Only
Quantity Price Temp Rain
Quantity 1
Price –0.14529 1
Temp 0.36255 0.079292 1
Rain 0.184057 –0.51944 0.247425 1
Irregular distributions. The mean, median, and mode are measures of central
tendency, and they can indicate other shape characteristics of a distribution. A
positively skewed3or “left steep” distribution results in the mean being larger than the
median. This means that this distribution does not have the symmetric bell shape of
the normal distribution. A lognormal distribution, however, describes the data in
Figure 4-1. A distribution is lognormal when the logarithms (base 10) of the observed
values are normally distributed. Lognormal distributions are typical for wealth-
related variables.
Peak-to-average use ratio. When considering capacity requirements and water-
rate design, knowing the ratio of peak-to-average water usage is often useful. This
helps inform utility officials about peak-capacity requirements and can lead to water
rates designed to reduce peak usage. Analysts compute the ratio using monthly, daily,
or even hourly data. For example, Figure 4-3 and Table 4-5 show the monthly peak-
to-average ratios for each month for the same water system shown in Figure 4-1. The
3. Skewness is a measure of the extent to which individual observations are not uniformly distrib-
uted around the mean value.
1.8
1.6
Peak-to-Average Water-Use Ratio
1.4
1.2
0.8
0.6
0.4
0.2
0
ar
y ary rch ril y ne ly us
t be
r er er er
u ru
Ma Ap Ma Ju Ju g m cto
b
mb mb
an eb Au pte e e
J F
Se
O
Nov Dec
Figure 4–3 Peak-to-Average Water-Use Ratio
Table 4–5 Peak-to-Average Water Use by Month
Month Quantity
January 0.66
February 0.62
March 0.73
April 0.92
May 1.16
June 1.50
July 1.56
August 1.19
September 1.29
October 0.93
November 0.77
December 0.66
lowest use month is February with only 62 percent of the four-year annual average.
The highest months are June and July with 150 and 156 percent of the average,
respectively. The seasonal shift in demand is dramatic, with July usage being 2½ times
as large as that in February.
SUMMARY
In this chapter, typical data on water-use rates were examined and key data issues
were discussed. If water-demand forecasting is to be effective, data must be collected
locally. Water-demand analysts must ensure that water-use records are suitably
archived, aiming for 30 years or more data on water use, preferably in a monthly
series. A balance must be struck between detailed descriptions of water use by
customer grouping and the stability of water-use patterns (which can be determined
for larger customer groupings).
The chapter also highlighted the basic types of data—cross-section and time-
series. Basic descriptive statistics, such as the mean, mode, median, variance, and
standard deviation, are discussed along with the problem of missing values and
outliers in data. Finally, sources of data for weather, population, personal income,
employment, production, and CI activity were examined.
The underlying purpose in this chapter is to demonstrate that having a good
understanding of the data used for analysis helps forecasters keep their feet on the
ground and avoid errors in analysis.
CHAPTER 5
Simple Forecasting Methods

and Reality Checks
There are two widely applied approaches for long-term water-demand forecasting: (1)
per capita water-demand forecasts and (2) sectoral or per customer water-demand
forecasts, disaggregated by major sectors such as residential, commercial, and
industrial. These forecast approaches are utilized primarily to inform design and
acquisition of water system capacity. The focus is on annual usage. Forecast horizons
extend out one to several decades.
Both approaches are simple in conception, although not necessarily in
execution. Water demand for a system (customer group) is projected as the product
of a per capita water use (or unit-use coefficient for that customer group) multiplied
by the population (number of customers in that group).
For per capita water-demand forecasts, the fundamental equation is
Qt = Ntqt (Eq 5-1)
where:
Qt = total system water use in time period t
Nt = population in the water-system service area
in time period t
qt = per capita water use in time period t
Utilities forecast total system water demand as the product of total population
in the service area and water use per capita (an average water use of all the customers
59
counted). Here, water use is often measured in terms of pumpage or what is

commonly called water production.
Sectoral water-demand forecasts are based on adding together forecasts of water
use for major water-demand components or water-using sectors. The demand
components include water use by customer group, such as single-family residential
and commercial and industrial (CI) customers, as well as system water losses and, in
some cases, bulk water sales to customers outside the utility service area.
The following is a typical example of a water-system-demand equation in the
sectoral format:
Qt = SFRESIDENTIALt + MFRESIDENTIALt + COMM/INDt + PUBLICt

+ OUTSIDESALESt + WATERLOSSt (Eq 5-2)
where:
Qt = total system water use in time period t
SFRESIDENTIALt = water sales to single-family residential water cus-
tomers in time period t
MFRESIDENTIALt = water sales to multifamily residential water cus-
tomers in time period t
COMM/INDt = water sales to CI and institutional customers in
time period t
PUBLICt = water sales or deliveries to public customers in
time period t
OUTSIDESALESt = water sales to customers outside the water utility’s
service area in time period t
WATERLOSSt = water losses, measured as the difference between
water production and sales in time period t
This chapter outlines steps for per capita and sectoral water-demand forecasts.
Thirty year water-demand forecasts are developed with coordinated data on popula-
tion growth, numbers of customers, water usage, and other factors.
The first major section considers per capita water-demand forecasts, including
a simple method for identifying a year of normal weather.
The second major section focuses on sectoral water-demand forecasts. The
discussion considers customer use rates and customer counts, savings from a
“mandate and rebate” conservation program, programs targeting water losses, and
forecasting options.
The third major section lists reality checks on the forecasts. In general, reality
checks on forecasts include checking the consistency of assumptions, monitoring
near-term water sales and drivers of water usage, and risk analysis. End-use models of
water use are especially useful in checking the “reasonableness” of per capita and per-
customer use rates. A theme of this chapter is that sectoral water-demand models are
more complex but offer sizeable compensations.
In addition to increasing forecast accuracy, forecasts of water use by customer
group create a framework to consider various what-ifs. For example: What if
SIMPLE FORECASTING METHODS AND REALITY CHECKS 61
customer incentives and rebates are developed for special appliances such as front-
loading washing machines? What if sales to wholesale customers outside the system
increase faster than sales to customers in the utility service area? What if household
sizes shrink in the future, as the average age of the population in the service area
increases? What if substantially greater funds are invested into leak detection and
control? Evaluating impacts from such programs and developments usually requires
more detail than is typically used for per capita water-demand forecasting. Although
it is possible to adjust per capita use rates after analyzing such programs and
developments, doing so usually means creating the type of data that would support a
sectoral water-demand forecast.
Sectoral forecasts also create a better framework for tracking water demand in
the near term. Of course, metering the water use of at least some customers is a
prerequisite for developing sectoral water-demand forecasts. Without full metering,
the water use of unmetered customers must be disentangled from system water losses
by some process of imputation and guesswork.
The bottom-line is that developing and maintaining a sectoral water-demand
forecast is good business practice in the sense of knowing the customer.
PER CAPITA WATER-DEMAND FORECASTS

Table 5-1 illustrates a simple per capita water-demand forecast. In this example, the
forecast horizon is 30 years. Forecasts for total system water use, along with peak-day
use, are in million gallons per day (mgd). Forecasts are shown for 5, 10, 20, and
30 years from the base year (2007), which is assumed to be the most current data
available to planners.
Table 5–1 Simple per Capita Forecast Approach
Base year population (2007) . . . . . . . . . . . . . . . . . . . . . . . . 100,000

2007 water demand/production, mgd . . . . . . . . . . . . . . . . . 28.3
2007 per capita water use, gpcd . . . . . . . . . . . . . . . . . . . . . 282.7
2007 year peak-to-average day ratio . . . . . . . . . . . . . . . . . . . 1.75
Year Population Water Demand Forecasts

Forecast
Average Day (mgd) Peak Day (mgd)
2012 110,408 31.2 54.6
2017 121,899 34.5 60.3
2027 148,595 42.0 73.5
2037 181,136 51.2 89.6
The 2007 population in the utility service area is 100,000. The total water-
system demand or production per day in this base year averages 28.3 mgd. This works
out to an average of 282.7 gallons per capita per day (gpcd) in 2007, and is taken as
the per capita use rate for the average-day forecast.
Population growth is projected at 2 percent per year over the next 30 years. This
conservative figure is chosen because urban areas that have reached a certain
threshold of total population tend to show slower growth rates into the future than in
the past.
In 2012, the population forecast is 110,408, and at the end of the 30-year
forecast horizon, this number rises to 181,136. Typically, population forecasts for the
utility service area are adapted from forecasts developed by local demographic
authorities in the municipality, county, or state.
These population forecasts, together with the 282.7-gpcd use rate, lead to the
water-demand forecasts in the lower part of Table 5-1. It can be determined that
average-day water demand/production is projected to rise from 31.2 mgd 5 years into
the forecast horizon to approximately 51.2 mgd at the end of the 30-year forecast
period in 2037.
Peak-day demand forecasts are developed with the ratio of peak-to-average-day
water demand in the base year of 1.75.
Identifying a Year of Normal Weather. Per capita water-use numbers in

many urban water systems show substantial year-to-year variability. This suggests the
need for more in-depth analysis, based on historical data on water usage and weather.
A weather index can help identify years of normal weather. Table 5-2 gives the
historical data supporting the analysis presented in this section. In Table 5-2, system
water production, measured in million gallons per day, is given in Column 1. Column
2 shows the population of the service area. Per capita water use in Column 3 is
computed by dividing total system water production by population. Columns 4 and 5
list average annual temperature and total annual precipitation.1
Columns 6 and 7 are standardized values for average annual temperature and
total annual precipitation. These values are created by subtracting the mean value of
the variable in question from each observation, and dividing the result by the
standard deviation. Table 5-2 takes the long-term average temperature to be 66.5°F
with a standard deviation of 1.5. Long-term averages and standard deviation for total
annual precipitation are 17.4 and 7.01 inches, respectively.
1. Forecasters should take climatic factors into consideration when selecting variables to identify a
year of normal weather. In some northern latitudes in North America, for example, winter is routinely
associated with high precipitation but very low temperatures. In such situations, it may be best to focus
on average temperature and total precipitation in the summer irrigation season. In other locales, usu-
ally with higher winter temperatures, ground moisture is significantly affected by year-round precipi-
tation. Annual figures for temperature and precipitation are highly defensible in this context.
Table 5–2 Database for Analysis of per Capita Water Demand
Average Standardized Standardized

Average Daily Population Annual Total Annual Average Annual Total Annual Long–Term
System Water of the Temperature Precipitation Temperature Precipitation Weather
Use (mgd) Service Area gpcd (°F) (in.) (°F) (in.) Index
Year (1) (2) (3) (4) (5) (6) (7) (8)
1993 19.7 71,191 276.8 65.1 26.6 –0.933 1.314 2.6

1994 20.0 71,903 277.8 65.4 20.1 –0.733 0.385 0.7
1995 20.7 72,622 285.0 65.5 13.4 –0.667 –0.575 0.8

1996 20.9 74,437 280.6 64.5 7.4 –1.333 –1.434 3.8
1997 21.2 76,298 277.5 64.0 19.6 –1.667 0.314 2.9
1998 22.1 78,587 281.6 65.1 11.9 –0.933 –0.780 1.5
1999 22.7 81,338 279.0 64.2 8.1 –1.533 –1.332 4.1

2000 23.5 83,778 280.5 66.1 27.9 –0.267 1.491 2.3
2001 25.2 87,129 288.8 68.3 10.7 1.200 –0.957 2.4
2002 26.3 91,485 287.7 67.6 17.8 0.733 0.050 0.5

2003 26.7 94,230 283.4 67.8 23.7 0.867 0.892 1.5
2004 28.1 95,172 295.0 67.6 8.7 0.733 –1.247 2.1
2005 27.2 95,648 284.5 67.2 23.5 0.467 0.870 1.0

2006 28.6 98,561 290.1 68.0 22.7 1.000 0.749 1.6
2007 28.3 100,000 282.7 65.9 15.6 –0.400 –0.258 0.2
NOTE: gpcd includes both residential and CI water use, divided by total population.
Table 5-2 uses the formulas 5-3 and 5-4 to standardize temperature (t) and
precipitation (p):
(t – t ) -
t∗ = ----------------- (Eq 5-3)
stdev ( t )
(p – p)
p∗ = ------------------- (Eq 5-4)
stdev ( p )
Here, t stands for average annual temperature over the period of record, and
p is average long-term precipitation. The standard deviation of the annual tempera-
ture series is denoted by stdev(t), while stdev(p) indicates the standard deviation of the
total annual precipitation series.
Standardizing average temperature and total precipitation transforms these
values into numbers clustering around zero. In this example, all the historical values
are within ±2 standard deviations from their long-term mean values. This serves the
purpose of putting both weather variables in a similar frame of comparison.
Column 8 in Table 5-2 lists the long-term weather index w, calculated by
formula 5-5 as the sum of the squared values in Columns 6 and 7:
w = t∗ 2 + p∗ 2 (Eq 5-5)
This formula accords temperature and precipitation equal weights in the

weather index. The idea is to take variables, each of roughly similar magnitude, and
square them so that large negative values count the same as large positive value. Then,
lower values of the sum or index w are associated with weather patterns closer to the
long-term averages of the weather variables in the index.
In Table 5-2, 2007 is the lowest, having an index w of 0.2. This weather index w,
therefore, suggests that 2007 is the year in the historical record in which temperature
and precipitation values are closest to the long-term norm.
Trends in per Capita Use. These temperature and precipitation values and
a linear trend component can be regressed against this per capita water-use series.
This exercise, which is included on the accompanying data CD, shows no statistically
significant trend. Annual temperatures in these data increase in the early years of the
data series, boosting per capita water use. The bottom-line is that regression analysis
also suggests that “weather-normalized” per capita water use is similar to the 2007 per
capita number.
When a time trend is discovered in water-use data, it makes a difference, in
practical terms, whether the trend is positive (increasing) or negative (decreasing).
When the historical trend in per capita water use is positive, this trend usually
is projected to continue over the forecast horizon.
Matters become more problematic when per capita water use exhibits a negative
trend. Many larger urban water systems currently report declining per capita water
use. There are limits to the forward projection of a negative trend in per capita water
use, however. Questions arise such as: How low is too low? At what point will per
capita water use stabilize? The answers are largely a judgment call, although an end-
use analysis can be helpful. Per capita water-use rates in this situation will generally be
determined to level off after some date. The gradual replacement of older faucets,
showerheads, and toilets with newer, more water-efficient models may justify
extrapolating a negative trend forward for some years.
Conservation and Rate Impacts. Conservation and rate impacts usually

require another level of detail in the analysis. When conservation programs are
developed primarily for residential customers, for example, it is necessary to go some
part of the way toward a sectoral water-demand analysis to estimate water savings.
Similarly, sectoral water use information is necessary to evaluate rate changes that are
differentiated by customer class. Such analysis by parts then can be worked back into
a per capita water-demand forecast.
Across-the-board rate changes that affect all customers can be integrated

directly into per capita water-demand forecasts. For example, consider a 20 percent
across-the-board rate increase, which has an impact on all customers of a water
system. In this case, research on system price elasticities provides a quick way to make
adjustments. By definition, a price (or income) elasticity indicates the percentage
change in water sales that is likely to result from a given percentage change in price.
Accordingly, a system price elasticity of –0.15 means that, other things being equal, a
20 percent price increase will result in a 3 percent reduction in per capita water use.
(See chapter 9 and Tables 9-5 and 9-9 through 9-12, which summarize information
on urban water price elasticities.)
SECTORAL WATER-DEMAND FORECASTS

The fundamental equation of sectoral water demand is a sum of the water use of the
various customer classes, plus sales outside the primary service area and water losses.
This can be written as
Q t = Q lt + Q 2 t + … + Q kt + X t + L t (Eq 5-6)
where:
Qjt = total water demand of customer group j in time period
j = 1, 2,...,k
Xt = sales of water to customers outside the utility service
area in time period t
Lt = overall system water losses in time period t
Many of the demand components in Equation 5-6 are forecast as a use rate
multiplied by the number of customers, as in the following formula:
Q jt = N jt q jt (Eq 5-7)
where:
Qjt = total water demand of customer group j in time period t
Njt = number of customers in customer group j in time period t
qjt = water-use rate or unit-use coefficient of customer
group j in time period t
Forecasts of water sales to customers outside the primary water utility’s service
area, Xt , usually require special analysis. Such sales can be contractually determined,
or they may require independent forecasts based on expected growth of outside-city
customers and their demands.
System water losses, Lt , have traditionally been calculated as a percentage of

total production or demand, but competing methods exist.2 In metered water
systems, forecasters can estimate water losses using utility records that show the
difference between water production and billed water sales. Frequently, water losses
vary between 10 and 20 percent of total water volumes passing through the
distribution system. Water losses include both water leakage from the distribution
system and water delivered to and used by customers without producing revenue
because of metering errors, accounting glitches, or other issues. Such water losses are
sometimes referred to as unaccounted-for water.
Sectoral water-demand forecasts are usually developed in two stages, as Figure
5-1 illustrates in simplified terms. In the first stage, analysts project current use rates
based on current conservation loading and rate levels. This produces a reference water-
demand forecast, which is usually a straightforward extrapolation of current condi-
tions in the water system. Earlier water planning studies call this the “unconstrained
forecast.”
The second stage leads to the final or adjusted sectoral demand forecast. Working
off the reference forecast, the forecaster adds (or subtracts) adjustments, such as water
savings from future conservation or impacts of changes in rate levels. For water
conservation, estimates of water savings are usually informed by end-use models of
indoor water use. Impacts of rate changes are typically calibrated with water-demand
price elasticities. In this second stage, the analyst may also consider adjustments to
other basic parameters in the reference water-demand forecast.
This two-step or phased procedure is mainly a convenience. The number of
parameters to be estimated grows rapidly with the number of customer groups, utility
programs, and policies. The two-step procedure establishes an initial benchmark
forecast based on the simplest setup, with programs and policies being added in the
second step. This helps identify programs and developments with the biggest impacts.
Example Data for a Sectoral Water-Demand Analysis. Table 5-3 presents
data for an example water system with two customer classes—residential and CI. The
period of record extends from 1993 to 2007.
The information is organized in three general blocks under the headings “Use
Rates,” “Customers,” and “Water Demand.” The residential and CI water-use rates
listed in Columns 1 and 2 are measured in gallons per customer per day (gcd). For
residential customers, this measures total household use in contrast to gallons per
capita per day (gpcd), which is gcd divided by the number of household members.
For example, if the average family has three members, residential consumption might
by 348 gcd, which is 116 gpcd. The use rates in Table 5-3 are ratios of total annual
2. Reporting simple percentages as a measure of water losses has been criticized on the grounds that
superior volumetric measures, taking into account the specific infrastructure setup of a water system,
can now be developed (AWWA Water Loss Control Committee 2003). The idea is that unavoidable
water losses are related to the length and type of water-transmission lines, connections, and other fea-
tures of the water-system infrastructure that may not be comparable across the topographies and geo-
graphic extent of different urban areas.
Water sales Customer

by customer counts Conservation programs
class (history) (history)
Project
Calculate
number of
use rates
customers
New
Retrofit
construction
Project
use rates
Estimate first round

total water use Revised forecast of water use
per customer class per cutomer class
Figure 5–1 Simplified Flow of Work Diagram for Sectoral Water-Demand Forecast
water sales to these customer classes, divided by the average annual number of
customers in each class. This information is presented in columns 3, 4, 5, and 6. The
residential water-use rate, for example, equals total residential water sales divided by
the number of residential customers.
Columns 3 and 4 list the average number of customers by year in these
customer groupings. The final data block (Columns 5 through 9) lists total water
demand by major demand component, subtotals, and totals.
Simplifying assumptions—including universal metering and a residential cus-
tomer class comprising only single-family residences—keep this discussion manage-
able. In addition, CI customers are combined into a single customer group, assumed
to also include institutional and governmental water customers.
Our example water utility is growing fairly rapidly, with a compound average
growth rate (CAGR) for population of 2.4 percent over the period from 1993 to 2007.
At 41 percent and 45 percent, respectively, there are roughly equal percentages of
residential and CI water use in ratio to total annual water usage in 2007. Water losses
make up 14.1 percent of total annual water use in 2007.
Data in Table 5-3 are from the water utility producing the gpcd numbers and
the populations displayed in Table 5-2. So, for example, the total system water usage
in 2007 is 28.3 mgd, implying a 282.7 gpcd when divided by the 2007 population of
100,000.
Table 5–3 Data for Sectoral Water-Demand Forecast
Date Use Rates Customers Water Demand

Subtotal
Residential Commercial/ Commerical/ Metered Total System
Use Rate Industrial Use Commerical/ Residential Industrial Usage Leakage Demand
(gcd) Rate (gcd) Residential Industrial (mgd) (mgd) (mgd) (mgd) (mgd)
Year (1) (3) (4) (6) (7) (8) (9)
(2) (5)
1993 245.9 3,319.3 27,381 3,306 6.7 11.0 17.7 2.0 19.7
1994 245.0 3,298.7 27,923 3,382 6.8 11.2 18.0 2.0 20.0
1995 246.4 3,264.7 28,479 3,459 7.0 11.3 18.3 2.4 20.7
1996 250.1 3,235.4 29,480 3,535 7.4 11.4 18.8 2.1 20.9
1997 235.0 3,168.3 30,519 3,612 7.2 11.4 18.6 2.6 21.2
1998 244.0 3,157.4 31,752 3,688 7.7 11.6 19.4 2.7 22.1
1999 244.5 3,141.9 33,199 3,764 8.1 11.8 19.9 2.8 22.7
2000 240.6 3,104.0 34,548 3,841 8.3 11.9 20.2 3.3 23.5
2001 269.4 3,134.6 36,304 3,955 9.8 12.4 22.2 3.0 25.2
2002 252.4 3,075.4 38,520 4,031 9.7 12.4 22.1 4.2 26.3
2003 245.8 3,026.7 40,098 4,107 9.9 12.4 22.3 4.4 26.7
2004 270.5 3,064.8 40,934 4,119 11.1 12.6 23.7 4.4 28.1
2005 258.8 3,007.4 41,586 4,164 10.8 12.5 23.3 3.9 27.2
2006 257.1 2,984.9 42,884 4,241 11.0 12.7 23.7 4.9 28.6
2007 258.5 3,012.1 44,444 4,232 11.5 12.7 24.2 4.0 28.3
Two-Step Computation Method. Following Figure 5-1, the first step (which
leads to the reference forecast) is to project water use by customer grouping, given
current or average levels of water use and existing trends in the numbers of customers.
At this point, the analyst projects water losses—not shown in Figure 5-1—and any
other components, such as outside sales. Existing conservation is not subject to a
separate analysis, but is implicitly included because it is reflected in the numbers used
for average levels of water use. Water commodity charges and fees are projected
forward at their current level.3
In a second round of analysis, the forecaster incorporates new or expanding
programs and policies, as well as any adjustments to basic parameters. This leads to
the final or adjusted sectoral demand forecast. A major objective of the adjusted
demand forecast is to determine which changes in programs, policies, and parameters
make the biggest difference to the system water-demand forecast.
3. Because annual inflation has ranged from 2 to 6 percent over the last two decades, rates must be
projected as constant in real, inflation-adjusted terms. Otherwise, analysts must include adjustments to
water-demand levels in the analysis because consumers see a fixed or nominal rate for water having less
and less impact on the cost of living. In short, maintaining a constant real or inflation-adjusted water
rate implies periodic increases in nominal rates to catch up with inflation.
Accordingly, some results in this second round of analysis are summarized in

terms of water savings. Identifying water savings from conservation programs helps
answer questions such as: How much additional water over the planning horizon
should be forthcoming from a proposed, new water conservation program? What
types of changes in water demand would be implied by a proposed increase in the real
cost of water to customers in the system?
The Reference Forecast. Table 5-4 shows the reference forecast, developed
with simple assumptions. This table extends the data to include the year-by-year
forecasts from 2008 to 2037.
A quick glance at the table shows that the residential and CI use rates are fixed
at 258.2 gcd and 3,019.1 gcd from 2008 forward. These are averages for the last five
years of the historical data (2003–2007). Averaging over several years tends to smooth
special weather effects such as the impacts of unusually hot and dry or cool and wet
years. The focus on the recent past also ensures that more contemporary develop-
ments and features predominate in the numbers used for forecasting.
The growth in residential customers from 2008 to 2037 is projected at the
population growth rate of 2 percent. The numbers of CI customers increase from
1993 to 2007, but at a rate less than the population growth rate. The simplest
assumption, therefore, is that this trend, which is consistent and essentially linear,
continues into the future. Accordingly, the forecaster projects numbers of CI
customers in the reference forecast with a linear trend, based on their pattern over the
period of historical record. Later in this section, we discuss other options, such as
predicting water use of the CI customer class based on employment projections.
From 2008 forward, the analyst projects water losses at a 15 percent rate, based
on management initiatives that will target this level of performance.
Column 9 of Table 5-4 gives the system-wide water-demand forecast. Beginning
at 29.4 mgd in 2008, the reference forecast projections rise to 47.4 mgd in 2037. This
2037 total is about 8 percent less than the 2037 total of 51.2 mgd from the per capita
forecast. The ratio of residential and CI customers shifts from 1993 to 2007 and
continues to decline over the forecast horizon. Clearly, more residential customers are
being served and employed in this community by a less rapidly growing number of CI
customers. This might indicate the spread of bigger stores and industrial concerns,
which must—at the same time—become more efficient water users because CI use
rates are projected ahead at a fixed value (3,013.6 gcd).
The Adjusted Sectoral Forecast. The following presents a relatively detailed
discussion of the adjusted sectoral forecast, followed by a list of further refinements
that seem plausible in this forecasting situation. The reader can follow the arguments
and computations as an example of the kind of thinking that typically goes into
formulating long-range sectoral water-demand forecasts. Table 5-5 shows the final or
adjusted sectoral water-demand forecast for our example utility.
Table 5-5 involves (1) estimating water savings from a generic conservation
program and (2) incorporating plans for more aggressive control of water losses, with
a target of no more than 12 percent a few years into the forecast horizon. These new
programs lead to water savings and demand reductions of approximately 8.4 percent,
compared with the reference forecast. Additionally, the 2037 total system water
demand in the final sectoral forecast is more than 17 percent less than the demand in
the per capita forecast.
There now are 10 numbered columns, in addition to the calendar year. The first
column of Table 5.5 lists persons per household, which can be calculated based on
population and numbers of residential customers, because residential customers are
assumed to be all single-family residences in this system. Column 6 now tabulates the
projected impact of a mandate and rebate conservation program, which we describe
in the sections that follow. After 2008, the entries in this column are negative because
they represent water savings rather than water demand.
Note that including Column 6 renders Column 5 a gross-instead of net-value
for residential water demand. After 2007, total residential water demand or usage is
calculated as the sum of Columns 6 and 7. We have not presented this sum as a
subtotal in Table 5-5 to economize on display space on the page.
Estimating water savings from a generic water conservation program. Some water
conservation is automatic because it is related to national water-efficiency require-
ments in the United States.4 For communities with a large customer base in 1994 and
low growth since then, continued retrofitting or replacement of old appliances and
fixtures may exert continuing downward pressure on water use for years into the
future.
In this example, the water system has grown substantially since 1994, when it
served 27,923 residential customers and 3,382 CI customers. By 2008, we assume that
updating older structures will have caused almost all customers to replace their older
plumbing fixtures, based on evidence from building permits. If a residual exists, we
can estimate its drawdown using an approach similar to that described in Table 5-6.
In general, water conservation is included in sectoral water-demand forecasts
through information produced by end-use models and forecasts. Future water savings
often accrue through retrofits and new construction. Existing customers can be
encouraged or may be obliged to install (retrofit) more water-efficient plumbing,
fixtures, and equipment, and to cultivate less water-intensive landscaping. Water-use
efficiency of new construction, on the other hand, may be enhanced by codes and
requirements.
To illustrate the basic approach, Table 5-6 is a spreadsheet that analyzes a
generic water-conservation program for residential customers, estimated to save
35 gallons per residential customer per day (gcd). We assume that this water utility
launches this mandate and rebate program in 2008. New residences are mandated to
install the new water-saving technology. In addition, rebates serve as incentives for
existing customers to retrofit their households with this new conservation technology.
4. The Energy Policy Act of 1992 (EPAct) establishes maximum water-use rates for toilets (1.6 gal-
lons per flush), urinals (1.0 gallons per flush), showerheads (2.5 gallons per minute [gpm]), and faucets
(2.5 gpm). These standards, which apply to plumbing fixtures in new and renovated residential and
nonresidential facilities, became effective in 1994, operating through manufacturers on the supply side.
Table 5–4 Data For Sectoral Water-Demand Forecast, Reference Forecast
Date Use Rates Customers Water Demand

Subtotal
Residential Commercial/ Commerical/ Metered Total System
Use Rate Industrial Use Commerical/ Residential Industrial Usage Leakage Demand
(gcd) Rate (gcd) Residential Industrial (mgd) (mgd) (mgd) (mgd) (mgd)
Year (1) (3) (4) (6) (7) (8) (9)
(2) (5)
1993 245.9 3,319.3 27,381 3,306 6.7 11.0 17.7 2.0 19.7
1994 245.0 3,298.7 27,923 3,382 6.8 11.2 18.0 2.0 20.0
1995 246.4 3,264.7 28,479 3,459 7.0 11.3 18.3 2.4 20.7
1996 250.1 3,235.4 29,480 3,535 7.4 11.4 18.8 2.1 20.9
1997 235.0 3,168.3 30,519 3,612 7.2 11.4 18.6 2.6 21.2
1998 244.0 3,157.4 31,752 3,688 7.7 11.6 19.4 2.7 22.1
1999 244.5 3,141.9 33,199 3,764 8.1 11.8 19.9 2.8 22.7
2000 240.6 3,104.0 34,548 3,841 8.3 11.9 20.2 3.3 23.5
2001 269.4 3,134.6 36,304 3,955 9.8 12.4 22.2 3.0 25.2
2002 252.4 3,075.4 38,520 4,031 9.7 12.4 22.1 4.2 26.3
2003 245.8 3,026.7 40,098 4,107 9.9 12.4 22.3 4.4 26.7
2004 270.5 3,064.8 40,934 4,119 11.1 12.6 23.7 4.4 28.1
2005 258.8 3,007.4 41,586 4,164 10.8 12.5 23.3 3.9 27.2
2006 257.1 2,984.9 42,884 4,241 11.0 12.7 23.7 4.9 28.6
2007 258.5 3,012.1 44,444 4,232 11.5 12.7 24.2 4.0 28.3
2008 258.2 3,019.2 45,333.3 4,396 11.7 13.3 25.0 4.4 29.4
2009 258.2 3,019.2 46,240.0 4,467 11.9 13.5 25.4 4.5 29.9
2010 258.2 3,019.2 47,164.8 4,538 12.2 13.7 25.9 4.6 30.4
2011 258.2 3,019.2 48,108.1 4,609 12.4 13.9 26.3 4.6 31.0
2012 258.2 3,019.2 49,070.3 4,680 12.7 14.1 26.8 4.7 31.5
2013 258.2 3,019.2 50,051.7 4,751 12.9 14.3 27.3 4.8 32.1
2014 258.2 3,019.2 51,052.7 4,822 13.2 14.6 27.7 4.9 32.6
2015 258.2 3,019.2 52,073.8 4,893 13.4 14.8 28.2 5.0 33.2
2016 258.2 3,019.2 53,115.2 4,964 13.7 15.0 28.7 5.1 33.8
2017 258.2 3,019.2 54,177.5 5,035 14.0 15.2 29.2 5.2 34.3
2018 258.2 3,019.2 55,261.1 5,106 14.3 15.4 29.7 5.2 34.9
2019 258.2 3,019.2 56,366.3 5,177 14.6 15.6 30.2 5.3 35.5
2020 258.2 3,019.2 57,493.6 5,248 14.8 15.8 30.7 5.4 36.1
2021 258.2 3,019.2 58,643.5 5,319 15.1 16.1 31.2 5.5 36.7
2022 258.2 3,019.2 59,816.4 5,389 15.4 16.3 31.7 5.6 37.3
2023 258.2 3,019.2 61,012.7 5,460 15.8 16.5 32.2 5.7 37.9
2024 258.2 3,019.2 62,233.0 5,531 16.1 16.7 32.8 5.8 38.5
2025 258.2 3,019.2 63,477.6 5,602 16.4 16.9 33.3 5.9 39.2
2026 258.2 3,019.2 64,747.2 5,673 16.7 17.1 33.8 6.0 39.8
2027 258.2 3,019.2 66,042.1 5,744 17.0 17.3 34.4 6.1 40.5
2028 258.2 3,019.2 67,362.9 5,815 17.4 17.6 34.9 6.2 41.1
2029 258.2 3,019.2 68,710.2 5,886 17.7 17.8 35.5 6.3 41.8
2030 258.2 3,019.2 70,084.4 5,957 18.1 18.0 36.1 6.4 42.4
2031 258.2 3,019.2 71,486.1 6,028 18.5 18.2 36.7 6.5 43.1
2032 258.2 3,019.2 72,915.8 6,099 18.8 18.4 37.2 6.6 43.8
2033 258.2 3,019.2 74,374.1 6,170 19.2 18.6 37.8 6.7 44.5
2034 258.2 3,019.2 75,861.6 6,241 19.6 18.8 38.4 6.8 45.2
2035 258.2 3,019.2 77,378.9 6,312 20.0 19.1 39.0 6.9 45.9
2036 258.2 3,019.2 78,926.4 6,382 20.4 19.3 39.6 7.0 46.6
2037 258.2 3,019.2 80,505.0 6,453 20.8 19.5 40.3 7.1 47.4
NOTE: gcd is water use by user class divided by number of billed customers in that class.
Table 5–5 Data for Sectoral Water-Demand Forecast, Adjusted Forecast

Use Rates Customers Water Demand
Subtotal Total
Residential Commercial/ Commerical/ Metered System
Use Rate Industrial Use Commerical/ Residential Water Industrial Usage Leakage Demand
Persons/ (gcd) Rate (gcd) Residential Industrial (mgd) Savings (mgd) (mgd) (mgd) (mgd)
hh Year (1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
2.6 1993 245.9 3,319.3 27,381 3,306 6.7 11.0 17.7 2.0 19.7
2.575 1994 245.0 3,298.7 27,923 3,382 6.8 11.2 18.0 2.0 20.0
2.55 1995 246.4 3,264.7 28,479 3,459 7.0 11.3 18.3 2.4 20.7
2.525 1996 250.1 3,235.4 29,480 3,535 7.4 11.4 18.8 2.1 20.9
2.5 1997 235.0 3,168.3 30,519 3,612 7.2 11.4 18.6 2.6 21.2
2.475 1998 244.0 3,157.4 31,752 3,688 7.7 11.6 19.4 2.7 22.1
2.45 1999 244.5 3,141.9 33,199 3,764 8.1 11.8 19.9 2.8 22.7
2.425 2000 240.6 3,104.0 34,548 3,841 8.3 11.9 20.2 3.3 23.5
2.4 2001 269.4 3,134.6 36,304 3,955 9.8 12.4 22.2 3.0 25.2
2.375 2002 252.4 3,075.4 38,520 4,031 9.7 12.4 22.1 4.2 26.3
2.35 2003 245.8 3,026.7 40,098 4,107 9.9 12.4 22.3 4.4 26.7
2.325 2004 270.5 3,064.8 40,934 4,119 11.1 12.6 23.7 4.4 28.1
2.3 2005 258.8 3,007.4 41,586 4,164 10.8 12.5 23.3 3.9 27.2
2.30 2006 257.1 2,984.9 42,884 4,241 11.0 12.7 23.7 4.9 28.6
2.25 2007 258.5 3,012.1 44,444 4,232 11.5 12.7 24.2 4.0 28.3
2.24 2008 258.2 3,019.2 45,333.3 4,396 11.7 (0.1) 13.3 24.9 4.4 29.3
2.23 2009 258.2 3,019.2 46,240.0 4,467 11.9 (0.2) 13.5 25.3 4.5 29.7
2.22 2010 258.2 3,019.2 47,164.8 4,538 12.2 (0.2) 13.7 25.6 3.5 29.1
2.21 2011 258.2 3,019.2 48,108.1 4,609 12.4 (0.3) 13.9 26.0 3.5 29.6
2.2 2012 258.2 3,019.2 49,070.3 4,680 12.7 (0.4) 14.1 26.4 3.6 30.0
2.19 2013 258.2 3,019.2 50,051.7 4,751 12.9 (0.5) 14.3 26.8 3.7 30.5
2.18 2014 258.2 3,019.2 51,052.7 4,822 13.2 (0.5) 14.6 27.2 3.7 30.9
2.17 2015 258.2 3,019.2 52,073.8 4,893 13.4 (0.6) 14.8 27.6 3.8 31.4
2.16 2016 258.2 3,019.2 53,115.2 4,964 13.7 (0.7) 15.0 28.0 3.8 31.8
2.15 2017 258.2 3,019.2 54,177.5 5,035 14.0 (0.7) 15.2 28.4 3.9 32.4
2.14 2018 258.2 3,019.2 55,261.1 5,106 14.3 (0.8) 15.4 28.8 3.9 32.8
2.13 2019 258.2 3,019.2 56,366.3 5,177 14.6 (0.9) 15.6 29.2 4.0 33.2
2.12 2020 258.2 3,019.2 57,493.6 5,248 14.8 (1.0) 15.8 29.7 4.0 33.7
2.11 2021 258.2 3,019.2 58,643.5 5,319 15.1 (1.0) 16.1 30.1 4.1 34.3
2.1 2022 258.2 3,019.2 59,816.4 5,389 15.4 (1.1) 16.3 30.5 4.2 34.8
2.09 2023 258.2 3,019.2 61,012.7 5,460 15.8 (1.2) 16.5 31.0 4.2 35.3
2.08 2024 258.2 3,019.2 62,233.0 5,531 16.1 (1.3) 16.7 31.4 4.3 35.8
2.07 2025 258.2 3,019.2 63,477.6 5,602 16.4 (1.3) 16.9 31.9 4.3 36.3
2.06 2026 258.2 3,019.2 64,747.2 5,673 16.7 (1.4) 17.1 32.4 4.4 36.9
2.05 2027 258.2 3,019.2 66,042.1 5,744 17.0 (1.5) 17.3 32.9 4.5 37.3
2.04 2028 258.2 3,019.2 67,362.9 5,815 17.4 (1.5) 17.6 33.4 4.5 40.0
2.03 2029 258.2 3,019.2 68,710.2 5,886 17.7 (1.6) 17.8 33.9 4.6 38.5
2.02 2030 258.2 3,019.2 70,084.4 5,957 18.1 (1.7) 18.0 34.4 4.7 39.1
2.01 2031 258.2 3,019.2 71,486.1 6,028 18.5 (1.7) 18.2 34.9 4.8 39.7
2 2032 258.2 3,019.2 72,915.8 6,099 18.8 (1.8) 18.4 35.5 4.8 40.3
1.99 2033 258.2 3,019.2 74,374.1 6,170 19.2 (1.8) 18.6 36.0 4.9 40.9
1.98 2034 258.2 3,019.2 75,861.6 6,241 19.6 (1.9) 18.8 36.5 5.0 41.5
1.97 2035 258.2 3,019.2 77,378.9 6,312 20.0 (1.9) 19.1 37.1 5.1 42.2
1.96 2036 258.2 3,019.2 78,926.4 6,382 20.4 (2.0) 19.3 37.7 5.1 42.8
1.95 2037 258.2 3,019.2 80,505.0 6,453 20.8 (2.0) 19.5 38.2 5.2 43.4
Table 5–6 Water Savings From Mandate and Rebate Conservation Program
Water Savings
From New Customers
New Residential Participating Water Savings
Residential Customers in Retrofit From Retrofit Total Water
Customers (gpd) Program (gpd) Savings (mgd)
Year (1) (2) (3) (4) (5)
2008 888.9 31,111.1 1,333.3 46,666.7 0.08
2009 1,795.6 62,844.4 2,626.7 91,933.3 0.15
2010 2,720.4 95,212.4 3,881.2 135,842.0 0.23
2011 3,663.7 128,227.8 5,098.1 178,433.4 0.31
2012 4,625.8 161,903.5 6,278.5 219,747.1 0.38
2013 5,607.2 196,252.7 7,423.5 259,821.3 0.46
2014 6,608.3 231,288.8 8,534.1 298,693.4 0.53
2015 7,629.3 267,025.7 9,611.4 336,399.2 0.60
2016 8,670.8 303,477.3 10,656.4 372,973.9 0.68
2017 9,733.1 340,658.0 11,670.0 408,451.4 0.75
2018 10,816.6 378,582.3 12,653.3 442,864.5 0.82
2019 11,921.9 417,265.0 13,607.0 476,245.2 0.89
2020 13,049.2 456,721.4 14,532.1 508,624.5 0.97
2021 14,199.1 496,967.0 15,429.5 540,032.5 1.04
2022 15,371.9 538,017.4 16,299.9 570,498.2 1.11
2023 16,568.3 579,888.9 17,144.3 600,049.9 1.18
2024 17,788.5 622,597.8 17,963.3 628,715.0 1.25
2025 19,033.2 666,160.8 18,757.7 656,520.3 1.32
2026 20,302.7 710,595.2 19,528.3 683,491.3 1.39
2027 21,597.7 755,918.2 20,275.8 709,653.2 1.47
2028 22,918.5 802,147.6 21,000.9 735,030.3 1.54
2029 24,265.8 849,301.7 21,074.2 759,646.1 1.61
2030 25,640.0 897,398.9 22,222.2 777,777.8 1.68
2031 27,041.7 946,457.9 22,222.2 777,777.8 1.72
2032 28,471.4 996,498.2 22,222.2 777,777.8 1.77
2033 29,929.7 1,047,539.3 22,222.2 777,777.8 1.83
2034 31,417.2 1,099,601.2 22,222.2 777,777.8 1.88
2035 32,934.4 1,152,704.3 22,222.2 777,777.8 1.93
2036 34,482.0 1,206,869.5 22,222.2 777,777.8 1.98
2037 36,060.5 1,262,118.0 22,222.2 777,777.8 2.04
NOTE: Assumed percent of existing customers opting to participate in the conservation program each
year—3%. Ceiling to percentage of existing customers who will participate—50%.
A key parameter in Table 5-6 is the percentage of existing residences (as of

2007) that will opt to participate in the rebate program each year. Based on market
research, we assume 3 percent per year. The survey research also suggests that only a
certain percentage of existing residential customers—50 percent—will ever consider
participating in the conservation program. This establishes a ceiling to participation
of existing customers in this mandate and rebate program.
Column 1 of Table 5-6 was developed from the assumption of 2 percent

population growth and the 2007 total of 44,444 residential customers. Each year after
the beginning of the generic conservation program in 2008, we add more residential
customers to the total new residential customers from the previous year to arrive at
the totals in Column 1.
Column 2 is simply the estimated conservation effectiveness of 14 gcd
multiplied by the total number of new residential customers.
Column 3 is an estimate of the numbers of customers participating in the rebate
program, which was computed using a simple recursion formula in the spreadsheet.
An estimated 1,333 customers elect to participate in the rebate program in 2008, the
first year of the new conservation program. This represents three percent of the 44,444
residential customers in 2007. For every subsequent year, the total number of
customers participating in the rebate program during the previous year is subtracted
from 44,444. The difference is the pool of customers eligible in that year for the rebate
program. In 2007, this pool is composed of existing customers who have not already
participated in the rebate program. This pool of customers is then multiplied by
3 percent and the product is added to the number of customers participating in the
previous year. In 2009, then, three percent of (44,444 – 1,333) = 1,293 customers, and
1,293 + 1,333 = 2,627 (after allowing for the fractional parts of these numbers and
rounding up). Interestingly, this is a standard method for projecting a diffusion
process, which we discuss in more detail in chapter 11.
Table 5-6 highlights the gradual impact of these programs. It shows that water
technology that saves 35 gallons per residential customer per day achieves 13.5 percent
water savings against the base residential use rate of 258.2 gcd. For households
enjoying this new and more water-efficient equipment, this is a significant percentage.
Note, however, that the overall water savings at the end of the forecast horizon
are only 10.6 percent of total residential water usage. This is calculated as 2.0/(20.8 –
2.0), transformed into a percentage.
Adding control of water losses. From 1993 to 2007, the percentage of total system
water production going to water losses increases from around 10 to 15 percent or
more. For the forecast, we assume that the management of this water utility is
committed to leak detection and repair that is designed to constrain annual system
water losses to no more than 12 percent. We assume that this program becomes fully
effective in 2010.
Including rate impacts in the sectoral forecast. Impacts of planned changes in
water commodity charges or rates are also incorporated in this second stage of
analysis, although we did not calibrate any for this example. The analysis of rate
impacts depends on whether the rate change is across the board for all customers, or
whether it applies to, for example, only residential customers.
The rate changes explicitly included in a long-term water-demand forecast are
those that are designed to reflect long-term changes in the real, inflation-adjusted
costs faced by customers. The basic economics principle here is that consumers
respond to real, inflation-adjusted prices. Because some inflation is almost always in
play, the nominal or dollar value of water rates must be adjusted periodically over the
planning horizon in order to keep the real costs of water constant. Otherwise,
customer water use will inch up, as the real costs of water erode in inflation-adjusted
terms.
Consequently, explicit inclusion of an increase in utility rates in a long-term
water-demand forecast is likely to be associated with the water system moving into a
new cost structure, such as might be associated with depletion of one supply source
and substitution of another. Another option in some water systems is moving to
higher real water rates to encourage water conservation and manage demand. In
chapter 9, we discuss how to adjust residential demand estimates with elasticities
derived from the extensive water-demand literature.
Other options and adjustments. The preceding discussion presents a

“plain vanilla,” “bare bones” sectoral water-demand forecast model. Although this
serves our exposition here, utilities might well consider additional options and
adjustments
• Trending residential water-use rates
• Trending CI water-use rates
• Breaking out and separating estimates of large CI customers’ future water
use
• Trending CI water use based on employment forecasts
• Incorporating forecasts of changing household composition in the form of
changes in household size or persons per household
We give more detail on these options in the subsections that follow Figure 5-2.
Graphing the data is always a good idea in evaluating forecasting tactics. Figure
5-2 charts the residential and CI water-use rates from 1993 to 2007. The graph has
two vertical axes, one for CI rates (left-hand axis) and the other for residential use
rates (right-hand axis). CI water-use averages over 3,000 gcd while residential use
ranges from 210 to approximately 280 gcd.
Trending residential water-use rates. At first glance, there appears to be a

positive trend to the residential use rate and a negative or downward-sloping trend to
the CI use rate. Further analysis, however, suggests that changes in the annual averages
for temperature and precipitation largely explain the apparent positive trend in
residential use rates (see Example 5-1 on the data CD). There could even be a basis for
trending the residential water-use rate downward over the forecast horizon—if the
demographic experts predict continuing drops in persons per household in this
system. Other things being equal, reduced household size will reduce indoor water use
along the lines suggested by end-use modeling.
Trending commercial/industrial use rates. Regression analysis confirms a

highly significant negative trend to CI water use (1993–2007). This analysis also shows
that weather influences on CI water use are weak to nonexistent.
3,400 280
3,300 270
260
3,200
250 Commercial/Industrial
Rate
3,100
Residential Use Rate
240
3,000
230
2,900 220
2,800 210
93
95
97
99
01
03
05
07
19
19
19
19
20
20
20
20
NOTE: Left vertical scale = CI gallons per day per connection.

Right vertical scale = residential gallons per capita per day.
Figure 5–2 Historical Residential and CI Water-Use Rates
Before we project this negative trend out 30 years, it is a good idea to explore its
basis. For example, it was discovered that several large, water-intensive processing
facilities left this service area during the period of record, and that the remaining
concerns are installing more efficient (and profitable) water-using technologies.
Additionally, continuing gains in water-use efficiency are certainly possible and are
likely to occur in this customer group.
Accordingly, we can project the CI use rate with a trending equation selected
from the options contained in Microsoft Excel. In this case, the equation of best fit is
nonlinear. It is a power relationship of the form
COMM/INDt = 3,385.8t – 0.042 (Eq 5-8)
where:
COMM/INDt is the CI water-use rate and t is the time index,
which ranges from 1 in 1993 to 45 in 2037. This generates a small
reduction—on the order of 1 percent—in the CI water-use rate
Over The 30-year Forecast Horizon.
Breaking out and separating estimates of large CI customers’ future

water use. In many systems, a few CI customers account for very large water
sales. It is good forecasting practice to break these out and to attempt to develop
independent forecasts of their future water usage. In some cases, engineers and
managers at these concerns will be the best source of information on planned water
use in the future, including in-house investments in water efficiency and recycling.
Trending CI water use based on employment forecasts. The forecasts
of CI water use or its major components are based on employment forecasts. This
usually involves calculating CI water use per employee from historical data and
coupling water use per employee with employment projections.
This approach is most suitable for larger metropolitan areas, where it has seen
recent application. Usually, an area planning organization is charged with developing
population forecasts for jurisdictions of the metropolitan area. These allocations of
metropolitan-area growth support applications for federal and state grants for
infrastructure projects by jurisdictions within the metro area. Usually, these area
planning organizations develop employment forecasts in addition to population
forecasts for subregional areas.
Research on CI use suggests that water use per employee is highly variable, and
subcategories, such as hotels and motels, use water in patterns essentially independent
of the number of employees. For these reasons, higher levels of disaggregation of CI
use—down to levels indicating very specific types of businesses—are often recom-
mended. As an alternative, in large metropolitan urban areas, the analyst can invoke
the law of large numbers to argue that this variability tends to “average out” over the
longer run.
Incorporating forecasts of changing household composition in the form
of changes in household size, or persons per household. Finally, forecasts
of household size can be incorporated. Demographic authorities project aging of the
US population while millions of new, younger immigrants are arriving to fill out the
younger age categories. With an older population, as well as with recent patterns that
favor single-parent and single-person households, expectations are that the average
persons per household will decline in many urban areas. If household size is forecast
to decrease, one issue takes on immediacy—making such household forecasts
consistent with the overall population forecast
Comparisons. Table 5-7 compares the per capita system-demand forecast;
the reference sectoral water-demand forecast; and the adjusted, final sectoral water-
demand forecast. Total system water use rises from around 29 mgd in 2008 to
47.5 mgd in 2037 (the year of the final forecast). This adjusted, final sectoral water-
demand forecast is approximately 9 percent less than the simpler reference forecast,
and more than 17 percent less than the per capita forecast for 2037.
These lower water demand totals are based on
• A mandate-and-rebate type conservation program launched in 2008
• Projections of more aggressive water-loss detection and control anticipated
to reduce water losses to an average of 12 percent per year
• Growth of the number of CI customers at rates lower than the population
growth rate
Table 5–7 Comparing the Forecasts

Reference Adjusted, Final
Per Capita Sectoral Sectoral
Forecast Forecast Forecast
(mgd) (mgd) (mgd)
Year (1) (2) (3)
2008 28.8 29.4 29.3

2009 29.4 29.9 29.7
2010 30.0 30.4 29.1
2011 30.6 31.0 29.6
2012 31.2 31.5 30.0
2013 31.8 32.1 30.5
2014 32.5 32.6 30.9
2015 33.1 33.2 31.4
2016 33.8 33.8 31.8
2017 34.5 34.3 32.4
2018 35.2 34.9 32.8
2019 35.9 35.5 33.2
2020 36.6 36.1 33.7
2021 37.3 36.7 34.3
2022 38.0 37.3 34.8
2023 38.8 37.9 35.3
2024 39.6 38.5 35.8
2025 40.4 39.2 36.3
2026 41.2 39.8 36.9
2027 42.0 40.5 37.4
2028 42.8 41.1 38.0
2029 43.7 41.8 38.5
2030 44.6 42.4 39.1
2031 45.5 43.1 39.7
2032 46.4 43.8 40.3
2033 47.3 44.5 40.9
2034 48.3 45.2 41.5
2035 49.2 45.9 42.2
2036 50.2 46.6 42.8
2037 51.2 47.4 43.4
Total savings from the generic water-conservation program at the end of the
forecast horizon run at 2 mgd. In 2037, savings from more aggressive water-loss
control total 1.9 mgd. Straight-line trending of the growth in CI customers, consistent
with a strongly linear historic trend, accounts for additional differences in water use
of 3.8 mgd in 2037.
REALITY CHECKS
The challenging question is whether a water-demand forecast is “right.” This can be
difficult to immediately answer; however, it may become evident as time passes.
In the near term, there are several reality checks that help validate water-
demand forecasts. These include examining the general adequacy of the analysis, plus
the consistency of assumptions. It also is advisable to conduct some sensitivity
analysis to explore the types of change in water demand that are likely to accompany
modification or change in the basic assumptions of the analysis. It is preferable to
monitor changes in population, the numbers of customers, and water-demand levels
to determine whether the numbers are in agreement with the assumptions and
projections used in the water-demand forecast.
Adequacy of Analysis and Consistency of Assumptions. The water-
demand forecast should be comprehensive—in the sense of accounting for major
anticipated developments—and consistent. Toward this end, the cross-checking of
assumptions is an essential step. For example, arithmetic linkages can exist among
population numbers, numbers of customers, and numbers of persons per household.
Therefore, in a water system comprised largely of single-family residences, it should
be possible to estimate fairly closely the number of residential customers by dividing
the population forecast by the persons per household.
Other relevant links can be between employment forecasts and the number of
CI and institutional water customers. Finally, projections of indoor water-
conservation savings should be calibrated to assumptions about changes in average
household size over the forecast horizon.
Other checks can relate to more fundamental information about water-using
behavior. For example, are the projected conservation savings consistent with
commonly cited end-use numbers and the information on behavior relating to indoor
water use (i.e., the number of uses of an indoor appliance or fixture per person per
day)?
Sensitivity Analysis. Many engineering studies suggest a sensitivity analysis.
To conduct such an analysis, the forecaster might allow all the input parameters to
vary by about 15 percent and then study the impacts on the output variable in the
analysis—in this case, water-demand forecasts. A sensitivity analysis conducted along
such lines serves an important purpose. It highlights which input variables can cause
the biggest impacts on the output variable.
In chapter 13, a more evolved type of sensitivity analysis called risk or Monte
Carlo simulation is outlined. This methodology invites the analyst to specify or
determine the potential variability of key input factors in the water-demand forecast.
The forecaster can establish these ranges of variation by consulting with experts or
examining history, or through group processes (Delphi methods). Then, essentially,
hundreds or thousands of scenarios are calculated, based on random sampling of the
values of the input parameters. This gives rise to a risk distribution—a curve of the
potential variations of water demands in various periods of the planning horizon.
Risk analysis is a valuable reality check on the long-term water-demand

forecast. Basic input variables and parameters can vary over time. The forecaster
should ask what are the possible ranges of water demand in future years, consistent
with the understood potential for the input variables of the demand model to vary.
Monitoring Population, New Customers, and Water Demand. Changes
in population forecasts and unanticipated numbers of new customers are often the
largest sources of error in long-term water-demand forecasts. Retrospective studies of
urban water demand show that long-term population forecasts dating from the 1970s,
the 1980s, and the early 1990s were consistently higher than actual populations in
major metropolitan areas.
Water-use rates, which are the other component of the demand equation, can
be directly monitored. Demand monitoring is especially advisable when the water
system is facing a target or constraint in demand, such as commitments to achieve a
target level of water savings by a certain date, or a soft or hard limit on water capacity.
Many monitoring approaches are available to utilities. Most involve statistical
analysis of higher frequency water-use data and the consideration of weather impacts.
Segmenting demand in a water system in a sectoral forecast model helps to
focus demand monitoring. For example, if residential water-use rates shoot up several
years running, something is probably wrong with the conservation or trending
analysis of the residential water-use rates.
Specific, focused studies can also be supported at this stage. Many urban water
systems, for example, are interested in whether the typical new single-family
residences use more water than older homes. The evidence seems mixed, and
probably depends on circumstances (such as typical lot sizes) and on climatic factors
that are location based. This type of question, therefore, is ideal for a focused on-site
study in a utility service area, and the results could have great import for overall
water-demand forecasts.

This chapter outlines key steps for per capita and sectoral water-demand forecasts.
Sample 30-year water-demand forecasts were developed using coordinated data on
population growth, numbers of customers, water usage, and other factors, allowing
results to be compared.
In general, sectoral water-demand models are more complex but also have
several advantages. They provide opportunities to increase forecast accuracy and
create a better framework to consider what-ifs. Almost all evaluation of utility
programs and policies requires the analyst to drill down to detail greater than is
customary for per capita water-demand forecasting. Sectoral water-demand forecasts
also provide a superior framework for tracking water demand in the near term. In
general, maintaining a sectoral water-demand forecast is good business practice.
Nevertheless, as a first cut and for smaller water systems, per capita water-demand
forecasts yield useful information for long-term planning.
The water utility in this example has a population of between 50,000 and
200,000 over the period of study and the forecast horizon. As a simplification, only
two customer classes are considered—residential and CI. Furthermore, all residential
customers are assumed to consist of single-family residences. The example utility is
implementing a major water-conservation program and plans to improve its water-
loss record.
The discussion suggests developing sectoral water-demand forecasts in two
steps. The first step, which leads to the reference forecast, involves extrapolating
current customer use rates based on current conservation loading and rate levels. The
second step, which leads to the final demand forecast, adds (or subtracts) adjust-
ments, such as estimates of water savings from future conservation or impacts of
changes in rate levels. For water conservation, estimates of water savings are usually
informed by end-use models of indoor water use. Analysts typically calibrate impacts
of rate changes with water-demand price elasticities. In the second stage, the
forecaster may also consider the impact of adjustments to other basic parameters
from the first stage.
Data used in a sectoral water-demand forecast include historical data on water
sales to customers and historical customer counts, as well as information on the end-
use savings from conservation programs. Socioeconomic and demographic informa-
tion can also be important.
A major objective of the second analytical round is to determine which changes
in programs, policies, and parameters make the biggest difference to the system water-
demand forecast. The second stage is designed to answer the question—how does the
dynamic of the water system change when major players and forces change?
Sectoral water-demand forecasts often generate different results than per capita
forecasts. In this example, the differences between the two grow to nearly 17 percent
by the end of the forecast horizon. In our example, this growing gap is attributable to
a combination of factors, including a mandate-and-rebate conservation program,
more aggressive water-loss management, and trending assumptions applied to use
rates and customer-growth numbers. Only some of these factors are readily
incorporated in the per capita water forecast by, for example, deducting estimated
water savings from a residential water-conservation program or recalculating the
gallons per capita per year with assumed changes in the underlying rate of system
water losses.
REFERENCES
AWWA Water Loss Control Committee. 2003. Committee Report: Applying Worldwide BMPs in Water
Loss Control. Journal AWWA 95(8): 8.
Vickers, A. 2002. Handbook of Water Use and Conservation. Amherst, Mass.: WaterPlow Press.
CHAPTER 6
Forecasting Seasonal and

Peak Water Demand
This chapter focuses on techniques for identifying peak daily demands and the
seasonal and monthly pattern of water use.
Information about seasonal and peak demands has many practical applications.
Foremost among these is sizing treatment capacity and engineering the water
distribution system. Information about monthly and seasonal water use is also
valuable for cash-flow analysis, budgeting, and utility financial planning. In addition,
monthly or seasonal water-demand models support policy evaluation, helping utilities
identify the impacts of programs and policies on water usage.
The methods in this chapter run the gamut from simple and moving averages
to multivariate regressions based on weather and other variables. Formal statistical
models usually focus on monthly or seasonal variation, with analysts using relatively
simple methods to identify the peak-to-average day demand ratio. Information about
seasonal and monthly patterns, however, provides a valuable guide as to when peak-
day events are likely to occur.
Every water-system analyst’s tool kit should contain the simpler methods
discussed in this chapter. These include (1) calculating typical monthly proportions of
annual water usage, (2) estimating indoor and outdoor water use from summer and
winter water demands, when possible, and (3) identifying interannual trends in total
system water use with moving averages.
Regression modeling requires more training and time, but can provide valuable
support for policy analysis. Weather, of course, cannot be predicted with much
accuracy, except in terms of monthly, daily, or hourly averages. This means regression
models developed with real-time weather information may do a good job of
83
explaining past variation in water use without being able to achieve much success in
predicting future water use—except in terms of broad averages. Nevertheless, these
regression models of seasonal, monthly, or daily water use can play a role in policy
evaluation. For example, regression models incorporating weather variables can help
the analyst address how, on average, system water usage can be expected to change
with higher average temperatures projected by climate change.
The first major section of this chapter discusses issues in the computation of the
peak-to-average day demand ratio for a water system. Figure 6-1 illustrates a data
series for such a demand ratio.
2.8
2.6
Ratio of Peak-to-Average Day Demand
2.4
2.2
2.0
1.8
1.6
1.4
1.2
1.0
60
63
66
69
72
75
78
81
84
87
90
93
96
99
02
19
19
19
19
19
19
19
19
19
19
19
19
19
19
20
NOTE: Data collected from 1980 to 1999 in Calgary, Alberta.
Figure 6–1 Historical Peak-to-Average Day Demand
The next section considers simple and moving averages with monthly total
system water-use data. Topics include determining monthly proportions of total
annual water usage, distinguishing between outdoor and indoor water use, and
identifying trends with moving averages.
Figure 6-2 depicts the month-by-month proportions of annual water sales or
production and their historical upper and lower bounds in a large urban water
system. In the example system, July is usually the month of peak water usage.
This average annual pattern of water use for this system can be compared with
the more complex monthly water production series shown in Figure 6-3. These data,
a synthetic series based on features of water use in a Canadian city, show strong
seasonality and a slightly positive trend.
FORECASTING SEASONAL AND PEAK WATER DEMAND 85
0.1400
0.1200
0.1000
Lower Bound Proportions
Proportion
0.0800
Average Proportions of Total
Annual Water Production
0.0600 Upper Bound Proportions
0.0400
0.0200
0.0000
ne
Ap h
ce ber
M ry
r
ay
Se ug ly
br y
ril
Oc ber
De em r
em t
be
v e
pt us
Fe uar
Ju
ua
No tob
ar
M
Ju
m
n
Ja
Month
Figure 6–2 Average Monthly Proportions of Annual Water Production
80
75
70
65
Million Liters
60
55
50
45
40
35
30
:1 :8 :3 10 :5 12 :7 :2 :9 :4 11 :6 :1 :8 :3 10 :5 12
95 95 96 6: 97 7: 98 99 99 00 0: 01 02 02 03 3: 04 4:
19 19 19 199 19 199 19 19 19 20 200 20 20 20 20 200 20 200
Year and Month
Figure 6–3 Monthly Total System Water Use
The third major section of this chapter focuses on using statistical regression to
analyze monthly and daily summer water use. This is illustrated with examples from
a major urban water system. Regression components and their interpretation are
discussed, highlighting the importance of long-term average weather in explaining the
seasonal pattern of water use. The examples suggest that months of unusually high or
low water demand are largely explained by departures of temperature and precipita-
tion from their long-term norms.
To illustrate very-short-term forecast models, a daily water-use model is
presented. Such models, which have application to day-by-day system optimization,
can produce forward-looking information on daily peak demand.
The discussions of monthly and daily water-demand regressions in this chapter
are exploratory and are designed to highlight modeling issues and tactics. These
modeling tactics can be researched further in several of the references cited at the end
of the chapter.
Research on short-run water-demand modeling is increasingly “analysis-
intensive.” In addition to averaging and regression, there is a growing literature on
neural network, fuzzy logic, and ARIMA (autoregressive integrated moving average)
models applied to high-frequency water-use data. While these techniques are beyond
the scope of the discussion here, averaging and regression approaches are valuable for
exploratory data analysis in prelude to calibrating these more complex approaches.
PEAK-TO-AVERAGE DAY DEMAND RATIOS

One of the challenges in urban water-demand analysis is identifying an appropriate
value for the peak-to-average day demand ratio in a water system. This is an
important design parameter in water treatment and distribution. Analysts typically
focus on the peak-to-average day ratio rather than higher frequency peak-to-average
demand ratios (such as a peak-to-average hour ratio). Given adequate sizing, most
urban water systems possess potable water-storage capability—in the form of water
tanks, towers, or water reservoirs—to supply water even on the hour of highest use in
the year.
Examining the historical record is the common method of pinning down the
peak-to-average day ratio.
When the record shows the peak-to-average day ratio fluctuating around a
given level, the analysis is easy. The forecaster can calculate an average over several
recent years, adding a “cushion” to guarantee that the treatment system has buffer
capacity during peak-demand events. More difficult cases involve historic information
that reflects an upward-sloping or downward-sloping trend in the peak-to-average
day demand ratio.
Figure 6-1 shows the peak-to-average day ratio for 1980–1999, based on daily
water-use data from the Calgary, Alberta, water system. Note that each data point in
Figure 6-1 represents the ratio of the maximum-day water usage for that year to the
average-day water usage for the year in question. As such, these ratios are absolute
numbers—they are not stated in liters or gallons or any other physical units of
measurement. The average peak-to-average day ratio for this twenty year period is
1.86. The maximum peak-to-average day ratio of 2.12 occurs in 1985. The minimum
peak-to-average day ratio is 1.51 and occurs in 1999, the end of the data series.
Because there is a downward trend, stabilizing somewhat in recent years, the
series in Figure 6-1 suggests that the higher peak-to-average ratios at the beginning of
this historic record may not recur. This downward trend may be related to the
growth—in absolute numbers and as a proportion of residential customers—in the
number of metered customers in this water system. Metering customer usage tends to
reduce both peak and average summer or outdoor water use because water bills can
be noticeably higher in the summer months.
Examining the data back to 1950, the decline in peak-demand ratios in this
system supports using more recent periods instead of all the available data to develop
a forward-looking peak to average demand ratio for planning purposes. The time
depth selected for analysis, however, needs to be long enough to allow anomalous
weather conditions to average out.
Although peak-to-average ratios are important planning parameters, relatively
little research addresses their statistical modeling. In part, this may be because of the
large role that short-term, inherently unpredictable weather plays in generating peak-
demand events. Peak-hour and peak-day demands are often linked to and are
triggered by runs of hot, dry weather.
The short-term regression models discussed later in this chapter can yield
forward-looking information on when peak-day demand might occur. The analyst,
for example, can link daily and monthly water-demand forecast models to establish
when a predicted peak is likely to be the maximum for the year.
AVERAGING FOR SEASONAL PATTERNS

AND INTERANNUAL TRENDS
In this section, methods are determined for (1) calculating monthly proportions of
annual water usage, (2) graphing summer and winter water use and separating indoor
and outdoor water use, and (3) identifying water use trends with moving averages.
These are straightforward, robust techniques. For data, either total system water sales
or production data can be used.
Monthly Proportions of Annual Water Usage. Averaging ratios of monthly
to total annual water use can produce robust estimates of the seasonal pattern of water
use. Such average monthly proportions are readily transformed into monthly water
use factors relative to the average level of annual water usage in the water system.
Computation of monthly ratios and factors is facilitated by organizing the data
into an array, as in Table 6-1. Here, monthly total system water-use figures are placed
in columns by year. Thus, Table 6-1 (labeled “Water Use by Month”) lists monthly
total system water production from January of 1980 to December of 1989. Water sales
are totaled for each year at the bottom row of the table. These total annual water-use
numbers become the denominator of the monthly ratios in Table 6-2. Thus, the ratios
in Table 6-2 are the monthly proportions of total annual water usage in this example
water system.
Averaging the monthly proportions over this 10-year period identifies the
overall average seasonal water-use pattern. These estimates are summarized at the
right side of Table 6-2 in the column labeled “Average.”
Table 6–1 Data Array for Calculating Monthly Proportions of Annual Water Usage
Year
1980 1981 1982 1983 1984 1985 1986 1987 1988 1989
Month Water Use by Month
January 36.73 38.45 38.80 39.92 40.92 41.52 40.57 42.05 43.55 42.82
February 36.33 40.00 39.31 40.08 40.80 41.76 40.97 41.23 42.31 42.98
March 36.81 39.45 39.16 39.81 41.12 42.64 41.74 42.48 42.69 43.24
April 38.04 40.36 40.86 42.10 44.04 44.71 44.08 42.52 43.09 45.87
May 42.36 41.41 47.87 52.30 48.63 54.42 58.05 46.10 46.98 46.95
June 47.61 52.24 46.51 46.49 50.84 53.21 50.49 57.06 53.95 50.94
July 42.67 57.10 61.10 56.11 49.94 62.81 62.37 66.93 65.13 54.09
August 45.42 58.93 53.70 67.56 50.38 58.45 74.18 51.42 63.25 52.37
September 44.50 45.38 46.68 55.24 50.81 48.31 59.20 48.43 52.17 45.70
October 38.63 39.47 42.39 44.97 44.04 44.47 46.43 43.76 45.83 43.26
November 37.27 38.53 40.27 41.64 42.15 42.74 44.18 42.95 42.66 41.18
December 37.55 37.96 39.84 40.99 41.85 41.35 41.49 44.36 42.00 41.44
TOTALS 483.91 529.28 536.48 567.21 545.52 576.41 603.74 569.29 583.61 550.85
Table 6–2 Monthly Proportions of Annual Water Usage

Year
1980 1981 1982 1983 1984 1985 1986 1987 1988 1989
Month Monthly Proportions Average
January 0.0759 0.0726 0.0723 0.0704 0.0750 0.0720 0.0672 0.0739 0.0746 0.0777 0.0732
February 0.0751 0.0756 0.0733 0.0707 0.0748 0.0724 0.0679 0.0724 0.0725 0.0780 0.0733
March 0.0761 0.0745 0.0730 0.0702 0.0754 0.0740 0.0691 0.0746 0.0732 0.0785 0.0739
April 0.0786 0.0763 0.0762 0.0742 0.0807 0.0776 0.0730 0.0747 0.0738 0.0833 0.0768
May 0.0875 0.0782 0.0892 0.0922 0.0891 0.0944 0.0961 0.0810 0.0805 0.0852 0.0874
June 0.0984 0.0987 0.0867 0.0820 0.0932 0.0923 0.0836 0.1002 0.0924 0.0925 0.0920
July 0.0882 0.1079 0.1139 0.0989 0.0915 0.1090 0.1033 0.1176 0.1116 0.0982 0.1040
August 0.0939 0.1113 0.1001 0.1191 0.0924 0.1014 0.1229 0.0903 0.1084 0.0951 0.1035
September 0.0920 0.0857 0.0870 0.0974 0.0931 0.0838 0.0981 0.0851 0.0894 0.0830 0.0895
October 0.0798 0.0746 0.0790 0.0793 0.0807 0.0771 0.0769 0.0769 0.0785 0.0785 0.0781
November 0.0770 0.0728 0.0751 0.0734 0.0773 0.0742 0.0732 0.0754 0.0731 0.0748 0.0746
December 0.0776 0.0717 0.0743 0.0723 0.0767 0.0717 0.0687 0.0779 0.0720 0.0752 0.0738
This computation indicates that July is typically when peak monthly water
demand occurs in this system. Information in the table indicates that July demand is,
on average, 10.4 percent of annual water usage. Accordingly, July water demand in this
system is, on average, 42 percent greater than January usage (0.104/0.0732).
These average monthly proportions are readily transformed into monthly

water-use factors relative to average annual water utilization or production. The
monthly factors relative to average daily demand are calculated by dividing the
number of days in a year by the number of days in any specific month and
multiplying this term by the average monthly proportion for that month. For
example, the average August factor relative to average daily demand is 0.1034
multiplied by 365/31, or 1.215. Average August demand is 21.5 percent greater than
average daily demand for the year in this system.
Figure 6-2 charts the typical pattern of seasonal variation, along with the
maximum and minimum monthly proportions for each month in the historical data.
The lower curve graphs the historical minimum monthly proportion and the upper
curve graphs the historical maximum monthly proportion. The middle curve is
generated by the average proportions developed by the computations in Table 6-2.
The application of these monthly proportions and factors to forecasting is
straightforward. Suppose, for example, that system demand in 2015 is forecast to total
710 megaliters (ML). After calculating the average monthly proportions in Table 6-2,
the information predicts that the July 2015 system water usage will be 73.84 ML =
0.104 × 710. Similarly, the August water use factor of 1.215 supports a quick
computation of future August water demand, given future gallons per capita per day
(gpcd) and population projections.
Data requirements. Data requirements for calculating monthly proportions
depend on factors such as the size of the water system, its policies, and local climatic
patterns. Larger water systems tend to have more stable seasonal patterns, largely
resulting from a statistical relationship called the law of large numbers.1 Prolonged
droughts or shifting timing of peak-temperature events and wet periods can influence
the volatility of monthly water usage. Utility policies, such as water-use restrictions,
also exert an effect. For example, ratios computed over a five-year period that includes
two years of summer water-use restrictions may not accurately represent normal
seasonality.
To develop seasonal factors, several years of monthly data are needed. In the
data set of Table 6-1, average monthly proportions stabilize with three or four years of
data, showing little change as more years are added.
Generally, when seasonal ratios are developed for the water system, total water
production by month proves to be an adequate source of data.
If there is interest in seasonal or monthly ratios by customer category, analysts
will need to compile data from billing records into total water use by customer
category. The resulting sectoral picture can be desirable in developing a cost-
allocation model to apportion peak-delivery costs to each customer category. For
example, industrial demand might be approximately level through the year, with
residential, other commercial, and governmental customers accounting for the bulk of
the summer peak demand in the system.
1. The idea for this application is that as the number of water customers increases, the effect of a
few unusual data points becomes insignificant in computing totals or averages.
Summer/Winter and Indoor/Outdoor Water Use. Water-demand analysts

often use a standard tactic to estimate outdoor water use—subtracting average winter
use from summer water use. Although the logic is straightforward, it does not apply
in all climatic zones. When temperatures are cold enough, winter water use is largely
associated with water demand for indoor purposes. A simple computation, then,
supplies data on the percentage of total system demand used for indoor and outdoor
purposes.
As an example, because of this system’s northerly latitude, winter levels of water
use are largely for indoor purposes. Averaging water use from November through
February generates a good estimate of average monthly indoor water use. Extrapolat-
ing this calculation for a whole year implies that indoor water use is approximately
80 percent of total annual water production in this system. Figure 6-4 charts the result
of this computation for several years. Outdoor water use in the summer months is
represented by the somewhat irregular pulses in the chart.
50
45
40
35
Million Liters
30
25 Indoor Water Use
20
Outdoor Water Use
15
10
5
0
19 10
19 10
20 10
19 5:1
19 :7
19 :4
19 8:1
20 :7
20 :4
20 1:1
20 :7
20 :4
:1
96
97
99
00
02
03
04
:
:
95
98
01
9
0
19
Year and Month

Figure 6–4 Indoor and Outdoor Water Use
Identifying Trends With Moving Averages. Averaging procedures also can

identify trends. Moving averages can cut through seasonal variations to reveal the
underlying trend in monthly water use. The key is to develop a rolling series of
moving averages, each calculated over 12 months of consecutive data. Suppose June,
July, and August are peak months of water demand in a public water supply. Any
series of 12-month moving averages then always includes these months—although
some may be in one calendar year and others in a different calendar year. Accordingly,
a 12-month moving average can smooth

MOVING AVERAGES seasonal variation, leaving a time series
To calculate a moving average, take the composed chiefly of the underlying trend
arithmetic average of a fixed number, and purely random or irregular compo-
such as k, of consecutive terms of a time nents.
series. The number of consecutive terms Centered moving averages are a spe-
is referred to as the time window of the cial type of moving average that can
moving average. provide early signals of turning points in
By sliding the time window across or trends. For example, consider a time series
down the time series, one can generate a beginning with the numbers (3, 2, 1, ...).
series of individual averages, which taken Construct a centered moving average of
together constitute a moving average. A
three terms, with the first being
moving average of 12 terms starts by
averaging the first 12 observations or
(3 + 2 + 1)/3 = 2. Note that this value is
numbers in a time series. The second associated with the second period of the
average starts with the second observa- time series. A simple moving average with
tion of this time series and averages it a time window of three, however, calcu-
with the 11 succeeding terms. This pro- lates exactly the same value for the mov-
cess continues until the last number, ing average but associates it with the third
term, or observation of the time series is period in the time series.
included in a moving average. Accordingly, the centered moving
average yields forward-looking informa-
tion on the direction of the trend.
Suppose, for example, that the time series goes on as (3, 2, 1, 2, 3, 4, ...). A
centered moving average for a time window of three terms signals an increasing trend,
with the third moving average (1 + 2 + 3)/3 associated with the fourth period of the
time series. A simple moving average, on the other hand, takes one additional period
to indicate this upturn in trend because (1 + 2 + 3)/3 is associated with the last
position in the time window or the fifth period in the time series, as shown in Table
6-3.
Table 6–3 A Simple Moving Average

Time period 1 2 3 4 5 6
Original data 3 2 1 2 3 4
Centered moving average — 2 5/ 2 3 —
3
Simple moving average — — 2 5/ 2 3

3
There is an apparent trade-off, in that the centered moving average does not
extend to the end of the time series. This can be remedied partially by applying one of
several methods to fill in the missing value, as discussed in following sections.
Table 6-4 illustrates the computation of a centered moving average with a
rolling time window of 12 months. Each moving average is “centered” at month 6.
Table 6–4 Calculating a Centered Moving Average (CMA)
Year Q* MATREND† Year Q* MATREND†
1995:5 42.36 2000:5 54.42 48.08

1995:6 47.61 40.33 2000:6 53.21 48.03
1995:7 42.67 40.47 2000:7 62.81 47.95
1995:8 45.42 40.77 2000:8 58.45 47.89
1995:9 44.50 40.99 2000:9 48.31 47.81
1996:5 41.41 44.07 2001:5 58.05 50.30
1996:6 52.24 44.11 2001:6 50.49 50.31
1996:7 57.10 44.14 2001:7 62.37 50.43
1996:8 58.93 44.08 2001:8 74.18 50.46
1996:9 45.38 44.05 2001:9 59.20 50.52
1997:5 47.87 44.55 2002:5 46.10 47.20
1997:6 46.51 44.71 2002:6 57.06 47.44
1997:7 61.10 44.80 2002:7 66.93 47.57
1997:8 53.70 44.86 2002:8 51.42 47.66
1997:9 46.68 44.92 2002:9 48.43 47.67
1998:5 52.30 47.17 2003:5 46.98 48.83
1998:6 46.49 47.27 2003:6 53.95 48.63
1998:7 56.11 47.35 2003:7 65.13 48.57
1998:8 67.56 47.41 2003:8 63.25 48.63
1998:9 55.24 47.52 2003:9 52.17 48.68
1999:5 48.63 45.39 2004:5 46.95 45.95
1999:6 50.84 45.46 2004:6 50.94 45.90
1999:7 49.94 45.51 2004:7 54.09
1999:8 50.38 45.59 2004:8 52.37
1999:9 50.81 45.72 2004:9 45.70
*Quantity
†Moving average trend
Figure 6-5 graphs the data in Table 6-4. There is a small positive trend in the
moving average in the early years from 1995 to 2001, giving way to a flattening and
eventually slightly declining trend in 2003 and 2004.
Extrapolating Moving Average Trends. Analysts rely on two common

methods for extrapolating a moving average trend line (See Figure 6-5 and Table 6-4).
The moving average trend (MATREND) can be extrapolated with a linear trend, such
as that available in Microsoft Excel in the TREND(.,.,.) function. As an alternative, the
moving average can be extrapolated with a two-parameter (Holt) exponential
smoothing model, available in Excel and many other programs.
80
70
60
Million liters
50 Monthly Total System
Water Production
40
Centered Moving
30 Average Trend
20
10
0
19 10
19 10
20 10
0
19 5:1
19 :7
19 :4
19 8:1
20 :7
20 :4
20 :1
20 :7
20 :4
20 :1
:1
96
97
99
00
01
02
03
04
:
:
95
98
01
04
9
9
19
Year and Month

Figure 6–5 Trend Indicated by a Centered Moving Average (CMA)
REGRESSION APPROACHES
In establishing a regression analysis, three basic questions are explored:
1. What is the dependent variable?
2. What are the explanatory variables?
3. What is the functional form of the regression?
After answering these questions, the regression is “estimated”—producing the
specific values of the coefficients of the explanatory variables and other information
valuable in diagnostic testing. The “goodness of fit” of the regression to the data is a
primary number used in evaluating the regression. Other diagnostics are derived from
the statistical significance of the individual coefficients, and tests of adequacy relate to
the regression’s “residuals.”2 Finally, there is the “acid test”—whether the regression
model works satisfactorily to predict “out-of-sample” results.
Although they are substantiated by the literature, the water-demand models
presented in this section are exploratory. The focus is on rationale and interpretation.
Monthly Water-Use Models. Monthly water-demand models that specify
weather variables can demonstrate the importance of climate and weather in
determining the average seasonal pattern of water use and departures from this
pattern.
2. Regression residuals are what is left over after the predicted values of the regression relationship
are subtracted from the actual values, sample observation by observation.
In this section, three different but related monthly water-use regression models
are developed. The models are “nested,” in the sense that the list of explanatory
variables for the first regression is included, but augmented, in the second regression.
The third regression includes the explanatory variables in the first and second
regressions, plus a variable that reflects the underlying trend. The discussion
conceptually highlights the importance of climate and weather in explaining within-
year water use.
Dependent variable. In a regression, the dependent variable is explained by the
explanatory or independent variables (refer to Figure 6-5, which charts the dependent
variable for this monthly regression discussion). These monthly data on quantity (Q)
show strong seasonality. As confirmed by the centered moving average in Figure 6-5,
the trend is modest, appearing initially up then moving slightly downward.
More generally, total system water use can manifest strong trends over time, as
this water system did in earlier decades. Scaling the total water-use data by dividing by
population or the number of customers is often valuable. Growth in customers served
by the water system contributes a positive component to the trend in total water use.
Adjusting to water use per customer damps a positive trend, helping the regression to
focus on the underlying structural relationships.3 A moderating effect on a downward
total system trend is also achieved by scaling by population or customers, in systems
serving fewer customers over time.
Often, series such as those shown in Figure 6-5 are scaled. Residential water-use
series are readily scaled to a per-customer basis. It makes less sense to scale
commercial, industrial, or total system water use by numbers of customers, however,
since customer water use may be highly heterogeneous in these categories. Scaling
with population is an alternative tactic. The problem is that, typically, population
estimates are produced or updated only annually or at intervals of every few years.
One remedy is a straight-line extrapolation or interpolation between available
population estimates.
Explanatory variables. Several terms are used for the explanatory variables of a
regression, including independent, regressor, driver, and sometimes control variables.
The entire list of explanatory variables in a regression is called the specification. In the
monthly regressions presented here, three types of explanatory variables are included:
(1) weather variables, (2) dummy monthly variables, and (3) the trend variable.
Weather variables. The research literature and applied planning studies often
report difficulties in obtaining statistically significant coefficients when regressing
3. Note that, if we dip below this high level of aggregation, these data have many components. For
example, there are different types of customers, including residential and “general-service” customers
(a category that includes CI concerns), along with larger multifamily customers. In addition, some
single-family residential customers in this system are metered and others are billed at a flat rate. During
this period, the utility instituted programs for installing meters and encouraging flat-rate customers to
convert to metering. Our analysis suggests, then, that the increased number of customers in the system
and the concomitant rise in water use were offset by moving several thousand single-family residential
users from the flat rate to the metered category.
temperature and precipitation onto water use. This problem is found in the example
data also. When average monthly temperature and total monthly precipitation are
regressed against total monthly water production in this system, only the temperature
variable is statistically significant. This is counterintuitive because rainfall is impor-
tant in summer water use in this high-prairie environment.
Often, the problem is related to the mutual correlation of temperature and
precipitation. Essentially, these variables can be correlated to such a degree that the
influence of one cannot be separated from the other by the regression procedure.
Generally, this indicates multicollinearity in regression, which is discussed in more
detail in chapter 12.
One solution for the problem of entangled temperature and precipitation values
is to create new variables that measure differences between actual month-to-month
temperature and precipitation from their long-term monthly averages. These trans-
formed variables, DELTEMP and DELPRECIP, are defined as follows:
DELTEMP = average temperature for the specific month minus

the long-term average temperature for that month
DELPRECIP = total precipitation for a specific month minus the
long-term average precipitation for that month.
The conceptual thrust of these transformations is that unusual weather is often

the primary cause of unusual levels of water use. Otherwise, monthly water demand
tends to follow a seasonal pattern that is consistent from one year to the next.
Dummy monthly variables: Monthly dummy variables are binary variables that
assume value 1 for the month in question and value 0 during all other months. These
are illustrated in Table 6-5, which shows several years of data used in the following
regressions. The table includes the dependent variable, total system water use
measured in millions of liters (Q); DELTEMP; DELPRECIP; and dummy variables for
June, July, August, and September. The table shows a time-related trend (MATREND)
as well.
To estimate the model, one monthly dummy variable must be omitted, so there
is no value for May. In effect, the constant of the estimated equation becomes the
coefficient for May, and each of the other estimated dummy monthly variable
coefficients shows the difference between that month’s water use and the level of water
use of the omitted month of May (the “base”).
The formal definitions of the monthly dummy variables follow:
JUNE = 1 if the observation is for June; 0 otherwise

JULY = 1 if the observation is for July; 0 otherwise
AUGUST = 1 if the observation is for August; 0 otherwise
SEPTEMBER = 1 if the observation is for September; 0 otherwise
Table 6–5 Sample of Regression Database

Q DELTEMP DELPRECIP JUNE JULY AUGUST SEPTEMBER MATREND
Year (1) (2) (3) (4) (5) (6) (7) (8)
1995:5 42.36 0.37 17.23 0 0 0 0

1995:6 47.61 1.31 (50.56) 1 0 0 0 40.33
1995:7 42.67 (2.28) 64.92 0 1 0 0 10.47
1995:8 45.42 (3.30) (11.16) 0 0 1 0 40.77
1995:9 44.50 1.01 (6.30) 0 0 0 1 40.99
1996:5 41.41 (3.63) (3.17) 0 0 0 0 44.07
1996:6 52.24 0.61 (34.76) 1 0 0 0 44.11
1996:7 57.10 (0.18) (26.58) 0 1 0 0 44.14
1996:8 58.93 1.80 (24.36) 0 0 1 0 44.08
1996:9 45.38 (3.19) 12.20 0 0 0 1 44.05
1997:5 47.87 0.37 46.03 0 0 0 0 44.55
1997:6 46.51 (0.09) 44.44 1 0 0 0 44.71
1997:7 61.10 (0.18) (51.58) 0 1 0 0 44.80
1997:8 53.70 0.00 12.44 0 0 1 0 44.86
1997:9 46.68 2.71 3.60 0 0 0 1 44.92
1998:5 52.30 3.97 31.73 0 0 0 0 47.17
1998:6 46.49 (1.19) 16.44 1 0 0 0 47.27
1998:7 56.11 0.52 63.72 0 1 0 0 47.35
1998:8 67.56 2.10 (27.36) 0 0 1 0 47.41
1998:9 55.24 1.91 (8.20) 0 0 0 1 47.52
1999:5 48.63 (0.03) (1.87) 0 0 0 0 45.39
1999:6 50.84 (1.09) 1.44 1 0 0 0 45.46
1997:7 49.94 (3.58 35.32 0 1 0 0 45.51
1999:8 50.38 (1.30) 43.84 0 0 1 0 45.59
1999:9 50.81 (0.49) (25.10) 0 0 0 1 45.72
2005:5 54.42 0.07 (25.87) 0 0 0 0 48.08
NOTES: Q = quantity
DELTEMP = average temperature for the specific month minus the long-term average
temperature for that month
DELPRECIP = total precipitation for a specific month minus the long-term average precipitation
for that month
MATREND = centered moving average trend
Trend variable. The third type of variable represents the underlying trend in
monthly total water use. For this analysis, MATREND, the 12-month centered moving
average, was selected.
The concept of the centered moving average was described earlier in the chapter
(refer to Tables 6-3 and 6-4 for illustrations). One advantage of using a moving
average trend is that it accommodates nonlinearities in the underlying data. In
contrast, a simple time trend depicts a totally linear trend line.
Functional forms. Mathematics tells us that any continuous function can be
approximated by a linear function. Thus, around a point (X1, X2, …Xk) in a
generalized k-dimensional space, a function Q =f(.) can be approximated by a linear
estimating equation such as
Q = a0 + a1 X1 + a2 X2 + … ak Xk (Eq 6-1)
where the ai, i=1,2,…k are the regression coefficients to be estimated, and a0 is
the constant of the regression.
Data transformations are also possible, and sometimes improve the ability of
the estimating function to fit the data. For example, the dependent variable and at
least some of explanatory variables can be logarithmically transformed.4 Details on
this tactic are described in Chapter 12, and while this discussion considers only linear
forms for the estimating equation.
The results are presented for the following three regressions:
Q = a 0 + a 1 DELTEMP + a 2DELPRECIP (Eq 6-2)
Q = b 0 + b 1 DELTEMP + b 2 DELPRECIP + b 3 JUNE (Eq 6-3)

+ b 4 JULY + b 5 AUGUST + b 6 SEPTEMPER
Q = c 0 + c 1 DELTEMP + c 2 DELPRECIP + c 3 JUNE + c 4 JULY (Eq 6-4)

+ c 5 AUGUST + c 6 SEPTEMPER + c 7 MATREND
Each of these regressions is statistically estimated with summer water-use data.

Analysts often develop separate water-demand regressions for winter and
summer because seasonal peaking is usually driven by outdoor water use.
Looking at the pieces or parts of monthly demand models in Equations 6-2, 6-3,
and 6-4 highlights a key question—under what conditions are regression coefficients
stable over different regression specifications? Or, to put it another way, under what
conditions is it possible to add or remove an explanatory variable from the regression
specification without impacting the estimated coefficient values of the other
explanatory variables? The following example highlights the importance of correla-
tions of the explanatory variables in accounting for this stability.
Regression results of monthly demand models. Regression results can be presented
in various formats. For example, the regression result can be written on one line in
equation form with the estimated coefficients and diagnostic statistics, such as the
t-statistics of the coefficients along with the coefficient of determination or R2. The
regression of DELTEMP and DELPRECIP onto Q, results in the following specific
parameters and diagnostics:
ˆ = 52.8 + 2.02DELTEMP – 0.065DELPRECIP

Q (Eq 6-5)
(68.2) (4.62 ) (2.12) R2 = 0.44
4. Logarithmic transformations require that variables be nonnegative and nonzero. Sometimes this
can be satisfied by transforming variables, such as adding a constant to all values.
Total water use Q becomes Q ˆ signifying that the right side of the equation
produces an estimate of Q.
The information in Equation 6-5 shows us that when temperature is higher
than normal for a month, total monthly water use increases because the coefficient of
DELTEMP is positive. Similarly, when precipitation is higher than normal, Q
decreases. Accordingly, this regression passes what is sometimes known as the sign
test—the signs of the estimated coefficients conform to expectations.
The t-statistics are given in parentheses underneath the estimated coefficients.
Note that, while, technically, the t-statistic has the same sign as the associated
coefficient, only its absolute magnitude is important. For sample sizes of 30
observations or more, a rule of thumb is that a t-statistic greater than 2 indicates that
a regression coefficient is significant at the 95 percent level. In other words, there is a
probability of 0.95 or greater that the estimated coefficient is not 0 or of opposite sign.
Accordingly, Equation 6-5 has regression coefficients that are significant from a
statistical standpoint at a 95 percent level or greater.
The other piece of information found in Equation 6-5 is the R2 or coefficient of
determination. This is a measure of the goodness of fit of the estimating equation to
the data. In this case, the R2 is 0.44, indicating that this simple regression equation
with these two weather variables explains about 44 percent of the variation in the
values of Q.
ˆ = 48.51 + 2.02DELTEMP – 0.065 DELPRECIP + 2.4JUNE

Q (Eq 6-6)
+ 9.3JULY + 9.1AUGUST + 1.1SEPTEMBER
Equation 6-6 gives the results of estimating the more complete model of Eq. 6-3.
These results are also presented in Table 6-6, which is an image of the output
generated by the regression routine in Microsoft Excel. The explanatory variables are
DELTEMP, DELPRECIP, and the four dummy variables, JUNE, JULY, AUGUST, and
SEPTEMBER. Q is the dependent variable.
The table contains a great deal of information, but essentially, with the same
elements: goodness of fit; significance of the individual coefficients, and, in this case,
an F-test that indicates the significance of the whole regression specification.
The column in the lower part of the display (labeled Coefficients) presents the
estimated regression coefficients—the bi in Equation 6-3. Table 6-6, then, shows the
specific form of the estimating Equation 6-3:
Total water use Q becomes Q ˆ , signifying that the right-hand side of the
equation produces an estimate of Q. Again, the signs of the estimated coefficients are
consistent with expectations. The coefficient of DELTEMP is positive and the
coefficient of DELPRECIP is negative. The values of these coefficients are essentially
the same as the ones estimated in the simpler Equation 6-5.
Each monthly dummy shows the difference between that month’s water use and
the base level of water use of the omitted month (May in this case). Each dummy
variable is positive, indicating that water use in June, July, August, and September in
this system is typically higher than water use in May.
Table 6–6 Regression Output for Estimating Equation 6-3
SUMMARY OUTPUT
Regression Statistics
Multiple R 0.870
R Square 0.757
Adjusted R Square 0.723
Standard Error 3.787
Observations 50
ANOVA
df SS MS F Significance F
Regression 6 1,919.169 319.861 22.301 9.79E-12
Residual 43 616.743 14.343
Total 49 2,535.912
Coefficients Standard Error t Stat P-value Lower 95% Upper 95%

CONSTANT 48.506 1.198 40.502 0.000 46.091 50.921
DELTEMP 2.023 0.302 6.703 0.000 1.414 2.632
DELPRECIP (0.065) 0.021 (3.070) 0.004 (0.108) (0.022)
JUNE 2.427 1.694 1.433 0.159 (0.989) 5.842
JULY 9.319 1.694 5.502 0.000 5.903 12.735
AUGUST 9.059 1.694 5.348 0.000 5.643 12.474
SEPTEMBER 1.137 1.694 0.671 0.506 (2.279) 4.553
The t-statistics, however, show that not all the dummy variables are statistically
significant at the 95 percent level. June and September have t-statistics less than 1.97.
This result is not particularly surprising, and simply suggests that June and September
water use is not all that different from May water use. July and August, on the other
hand, are significantly different. Related information in the regression printout
encompasses the standard errors of the coefficient estimates. These are valuable in
constructing confidence intervals for the regression coefficients (see our discussion in
chapter 12). With the added variables, the regression now explains approximately
76 percent of the variation in Q.
To improve the goodness of fit of the regression further, a variable is included
to represent the trend in this system’s water use. The results of estimating Equation 6-5
are presented in Table 6-7.
This regression provides a close fit to the sample data, explaining 89 percent of
the variation in the sample data. Most of the regression coefficients are statistically
significant at a 95 percent level or higher, the exceptions being CONSTANT and the
dummy variables for JUNE and SEPTEMBER. One explanation for this may be that,
once the trend variable is added, water use in May, June, and September is not that
different from the underlying trend. In the previous models without trend, the
constant picked up the effect of trend.
The closeness of fit of this regression model to the sample data is illustrated in
Figure 6-6. Note that the predicted summer monthly demands are close to the
historical values, and the predictions of the model correctly identify the timing of
maximum monthly demand in each of the sample years.
Table 6–7 Regression Output for Estimating Equation 6-5
Multiple R 0.944 Standard Error 2.607
R Square 0.891 Observations 46
ANOVA
Degrees of Sum of Mean
Freedom Squares Squares F-test Significance F
Regression 7 2,106.80 300.97 44.27 2.3966E-16
Residual 38 258.33 6.80
Total 45 2,365.13
Standard Lower Upper

Coefficients Error t Stat P-value 95% 95%
CONSTANT 1.824 7.202 0.253 0.801 (12.755) 16.403
DELTEMP 1.531 0.229 6.686 0.000 1.067 1.994
DELPRECIP (0.074) 0.015 (4.975) 0.000 (0.104) (0.044)
JUNE 1.748 1.198 1.459 0.153 (0.678) 11.515
JULY 9.024 1.231 7.333 0.000 6.533 11.515
AUGUST 8.278 1.231 6.726 0.000 5.787 10.770
SEPTEMBER 0.161 1.233 0.131 0.897 (2.334) 2.657
MATREND 1.025 0.155 6.620 0.000 0.711 1.338
Stability of regression coefficients. The coefficients of DELTEMP and DELPRE-

CIP are essentially identical in Equations 6-5 and 6-6. This is because there is virtually
zero correlation among DELTEMP; DELPRECIP; and the dummy variables JUNE,
JULY, AUGUST, and SEPTEMBER. This is no surprise because there is no reason to
expect deviations of temperature and precipitation from their seasonal averages to be
correlated with those seasons. In contrast, temperature and precipitation are likely to
be highly correlated with the dummy monthly variables.
Another way of stating this is that, ideally, explanatory variables in a regression
specification will have zero mutual correlation or be “orthogonal.” Then, deleting any
explanatory variable from the specification does not affect new estimates of the
remaining explanatory variables when the reduced equation is reestimated. By the
same token, when there are nonzero correlations between the explanatory variables,
adding or deleting variables from the regression specification can affect the value of
the remaining coefficient estimates.
The final estimate of DELTEMP in Table 6-7, then, differs somewhat from those
in Equations 6-5 and 6-6, although the value for DELPRECIP is closer to the value
estimated in the first two regressions. The estimates of the monthly dummy variables
are roughly the same in Equation 6-6 and Table 6-7.
If all the independent variables were exactly orthogonal5 to each other, their
estimated coefficients would stay the same regardless of which ones were included or
80
75
70
65
Actual Summer
Million Liters
60 Monthly Total System

Demand
55
Predicted Summer
50 Monthly System
45 Demand
40
35
30
19 :7
19 :6
19 :5
19 :9
19 :8
20 :7
20 :6
20 :5
20 :9
20 :8
20 :7
:6
95
96
97
97
98
99
00
01
01
02
03
04
19
Year and Month
Figure 6–6 Comparison of Summer Water Use and Values Predicted by Monthly Water
Demand, Estimating Equation 6-4
omitted from the model specification. The extent to which they change when
variables are added reflects the degree of correlation among the independent
variables.
Applications of regression models. Suppose that average summer temperatures
increase 2° in this example water system. Going back to the database used to develop
these monthly regression models and substituting in an increase of 2° in DELTEMP
would be straightforward. Assuming that these higher temperatures are not associated
with change in DELPRECIP, these regression models suggest that total monthly water
use May through September will increase six to nine percent.
These regression models can be used to estimate impacts of water restrictions.
Rather dramatic “squashing” of seasonal peaks can be associated with the imposition
of emergency water restrictions. To model this, a regression can be specified with a
dummy variable that assumes the value of 1 for those months in which there are
watering restrictions in the historical data, and a value of 0 otherwise. Dummy
variables that signify special events are sometimes called “intervention” variables.
Daily Water-Demand Regressions and Peak-Day Demand. The focus of

daily water-demand models is typically on near-term daily water use during the
summer watering season. Combined with short-term weather forecasts, these daily
water-demand models can be applied to maintenance scheduling and system
5. Variables are orthogonal when their correlation is zero.

optimization. In addition, they can give utilities a “heads up” about peak-demand
events. Statistical modeling of daily water demands offers a key advantage over
judgmental methods and rules of thumb in this context. This is the capability of
statistically based models to support probabilistic assessments, such as the likelihood
that daily demands will reach or exceed a given level within a forecast horizon of one
to a few days. This capability can be especially important for water systems with
storage limitations.
Daily water-demand models generally have two major structural components.
First, all these models proceed from recent events and demand levels to near-term
developments. As a result, there is some term or construct in the model designed to
capture the recent level of water use.
Second, current or near-term weather conditions are integrated into the model.
This discussion focuses on a simple implementation of this type of model. The sample
data include daily water production and weather data for the summer months of May
through September (1996–1998). The regression relationship estimated on this daily
sample data is then tested for out-of-sample performance on daily data from 1999.
The demand specification includes a lagged dependent variable DAILYQLAG1,
weather variables ADTEMP and DPRECIP that are contemporaneous with the
dependent variable DAILYQ, and lagged weather variables, ADTEMPLAG1 and
DPRECIPLAG1.
The functional form of the regression equation is
DAILYQ = CONSTANT + a 1 ADTEMP + a 2 DPRECIP (Eq 6-7)

+ a 3 ADTEMPLAG1 + a 4 DPRECIPLAG1 + a 5 DAILYQLAG1
where:
DAILYQ = total system water production in day t
ADTEMP = average daily temperature for day t
DPRECIP = total daily precipitation for day t
ADTEMPLAG1 = average daily temperature for day t-1
DPRECIPLAG1 = total daily precipitation for day t-1
DAILYQLAG1 = total system water production in day t-1
CONSTANT, a1, a2, a3, a4, and a5 are parameters to be estimated by ordinary
least squares (OLS) regression.
Because this database includes only the summer months of May through
September, the lagged dependent variable DAILYQLAG1 must be determined
correctly.6
The diagnostics of this regression model are mixed. On the one hand, the
regression explains about 76 percent of the variation in the daily summer water-use
6. In other words, the observation for May 1 of each year is lost to make sure DAILYQLAG1 pre-
cedes DAILYQ by one day, allowing for the seven month gaps in the data (September 30 from the pre-
vious year to May 1 of the succeeding year). The results of this regression are presented in Table 6-8.
Table 6–8 Daily Demand Regression Model
SUMMARY OUTPUT
Multiple R 0.871
R Square 0.758
Standard Error 55.762
Observations 455
Analysis of Variance
Regression 5 4,382,735 876,547 282 5.3729E-136
Residual 449 1,396,138 3,109
Total 454 5,778,873
Standard Lower Upper
Coefficients Error t Stat P-value 95% 95%
CONSTANT 99.471 13.260 7.502 0.000 73.412 125.529
ADTEMP 9.788 1.067 9.172 0.000 7.690 11.885
DPRECIP (1.460) 0.486 (3.003) 0.003 (2.415) (0.505)
ADTEMPLAG1 (5.132) 1.151 (4.459) 0.000 (7.394) (2.871)
DPRECIPTLAG1 (1.955) 0.494 (3.954) 0.000 (2.926) (0.983)
DAILYQLAG1 0.704 0.031 22.492 0.000 0.642 0.765
data over these three sample years. The CONSTANT or intercept term and all the
regression coefficients are statistically significant at the 95 percent confidence level or
higher. The signs of several of the coefficients conform to expectations. On the other
hand, the negative coefficient of ADTEMPLAG1, appears somewhat paradoxical. It is
expected that an especially high temperature the previous day would be associated
with higher current water use. One possibility is multicollinearity between lagged
water use and lagged temperature. In further refinement of this regression model, it
might be prudent to delete one of the two lagged variables.
Interestingly, the out-of-sample performance of this model on daily summer
data from 1999 is fairly solid, as shown in Figure 6-7. This highlights the fact that even
an imperfect model can provide solid performance in forecasting.
A Normal Probability Distribution. The residuals of the estimating Equa-

tion 6-7 are calculated as differences, on an observation by observation basis, between
actual and predicted daily water usage in the sample period. These residuals plot as a
roughly bell-shaped or normal distribution and have a standard deviation of 55.45.
Normally distributed residuals support the development of probabilistic
confidence intervals (error bands)7. Predicted daily demand one day out in the
forecast horizon has a minimum 95 percent confidence interval of ± 1.96 × 55.45 =
± 108.682 mld. Typical errors of one-day-ahead weather forecasts for average daily
730
680
630
Million Liters
580
Predicted
530 (out-of-sample)
Actual
480
430
380
330
7/ 999
15 99
12 99
9/ 999
16 99
22 9
29 9
8/ 999
19 9
26 9
9/ 999
23 9
30 9
9
7/ 199
7/ 199
8/ 199
8/ 199
9/ 199
9/ 199
99
7/ 19
8/ 19
9/ 19
1
1
/1
/1
/1
1/
8/
5/
2/
9/
/
/
/
/
/
/
7/
Figure 6–7 Projected and Actual Water Use, Eq. 6-7
temperature and total precipitation add to this error band because the water-use
forecast depends on these weather forecasts.
APPLICATIONS OF FORECASTING AND RISK EVALUATION

Statistical modeling of water demand offers a key advantage over judgmental methods
and rules of thumb. This is the capability of regression and other statistical models to
support probabilistic assessments.
Monthly regression models provide a rational basis for analyzing the potential
variability in total system water use as a result of characteristic weather patterns.
Analysts can use risk or Monte Carlo simulations to run “synthetic” weather series
through the model to study the range and distribution of variation in the peak month
values (See the discussion in chapter 13). Daily regression models can be applied
recursively8 to generate forecasts for several days in a row. Thus, the one-day forecast,
together with the two-day-out forecast of weather, is used to estimate daily water
demand two days ahead. Then, this forecast for day two, together with a three-day-out
weather forecast, might be used to produce a water-use forecast for day three. The
error bands for these recursive predictions can be calculated and widen quickly for
forecasts more than a few days ahead.
7. The terms error band and confidence interval are used interchangeably.
8. Recursive means that computations feed into each other. Thus, computations for one period use
the result of computations relating to prior periods.
Monthly and daily forecast models also can provide forward-looking informa-
tion on when peak-day events are likely to occur. Peak demand days tend to occur
during peak demand months. Thus, the monthly and daily water-demand models are
combined to assess the likelihood of peak water-use events a few days hence.
Another planning or evaluation application concerns estimates of the water
saved by imposition of emergency water restrictions. When an after-the-fact
evaluation is conducted, exact information on monthly weather variables and water
use is available. A summer month regression model can then determine, with a fair
degree of precision, what water demand would have been with and without the
conservation program. This type of application is discussed further in Chapter 10.
SUMMARY
In this chapter, tools and methods for analyzing and modeling seasonal and peak
water demands are considered. These include averaging, visual inspection of charts,
and multivariate regression.
The common method of pinning down the peak-to-average day ratio in a water
system involves examining the historical record. Issues can arise, though, in terms of
how much data to consider, how to choose the size of the data window, and how to
extrapolate trends.
Calculating average monthly proportions of water use against annual totals is
relatively straightforward. In climatic zones with freezing winter weather, determining
winter and summer water usage can lead to identifying levels of indoor and outdoor
water use. In milder climates, accurate separation of indoor and outdoor use requires
other methods, which might include data logging or direct surveys.
Moving averages are shown to be useful in highlighting trends in higher
frequency data.
Although the discussion of monthly and daily water-demand regressions is
more technically challenging, it can be reduced to six general lessons:
1. Analysts can use regression analysis to explain variations in system-wide
water use to a high degree of accuracy, basing the analysis on three types of
variables: (a) the monthly pattern of water use; (b) the underlying trend in
water use, and (c) weather variables, expressed in terms of their differences
from monthly long-term averages. Three to five years of monthly data are
recommended, at a minimum.
2. Developing the list of regression variables (specification) is an art, not a
science, but there are guidelines. Diagnostically, for example, two primary
tools are the sign test and the statistical significance of the estimated
coefficients. A good outcome is when the coefficients on the independent or
regressor variables have the expected signs and each coefficient is statisti-
cally significant. If, however, an explanatory variable has an unexpected
sign, the analyst may wish to experiment with dropping that variable from
the specification (or reconsider the underlying theory of the interrelation-
ships). If a variable with a counterintuitive sign also is highly correlated

with another explanatory variable, multicollinearity may be a problem.
Sometimes including only one of the multicollinear variables improves the
logic of the regression, without harming its ability to generate good
forecasts. The analyst may also consider dropping variables with very low
t-statistics.
3. The distinction between indoor and outdoor water use often is strategically
important in monthly or seasonal water-demand modeling. Precipitation
and temperature can have different effects on water usage in winter than in
summer. When winter temperatures are characteristically frigid, fluctua-
tions in winter weather variables can have virtually no effect on system
water usage. In milder climates, some landscape irrigation continues
during the winter, suggesting that winter precipitation can have an impact
on water use.
4. The high explanatory power often associated with monthly or daily water-
demand regressions makes these tools especially useful in policy evaluation
and retrospective studies.
5. Prospectively (or in terms of forecasting), though, the situation is less
promising. As discussed in chapter 8, the reliability of weather forecasts
beyond a few days is severely limited. Nevertheless, such models support
statements such as “if average summer temperatures increase by 1°, total
system water demand is likely to increase by X.”
6. Daily water-demand models share some general characteristics with
monthly water-demand models, especially the inclusion of terms or
variables designed to capture the recent level of water use and fluctuations
in the weather.
REFERENCES
Aly, A.H., and N. Wanakule. 2004. Short-Term Forecasting for Urban Water Consumption. Journal of
Water Resources Planning and Management 130(5): 405–410.
Brekke, L., M.D. Larsen, M. Ausburn, and L. Takaichi. 2002. Suburban Water-Demand Modeling Using
Stepwise Regression. Journal AWWA 94(10): 65–75.
Liu, J., T. Zhang, and S. Yu. 2004. Chaotic Phenomenon and the Maximum Predictable Time Scale of
Observation Series of Urban Hourly Water Consumption. Journal of Zhejiang University Science
5(9): 1053–1059.
Maidment, D.R., S.P. Miaou, and M.M. Crawford. 1985. Transfer Function Models of Daily Urban Water
Use. Water Resources Research 21(4): 425–432.
Makridakis, S., S.C. Wheelwright, and R.J. Hyndman. 1998. Forecasting: Methods and Applications. 3rd
ed. New York: John Wiley & Sons.
Michelsen, A.M., J.T. McGuckin, and D.M. Stumpf. 1998. Effectiveness of Residential Water Conservation
Price and Nonprice Programs. Denver, Colo.: Awwa Research Foundation (AwwaRF).
Gutzler, D.S., and J.S. Nims. 2005. Interannual Variability of Water Demand and Summer Climate in
Albuquerque, New Mexico. Journal of Applied Meteorology 44 (December): 1777–1787.
Zhou, S.L., T.A. McMahon, A. Walton, and J. Lewis. 2000. Forecasting Daily Urban Water Demand: A
Case Study of Melbourne. Journal of Hydrology 236:153–164.

CHAPTER 7
Population, Employment,
and Technology Forecasts
Water-demand forecasts are based on forecasts of population and employment, which

usually are developed outside water agencies in city, metropolitan, and state planning
organizations. Over time, the technology for using water changes, and housing trends
and business conditions can also be important in forecasting water demand. This
chapter reviews basic facts about these key factors and how they relate to forecasting.
In the first section, we focus on high-profile population trends and develop-
ments in the United States. Trends include the movement of the “center” of the US
population west and south, along with the impacts of immigration. We also look at
the fastest growing states and metropolitan areas and consider methods for
forecasting regional and local populations.
Next, significant developments in national employment numbers are reviewed
in a section on employment forecasts. We consider the occupations that, according to
the US Bureau of Labor Statistics (US BLS), are likely to be the fastest growing, as well
as potential future impacts of continued outsourcing of manufacturing and service
jobs overseas.
The section on regional economic forecasts presents information on personal
income, effects of inflation, income trends, and housing composition. The section on
technology forecasts discusses the acquisition of new technologies as a diffusion
process, technology and industrial water use, and appliances. We conclude this
chapter with checklists that offer suggestions for evaluating population, employment,
and regional economic forecasts.
109
POPULATION FORECASTS
The population of the United States grew by approximately 33 million during the
1990s, and these growth trends continued through the first half of the first decade of
the twenty-first century. Current interim US Bureau of Census population forecasts
project that this pattern of increase will continue through 2050.1 The interim
projection has US population growing by 50 percent over 50 years—from 282 million
recorded in the 2000 census to 421 million in 2050, an increase of 139 million. This
represents a substantial revision to US Census projections from the 1990s, which, in
some variants, forecast leveling of total population in the 2020 to 2030 time frame.
Figure 7-1 plots historical population on the left-hand scale and growth rates
(in percentages) on the right-hand scale from 1960 through 2000. The figure shows
currently available interim population projections and the corresponding decade-by-
decade growth percentages for 2000 through 2050. The population forecast is
essentially a straight-line extension of the population curve from 1960 to 2000, except
for the growth surge in the 1990s. Decade-by-decade population growth rates fell
significantly between 1960 and 1990, and these rates are projected to drift downward
following the sharp increase during the 1990s.
This forecast includes significant developments relating to the ethnic and age
composition of the US population. The Hispanic component of the US population is
projected to double from about 12.6 to 24.4 percent of total population between 2000
and 2050. The percentage of individuals who are 65 and older is also expected to
nearly double, from 12.4 percent of the total population in 2000 to 20.7 percent of the
population over the next 50 years.
Population growth in the United States is closely linked with job availability but
has a dynamic of its own related to adjustments in migration, fertility, mortality, and
age structure. As a consequence, census forecasts are generally disaggregated into
births, deaths, and migration, and also by age group (cohorts), gender, and ethnic/
racial groupings (Day 1993).
NATIONAL POPULATION TRENDS

Major changes in US demographic patterns include:
• Movement of the center of US population west and south
• Increases in the median age of the population
• Rapid population growth in smaller cities and metropolitan areas in Florida,
elsewhere in the South, in the Southwest, the West, and California
1. Unexpected immigration and associated change in demographic factors appear to have delayed
publishing long-range national population forecasts with high- and low-growth scenarios following
the 2000 US Census (US Census Bureau 2004). Alternative methods of including uncertainty in census
population forecasts, such as issuing stochastic forecasts with probability bounds or confidence inter-
vals, are being considered.
POPULATION, EMPLOYMENT, AND TECHNOLOGY FORECASTS 111
450 20
400 18
350 16
14
300
Million Persons
12
% Growth Rate
250
10
200
8
150
6
100 4
50 2
0 0
1960 1970 1980 1990 2000 2010 2020 2030 2040 2050
Year
Source: Compiled from US Bureau of Census reports.
NOTE: Total population in blue along left-hand scale; growth rates (percentages)
in pink along right-hand scale.
Figure 7–1 Historical Population From 1960 to 2000 and Projections to 2050
• Rapid increases in the Hispanic population
• Growth in single-parent households and households composed of unrelated

individuals.
A major development has been the surge in immigration to the United States.
More than one-third of the population gain in the 1990s resulted from foreign-born
persons entering the United States (about 12 million people). Many of these new
residents were of Hispanic origin. Typically younger than the resident population,
immigrant groups often share distinctive socioeconomic and cultural characteristics.
Current fertility rates of Hispanic families, for example, are higher than those of other
major ethnic groups in the US population. The result is a boost in US birth rates,
sustaining the number of individuals anticipated in the younger age categories.
Residence patterns or living arrangements are also significant for water
planning, because there are “economies of scale” in household water use. Two one-
person households use as much as 60 percent more water than one two-person
household. Accordingly, trends toward greater numbers of single-person households
contribute to increases in water use, assuming that housing patterns remain relatively
constant. The number of single-person households surged from 17 percent of all
households in 1970 to 26 percent in 2000. Married couples represented 71 percent of
US households in 1970, but only 52 percent of the 105 million households in the
United States in 2000.
REGIONAL POPULATION TRENDS

Two major trends that are affecting regional populations in the United States are (1)
continuing reduction in the population of core city areas and (2) expansion of
peripheral suburban areas, including an increase in the number of group retirement
homes and communities. Communities in the South and the West continue to grow
rapidly. The 10 states with the highest percentage increases during the 1990s were
Nevada (66.3); Arizona (40.0); Colorado (30.6); Utah (29.6); Idaho (28.5); Georgia
(26.4); Florida (23.5); Texas (22.8); North Carolina (21.4); and Washington (21.1).
Table 7-1 shows the 30 most rapidly growing metropolitan areas for the period
from 1990 to 2000. Most of these are located in the Rocky Mountain West, the
Southwest, California, and the eastern seaboard of the South.
Table 7–1 Most Rapidly Growing Metropolitan Areas, 1990–2000
Population
Percent
2000 Change in
Metropolitan Statistical Area With 2005 2004 Estimates Population
Metropolitan Division July 1 July 1 Base 1990–2000
St. George, Utah 118,885 110,425 90,354 31.6
Greeley, Colo. 228,943 219,961 180,861 26.6
Las Vegas-Paradise, Nev. 1,710,551 1,648,524 1,375,738 24.3
Cape Coral-Fort Myers, Fla. 544,758 514,923 440,888 23.6
Bend, Ore. 141,382 134,618 115,367 22.5
Naples-Marco Island, Fla. 307,242 296,675 251,377 22.2
Provo-Orem, Utah 452,851 443,109 376,778 20.2
Riverside-San Bernardino-Ontario, Calif. 3,909,954 3,785,883 3,254,821 20.1
Port St. Lucie-Fort Pierce, Fla. 381,033 364,803 319,426 19.3
Raleigh-Cary, N.C. 949,681 914,963 797,026 19.2
McAllen-Edinburg-Mission, Texas 678,275 657,310 569,463 19.1
Gainesville, Ga. 165,771 160,788 139,315 19.0
Phoenix-Mesa-Scottsdale, Ariz. 3,865,077 3,713,291 3,251,876 18.9
Prescott, Ariz. 198,701 190,737 167,517 18.6
Stockton, Calif. 664,116 649,241 563,598 17.8
Orlando-Kissimee, Fla. 1,933,255 1,863,086 1,644,563 17.6
Coeur d'Alene, Idaho 127,668 122,447 108,685 17.5
Ocala, Fla. 303,442 291,768 258,916 17.2
Boise City-Nampa, Idaho 544,201 524,789 464,840 17.1
Fayetteville-Springdale-Rogers, Ark.–Mo. 405,101 390,944 347,045 16.7
Laredo, Texas 224,695 218,806 193,117 16.4
Austin-Round Rock, Texas 1,452,529 1,411,199 1,249,753 16.2
Madera, Calif. 142,788 138,895 123,109 16.0
Atlanta-Sandy Springs-Marietta, Ga. 4,917,717 4,796,268 4,248,018 15.8
Myrtle Beach-Conway-North Myrtle Beach, S.C. 226,992 217,635 196,629 15.4
Kennewick-Richland-Pasco, Wash. 220,961 215,552 191,822 15.2
Reno-Sparks, Nev. 393,946 384,343 342,885 14.9
Merced, Calif. 241,706 236,857 210,554 14.8
Wilmington, N.C. 315,144 303,258 274,550 14.8
Bakersfield, Calif. 756,825 734,077 661,653 14.4
Other areas experienced population declines. Older core metro city areas—such
as in Baltimore, Detroit, and Philadelphia—lost population. The reasons are complex
but can involve flight to the suburbs, loss of employment base, and changes in the
economic competitiveness of the region compared to other areas both nationally and
globally.
REGIONAL AND LOCAL POPULATION FORECASTS

Regional and local population forecasts are the major focus for local utility planners.
Errors in projected birth rates plague national population forecasting. On the regional
and local scale, errors in migration assumptions often result in large errors. Analysts
can base regional and local population forecasts on several approaches:
• Extrapolating the current average growth rate, perhaps including some
adjustment factor(s) throughout the forecast period
• Applying cohort-component techniques based on assigning different fertili-
ty, migration, and death rates to various age groups or cohorts (Pittenger
1976)
• Determining land-use ratios, which allow analysts to infer the eventual
build-out population from the land-use plan and zoning ordinances
• Using economic-base models, which link population growth to growth in
economic base industries and employment
• Employing structural regression models that use causal variables in which
population growth is the left-hand (dependent) variable “explained” by a set
of independent variables representing social and economic conditions
Virtually all population forecasts are based on imprecise base population
numbers. Even official government census results contain errors and omissions, which
are likely to be larger in central city areas with higher population densities, lower
employment rates, a large proportion of transitory or homeless population, and
perhaps a larger proportion of undocumented individuals. Nevertheless, the official
census counts are generally considered to be reasonably accurate, so analysts use these
figures as the basis for their population estimates and forecasts.
Methods for estimating the existing population in a noncensus year start with
the official census year counts and then draw on information such as county and state
birth and death records; new (and discontinued) telephone, electric, and water
hookups; and school enrollments to estimate inter-census-year populations. Usually,
the most important factor affecting population change in a region is the economic
situation—employment opportunities are a magnet for new migration. Expanding
employment opportunities also support family formation and, accordingly, the
increase of the population. When job growth slows or declines, individual communi-
ties and entire states often experience depopulation.
EMPLOYMENT FORECASTS
Total employment2 in the United States grew by approximately 18.1 million workers
during the 1990s, or by about 1.8 million workers per year. Job growth slowed from
2001 to 2005, averaging 600,000 to 1,000,000 additional jobs per year. The years 2001
and 2002 saw net job losses in many occupations and industries because of a recession
that was especially sharp for information technology (IT) and electronics.
After each census, the US BLS develops national employment forecasts, which
are medium-range forecasts by major occupation. These numbers, in turn, are broken
down by major geographic region by various private forecasting services. The
occupational forecasts issued in 2004 extend out to 2014 and indicate continuation of
employment growth in the United States at an average of about 1.3 percent per year.
This is essentially the same rate of job creation as seen in the period from 1994 to
2004.
Table 7-2 summarizes these US BLS occupation outlook projections by major
industry. The combination of health care and social assistance is projected to offer the
greatest numbers of new jobs—4.3 million new jobs adding to the approximately
14 million employed in this sector as of 2004.
Another classification where big gains in jobs are anticipated is the catchall
labeled “administrative and support and waste management and remediation
services.” By 2014, an additional 2.5 million jobs are projected to be offered in this
category, which includes diverse services such as:
• Office administrative functions, such as facilities support, employment
services, business support services, telephone call centers, and collection
agencies
• Travel arrangement and reservation, including travel agencies, tour opera-
tors, and all other travel services
• Investigation and security services, including security systems services
• Services to buildings and dwellings, such as extermination and pest control
and landscaping services
• Janitorial and all other services to buildings and dwellings
• Waste management and remediation
State and local government, professional/scientific/technical services, accom-
modation and food services, and transportation and warehousing are other major
sectors projected to add one to two million additional workers, or more, from 2004 to
2014. Manufacturing employment, on the other hand, is anticipated to continue to
lose jobs, although at a reduced pace compared with the period from 1994 to 2004.
2. These numbers refer to the noninstitutional civilian employment of individuals 16 years of age
or older.
Table 7–2 Employment and Output by Industry, 1994, 2004, and Projected 2014
Thousands of Jobs Change

Industry Title 1994 2004 2014 1994–2004 2004–2014
Nonagriculture wage and salary
Mining 114,984 132,192 150,877 17,208 18,685
Utilities 577 523 477 –53 –46
Construction 689 570 563 –119 –7
Manufacturing 5,095 6,965 7,757 1,870 792
Wholesale trade 17,020 14,330 13,553 –2,691 –776
Retail trade 5,248 5,655 6,131 407 476
Transportation and warehousing 13,491 15,035 16,683 1,543 1,649
Information 2,739 3,138 3,502 400 364
Finance and insurance 5,135 5,966 6,462 831 496
Real estate and rental and leasing 1,732 2,086 2,440 354 353
Professional, scientific, and technical services 4,844 6,762 8,684 1,918 1,922
Management of companies and enterprises 1,666 1,718 1,900 52 182
Administrative, support, waste management 5,664 7,934 10,396 2,270 2,462
and remediation services
Educational services, private 1,895 2,766 3,665 872 898
Health care and social assistance 10,912 14,187 18,482 3,275 4,295
Arts, entertainment, and recreation 1,376 1,833 2,293 458 460
Accommodation and food services 8,724 10,646 12,401 1,922 1,755
Other services (except government) 5,202 6,210 6,944 1,008 734
Federal government 3,018 2,728 2,771 –291 43
State and local government 16,257 18,891 21,019 2,634 2,128
Owner-occupied dwellings — — — — —
Agriculture, forestry, fishing, and hunting 2,890 2,140 1,910 –750 –230
Nonagriculture self-employed and unpaid family workers 9,360 9,556 10,012 196 456
Secondary wage and salary jobs in agrigulture and private 182 138 127 –44 –11
household industries
Secondary jobs as a self-employed or unpaid family worker 1,830 1,587 1,614 –243 28
Total 129,246 145,612 164,540 16,366 18,927
Source: Berman 2005.
Offshoring/Outsourcing. Manufacturing employment in the United States

fell by several million jobs over the 1990s, in part because of the movement of
manufacturing facilities offshore to low-wage areas such as China, Mexico, and
Eastern Europe. More recently, the offshoring (also called outsourcing) trend has hit
service jobs, such as customer call centers and computer programming. According to
experts, these trends are likely to continue to exert pressure on US jobs, having an
impact on both white- and blue-collar employment.
Research by the US BLS identifies the type of jobs that are likely to be most
susceptible to outsourcing. According to US BLS Occupational Employment Projec-
tions (US BLS 2006), from 2002 to 2012, the work most susceptible to outsourcing or
offshoring “can be digitally transmitted, is Internet enabled, includes repetitive tasks,
has clear requirements with few nuances, has little face-to-face interaction with end
users or clients, is not particularly time-sensitive, and is not multidisciplinary.”
By contrast, work relatively secure from offshoring “crosses many disciplines,
requires considerable interaction, includes much uncertainty about specifications,
involves nuances or a deep cultural understanding, and depends on creativity and
innovation.” Continued outsourcing of US jobs is likely to affect the distribution of
pay, because “occupations with median wages within the middle two wage quartiles
[are] more susceptible to offshoring than those in the top and bottom quartiles” (US
BLS 2006). A rough estimate is that the total number of current US service-sector jobs
susceptible to offshoring over the next 10 years is two to three times the number of
current manufacturing jobs, which is about 14 million (Blinder 2006).
Regional and Local Employment Forecasts. Experts frequently take a

top-down approach to generating employment forecasts. They first allocate national
employment growth forecasts to the states and then to metropolitan areas. This
process can employ several regional economic analysis tools such as input-output
tables and employment multipliers for growth in specific types of business and
industry. A careful forecasting process also compares employment and population
forecasts in a metro area to ensure consistency between assumptions of total
population and labor force size. When inconsistencies arise, the experts often adjust
the population forecast by modifying assumptions about migration into the metro
area.
The presumption that employment is more volatile than population is often
reflected in their forecasts. When job losses hit certain areas, younger workers and
their families tend to move to find better employment opportunities. Older workers,
however, are more likely to stay put and wait for new opportunities, perhaps owing to
inertia, sunk investments in housing, strong local ties, or other factors. Some areas are
well-known for lack of outmigration even after years of low economic opportunities.
The 2001 recession is a recent example of the pitfalls of employment forecasts.
In 2001–2003, the areas that had been the fastest growing—because of the electronics
and IT industries—suffered substantial job losses.
REGIONAL ECONOMIC FORECASTS

George Bernard Shaw once joked, “If all the economists in the world were laid end to
end, they still could not reach a conclusion.”
Yet longer-range private and public sector economic forecasts, which can range
from a few to 20 years, often look quite similar. These forecasts usually project an
inflation series such as the consumer price index (CPI), and production and income
quantities such as gross domestic product (GDP), personal income, and disposable
income.
The Federal Reserve Bank of Philadelphia maintains a valuable resource relating
to national economic forecasts, the Survey of Professional Forecasters (available as a
download from the Philadelphia Federal Reserve Web site at www.phil.frb.org/). Once
a quarter, the survey lists forecasts of growth of GDP, inflation, interest rates, and
other key macroeconomic variables for up to two-year horizons. These forecasts are
culled from contributing experts at leading banks, federal agencies, and corporations
and then averaged. Also, private services allow online access to forecasts and
associated databases with an annual subscription.
Staking out a position on the longer-term growth possibilities for the US

economy almost inevitably involves scenario construction. Before striving for
quantitative precision, the planning community must identify plausible scenarios
about the future. Internal consistency and linkage with currently perceived and
emergent tendencies are the primary criteria for realistic scenarios. Scenarios can be
valuable in clarifying how long-term water projects may play out under the variety of
conceivable future conditions.
Personal Income. Many studies reveal a positive relationship between water

use and personal income per capita or per household. An analysis of 27 different
statistical studies estimated the average increase in urban water demand resulting
from a 10 percent increase in real income to be 4.5 percent. Although analysts
generally prefer to use income per household rather than income per capita in a
forecasting model, this variable is not always readily available.
Although this response is less than one-for-one, the cumulative effect can be
significant. For example, if real income per capita (after adjusting for inflation) rises
by 2 to 3 percent per year (as it did after World War II), per capita water demand can
be expected to increase by 20 to 30 percent over 20 years. Ignoring the effect of such
changes clearly introduces significant errors in the long-term water-demand forecast.
Water use goes up with increases in real income because additional water-using
items such as clothes washers, dishwashers, garbage disposals, luxurious landscaping,
and swimming pools become affordable. In addition, as incomes rise, people may be
less concerned with their water bills because these bills typically absorb a smaller
proportion of their income. If these “high-income” water uses approach saturation in
a community, further income increases may have much less effect on water use. When
the use rates of these water-using home features are included in the specification of a
demand model, the estimated size of the income elasticity of demand will diminish
dramatically. The acceptance rates of these water-using features frequently are highly
multicollinear with income and therefore omitted from a demand model specifica-
tion. In this case, the income variable represents the use rates of these home features.
Effect of Inflation. It is appropriate to stress the distinction between real

and nominal monetary variables in economic parlance. In the United States, the US
BLS periodically surveys the prices of a wide variety of goods and services in order to
compute price indexes as measures of inflation. The CPI, based on the cost of buying
a typical consumer’s basket of goods and services, including housing, is the most
widely used measure of inflation. In other words, when the CPI shows an increase, it
takes more money to buy the same market basket of consumer goods.
Consumers may notice how much inflation reduces the purchasing power of
their income. For such consumers, their real, or inflation-adjusted, income is the key
factor in their buying decisions rather than their nominal income, i.e., their wages,
salaries, retirement income, or other compensation for services rendered. This effect
is most obvious over multiple years. Generally, over longer time spans, wage rates
keep up with or exceed inflation rates so that real incomes gradually rise. Many
individuals, however, do not report feeling “richer,” perhaps because the expectations
of living standards also increase.
Income Trends. Analysts must examine individual community characteris-

tics and the pattern of change in the utility service area to evaluate the potential for
income growth. Both the number of jobs and average salary levels are important.
Low-paying service jobs replacing high-paying manufacturing jobs, or the reverse,
would obviously change community water demands if this change occurred on a
sufficiently large scale.
Whether real per capita income will continue to increase at its historical rate of
between 2 and 3 percent is an open question. Real per household incomes have
increased over the past two decades largely because of the emergence of the two-wage-
earner household. The purchasing power of the average paycheck, however, has
declined simultaneously with the increased labor-force participation rate. Because of
the growth in two-wage-earner households, real income improvement is likely to
plateau. Local area income trends are of primary importance to the utility analyst,
however, and these may vary substantially from the national trend. Another trend is
the increasing inequality of incomes. An AWWA committee noted this as one of “Ten
Primary Trends,” and this trend has implications for provision of water services
(Means 2005).
Housing Composition. The housing stock is a major influence on residential

water usage. In general, housing units increase at approximately the same rate as the
population, with some differences based on the average number of individuals per
household. From 1990 to 2000, for example, the US population grew from 226.5 to
282 million, and (occupied) housing units grew from 94.2 million to 105.7 million,
implying an increase in household size from 2.63 in 1990 to 2.67 by 2000.
Some general facts and figures are helpful as background information. In 2003,
the United States had an estimated 121 million housing units, approximately
75 million of which were single-family detached structures. The proportion of single-
family detached homes has remained fairly constant at approximately 60 percent of
the total stock for several decades. Single-family detached homes were at their highest
level in 1960, making up more than two-thirds of the total housing inventory.
Single-family attached houses (row houses and townhouses) were 6 percent of
the total stock of housing in 2003. Their share of the housing inventory was highest in
1940, at 7.6 percent. Apartment housing with two to four units in the building was at
its highest level in 1950, when it made up almost 20 percent of the total housing stock.
By 2003, these smaller apartment complexes dropped to 8.3 percent of the inventory.
The percentage of larger apartment buildings of five or more units increased
dramatically over the past 40 years, from 11 percent in 1960 to more than 16 percent
of the housing stock in 2003. Mobile homes also experienced significant growth for
many decades, comprising 7.4 percent of the total housing stock in 2003.
TECHNOLOGY FORECASTS
One challenge facing water forecasting is projecting the acceptance, acquisition, or
assimilation of water-saving and water-using technologies, appliances, conveniences,
and devices. Several products on the drawing board or in prototype have the potential
to substantially affect future urban water demand. For example, research in the field
of genetic engineering may produce extremely low water-demanding lawns compara-
ble in appearance and feel to existing lawns. In some parts of the country, increasing
numbers of hot tubs, saunas, and swimming pools are increasing water consumption.
Product acquisition or penetration in a community is a diffusion process. The
introduction of a product may be followed by a period of slow acceptance. When 10
or 20 percent of the population acquire the good, the acquisition rate is likely to
accelerate. After more of the population acquires the product, the acquisition rate
slows again, suggesting a type of saturation phenomenon. Eventually, market
penetration reaches a more or less stationary plateau (not necessarily 100 percent).
For costly durable goods such as home appliances, inadequate income or housing
space is the primary barrier to acquisition. Graphically, the diffusion process results in
an S-curve.
The trick in predicting the acceptance rate of some new water-using or water-
saving appliance or device is to figure out the shape of the S-curve over time. The data
disk shows how this S-curve, or sigmoid curve, bends and changes slope as the
underlying diffusion parameters are modified. Expert judgment plays the major role
in establishing a reasonable value for such diffusion parameters. Methods such as the
Delphi process (see appendix A) formalize the process of recording these judgment
calls and continue until a convergent estimate of acceptance rate and penetration is
reached. Field research on diffusion rates is expensive because it requires tracking over
many years, and little information is available in the published literature.
Technology and Industrial Water Use. One major trend in water demand
has been the increasing efficiency of industrial water use. Much of the industrial water
used is self-supplied, but a common challenge of industrial users of potable and self-
supplied systems is the cost of treatment to meet effluent standards. Water use
coefficients or patterns of specific industries vary widely from place to place, from
industry to industry, and from firm to firm.
For individual firms granted special industrial water-use rates, a condition of
service could be firms’ periodic water-use forecasts given to the water utility to aid in
system planning. These multiyear forecasts aid the utility in preparing for changes in
industrial demands and help avoid either excessive or insufficient system capacity.
Space cooling generally uses a large amount of electricity and water (both
directly and indirectly through the generation of electricity) in all sectors of the
economy. Such water use is primarily a function of the size of the cooled space and the
climate. The technology involved generally evolves rather slowly; there is no reason to
expect sudden changes. There is, however, always the possibility of a cost-reducing
technological breakthrough that may be either water saving (e.g., air-cooled compressors)
or water using (e.g., multistage evaporation). Treated effluent could also be used
instead of potable water for some applications.
Electric power generation is a large user of water, most of which is self-supplied.
Since solar or wind power use no water to generate electricity, their increasing share
of electric production has a dramatic impact on future water requirements.
Significant opportunities also exist to substitute treated effluent for potable water in
electric generation. For example, the largest nuclear plant in the United States, the
Palo Verde Generating Station northwest of Phoenix, Arizona, uses effluent rather
than potable water in its generators.
Appliances. The proportion of households with water-using appliances is

only partly a function of income: it is also related to cultural factors, technological
improvements, housing characteristics, and appliance price.
From the late 1940s to the early 1960s, a pronounced change occurred in water-
use habits of the US public. In the years following World War II, average residential
water use per customer climbed steadily as the public adopted automatic clothes
washers, dishwashers, garbage disposals, and other water appliances. The construction
boom in single-family homes also promoted residential water use as new households
watered and cared for their new, and usually large, lawns. Growth of the two-car
family status increased domestic as well as commercial auto-washing requirements,
and widespread use of evaporative air-conditioning in western areas also contributed
to increased water demands (Solley and Pierce 1988).
Although some appliances such as water softeners, garbage disposals, reverse-
osmosis water filters, and evaporative coolers increase community water use, other
appliances, including clothes washers and dishwashers, do not necessarily increase
total water use. Following is a brief review of appliance characteristics related to water
usage.
Clothes washers. Clothes washers are an obvious convenience. Since the wash
must be done one way or another, home washing machines may or may not change
overall community water use, although clothes may be washed more often by those
with the convenience of their own equipment. The water-use efficiency of clothes
washers has increased markedly in recent years, especially with the introduction of
front-loading machines. Water use per capita for clothes washing is likely to drift
downward as older machines are replaced. These are discussed in detail in chapter 10.
Dishwashers. Dishwashers may either increase or decrease household water use,
depending on dishwashing habits. For those who prewash dirty dishes before putting
them into the dishwasher, total water usage is likely to increase. For people who would
otherwise wash everything under running water, but instead simply put it in the
machine, the use of the machine saves water. A machine with a water-conserving
option does not guarantee that people will use it. Many people believe that reducing
water use increases the chance of leaving some dishes imperfectly cleaned.
Garbage disposals. Using sink garbage disposals increases household water use
because it replaces a waterless disposal method, i.e., putting garbage in the trash sack,
with a water-using technology. Community mores and regulations may significantly
affect individual household choices of using or not using the disposals. Some
communities may prohibit disposal use, and others may effectively require their use
by prohibiting the disposal of wet garbage in the regular trash.
Community efforts to encourage or mandate recycling of metal cans, glass
bottles, and so on, may also affect the use of disposals. If a program to institute the
frequent pickup of recyclable materials reduces the frequency of trash collections,
households that previously disposed of garbage in the trash may switch to disposals.
Specifically, if a community switches from twice-weekly trash collection to once
weekly, some water customers may install disposals to reduce the accumulation of
garbage and its odor. Trash-reduction programs that institute a charge per can of
trash pickup may similarly encourage disposal use as households attempt to reduce
both the volume and frequency of the service.
Reverse-osmosis water filters. As households, particularly those with higher
incomes, express increasing concern about the quality of their drinking water, water
filter use is likely to increase. The most efficient current technology to remove
dissolved chemicals appears to be reverse osmosis, which produces a stream of slightly
brackish wastewater. Reverse osmosis uses 5 to 10 gallons of water for each gallon of
drinking water produced.
Water softening. Conventional water softening produces wastewater with a high
salt content. This can increase water use, and the salt added to the wastewater reduces
its value for reuse.
Evaporative coolers. Many homes and small businesses in hot and arid climates
use evaporative coolers as a lower-cost alternative to air-conditioning. Unlike the
other appliances we discuss, evaporative (or “swamp”) coolers have a negative income
elasticity of demand. As incomes rise, households have a tendency to substitute the
much more expensive electric air-conditioning. This substitution, however, is
sensitive to electric rates, leading higher income households to install dual systems
that use coolers when the relative humidity is low and air-conditioning when it is
high. Where evaporative coolers are commonly used, their falling use with rising
incomes may offset increased use of water for other purposes as incomes rise.
CHECKLISTS
Given the need for and volatility of long-range population projections, utility analysts
should adopt a due diligence approach to using population forecasts. A population
forecast meets the due diligence test when the answers to the following questions are
yes:
• Does the population forecast offer high, medium, and low variants or a
confidence interval around the standard case projection?
• Does the population forecast justify the estimate of the existing population
based on information from sources such as birth and death records, school
enrollment, and utility connections?
• Is the forecast consistent with other regional and national forecasts?

• Does the population forecast acknowledge developments such as aging of
the population, projected changes in migration, and possible changes in
birth rates?
• If not, does the population forecast explain why these commonly perceived
trends and developments are not relevant in the utility service area?
• If the population forecast is for a smaller population area, does the forecast
acknowledge the relatively higher potential for variability?
Population forecasts that meet these criteria are probably reasonably objective
estimates. Otherwise, utility planners are well-advised to develop their own popula-
tion growth scenarios, which are useful in exploring the sensitivity and vulnerability
of capacity expansion plans.
Accuracy Issues. Typically, the smaller the region or area of the forecast, the
larger the expected error, because errors tend to net out to a greater degree in
predictions of total population or economic activity for larger areas.
Political pressures can affect employment and population forecasts in metropol-
itan areas. For example, some metro area planning authorities issue population and
employment projections that area municipalities and jurisdictions then use in
applications for federal assistance for highways and other infrastructure construction.
Consequently, an upward bias tends to diminish the accuracy of these projections.
One of the few retrospective studies of forecast accuracy of long-range water-demand
forecasts shows consistent upward bias in population forecasts for the Washington,
D.C. metropolitan area (Hagen et al. 2005).
A Checklist for Regional Economic Forecasts. Analysts should scruti-
nize the assumptions underlying economic forecasts because faulty assumptions will
result in unreliable forecasts. The following questions or issues might be considered
when analyzing economic forecasts:
• What industries and sectors fueled growth in the past?
• Are these industries likely to continue playing an important role in the
future, or will they become “declining sectors”?
• If growth has been led by firms producing goods for sale outside the area, is
the national or international market likely to continue expanding?
• Are local producers competitive?
• If growth has been supported by in-migration, are the factors that made the
region desirable still in place?
• If the region has lost major employers or industries, is there sufficient size
and complexity in the local economy to attract new firms and industries?
Although various long-range economic forecasts often tell basically the same
story, these forecasts are subject to errors over time comparable to population
forecasts. This is not surprising given linkages between the economic and
demographic factors.
Since 1968, the American Statistical Association, the National Bureau of
Economic Research, and now the Philadelphia Federal Reserve Bank have tracked
several dozen macroeconomic forecasting operations. The forecasts, and the actual
numbers, show that predicting turning points in the business cycle is challenging.
Business cycles are not consistent by region and industry: some industries may be
contracting, even dramatically, as others continue growing. The automotive industry
employment in the upper Midwest has declined as employment in other areas of the
United States increased.
The most complex models are not always the ones with the best record in
forecasting. This finding emerges from various “forecasting competitions,” such as the
now-famous “M-Competition” organized by Spyros Makridakis. Another finding is
that many forecasting techniques lose accuracy after about five periods out in the
forecast horizon, regardless of whether the model is developed with weekly, monthly,
quarterly, or annual data (Makridakis 1986; Ashley 1988).

The current “interim” US Bureau of Census population forecast projects straight-line
growth of the US population through 2050. This forecast includes significant
developments relating to the ethnic and age composition of the US population. US
BLS occupational forecasts indicate continuation of employment growth in the
United States at an annual average of about 1.3 percent, essentially the same rate of
job creation as in the period 1994–2004.
A positive relationship exists between water use and personal income per capita
or per household. As real income increases, water-using items such as clothes washers,
dishwashers, garbage disposals, luxurious landscaping, and swimming pools become
affordable. And as incomes rise, water bills become less important because they
absorb a smaller proportion of income.
Plumbing codes requiring greater water-use efficiency for appliances, fixtures,
and various industrial processes continue to exert downward pressure on water use
per capita and per household. Some utility areas are experiencing a significant
downward trend in total water use, reducing the need to increase system capacity.
Analysts can supplement the checklist approach to population and employment
forecasts with scenario development. Instead of striving for quantitative forecasts,
they can combine anticipated elements in a plausible “story” about the future. A
reasonable scenario must be internally consistent and have plausible points of
attachment with current trends. Scenarios can be valuable in clarifying how long-term
water projects may play out under the variety of conceivable future conditions.
Scenarios for the future can and should focus on the complex linkage of
socioeconomic and urban land use factors influencing water consumption. For
example, urban residential land use ranges from low-density areas with single-family
houses on large lots, to medium-density areas with smaller single-family lots and
some multifamily units, to high-density development consisting of larger apartment

or high-rise complexes.
Although indoor water use depends primarily on household size and income,
outdoor use depends on climate and the local consensus about appropriate
landscaping choices. A shift from single-family housing on large lots to apartment or
townhouse living will reduce water use per capita, but may increase water use per
square mile as population densities increase.
Less dramatically, a gradual shift to smaller urban lots or the development of a
social consensus that discourages lush green landscaping may reduce water use both
per household and per acre. Conversely, in an area with rising incomes, social
consensus may push strongly toward larger lots with more extensive landscaping and
thus increased outdoor water use. Sorting out the exact balance of these factors for
periods some years into the future is difficult. It is, however, extremely constructive to
conceptualize how the various elements might fit together. Scenario building offers a
check on forecasts developed by other methods.
REFERENCES
Ashley, R. 1988. On the Relative Worth of Recent Macroeconomic Forecasts. International Journal of
Forecasting 4: 363–376.
Blinder, A.S. 2006. Offshoring: The Next Industrial Revolution? Foreign Affairs 85(2).
Berman, J.M. 2005. Industry Output and Employment Projections to 2014. Monthly Labor Review Online
128(11).
Day, J.C. 1993. Population Projections of the United States, by Age, Sex, Race, and Hispanic Origin: 1993
to 2050. Current Population Reports P25-1104. Washington, D.C.: US Department of Commerce,
Economics and Statistics Administration, Bureau of the Census.
Hagen, E.R., J.K. Holmes, J.E. Kiang, and R.C. Steiner. 2005. Benefits of Iterative Water Supply
Forecasting in the Washington, D.C., Metropolitan Area. Journal of the American Water Resources
Association 41(6): 1417–1430.
Hecker, D. 2005. Housing and Household Economic Statistics Division. Last revised December 16.
Washington, D.C.: U.S. Census Bureau.
Hobbs, F., and N. Stoops. 2002. Census 2000 Special Reports, Series CENSR-4, Demographic Trends in the
20th Century. Washington, D.C.: U.S. Census Bureau, U.S. Government Printing Office.
Makridakis, S. 1986. The Art and Science of Forecasting: An Assessment and Future Directions.
International Journal of Forecasting 2(1): 15–39.
Means, E.G. III, L. Ospina, and R. Patrick. 2005. Ten Primary Trends and Their Implications for Water
Utilities. Journal AWWA 97(7): 64–77.
Pittenger, D.B. 1976. Projecting State and Local Populations. Cambridge, Mass.: Ballinger.
Solley, W.B., and R.R. Pierce. 1988. Trends in Water Use in the United States, 1950 to 1985. In Water Use
Data for Water Resource Management, Proceedings of a Symposium. Bethesda, Md.: American Water
Resources Association.
US Bureau of Labor Statistics. 2006. Occupational Outlook Handbook, 2006–07 Edition. Washington,
D.C.: US Government Printing Office.
US Census Bureau. 2004. 2000 Census of Population and Housing, Population and Housing Unit Counts
PHC-3-1, United States Summary. Washington, D.C.: U.S. Department of Commerce Economics
and Statistics Administration, U.S. Census Bureau. Available online at www.census.gov/prod/
cen2000/phc3-us-pt1.pdf.

CHAPTER 8
Weather and Climate
A 2006 report on climate change and water resources explained the distinction
between weather and climate as follows:
Climate is what you expect, and weather is what you get. In other words,
weather describes the evolution of the current state of the atmosphere, while
climate is a measure of the typical weather for a particular place, hour of
day, and time of year. (Miller and Yates 2006, p. 19)
Climate—basically recurring patterns in temperature, precipitation, cloud
cover, and wind—is a primary driver of seasonal variation in urban water use.
Weather—day-to-day fluctuations in temperature, precipitation, cloud cover, and
wind—is closely linked with short-term variation in water demand. Factors such as
population growth or growth in numbers of customers, on the other hand, exert more
influence on across-year trends in water usage.
This chapter focuses on the features of weather and climate that are directly
relevant to urban water use, and on “forecastability.”
The American Meteorological Society (AMS), the professional association of
weather forecasters, has gone on record to state that absolute limits to the
predictability of weather are reached at forecast horizons of about two weeks.
Essentially, chaotic process dominates the weather pattern after several days of
movement in weather fronts and atmospheric conditions.
The situation is more promising for longer-term weather forecasts, at least on a
probabilistic basis. Major developments such as El Niño can improve the forecasting
of average seasonal temperature and precipitation a year or more forward.
Several water utilities are developing long-term plans based on scenarios related
to “global warming” or “global climate change.” The topic invites controversy, but,
increasingly, major scientific bodies, government-sponsored commissions, and leading
127
scientists project increases in average global temperature for the rest of this century.
These warmer temperatures will trigger more evaporation and greater precipitation.
However, the amount of precipitation, as well as its intensity and distribution, are
likely to change in time. Research on Rocky Mountain watersheds, for example,
predicts changes in the proportion of winter precipitation falling as rain or snow, and,
hence, the timing of runoff. Climate change and global warming are topics that
should be followed closely by urban water managers and their technical staff.
Weather is the first major topic in this chapter. The discussion focuses on
weather patterns and their variability, weather variables, sources of data, statistical
analysis of weather data, simulation of weather data, and weather forecasts. This
section also discusses extreme events such as droughts and heat waves, and so-called
weather cycles.
Climate is the focus of the second major section, which considers some of the
basic science behind the projections of warming, global climate models (GCMs),
downscaling of GCMs, and climate change scenarios integrated into long-range
planning.
WEATHER
Weather Variables. Every day, the US Weather Bureau daily collects data
such as maximum and minimum temperatures, precipitation, snowfall, snow depth,
maximum and minimum relative humidity, peak gust speed, daily total sunshine,
heating and cooling degree days, average daily station pressures, and wind speed.
These data are available for many towns and cities in the United States. Some US
Weather Bureau stations offer hourly data, although most collect only daily and
monthly data.
The National Climate Data Center (NCDC) in Asheville, North Carolina, is the
major repository for US climate data. In addition, six National Oceanic and
Atmospheric Administration (NOAA), Regional Climate Centers produce climate
information relevant to regions of the country. Most states also have a state
climatologist with access to local climate records, and the National Weather Service
(NWS) is another source of information from local offices.
Environment Canada has comparable statistics for Canadian weather data
collection sites.
Average temperature and total precipitation. Urban water-demand models often
are based on average daily or monthly temperature and total daily or monthly
precipitation. Maximum temperatures also play a role, especially in modeling peak
water demands.
Figures 8-1 and 8-2, for example, show average daily temperature and daily
precipitation, respectively, for Atlanta, Georgia, from 1985 to 2006. These high-
frequency data illustrate the variability of weather events over time as well as the
existence of underlying patterns. Statistical analysis suggests that daily temperatures
are conditional on precipitation events.
WEATHER AND CLIMATE 129
100
90
80
Temperature (°F) 70
60
50
40
30
20
10
0
86 :1
19 987 :1
19 :4: :1
20 :4: 8
05 :26
8 30
90 :30
92 :30
93 29
94 29
19 5:4 9
19 :4 9
20 :4 7
6
06 26
9 30
97 :28
9 28
20 0:4 8
02 :27
0 27
04 :27
99 :2
19 :5
1 :5
88 :5
9 2
96 :2
01 :2
:2
0 2
19 :4:
19 :4:
19 :4:
20 :4:
19 :4:
19 :4:
20 :4:
19 8:4
85
20 :4
19 9:4
19 1:4
:4
20 3:4
19
Year:Month:Day
Source: National Climatic Data Center (NCDC), Asheville, N.C.
Figure 8–1 Average Daily Temperatures, Atlanta, Ga.
7
Precipitation Depth (in.)
0
:1 6 1 5 0 :5 1 5 0 5 0 :4 7 2 7 2 :6 :1 7 2 6 1
:5 4:2 4:2 4:1 4:1 :4 3:3 3:2 3:2 3:1 3:1 :3 2:2 2:2 2:1 2:1 :2 :2 1:2 1:2 1:1 1:1
85 6: 7: 8: 9: 90 1: 2: 3: 4: 5: 96 7: 8: 9: 0: 01 02 3: 4: 5: 6:
19 198 198 198 198 19 199 199 199 199 199 19 199 199 199 200 20 20 200 200 200 200
Year:Month:Day
Source: National Climatic Data Center (NCDC), Asheville, N.C.
Figure 8–2 Daily Precipitation, Atlanta, Ga.

Special variables. Analysts often divide standard weather variables, in terms of

their effects on urban water demand, between the irrigation months and the
remainder of the year.
Some studies count only rainfall above a threshold, such as 0.1 in. or the
number of significant days of rain per month.1 This latter refinement might prove to
be beneficial, because three rain days, each delivering 0.5 in. of precipitation, may
reduce the need to irrigate more than one storm delivering 1.5 in.
The effect of temperature or rainfall may also be nonlinear, so a better fit might
be obtained with a two-part variable such as temperatures above and below 86°F.
Adding a quadratic term for temperature or precipitation is another way to deal with
nonlinear effects.
Analysts can use additional weather variables such as humidity, wind speed,
daylight hours, and pan or pond evaporation.
Potential or measured evapotranspiration. Evapotranspiration, the combined
water loss from the soil and the leaf surfaces of plants, is a key element of the
hydrologic cycle. Because both vegetation and soil type control evapotranspiration, it
is difficult to measure. Given the limited availability of field measurements, many
researchers focus on a theoretical value referred to as “potential evapotranspiration,”
which can be computed using National Weather Service climatic and weather data.
Potential evapotranspiration is the water loss from a vegetation cover that never
suffers from a lack of water; it is primarily a function of the energy available for
evaporation.2 Thornthwaite and Mather (1955) fitted evaporation data from water-
sheds and irrigation plots to air temperature, obtaining the following expression for
unadjusted potential evapotranspiration in centimeters per month:
10t a
E = 1.6 ⎛ ------- ⎞ (Eq 8-1)
⎝ I ⎠
where:
t = average monthly temperature, in °C
I = an annual heat index calculated by summing the 12 monthly heat
index values i (see Palmer and Havens [1958] for a table of the
monthly heat indexes)
1. An interesting class of time-series models incorporates “state-dependent” effects (see Miaou

1990). State-dependent effects occur when, for example, the effect of a 1°F increase in average (daily,
weekly, monthly, or annual) temperature is not assumed to be independent of other climatic and
socioeconomic conditions at the time it is experienced. Advanced estimation techniques requiring
numerical optimization and maximum likelihood usually are required. The Miaou application
concludes that the number of days of rainfall per month is a significant weather variable.
2. Water-filled pans such as the US Weather Bureau class A evaporation pan may only approximate
evapotranspiration. For a detailed discussion of the measurement problem and other issues related to
evapotranspiration (see Mather 1984).
a = a function of the heat index as indicated in Eq 8-2:
3 2
I
a = ------------------- I
- + ------------------- I -
- + ------------------- (Eq 8-2)
7 5 2
6.75 ( 10 ) 7.71 ( 10 ) 1.79 ( 10 )
The classic Howe and Linaweaver (1967) research conducted on a nationwide

sample of residential customers uses the Thornthwaite method of estimating
evapotranspiration.
Some research suggests simpler weather variables perform as well or better than
potential evapotranspiration.3 Direct field measurements of evapotranspiration are
sometimes available, and previous measured values can be adjusted for actual
temperature and rainfall after the fact to provide an effective way to model the
weather.4
Extreme weather events. Extreme deviations of weather from the norm include
events such as floods and drought; yet even these extreme events evince patterns.
Figure 8-3, for example, maps subregions of the United States to highlight various
proportions of time in severe and extreme drought, covering the period from 1895 to
1995. In general, areas in the Rocky Mountain West spent the most time in drought
during the 110-year period (Palmer 1958, 1965, 1968).5
Statistical Analysis of Weather Data. Statistical analysis of weather data
starts with basic descriptive statistics, looking at, for example, annual and monthly
averages for a weather variable and variances or standard deviations. Checking
correlations between weather variables is the next logical step. This correlation
analysis also has an important component pertaining to time dependencies within
weather time series—called autocorrelation or serial correlation.
The statistical analysis of weather data offers several basic lessons for urban
water-demand studies:
• Temperature variables (minimum temperature, maximum temperature,
average temperature) usually follow a seasonal pattern.
• Precipitation (rainfall, snowfall) also often follows a seasonal pattern,
although the climographs (in Figures 8-4 through 8-11 show some instances
of relatively even precipitation year-round.
• Subtracting the long-term average for a weather variable for a particular
month or day extracts the seasonal or annual pattern, leaving a seemingly
random series of departures above and below the zero line.
3. Using data for 33 cities in Southern California for 1970, Douglas and Smolen (1976) compared
the performance of potential evapotranspiration with average monthly temperature and precipitation.
4. Billings and Agthe (1980) used a normalized monthly evapotranspiration series, adjusted for
actual temperature and rainfall.
5. The Palmer Drought Severity Index is a soil moisture algorithm applicable for regions with
relatively homogenous features utilized by many US and state agencies.
Percentage of Time
PDSI <_ –3
Less than 5
5–9.99
10–14.9
15–19.9
20 or greater
Source: National Drought Mitigation Center, University of Nebraska, Lincoln.
Figure 8–3 Palmer Drought Severity Index (PDSI; Percentages of Time in Severe and Extreme
Drought, 1895–1995)
• These residuals often are “normally distributed.” Thus, collecting all these
residuals for the same date and charting them in a histogram usually
produces a distribution that is approximated by the normal probability
distribution.
• Viewed over successive months or days, these residuals usually exhibit
positive first order autocorrelation. In ordinary language, if the residual for
one month or day is above the zero line, there is a definite probability that
the residual for the following month or day also will be above the zero line.
This means, for example, that unusually hot days tend to be followed by
other unusually hot days or that, in many locales, rainy weather tends to
occur in clusters of days.
• Daily observations on weather variables such as temperature and precipita-
tion are correlated.
Weather Cycles. One way of looking at weather cycles is to tally heat waves.
Ruffner (1987) explains:
“There is nothing in American climatological annals to touch the heat waves
which came with the Dust Bowl droughts of the 1930s. The years 1930, 1934, and
1936 brought progressively more severe summer weather. Record highs of 121 degrees
in North Dakota and Kansas, and 120 degrees in South Dakota, Oklahoma, Arkansas,
and Texas were observed in the ugly summer of 1936. July and August of that year saw
record highs of 109 degrees or better tied or broken in Indiana, Louisiana, Maryland,
Michigan, Minnesota, Nebraska, New Jersey, Pennsylvania, West Virginia, and
Wisconsin. … The heat wave of July 1966 covered much of the eastern and middle
continent with high temperatures and very high humidity. ... The highest death toll
for a year since the 1930s belongs to 1980, according to the National Center for Health
Statistics. That year saw unusually high temperatures hold the central and southwest-
ern United States in their grip for the best part of 15 weeks. … The 1980 heat wave ...
began about June 10, and except for remissions from about July 21 to August 10, and
August 17 to September 1, lasted continuously through September's third week. As an
example of its intensity, the first week of September saw temperatures of 9 degrees
above normal over much of the central and eastern states.”
North American weather patterns are linked with the position and orientation
of the jet stream. As winter abates and the continent warms, the jet stream usually
shifts northward, moving thundershowers to the Plains and humidity to the Midwest.
A “dust bowl pattern” exists if the prevailing winds from the Southwest sweep farther
north than usual in a more or less self-perpetuating circulation pattern.
Heat waves and drought events appear to be periodic. The climatologist
J. Murray Mitchell conducted a statistical study of tree rings and found a 20- to 22-
year drought cycle, persistent since 1700 (Schneider and Londer 1984). Local
exceptions usually exist with respect to area-wide heat waves and droughts. Studies of
the cycle of extremely dry, hot weather for specific localities often suggest the
existence of a tendency for recurrence at a more or less fixed cycle. Such tendencies,
however, usually explain only a fraction of droughts and predict drought periods
when none has occurred in the historical record (Kane and Teixeira 1991).
Weather cycles, therefore, do not appear to be sufficiently predictable for
inclusion in long-term water-demand forecasts. Analysts should, however, examine
the historical record of temperature and precipitation to estimate the likely range of
variability of these key factors.
Weather Forecasting. Weather is notoriously difficult to predict. Evidence

points to the existence of a “limit of predictability” for deterministic weather forecasts
of only a few days into the future. Beyond about two weeks, the chaotic aspect of local
atmospheric conditions overwhelms regularities used in developing the forecast.
Thus, the current location and condition of weather fronts give near-term
information that is predictive or determinative. These weather fronts, however, can
evolve wholly unpredictably over one to two weeks.
Mathematically, this is a chaotic process. In the early years of climate
simulation, the MIT meteorologist Edward Lorenz developed computer models
whereby the flap of a butterfly in one location could trigger tornados on the other side
of the world—giving rise to the famous “butterfly effect” in chaos theory. The point
is that chaotic processes evolve unpredictably from initial conditions. Small
differences in current wind speeds, extent of the front, direction of its movement, and
other physical characteristics can—through chaotic localized processes—lead to
wholly unpredictable results a few weeks later.
The American Meteorological Society, the professional association of weather
scientists, has gone on record that “theoretical and experimental studies” set the limit
of predictability of day-to-day weather changes at one or two weeks:
“While great progress has been made on [weather] forecasts up to almost a
week, all current applied and theoretical studies lead to the conclusion that the day-
to-day fluctuations in weather are not predictable beyond one to two weeks. Beyond
that time, errors in the initial conditions (i.e., the data that define the state of the
atmosphere at the start of the forecast period) no matter how small they are assumed
to be, grow to the point that they overwhelm whatever valid information that the
forecast might contain.”
The AMS further asserts that “claims of skillful predictions of day-to-day
weather changes beyond this limit have no scientific basis and are either misinformed
or calculated misrepresentations of true capabilities” (AMS Council 2001, p. 701).
Precision in short-term water-demand forecasts, accordingly, is limited to
short-range applications—a matter of hours or a week or so. Utilities might use such
forecasts to fine-tune operations of pumps and reservoir gates.
Interannual weather forecasts. Although there are limits to short-term weather
forecasts, weather forecasts extending across adjacent years are increasingly successful.
These must be cast in terms of probabilities, however, rather than deterministic
projections.
The atmosphere is especially sensitive to tropical sea surface temperature
anomalies such as El Niño or, to use its full name, the El Niño/Southern Oscillation
(ENSO). This creates a scientific basis for predicting seasonal mean climate anomalies,
or departures from normal of averages and other statistics of weather over periods of
a season or longer. The atmosphere fluctuates rapidly day to day, but is tied to more
slowly evolving components of Earth’s system. The interaction, or coupling, between
the oceans and the atmosphere at seasonal timescales was first appreciated and
understood in the context of the ENSO phenomenon: the episodic warming and
cooling of the sea surface temperature in the equatorial central and eastern Pacific
(popularly known as El Niño and La Niña events). In addition to the surface and
subsurface of the equatorial Pacific Ocean, the ENSO also encompasses the ocean's
interactions with the global atmosphere on timescales of several seasons.
Simulating Weather. Weather simulations have several applications to water

utility programs and policies, including water-demand forecasts. One important
application is in developing more realistic confidence intervals for peak monthly and
peak daily demands. In general, when water-demand models incorporate weather
variables, weather simulations can produce information about the likely range of
water usage for particular periods of time.
Computer programs called stochastic weather generators, for example, simulate
temperature, precipitation, and other weather variables based on patterns in the
historical data. Generally, this software keys off the fact that high-frequency
observations of weather variables exhibit first order autocorrelation. That is, higher
(or lower) than normal values for a particular time period have a statistical tendency
to be followed in the following time period by other higher (or lower) values
compared to the norm. In addition to these autocorrelation patterns, there also are
correlations between weather variables. Precipitation often is pivotal, for example. As
a consequence, many weather simulators first produce numbers for precipitation
occurrence and amount and then generate values for other weather variables based on
conditional probability distributions. That is, other weather variables are generated
for each weather station using probability distributions that are conditional on the
amount of precipitation for that day.
One problem with this approach is that causation is not exclusively one-way.
Precipitation occurrence and amount for a particular day may be related to the values
of other weather variables on that day. There also can be correlations between other
weather variables, e.g., maximum temperature and mean temperature.
Nonparametric methods generate weather sequences by resampling from the
historical record, which may improve on these parametric methods. These resampling
methods assume past meteorological conditions are representative of what may occur
in the future. They compare available information on weather variables for a given
day, t, with comparable information from similar dates in the historical record. For
example, the k-most similar days can be taken as the k-nearest neighbors, where k =
square root of the number of similar dates used in the comparison. One of these
neighbors is randomly selected, and the day following the selected neighbor is taken
as the next simulated day (day t+1). By resampling from all the data in the historical
record, analysts can preserve unique spatial and temporal dependencies and make
allowance for simple and effective multivariate, multisite weather generation (Raja-
gopolanm and Lall 1999).
For longer periods, of a decade or more, climate change is likely to alter basic
parameters of weather in almost any locale, including those parameters that underpin
simulations of weather variability. There appears to be no easy fix for adjusting
historical temperature and precipitation series for these changes.
CLIMATE
Climographs, shown in Figures 8-4 through 8-11, help characterize the climate of
urban regions. These graphs, prepared by the National Drought Mitigation Center,
show average monthly temperature and precipitation. Significant features include the
level and variation of summer temperature and precipitation. Phoenix, Arizona, is hot
and dry, and Tampa, Florida, is warm and wet. Seattle, Washington, has a cooler, wet
climate. San Francisco experiences only minor differences between summer and
winter average monthly temperatures, which never fall below or even approach
freezing.
Average temperatures in Chicago and Kansas City fluctuate between freezing
and significant increases in summer months. Summers are humid, because peak
precipitation occurs with higher temperatures in late spring and summer. Climo-
graphs for Atlanta, Austin, and Boston are included, highlighting subtler differences
and patterns.
Utilities should review climographs with data for monthly average and peak
water usage for an urban area; such comparisons can highlight a complex interplay
between climate and water use as well as significant discontinuities in outdoor water
use through the year.
14
100
12
80 10
Precipitation (in.)
Temperature (°F)
60 8 Average Monthly
Precipitation
6 Average
40 Temperature
4
20
2
0 0
M y
ne
ch
Fe ary
ay
Oc ber
pt st
ce er
A ly
r
ve er
ril
r
be
ua
Ju
Se ugu
M
De mb
ar
Ap
Ju
No tob
nu
em
m
br
Ja
Source: NCDC database.
Figure 8–4 Climograph for Austin, Texas
90 14
80
12
70
10
Precipitation (in.)
60
Temperature (°F)
Precipitation
40 6 Average
Temperature
30
4
20
2
10
0 0
ve er
ne
M ry
ce er
r
ch
Fe ary
Oc ber
pt st
A ly
ay
ril
be
ua
Ju
No tob
De mb
Se ugu
Ju
ar
Ap
M
nu
m
em
br
Ja
Figure 8–5 Climograph for Boston, Mass.

90 14
80
12
70
10
60
Precipitation (in.)
Temperature (°F)
50 8
Average Monthly
Precipitation
40 6 Average
30 Temperature
4
20
2
10
0 0
ch
ne
br y
r
ve er
M y
pt st
ce er
Oc ber
ay
ril
Au ly
be
Fe uar
r
ua
Ju
Se g u
No tob
ar
Ap
De mb
Ju
M
m
em
n
Ja
Figure 8–6 Climograph for Chicago, Ill.
14
100
12
80 10
Precipitation (in.)
Temperature (°F)
Precipitation
6 Average
40 Temperature
4
20
2
0 0
ne
ch
pt st
ay
r
A ly
Oc ber
ce er
ve er
M y
ril
Fe ary
be
r
Ju
Se ugu
ua
ar
De mb
Ju
Ap
No tob
m
nu
em
br
Ja
Figure 8–7 Climograph for Phoenix, Ariz.

14
100
12
80 10
Precipitation (in.)
Temperature (°F)
Precipitation
6 Average
40 Temperature
4
20
2
0 0
ne
ce er
em t
ay
ly
ve er
M ry
r
ril
ch
Fe ary
Oc er
pt us
be
Ju
De mb
ua
Ap
No ob
M
Ju
b
ar
Se Aug
nu
m
br
t
Ja
Figure 8–8 Climograph for Kansas City, Mo.
14
100
12
80 10
Precipitation (in.)
Temperature (°F)
Precipitation
6 Average
40 Temperature
4
20
2
0 0
Oc ber
ne
A ly
ve er
M ry
ch
r
Fe ary
pt st
ce er
ay
ril
be
Ju
ua
No tob
Se ugu
Ju
De mb
ar
Ap
em
nu
m
br
Ja
Figure 8–9 Climograph for Los Angeles, Calif.

14
100
12
80 10
Precipitation (in.)
Temperature (°F)
Precipitation
6 Average
40 Temperature
4
20
2
0 0
ne
ch
A ly
Oc ber
ay
ce er
r
ve er
M ry
ril
Fe ary
pt st
be
Ju
ua
ar
Ju
Ap
De mb
No tob
M
Se ugu
em
nu
m
br
Ja
Figure 8–10 Climograph for San Francisco, Calif.
14
100
12
80 10
Precipitation (in.)
Temperature (°F)
Precipitation
6 Average
40 Temperature
4
20
2
0 0
ch
ne
Fe ary
Oc ber
A ly
pt st
ay
r
ve er
M y
ce r
ril
be
De mbe
r
Ju
Se ugu
ua
ar
Ap
No tob
Ju
M
nu
em
m
br
Ja
Figure 8–11 Climograph for Tampa, Fla.

Medium-Term Predictions of Climate. Scientifically based seasonal cli-

mate prediction (e.g., predicting the coming season’s average temperature and
precipitation) has come of age over the past two decades. Until 20 years ago, seasonal
climate predictions were based exclusively on empirical and statistical techniques that
provided little understanding of the physical mechanisms responsible for relationships
between current conditions and the climate anomalies (departures from normal) in
subsequent seasons. Researchers still use empirical and statistical tools for seasonal
prediction, but now within the context of a greatly improved understanding of the
underlying physical processes.
Researchers are also increasingly using mathematical models analogous to those
used in numerical weather prediction, but including representation of atmosphere–
ocean interactions, in conjunction with or as an alternative to empirical methods.
Coastal areas may have an advantage in translating forecasts of global climate
models (GCMs) down to smaller geographic areas, such as a watershed, because of the
proximity of the ocean and its temperature effects. For many parts of the nation,
however, these elaborate computer models are not reliable when it comes to
predicting precipitation over local areas.
Heat-Island Effects. Climatic trends such as heat-island effects and global
warming encompass potential effects that could be included in long-run urban water-
demand forecasts. The issue is whether these factors are understood well enough to be
used. Heat-island effects are increases in average temperature throughout the year in
the microclimate of larger metropolitan regions. This phenomenon is well
documented and could be included in long-term water-demand forecasts.
Temperature trends can be identified in many major urban areas (see Cayan
and Douglas 1984; Maidment and Djokic 1990). Many areas of Los Angeles,
California, for example, experience heat-island effects, with annual temperatures
rising on average 0.5°F per decade.
Because temperature trends resulting from heat-island effects are statistically
significant, they appear to be suitable for inclusion in long-term forecasts. Their use
requires predicting the urban growth rate. Because a growth rate forecast of
population, commercial development, and industrial location is central to every long-
term water-demand forecast, adding heat-island effects does not add significant
complexity to the forecasting task and may slightly increase forecast accuracy.
Climate Change. Climate change has the potential to significantly affect
urban water demand as well as watershed and other supply conditions of public water
supply systems. The recent National Research Council (NRC; 2006) study concludes
that there is an increasing likelihood that warming is occurring on a global scale and
that anthropomorphic sources are a major contributor.
Carbon dioxide, a by-product of combustion, is the major culprit, although
other greenhouse gases, such as methane, are significant contributors. Historical data
on atmospheric carbon dioxide concentrations going back 650,000 years or more have
been obtained from ice core drillings in the Antarctic and Greenland ice sheets.
Coupled with other information, such as tree ring data and isotopic concentrations in
ice bubbles, scientists can piece together historic data series on carbon dioxide and
global temperature.
Carbon dioxide forms a sort of blanket in the atmosphere, trapping re-radiation
of solar energy from Earth’s surface—the “greenhouse effect.” Carbon dioxide
concentrations are projected to double and triple because of general economic growth
and regional factors such as dozens of new coal-fired power plants in China and India.
There is a significant probability that major climate change will follow on this
movement of CO2 concentrations to significantly above their highest previous values
in the past 1-million-year span of history.
Even if the global community moves to rapidly reduce carbon dioxide
emissions, global average temperatures are forecast to continue to increase for many
decades.
Some of the most dramatic effects of global warming on water demand will be
via the number of extremely hot days. A 2°F hike in average temperature results in a
substantial increase in the probability of long runs of hot days. Global warming also
intensifies the hydrologic cycle; evaporation increases with increases in average
temperature, and so, also, does rainfall. The expectation is that weather events will
become more extreme as more energy is pumped into the global climate system.
For the Rocky Mountain West and the West Coast, there is a presumption that
watersheds will experience significant changes, affecting river runoff. There may be
less snowfall and more rainfall, changing the pattern of surface runoff in the spring,
which feeds many reservoirs throughout the West. The impacts on water demand
relate especially to temperature effects. Runs of hot days are increasingly likely, with
the potential to push up peak demand.
Global Climate Models. Basically, modern Global Climate Models (GCMs)

involve an array of models—models for the atmosphere; models for the oceans,
including surface temperature and deeper circulation patterns; models of evaporation,
soils, vegetation, and topography on land areas; and polar ice models. These systems
are coupled together with supercomputers to trace the interactions predicted by
physics. One issue is that the underlying partial differential equations refer to
mathematical points. Implementing the governing equations in GCMs, however,
requires reference to discrete volumes. Currently, these atmospheric cells can be as
large as several hundred kilometers on a side. Resolving to smaller grids is a
continuing objective of global climate modeling.
The problem for urban water utility planners is translating from GCMs to the
specific circumstances of watersheds and the utility service area. As mentioned, GCMs
currently operate with volumes and areas that are scores or hundreds of miles
(kilometers) on a side. Translation from these large cells to regional and local climate
impacts is called “downscaling.” Downscaling to areas more relevant to public water
supply systems often produces mixed signals about precipitation. Nevertheless,
researchers are reducing uncertainty for certain regions, and water utilities are quick
to incorporate these results in long-range planning.
Implications for Demand Forecasts and Water Utility Planning. Climate

change is likely to affect urban water demand and water utility planning for future
water resources. More and more water utilities are conducting research on potential
impacts of climate change and global warming. This requires downscaling GCMs to a
regional basis, a problematic procedure given their current granularity. Global
warming means intensification of the hydrologic cycle—more evaporation, more
rainfall, and, probably, more extreme weather events. Areas in the continental United
States will become warmer and dryer. Other locations will see warmer, wetter
conditions. Still other places in North America may experience cooler conditions,
which again may be either wet or dry. Generally, downscaling temperatures has been
more decisive than predicting precipitation. Nevertheless, precipitation is crucial for
water utility planning.
On the demand side, computations in chapter 6 show typical results. Summer
water use in an urban water system is likely to increase only a few percentage points
with an increase in average summer temperatures of one to a few degrees. As Boland
(1997) notes, this can be interpreted to mean that demand effects of climate change
are “small” relative to business growth and effects of water conservation. Such
conclusions must be qualified, however, in view of the longer runs of extremely hot
weather that, in many locales, are likely to accompany increases in the average
temperature.
More fundamentally, increased uncertainty is a basic fact that must be dealt
with, on both the demand and supply side. Miller and Yates (2006) compile case
studies illustrating approaches undertaken by water utilities to come to grips with
impacts of climate change.
The recommended solution at present appears to be developing plans for
infrastructure and management of current programs that are cost-effective across a
range of future scenarios. This may be accompanied by “trigger points” in planning,
where decisive action can be taken when the future mix of temperatures and
precipitation comes more clearly into view.

A generic water-demand equation for a residential customer might be characterized
by an equation, such as:
q = f ( R, H , D , Z ) (Eq 8-3)
where:
R = vector of water rates and charges
H = vector of household characteristics
D = vector of demographic chacteristics
Z = vector of weather and other environmental variables.
Predictions of future values of R, H, and D usually are possible. As documented

in this chapter, however, there are significant limitations on capabilities for forecasting
weather, at least in the near term. Climatic variation over seasons is relatively
predictable, as the several climographs for major urban areas in the United States
suggest. Interannual forecasts of a probabilistic sort may be possible by year—sort of
the modern equivalent of the Farmer’s Almanac.
Demand models can be constructed with hindsight. These models can be fitted
to monthly or daily variations in temperature and precipitation, explaining a high
percentage of the variation in the water usage data.
Prospectively, however, it usually is possible to forecast only the average pattern,
as in the monthly average pattern. It is difficult and, in many cases, impossible to
forecast deviations of average temperature or precipitation from the seasonal or
monthly norm.
Short-term demand models—operating on a scale of days—may key off of
weather forecasts for a few days forward and provide forward-looking information
about total system demand during peak demand periods.
Utilities can use water-demand models operating on other time scales,
incorporating weather variables, in risk analysis. For this reason, it is relevant to
understand how to simulate weather.
Climate change has the potential to significantly affect urban water demand, as
well as watershed and other supply conditions of public water supply systems.
Currently, however, GCMs are not consistent in their capability for forecasting
precipitation for local areas, even though their general temperature forecasts appears
to be more sound.
The bottom line is that water managers and planners must plan for the long-
term future in the context of considerably greater uncertainty.
The best current solution appears to be developing plans for infrastructure and
management of current system that are cost-effective across a range of future
scenarios.
REFERENCES
Aly, A.H., and N.W. Wanakule. 2004. Short-Term Forecasting for Urban Water Consumption. Journal of
AMS (American Meteorological Society) Council. 2001. Statement on Seasonal to Interannual Climate
Prediction, Adopted 14 January 2001. Bulletin of the American Meteorological Society 82:701.
AMS Council. 1998. Weather Analysis and Forecasting, Adopted 17 August 1998. Bulletin of the American
Meteorological Society 79:2161–2163.
Barnston, A.G., A. Leetmaa, V.E. Kousky, R.E. Livezey, E.A. O’Lenic, H. van den Dool, A.J. Wagner, and
D.A. Unger. 1999. NCEP Forecasts for the El Niño of 1997–98 and Its U.S. Impacts. Bulletin of the
American Meteorological Society 80:1829–1852.
Billings, R.B., and D.E. Agthe. 1980. Price Elasticities for Water: A Case of Increasing Block Rates. Land
Economics 56(2): 73–84.
Billings, R.B., and C.V. Jones. 1996. Forecasting Urban Water Demand. Denver, Colo.: American Water
Works Association.
Boland, J.J. 1997. Assessing Urban Water Use and the Role of Water Conservation Measures under
Climatic Uncertainty. Climatic Change 37(1): 157–176.
Cayan, D.R., and A.V. Douglas. 1984. Urban Influences on Surface Temperatures in the Southwestern
United States during Recent Decades. Journal of Climate and Applied Meteorology 23(11): 1520–
1530.
Changnon, S.A. 2002. Impacts of the Midwestern Drought Forecasts of 2000. Journal of Applied
Meteorology 41(10): 1042–1052.
Douglas, M.w., And J.c. Smolen. 1976. Climatic Indicators In The Estimation Of Municipal Water
Demand. Water Resources Bulletin 12(3): 511–518.
Edwards, D.C., and T.B. McKee. 1997. Characteristics of 20th Century Drought in the United States at
Multiple Time Scales. In Climatology Report Number 97–2. Fort Collins, Colo.: Colorado State
University.
High Plains Regional Climate Center. Albers Equal Area Projection. Lincoln, Neb.: University of
Nebraska. Online at http://hprcc.unl.edu/index.php.
Howe, C.W., and F.P. Linaweaver Jr. 1967. The Impact of Price on Residential Water Demand and Its
Relation to System Design and Price Structure. Water Resources Research 3(1): 13–32.
Kane, R.P., and N.R. Teixeira. 1991. Power Spectrum Analysis of the Annual Rainfall Series for
Massachusetts (NE, USA). Climate Change 18(1): 89–94.
Lorenz, E. 1963. Deterministic Nonperiodic Flow. Journal of Atmospheric Science. 20:130–141.
Maidment, D.R., and D. Djokic. 1990. Weather-Related Water Demand Variability in Metropolitan Water
District Service Area. Consulting report submitted to the Metropolitan Water District of Southern
California.
Mason, S.J., L. Goddard, N.E. Graham, E. Yulaeva, L. Sun, and P. A. Arkin. 1999. The IRI Seasonal
Climate Prediction System and the 1997/98 El Niño Event. Bulletin of American Meteorological
Society 80:1853–1873.
Mather, J.R. 1984. Water Resources: Distribution, Use, and Management. New York: John Wiley & Sons.
Miaou, S.P. 1990. A Class of Time Series Urban Water-Demand Models with Nonlinear Climatic Effects.
Water Resources Research 26(2): 169–178.
Miller, K., and D. Yates. 2006. Climate Change and Water Resources: A Primer for Municipal Water
Providers. Denver, Colo.: Awwa Research Foundation (AwwaRF).
NCDC (National Climatic Data Center). Climatography of the United States, No. 20 1971-2000. NCDC,
Asheville, N.C.: NCDC. Available online at www.drought.unl.edu/whatis/climographs/pdf/Seat-
tle.pdf
NOAA (National Oceanic and Atmospheric Administration). 1991. Local Climatological Data, Annual
Summary for 1990, Los Angeles, Calif. Ashville, N.C.: NOAA.
NRC (National Research Council). 2006. Surface Temperature Reconstructions for the Last 2,000 Years.
Washington, D.C.: National Academy Press.
NWS (National Weather Service). 1988. A Result for Temperature and Precipitation. NOAA Technical
Memorandum, NWS FCST 31, August.
Palmer, W.C. 1965. Meteorological Drought. Research Paper No. 45. Washington, D.C.: US Department of
Commerce Weather Bureau.
———. 1968. Keeping Track of Crop Moisture Conditions, Nationwide: The New Crop Moisture Index.
Weatherwise 21:156–161.
Palmer, W.C., and A V. Havens. 1958. A Graphical Technique for Determining Evapotranspiration by the
Thornthwaite Method. Monthly Weather Review (April):123–128.
Rajagopolanm, B., and U. Lall. 1999. A k-Nearest Neighbor Simulator for Daily Precipitation and Other
Weather Variables. Water Resources Research 35:3089–3101.
Ruffner, J.A., and F.E. Bair, eds. 1987. The Weather Almanac: A Reference Guide to Weather, Climate, and
Air Quality in the United States and Key Cities. 5th ed. Detroit, Mich.: Gale Research Company.
Schneider, S.H., and R. Londer. 1984. The Coevolution of Climate and Life. San Francisco: Sierra Club
Books.
Stickel, L., 2006. The Challenges for Incorporating Climate Change and Variability Information into
Water Resources Decision Making. Paper presented at the AWWA Sources Conference, Albuquer-
que, New Mexico, February 6.
Thornthwaite, C.W., and J.F. Mather. 1955. The Water Balance, Publications in Climatology. South
Centerton, N.J.: Drexel Institute of Technology, Laboratory of Climatology.
VanRheenen, N.T., R.N. Palmer, and H.A. Hahn. 2003. Evaluating Potential Climate Change Impacts on
Water Resources Systems Operations: Case Studies of Portland, Oregon and Central Valley,
California. Water Resources Update 124:35–50.
Wagner, J.A. 1989. Medium and Long-Range Forecasting. Weather and Forecasting 4(3): 413–426.
White, D.H., and B. O’Meagher. 1995. Coping with Exceptional Droughts in Australia. Drought Network
News 7(2): 13–17.
Wilhite, D.A. 1995. Developing a Precipitation-Based Index to Assess Climatic Conditions across Nebraska.
Final report submitted to the Natural Resources Commission, Lincoln, Neb.
Wilhite, D.A., and M.H. Glantz. 1985. Understanding the Drought Phenomenon: The Role of
Definitions. Water International 10(3): 111–120.
Willeke, G., J.R.M. Hosking, J.R. Wallis, and N.B. Guttman. 1994. The National Drought Atlas. Institute
for Water Resources Report 94–NDS–4. Fort Belvoir, Md. US Army Corps of Engineers. Available
online at www.iwr.usace.army.mil/iwr/atlas/Atlasintro.htm.

CHAPTER 9
Price Effects
When the well is dry, we learn the worth of water.

—Benjamin Franklin
What is water worth? In a society that often seems to be dominated by money—how

to get it, spend it, and save it—we may think of the value of water as the price we pay
for using it. On reflection, however, it becomes obvious that the intrinsic value of
water depends on how it is used. Water used to irrigate a field or fill a fishing stream,
although valuable, has far less value per unit than water for most household uses. The
extreme example of this difference is the price many people willingly pay for bottled
drinking water, when tap water in many areas is of equal or higher quality. Although
most residents of North America and Europe are accustomed to seemingly unlimited
supplies of high-quality water at a low cost, households in many poor countries spend
a considerable proportion of their total resources for water—sometimes walking miles
to obtain enough for drinking and cooking. An analysis by Brown (2006) of 622 water
sales in the western United States from 1990 to 2003 found a median price of
$1,955/ML,1 with some upward trend. This chapter will not resolve the question of
the value of water2; it will, however, carefully examine the effect of water prices on
water-use decisions by households, business, and government.
1. One ML = 0.81 acre-foot.

2. For discussions of the value of water, see Merrett 2005, and Young 2005.
147
Water customers cut back water purchases when water rates increase, with
effects noted in the short term (weeks or months) and long term (years). Over short
time periods, customers first reduce water uses that are easy to modify and that they
value least, such as washing sidewalks, buildings, and cars, and, to an extent, irrigating
landscape. They may also simply stop watering their grass, or make personal
adjustments in water use such as taking shorter showers. Over longer time periods,
customers may replace water-using appliances, devices, and equipment with more
water-efficient models, adopt low-water-use landscaping innovations, and make other
investments that result in increased efficiency in water use.
These effects often are subtle, given the typically small water bills and minimal
increases in water rates, and are ensured only if water rates keep pace with inflation.
Customer responses to water rates and fees, however, translate directly to the bottom
line—water utility revenues. It is important, then, to strive for accuracy in projecting
the effect of changes in water rates.
This chapter presents an overview of common types of water rate and fee
schedules, with examples of price elasticities of demand and revenue. The discussion
includes (1) effects of water metering, (2) controlled comparisons of communities
with similar socioeconomic and weather conditions but different levels of water rates,
and (3) econometric studies of water demand. The chapter provides rules of thumb
for including price effects in the demand forecast and the effect on revenue of changes
in water sales under each type of rate schedule.
DEMAND CURVES FOR WATER

Economists often speak of the demand for water, assuming their audience knows what
this means. The simplest demand function includes only price and quantity, and
shows the amount of water that would be purchased at each possible price. All other
influences on water purchases, such as personal income, temperatures, and rainfall,
are assumed to be constant. The demand curve could describe the possible purchase
of water by one individual, a neighborhood, or an entire utility service area. Figure 9-1
shows two hypothetical demand curves for water, one linear and the other
exponential (or logarithmic).3 Either curve would adequately describe a customer’s
response to price variations between about $1.00 and $1.80 per unit of water. Beyond
this range, the curves diverge enough so that the choice of the form of the demand
function matters. (Note: Figure 9-1 shows a much broader range of prices than those
that would be enacted by an actual utility.)
When the price of water is very low, the linear curve may be more realistic,
because this example says that there is a definite limit on how much water people
would use, even at a zero price. Specifically, it predicts that at a zero price, a little more
than 18 units of water would be used. This matches well with various studies of
residential water use cited later in this chapter and in chapter 10.
3. The equations for these curves are Q = 18.2 – 3.7143 × P and Q = 0.93842 × P–0.4.
PRICE EFFECTS 149
6
Price in Dollars per Unit
5
0
0 2 4 6 8 10 12 14 16 18 20
Quantity in Units per Month
Linear Exponential
Figure 9–1 Typical Linear and Exponential Demand Curves
When the price of water is much higher, however, the logarithmic curve may be
more realistic, because it shows some water use at every price. In contrast, the linear
curve shows water use going to zero when the price rises to about $5.00 per unit. This
is unrealistic, at least in North America and Europe, considering that even at $5.00 per
unit, water would absorb only a very small part of most consumers’ budgets and that
most urban consumers do not have a feasible alternative to water delivered by the
utility. In low-income countries, however, even a modest increase in water prices may
prevent many consumers from buying water from the utility, and leave them
dependent on often contaminated and limited surface water supplies. Commercial
bottled water is often used for drinking, but other household water uses generally
make this substitution trivial in its effect on utility water sales. Increasing bottled
water sales may, however, decrease consumers’ willingness to absorb any water rate
increases necessary to maintain or improve delivered water quality.
SCHEDULES OF RATES AND FEES

Many types of rate schedules and fee arrangements exist in the potable water industry,
including the following:
• Flat rate, equal for all customers, or based on property characteristics
• Monthly service fee and constant volume charge
• Decreasing block rate schedule
• Increasing block rate schedule and lifeline rates
• Minimum use included with the service charge

• Emergency drought rates to ration shortages
• Seasonal rates
• Step rates
• Marginal cost and rebate
• Customized rates based on historical usage or family size
• Negotiated rates with very large volume nonresidential customers
Flat Rate. Some utilities do not meter water use by individual customers, so
they charge a fixed periodic fee for water use. This may be the same for all customers
or be determined by the size of the customer’s connection. The fee could also vary
with characteristics of the property, such as size of building or lot, number of
bathrooms, zoning, or use of property. Commercial customers might be charged
based on their business classification, sales volume, number of employees, or other
observable features. The most important fact about any of these fee arrangements is
that there is no direct connection between the customer’s bill and water use. These fee
schedules are appropriate when water supplies are plentiful and delivered cost is low.
Monthly Service Fee and Constant Volume Charge. This schedule com-
bines a monthly service fee for each metered connection—which often varies with the
size of the connection—with a constant volume charge for the amount of water used.
Figure 9-2 shows a typical schedule with a marginal price (the price for the last unit
sold) of $2.80 per unit, regardless of the quantity purchased. The average price of
water for each quantity is also shown, assuming a monthly availability of service fee of
$7.00.
Decreasing Block Rates. Decreasing block rate charges, such as the three-
part rate schedule in Figure 9-3, are widespread. Initially known as the “Hopkinson”
rate schedule in the electric power industry, the decreasing or declining block rate
schedule has lower block rate charges for higher quantities of water usage or purchase.
Figure 9-3 shows a rate schedule in which $3.50 per unit is charged for the first 3 units
of water per billing period; purchases of more than 3 units and up to 12 units in a
billing period are charged at $2.60 per unit. Finally, purchases of greater than 12 units
are priced at $2.10 per unit. Thus, for a customer purchasing more than 12 units per
month, water in the first and second blocks is priced higher than all remaining water.
The customer must purchase some water at the higher prices to purchase additional
water at lower prices. The average price for each use rate is also shown, assuming a
$7.00 fixed fee for the service hookup.
The decreasing block rate schedule is often considered a marketing tool, i.e., a
tactic to increase water sales. Other rationales for the decreasing block rate schedule
include the ability to specify a rate schedule that will charge lower rates to industrial
customers, who presumably use much larger volumes of water, than to residential or
smaller commercial customers. This objective can often be more effectively achieved
PRICE EFFECTS 151
12.00
10.00
Price per Unit ($) 8.00
6.00
4.00
2.00
0.00
0 2 4 6 8 10 12 14 16 18 20
Quantity
MP AP
NOTES: MP=marginal price; AP=average price.
Figure 9–2 Uniform Commodity Rate and Average Price With a $7.00 Monthly Charge
12.00
10.00
Price per Unit ($)
8.00
6.00
4.00
2.00
0.00
0 2 4 6 8 10 12 14 16 18 20
Quantity
MP AP
NOTES: MP=marginal price; AP=average price.
Figure 9–3 Decreasing Block Rates and Average Price With a $7.00 Monthly Charge
by designing separate industrial and commercial rate schedules or by negotiating

specific rates with very large volume commercial or industrial customers. Decreasing
block rates are appropriate when a utility has excess water available and wishes to
increase sales.
Increasing Block Rates and Lifeline Rates. Many utilities have changed
to increasing block rates to provide incentives for larger volume customers to conserve
water. Sometimes referred to as an inverted rate schedule, an increasing block rate
schedule reverses the pattern of a decreasing block rate schedule, charging higher rates
for greater quantities of water consumption. Figure 9-4 shows an increasing block rate
schedule with the same block rate boundaries as the decreasing block rate schedule of
Figure 9-3. Again, the average price of water at each use level is shown, including a
$7.00 per month service fee.
Table 9-1 presents these three rate schedules as lists of rates and fees. For this
illustration, it does not matter if the units of water are 1,000 gal, 100 ft3, or 4,000 L.
10.00
9.00
8.00
Price per Unit ($)
7.00
6.00
5.00
4.00
3.00
2.00
1.00
0.00
0 5 10 15
Quantity
MP AP
NOTES: MP=marginal price; AP=average price
Figure 9–4 Increasing Block Rates and Average Price With a $7.00 Monthly Charge
Table 9–1 Water Rate Schedules
Volume Range Declining Constant Increasing
0–3 3.50 2.80 1.80

4–12 2.60 2.80 3.10
13 + 2.10 2.80 4.40
A lifeline feature can be incorporated into the rate schedule, with an initial low
volume block as a way to ensure “necessity” quantities of water to all customers at a
low cost. This initial use block sometimes features a zero marginal price, thereby
PRICE EFFECTS 153
effectively giving each customer a necessity quantity of water covered by the monthly
connection fee.
The rate in the final block of an increasing block rate structure could be very
high (punitive) if the need is to curtail water use because of a drought or a long-term
supply limitation. Utilities often impose such punitive rates as part of a crash
conservation program, as discussed in chapter 11. One Utah system charged a penalty
rate of $10.00/1,000 gal for usage exceeding the mandatory limit and experienced a
50 percent overall drop in water use. Some California utilities saw reductions of 80 to
95 percent in their top use block in response to high penalty rates.
To be effective, an increasing block rate schedule should have no more than
three or four blocks, with substantial jumps in the commodity charge between blocks.
Block rate schedules are particularly useful for residential and nonwater intensive
commercial customers. Alternative rate designs, or block rates with much wider
blocks based on best available technology, are appropriate to accommodate the use
levels of commercial activities such as restaurants and car washes and many industrial
activities, including power generation. A carefully tailored rate structure would
encourage these firms to use water efficiently without imposing excessive cost burdens
on them.
The first block in a residential or small commercial rate schedule should include
an amount of water that is thought necessary for normal household operation,
including drinking, bathing, toilet use, cooking, and clothes and dish washing. This
can be estimated from end-use studies, and is likely to be no more than 3,000 gal
(11,356 L) per household per month. The second block would be broader, including
the typical usage of utility customers. The third and final block includes all water use
above the first two blocks, and ideally its commodity charge would be equal to the
marginal long-term social cost of providing water services. The marginal long-term
social cost not only includes all the operating costs incurred by the utility but also
encompasses the expected cost of increasing the supply of water; it would, ideally,
include the lost value to future generations if the utility is depleting its aquifer or
otherwise decreasing future supplies of potable water.
Step Rates. A few utilities have adopted rate schedules with numerous steps
that could be either ascending or descending. In contrast to increasing or decreasing
block rates, with step rates, each customer pays a single price per unit of water. The
price paid, however, depends on the volume purchased. Figure 9-5 shows a typical
increasing step rate, with the rate increasing by 10 cents for each three units
purchased, starting at $1.60 per unit for the first three units. Step rates are designed so
that the customer does not experience a large jump in their utility bill when they
move to a higher step. Therefore, the steps are small, as is the price increase between
steps. Table 9-2 shows a typical step rate schedule, the customer’s water bill for each
quantity, and the (implicit) marginal price paid for each unit of water. Because this is
a step function, the marginal price is higher for the first unit on each step and then
drops back to the price for that step. Even if the steps were all one unit wide, there
would still be a discrepancy between the price at each step and its associated marginal
price, because with every step, the consumer pays the new marginal price for all units
of water, including those that were previously priced at lower steps in the schedule.
With steps one unit wide, however, both the steps and the implicit marginal price at
each step would move steadily higher.
2.50
Price per Unit ($)
2.00
1.50
1.00
0.50
0.00
0 2 4 6 8 10 12 14 16 18 20
Quantity
Figure 9–5 Typical Increasing Step Rate
RATE DESIGN AND REVENUE STABILITY

The Organisation for Economic Co-operation and Development (OECD; 1987)
identified the following objectives that could be furthered by water pricing policy,
only some of which are directly relevant to this work:
• Equity
• Public health
• Environmental efficiency
• Financial stability
• Simplicity
• Public acceptability
• Transparency
Adequate and stable revenue is important because a large part of the cost of
providing water is a fixed cost over considerable time periods. Physical plants,
including wells, dams, pipelines, treatment facilities, distribution pipes, and meters,
must be paid for regardless of the amount of water used. Also, most personnel costs
do not vary with the volume of water sold. Thus utilities often favor a rate structure
with high monthly service charges and low volume rates because it provides a highly
predictable revenue stream. In contrast, with an increasing block or step schedule,
PRICE EFFECTS 155
Table 9–2 Typical Step Rate Schedule, Customer Bill, and Marginal Cost
Rate Bill Marginal Cost
Quantity ($/unit) ($/month) ($/unit)
1 1.60 1.60 1.60

2 1.60 3.20 1.60
3 1.60 4.80 1.60
4 1.70 6.80 2.00
5 1.70 8.50 1.70
6 1.70 10.20 1.70
7 1.80 12.60 2.40
8 1.80 14.40 1.80
9 1.80 16.20 1.80
10 1.90 19.00 2.80
11 1.90 20.90 1.90
12 1.90 22.80 1.90
13 2.00 26.00 3.20
14 2.00 28.00 2.00
15 2.00 30.00 2.00
16 2.10 33.60 3.60
17 2.10 35.70 2.10
18 2.10 37.80 2.10
19 2.20 41.80 4.00
revenue changes much more rapidly with sales volume. Exceptionally wet summers
can reduce water sales for irrigation purposes, thereby cutting into revenue, and a hot,
dry summer may spur sales, creating substantial surpluses if sufficient supplies of
water are available. Effective nonprice conservation programs, as during a drought
when supplies are limited, reduce revenue.
Table 9-3 illustrates the effect of a drop in sales under each of the rate schedules
presented in Table 9-1 and illustrated in Figures 9-2, 9-3, and 9-4. Water sales could
have decreased because of reduced irrigation during a wet, cool summer or because of
an effective nonprice conservation program. The quantities and revenues are
presented for an average customer of a mostly residential utility. This illustration
assumes that under initial conditions, consumers had previously adjusted to their
marginal price for water, so that initial usage and revenue are highest under a
declining block rate where the marginal price is $2.10 and lowest under increasing
block rates when marginal price is $4.40. The illustration also assumes that with these
values, the utility is covering its operating costs under each rate schedule.
The drop in revenue resulting from a 10 percent volume reduction is
significantly different under each rate schedule. Under declining block rates, the
10 percent reduction in sales cuts revenue by 7.07 percent. With constant commodity
charges, revenues fall by 8.65 percent and under this increasing block rate schedule, by
Table 9–3 Effect on Revenue of a 10 Percent Drop in Sales by Rate Structure*

Type of Rate Structure
Declining Constant Increasing
Initial Conditions
Quantity 18.00 16.00 14.00
Marginal price ($) 2.10 2.80 4.40
Revenue ($) 53.50 51.80 49.10
Reduced Use
Quantity 16.2 14.4 12.6
Revenue ($) 49.72 47.32 42.94
Decrease ($) 3.78 4.48 6.16
Percentage 7.07 8.65 12.55
*Using the rate schedules from Table 9-1, along with a $7.00 per month service fee
12.55 percent. Thus, in this example, the fall in revenue is 45 percent larger under
increasing block rates than under decreasing block rates.
A good way to deal with revenue uncertainty is to establish a sinking fund so
that surplus revenue in times of high water sales offsets reduced revenue when sales
fall because of a conservation program or the weather. To make a sinking fund
politically acceptable, utilities must include some mechanism to refund any excess
accumulations. The refund mechanism then becomes a critical aspect of rate design.
Assuming that utility plans include conservation, the refund should not effectively
lower consumers’ marginal price of water or induce them to increase water use. We
suggest a once-a-year refund to all customers when excess funds accumulate, sent by
check independent of the water bills. The object is to refund excess revenue in such a
way that most consumers view the refund as a windfall and not as a reduction in their
water bills. If conservation is not an objective for the utility, however, surplus sinking
fund revenue could be used to directly reduce bills once a year, understanding that
this might increase water sales.
Marginal Cost and Rebate. Griffin (2006) recently proposed an innovative

rate design that incorporates both marginal cost pricing and revenue neutrality. This
is an ideal rate schedule for utilities choosing to set a single marginal price equal to the
long-term marginal cost of providing water service or advancing water use efficiency
in scarce water environments. In Griffin’s view, this rate schedule is a good alternative
to typical pricing policies that rely on setting p below efficient levels to balance utility
budgets, thereby fostering excessive water use. Each consumer’s bill is computed as
follows:
Bill n = M + p × ( w n – W ) (Eq 9-1)

PRICE EFFECTS 157
where:
Billn = nth customer’s water bill
M = periodic meter charge
p = marginal price per unit of water
wn = nth customer’s water use per period
W = bill threshold.
The water bill for a consumer using exactly W units of water is M, the periodic
meter charge. If more water than W is used, each unit above W is charged the
marginal price p. The innovation in this structure applies to customers using less than
W: their bills would be reduced by p for each unit by which their use falls short of W.
If these credits for not using water exceed M, the customers would receive a refund.
The purpose of M is the same as the fixed fee in any other rate schedule: it covers
metering and other costs of having each connection. The utility would compute its
marginal cost of providing water, including all the usual cost elements plus the value
of natural (unprocessed) water when applicable, and set p equal to marginal cost.
Such a system has favorable implications for low- or fixed-income user groups while
maintaining efficient pricing signals.
The parameter W is used to balance the utility budget. For a utility facing
increasing marginal costs of providing water, which is becoming more common as
readily available supplies of raw water become scarce, the price per unit, p, is likely to
exceed average cost; W would allow the utility to avoid surpluses. If marginal cost
were close to average cost for the utility, W would be small.
In this rate structure, all customers face the same marginal price for water.
Large-volume users pay the marginal price for each unit above W that they use, and
low-volume customers receive a credit for each unit below W that they do not use. A
utility facing a revenue shortfall would lower W, and one facing a surplus would raise
W. Utilities could include summer surcharges in this rate structure if the surcharges
reflect the increased cost to the utility of providing extra capacity to meet peak
demands. The schedule could also be differentiated by customer class, reflecting
differences in the cost of service.
Table 9-4 is an example of this rate schedule; it shows the effect of a long-term
downward trend in water use that reduces average usage by 10 percent. The marginal
price is set at $2.80 per unit, the monthly service charge is set at $7.00, and W is set at
3 units of water. For an average usage of 16 units, the average bill is thus $43.40. A
customer using only 2 units of water, which is less than W, receives a credit for 1 unit
of water not consumed, which is subtracted from the monthly service charge, leaving
a water bill of $4.20. When use falls by 10 percent, the utility could reduce W to
1.4 units of water to maintain the same revenue. For the low-volume user, the drop in
W is only slightly offset by a 10 percent drop in volume, which increases the bill to
$8.12, an increase of $3.92 ($2.80 × 1.6 – $2.80 × 0.2). Because the high-volume water
user uses far less water, his water bill falls. In contrast, if some customers in each
category do not change their water use as system-wide average use falls by 10 percent,
they receive the new bills shown at the bottom of the table.
Table 9–4 Effect of a Drop in Use, Marginal Cost, and Rebate Schedule
Consumption Rates
Low Average High
Marginal price 2.80 2.80 2.80
Quantity 2.0 16.0 24.0
Bill ($) 4.20 43.40 65.80
Monthly charge ($) 7.00 7.00 7.00
W 3.0 3.0 3.0
With 10% Reduction in Use
New quantity 1.8 14.4 21.6
New bill ($) 8.12 43.40 63.56
Monthly charge 7.00 7.00 7.00
New W 1.4 1.4 1.4
For individuals with Unchanged Use
Old quantity 2.0 16.0 24.0
New bill ($) 8.68 47.88 70.28
WHAT IS THE “PRICE” OF WATER?

When determining the price of water to be used in statistical analysis, analysts can
divide the rate schedules and fee arrangements previously listed into two categories:
unmetered service and metered service.
Unmetered Service. A constant bill, regardless of how much water a
customer uses, is the flat rate charge used in areas without water meters. Flat rate
charges give no incentive to conserve or economize on water use, and, as noted later
in this chapter, typically result in higher levels of water usage than when usage is
metered. With flat rates, the price of additional quantities of water to an individual
customer is always zero, although there is a cost of obtaining service and an obvious
cost to the utility supplying water. If a utility charges groups of customers based on a
share of total group usage, some incentive exists to use less water to reduce the
common water bill. The strength of this incentive effect depends on the size and
cohesiveness of the group. In a large, heterogeneous group such as a large apartment
complex or a utility service area, individuals will pay little or no attention to
conserving to lower the water bill, because one person’s consumption has little effect
on the total cost of providing water. In a small, cohesive group, however, everyone
may have a collective desire to reduce water use and thus bills through goodwill or by
peer pressure.
Metered Service with Commodity Charges. If the water bill contains
components that vary with the amount of water purchased—a commodity charge—
consumers are likely to react to increases in their water bill by reducing their water
purchases. Charges for sewerage usage, when based on actual measured flows or on
PRICE EFFECTS 159
measured water use, should also be considered in any analysis of the effect of price on
water use because the total cost of using water drives usage decisions. Strong evidence,
cited in chapter 10, shows that increases in the effective price of disposing of effluent
often trigger industrial firms to take water-conserving actions.
Analysts can forecast future water use or even explain the factors driving
historical use with a regression or other statistical analysis to estimate consumer price
effects. This requires translating the schedule of rates and charges applying to water
sales to a number or numbers representing the price of water and raises the issue of
specifying the price or rates. How to best represent water rates statistically figures
importantly in many published studies of urban water price impacts. The analyst
must determine which rate or combination of rates triggers—and is quantitatively
linked with—the consumer’s purchasing decisions.
The literature is largely concerned with the following price specifications: (1)
average price, (2) marginal price, (3) marginal price plus “difference”, and (4) total
bill.
Average price is the consumer’s total periodic water bill divided by the quantity
of water used. Thus if a consumer used 10 units of water and paid $25 for water
service, the average price is $2.50 per unit. Average price can be easily computed
regardless of the form of the rate schedule. When water use determines the cost of
sewerage usage, the analyst should include the combined water and sewer bill when
computing average price. This is generally advisable even if they are billed separately,
although this assertion can readily be checked empirically by developing regression
models with and without sewer use charges. This suggests that an increase in sewerage
usage fees may reduce water utility sales and revenues somewhat. Typically, the change
from year to year in sewerage usage fees is small enough that this effect is not noticed
because of weather-related changes and other influences on sales. The effect could be
significant for individual consumers without having a large effect on utility-wide
averages.
Marginal price is the price per unit of water for the last unit used. Figures 9-2,
9-3, and 9-4 show both marginal and average price and their relationship for various
quantities of water purchased. If a consumer facing the increasing block rate schedule
of Figure 9-4 used 14 units of water, his marginal price would be $4.40. His total water
bill would be $7.00 + 3 × 1.80 + 9 × 3.10 + 2 × 4.40 = $49.10. His average price would
be $49.10/14 = $3.51. In this example, average price is lower than marginal price
because of spreading the lower prices for the first 12 units of water over the 14 units
of water used. Difference is defined as the consumer’s total water bill minus marginal
price times total quantity used. In this example, difference = $49.10 – 14 × $4.40 =
–$12.50. A positive value shows that the consumer would have paid less if there were
no service charge and all water were billed at marginal price. Average price will
become close to the marginal price as the quantity used becomes larger. If the
monthly service charge was raised to $19.50 (the original service charge of $7.00 plus
the difference between the highest marginal price and the marginal prices in the first
two use blocks times their respective quantities), marginal and average price for
customers using 12 or more units of water would be identical. Consumers billed
under increasing block rates who use relatively small quantities of water often face an
average price that exceeds their marginal price. Chapter 13 examines these issues of
specification in greater detail.
Fortunately for the utility analyst preparing a forecast of water use, the results
of prior studies, including the estimated price elasticities of demand, can generally be
used regardless of which price specification was used to derive them. The price
specification becomes critically important only when rate structures are being
changed. For proportional changes in all elements of a rate structure (e.g., monthly
service charges and marginal prices), all the price specifications predict similar effects
on water usage and revenue. For such rate increases, therefore, the analyst can rely on
price elasticity estimates obtained from diverse price specifications.
When customers pay for sewer services based on water volume, these charges
should be included in any analysis of price effects. A totally rational consumer might
cut back on all water use during the two or three months used in computing their
sewer bills. Previous research, however, has shown that most customers do not react
this way, but consider sewer usage charges to be part of each month’s cost of water,
which is the way they are often billed.
If consumers respond primarily to their total water bill rather than to marginal
price, a change in rate structures that does not significantly change most customers’
water bills will have little or no effect on water use decisions. This criterion is unlikely
to be realized for any real-world shift in rate structures, such as from decreasing block
rates to increasing block rates. A change in the form of the rate structure is likely to
have a significant effect on both aggregate demand and utility revenue, so it cannot be
ignored for forecasting purposes. Imposing a very high, or punitive, marginal price
for usage—above some selected quantity—will have a dramatic impact on demand.
When utilities modify rate structures, they should disaggregate water customers by
volume use categories to intensively analyze those who will be receive significantly
higher or lower water bills. Some evidence has shown that using average price in a
regression model produces stronger results when water bills are a low percentage of
income, and marginal price works better when bills exceed 2 percent of income
(Billings and Day 1989). This issue is further discussed in Chapter 10.
PRICE ELASTICITIES OF DEMAND AND REVENUE

The concept of price elasticity is often used by water analysts. The following sections
introduce this concept and its application in water economics.
Definitions, Formulas, and Computations. The price elasticity of demand
measures the responsiveness of the quantity demanded to a change in price. It is
computed as the percentage change in quantity of water sales or usage divided by the
percentage change in price/rates, when other factors, such as income and weather, are
constant or controlled in the estimating equation. It is often defined as the percentage
change in water purchases associated with a 1 percent change in water price. Price
elasticity is a pure number, independent of the dimensions in which water sales and
PRICE EFFECTS 161
prices are measured, because it is the ratio of two percentages. The price elasticity of
demand is always a negative number because of the Law of Demand, which states that
at higher prices less will be purchased.
Elasticity is a convenient, shorthand way to describe the most important
information contained in a complete demand curve such as that shown in Figure 9-1.
The equation used to obtain Figure 9-1 contains more information, but it is more
complex and harder to describe than a single elasticity value. The equation also
requires a careful statement of the units of measure employed, such as gallons and US
dollars or liters and Euros. In contrast, elasticity is a pure number—not needing units
of measure. Its meaning is universally understood among economists and water
analysts, and it is easily compared among systems, across locations, or over time.
As long as the demand curve for water has a negative slope (i.e., people buy less
water when the price rises), the price elasticity of demand will be numerically negative
(less than zero). Although some individuals may ignore a price increase and continue
with their usual purchases, it is highly unlikely that anyone would buy more water in
response to a price increase.
The price elasticity of revenue shows the responsiveness of total revenue to a
change in price. It is defined as the percentage change in total revenue divided by the
percentage change in price. The price elasticity of total revenue is always equal to 1
plus the price elasticity of demand; thus, the price elasticity of revenue can be either
positive or negative.
Relatively inelastic. Measured price elasticities of demand between 0 and –1 are
referred to as relatively inelastic. That is, quantity is relatively unresponsive to a change
in price because the percentage change in quantity is less than the percentage change
in price. The corresponding price elasticity of revenue is between 1 and 0.
For example, if a water utility is considering raising water rates by 15 percent,
and an estimate of the system-wide price elasticity of demand is –0.3, the effect on
both water sales and system revenue can be directly computed. The projected price
increase will reduce water sales by 4.5 percent (the elasticity of –0.3 times the price
increase of 15 percent).
The price elasticity of total revenue in this example is 0.7 = 1 – 0.3. Revenue
increases by 10.5 percent (0.7 × 15 percent) when price is increased by 15 percent,
assuming nothing else changes. Thus, although water rates increased by 15 percent,
revenues increased by only 10.5 percent because customers responded to the price
increase by reducing water use by 4.5 percent.
Perfectly inelastic. A price elasticity of demand equal to 0 is known as perfectly
inelastic and indicates that water sales do not change in response to a change in price.
A perfectly inelastic demand results in a revenue elasticity of 1.0; total revenue
increases by the same proportion as price. Although some individuals may have a
perfectly inelastic demand for water, it is highly unlikely that this would be true for an
entire community unless water bills were a very small proportion of every customer’s
income. For some utilities, where water bills are substantially less than 1 percent of
most customers’ incomes, the elasticity may be low enough that any drop in sales
caused by a modest price increase is obscured by other influences on demand, such as
variations in sprinkling demand because of the weather. The effect of a modest price
increase will also be offset by continuing inflation in the general economy. For
example, in an economy with 4 percent inflation per year, a 5 percent price increase
“disappears” in about 15 months. Therefore, not surprisingly, many utility officials
have discounted the possible impact of price changes on demand, and, de facto,
believe that the demand for water is perfectly inelastic.
Relatively elastic. A price elasticity of demand less than –1.0 indicates that the
percentage change in quantity is greater than the percentage change in price, a
situation known as relatively elastic. A price increase when demand is relatively elastic
results in less revenue. For example, suppose the price elasticity of demand is –1.4.
The corresponding revenue elasticity is –0.4 (1 – 1.4). Thus, a 15 percent increase in
water price would cut sales by 21 percent and revenue by 6 percent. Fortunately for
the water utility industry, the price elasticity of demand for water is almost always
relatively inelastic, so the price elasticity of revenue is positive and raising prices
increases revenue.
Unitary elastic. A price elasticity of demand of –1.0 is called unitary elastic. This
means that the percentage changes in price and quantity are equal. The corresponding
revenue elasticity is exactly 0. In this unusual situation, a price increase translates into
a quantity decrease of exactly the same percentage, leaving total revenue unchanged.
Because the fall in volume delivered would cause some drop in utility cost, a price
increase would mean a small net revenue benefit to the utility.
Computation. Price elasticities are usually local estimates, applicable only within
a range around the price where they are initially estimated. For the linear demand
curve, the elasticities at a specific point on the curve are:
ε = b × P ∗ ⁄ Q∗ (Eq 9-2)
γ = ε+1
where:
ε = price elasticity of demand
γ = price elasticity of revenue
b = the inverse of the slope of the demand curve
P∗ = a specific price on the demand curve
Q∗ = the specific quantity corresponding to P∗
Taking the values illustrated in Figure 9-1, and using P∗ = $1.40, Q∗ = 13, and
the inverse of the slope of the linear demand curve, b = –3.7, the price elasticity of
demand is –3.7 × 1.40/13 = –0.40, which is relatively inelastic. The corresponding
revenue elasticity is (1.0 – 0.40) = 0.60, meaning that a 10 percent price increase
would decrease sales by 4 percent and increase revenue by 6 percent.
The elasticity along the linear demand curve in Figure 9-1 at a higher price such
as $3.02 and corresponding volume of seven units of water would be –3.7 × 3.02/7 =
–1.60. In this situation, a 10 percent price increase would cut sales by 16 percent and
PRICE EFFECTS 163
decrease revenue by 6 percent. Generally, elasticities are assumed to be symmetrical,

so that for these values, a price decrease of 10 percent would increase sales by
16 percent and increase revenue by 6 percent. An elasticity of this size is improbable
for total system demand, although some large-volume commercial or industrial
customers might respond to a price increase by investing in water saving capital
improvements, dramatically reducing their water use. In this case, a later price cut
would not return water sales to their previous level.
The exponential demand curve in Figure 9-1 exhibits a constant price elasticity
of demand of –0.4 regardless of the price-quantity combination chosen. For price
increases greater than 10 percent, using values from a curve such as this will create
more accurate forecasts than a simple computation using the elasticity value. Analysts
can also obtain a reasonably accurate result using elasticity by computing successive
10 percent price changes, as explained later in this chapter.
It is important to note that the estimates using elasticities are sensitive to
inflation. Consumers typically take inflation into account when they decide what and
how much to buy. After one year of inflation at 4 percent, a 15 percent water rate
increase deteriorates to 11 percent. After only four years of 4 percent price inflation,
the effect of the initial 15 percent rate increase erodes to a price reduction of 2 percent
in real terms (with compounding). This dissipation of rate effects is quicker for more
rapid inflation. When rate increases are smaller than the general rate of inflation,
water prices drop in real terms. All other things being equal (which they never are),
we could expect a utility with constant nominal rates (no change in the stated rates)
to experience slowly increasing demands for water as inflation reduces the real price
of water. One offsetting factor in the United States and other countries is national
plumbing codes that mandate low-flow indoor fixtures.
How Large Is the Price Elasticity of Demand? A review of numerous sta-

tistical studies of urban water demand suggests the following qualitative “facts” about
the price elasticity of demand for water:
• Demand is relatively inelastic. When prices rise, the percentage increase in
price is almost always larger than the resulting percentage drop in sales.
Thus utility revenue almost always increases when prices are raised.
• Demand is more elastic (or less inelastic) for outdoor (landscape irrigation)
than for indoor use.
• The response to price changes increases with time, as long as inflation does
not erode these changes. A temporary “shock effect” of a price increase may
result when water users overestimate the effect of the increase on their water
bills. For example, they may stop watering their lawns. Typically, the
adjustment to much higher (or lower) water prices will occur over an
extended period of time as users alter fixtures and landscaping.
• Price elasticity increases with the ratio of water bills to income. This can be
expressed in the following three corollaries:
1. Demand is more elastic (less inelastic) at higher prices, because at

higher prices water bills take a larger proportion of income.
2. Demand is more elastic for higher volume uses. Within any well-
defined category of customers, those who use larger volumes of water,
other factors being equal, will find water bills more burdensome than
their counterparts. It is also likely that those using more water have
adopted marginal uses for water that are avoided by those with
smaller usage levels. A price increase tends to encourage these users to
reduce or abandon some of these “extra” uses of water, such as hosing
down sidewalks or forgetting to turn off faucets. A rate increase may
also encourage larger-volume users to invest in water-conserving
technology such as low-flush toilets.
Commercial and industrial demand also tends to be more elastic for
those who use larger than average volumes. The demand for water by
an electric generating plant using a larger volume of water than the
industry average will have a higher (long-run) elasticity. Their short-
run elasticity may be near 0 if reducing water use requires new capital
investment, but given time for adjustment, the larger outlays for
water will create an incentive for water-saving changes in production
methods. Some users have discovered that investing in water-saving
technology actually increases profits, because the reduced cost of
water exceeds the amortized cost of the required investments.
3. Demand becomes less elastic as incomes rise. With higher incomes,
water bills take a smaller proportion of the household budget,
making customers less sensitive to their prices.
Numeric Rules of Thumb. Given the numerous available elasticity estimates
and their seemingly conflicting results, the forecaster may be at a loss about what
elasticity to use in a specific application. Griffin (2006) addressed this issue by
developing rules of thumb, which we have expanded on here. Table 9-5 presents these
shortcuts, which are a distillation of the numerous elasticity studies mentioned later.
Forecasters can use these values directly for forecasting the response of residential
water use to price changes as well as total system response for all but large-volume
industrial or commercial customers.
The starting point for these values is the average elasticity estimate from the
many available studies and research showing that elasticity is higher for increasing
block rates than for other rate forms. Additional research suggests that prominently
displaying the marginal price on water bills increases the response to changing
marginal price. Although we doubt that any system with declining block rates
prominently displays this information on water bills, we surmise that if they did, it
would reduce the absolute size of elasticity. Numerous studies have shown that winter
elasticity and, by extension, elasticity in wet, cold climates, is closer to 0 and elasticity
in the arid West is larger in absolute size. Areas where water bills are more
PRICE EFFECTS 165
Table 9–5 Rule-of-Thumb Elasticity Values

Marginal Price in Rate Structure
Declining Constant Increasing
Base case elasticity (annual) –.3 –.4 –.5

Additions or Subtractions
Marginal price on bills –.1 –.15
Wet/cold climate +.1 +.1 +.1
Arid West –.1 –.1 –.1
Winter (low irrigation season) +.15 +.15 +.15
Summer (high irrigation season) –.15 –.15 –.15
Bills >1½% of average income –.1 –.1 –.1
Bills <½ of 1% of average income +.2 +.2 +.2
Effective long-run conservation +.1 +.1 +.1
Source: Derived from Griffin 2006 and authors’ estimate.
burdensome for consumers are likely to be more responsive to price changes, and,
conversely, those where they are more trivial are likely to be less responsive. Renwick
and Archibald (1998) found that in Santa Barbara, Calif., elasticity varied significantly
by income group. For those with incomes under $20,000, price elasticity was
estimated at –0.53. For incomes between $20,000 and $100,000, the elasticity was
–0.21, and for those greater than $100,000, elasticity was only –0.11. Effective long-
run conservation programs result in elasticity being closer to 0, because some of the
water-saving opportunities are already in use. According to several studies, the
response to price tends to increase over time, so the long-term elasticity is greater than
the short-term elasticity. These rules of thumb are an effective way to include
reasonable elasticity values in a water-demand forecast when an individual utility does
not have detailed research results. They also serve as a cross-check on locally
generated elasticity estimates.
For example, suppose we need a winter price elasticity for a utility with a
declining block rate schedule and an effective, long-term conservation program. From
Table 9-5, the base case elasticity is –0.3, the adjustment for winter is +0.15, and the
adjustment for the conservation program is +0.1, giving an effective elasticity of –0.3
+ 0.15 + 0.1 = –0.05, which is the value Maddaus and Maddaus (2006) recommend.
Consider another utility with increasing block rates and predominantly low-income
customers needing a summer elasticity. Their base case elasticity is –0.5, the summer
adjustment is –0.15, and the low-income adjustment is –0.1, giving –0.75 as an
operational elasticity estimate.
USING PRICE ELASTICITIES IN DEMAND FORECASTS

Analysts can use price elasticities from previous studies or the rules of thumb
described previously to estimate the effect of a rate change on volume sales and
revenues. They also can be used to determine the rate change needed to achieve a
given revenue requirement. Three simple formulas summarize these computations,
illustrated by the following examples:
ΔP
ΔQ = ------- × ε × Q (Eq 9-3)
P
ΔP
ΔR = ------- × γ × R (Eq 9-4)
P
ΔP ΔR 1
------ = ------- × --- (Eq 9-5)
P R γ
where:
ΔQ = change in quantity
ΔP = change in price or rates
ε = price elasticity of demand
Q = original quantity
ΔR = change in revenue
γ = price elasticity of revenue = ( ε + 1)
R = original revenue
P = original price
Example Computation. A utility in Kentucky currently delivers 40 million

gallons per day (mgd) at an average price of $2.00/1,000 gal, yielding revenues of
$80,000 per day. This water agency wishes to increase revenues by 8 percent next year
to catch up with past inflation in the costs of supplies and labor.
Translating the revenue objective into a rate increase. Most rate and revenue
computations begin with an unconstrained demand forecast. This forecast is
unconstrained in the sense that it ignores consumers’ behavioral responses to higher
rates. In effect, it assumes that the price elasticity of water demand is 0. In this case,
the unconstrained forecast suggests an 8 percent increase in utility rates and charges is
needed to supply an 8 percent increase in utility revenues next year.
This should be understood to mean an 8 percent real (inflation-adjusted)
increase in water rates and charges. If the objective is to close the gap between labor
and supply costs, and revenues, then next year’s inflation must be taken into account
in determining the dollar amount of the rate increase. In nominal dollar terms, the
rate increase supplying 8 percent more in purchasing power to the utility next year
equals 8 percent plus next year’s anticipated rate of inflation. Utilities generally take
this into account by projecting next year’s budget expenditures, which include
expected wage rate and materials price increases.
PRICE EFFECTS 167
If 3 percent inflation is anticipated next year, rates and charges must rise by
11 percent to achieve an 8 percent real increase in revenue next year and a balanced
budget. When both 3 percent inflation and the most likely consumer responses to the
price increase are taken into account, the required increase in water rates is close to
13 percent. The required computations will be presented after the discussion of
selecting a price elasticity value.
Selecting the price elasticities. The analyst could estimate an elasticity for the
specific utility from a structural demand model (see chapter 12). A value could also be
obtained from one of the many available research reports such as those cited by
Dalhuisen (2003). The problem with these research results is their wide dispersion.
Although most estimates are between 0 and –1.0, this is a wide range. Utility
managers want to avoid both excessively large rate increases leading to unwanted
surpluses and insufficient increases leading to budget deficits. The ideal strategy is to
implement rates that adjust upward semiautomatically with general inflation in the
cost of running the utility. Relatively small annual increases avoid rate shock, keep the
utility budget balanced, and offer a continuing stream of evidence on the system price
elasticity of demand.
The rules of thumb in Table 9-5 give reasonable and easily accessed elasticity
values for use in demand forecasts when more specific local information is not
available; they also serve as a check on local estimates. In the absence of reliable, local
elasticity estimates, a utility would benefit from using the elasticity values from this
table rather than assuming the elasticity is 0. For this example, we use the elasticity
value of –0.2 suggested by Maddaus and Maddaus (2006). With some additional
information, the same value could be obtained from Table 9-5. Assume the utility uses
a constant marginal price, so its base elasticity is –0.4. Add 0.1 for a wet climate, and
0.1 for an effective, long-run conservation program to obtain a value of –0.2.
Estimating quantity impacts and the revenue target. With a price elasticity of
demand of –0.2, the 8 percent real price increase will decrease water demand by
0.64 mgd (= 40 mgd × 0.08 × –0.2). That is, the current usage is multiplied by the
proportion of the price increase times the elasticity (Equation 9-3). The new sales will
be 39.36 mgd, 1.6 percent less than the unconstrained forecast.
The corresponding increase in revenue is the price elasticity of revenue
multiplied by the proportion of the price increase times the current revenue
(Equation 9-4). The price elasticity of revenue is 0.8 (equal to 1 plus the price
elasticity of demand [1 – 0.2]). Thus, the increase in revenue is $5,120
(= 0.8 × 0.08 × $80,000). The new total revenue is $85,120 per day.4 This can be
compared to the unconstrained forecast of $86,400/day and revenues before the price
increase of $80,000 per day ($0.002/gal × 40 mgd). Instead of increasing by the
expected 8 percent or $6,400 per day (in real terms), revenues increased only 6.4
percent or $5,120, an error of 20 percent.
4. Because elasticities are exact for only very small changes in price, the new total revenue computed
using the revenue elasticity will differ somewhat from a computation using the new price and the new
quantity. As the size of the price change approaches 0, this difference will disappear. As a practical mat-
ter, this difference will be much smaller than the random error in the forecast of revenues and can be
ignored.
Given the price elasticity of revenue, the price increase required to obtain any
specified revenue increase (in real terms) can be directly computed. The required
percentage increase in price is the desired percentage increase in revenue divided by
the revenue elasticity coefficient (Equation 9-5). Therefore, to obtain an 8 percent
revenue increase in this example, divide 8 percent by 0.8, yielding a 10 percent increase
in real price. Assuming next year’s rate of inflation is 3 percent, a 13 percent nominal
increase in rates and charges can be projected to yield an 8 percent real increase in next
year’s revenues in this system.
Cautions in Applying Elasticities. Two qualifications should be noted

when using price elasticities. Price increase in this context means a proportional
increase in all rates and charges. This interpretation is dictated by problems
surrounding the relevant price specification and estimation of statistical price or rate
impacts. Even when price increases are not strictly proportional, however, elasticity
computations can still provide useful estimates.
Second, elasticities are strictly valid only for small changes in price. The larger
the proportional price change, the wider the confidence interval must be around the
estimated effects. Put another way, as the proportional price changes become larger,
our confidence in the results necessarily decreases. For an extreme example, using a
price elasticity of demand of –0.5 and a 200 percent price increase (e.g., price goes
from $0.30/unit to $0.90/unit), the simple computation outlined previously would
predict that water use would drop by 100 percent—to 0.
Part of the solution to this problem is to do the computations incrementally.
For that same 200 percent price increase, perform the elasticity-based computations
with a series of 10 percent increases. The first computation increases prices by
10 percent, with a resulting 5 percent drop in quantity. With compounding, the
second 10 percent increase results in a total increase of 21 percent in price and a total
drop in quantity of 9.75 percent. After 12 iterations, the total increase in price
amounts to 214 percent and quantity falls by 46 percent, as shown in Table 9-6, which
gives the detailed computations required. This procedure is approximately equivalent
to assuming a double logarithmic functional form for the demand curve.
When a utility changes the rate schedule in a nonproportional manner, or shifts
from one rate form such as constant commodity charges to increasing block rates,
estimating effects becomes much more complex. Chapter 10 describes an example of
such a shift in the context of using rate changes to encourage water conservation.
Essentially, these computations involve disaggregating customers by usage categories
and applying elasticities individually to each category. The most appropriate elasticity
for these forecasts may differ among the various groups. The utility forecaster would
need to make an informed judgment and then evaluate customer response with future
statistical research to determine how groups of individuals in different circumstances
responded. When the utility institutes large changes in price, or changes the structure
of rates, the revenue and demand forecasts should be frequently compared with
realized results to quickly alert management if the response to these price changes is
significantly different than expected.
PRICE EFFECTS 169
Table 9–6 Detailed Computations for a 200 Percent Price Increase
Iteration Price % Rise Starting Price ($) New Price ($) Starting Quantity New Quantity
1 10 1.20 1.32 30 28.5

2 10 1.32 1.452 28.5 27.075
3 10 1.452 1.597 27.075 25.721
4 10 1.597 1.757 25.721 24.435
5 10 1.757 1.933 24.435 23.213
6 10 1.933 2.126 23.213 22.053
7 10 2.126 2.338 22.053 20.950
8 10 2.338 2.572 20.950 19.903
9 10 2.572 2.830 19.903 18.907
10 10 2.830 3.112 18.907 17.962
11 10 3.112 3.424 17.962 17.064
12 5.14 3.424 3.60 17.064 16.625
ESTIMATING THE IMPACT OF METERING AND RATE CHANGES

Statistical studies and utility officials’ practical experience reveal that water use
decreases when utilities install water meters and impose commodity charges on water
customers. Meter installation typically reduces water use from 15 to 40 percent.
The Johns Hopkins Residential Water Use Project, funded by the Federal Home
Administration, developed a rigorous comparison of metered and unmetered water
use. Thirty-nine study areas in the United States, each including a relatively
homogenous residential district, were master-metered from 1961 to 1965; flows of
water were recorded every 15 minutes (Linaweaver et al. 1966). The study included
• Ten metered areas with public sewers in Oakland, Los Angeles, and San
Diego, Calif.;
• Five metered areas with septic tanks in Des Moines, Iowa; Great Falls, Mont.;
and Denver, Colo.; and
• Five apartment building areas without individual residential meters in San
Diego, Denver, and Washington, D.C.
Linaweaver and colleagues (1966) report dramatically lower water use per
household in areas with meters and commodity charges (Table 9-7). When water
meters were installed in Boulder, Colo., average residential water use per month fell
from 9,200 to 5,900 gal (Hanke 1970). Metering resulted in a dramatic drop in lawn
irrigation, from an estimated 165 percent of the ideal amount to only 84 percent, with
virtually all of the decrease occurring in the first year. These reductions in use were
long term, as verified by follow-up studies in the same areas of Boulder.
A large study of submetering in multifamily housing found that water use in

individually metered units averaged 3,686 gal per month compared to centrally
metered complexes with use of 4,349 per month, a difference of 15 percent (Mayer
et al. 2004). This research also found that allocating centrally metered water bills to
individual units had no effect on water use. Additional evidence on the effect of
metering is presented in chapter 10.
Controlled Comparisons. Although it may not be easy to exactly identify the

relevant price of water, customers do react to variations in water prices when the
amount of water purchased in a billing period is used to help determine the total
water bill.
Table 9-8 presents a controlled comparison of average water use per metered
single-family residence in two adjacent Arizona communities. These communities
have similar household income levels, household sizes, and other characteristics.
Table 9–7 Water Use in Metered and Flat Rate Areas, Western United States
Gal/Day/Dwelling
Unit Annual Average Metered Flat Rate
Leakage and waste 25 36

Interior use 247 236
Sprinkling 186 420
Total household use 458 692
Peak hour 2,481 5,170
Modified from Linaweaver et al. 1966.
Table 9–8 Controlled Comparison of Water Usage Rates

Tucson Water Flowing Wells
Average Monthly Water Sales per Average Monthly Water Sales per
Year Water Bill ($) Customer (gal) Water Bill ($) Customer (gal)
1975 6.50 12,300 3.90 16,500

1976 9.32 10,700 6.31 15,200
1977 11.04 9,500 6.37 13,700
1978 11.71 9,400 6.79 12,800
1979 11.61 9,200 7.06 12,600
1980 11.20 9,600 6.66 13,000
Source: Billings and Day (1982).

NOTE: The water bills used to measure the price of water were computed as the average
of the bills for 8,976 gal/month using winter rates and 14,960 gal/month using summer rates.
Bills are stated in 1974 dollars.
PRICE EFFECTS 171
Rainfall and temperature were the same in both areas, and account for most shared
year-to-year variation in water use. Over the six years recorded, the average price of
water in one district was 65 percent higher than in the adjoining area, and average
usage was 28 percent lower. This implies a long-run price elasticity of demand of –0.43.
ECONOMETRIC STUDIES
The third and most comprehensive type of evidence for the existence and size of
customer responses to water rates comes from an extensive collection of econometric
studies. These typically estimate a multivariate regression such as the linear
relationship (see sidebar).
Q = b0 + b1 P + b2 X2 + … + bk Xk (Eq 9-6)
where:
Q = water use
b0 = constant or intercept of estimated regression
b1, b2, … bk = coefficients estimated by the regression, which show the effect of
each independent variable on Q
P = the real or inflation-adjusted price of water
Xi, i = 2, 3,... k = other explanatory variables such as personal income, household
size, rainfall, and temperature.
The studies usually give estimated values of the price elasticity of demand in
addition to estimated regression coefficients and diagnostic statistics.
THE DOUBLE LOGARITHMIC FUNCTIONAL FORM

A popular alternative functional form is multiplicative, which is often referred to as double
logarithmic because it is typically estimated by taking natural logarithms of both sides of the
b0 b1 b2 bk
equation. Thus, Q = e P X 2 … X K becomes lnQ = b0 + b1lnP + b2lnX2 + …. bK lnXK,
where ln is the natural logarithm and e is the base of the natural logarithm. Multiplicative forms
often fit water data better than the linear form in Equation 9-3. There is the added bonus that
the estimated coefficients—the b1, b2, and so on—are directly interpretable as elasticities that
remain constant over the entire estimated demand curve.
A third functional form for the estimating equation was used to obtain the results shown in
Table 9-7. After taking logs of both sides, this semilogarithmic equation, which is often referred
to as inverse semilogarithmic, is lnQ = b0 + b1P + b2X2 + …. bKXK
The elasticity estimate from this model depends on price, and is ε = b1P. Thus, the elasticity
estimate is forced to increase in proportion to the price of water.
SINGLE-FAMILY RESIDENTIAL PRICE ELASTICITY OF WATER

DEMAND
A recent examination by Dalhuisen (2003) of 64 studies with 314 price elasticity
estimates gives some interesting insights into why the elasticity estimates vary among
studies. Table 9-9 shows selected results. For our purposes, the most important
findings are as follows:
1. Elasticity tends to be higher (further away from 0) for utilities using
increasing block rates.
2. Elasticity tends to be lower in Europe (where use rates are lower) than in
the United States.
3. Elasticity tends to be higher in the arid West than in the remainder of the
United States.
4. No significant time trend exist in the estimated elasticity values.
The first three results confirm conventional wisdom about elasticities. The
fourth, the lack of a time trend in elasticity estimates, means that results from studies
done even 30 years ago can be used with confidence today and in the future.
Customers’ responses to price increases have not changed over time. For the 314
estimates examined, the average price elasticity of demand is –0.41 and the median is
–0.35.
A cross-section study by Martin and Thomas (1986) of cities in arid regions of
the United States, Australia, and Kuwait found a price elasticity of –0.5 for an
extremely wide range of prices. Foster and Beattie (1979) estimated price elasticities
by region for the United States based on data from 218 US cities; they calculated
variations in water use per household as a function of (1) average price, (2) median
household income, (3) rainfall during the growing season, and (4) average number of
residents per household. The study used a functional form that makes the estimated
price elasticity proportional to price. Table 9-10 presents the price elasticity estimates
for three price levels.
The most general studies, such as those just cited, include a number of cities or
regions and compare use rates and prices among them. Estimates can be for outdoor
water use, indoor water use, seasonal water use, or annual water use.
Multifamily Housing Elasticities. Almost 40 percent of the US population
lives in some form of multifamily housing, ranging from condominiums to
apartments. These homes generally have very limited exterior irrigated landscaping
per unit compared to single-family residences. Typical water use in apartments is
much lower than in single-family homes, though most apartment dwellers face a
0 marginal price for water, because only 3.9 percent of multifamily complexes that
account for 10 percent of all multifamily units were individually metered in 2004. The
proportion of metered units is rapidly increasing as multifamily housing managers
install individual unit meters and charge a positive marginal price. In response to this
PRICE EFFECTS 173
Table 9–9 Selected Residential Price Elasticity Estimates
Reference Area Differentiator/Elasticity
Agthe et al. 1986 Tucson, Ariz. Long term: –0.62

Short term: –0.36
Bachrach and Vaughan 1994 Argentina –0.2 to –0.4
Danielson 1979 Kentucky At a very high price: –0.92
Deller et al. 1986 Rural Illinois,
nine districts –0.55 to –1.19
Gibbs 1978 Raleigh, N.C. Summer: –1.38
Winter: –0.31
Griffin and Chang 1990 Texas Summer: –0.37
Winter: –0.18
Hansen and Narayaman 1981 Salt Lake City, Utah –0.47
Hansen 1996 Copenhagen, Denmark –0.10
Hewitt and Hanemann 1995 Denton, Tex. Summer: –1.60
Howe 1982 United States Winter: –0.06
Eastern United States Summer: –0.57
Western United States Summer: –0.43
Maddaus and Maddaus 2006 United States Interior: –0.05
Exterior: –0.20
Martin and Thomas 1986 United States, Kuwait, –0.50
and Australia; five arid cities
Martinez-Espineira 2003 Northwestern Spain; –0.32 to –0.66
four towns
Moncur 1987 Honolulu, Hawaii –0.35
Narayanan 1985 Utah cities –0.07 to –0.09
Nauges and Thomas 1998 France –0.22
Nieswiadomy 1992 United States; 430 utilities –0.29 to –0.45
Palancia 1988 Manila, Philippines –0.287
Piper 2003 United States –0.32
Renwick and Archibald 1998 Santa Barbara, Calif. All income: –0.33
High income: –0.11
Middle income: –0.22
Low income: –0.53
Renzetti 2005 Canada Declining block rates: –0.10
Single marginal price: –0.16
Increasing block rates: –0.25
Rizaiza 1991 Saudi Arabia –0.40 to –0.78
Schneider and Whitlatch 1991 Columbus, Ohio –0.12
–0.26
Stevens 1992 Massachusetts; –0.10 to –0.69
85 communities
Taylor et al. 2004 Colorado; 34 Front-Range –0.30
utilities
Wang et al. 1999 Delaware –0.51 to –0.70
Weber 1989 Oakland, Calif. –0.10 to –0.25
Woo 1992 Hong Kong Special –0.38
Administrative Region
growth in metering, several firms offer billing services for a fee. Mayer et al. (2004)
estimated a price elasticity of demand for water in metered units of –0.275, a value
similar to that found for interior use in single-family homes.
Table 9–10 Estimated Price Elasticities by Region for Selected 2006 Prices
Price Levels (per 1,000 gal)

Region $1.00 $2.00 $3.00
Northeast –0.13 –0.26 –0.38
Midwest –0.09 –0.18 –0.27
South –0.10 –0.21 –0.31
Plains, Rockies –-0.24 –0.49 –0.73
Southwest –0.13 –0.27 –0.40
Northwest –0.29 –0.59 –0.88
Modified from Foster and Beattie 1979.
Commercial Water Price Elasticities. Most commercial accounts use water

in the same way as households, that is, indoor water use in bathrooms or for drinking
and outdoor water use for landscape irrigation. When residential and commercial
accounts have no irrigated landscaping, demand tends to be quite inelastic (between
0 and –0.3). Table 9-11 shows several estimated commercial price elasticities.
Commercial accounts with irrigated landscapes exhibit higher elasticities, up to –0.7.
Other commercial customers such as restaurants, hotels, and car washes tend to have
relatively less price response than individual households, although for some activities
they can recycle cooling water and implement other efficiency measures. Businesses
using large amounts of water for processing or cooling often can permanently reduce
water use through new capital investment in equipment to reuse water. Some
commercial firms may also be in a position to self-supply their water. Thus there may
be trigger points where a seemingly small price increase for water results in a large
drop in usage for certain customers and little response from many others. Aggregate
data hide changes such as this, so examining individual use records can reveal such
events. Thus, for a forecast of commercial water use, analysts should use elasticity
values such as those in Table 9-11, or the rules of thumb in Table 9-5, with caution.
Table 9–11 Commercial Price Elasticity Estimates
Elasticity 95% Confidence Interval

Reference Category Estimate Lower Limit Upper Limit
Moeltner and Stoddard 2004 Wholesale –0.27 –0.54 –0.001

Food and Drink –0.23 –0.33 –0.12
Banks –0.42 –0.82 –0.01
Auto Service –0.28 –0.48 –0.08
Amusement –0.63 –1.09 –0.16
Williams and Suh 1986 All –0.25 –0.36 –0.12
Hussain and Baker 2002 All –0.18 –0.31 –0.05
PRICE EFFECTS 175
Industrial Water Price Elasticities. There is good evidence that industrial

firms take the price of water into account when making investment and production
decisions. Table 9-12 shows results from several studies of the industrial demand for
water. Available research shows that the price elasticities vary dramatically by industry
and even by firm within the industry, reflecting varying price and production
conditions faced by firms in different regions. An examination of input-output tables
published by the US Department of Commerce suggests that even narrowly defined
industries include a wide range of actual production processes, reducing the value of
generalizations. These observations signal the need for well-designed studies of
industrial demand at the individual utility level to accurately include the most likely
price responses of industrial firms in forecasts. For a utility with only a few large-
volume industrial customers, developing a good relationship with their management
to determine the likely response of the firm to price changes, the opportunities for
water conserving actions and future water-related plans can lead to large payoffs.
Table 9–12 Industrial Price Elasticity Estimates
Reference Location Industry Elasticity
Babin et al. 1982 United States Food 0

Chemicals –0.74 to –0.66
Paper and wood –0.66
All –0.66 to 0
DeRooy 1974 United States Stone, clay –0.38
Cooling –0.9
Chemicals –0.89 to –0.74
Dupont and Renzetti 2001 Canada Intake water –0.77
Grebenstein and Field 1979 United States All –0.80 to –0.33
Harrington 1987 United Kingdom All –0.3
Hussain and Baker 2002 Sri Lanka All –1.3
ITRI 1997 Taiwan Cooling process –1.37 to –0.41
Malla and Gopalakrishnan 1999 Honolulu, Hawaii Food –0.4 to –0.3
Renzetti 1988 British Columbia, Forest –0.51
Canada Heavy industry –0.25
Light industry –0.54
Petrochemicals –0.12
Renzetti 1992 Canada All –0.59 to –0.15
Reynaud 2003 France All –0.79 to –0.10
Schneider and Whitlatch 1991 Columbus, Ohio All –1.16
Turnovsky 1969 Massachusetts All –0.5
Wang and Lall 2001 China Leather –1.2
Electric power –0.57
Williams and Suh 1986 United States All –0.97 to –0.44
Ziegler and Bell 1984 Arkansas All –0.08
Involving industry management in developing and evaluating industrial water-

demand forecasts is likely to greatly improve these efforts. Some industrial firms may
be in a position to substitute self-supplied water for utility supplied water, and will do
so if price increases call their attention to spending on water and the economics of
self-supply are favorable.
Reynaud (2003) found that 32 percent of the industrial plants in a French
survey used water in their production process, with 71 percent of these reporting use
for cleaning and 52 percent for cooling. Among these plants, 96 percent of production
cost was for labor, and of the expenditures for water, 64 percent was to purchase water
from a local utility, 28 percent for self-supplied water, and 8 percent for water
treatment at the facility. Reynaud estimated a range of elasticities for purchased water
from –0.10 to –0.79 with a mean of –0.29. The elasticity of water reuse with respect to
its cost to the firm was relatively elastic, however, with a range from –0.90 to –2.21
and a mean value of –1.42.
Renzetti and Dupont (1999) reported that if fees were initiated for industrial
water withdrawals in Canada, use would fall by between 1.2 percent when flat fees
were imposed and 8.27 percent with a volumetric fee of $5.00 Canadian per 1,000 m3.
Grebenstein and Field (1979), using aggregate US data, reported that the share
of cost for water varied from 1.2 percent to 1.9 percent. A survey by Liaw and Chen
(2004) revealed a wide variation in water reuse ratios, which suggests substantial
room for more industrial water reuse. Price increases certainly are a powerful
incentive for greater reuse. Reuse ratios varied from an average of 19.1 percent in
Taiwan to 49.9 percent in Japan to 65.2 percent in the United States. Water reuse
increases with improvements in reuse technology and with water prices. Reuse
becomes increasingly economical when water prices, including combined charges for
fresh water delivery and wastewater disposal, rise above about $1.00/m3. A linear
programming model of the electric generating industry in Poland (Stone et al.1982)
4
Price
0
0 10 20 30 40 50
Quantity
Source: Stone, et al 1982.
Figure 9–6 Demand for Water by Electric Power Generation

PRICE EFFECTS 177
resulted in the demand curve for water shown in Figure 9-6. This research shows a
dramatic reduction in water use once price increased to above 1 zloty/m3. Renzetti
(2002) collected numerous research reports on industrial water use.
SUMMARY
The diversity of rate schedules, including decreasing and increasing block rates,
suggests some problems with defining or identifying the price of water. The following
four main price specifications are noted in this discussion: average price, marginal
price, two-part price specification, and total bill. Evidence from (1) controlled
comparisons, (2) analysis of flat rate versus metered customers, and (3) econometric
studies suggests that final resolution of the price specification problem is not
necessary to demonstrate the existence of consumer responses to changes in water
rates.
Because price elasticities are low and inflation typically erodes the impact of
price increases, and because water is relatively cheap,5 water resource planning
discussions underplay and sometimes ignore rate effects. When water is unmetered or
rates are very low, modest price changes can probably be safely ignored (Houtakker
and Taylor 1970; Wong 1972). Nevertheless, rate or price effects are important
because water rates and charges are usually the primary source of revenue for the
water utility, and large rate increases can be counted on to cause a drop in sales.
The bottom line is that even relatively moderate price responsiveness in the
urban water system can produce large errors in estimates of revenue when price
effects are ignored. We suggest a relatively simple method for estimating price impacts
of rate changes on urban water demand. Although these computations will not
produce perfect accuracy, analysts can develop water-use forecasts that account for
consumer response to price changes, and these forecasts will be more accurate than
those that ignore such effects.
5. In 1990 residential water rates averaged less than $2.00/1,000 gal, ranging from $1.26 to $2.34/1,000
gal. At these rate levels, water service to single-family residences cost between $200 and $400/year, or
about $20 to $30/month.
REFERENCES
Agthe, D.E., R.B. Billings, J.L. Dobra, and K Raffiee. 1986. A Simultaneous Equation Demand Model for
Block Rates. Water Resources Research 22(1): 1–4.
Babin, F.B., C.E. Willis, and P.G. Allen. 1982. Estimates of Substitution Possibilities Between Water and
Other Production Inputs. American Journal of Agricultural Economics 64(1): 148–151.
Babin, F.B., C.E. Willis, P.G. Allen, and A.S. Vlachou. 1981. Substitution Possibilities for Water Inputs in
Selected Industries. NTIS PB81-209090. Washington, D.C.: National Technical Information Service.
Bachrach, M., and W.J. Vaughan. 1994. Household Water-Demand Estimation. Working Paper. Washing-
ton, D.C.: Inter-American Development Bank (IDB).
Economics 56(2): 73–84.
Billings, R.B., and W.M. Day. 1982. Price, Socioeconomic, and Climatic Influences on Residential Water
Use in Southern Arizona, Project Completion Report. Tucson, Ariz.: Office of Water Resources
and Technology.
———. 1989. Demand Management Factors in Residential Water Use: The Southern Arizona
Experience. Journal AWWA 81(3): 58–64.
Brown, T.C. 2006. Trends in Water Market Activity and Price in the Western United States. Water
Resources Research 42:W09402.
Dalhuisen, J.M. 2003. Price and Income Elasticities of Residential Water Demand: A Meta-Analysis. Land
Economics 79(2): 292–309.
Danielson, L.E. 1979. An Analysis of Residential Demand for Water Using Micro Time-Series Data. Water
Resources Research 15(4): 763–767.
Deller, S.C., D.L. Chicoine, and G. Ramamurthy. 1986. Instrumental Variables Approach to Rural Water
Service Demand. Southern Economic Journal 53(2): 333–346.
DeRooy, J. 1974. Price Responsiveness of the Industrial Demand for Water. Water Resources Research
10(1): 403–406.
Dupont, D. and S. Renzetti. 2001. Water’s Role in Manufacturing. Environmental and Resource Economics
18(4): 411–432.
Espey, M., J. Espey, and W.D. Shaw. 1997. Price Elasticity of Residential Demand for Water: A Meta-
Analysis. Water Resources Research 33(6): 1369–1374.
Foster, H.S., and B.R. Beattie. 1979. Urban Residential Water Demand for Water in the United States.
Land Economics 55:43–58.
Gibbs, K.C. 1978. Price Variable in Residential Water-Demand Models. Water Resources Research
14(1): 15–18.
Grebenstein, C., and B. Field. 1979. Substituting for Water Inputs in U.S. Manufacturing. Water Resources
Research 15(2): 228–232.
Griffin, R.C. 2006. Water Resource Economics: The Analysis of Scarcity, Politics, Projects. Cambridge, Mass.:
MIT Press.
Griffin, R.C., and C. Chang. 1990. Pretest Analysis of Water Demand in Thirty Communities. Water
Hanke, S.H. 1970. Demand for Water under Dynamic Conditions. Water Resources Research 6(5): 1253–
1261.
Hansen, L.G. 1996. Water and Energy Price Impacts on Residential Water Demand in Copenhagen. Land
Economics 72(1): 66–79.
Hansen, H.D., and R. Narayanan. 1981. A Monthly Time Series Model of Municipal Water Demand.
Water Resources Bulletin 17(4): 578–585.
Harrington, P. 1987. Pricing for Water Services. Paris, France: Organisation for Economic Co-operation
and Development.
Hewitt, J.A., and W.M. Hanemann. 1995. A Discreet/Continuous Choice Approach to Residential Water
Demand under Block Rate Pricing. Land Economics 71(2): 173–192.
PRICE EFFECTS 179
Houthakker, H.S., and L.D. Taylor. 1970. Consumer Demand in America. 2nd ed. Cambridge, Mass.:
Harvard University Press.
Howe, C.W. 1982. The Impact of Price on Residential Water Demand: Some New Insights. Water
Howe, C.W., and F.P. Linaweaver Jr. 1967. The Impact of Price on Residential Water Demand and Its
Relation to System Design and Price Structure. Water Resources Research 3(1): 13–32.
Hussain, I.S.T., and R. Baker. 2002. Economic Analysis of Residential, Commercial and Industrial Uses of
Water in Sri Lanka. Water International 27(2): 183–193.
ITRI (Industrial Technology Research Institute). 1997. A Report on the Elasticity of Industrial Water
Demand in Taiwan. Taipei: Taiwan Department of Industrial Technology.
Liaw, C.H., and L.C. Chen. 2004. Rational Industrial Water Reuse Ratios. Journal of the American Water
Resources Association 40(4): 971–979.
Linaweaver, F.P. Jr., J.C. Beebe, and F.A. Skrivan. 1966. Data Report of the Residential Water Use Research
Project. Baltimore, Md.: Johns Hopkins University, Department of Environmental Engineering.
Maddaus, W.O., and L.A. Maddaus. 2006. Water Conservation Programs—A Planning Manual. Denver,
Colo.: AWWA.
Malla, P., and C. Gopalakrishnan. 1999. The Economics of Urban Water Demand: The Case of Industrial
and Commercial Use in Hawaii. Water Resource Development 15(3): 367–374.
Martin, W.E., and J.T. Thomas. 1986. Policy Relevance in Studies of Urban Residential Water Demand.
Water Resources Research 22(Dec.): 1735–1741.
Martinez-Espineira, R. 2003. Estimating Water Demand Under Increasing Block Tariffs Using Aggregate
Data and Proportions of Users per Block. Environmental and Resource Economics 26:5–23.
Mayer, P.W., E. Towler, W.B. DeOreo, E. Caldwell, T. Miller, E.R. Osann, E. Brown, P.J. Bickel, and
S.B. Fisher, 2004. National Multiple Family Submetering and Billing Program Study. San Francisco,
Calif., Aquacraft, Inc. and East Bay Municipal Utility District.
Merrett, S. 2005. The Price of Water. La Vergne, Tenn.: Lightning Source Inc.
Moeltner, K., and S. Stoddard. 2004. A Panel Data Analysis of Commercial Customers’ Water Price
Responsiveness under Block Rates. Water Resources Research 40(WO1401): 1–9.
Moncur, J.E.T. 1987. Urban Water Pricing and Drought Management. Water Resources Research 23(3):
393–398.
Narayanan, R., D.T. Larson, and T.C. Hughes. 1985. Effectiveness of Drought Policies for Municipal
Water Management. Water Resources Bulletin 21(3): 407–416.
Nauges, C., and A. Thomas. 1998. Privately Operated Utilities, Municipal Price Negotiation, and
Estimation of Residential Water Demand: The Case of France. Presentation at European
Association of Environmental and Resource Economists (EAERE) annual meeting, Venice.
Nieswiadomy, M.L. 1992. Estimating Urban Residential Water Demand: Effects of Price Structure,
Conservation, and Education. Water Resources Research 28(3): 609–615.
OECD. 1987. Pricing of Water Services. Paris, France: OECD.
Palencia, L.C. 1988. Residential Water Demand in Metro Manila. Journal of the American Water Resources
Association 24(2): 275–279.
Piper, S. 2003. Impact of Water Quality on Municipal Water Price and Residential Water Demand and
Implications for Water Supply Benefits. Water Resources Research 39(5): 1–9.
Renwick, M.E., and S.O. Archibald. 1998. Demand-Side Management Policies for Residential Water Use:
Who Bears the Conservation Burden? Land Economics 74(3): 343–359.
Renzetti, S. 1988. An Econometric Study of Industrial Water Demands in British Columbia, Canada.
———. 1992. Estimating the Structure of Industrial Water Demands: The Case of Canadian
Manufacturing. Land Economics 68(4): 396–404.
———. ed. 2002. The Economics of Industrial Water Use. Northampton, Mass.: Edward Elgar.
Renzetti, S., and D.P. Dupont. 1999. An Assessment of the Impact of Charging for Provincial Water Use
Permits. Canadian Public Policy 25(3): 361–378.
Reynaud, A. 2003. An Econometric Estimate of Industrial Water Demand in France. Environmental and
Resource Economics 35:213–234.
Rizaiza, O.S.A. 1991. Residential Water Usage: A Case Study of the Major Cities of the Western Region of
Saudi Arabia. Water Resources Research 27(5): 667–671.
Schneider, M., and E. Whitlatch. 1991. User-Specific Water-Demand Elasticities. Journal of Water
Resources Planning and Management 117(1): 52–73.
Stevens, T.H., J. Miller and C.E. Willis, 1992. Effect of Price Structure on Residential Water Demand.
Journal American Water Resources Association 28(4): 681–685.
Stone, J.C., F.D. Singleton, Jr., A. Saleqicz, M. Gadkowsaki, and W. Sikorski. 1982. Water Demand for
Generating Electricity: A Mathematical Programming Approach With Application in Poland. RR-82-
16. Laxenburg, Austria: IIASA.
Taylor, R.G., R.A. Young, and J.R. McKean. 2004. Alternative Price Specifications for Estimating
Residential Water Demand with Fixed Fees. Land Economics 80(3): 463–475.
Turnovsky, S.J. 1969. The Demand for Water: Some Empirical Evidence on Consumers’ Response to a
Commodity Uncertain in Supply. Water Resources Research 5(2): 350–361.
Wang, H., and S. Lall. 2001. Valuing Water for Chinese Industry: A Marginal Productivity Analysis.
Applied Economics 34(6): 759–765.
Weber, J.A. 1989. Forecasting Demand and Measuring Price Elasticity. Journal AWWA 81(5): 57–65.
Williams, M., and B. Suh. 1986. The Demand for Urban Water by Customer Class. Applied Economics
18:1275–1289.
Wong, S.T. 1972. A Model of Municipal Water Demand: A Case Study of Northeastern Illinois. Land
Economics 48(1): 34–44.
Woo, C.K. 1992. Drought Management Service Interruption and Water Pricing: Evidence From Hong
Kong. Water Resources Research 28(10):2591–2595.
Young, R. 2005. Determining the Economic Value of Water: Concepts and Methods. Washington, D.C.:
Resources for the Future Press.
Ziegler, J.A., and S.E. Bell. 1984. Estimating Demand for Intake Water by Self-Supplied Firms. Water
CHAPTER 10
Long-Term Water
Conservation
One of the challenges facing utility managers is predicting the most likely effect of
conservation programs on water use and utility revenue. The task is complicated by
the fact that utilities frequently implement several water-efficiency programs and
measures simultaneously; thus synergism and competition among programs can
become an issue in projecting system water savings.
The goal of a conservation program may be to bring about a specified
percentage reduction in water use, to keep water use from exceeding a predetermined
amount, to reduce the rate of growth of water demand to postpone water treatment
projects, or to prepare for drought. Conservation is often the least-cost method of
accommodating the demands of a growing community. Thus successful long-term
conservation can keep water rates below what they would otherwise need to be,
because it may postpone or even eliminate the need for additional supplies of raw
water and expanded treatment and transmission facilities. Ideally, the water savings
from conservation would occur in step with new water demands in the utility service
area, maintaining a balance between supply and demand. Indoor water conservation
will also reduce sewer flows and the need to expand wastewater treatment facilities.
Conservation may provide significant long-term savings on both water and sewer use
bills; median monthly water bills for single-family residences in the United States in
2004 were $19.08 compared to wastewater charges of $22.23 (Raftelis 2004).
This chapter discusses a variety of conservation initiatives, along with engineer-
ing estimates of their potential for saving water, and verified water conservation
results from careful analysis of the experience of numerous utility districts. Table 10-1
summarizes these results. Chapter 11 discusses additional measures that are short-term
181
Table 10–1 Anticipated Water Savings by Device or Program.

SDFSDFSDF
Water Life Water Savings

Description Use in Years (gpcd) Percentage Comments
Bathroom
Showerhead 2.5 gpm 5–10 0–2.4 0–21 Use habits
Faucet aerator 2.2 gpm 5 0–1.6 Use habits
Toilet (new) 1.6 gpf 20 + 9–20 50–70 Replaced unit age
Kitchen
Faucet aerator 2.5 gpm 5 0–1.3
Insulate hot pipes 20 + 20+
Dishwasher 7 gal 10–15 0.3–0.6 30–60 Replaced unit age
Household other
Clothes washer
(front-loading) 30 gal 15–20 5.5 30 Replaced top loader
Pressure reducer ≤ 80 psi 20 + 3-6
Submeter apartments 20 + 16
Leak repair 3+ 5+
Landscape
Rain sensor 10
Evapotranspiration
controller 10–15 18–22
Native plants 5–20
Commercial
Urinals 1 gpf 20+ 3
Urinals, waterless 0 20+ 6
Sources: Mayer et al. 1999; Maddaus and Maddaus 2006; Vickers 2001.
or emergency in nature, such as appeals to take shorter showers and temporary

limitations on irrigation.
This chapter also estimates the most likely effect of rate increases and changes
in the structure of rates on both water sales and revenues. When time and resources
permit, on-site statistical research is often helpful in pinning down the effects of price
changes, either through control group comparisons or with regression models, using
a scientific sample of data on customer water use. Utilities sometimes conduct small-
scale pilot programs to fine-tune a program and evaluate its effectiveness and cost
before deciding to expand the program to all customers. This tactic has been used
with various retrofit and rebate programs. Detailed discussions of these approaches to
estimating water savings potential are presented in this chapter, along with the results
from several careful evaluations of the effectiveness of various programs. Crucial
issues surrounding conservation programs include market penetration, the likely
range of effectiveness, persistence over time, and costs to both the utility and its
customers.
A survey of major utilities in the United States in 2006 showed widespread
implementation of both long- and short-term water-conservation programs, as
LONG-TERM WATER CONSERVATION 183
shown in Table 10-2. The most common initiative is public educational efforts,
employed by 55 percent of the surveyed utilities. Their widespread use is readily
understood, because virtually every conservation program has a significant educa-
tional element and no program can be expected to be successful without using some
educational and publicity components. Increasing block rates have also become
increasingly common, as utilities try to reduce usage through the price mechanism
without angering many voters and customers and without the sometimes large budget
surpluses that could result from higher uniform rates.
Table 10–2 Water Conservation Programs, US Utilities, October 2006
Which (if any) of the following water conservation programs are being
implemented in your water system? Please mark all that apply.
Response Count Percentage
Public education programs 365 55.3

Rebates for installation of low-flow
or ultra low-flow toilets 67 10.2
Free distribution of shower-flow restrictors
and toilet dams 117 17.7
Rebates for front-loading washing machines 49 7.4
Rebates for other water-efficiency appliances,
hardware, or fixtures 28 4.2
Conservation rates, such as increasing block rates,
seasonal rates, water budgets 201 30.5
Programs to affect landscaping 147 22.3
Emergency water restrictions 209 31.7
None of the above 176 26.7
Source: AWWA Water Demand Survey, October 2006 (see appendix B).
EVALUATING WATER CONSERVATION PROGRAMS

The focus of conservation programs may be to give customers a motive for saving
water, methods of saving water, or, in most cases, some combination of motives and
methods. Nevertheless, significant water savings are being achieved by plumbing
codes that set limits on the water-use rates of toilets, faucets, showerheads, and water-
using appliances, often without the active participation of consumers. Many
consumers using these more water-efficient fixtures are not even aware of the
differences in water use. To achieve additional water conservation, customers must be
motivated to save water. Simply having water-saving technologies available does not
ensure that they will be adopted. Customers may also engage in offsetting behavior
such as installing a decorative fountain using the water saved by installing a low-flush
toilet. Offsetting behavior is least likely when water users are convinced that
conserving water is highly desirable in and of itself.
Campbell and colleagues (2004) suggest a division of water conservation
programs into five categories:
1. Regulations
2. Pure engineering, which includes a variety of water saving devices and
technologies
3. Mediated engineering, which includes personal contact along with devices
4. Public information (including general publicity) and educational efforts
5. Price-related strategies
The following lists illustrate actual or proposed programs falling into each of
these groups:
Regulations
• Plumbing codes specifying low-water-use fixtures
• Landscape ordinances requiring low-water-use plantings
• Water waste prohibition
• Requirements for raw or nonpotable water for certain outdoor purposes
• Requirements for retrofit on resale
Pure Engineering
• Free distribution or subsidies for new toilets
• Free distribution of low-flow showerheads and faucet aerators
• Subsidies for low-water-use clothes or dishwashers
• Subsidies for landscape conversion and sprinkling system upgrades
• Subsidies for rainwater harvesting and use on landscapes
• Subsidies for on-site graywater reuse on landscaping
• Extension of a reclaimed water delivery system
• Leak-detection programs
Mediated Engineering
• Any engineering solution with personal interaction
• Water-use audits directed at customers with high water use
Public information
• General publicity: billboards, mailings, water bill inserts
• Public forums
• Coverage in newspapers, radio, TV, and the Internet

• Educational programs in elementary and high schools
• Workshops to promote efficient irrigation systems
• Workshops to promote water-wise landscaping (xeriscape)
• Adding landscape irrigation needs to daily weather reports
Price-related strategies
• Meters with commodity charges
• Universal metering of water deliveries
• Prominent display of the marginal price of water on bills
• Increase in the price of water
• Higher water prices for peak demand periods
• Increasing block rates for water
• Water budgets
• Subsidy of reclaimed water to make it economically attractive
Utilities have instituted many programs in a variety of ways, with results that
often vary significantly. A successful program requires careful planning. A program
proposal should include estimates of cost as well as potential water savings and
expected variations in those savings over time. For example, for a toilet retrofit
program, how many customers per month can be expected to respond to the
program, and how long should the program last? When a program has ended, the
utility should carefully analyze the results to give guidance for the next round of
programs. For example, the success of a leak-detection program in which the utility
conducts audits of water use by individual customers can be determined after the fact
by how many of the targeted customers participated, and by measured decreases in
their water use compared to what it would have been without the program. This
determination can, however, be a challenge, because no one can know with certainty
what would have happened without the program. The persistence of the water savings
depends on the rate at which the repairs undertaken deteriorate to pre-audit
conditions. This is affected by the willingness and ability of the customer to
incorporate leak detection and repair into their regular maintenance procedures. For
other programs, such as low-flush toilet rebates, the water savings can be expected to
be permanent.
The long-term potential for water conservation is signaled by the wide range of
water-use rates in cities around the world (see Figure 10-1). This is not to suggest that
a city like Phoenix, Arizona, could readily adapt to the use rates in Coober Pedy,
Australia, but rather that multiple conservation opportunities exist, and water use
could drop dramatically if needed.
700
Liters per Capita per Day

600
500
400
300
200
100
0
lia
lia
S
it
ZU
wa
tra
tra
ZU
Ku
us
us
,A
A
,A
,A
n,
ni x
so
dy
r th
oe
c
Pe
Pe
Tu
Ph
er
ob
Co
Source: Martin and Thomas 1986.
Figure 10–1 Water Use in Cities
Synergism, Competition, and Independence. Before analysts can use

conservation effects in forecasting the effectiveness of a program, the potential
interactions among programs implemented must be analyzed. Often, a utility will
implement several water-conservation or efficiency programs at the same time, e.g., a
publicity campaign, a rebate program, and a rate increase. This requires considering
synergism and competition (antagonism) between the measures implemented to
derive reliable estimates of overall water savings.
A synergism is when the total water savings from implementing several water-
efficiency or conservation measures is greater than the sum of the expected water
savings from each measure by itself. An example is the interaction of pricing with
other conservation measures. Higher water prices encourage consumers and produc-
ers to take advantage of the full potential for water savings from available measures or
devices. Estimating price and nonprice measures independently is expected to provide
a low estimate of conservation for programs including both measures, provided that
both the price elasticity and the nonprice effects are measured in an appropriately
specified model. Insufficient research has been carried out, however, to ascertain the
size of these effects with any certainty.
Competition or antagonism between water-conservation measures, practices,
programs, or policies exists when total water savings are less than would be indicated
by adding up the expected individual water savings. For example, mandatory
restrictions and price increases may be antagonistic because the conservation that
would be induced by a price increase may be effectively mandated by restrictions on
irrigation or other outdoor water uses (Goemans 2007). In this case, the price increase
may have no effect on water use beyond that caused by the regulations, so its impact
is to increase utility revenue compared to what it would otherwise have been. If the
mandatory restrictions are later cancelled, however, because of an improving supply

situation, the higher price will continue to support water conservation as well as
provide needed revenue.
Benefit-Cost Comparisons. Most analysts suggest that water-conservation

or efficiency programs, policies, practices, or measures should be evaluated using a
“net benefits” criterion to determine if the program is worth its cost.1 A comprehen-
sive evaluation of urban water conservation programs (Planning and Management
Consultants et al. 1991) lists seven standard methods of economic analysis that can be
performed to evaluate relative costs and benefits of urban water-conservation
programs. Computing a benefit-to-cost ratio can be used for every program, and
most of the other methods are appropriate for programs involving capital outlays and
long-term program effects. The methods include
• Benefit-to-cost ratio
• Net present value
• Internal rate of return
• Discounted payback
• Life-cycle revenue impacts
• Levelized costs and benefits
• Number of years to recover capital outlay
Every evaluation approach requires careful estimates of the anticipated reduc-
tion in water use from the program and the resource costs to both the utility and its
customers. The evaluations also require estimates of the nonmonetary amenity cost to
consumers of reduced water use. Some programs, such as subsidies to consumers to
install low-flush toilets, share the resource cost between the utility and consumer, with
essentially all of the cost being explicit. This program has no amenity cost. In contrast,
a program to limit lawn sprinkling to once a week will reduce water use, thereby
lowering utility revenue. It will save customers money in the form of reduced water
and perhaps fertilizer bills, but impose potentially significant amenity cost in the form
of less lush lawns.
To facilitate the process of choosing a program, analysts can rank measures with
expected positive net benefits according to social acceptability, based on local values,
practices, and experience. There is typically broad consensus about the value of
• Pursuing leak control in the water system and on customer’s premises
1. An early, landmark study of water conservation states, “The first axiom of the definition of water
conservation is that the beneficial effects of the reduction in water use (loss) must be considered greater
than the adverse effects associated with the commitment of other resources to the conservation effort”
(Bauman et al. 1978).
• Providing incentives to customers to install and use low-water-use appliances

and fixtures
• Installing in-house recycling equipment for industrial cooling and washing
systems
Even without any local program, national plumbing codes are resulting in
substantial sustained reductions in water use that will continue for years as obsolete
fixtures are replaced.
Pricing policies are often politically controversial, but can be effective. Many are
designed to reduce the peak level of demand, because peak demands often contribute
much more to escalating system capital and operating costs than to revenue. Direct
controls in the form of water restrictions or ordinances limiting new construction
usually are short-term solutions, because of public dislike of measures designed to
compel behavior.
STRUCTURAL CONSERVATION PROGRAMS

The proportion of water used for landscaping and the distribution of interior water
use suggest one measure of the potential for water conservation. Table 10-3 shows
exterior and interior water use for a select group of cities. As with the total water use
by city shown in Figure 10-1, there is a wide variation.
Exterior water use for landscaping varies with local climate and traditions, with
some cities in hot arid climates devoting more than half of all water to lush
landscaping and other cities, especially in cooler wet climes, using most water
indoors. Figure 10-2 shows a typical distribution of single-family residential interior
water use. Conservation programs targeting activities using larger percentages of total
water, such as toilets, showers, faucets, leaks, and clothes washers, offer the best
opportunities for water savings. Surprisingly, dishwashers use less than 2 percent of
residential interior water; thus even if they become more water frugal, the net effect
on total water use will be small.
Water Audits and Leakage Reduction. Mayer and colleagues (1999) found
that a small proportion of homes accounted for most leakage. Because leakage was
inferred from water-use logging, some of this “leakage” may have been intended
beneficial uses such as letting a hose drip on vegetation. Average daily leakage was
21.9 gal with a standard deviation of 54.1. Median leakage was 4.2 gal. The highest
100 homes (out of 1,188) leaked an average of 90.4 gal/day. Considering the sample of
individual homes from 12 cities, average leakage was highly correlated with total
average indoor usage. These results suggest that conservation programs targeted to
homes with exceptionally high water use during nonirrigation periods could be quite
effective.
North Marin County Water District audited single-family homes randomly
selected from those in the top quartile of water use, which averaged 41 percent above
all single-family residences in 1988 (Nelson 1992). The district invited 1,276
Table 10–3 Annual Indoor and Outdoor Water Use per Household
Indoor Outdoor
1,000 gal Percent 1,000 gal Percent
Ontario, Canada
Waterloo 67.7 89.7 7.8 10.3
Cambridge 71.2 90.1 7.8 9.9
Florida
Tampa 56.1 64.8 30.5 35.2
Northwestern United States
Seattle, WA 54.1 71.4 21.7 28.6
Eugene, OR 65.1 57.2 48.8 42.8
Colorado
Boulder 54.4 42.5 73.6 57.5
Denver 61.9 37.2 104.7 62.8
Southern California
Walnut Valley 76.3 39.9 114.8 60.1
San Diego 55.3 35.8 99.3 64.2
Lompoc 62.1 58.8 43.5 41.2
Las Virgenes 70.9 25.0 213.2 75.0
Arizona
Tempe 65.2 39.4 100.3 60.6
Scottsdale 60.1 27.7 156.5 72.3
Phoenix 70.8 30.4 161.9 69.6
Source: Mayer 1999; Table 5.14, p. 114.
homeowners to participate by letter and performed audits for the 237 homeowners
who accepted. The labor expended for each audit averaged two hours; the audit
included measuring flow rates, replacing showerheads if they used more than
3 gallons per minute (gpm) installing toilet tank bottles, repairing flappers, and
recommending effective irrigation practices. A careful evaluation involved pairing
each audited home with a control group home in the same area of town with the same
pre-audit water-use pattern. The statistical analysis dropped about 14 percent of the
audited homes because the owner had moved and another 14 percent because water-
use records indicated an obvious leak during the study period. An alternative would
have been to adjust the water use during the month when the leak occurred, but this
would raise the challenge of choosing the “correct” amount of water that would have
Clothes
Other, 2.2
Washer,
21.7
Toilet, 26.7
Dishwasher,
1.4
Bath, 1.7
Faucet,
Shower,
15.7
16.8
Leak, 13.7
Source: Adapted from Mayer et al. 1999; Figure 5.5, p. 87
Figure 10–2 Typical Single-Family Residential Indoor Water Use, by Percentage
been used without a leak. The sample size was sufficient, so this was unnecessary.
Actual water savings by the audited homes was 4.6 percent compared to the control
group.
Low-flow showerheads installed during the audit raised the percentage of
homes using them from 30 to 91 percent. The percentage of toilets with displacement
devices went from 11 (the remainder from a 1978 program) to 100 percent. At the
time of the audits, 41 percent of the toilets were 3.5 gal flush (which became
mandatory for new construction in California in 1978). Eighteen percent of the toilets
examined leaked, and the auditor was able to repair 81 percent of them, mostly by
installing new flapper valves. Repair recommendations were made for remaining
leaks. A survey of audited homes conducted 20 months later found that homeowners
had removed about 6 percent of the toilet tank displacement devices and new shower
heads. Ninety percent of respondents reported following up on repair suggestions,
and 55 percent reported using the irrigation schedule recommended by the auditor.
Seventy-six percent of the audited homes were using an average of 15.5 percent less
water than before the audit, but when compared to the control group, actual water
savings were 4.6 percent. Exterior water savings were far less than anticipated. Only
27 percent of the anticipated exterior savings actually occurred.
Each audit cost the utility $45.00 and, counting reduced water use and hot
water heating cost, saved the homeowner an average of $22.62 per year, more than
half of which was for heating shower water. Thus, homeowners would experience a
two-year payback period if they paid for the audit. For the utility, the payback period
from water saving was 30 years, signaling that this was an inefficient use of utility
funds. The utility might lower its cost by requesting homeowners pay for the audit,
either when it was completed or with their water bill for the next 24 months. The low
acceptance rate of this program of less than 20 percent suggests that additional
incentives would be necessary to achieve higher participation rates. An increase in the
marginal price in the top block of an increasing block rate structure would help
provide this incentive. In this case, including the conservation effect of a price
increase and the effect of audits in the forecast would overstate total conservation.
Educational efforts could increase the effectiveness of a “don’t waste water” message.
Toilet Rebate Program. Low-flow (1.6 gallon per flush [gpf]) toilets save
water without any negative impacts on users. The 1.6-gpf toilets have been well
received by consumers (Vickers 2001; Consumer Reports 2005) and there is no
evidence that double flushing has increased compared with higher volume toilets
(Mayer et al. 1999; Mohadjer, 20032). The potential water savings from their
installation depends almost entirely on the characteristics of the existing stock of
toilets in the utility service area. Federal law in the United States has mandated that
toilets sold starting in 1994 use no more than 1.6 gpf. Some other countries limit flush
volumes to 3.0 to 4.5 L (0.8 to 1.2 gal), and California has considered legislation for
this lower standard.
Figure 10-3 shows the frequency distribution of recorded flush volumes (Mayer
et al. 1999). Converting toilets to the newer 1.6-gpf units can reliably save anywhere
from 2 to 5 gpf; dual flush units, which are less common, would save even more in the
0.8-gpf liquid-waste-only mode. Toilet prices reported in August 2005 ranged from
$200 to $500, with Consumer Reports top-rated models averaging $285 and their best
buy coming in at $200. Almost all tested toilets did a good job of cleaning the sides of
the bowl, but more than half did only a fair to poor job of flushing solid waste. Thus
selection of effective models is important. Installation cost, which is likely to exceed
$200, is a consideration in planning a toilet replacement program or forecasting the
acceptance rate of such a program.
20
Relative Frequency
15
10
0
1 2 3 4 5 6 7 8
Flush Volume, in Gallons
Source: Adapted from Mayer et al. 1999; Figure 5-10, p. 97
Figure 10–3 Toilet Flush Volume
The high frequency of 3.5-gpf toilets in Mayer and colleagues (1999) is not
surprising, considering that all of the areas in the survey are high growth areas, with many
homes built when 3.5-gpf toilets were mandated. The secondary peak in Figure 10-3
between 1.5 and 2.5 gpf suggests the installation of 1.6-gpf toilets in homes
constructed since that standard was enacted, or homes in areas with retrofit programs.
2. Available at www.cuwcc.org/toilet_fixtures.
If every toilet performed according to specifications, the frequency distribution would

have very large values at 1.6, 3.5, and 5.0 gpf. The frequency distribution shows that
actual flush volume is often larger than the labeled volume because of variations in
manufacturing, improper adjustment, or leaks. Figure 10-3 suggests that predictions
of water savings from toilet retrofit programs depend primarily on the effectiveness of
the program and the age distribution of existing toilets.
In older, slow-growth areas that have not had toilet replacement programs,
many toilets probably use 5 to 7 gpf, and a rebate program would achieve greater
savings than in the high-growth western cities included in this report. The survey by
Mayer (1999) found that homes built before 1970 or after 1990 were more likely to
have 1.6-gpf toilets than those built in the middle years, suggesting that targeting
homes by age might be an effective strategy and needs to be included in the forecast.
Several utilities have instituted programs to provide free 1.6-gpf toilets or
rebates. New York City (Department of Environmental Protection, 1996) sponsored a
rebate program for residential and commercial customers between 1994 and 1997 that
resulted in replacing 1.3 million toilets, many of which were of the old 5- to 7-gpf
type. The estimated reduction is water use is 70 mgd, or 54 gpd per toilet. The
estimated savings in Los Angeles, which helped fund the replacement of 900,000
toilets, was smaller, at 32 gpd, probably reflecting the replacement of large numbers of
3.5 gpf units. A similar program in Tampa, Fla., replaced 15,300 fixtures, with an
estimated savings per toilet of 29 gpd. Reported savings from other cities include
79 gpd in Toronto, Canada; 41 gpd in Salt Lake City, Utah; and 22 gpd in Seattle,
Wash. (Mohadjer 2003; Vickers 2001).
Chesnutt and coauthors (1994) evaluated the effect of toilet replacement in
single-family homes with multiple toilets in Los Angeles. Savings in homes with one
toilet were 24.2 gpd. If the home had two, savings from the first was 33.7 and the
second saved 20.8. Homes with three realized savings of 45.9, 36.1, and 20.1 gpd,
respectively. In this study, the number of toilets was highly correlated with the
number of people in the household. In areas of large homes with small families, the
savings per toilet, especially for the second and third toilet, would be dramatically
lower, especially in homes with only adults who work outside the home.
Higher marginal prices for water and effective publicity campaigns increase the
incentive to replace toilets in these programs. The homes in the West Coast research
(Mayer et al. 1999) would save an estimated 8,650 gal/year by retrofitting with new
toilets. If the marginal price of water is $4.00/1,000 gal, this would save $35/year. At
an installed price of $400, the payback period is more than 11 years, with an internal
rate of return on this investment of 6 percent, assuming a 20-year life for the new
toilet. Although this is higher than most people could earn from bank certificates of
deposit, many consumers need some added incentive to make this investment. Using
the New York data, water savings are 19,710 gal/year; with the same marginal price of
$4.00, this would save $79/year, giving a payback of 5 years. This is an acceptable
payback period for a program in which the benefits can be expected to last for
20 years or more.
Part of the challenge in forecasting water savings from a toilet replacement

program is estimating the penetration rate and the actual savings per toilet. Toilets
have a life span of 25 to 50 years, although many are replaced after 15 to 20 years.
Mandatory programs such as retrofit-on-sale with serious enforcement can be
expected to have very high participation, approaching 100 percent. For voluntary
programs, penetration rates depend on local conditions and choices, including the
size of rebates, the ease of program participation, the intensity of publicity and
educational efforts, the civic-mindedness of targeted customers, and the age of the
stock of toilets.
Analysts can estimate the characteristics of the existing stock of toilets based on
the age of the housing stock and nonresidential buildings, prior toilet replacement
programs, and the estimated rate of bathroom remodeling in the area. Kitchen and
bathroom remodeling is a function of income, and is particularly prevalent in areas of
rapidly rising incomes such as neighborhoods where low-income residents are being
replaced by households with high incomes.
Field studies by the California Urban Water Conservation Council (2001) found
that few commercial, industrial, and institutional (CII) water users had ultra low-
flush (ULF) toilets in place in 1997, as shown in Table 10-4. Because manufacturers
shifted to 3.5-gpf toilets in 1977, and then to 1.6 gpf in 1992, the table suggests that
38.5 percent of these toilets are more than 20 years old and another 55.4 percent are
between 7 and 20 years old. Combining use rates with toilet ages, the study estimated
the potential water savings by sector of toilet replacement, as shown in Table 10-5.
This research also offers a method for estimating the total number of toilets in the CII
sectors based on readily available employment and production data. Results such as
this are a rich source of data that can be used in both planning a toilet replacement
program for the CII sector and forecasting the potential water savings of such a
program.
Table 10–4 Metropolitan Water District Toilet Inventory
gpf Estimate gpf Observed Range Number Percent
1.6 Up to 2 2,347 6.1

3.5 2.0 to 4.0 21,329 55.4
5 Over 4.0 14,840 38.5
Total 38,517 100
Source: California Urban Water Conservation Council 2001.
Urinals. Older commercial, industrial, and governmental buildings often

have a substantial potential for water conservation by retrofitting older 2- to 5-gpf
urinals with new valves providing 1 gpf or less. Older urinals could also be converted
to waterless. And if new fixtures are required, waterless urinals have received good
reviews from users and are becoming more common. Hills and Birks (2004) report
Table 10–5 Estimated Savings per Toilet
90% Confidence
Market Segment Estimated Savings (gpd) Interval (gpd)
Wholesale 57 19–94
Food store 48 37–59
Restaurant 47 36–58
Retail 37 33–42
Automotive 36 22–50
Multiple use 29 14–45
Religious 28 20–37
Manufacturing 23 15–32
Health care 21 13–28
Office 20 17–23
Miscellaneous 17 11–23
Hotel/motel 16 11–20
that when waterless urinals were installed in the Millennium Dome in London, the
expected water savings was achieved and the cleaning staff found the new urinals
easier to maintain and keep clean. One reason may be that with waterless urinals,
nonliquid waste such as tissues, paper, cigarette butts, and so on are not flushed into
the drain, where they often cause clogging. The savings in water and maintenance cost
make this a good investment where moderate to heavy usage is experienced.
When the stadium at the University of Arizona in Tucson was expanded and
fitted with conventional urinals, the sewer pipes could not handle the periodic peak
loads from the bathrooms. Installing waterless urinals solved the problem, saving the
university a huge plumbing bill as well as reducing water use.
Low-Flow Showers. Showers are the third largest indoor water use in the
United States. Many conservation programs have offered low-flow showerheads that
deliver no more than 2.5 gpm. Mayer and colleagues (1999) shows, however, the
median shower flow rate of surveyed homes was 2.02 gpm and the average flow was
2.2 gpm. More than 70 percent of recorded showers used less than the 2.5-gpm
standard, even though only 50.6 percent of those surveyed reported having low-flow
showerheads. One explanation for this is that many showers have controls for both
temperature and volume, and many people do not choose a larger volume. Some
homes and other facilities, however, such as athletic clubs, have only temperature
controls on their showers. For those customers with higher flow rates, the savings in
water heating costs could be far more significant than the cost of saved water. An
athletic club in Massachusetts replaced 35 showerheads and saved 328,000 gal of water
and $3,330 per year for water, sewer, and energy, achieving a payback of its conversion
cost within a month (Vickers 2001).
Longer showers offset part of the expected water savings from low-flow
showers, as shown in Table 10-6. Even with this offset, low-flow fixtures offer a
substantial benefit in water and energy conservation. This report is from micrometered
water use, and does not show the type of showerhead in use. The low cost of
showerheads makes their replacement a good investment.
Table 10–6 Daily Shower Use
Shower Use
Sample Shower Use per per capita
Size Household (gpd) per day (gcpd) Shower Duration
Shower
Flow
Category Number Mean Std. Dev. Mean Std. Dev. Mean Std. Dev.
Low 177 20.7 14.2 8.8 6.6 8.5 3.4

Mixed 702 32.2 20.9 11.8 8.2 8.0 3.8
High 289 34.8 24.7 13.3 10.3 6.8 3.1
Source: Mayer 1999; Table 5.21, p. 134.
Clothes Washers. Water and energy-efficient clothes washers provide another

conservation option. Typical clothes washers in the United States have become more
water efficient over the years, with average use for a full load dropping from 56 gal
before 1980 to 51 in the 1980s to 39 to 43 gal in the 1990s. Horizontal axis washers
using 27 gal or less have become more common since 2000, but their generally higher
price has limited their acceptance. Out of 22 front loaders reviewed by Consumer
Reports (2007), their “best buy” model was priced at $850, compared to their “best
buy” top loader priced at $380. Some users also object to having the washing machine
door close to the floor, but this objection can be offset by installing the machine on a
platform. For other consumers, the option of stacking a clothes dryer on top of the
washer is attractive. An additional selling point for front-loading washers is that they
produce less wear and tear on clothing.
The price of front-loading machines in 2006 ranged from $620 to $1,600, with
a median price of $1,025. In contrast, conventional top loaders ranged from $330 to
$500, with a median of $435. Unconventional top loaders were priced from $800 to
$1,000 but none received high rankings for water conservation (Consumer Reports
2007). All but two3 reviewed front loaders were rated excellent or very good for water
conservation, but top loaders were all rated good, fair, or poor. Because a typical
washer life is 10–14 years, most replacement opportunities within conservation
programs would replace conventional washers using about 40 gal per wash with front
3. The remaining two were rated good.

loaders using 27, a savings of 13 gal per wash, or 33 percent. New US energy standards
that went into effect in January 2007 may make washer replacement more attractive.
The incentive required to induce many customers to make the switch to front-
loading washers is unknown. The large price differential between conventional
washers and front-loading washers suggests that consumers may require substantial
incentives to make this substitution. Several utilities have offered rebates or even free
replacement of higher-use washers. For equipment installed in high-use situations
such as hotels and laundromats, the savings in water heating combined with water-use
reductions would make investments in front loading machines particularly attractive.4
Faucets. A typical home uses between 8.7 and 12.3 gpcd from all its indoor
faucets (Mayer et al. 1999). Maximum faucet flow rates in the United States ranged
from 3 to 7 gpm before 1980, decreasing to 2.75 between 1980 and 1994. The federal
government mandated a maximum flow rate of 2.5 gpm at 80 lb pressure or 2.2 gpm
at 60 lb in 1994, and since then at least some faucets have become available with a flow
rate of 1.5 gpm (Vickers 2001). Except when filling the kitchen sink or a kettle,
consumers typically use less than the maximum flow rate. Because faucets have a life
expectancy of 15 years or more, retrofitting faucets with aerators may be an effective,
low-cost way to reduce faucet water use. Repairing faucet leaks is also effective,
because washers and other parts often wear out or become damaged. Savings from
retrofitting a kitchen faucet will be greater in homes doing dishes by hand or where
people prerinse dishes before loading the dishwasher.
An experiment was conducted at the Millennium Dome in London to
determine the water savings and acceptance of water-saving bathroom fixtures (Hills
and Birks 2004). Three types of water-saving faucets were installed, as shown in
Figure 10-4. Surprisingly, the conventional handle faucets used less water than either
of the automatic shutoff types. Users reported that the push top faucets were easiest to
use and the infrared were hardest to use. Many individuals washing their hands with
soap triggered the automatic faucets two or more times. Some of the faucets were
installed with signs explaining how they worked and why they were installed. Users
reported a much more positive attitude toward these fixtures and were more satisfied
with their operation when signs were present.
The Carney Hospital in Dorchester, Mass., placed 1.5-gpm flow-control devices
on faucets in patient and examining rooms that had 5-gpm faucets. With an estimated
faucet use of 25 minutes per day, this saved an estimated 88 gpd per faucet. The
program cost $12 for each sink and saved $280 per year per sink in water and heating
costs (Vickers 2001, p. 110).
Faucet aerators and faucet repair/replacement may be where publicity and water
audits could have a significant effect on water use at low cost to the utility. As with
other water-saving fixtures, higher marginal prices for water make these investments
more attractive.
4. A list of rebate approved high-efficiency washers can be found at sdcwa.org/manage/pdf/

conservation.
1.5
Liters per Use

1
0.5
0
Infared Push Top Handle
Type of Faucet
Source: Hills and Birks 2004.
Figure 10–4 Water Use by Washroom Faucets
Offsetting Behavior. The results reported previously signal the clear poten-
tial for substantial reductions in indoor water use through new equipment. The net
effect of installing these items depends, however, on the presence or absence of
offsetting behavior. If people install water-saving toilets and other fixtures to use more
water for a Jacuzzi, decorative fountain, or grass, the actual water savings may be
small or even negative. Context is all important. In New York City, such offsetting
behavior is highly unlikely—most residents have no place to install such amenities
even if they wanted them. In the Phoenix metropolitan area, however, with large
residential lots and a history of unrestricted water use, low marginal water prices, and
green lawns, the situation may be different. Campbell and colleagues (2004) report on
a detailed study that shows that offsetting behavior may have negated any benefits to
the utility of distributing water-saving kits. They concluded that offsetting behavior
was absent when personal contact accompanied the kits, suggesting that if water
consumers believe water conservation is worthwhile, they are unlikely to engage in
offsetting behavior. They also found that when homes had high-efficiency toilets and
residents were unaware of this feature, there was no offsetting behavior. On the other
hand, when homeowners retrofit these toilets, they often engaged in offsetting
behavior.
Landscape Conversion. The potential for reductions in exterior water use

varies widely, depending on local climate, income, housing and commercial lot
characteristics, and community expectations about vegetation and yard care. For the
long term, municipal governments could place legal limits on the use and amount of
turf, or impose other exterior landscaping requirements. The Board of Supervisors of
Pima County, Arizona, recently considered an ordinance to prohibit decorative turf
for new commercial and industrial properties.
Table 10-2 shows a wide variation in outdoor water use for surveyed homes,
even among cities in the same metropolitan area, such as Phoenix, Tempe, and
Scottsdale. These data are highly accurate in dividing indoor and outdoor water use
because they are based on actual logging rather than a rule of thumb subtracting
average winter month use from monthly water use. Especially in warmer climates,
some exterior water use occurs year-round for car washing, pool maintenance, and
winter grass irrigation. Thus a utility must evaluate the potential for exterior water
conservation by community before launching a conservation program.
Some cities in arid regions have had success with programs that provide
incentives for customers to convert high-water-use landscapes such as grass to low-
water-use forms such as xeriscape. Water-wise landscaping, which includes Xeriscape
as a widely publicized subcategory, incorporates seven principles (see, for example,
Vickers 2006 and Sunset, 2005):
• Sound landscape planning and design
• Limitation of turf to functional areas such as playgrounds
• Soil analysis and use of soil amendments
• Use of water-efficient plants adapted to a site
• Efficient irrigation with drip systems and water harvesting
• Mulches
• Appropriate landscape maintenance
When landscapes are converted, researchers report water savings ranging from
20 to 53 percent of total usage in single-family homes. A research project in Las Vegas,
Nevada (Sovocool et al. 2006), recruited volunteers to convert their landscaping or to
be monitored without conversion, and compared these monitored groups with
unmonitored properties with similar characteristics. Those converting to xeriscape
were given an incentive of $0.45/ft2 for up to 2,000 ft2, with many homeowners going
beyond the incentive to convert more than 2,000 ft2. To be eligible for the incentive,
xeriscape plans had to include a canopy cover of at least 50 percent at maturity. In the
study group, 370 properties converted at least 500 ft2 of grass. The stay-in-turf group
consisted of 253 properties with an average of 2,462 ft2 of grass. Both groups initially
irrigated their grass using in-ground irrigation systems with controllers.
Because Las Vegas is extremely dry (receiving only 4.5 in. of rain per year) and
hot, virtually all vegetation requires irrigation year round. Although the monitored
turf properties applied an average of 117.2 in. of water per year, the xeriscape
properties used only 27.6. Figure 10-5 shows the monthly water application rates for
properties with turf and xeriscape. The striking features of these data are the reduced
peak-to-average ratio and the overall reduction in water-use intensity.
In mid-2000, Las Vegas made a general offer of $0.40/ft2, and experienced a
significant rate of conversions, as shown in Figure 10-6, especially among commercial
properties. In March 2003, in response to a drought on the Colorado River, several
conservation programs were initiated, including an increase in the incentive to $1.00/ft2.
Figure 10-6 shows the dramatic response. Analysts can include water savings such as
this directly in the forecast. These experiences also signal that the larger incentive of
$1.00/ft2 of converted turf is clearly above the tipping point where many water
customers find the incentive worthwhile. It seems likely that these conversion rates
14
12
Gallons/Ft2/Month
10
8 Turf
6 Xeriscape
4
2
0
t
ay
ar
n
p
g
r
v
b
Ju
Oc
Ap
No
Ju
Au
Ja
Se
Fe
M
Source: Adapted from Sovocool et al. 2006.
Figure 10–5 Water Use by Turf and Xeriscape
40
Converted Area Million Ft2
35
30
25
20
15
10
5
0
2000 2001 2002 2003 2004
Year
Source: Adapted from Sovocool et al. 2006.
Figure 10–6 Area Converted From Turf to Xeriscape, Las Vegas, Nev.
could be applied to other cities, but the savings in water use would need to be scaled
down to match local conditions.
For Las Vegas, the measured water savings from this conversion was 89.6 inches
per year, which converts to 7.47 ft3 per square foot of converted turf per year. At
$3.00/100 ft3 for water, this adds up a savings of $22.40/ft2 for water alone. In
addition, property owners reported reduced expenditures for landscape maintenance.
This is obviously a very cost-effective incentive for both the water utility and
customers.
Automatic Timers on Watering Systems. Although one would intuitively
predict installing timers on watering systems would lower water use, because the
timers reliably turn off water at the end of a cycle whereas with manual controls it
might be left on for excessively long periods from time to time, research has shown
the opposite. In a study of water use by apartments in Tucson, Arizona, Billings and
Longstreth (1989) found that water use was higher when these controls were present,
after accounting for other variables. Vickers (2001, p. 204) reports that automatic
sprinkler systems often deliver up to double the amount of water actually needed on
a landscape because people overestimate the amount of water required and don’t take
the time and effort to carefully adjust their timers. Chestnutt and McSpadden (1991)
found that homes in Los Angeles with automatic sprinkler systems used 11 percent
more water than those with manual systems. Likewise, Renwick and Archibald (1998)
found that lawns with sprinkler systems used 9 percent more water than those
without.
Automatic control systems can save large amounts of water. Vickers (2001,
p. 211) reports that four Southern California commercial sites that switched from
fixed schedules to evapotranspiration (ET)-based schedules reduced their water
application from 68 to 44 in. per year, when the theoretical ET requirement was 45 in.
Water savings averaged 1,550 to 4,600 gal per acre per day, with a monetary savings of
$1,500 to $4,600 per acre per year. For a commercial or governmental entity with
numerous irrigated areas, a central computer-controlled irrigation scheduling system
can significantly increase water-use efficiency and provide a very short payback
period.
These results suggest the potential benefits of finding ways to make it easier for
customers to efficiently program their irrigation timers, either by better design or
some automated adjustment system.
Rainwater Harvesting and On-Site Graywater Reuse. A tremendous po-

tential exists for reducing potable water use on landscaping by effective rainwater
harvesting. This can be totally passive, using contours, berms, and basins to collect
rainwater where plants can use it. Careful design can meet a large proportion of plant
requirements when combined with climate appropriate landscape material. Home-
owners can make more extensive use of rainwater by collecting roof runoff in cisterns,
and using this water over time as needed for irrigation.
Domestic graywater can often be used on landscaping, but this requires careful
management, including the use of bio-degradable soaps, rotation of where this water
is used, and compliance with local and state plumbing codes. Graywater is generally
defined as water originating from showers, baths, and clothes washing. Water from
toilets is always considered blackwater, unsuitable for on-site reuse. Water from
kitchen sinks and dishwashers likewise is usually considered to be blackwater,
although a few states include it in the definition of graywater. Local building codes
often discourage the on-site use of graywater for landscape irrigation, based on public
health considerations.
A relatively inexpensive graywater system uses the water as it is generated,
avoiding the need for storage and treatment. Where permitted, regulations generally
require that graywater irrigation be below the surface to prevent standing water from
this source. Although there is a significant potential for conserving potable water by
encouraging the use of graywater, we would be reluctant to project substantial water
savings from this source. In addition to regulations that discourage on-site use of
graywater, few existing structures are plumbed in a way that would facilitate its use.
Although homes and other buildings with basements could be retrofit, such a
conversion is impractical for single-floor structures built on concrete slabs, where all
the plumbing is imbedded in or below the concrete.
Innovators such as Lancaster (2006) have demonstrated that public health and
building codes can be honored while using both rainwater and graywater for
landscape irrigation. At his private residence in Tucson, Arizona, he has made such
extensive use of graywater and rainwater that water utility repair crews have checked
his water meter for possible undermeasuring several times.
The Natural Resources Defense Council office building in Santa Monica, Calif.,
collects both rainwater and graywater, which is treated, filtered, stored, and used for
toilets and irrigation. With the use of dual-flush toilets and this system, no potable
water is used on the adjoining landscape.5
The new south campus office development of Toyota Motor Sales in Torrance,
Calif., is reported to use 94 percent less water than would be expected for a
conventional set of structures. With a 40 acre site and 624,000 ft2 of building space,
the complex saves 11 million gal of water per year through a combination of harvested
rainwater and recycled graywater used for cooling, irrigation, and flushing.6
Industrial, Commercial, and Institutional Conservation. Large volume
customers have a great potential for water savings in addition to what can be saved by
measures previously discussed. Often, even a small price increase can shift the
economics of reuse and recycling of water such that innovations, often requiring
capital expenditures, become economically feasible. In many cases, water conservation
has not appeared on the radar of management, so opportunities that would lower
costs while saving water are not used. A survey by Plosser et al (1992) and the US EPA
and California Department of Water Resources (1997) found potential savings as
shown in Figure 10-7, ranging from 9 percent at military facilities to more than
30 percent at churches and nonprofits. Additional evidence of the potential savings in
nonresidential sectors is gained from a detailed analysis of Santa Clara, Calif., large-
volume water users (PPI 2004). The survey included 11 industrial firms, 2 paperboard
plants, 10 commercial facilities, and 4 institutional customers. Those included were
among the highest volume customers of the utility. Table 10-7 gives the combined
potential water savings that these customers could obtain, the investment required,
and the payback in years for each category.
Given the wide variety of organizations, processes, and products found among
nonresidential customers, analysts cannot directly apply findings such as these to the
most likely outcomes for any specific water utility service area.
The best strategy is to carefully examine the water usage of large-volume
customers to determine what types of conservation programs may be worthwhile and
acceptable. Many conservation opportunities are projected to have payback periods of
only a few years, making them an attractive investment. For example, if an
organization estimates that a project will have a 2-year payback and that the
5. US Green Building Council Web site; www.usgbc.org.

6. Toyota Motor Sales Web site; www.toyota.com.
Churches
Recreational Facilities
Restaurants and Bars
Car Washes
Turf Irrigation
Hospitals
Schools
Food Processing
Communications
Hotels
Beverage Processing
PC Board Manufacturing
Laundries
Prisons
Military Facilities
0 5 10 15 20 25 30 35
Potential Conservation Percent
NOTES: Purple and red bars show higher estimates; purple shows lower or only estimates.
Source: Derived from Plosser et al. 1992 and USEPA and California Department
of Water Resources 1997.
Figure 10–7 Industrial Conservation Potential by Sector
investment will last for 10 years and start providing benefits within a few months of
expending funds for the project, the internal rate of return is about 60 percent. Most
firms have few investment opportunities with this high a return. If, however, through
a miscalculation the payback period turns out to be 4 years, the rate of return drops
to a still respectable 24 percent. Uncertainty about actual returns is one reason many
organizations will undertake only investments with short payback periods. Another
reason is that some organizations may not have funds available to make capital
investments. In this case, a program of loans with repayments based on realized water
bill savings may be a good strategy for the utility. This, however, places the risk that
the firm will go out of business or not realize the expected savings and never repay the
loan on the utility.
Use of Reclaimed Water. Many irrigation applications could use reclaimed

water. Several industrial applications exist, although utility officials have a much
Table 10–7 Potential Water Savings and Payback for Santa Clara, Calif.
Water Total
Savings, Annual Cost Project Payback
Recommendation 1,000 gal/yr Savings Cost, $ Years
Water/wastewater recycling 2,366,450 1,767,648 1,928,400 1.3

Conductivity controllers 3,655 13,309 48,000 4.8
Dynamic seals 42,944 250,300 100,000 0.9
Faucet aerators 361 2,689 90 <0.1
Felt washer spray valves
(replacement) 18,980 3,780 0 <0.1
Food steamers 110 585 2,000 3.4
High-efficiency ice machines 18 97 4,000 41.2
High-efficiency urinals 186 1,253 120 0.1
High-efficiency washers 4,546 38,060 188,000 4.9
Low-flow high-pressure sanitary
spray nozzles and CHP systems 8,320 47,508 50,000 1.2
Industrial paper process
upgrades 59,445 321,539 90,000 4.6
Pre-rinse spray valves 527 4,075 350 0.2
Leaky faucet repair 46 244 15 0.1
Sprinkler replacement 58 79 70 0.9
Ultra low-flow toilets 7,596 39,671 124,150 2.4
Source: Extracted from PPI 2004; Table 3.21.
harder time identifying and developing this market. The following challenges face
utility managers attempting to increase the use of reclaimed water:
• Capital and operating cost of any required additional sewerage treatment
• Capital cost of building and extending the reclaimed water delivery system
• The need to motivate potential users to switch from potable to reclaimed
water
Often the marginal cost to the utility of meeting these challenges exceeds the
delivered marginal price of potable water, so developing this program requires
substantial subsidies. This is just one more signal that potable water is typically priced
far below its full cost.
The full cost of water to society includes all of the following, but only the first
two are typically direct costs to the utility, and even capital charges are not always fully
borne by utility revenue:
1. Operating and maintenance costs

2. Capital charges
3. Opportunity cost of a potentially higher-valued use forgone
4. Economic externalities
5. Environmental externalities
Item 3 refers to the possibility that some other use of water may have a higher
value, either today or in the future. Item 4 is the possibility that water extraction by
the utility may impose costs on other individuals that are not paid for. For example, if
water extraction lowers the water table, some other well owners may experience
higher pumping costs and even dry wells. Excessive extraction may also result in
subsidence, which the utility does not necessarily pay for. Environmental externalities
are associated with public health and ecosystem maintenance.7 Although it is difficult
to estimate values for items 3 through 5, ignoring them effectively sets their value
equal to zero, which is obviously incorrect. Obtaining the political will to place a
positive value on these costs and incorporate them into water rates is even more
daunting.
BEHAVIORAL CONSERVATION PROGRAMS

Education and Publicity. Educational efforts, including publicity about the
needs and benefits of water conservation, are an integral part of an effective program.
It is problematic, however, to anticipate any particular amount of conservation from
these efforts. Some communities have reported sizable impacts from various publicity
efforts, especially when dealing with short-term emergencies, as reported in chapter
11. Many others, however, have reported little or no impact. The effects have often
proven to be very difficult to find, even with careful analysis.
There is ample evidence that conservation programs, including conservation-
related price increases, will be more readily accepted when they are actively promoted
by government and utility officials having high credibility. It is also clear that publicity
must be continual to maintain its effect, and that government entities should lead by
example. There is nothing like seeing water running down the street from a municipal
irrigation system to lead water customers to conclude that all the talk about
conservation is simply hot air and can be ignored. Utilities should also examine their
infrastructure and implement programs to minimize system unaccounted-for water
before asking customers to conserve.
Changing Indoor Habits. Encouraging customers to use water frugally by
actions such as turning off water while brushing teeth and taking shorter showers and
running clothes washers with full loads or with appropriate water level adjustments is
7. This discussion is based on work by Dziegielewski 2003 and Rogers et al. 1998.
worthwhile. Hotels can be encouraged to ask guests to use the same sheets and towels
for several days instead of washing them every day. They could even offer a small
reward of some type for saving water. Offering information about water and energy
savings may have some effect, but raising water and sewer rates is more likely to elicit
the desired changes. Publicity and higher rates reinforce each other.
Changing Outdoor Habits. Emergency conservation programs, as discussed
in chapter 11, have had substantial success with various restrictions on exterior water
use. The challenge is to give customers a large enough incentive to increase the
efficiency of their irrigation actions. Some electric utilities have developed remote
systems that can automatically interrupt the service to contracted customers when
systems loads are excessive. ET controller technology has progressed to the level where
water utilities could offer to connect customers’ irrigation controllers to a central
computer that would adjust the duration and timing of irrigation to match seasonal
requirements as well as day-to-day weather conditions and individual vegetation
requirements.
As Vickers (2006) suggests, the irrigation and chemical industries continue to
try to convince consumers that they should have lush green lawns. Often the result is
overwatering, overfertilizing, and excessive use of weed and pest control chemicals.
These practices lead not only to wasted water and fruitless money spent for chemicals,
but also to chemical-laden runoff that pollutes water sources. The challenge, in the
face of extensive commercial advertising, is to convince consumers to both become
more efficient in their lawn-care practices and to expect less perfection from their
turf.
Experiments with Bermuda grass have shown that the more it is watered, the
more water it uses. Water use could be cut dramatically, as it has been in drought
situations, if people would accept a different standard. Proper watering schedules also
encourage vegetation to grow deeper roots, and thus require less frequent and
probably less total irrigation.
Effect of Metering of Water Use. Clear evidence supports the idea that
installing meters and charging for water based on the volume reduces water use
significantly. A recent study of water use in the United Kingdom (Kowalski and
Marshallsay 2005) found that the 25 percent of households with metered use
accounted for only 12 percent of water deliveries whereas the remaining 75 percent
(unmetered homes) accounted for 56 percent. Metering resulted in a 36 percent drop
in water use. Linaweaver et al. 1966 reported similar results for the western United
States, where homes in metered areas used 34 percent less water than those without
meters, while peak hour use fell by a very impressive 52 percent. Morris and coauthors
(1997) reported that when metering was introduced in Las Vegas Valley, Nev.,
household use fell from 437 gal/day to 225 gal per day, a drop of 48 percent. Thus
research and practical experience verify the economic theory of demand, which states
that when users consider a good to be free, they will use additional units up to the
level where an additional unit provides no additional value to the user. For example,
when water is considered a free good, users have no incentive to fix leaky plumbing.
The research cited clearly shows that shifting from flat rates to metered consumption-
based rates will reduce water use by anywhere from 30 to 50 percent.
Effective Water Bill Presentation Increases Price Response. Consumers’
response to a price increase may be greater when water bills prominently display
marginal prices; Gaudin (2005) found that the effect of a price increase may be
25 percent greater with this type of display. A survey of 383 large water utilities found
only 17 percent prominently displayed marginal price on their water bills, and
79 percent showed only the total amount due. This statistical research found a price
elasticity of demand of –0.35 among utilities with no marginal price information of
their water bill, contrasted with a price elasticity of –0.51 among those with marginal
price prominently displayed. Thus clearly displaying the marginal price on the bill is
a potentially effective strategy to increase the conservation effectiveness of a price
increase.
For example, a certain utility’s water bill, which does show the marginal price of
water, fails to display it prominently. The relevant parts of this bill, displayed in Figure
10-8 prominently show the total amount due, and the back of the bill shows that the
customer used 23 CCFs (1 CCF = 100 ft3). Few customers have any idea what a CCF
is, and very few are likely to take the time to find out. Where the marginal price is
displayed, the volume and price are reversed, adding to the obscurity.
Front of Water Bill Back of Water Bill
Last Bill Amount $ 77.08 Monthly Serv Chrg 5.35

Payment 77.08 VOL $ 1.03 @ 15.00 15.45
Sewer 12.40 VOL $ 3.60 @ 8.00 28.80
Water 50.52
Refuse 14.00
Other Charges 4.09
Total Due $81.01 WATER USED IN CCF 23
Figure 10–8 A Water Bill That Fails to Effectively Show the Marginal Price of Water
A bill that effectively displays the marginal cost of water might appear as in
Figure 10-9.
When utility officials want to encourage water conservation, final formatting of
water bills should not be left to accountants and engineering staff, but to conservation
or customer service officials who are knowledgeable about information responses to
billing information and have daily contact with the utility’s customers.
Raising Price—Proportional Increases in all Rates. The simplest rate
change to forecast is a proportional change in all rate categories, including the fixed
charge, when no other policies or practices change. In this case, the analyst can
directly apply the estimated price and revenue elasticities of demand to water usage to
estimate both conservation and revenue effects. We assume that the rate increase
analyzed here is in addition to an adjustment for inflation in the utility service area.
For example, if a utility wishes to induce a 7 percent drop in water usage by a
proportional rate increase, and has estimated that its system-wide price elasticity of
demand is –0.33, the required rate increase is calculated as follows:
(Required Percent Change in Price) =

(Desired Percent Change in Quantity)/ ε (Eq 10-1)
where:
ε = price elasticity of demand.
In this case, the required price increase is –7.0%/–0.33 = 21.2%. The
accompanying increase in revenue is given by:
(Percent Change in Revenue) = (1 + ε ) × (Percent Change in P)
The resulting revenue increase is therefore (1 – 0.33) × 21.2% = 0.67 × 21.2% =

14.2%. This is perfect if the utility needs a revenue increase of 14.2 percent.
Total Water Used This Month: 23 units. One unit = 748 gallons.
You could reduce your bill by $3.60 for each unit of water saved.
Total Bill Due $81.08 Water Charges

Water 50.52 Monthly Service 5.35
Sewer 12.40 15 Units @ $1.03 15.45
Refuse 14.00 8 Units @ $3.60 28.80
Other Charges 4.09 Other Charges .92
Last Bill Amount 77.08 Total 50.52
Payment Received 77.08
Figure 10–9 A Water Bill That Effectively Displays the Marginal Cost of Water
As with any forecast, it is unlikely that this utility would experience these exact
results. Suppose, for example, that the standard error of the elasticity estimate itself is
0.09 and that the elasticity estimate is normally distributed. Table 10-8 shows the
probabilities of the likely range of values for elasticity and the resulting changes in
water use and revenue. The chance of the realized values falling outside a given
confidence interval decreases as the size of the confidence interval widens. The
confidence intervals shown range from 70 percent, so there is a 30 percent chance that
the realized value will fall outside the given range, to 99 percent, which leaves only a
1 percent chance of excluding the true values. The first row shows that, in this
example, there is a 70 percent chance that the true value of the elasticity of demand is
between –0.237 and –0.423, with the drop in water sales falling between 5.0 and 9.0
percent, and revenue increases between 16.2 and 12.2 percent. At the other extreme,
Table 10–8 Confidence Intervals for Price Effects on Water Use and Revenue
Price
Probability Elasticity of % Price % Quantity % Revenue
(%) Demand Increase Decrease Increase
70 –0.33 ± 0.093 21.2 5.0–9.0 16.2–12.2

80 –0.33 ± 0.115 21.2 4.6–9.4 16.6–11.8
90 –0.33 ± 0.148 21.2 3.9–10.1 17.3–11.1
95 –0.33 ± 0.176 21.2 3.3–10.7 17.9–10.5
99 –0.33 ± 0.232 21.2 2.1–11.9 19.1–9.3
there is only a 1 percent chance that these values will be outside the ranges shown in
the last row of the table.
Aside from these statistical considerations, factors outside the utility’s control,
such as rainfall and average high temperatures, may cause results different than those
predicted. This large a change in price, however, is very unlikely to be overwhelmed by
variations in the weather. Smaller changes, however, are often overshadowed by
unusually wet and cool or dry and hot conditions, especially during the irrigation
season.
We recognize that many utilities in a situation requiring reduced water use of
this size would not be in a position to raise rates by this amount, because of either
regulatory restraints in the case of an investor-owned utility or political consider-
ations in the case of a government-run utility. The utility may not need or be allowed
to increase revenue by this large a percentage. A recent survey of utility officials,
however, found that financing the replacement of aging infrastructure is a top concern
(Runge and Mann 2006). Many water utilities have postponed replacing aging
infrastructure to such an extent that, in the minds of many officials, the aging
infrastructure has become the “failing infrastructure.” This concern, combined with
the observations that consumers in the United States may be spending as much on
bottled water as is spent for utility service, the typically small percentage of household
income spent for utility-supplied water, and the proven effectiveness of higher water
rates in inducing water conservation, suggest a powerful rationale for higher marginal
prices for water. If raising marginal prices creates excess revenues, the excess could be
deposited in a sinking fund dedicated to infrastructure replacement or returned to
customers in a way that does not lower water bills.
Increasing Marginal Price and Decreasing the Monthly Service Charge.

This change in the rate schedule is likely to somewhat reduce water use,
especially if customers are made aware that they now have greater control over the size
of their water bills. The size of the effect will depend on the size of the decrease in the
fixed monthly service charge and the way in which the marginal price is increased as
well as the relevant price elasticity of demand. Such a shift will also make utility
revenue less stable, because it will depend more heavily on water volume and less on
the number of customers.
For example, suppose a utility serving primarily residential customers using an
average of 20 units of water per month at $1.50 per unit8 decides to reduce its
$8.00 per month service charge to $5.00. A first-cut computation, which assumes no
change in water use, is that an increase in marginal price of $3.00/20 = $0.15, or 10
percent ($0.15/$1.50), would be required to keep revenue constant at $38.00 per
customer. Thus the new first-cut marginal price is $1.65 per unit.
A second-cut computation includes the effect of the increase in marginal price
on water sales. Assume a price elasticity of demand of –0.28, which makes the price
elasticity of revenue equal to 0.72, and for simplicity assume no customer response to
the decrease in the monthly service charge.9 Equation 9-5 shows the required price
increase to replace the lost revenue of $3.00 per customer per month from cutting the
service charge. The needed percentage increase in marginal price equals the required
percentage increase in revenue divided by the price elasticity of revenue. Thus, the
required price increase is 13.89% = 10%/0.72. The new marginal price is $1.71 and
the new quantity is computed using Equation 9-3: the percentage change in quantity
equals the percentage change in price times the price elasticity of demand. Thus, the
change in quantity = 3.889% = 13.89% × –0.28, which is –0.778 units, making the
new quantity 19.222 units. Average revenue per customer is now equal to
$5.00 + $1.71 × 19.222 = $37.87.10
Is this second-cut computation realistic? If customers are only aware of their
total water bill, changing the way it is computed will have no effect. For this change in
the rate schedule, those using less than the average amount of water would receive a
reduced water bill, which might encourage them to use more water, but those using
more than the average would receive an increased water bill, and many will respond
by using less. In the preceding example, customers continuing to use only 5 units per
month would see a drop in their water bill from $8.00 + $1.50 × 5 = $15.50 to
$5 + $1.71 × 5 = $13.55. A high volume user at 40 units per month would find a bill
increase from $8.00 + $1.50 × 40 = $68.00 to $5.00 + $1.71 × 40 = $73.40. Carefully
examining the use distribution of customers might offer some clues on the most likely
outcome of reducing the monthly service charge and increasing the uniform marginal
price.
Changing to Increasing Block Rates. Increasing block rates and increas-

ing step rates are often touted as effective conservation tools for a water utility. The
structure of these rates is more fully discussed in chapter 9. To achieve revenue
neutrality when shifting from decreasing block rates or a single uniform rate to
8. This example assumes the rate schedule uses a uniform marginal price.
9. Economic theory states that the effect of a change in the service charge should be the same as a
change in income. In effect, only the marginal price matters.
10. The new revenue is not exactly $38 because elasticities are perfectly accurate only for very small
changes.
increasing block rates, low-volume customers would experience a rate decrease and
high-volume customers would be subjected to an increase. Table 10-9 provides a
hypothetical example of how this shift might work. The example assumes that all
customers have the same price elasticity of demand. At the initial average quantities
Table 10–9 Shift From Constant to Increasing Block Rates, Uniform Elasticity
Customer Group Low Use Average Use High Use Total
Usage range 1–10 11–20 21 and over

Price elasticity –0.40 –0.40 –0.40
Number of customers 30 50 20 100
Uniform rate, $ 2.80 2.80 2.80
Average use 5 15 27 1,440
Revenue, $ 420 2,100 1,512 4,032
Block rate prices (1), $ 1.00* 2.80 3.30
Revenue (initial Q), $ 150 2,100 1,782 4,032
Expected new Q 7.34* 15 25.07 1,472
Expected revenue, $ 220 2,100 1,655 3,975
Block rate prices (2), $ 1.40 2.80 3.39
Expected new Q 6.00 15 24.80 1426
Expected revenue, $ 252 2,100 1,681 4,033
* See Table 10-10 for computations required to obtain these values.
for each usage category, total revenue (for 100 customers) is $4,032 and total usage is
1,440 units per month. The shift to increasing block rates is designed to produce the
same amount of revenue, assuming no change in usage. When elasticities are taken
into account, however, usage in the first category rises and that in the third falls. The
example assumed, for simplicity, that the rates in the middle category do not change.
The first category, involving a huge drop in price, is computed by 10 percent
increments, as shown in Table 10-10. An example in chapter 9 fully explains this
procedure. Because the initial forecast of usage after the rate shift was for no change,
a further computation was required to regain the initial level of revenue. Although
many rate combinations would achieve the desired result, the illustration includes a
small increase in both the low- and high-use categories from their initially changed
amounts. With these further changes, total revenue returns to its original level and
water usage is slightly below the initial level.
Many possible initial assumptions could be incorporated into the example, such
as a lower elasticity for low-use customers. Some empirical evidence points to
Table 10–10 Iterative Computation of Large Price Effects on Quantity
Iteration Change in Price (%) Price ($) Quantity
0 Initial Values 2.80 5.00

1 10 2.52 5.20
2 10 2.27 5.41
3 10 2.04 5.62
4 10 1.84 5.85
5 10 1.65 6.08
6 10 1.49 6.33
7 10 1.34 6.58
8 10 1.21 6.84
9 10 1.08 7.12
10 8 1.00 7.34
differences in elasticity among low-, middle-, and high-volume water customers. As

pointed out in chapter 9, we would expect the elasticity to be higher among those
using higher quantities, simply because they have most likely expanded their use to
include activities with less value. For example, some people routinely wash down their
sidewalks, patios, and driveways when they could be swept or blown clean. To
illustrate this point, the example is reworked with a more realistic elasticity of demand
of –0.2 for the low-use customers and –0.6 for the high-use customers. The previous
elasticity of –0.4 is retained for the larger middle category, which is unaffected by the
shift in rate schedules. Using the same procedures, the results are shown in Table 10-
11. For the low-use category, price drops from $2.80 to $1.65 and average usage
increases from 5 to 6 units per month. Price for the high-use category increases from
$2.80 to $3.35, causing average use to fall from 27 to 23 units. Prices were chosen to
keep total utility revenue constant. With these differences in price elasticity between
high-use and low-use customers, total water deliveries dropped from 1,440 to 1,390,
which is a 3.5 percent reduction. This analysis could be further refined by looking at
the distribution of customers within the first and third block of the new rate schedule
to account for some customers who were initially in the first block moving up to the
second block, and some who were in the third block dropping back to the second
block. These examples illustrate the challenge of making good forecasts of the effect of
a change from declining or constant commodity charges to an increasing block rate
schedule. After making such a change, the utility would benefit from carefully
tracking the response of customers by initial use rates, to observe and react to these
changes if that becomes necessary.
Table 10–11 Shift From Uniform to Increasing Block Rates: Differentiated Elasticities
Usage Range 1–10 11–20 21 and Over Totals
Price elasticity –0.2 –0.4 –0.6

Number of customers 30 50 20 100
Uniform rate, $ 2.80 2.80 2.80
Average use 5 15 27 1,440
Revenue, $ 420 2,100 1,512 4,032
New Q forecast 6.09 15.00 24.11 1,415
Revenue forecast, $ 183 2,100 1,591 3,874
New Q forecast 6 15 23.01 1390
Revenue forecast, $ 297 2,100 1,634 4,031
INCORPORATING CONSERVATION INTO THE FORECAST

The basic method of including water conservation and efficiency savings in demand
forecasts is to deduct the estimated savings associated with each measure or program
from system water demand. In a regression-estimated structural/causal model, the
conservation measures are represented by independent variables and their effects are
directly incorporated into the results. Savings estimates can be obtained from
engineering end-use studies or local statistical analysis of water conservation and
efficiency programs. In some instances, total savings may be less or greater than the
sum of the individual water savings associated with each conservation measure,
practice, or program implemented by the water utility. One task in integrating
conservation savings estimates into forecasts, therefore, is to make reasonable
allowance for overlapping and complementary aspects of water efficiency programs.
Estimating Water-Savings Potential. Table 10-1 presented estimated water

savings for a variety of conservation measures. Many of these water-savings estimates
rely on the technical characteristics of the devices or programs and anticipated
customer participation rates. These estimates support what is often called an
engineering, or end-use, approach to estimating water savings from a conservation
program. Total water savings are represented in this approach as a product of several
factors, including (1) the water saving per use of a device, appliance, or fixture; (2) the
frequency of use of the device, appliance, or fixture per time period; (3) the potential
market for acceptance of measures in terms of the number of water customers; and
(4) the fraction of water customers who are likely to accept the measure and install the
device in each time period to be analyzed.
Thus, end-use estimates, although founded on technical characteristics of
water-using devices, require information about how people behave in connection
with this equipment. Also required is a projection of how the equipment will be
accepted in the service area of the water utility. Finally, the estimated savings depend
on customers refraining from offsetting changes in behavior, such as using the water
saved by a low-flush toilet to irrigate additional flowers or trees.
Water Conservation Goals. When analysts estimate water conservation
results before the programs are implemented, actual results may vary significantly
from what was expected. A specific long-term conservation goal can be achieved,
however, when utility officials commit themselves to that goal and retain the flexibility
to implement a variety of programs as needed to achieve it. To ensure success, a
system for monitoring progress should be incorporated into the conservation
program.
Including Conservation in a Wholesale Demand Forecast. Lev in and
coauthors (2006) report on a project to incorporate conservation savings into a long-
term forecast of wholesale and retail water demand by the San Francisco Public
Utilities Commission (SFPUC). SFPUC provides water either directly or through its
wholesale customers to about 2.4 million people. The modeling approach used the
following steps.
1. Develop baseline water demand from current water use broken down by
customer class into end uses such as toilet flushing, showers, and
irrigation.
2. Forecast the number of accounts by account type, using published
population and employment projections by area.
3. Research current and past conservation activities to determine their effects
on baseline water demand.
4. Apply the end-use model to forecast water demand, including the effects of
plumbing codes and past conservation efforts.
5. Incorporate conservation practices into the end-use model demand
forecasts to assess the benefits in saved water and costs of possible future
conservation measures.
All wholesale customers of SFPUC are committed, either directly or indirectly
through the Bay Area Water Supply and Conservation Agency (BAWSCA), to
conservation programs incorporating best management practices (BMPs). These
include
• Interior and exterior water audits and incentives for residential customers
• Residential plumbing retrofits and toilet rebates
• Horizontal-axis washing machine rebates

• System water audits, leak detection and repair
• Metering and commodity charges, including conservation pricing
for all users
• Conversion programs for large landscapes
• Public information and school education programs
• Commercial, industrial, and institutional water conservation
• Technical and financial assistance to wholesale customers
• Employment of a conservation coordinator
• Water waste prohibition
The resulting baseline forecast, including conservation programs already in
place, estimated an overall increase of 19 percent in water demands between 2000 and
2030, matching the anticipated 19 percent increase in area population, but falling far
below the anticipated 31 percent increase in employment. The model predicts that
residential water use will fall from 62 percent of system deliveries in 2001 to 58
percent by 2030, with other uses picking up the difference. Plumbing codes already in
place are projected to depress demand by 7.8 percent between 2000 and 2030
compared to what it would otherwise have been.
The forecast process also involved evaluating 75 potential conservation mea-
sures for benefits and costs to both customers and utilities. Utility benefits in reduced
need for additional supplies were evaluated at $3.033/1,000 gal. Consumer benefits
included reduced energy for water heating, which is often larger than any savings in
water bills. Thirty-two of the programs were found to be cost-effective. SFPUC and
each of the wholesale customers then selected conservation measures to be imple-
mented during the forecast period. These were developed into conservation programs
containing between 5 and 15 measures for each utility. Table 10-12 summarizes the
forecast of potential water conservation under each of three programs. All three are
highly worthwhile for the utilities involved, having benefit-cost ratios of 1.95, 2.35,
and 2.50 for programs A, B, and C respectively. Although program C is projected to
Table 10–12 San Francisco Wholesale Demand Forecast and Conservation Savings
Percent of
Total Water Use New Conservation, 2030 Added Use
Base Year Forecast

2001 for 2030 Added Use Plan A Plan B Plan C Plan A Plan B Plan C
272 324 52 7.65 14.53 19.59 15 28 38
Source: Levin et al. 2006.

save 38 percent of new water demands, this is only 6.05 percent of the baseline
forecast. The cost to the utilities of saving water ranges from $0.69 to $0.86/1,000 gal,
far below their expected cost of augmenting present supplies of $3.033/ 1,000 gal in
2015.
MANAGEMENT OF CONSERVATION PROGRAMS

Keyes and colleagues (2004) compared selected water districts to identify organiza-
tional attributes of effective demand-management programs. Their principal conclu-
sion was that successful programs shared all seven of the following characteristics:
1. Political leadership
2. Stakeholder involvement in planning and implementation
3. A detailed policy outlining goals and conservation measures
4. Detailed water-use data, demand forecasting, and monitoring
5. Stable funding sources for water-conservation measures
6. Sufficient staff and technical assistance to implement the program
7. Broad-based education and outreach
This section deals primarily with the fourth characteristic, with limited
attention to the third, fifth, and seventh.
Estimating Conservation Effects With Local Studies. Variability in the
reported effects of conservation programs suggests the need for local statistical study
of water conservation and efficiency programs, and continued monitoring of their
effectiveness.11 The simplest approach compares aggregate weather-normalized water
use before and after implementation of a conservation program. For more detailed
information, the response of a sample of individual customers to a conservation
program can be evaluated using control group comparisons or a regression-estimated
demand model. To ensure accurate monitoring of a conservation plan, all three
methods mentioned require careful planning before implementing the conservation
program, and implementation of a tracking model starting when the program is
enacted, continuing for several years after its completion.
Aggregate Before-and-After Analysis. When one or more conservation
programs are implemented during a relatively short period of time, the analyst can
evaluate their impact with an aggregate water-demand model. Water use before the
program can be normalized for weather and projected into the future to be compared
with actual water use after the program. This works effectively for relatively short time
intervals, because the analyst can ignore changes in other longer-term factors such as
11. For further discussion of low-cost methods, see Pekelney and Mitchell 1996.
income, housing patterns, and commercial activity. If some nonconservation factor is

known to have changed, it should be built into the model. Chapter 6 presents
methods of weather normalization.
Normalized water-use data from before and after a conservation program can
be used to determine the size of any resulting change in water use and the chance that
this change did not result from the program but from random chance. The
significance of a measured change in water use can be computed using a t-statistic.
Suppose that water use is measured and normalized for seasonality, weather,
and trend (if any) during 12 months before and another 12 months after a new
conservation program. Assume that normalized water use averages 4,500 million gal
per day (mgd) before and 3,900 mgd afterward. From these data, the standard
deviation before (S1) is computed as 450 mgd and afterward (S2) at 520 mgd. With 11
(n-1) degrees of freedom, the critical t-statistic at the 5 percent level for a one-tailed
test is 1.796. The test statistic for the change in water use is
2 2 2 2
S S 450 - + ---------
520 - = 356.5
t ----1- + ----2- = 1.796 --------- (Eq 10-2)
n1 n2 12 12
Because mean normalized water use changed by 600 mgd, the change in water
use is statistically significant at the 5 percent level. This means there is less than a 1 in
20 chance that the observed drop in water use was caused by random change rather
than the program. In this example, there is actually less than a 1 in 100 chance that the
drop in water use was due to random chance, because the critical t-statistic at
1 percent is 2.718.
This aggregate before-and-after method of evaluating a conservation program is
the simplest and least expensive of the three approaches discussed. It is especially
valuable for assessing the overall effect of a conservation program, but provides no
information on individual responses to the program. Detailed data on individual
customers, which would be collected for the other approaches, may be valuable in
designing program changes.
Control Group Comparisons. Control group comparisons involve a “with
and without” analysis. One group of water customers, the treatment group, is selected
to use a water-conservation device, or otherwise become actively included in a specific
water-efficiency program. A second group, called the control group, is excluded from
these initiatives. Differences in the water use of the treatment and control groups are
measured before and after participation in the program. Using random sampling to
select the comparison groups creates a basis for standard tests to determine (1)
whether observed differences in the water consumption of the treatment and control
groups are caused by the program under study or (2) whether such differences could
be explained as chance variations.
Suppose that a 2,000-customer treatment group averaged 60,000 gal/household
before treatment and 55,750 during the study period. Meanwhile, the 2,000-customer
control group average was 56,000 gal/household before and 52,200 during the
experiment. Thus, water use fell by an average of 4,250 gal in the treatment group, but
by only 3,800 gal in the control group. Is the 450-gal difference in these changes
statistically significant or can it be explained as a chance variation?
The statistical test is based on the t-statistic. The test accepts or rejects the null
hypothesis, which is that the observed difference in the group averages is due to
chance. A one-tailed test is appropriate in this context, because we are interested in
whether or not the difference represents a true water savings; i.e., water savings
generated by the treatment are significantly greater than 0. The critical t value for the
one-tailed test, taking a 5 percent chance of rejecting the null hypotheses when it is
true, is 1.645. The test relies on the standard deviations estimated from the sample
data. For this example, assume the standard deviations of water-use reduction for the
treatment and control groups, each of which has 2,000 members, are respectively
2,800 and 2,100 gal. The test statistic is:
2 2 2 2
S S 2, 800 - + --------------
2, 100 - = 128.74
t = ----1- + ----2- = 1.645 -------------- (Eq 10-3)
n1 n2 2, 000 2, 000
Because the difference in the water-use reduction between the treatment and
control group is 450 gal, which is greater than the critical test statistic of 128.74 gal,
we can reject the null hypothesis of no difference and accept the alternative
hypothesis, that the treatment did result in a statistically significant reduction in water
use.
Setting Up a Controlled Comparison. Suppose a water utility with 9,000

residential customers is considering plans to distribute and install low-flow shower-
heads at no cost to customers and wants to run a pilot study to determine the likely
water savings. Suppose also, as in the summer residential water-use distribution
presented in chapter 4, the population standard deviation of summer residential water
use is approximately 5,000 gal for the July–August billing period and the average
customer water use is 8,000 gal. End-use estimates suggest that low-flow showerheads
can save 300 gal in a typical household over a two-month period. Assuming all
treatment group households elect to install the showerheads, and that none of the
sample households are already using these devices, what sample size is necessary to
estimate, at a 95 percent confidence level, the expected decrease in average water use?
The key equation involves the tolerance, which is 300 gal; the standard
deviation of water use, which is 5,000; and a t-statistic of 1.645, which was obtained
from a table for the normal distribution for a 95 percent one-tailed confidence
interval. The required sample size is
tS 2 ( 1.645 ) ( 5000 ) 2
N = ⎛ -----------------⎞ = ⎛ -------------------------------⎞ = 752 (Eq 10-4)
⎝ tolerance⎠ ⎝ 300 ⎠
This sample size of 752 can be slightly reduced with the “finite sample
correction” (FSC), which is
population – sample size 9, 000 – 752

FSC = ------------------------------------------------ = -------------------------- = 0.957 (Eq 10-5)
population – 1 9, 000 – 1
This results in a slight reduction of the sample size to 0.957 × 752 = 720 for the
treatment group receiving the device and another 720 in the control group, to obtain
a controlled comparison of average water use with and without the device.
Once the sample size is determined, the analyst selects a random sample of
customers and randomly assigns them to the treatment and control groups. The
conservation devices would then be installed in the houses of those assigned to the
treatment group. Households that already have water-saving showerheads or flow
restrictors would be excluded from the sample and replaced with additional randomly
selected households. Ideally, placebo devices would be installed in the homes of the
individuals assigned to the control group, and members of the groups would not be
told which device they received. This procedure may seem cumbersome, but it avoids
the all-too-real problem of contaminating the results because the individuals receiving
the devices want to please the research team and change their behavior during the
study period to obtain the anticipated results.
An obvious problem is that some households will refuse to participate in the
experiment. At best, this requires starting with a larger sample to achieve the target
sample size. At worst, it partially invalidates the results, because nonparticipants may
differ in their water-use behavior from participants. The number of refusals does,
however, produce valuable information on the likely acceptability of a utility-wide
program. There are many provisos to setting up and conducting a reliable controlled
comparison (see Planning and Management Consultants 1992).
The water use of each customer in both the treatment and control samples is
measured before, during, and after the experimental period. Then, it is possible to
compute the sample standard deviations for the treatment and control groups, and
t-tests for the difference of means can be applied. This experiment would provide an
estimate of the mean and standard deviation of water saved by the device, as well as
an estimate of the penetration rate that could be expected for the entire utility given
that every customer was requested to install the device with the same intensity of
exhortation as the sample. The effort to enlist customers from the selected sample
would also give the utility a good estimate of the proportion of customers who already
were using low-flow showerheads or flow restrictors. A satisfaction survey could be
added to this experiment to determine how satisfied the customers in the experimen-
tal group and the control group were with their “new” showerheads.
Disaggregated Causal/Structural Model. Figure 10-10 shows a compre-

hensive model that can be used for estimating and forecasting water-conservation
effects. This model, proposed by Bruvold and Smith (1988), incorporates five major
components:
1. Conservation programs
2. Decision makers’ processing of these programs
3. Water-conservation actions
4. Observable influences on water use (weather, household, and community)
5. Outcomes: water use
Growing
Public Season
Education Temp
Beliefs
Growing
Behavioral Season
Responses Rain
Pricing
Knowledge
HH Income
Water
Use
Per
HH Size
Day
Use
Restrictions House Age
Structural
Lot Size
Building Responses
Code
Local Area Mores
Adapted from Bruvold and Smith 1988.
Figure 10–10 Residential Water Conservation Model
“Belief variables” refer to the decision maker’s perceptions of the need or

desirability of water conservation, the responsibility of individual businesses or
government to do something about conservation, and conceptions about the benefits
and costs of conservation, both monetary and psychological. Knowledge concerns
objective facts such as the amounts of water used for various purposes by the
individual and the monetary cost of water. “Behavioral responses” are changes such as
turning off the faucet while brushing teeth, taking shorter or less frequent showers, or
watering the lawn less often. “Structural responses” include fixing leaks or installing
water-saving fixtures.
Bruvold and Smith (1988) collected information to estimate this disaggregated
model with a sample survey of single-family residences in nine San Francisco Bay
water districts. They obtained additional data directly from the water districts. A
cross-section model of daily use per capita was estimated, using data from 641
individual households.
Q = – 252.2∗ – 28K∗ – 0.02C – 21.1P∗ – 8.0D + 7.4T∗ (Eq 10-6)

– 22H∗ + 3.5I∗
2
R = 0.18∗
where:
✻
= statistically significant
Q = household water use per capita per day for August 1982
K = knowledge: 0 = none, 1 = some
C = conservation beliefs: from survey, values from 20 to 100
P = marginal price
D = difference = P✻ × Q – actual water bill
T = average temperature
H = household size in persons
I = income category.
Because this is a cross-section model of individual consumers, the low R2 is

expected, because the model was unable to measure many of their specific water-use
behaviors. As usual, in models such as this, not all the independent variables are
statistically significant. The question facing the analyst then is whether to omit the
insignificant variables. They are often retained because they account for part of the
variation in water use, even if little confidence can be placed in their specific values.
The model shows that water use decreased with knowledge of water use and
rates, but was unaffected by the measures of belief revealed by the survey. As expected,
water use was lower where marginal price was higher. Average water use ranged from
a low of 97 gal per person per day in San Francisco to a high of 205 in Napa.
Individual household use varied much more widely, from less than 50 to far more
than 300. As expected, water use increased with temperature and household income.
Water use per capita decreased with household size, indicating some economies of
scale as more individuals share the same home and yard. The major challenge facing
the analyst is obtaining good measures of beliefs and knowledge, as well as water-
conservation programs.
The model of Figure 10-10 can also be applied to other categories of water
users. For example, Billings and Longstreth (1989) did a study of monthly summer
water use at apartments with grass in Tucson, Ariz., estimated the following linear
demand model by multiple regression:
Q = 6.22 + 0.85YARD + 1.00SHOWER – 2.09FAUCET (Eq 10-7)

+ 0.70KNOW – 1.75WC + 0.12VACANT – 0.014SIZE
+ 0.71POLICY
where:
Q = average summer water use per occupied apartment per month
in Ccf. Mean water use was 10.5 Ccf/apartment.
YARD = type of yard, ranging from 1 = desert to 5 = lush vegetation
SHOWER = use of water-saving showerhead = 1, regular showerhead = 0
FAUCET = low-flow faucets = 1, otherwise = 0
KNOW = rate structure knowledge, ranging from 0 = none to 5 =
maximum
WC = water-saving toilet or toilet-tank dam = 1, otherwise = 0
VACANT = vacancy rate, in percent
SIZE = size of complex; varied from 2 to 456 units
POLICY = managers have authority to implement water-conserving
changes = 1, no authority = 0.
Water use is expected to increase with increasing values of YARD. The

coefficients on each of the three water-saving fixtures are expected to be negative,
indicating that their use does reduce water consumption. Water use is expected to
decrease with manager knowledge of water rates, so the coefficient on KNOW should
be negative. Likewise, the expected sign on POLICY is negative, because apartments
where managers have authority to institute water-saving changes can be expected to
use less water. The coefficient on VACANT is expected to be positive, because water
use is defined per occupied apartment. Thus, irrigation water use is spread over fewer
apartments when there are numerous vacancies. Landscaping uses the same amount
of water regardless of occupancy. If water use were defined as water use per apartment
(vacant or occupied), the sign on VACANT would be negative. Finally, the sign on
SIZE, the number of apartments in a complex, might be expected to be negative,
reflecting economies-of-scale in water use in apartment complexes—perhaps because
of the decrease in the irrigated yard area per apartment in larger complexes.
The model was estimated using data from 178 apartment manager interviews
done in January 1989 and water-use data from the water utility for the summer of
1988. Since they all paid the same price for water, no price variable is included in this
model. Additionally, because the apartments are all located in Tucson, they all
experienced the same weather conditions, allowing for the omission of weather and
climate variables. The data for this regression are on the data disk, so the interested
reader can replicate this result as well as explore the effect on the model of changing
the model specification by omitting individual variables. The reader can also explore
the effect of alternative functional forms for the model, such as double logarithmic.
The two most interesting predictions from the model are as follows:
• Apartments with low-flush toilets will use 1.75 Ccf of water less per month
in the summer than those without.
• Low-flow faucets save 2.09 Ccf of water per month.
Before using results such as these, however, the analyst must evaluate the
reliability of the model and its results. Equation 10-7 has an R2 of 0.136, indicating
that the model explains 14 percent of the variation of water use from its mean value.
Although this is relatively low, even for a cross-section model, the regression may still
provide some useful results. The F statistic is 3.327, which is significant at 0.0015,
meaning that the equation is highly significant from a statistical viewpoint. The
standard error of the estimate is 6.046 Ccf. Unfortunately, the estimated savings from
the water-saving devices are smaller than the standard error of the estimate, which
means that the estimated savings are less than the typical error in the estimates.
The analyst should attempt to reduce the size of this standard error. The large
standard error may be caused by errors in reporting by apartment managers or other
causes. If the analyst could physically measure the flow rates of installed showerheads
and faucets, and measure the gallons required per flush for toilets, these potential
errors could be eliminated. Resources seldom permit this type of detailed data
gathering.
All of the individual estimated coefficients, with the exception of SHOWER and
POLICY, were significant at 5 percent or better. Thus, there is less than a 1 in 20
chance that the true value of the coefficient is 0 or has the “wrong” sign (using a one-
tailed test, because there is a stated hypothesis on the expected sign of each variable).
The estimated coefficients on SHOWER and POLICY have unexpected positive
signs and are not statistically significant. These results may be caused by errors in
reporting by apartment managers or other unspecified causes. Additionally, the
estimated coefficient on KNOW, manager knowledge of water rates, has an unexpected
positive sign. This variable is statistically significant at a 5 percent level using a two-
tailed test, which would be used if the initial hypothesis about the direction of its
effect were undetermined.
Analysts often drop insignificant variables from their models, and the reader
can verify that dropping the two insignificant variables from this specification lowers
the R2 only slightly and slightly increases the adjusted R2. Interestingly, when
SHOWER and POLICY are dropped from the regression, the F statistic rises to 4.18.
The relatively minor changes in the other coefficients when these two variables are
omitted from the model increases our confidence in these results. The “rule of
parsimony” (i.e., using the simpler model), supports dropping SHOWER and POLICY
before using a model such as this for forecasting applications.
This regression illustrates the type of information available from collection and
analysis of firsthand sample data on water use and conservation. The regression
model validates using certain conservation savings estimates such as those for low-
flow toilets and water-saving faucets. In other respects, the regression analysis suggests
certain problems and issues for further investigation. For example, why would more
knowledge about rates lead to increases in water use? Do the positive coefficients on
SHOWER and POLICY, together with the low statistical significance of these
estimates, mean that low-flow showers and managerial authority to implement water
conservation are just not important in the overall water use picture?
Clearly, validating some end-use estimates and raising questions about other
effects are potentially more valuable than simply using a list of end-use estimates for
water conservation developed from national data or other water systems and not
attempting to understand local conditions or attitudes about water use.
SUMMARY
Water conservation or efficiency programs are increasingly important in urban water
systems. Millions are served by systems with multiple water-conservation efforts,
including publicity about why conservation is important. The potential for water
conservation is enormous, because only a small percentage of delivered water is
actually used directly by people for personal uses A simple way to include water-
conservation and efficiency program effects in demand forecasts is to deduct the
estimated savings of the conservation program from the unconstrained (without
conservation) system water-demand forecast. A more sophisticated approach would
include conservation impacts as independent variables along with the price of water,
personal income, business activity, weather, and so on in a multivariable forecasting
model.
The broad range of estimates of conservation savings signals the difficulty of
estimating the effectiveness of particular water-conservation/efficiency programs
before they are implemented. This difficulty is compounded by synergistic and
competitive interactions among individual conservation program elements. These
uncertainties suggest the need for local statistical studies of water conservation
programs, as well as continuing monitoring of their effectiveness.
The generally accepted approaches for studying the effectiveness of water-
conservation programs include (1) control group comparisons and (2) multivariate
models estimated by regression analysis. Utility officials determined to use any or all
available conservation program elements can achieve specific water-use reductions in
a range of 5 to 20 percent. In this case, the water-conservation plan may become a
highly accurate forecast (Texas Water Development Board 2006; USEPA and
California Department of Water Resources 1997).
REFERENCES
Bauman, D.D., J.J. Boland, J.H. Sims, B. Kranzer, and P.H. Carver. 1978. The Role of Conservation in Water
Supply Planning. Contract Report 78-2. US Army Corps of Engineers Institute for Water
Resources.
Billings, R.B., and M. Longstreth. 1989. Water Use and Conservation in Multiple Family Dwellings in
Tucson, Arizona. Project Report. Tucson, Ariz.: Tucson Water.
Bruvold, W.H., and B.R. Smith. 1988. Developing and Assessing a Model of Residential Water
Consumption. Water Resources Bulletin 24(3): 661–669.
California Urban Water Conservation Council. 2001. Commercial, Industrial, Institutional Ultra Low Flow
Toilet Savings Study. 2nd ed. Sacramento, Calif.: California Urban Water Conservation Council.
Campbell, H., R.M. Johnson, and E.H. Larson. 2004. Prices, Devices, People, or Rules: The Relative
Effectiveness of Policy Instruments in Water Conservation. Review of Policy Research 21(5): 637–
662.
Chesnutt, T.W., and C.N. McSpadden. 1991. A Model-Based Evaluation of Westchester Water Conservation
Program. San Diego, Calif.: A & N Technical Services.
Chesnutt, T.W., C.N. McSpadden, and A. Bamezail. 1994. Ultra Low Flush Toilet Programs: Evaluation of
Program Outcomes and Water Savings. Los Angeles, Calif.: Metropolitan Water District of Southern
California.
Consumer Reports. 2005. Toilets: Comfortable and Efficient. August: 33–35.
Consumer Reports. 2007. Washers and Dryers: Savings at a Price. January: 39–41.
Dziegielewski, B. 2003. Water Supply Economics. In Managing Urban Water Supply. Edited by D.E. Agthe,
R.B. Billings, and N. Buras. Dordrecht, Netherlands: Kluwer Academic Publishers.
Gaudin, S. 2005. Using Water Bills to Reinforce Price Signals: Evidence from the United States.
International Water Association International Conference on Water, Economics, Statistics, and
Finance.
Goemans, C.G. 2007. Optimal Policy Instruments for Utilities Using Increasing Block Rate Pricing
Structures, unpublished manuscript.
Hills, S., and R. Birks. 2004. Washroom Behavior and User’s Perception of “Novel” Water Efficient
Appliances. Water Science and Technology: Water Supply 14(3): 13–23.
Keyes, A.M., M. Schmitt, and J.L. Hinkle 2004. Critical Components of Conservation Programs That Get
Results: A National Analysis. AWWA Water Conference Proceedings.
Kowalski, M., and D. Marshallsay. 2005. Using Measured Micro Component Data to Model the Impact of
Water Consumption Strategies on the Diurnal Consumption Profile. Water Science and Technology:
Water Supply 5(3): 145–150.
Lancaster, B. 2006. Rainwater Harvesting for Drylands, Volume 1: Guiding Principles. Tucson, Ariz.:
Rainsource Press.
Levin, E.R., W.D. Maddaus, M.M. Sandkulla, and H. Pohl. 2006. Forecasting Wholesale Demand and
Conservation Savings. Journal AWWA 98(2): 102–111.
Linaweaver, F.P., J.C. Beebe, and F.A.L. Skirvan. 1966. Data Report of the Residential Water Use Research
Project. Baltimore, Md.: Johns Hopkins University, Department of Environmental Engineering.
Maddaus, W.O., and L.A. Maddaus. 2006. Water Conservation Programs—A Planning Manual, M52.
Denver, Colo.: AWWA.
Martin, W.E., and J.F. Thomas. 1986. Policy Relevance in Studies of Urban Residential Water Demand.
Mayer, P.W., W.B. DeOreo, E.M. Opitz, J.C. Kiefer, W.Y. Davis, B. Dzieglelewski, and J.O. Nelson. 1999.
Residential End Uses of Water. Denver, Colo.: Awwa Research Foundation.
Mayer, P.W., E. Towler, W.B. DeOreo, E. Caldwell, T. Miller, E.R. Osmann, E. Brown, P. Bickel, and S.
Fisher. 2004. National Multiple Family Submetering and Allocation Billing Program Study. Aquacraft
Inc. and East Bay Municipal Utility District.
Mohadjer, P. 2003. Residential Ultra-Low Flush Toilet Replacement Program. West Jordan, Utah: Jordan
Valley Water Conservancy District.
Morris, R.L., D.A. Devitt, A.M. Crites, G. Borden, and L.N. Allen. 1997. Urbanization and Water
Consumption in Las Vegas Valley, NV. Journal of Water Resources, Planning, and Management
123(3): 189–195.
Nelson, J.O. 1992. Water Audit Encourages Residents to Reduce Consumption. Journal AWWA 10(10):
59–64.
New York City Department of Environmental Protection. 1998. Evaluation of New York City’s Toilet
Rebate Program. New York: New York City Department of Environmental Protection.
Pekelney, D.M., and D.L. Mitchell. 1996. Cost-Effective Cost-Effectiveness: Quantifying Effectiveness on
the Cheap. Presented at AWWA Conference, June, Toronto, Canada. Available online from
California Urban Water Conservation Council at. www.cuwcc.org.
Planning and Management Consultants, 1992. Evaluating Urban Water Conservation Programs: A
Procedures Manual. California Urban Water Agencies. Available at www.cuwcc.org.
Plosser, J.H., C.W. Pike, and J.D. Kobrick. 1992. Nonresidential Water Conservation: A Good Investment.
Journal AWWA 10(10): 65–73.
PPI (Pollution Prevention International, Inc.). 2004. Commercial, Institutional and Industrial Water Use
Survey Program, Santa Clara Water District, California. Irvine, Calif.: PPI.
Raftelis, G. 2006. 2004 Rate Survey. Available from www.raftelis.com/ratessurvey.html.
Renwick, M.E., and S.O. Archibald.1998. Demand Side Management Policies for Residential Water Use:
Rogers, P., R. Blatia, and A. Huber. 1998. Water as a Social and Economic Good: How to Put the Principle
into Practice. Stockholm: Global Water Partnership.
Runge, J., and J. Mann. 2006. State of the Industry Report 2006: An Eye to the Future. Journal AWWA
98(10):64–72.
Sovocool, K.A., M. Morgan, and D. Bennett. 2006. An In-Depth Investigation of Xeriscape as a Water
Conservation Measure. Journal AWWA 98(2): 82–93.
Sunset Magazine, 2005. Backyards from the Ground Up. (May 1)
Texas Water Development Board, 2006. 2007 State Water Plan. Available at www.twdb.state.tx.us/
Publications/Reports.
USEPA and California Department of Water Resources. 1997. Study of Potential Water Efficiency
Improvements in Commercial Businesses. Final Report. Sacramento, Calif.: USEPA, California
Department of Water Resources.
Vickers, A. 2001. Handbook of Water Use and Conservation. Amherst, Mass.: Waterplow Press.
Vickers, A. 2006. Perspectives—New Directions in Lawn and Landscape Water Conservation, Journal
AWWA 98(2): 56-61, 156.

CHAPTER 11
Short-Term
Emergency Conservation
The Hydro-Illogic Cycle
Awareness Concern
Panic: Water
Drought Conservation
Apathy Rain
Source: Adapted from The Cross Section 52(8) August 2006.
227
Some water utilities, especially in arid regions, occasionally face emergency shortages
of deliverable water. Utility managers respond with a variety of programs designed to
temporarily reduce water use. Many of these programs are politically and legally
acceptable only because they are temporary. Although many programs can reduce
water demand in both the short and long term, this chapter focuses on programs
intended to produce significant water savings in a short time period. Some of the
most severe programs are temporary, and others, such as general education, are often
part of longer-term efforts. Often these programs lose their effectiveness or are
discontinued once the crisis passes. Long-term or open-ended conservation programs
are analyzed in chapter 10.
This chapter starts with a general summary of the water savings that have been
experienced with the use of each program, and then discusses the revenue effects of
short-term conservation. This is followed by detailed discussions of the most likely
range of conservation that can be obtained from each program element. A brief section
on conservation program planning with reference to detailed planning manuals follows.
The final section describes several case studies of short-term conservation programs as
implemented by various utilities. Even though some of the case studies are somewhat
dated, each is an example of an actual emergency conservation program, thereby
demonstrating the range of effectiveness of various program elements.
Droughts cause most shortages in North America and Western Europe, reducing
stream flows and storage in reservoirs. Shortages can also result from aquifer drawdown,
mechanical failure in the water system, or failure to expand the supply system to keep up
with area growth. Shortages could also be the result of sabotage or warfare.
EFFECTIVENESS OF SHORT-TERM CONSERVATION PROGRAMS

The acceptability and the effectiveness of various conservation measures depend on
the cause and severity of the shortage. One of the challenges facing utility managers is
choosing a set of conservation measures and predicting their effect on water use and
utility revenue. Table 11-1 gives estimates of the potential water savings from a variety
of mostly short-term programs. The wide range of savings for some programs
suggests the difficulty of developing a good forecast of savings ahead of program
implementation, and the importance of program details and local effects. Neverthe-
less, these values can be used as a starting point in developing a forecast of program
savings, and as a check on final forecast results.
Often the initial goal is to reduce water use by a given percentage, or to keep
water use from exceeding a specified amount. The goal may turn out to be a moving
target, especially if a drought causes the shortage, because the severity and duration of
droughts cannot be adequately predicted. A drought can end at any time, but the
longer it has persisted, the more likely it is to continue. Tree ring research has shown
that some droughts in arid regions such as the southwestern United States have
persisted for decades.
SHORT-TERM EMERGENCY CONSERVATION 229
Table 11–1 Projected Savings by Program

Savings,
Program Comments
% of Total
Publicity/education 0 to 5 Results are program specific
Water audits, leak repair 3 to 5
Faucet aerator distribution 1 to 2
Showerhead distribution 1 to 3
Toilet tank displacement device distribution 1
Rebates for 6 - L toilets 3 to 5
Front-loading clothes washer, rebates or distribution 1 to 3
Pressure reduction 2 to 4
Voluntary outdoor use limits 0 to 10
Mandatory outdoor use limits 10 to 40 Varies with severity of limits
Shift to increasing block rates 1 to 5 Keeping revenue constant
Marginal price increase of 10% 1.5 to 7 See Table 9 - 4
Mandatory water budgets 10 to 60
Source: Author’s estimates, adapted from Mayer et al. 1999 and Maddaus and Maddaus 2006.
Conservation Effect on Revenue. When designing and implementing pro-

grams to deal with water shortages, managers cannot ignore the effect of reduced
water use on utility revenue, especially if they do not plan to provide subsidies to
water production from other sources such as budget reserve funds or sales and
property taxation.
Drought fund. A creative, forward-looking way to deal with emergency revenue
shortfalls is to develop a “rainy day fund,” sometimes referred to as a sinking fund (in
this case a drought fund), into which periodic deposits are made and which is
available to cover revenue shortfalls beyond the control of the utility. This strategy is
like the rainy day funds established by some state legislatures to help maintain
program expenditures when revenue falls during a recession. The difficult challenge is
to convince the public, policy makers, and governing officials that the utility should
run surpluses during normal times. The pressure to avoid the rate increases that
would be required to fund this program is often quite strong. Forecasting how much
might be needed is difficult and open to differing opinions. A comprehensive utility
forecasting model could be invaluable in determining a target for the fund and an
acceptable rate of accumulation for it.
Funding sources. The effect of a successful nonprice conservation program on
utility revenue depends on how the utility is funded. If utility costs are funded entirely
from property taxes, bond issues, and fixed periodic fees paid by water users,
conservation obviously does not affect utility revenue. This describes situations in
which water use is not metered, and hence individual customer water use is not
constrained by the price of water.
A successful nonprice conservation program will result in reduced water
deliveries and therefore reduced revenue when water use is metered and the utility
imposes a commodity charge. The size of the decrease in revenue depends on the rate
structure. A structure with a high fixed monthly service charge and a low commodity
price for delivered water will suffer less than a system with a small monthly charge
and increasing block rates. Decreased deliveries of water will also likely reduce cost
somewhat. The utility may be forced, however, to acquire new or more expensive
sources of water, so that overall less water is available, but its unit cost of acquisition
is greater. The utility may also by necessity turn to lower quality intake water,
incurring increased treatment cost. If revenue from water bills is supplemented by tax
revenue, it may be possible to meet the revenue decrease from this source. Although
increasing water rates is never politically popular, the one thing to be avoided if at all
possible is to institute a conservation program and, after it succeeds in reducing water
use, raise rates. Consumers will feel that they have been rewarded for their
conservation efforts by being stabbed in the back. It is far better to take the initial heat
of raising rates at the beginning of the conservation program. To allay concerns about
the effect of rate increases on low-income customers, a lifeline rate block can be
maintained or added to the rate structure. This is a low or zero commodity charge for
the initial amount of water used each month, based on what is required for drinking,
cooking, and sanitary needs.
A rate increase can be justified on several related grounds:
1. Additional revenue is needed to fund the emergency conservation program.
2. The utility will incur higher costs for acquiring and treating new sources of
water during the emergency.
3. New investments will have to be made to increase future supplies of water,
and these require additional revenue, both to repay bonds and to maintain
a favorable bond rating to minimize interest cost.
4. Higher rates will give individuals and businesses an incentive to use less
water.
5. With reduced water sales, a higher price will be required to enable the
utility to meet its fixed operating costs and bond repayment schedules.
When increases in marginal price are an integral part of a conservation
program, revenue shortfalls can be avoided. As discussed later in this chapter and in
chapter 10, higher water rates reinforce and interact with other conservation
programs, potentially making it easier to achieve conservation goals.
CONSIDERATIONS IN DEVELOPING A CONSERVATION PROGRAM

A variety of practices and political realities must be evaluated by policy makers in
designing a conservation program.
Incentives and Techniques for Conservation. It is helpful to distinguish
between programs that offer incentives to conserve water and those that give
consumers techniques for saving water. The prime example of an incentive to save
water is a very large increase in the marginal price of water, often referred to as a
penalty rate. An example of a technique to save water is the distribution and
installation of reduced-flow showerheads. Even though the concepts of incentives and

techniques are quite distinct, many conservation program elements contain aspects of
both. Incentives (motives) for short-term emergency conservation include:
• Appeal to civic pride
• Publicize the need to conserve
• Shift from constant or decreasing to increasing block rates
• Raise the marginal price of water
• Vary marginal price directly with the severity of the drought
• Impose penalty rates for use above a predetermined amount
• Threaten water cutoff for excessive use
Methods for short-term (emergency) conservation include:
• Reduce water pressure
• Offer free water audits to high-usage customers
• Offer free or low-cost repair of simple leaks
• Prohibit new water services
• Distribute and possibly install low-flow showerheads
• Encourage shorter showers
• Encourage turning off faucet while brushing teeth
• Distribute and possibly install faucet aerators
• Encourage dishwasher use with full loads and not prewashing dishes
• Encourage matching clothes washer water level with load
• Distribute and possibly install toilet tank displacement devices
• Institute a high-efficiency toilet replacement program
• Encourage hotel/motel guests to reuse towels and bedding for more than
one day
• Discourage restaurants from automatically serving water
• Institute voluntary irrigation restrictions
• Offer real-time information on irrigation requirements of vegetation
• Institute mandatory irrigation restrictions
• Prohibit water runoff on sidewalks, streets, and parking lots
• Prohibit irrigation during rain
• Prohibit daytime automatic sprinklers
• Limit duration and times per week of automatic sprinklers
• Permit watering with only a handheld hose

• Prohibit hosing off sidewalks, patios, driveways, or other paved areas
• Prohibit car washing except at carwashes that recycle water
• Set up water budgets for each customer
Choice of Program. How effective is each program likely to be? How long
will that effectiveness last? What is politically and legally acceptable? How do multiple
programs interact? These are difficult questions, and the answers will necessarily vary
by state, locality, severity of the shortage, and other factors.1 The available programs
are also likely to be different for government-owned utilities than for state-regulated
investor-owned utilities. How much authority utilities have to request or mandate
conservation and to spend funds for conservation programs ultimately depends on
the laws and constitution of each state.
The effectiveness of individual emergency conservation programs depends on
many factors, only some of which can be influenced or controlled by utility officials.
The imposition of multiple programs at the same time, as well as variations in
weather conditions such as rainfall and daily high temperature, complicate program
evaluations.
Metropolitan Area Coordination. Several studies have demonstrated the

value of a unified conservation program within the same generalized area. When
adjoining water utility service areas served by the same local news sources have
different programs, water customers become confused about what is expected of
them, which may dramatically reduce the effectiveness of otherwise well-designed
programs. Consumers in the San Francisco Bay area reported receiving conflicting
messages regarding the severity of the supply situation during droughts.
Demand Hardening—Reality or Myth? Some utility officials and consult-

ants have expressed concern about possible “demand hardening.” The California
Urban Water Agencies (Tabors Caramanis and Associates 1994, p. 11) defines demand
hardening as “the diminished ability or willingness of a customer to reduce demand
during a supply shortage as a result of having implemented long-term water
conservation measures.” California water officials participating in a survey had this
concept explained and were then prompted to rate their level of concern. The results
are shown in Table 11-2. In a follow-up question, none of the officials indicated that
this phenomenon had any significant effect on their long-term conservation plans.
One of the thoughts behind this concept is that short-term conservation is
primarily focused on changes in behavior whereas long-term conservation stems from
changes in fixtures, equipment, and landscape choices. Although it is true that
behavior can be altered quickly, various fixtures such as low-flow showerheads are
often part of an emergency conservation plan. Conversely, some changes in behavior
1. Chapter 10 examines methods of analysis of effectiveness.

Table 11–2 Survey Concern Level for Demand Hardening
Level of 0 1 2 3 4 5 6
Concern None Very
High
Number of 0 7 0 4 1 0 0
Responses
Source: Based on Tabors Caramanis and Associates 1994.
may be long-lasting, such as turning off the faucet while brushing teeth. Certainly,
replacing 5 gal per flush (gpf) toilets with 1.6 gpf units severely reduces the potential
for additional conservation from this source. Likewise, changing faucet-use behavior
has less impact if faucets deliver less water than previously. Every conservation
forecast must take into account the existing stock of water-using fixtures and features
of the utility area. In this sense, effective long-term conservation programs reduce the
potential for further conservation. In some utility areas, however, part of the success
of emergency programs has been attributed to the ongoing presence and acceptance
of the need for long-term conservation. As indicated in Figure 10-1, local habits and
expectations determine a large proportion of water use, and that usage can be
dramatically reduced if absolutely necessary.
Synergy and Competition Among Conservation Programs. Forecasts of
anticipated water conservation resulting from multiple programs must be adjusted for
synergy and competition. For example, although publicity by itself may not result in
much conservation, when combined with most other programs, it makes those
programs more effective. Conversely, the expected savings from higher rates cannot
simply be added to those from other effective programs, because price increases
provide incentives for conservation and methods such as moisture sensors on
irrigation systems or faucet aerators provide a method for conservation. One of the
cautions in applying price elasticity estimates to forecast the effect of a price increase
on water use is to avoid double counting.
PROGRAM ELEMENTS
A wide variety of specific policies have been used by water utilities to encourage water
conservation.
Education and Publicity. Every conservation program will include several
elements to educate the population about the water supply emergency and to
publicize the need for conservation efforts. To be effective, these efforts should
provide accurate, timely information on the causes and severity of the problem, and
on what measures are being taken to solve the problem. It is highly beneficial if all
stakeholders can be brought into the discussion and the decision-making process.
These include utility management; city, county, and perhaps state office holders; and
media outlets such as newspaper, radio, and television. If all of these information
sources can reach a consensus on the needed programs and speak with a single, clear
message, the programs are much more likely to be successful. Sometimes the most
effective educational efforts are carried out through the school system, which can
incorporate water-conservation benefits and methods into its environmental and
citizenship curriculum.
Although education and publicity are an essential element of any conservation
program, forecasting the potential water-use reductions from publicity is problematic.
A few of the case studies presented found water savings from publicity, but most did
not. A conservative forecast would attribute only minimal water savings of between 0
and 5 percent to publicity alone; however, publicity is synergistic with most other
conservation programs, which would be far less effective without it.
Flow-Reducing Device Distribution. Many utilities, especially in the arid

West, have distributed faucet and shower flow restrictors and toilet tank displacement
bags or bottles. These have had mixed results, from no effect because few people
bothered to install them to modest water savings, which tended to deteriorate over a
period of a few years as the devices were removed by those who used them. To be
successful, the distribution of these devices must be accompanied by extensive
publicity on why their use is important and how to install them. There is also some
evidence that distribution of devices by personnel trained to install them is likely to
substantially increase their utilization.
These devices have been made moot in many communities by the widespread
installation of lower-flow faucets and showerheads along with 1.6 gpf toilets. Front-
loading washing machines and 6-L (1.6-gal) toilets are discussed in chapter 10.
Distributing these devices is likely to be most effective in areas of older housing
where they have not been distributed before and where little retrofitting has occurred.
Communities have experienced overall water savings of anywhere from 0 to 4 percent.
Low-Flow Showerheads and Faucet Aerators. Many utilities have dis-

tributed faucet aerators and reduced-flow showerheads that, if installed and main-
tained, measurably reduce water use. The challenge in distributing these items is to
inspire people to install them. Some programs have been relatively successful, yet
others have found that simply distributing these items is ineffective because they are
not used. At substantially increased expense, the utility can offer free installation.
Leak Detection and Repair. Although there may not be enough time dur-
ing an emergency water shortage to carry out a complete leak detection program, leak
detection and repair should be part of any comprehensive conservation program, and
will result in predictable water conservation. The beginning point is to compare
accounted-for final deliveries of water to system pumpage. If there is a sizable
proportion of unaccounted-for water, several relatively new techniques can determine
which mains and distribution pipes may be leaking, and the community can be
enlisted in an effort to report suspected leaks. To make this effective, the utility must
be responsive to leak reports.
The Residential End Uses of Water (Mayer et al. 1999) report published by
AwwaRF found that almost 14 percent of all indoor residential water was lost to
leakage beyond the water meter. A relatively few customers, however, accounted for
the majority of water lost to leakage. Within the study sample, 67 percent of the
homes had leakage of 10 gal per household per day or less, and 5.5 percent of the
homes had leakage greater than 100 gal per day. Thus the utility could have a
significant impact on total water use by offering (or requiring) water audits for
residential customers with especially high water usage, and emphasizing leak
prevention in its publicity and education efforts. These findings can be used to
forecast the water savings from a leak detection and repair program.
Voluntary Conservation. Many communities have requested that water

customers voluntarily reduce lawn sprinkling, take shorter showers, limit car washing,
avoid washing off sidewalks and driveways, and so on. These programs have had
mixed results. In San Diego, voluntary programs were highly successful, apparently
because of an existing extensive array of conservation programs that had customers
already thinking about conservation. The success was also attributed to intense
competition between San Diego with its voluntary program and Los Angeles with its
mandatory program. The competition engendered an extremely large volume of buzz
in the media, and, no doubt, among individuals.
In contrast, voluntary programs in many communities have had only negligible
effect, as discussed in the case studies at the end of this chapter. Requests for voluntary
restraint in water use often fall on deaf ears, and in some instances, consumers who
were surveyed after the programs had been implemented were unaware of the
programs’ existence even though they had been exposed to water bill inserts and
media advertising.
Mandatory Limits on Outdoor Water Use. Many communities facing a

water shortage have used mandatory outdoor limits. They are effective and can
substantially reduce water use during the lawn-watering season. The literature says
little about enforcement efforts, although some urban areas include “water cops” in
their programs. Presumably a city implementing such a program would need a few
water cops to watch for violators and respond to reports of violations. Most likely the
primary method of enforcement is examination of water bills, which could be
automated by computer. It is unclear how many years such a program can be
successfully extended. Most of the case studies considered here used the program for
only one or two years. If the drought or other supply shortfall is longer lasting, more
permanent measures are called for, as discussed in chapter 10. If the restrictions are
severe enough that lawns, ornamental plantings, and even trees die from lack of water,
the cost to property owners is high, and justified only under conditions of extreme
supply limitations. Nevertheless, there is ample evidence that many property owners
irrigate their vegetation much more than is required for maintenance, so that
cutbacks can be enacted without inflicting permanent damage.
Water Budgets. Setting an individual water budget for each customer can be
an effective method of inducing water conservation: use above the budget reaps a high
penalty. The utility allocates water to each customer based on their presumed needs
and the overall balance between supply and demand in the utility service area. Ideally,
utilities develop the allocations using best management practices, which may include
water-saving indoor fixtures and efficient outdoor irrigation schedules. Actual
allocations necessarily work with existing mixes of technology. The budgets may be
partially set based on historical water-use rates, although many customers find this
basis to be unfair, because those who conserved in the past are effectively penalized
compared to the profligate water users. Budgets for residential and small commercial
customers can generally be set based on household size, number of employees, and
type of business. Those for larger business firms may need to be set by individual
negotiation.
Increased Water and Sewer Use Rates. Raising the price of water is a
proven method of getting a customer’s attention and creating an incentive to
conserve. When sewerage use fees are based on water use, the two should be
combined to consider their effect on water use. If the water system does not meter
individual users, no direct link exists between water use and payments for water. In
this case, the utility might decide to begin installing meters, but price increases are not
available as a conservation strategy until the metering program is substantially
completed. Most US water systems without meters are very small, but some large
cities in other countries do not meter water deliveries.
Raising the marginal price of water can engender significant political objec-
tions. One way to raise customers’ ire is to request or mandate conservation, and then
when water deliveries and revenues fall, raise rates to cover the shortfall. It is far better
to announce that water rates are being raised to encourage water conservation as part
of a comprehensive water conservation program. The utility should estimate the
anticipated drop in water sales, taking into account both the price and nonprice
conservation likely to result from the program, and adjust rates appropriately.
The utility could mitigate the effect of a price increase on low-income
residential customers, who are generally also low-volume users, by instituting or
continuing a low marginal price for an initial amount of water used each month, as
discussed more fully in chapter 9. Subsidized leak repairs and other programs to
facilitate reduced water use could also be targeted to low-income residents. Price
increases and other conservation measures such as mandatory restrictions are often
antagonistic, in that the sum of conservation that could be anticipated from each
alone is greater than that achieved when they are used together. To illustrate this,
consider that a large price increase by itself is likely to induce cuts in irrigation.
Mandatory limits on irrigation will also reduce irrigation water use. Therefore, when
both are used together, although they reinforce each other, the total savings are likely
to be only slightly larger than for either program taken by itself.
Water rates as a function of supply constraints. An intriguing proposal that uses
water rates to manage water use during a drought is to vary rates with the severity of
the drought situation. Using a predetermined sliding scale, utilities could increase
water rates when the balance between available supplies of water and deliveries
worsens, and decrease them when people conserve more, thereby improving the
balance between supplies of water and deliveries. This has the great advantage of
directly rewarding customers for increased conservation. Because few, if any, utilities
have metering systems that would permit the system to operate on a real-time basis,
the rate changes would lag the changes in water use by a month or more. Utilities
would need to announce rate changes in a timely manner so that people were aware
of them before making water-use decisions each month. They would need to carefully
calibrate water rates for this program to avoid revenue shortfalls or have sufficient
reserve funds to cover any shortfalls. To be effective, the program would need to be
widely publicized, including clear statements on monthly water bills of the current
situation, and the possible rate changes that depend on conservation effectiveness. We
are not aware of any utility that has tried this idea.
Although government-owned water utilities can usually adjust water rates
quickly by act of their governing board, investor-owned utilities typically must receive
permission from a regulatory agency to raise water rates. To receive permission, the
utility needs a well-developed conservation plan and a defensible forecast of the effect
of conservation on utility sales and revenue. The additional constraint in this
situation arises from utility commission mandates. Private utilities also lack policing
power, so some of the conservation programs require implementation by local or
regional governmental authorities. Investor-owned utilities must therefore coordinate
their conservation plans with local governmental officials as well as their own boards
of directors and state regulatory bodies.
Water rates and inflation. Increased water rates create an incentive for every
customer to reduce water use.2 By increase, we mean that rates rise by more than the
general level of prices, as measured by the consumer price index (CPI). Thus, if over
a one-year period the CPI increased by 4.2 percent and water rates increased by 5.9
percent, the real (inflation-adjusted) price of water would go up by only 1.7 percent.
The usage reduction resulting from increased water rates depends on the rate
structure, the importance of water bills in the customer’s budget, and other factors.
The potential effect is most easily described by the price elasticity of demand for
water. If the utility has a demand forecasting model that includes the price of water, it
will already have access to this statistic. If not, the utility can make a good guess as to
its price elasticity based on the methods discussed in chapter 9. There is some
evidence that consumer response to a price increase will be greater if the marginal
price of water is prominently displayed on their periodic water bill. If water rates are
very low, however, this tactic could backfire when customers look at the rate and
respond, “If water is that cheap, there is no need to conserve.”
Computation of a required price increase. Given a desired percentage decrease in
water use and the price elasticity of demand, computing the required general rate
2. Some customers will not respond to a price increase, but simply pay their higher water bill. This
may upset some individuals who do not favor market solutions to water shortages.
increase is straightforward. Suppose the price elasticity has been estimated as –0.24,
which means that a 10 percent increase in water rates results in a 2.4 percent drop in
water usage. If the utility wanted to decrease water use by 10 percent using water rates
alone, it could increase water rates as follows, using the basic elasticity formula from
Chapter 9:
Required percent price increase = Desired percent change in quantity ÷ ε (Eq 11-1)
where:
ε = price elasticity of demand.
In this example, the required increase is –10% ÷ –0.24 = 41.7%.
It must be emphasized, however, that little information exists on the most likely
response to price increases of this magnitude. The response may be greater than
shown in this example if the price elasticity of demand increases as people become
more aware of the price and as their water bills increase. Appropriate publicity and
educational efforts can also increase the effect of price increases. The effect on water
use could also be smaller, especially if water bills remain less than 1 percent of income
for most households. An analyst could develop a range of possible values rather than
a single most likely value for the forecast using a confidence interval for the estimated
price elasticity of demand.
Price Penalties for Excessive Use. Price penalties for excessive use is a
conservation strategy that has proven to be highly effective in getting the attention of
customers. To be effective, the penalty must be large, 5 to 10 times the high marginal
price in the rate schedule, so that few customers would be willing to pay it. As with any
price strategy, it requires widespread publicity. The penalty rate should be prominently
displayed on all water bills, with a clear statement of how much water, if any, was billed
at the penalty rate. If an individual’s usage is somewhat close to the penalty level, that
information should also be prominently displayed. Experience shows that most
customers will reduce water use sufficiently to avoid these penalty rates.
Changes in the Form of the Rate Structure. A change from declining
block rates or constant unit commodity charges to increasing block rates increases the
effectiveness of prices as a conservation measure. There is no easy way to generalize
the expected reduction in water use from making such a change. Nevertheless, such a
change signals to consumers that water is an increasingly valuable resource, and that
higher use levels will result in higher marginal prices. This message gains increasing
punch when utilities prominently display the customer’s marginal price for current
use along with either the reduced marginal price if usage drops or the increase in
marginal price for higher usage. The effect will depend on the details of the rate
structure such as the placement of the use block boundaries, the size of the steps
between blocks, and the overall level of water rates. The response may also vary with
the characteristics of the community, including the income distribution of residents
and historical expectations about outdoor water use. Chapter 9 gives a more detailed
treatment of increasing block rates. Chesnutt and colleagues (1997) wrote a handbook
on conservation rate design and implementation, giving very detailed suggestions. A
conservative forecast is that changing to an increasing block rate structure, with no
intended increase in revenue, would reduce water use by 1 to 5 percent.
DROUGHT PLANNING
During an emergency, utility officials will be continually monitoring the balance
between their supplies of water and the amount being used. Figure 11-1 is an example
of a conservation program process. In the last few decades, interest in planning for
drought has increased at all levels. In 1980, only three states (New York, South Dakota,
and Colorado) had drought plans. Today, most states either have some type of plan or
are developing a plan. The tremendous economic, social, and environmental costs
associated with drought impacts are the primary impetus for development of these
plans.
With increased interest, the drought planning process has evolved through trial
and error and careful research. Although the prospect of drought planning can still be
daunting, there are now a plethora of useful examples and source materials.
One excellent source for planning water conservation, among others, is
Maddaus and Maddaus (2006) Water Conservation Programs—A Planning Manual,
published by AWWA. The Texas Commission on Environmental Quality (TCEQ) has
prepared model drought contingency plans for wholesale and retail public water
suppliers, irrigation districts, water supply corporations, and investor-owned utili-
ties.3 The California Urban Water Conservation Council also published several
conservation manuals, some of which are available for free download on its Web site4.
Valuable material is also available from agencies in other states and other countries.
Each drought contingency plan for a retail water utility should include the
following elements, many of which will be discussed in the next major section of this
chapter:5
• Specific, quantified targets for water-use reductions (new requirement)
• Drought response stages
• Triggers to begin and end each stage
• Supply management measures
• Demand management measures
3. Listed in references. If you use the model plan, adapting it to fit your customer base and your
water system, you'll be sure to meet the TCEQ’s minimum requirements. These plans are available
from the TCEQ. You can receive a print copy of the model plan by calling 512-239-4691, or by sending
an e-mail to wras@tceq.state.tx.us.
4. cuwcc.com/publications.
5. Texas Administrative Code Title 30, Chapter 288, Subchapter B.
Objective: Balance water supply and

demand and maintain adequate revenue
Evaluate: Compare expected

cost/benefits of new programs
Choose: Determine water efficiency

goal and needed revenue
Select program: Determine conservation,

efficiency, and revenue measures to
implement
Implement program
Evaluate results: Evaluate water savings

from programs and adequacy of revenue
Question: Are goals achieved? NO
YES
Figure 11–1 Flow Diagram of Goal Achievement Process
• Descriptions of drought indicators

• Notification procedures
• Enforcement procedures
• Procedures for granting exceptions
• Public input to the plan
• Ongoing public education
• Adoption of plan
• Coordination with regional water planning group

Evaluating Program Effectiveness. The obvious indicator of the effective-
ness of conservation programs is total system deliveries on a daily and peak basis. If
deliveries fall sufficiently, the conservation program is usually deemed successful, and
if not, additional steps must be taken. A simple comparison of daily deliveries
compared to the same date a year ago or to a week ago, however, does not tell the
entire story of conservation effectiveness.
One challenge in evaluating the effectiveness of any conservation program is to
estimate how much water would have been used without the conservation program.
Several models can predict water use from one to several days ahead using weather
forecasts that incorporate recent and anticipated precipitation and temperatures. Such
models produce a better measure of the effectiveness of conservation programs,
because water-use rates depend on seasonal factors and immediate and recent weather
conditions. A good soaking rain may reduce sprinkling demand for several days or
even a week, and extremely hot weather will increase those demands., Evaluation
models using realized weather data can provide relatively accurate estimates of what
water use would have been without conservation programs.
As long as water use is less than what it would have been without the program,
the conservation efforts were not entirely unsuccessful. Of course, if water-use goals
are not met, additional measures are needed.
Emergency Conservation Experience. Utilities and communities have insti-
tuted a wide variety of short-term conservation programs. Fortunately, many have been
carefully analyzed. Each of the following case studies contains a careful description of
the situation of the community experiencing shortages, the programs implemented, and
the resulting changes in water use. These case studies are arranged in order of their
estimated value in understanding, developing, and analyzing short-term conservation
programs. Each case study starts with a few key words describing its most significant
contribution to understanding and forecasting program effectiveness.
Denver, Colorado, and Other Front Range Cities, 2002.
• Voluntary versus mandatory restrictions
• Coordination among adjacent communities
Background and program. The Front Range (eastern side of the Rocky
Mountains) houses several cities, with various surface water sources and rights. A
severe drought in 2002 prompted some to implement conservation programs,
focusing on publicity about drought conditions, mailings, educational efforts, and
outdoor water-use restrictions. Although the restrictions focused primarily on lawn
watering, car washing and other uses were also limited.
Results. Voluntary restrictions had little effect on water use. Mandatory limits
on lawn watering to once every three days resulted in an average water saving of
22 percent. Restricting watering to two times per week saved 33 percent, and limiting
watering to once per week saved 56 percent. Residents registered some confusion
about these regulations because all of the cities are served by the same media outlets,
and people heard about the whole range of restrictions rather than only those that
they were expected to follow. An effort to achieve coordination among these cities in
their conservation programs came to an abrupt halt when the drought ended (Kenney
et al. 2004).
Central California, 1987–1992.
• Water rate increase
Background and program. A prolonged drought of varying intensity persisted in
central California for at least five years. Pint (1999) reports on several utility districts
that relied primarily on price increases to reduce demand. He gives a detailed analysis
of the price increases and responses in Alameda County, near San Francisco. His
method is discussed in chapter 13.
The research used detailed information including variables such as lot and
house size on 599 individual single-family households. The study covered five
predrought years and five drought years. The local utilities imposed increasing block
rates starting in July 1991. Water rates increased in real terms by about 41 percent over
the years of uniform marginal prices, with dramatic increases for higher-volume users
after July 1991.
Results. The following system demand reductions were obtained: Alameda
County, 27 percent; East Bay Municipal Utility District, 33 percent; Santa Barbara,
62 percent; and Long Beach, 20 percent. When increasing block rates were imposed by
Alameda County, the percentage of customers using larger quantities in each of the
newly enacted higher-use blocks fell dramatically, as shown in Table 11-3. Even
though the highest rates were reduced the following year, the proportions of
customers in these upper blocks of the rate schedule did not return to their prior
levels during the period under study.
Table 11–3 Distribution of Customers by Water Use Under Uniform and Increasing Block Rates,
Alameda County, California
Uniform Rates With Increasing Block Rates
Percent of Percentage of Increase Over
Residential Uniform Rate Percentage of
Use Block Customers July 1991 July 1992 Customers
First 60 9 11 78
Second 18.3 100 35 14
Third 10 200 35 5
Fourth 11.7 300 71 3
Source: Pint 1999.
Southern California, 1991.

• Effect of publicity
Background and program. By early 1991, storage levels in the large northern
California reservoirs declined to only 5 to 20 percent of normal and snowpacks were
alarmingly low. Wholesale suppliers ordered Los Angeles (LA) and San Diego (SD) to
reduce water use. These cities have quite similar demographics and climate. LA
imposed mandatory watering restrictions and penalty rates for excessive water use; SD
chose voluntary restrictions. Extensive publicity surrounded both programs, initiated
by the utilities, the local press, and the wholesale supplier. SD also received extensive
and ongoing accusations that it was not doing enough. Both cities had extensive
predrought conservation programs, and both made sustained use of publicity to
encourage water conservation during the summer of 1991. Although much discussion
centered on the fact that LA had a mandatory program while SD had a voluntary
program, no information was reported on enforcement activities in LA. Shaw and
others (1992) described the program details, which are reproduced in Table 11-4.
Table 11–4 Conservation Programs, Los Angeles (LA) and San Diego (SD), 1990–1991
Program Description Used By
Long-Term Continued Activities LA SD
Water-efficient landscape training and education ✔ ✔
Xeriscape requirements for new construction ✔
Mandatory 10% reduction in turf >3 acres ✔
Residential water surveys ✔ ✔
Efficient bathroom fixtures required on remodel or sale ✔
Ultra low-flow toilet rebates ✔ ✔
Water-conservation kits ✔
Low-flow showerhead distribution ✔
Business and industry program ✔
Incentives for public schools ✔
Low-interest loans ✔
Water-waste
- complaint procedure ✔
Public information, advertising, and bill inserts ✔ ✔
Summer water-rate
- surcharge ✔
Drought Response Measures
Increased advertising and information ✔ ✔
Prohibited watering of sidewalks and driveways ✔
Required restaurants to serve water only on demand ✔
Restricted lawn watering to the hours between 10 a.m. and 5 p.m. only ✔
Levied 15% to 25% surcharge on use exceeding 1986 base level ✔
Expanded conservation hotline ✔
Worked with high-volume users to reduce consumption ✔
Source: Extracted from Shaw et al. 1992.
Results. Both cities experienced water-use reductions of 22–25 percent com-

pared to what would have been expected without the conservation programs. We are
not impressed by the LA penalty rates, because they kicked in only when individual
users exceeded their use in 1986 and were modest in size. We believe that the similar
reductions in water use between LA and SD were the result of the large, continuing
controversy about who was conserving better, and publicity about very low reservoir
levels.
San Francisco Area, California, 1977.
• Coordination among districts
• Mandates—rigorous, moderate, or mild
• Customer perceptions of equity
Background and program. The San Francisco area experienced a two-year
drought, which received widespread publicity throughout the region. For the year
ending in July 1976, precipitation was less than half its long-term average, and less
than one-third the following year. The drought problem was not recognized until
after normal precipitation did not occur in the spring of 1977. Bruvold (1979) reports
on survey-based research among nine major water supply districts. Rigorous
programs aimed for reductions of 30 percent or more using mandatory restrictions
and penalties along with publicity. Moderate programs aimed for reductions of 25 to
30 percent using milder regulations and penalties. Mild reductions of less than
25 percent were sought by other communities using voluntary programs. All
programs used publicity and education. The wide variety of desired water savings
reflected the varying supply conditions of the water utilities. Bruvold’s survey was
carried out on a sample of consumers from nine districts, each having more than
10,000 customers and falling clearly into one of the three conservation program
categories. Hoffman and colleagues (1979) reported on an overlapping set of districts.
Table 11-5 shows the districts included in each report and their chosen policy
approaches. The two reports agree except for one district as noted in the table.
Results. Reductions in consumption averaged 50 percent with rigorous pro-
grams, 37 percent with moderate programs, and 20 percent with the mild programs
(Bruvold 1979). Water use in San Rafael among single-family residences actually fell
by 75 percent and apartment dwellers reduced use by 45 percent. Water use generally
fell by more than the requested percentages for all districts studied (Hoffman et al.
1979). Consumer responses to these programs were generally favorable, presumably
because they recognized the necessity for conservation to avoid more severe
consequences later. Media coverage of empty reservoirs and the obvious lack of
precipitation reinforced people’s belief in the severity of the drought. Most negative
comments about fairness focused on required reductions that were based on a
percentage of prior years’ usage. Respondents preferred per capita based use limits.
Those subject to modest programs suggested that programs would be more
effective if made mandatory. Respondents also wanted better education and commu-
nication. Communication was most likely muddied by the close proximity of districts
Table 11–5 Service Providers and Programs, San Francisco Area
Program Description
Percentage
Reduction Allocation
Provider Hoffman et al. Bruvold Sought Approach
San Rafael, Marin County Mandatory Rigorous 57 Per capita

East Bay Municipal Utility Mandatory Rigorous 35 Per capita
Contra Costa County Mandatory Rigorous 30 Prior use
San Francisco Water Mandatory Moderate 25 Prior use
Palo Alto Moderate 25 Prior use
San Mateo Moderate 25 Prior use
Sunnyvale Water Voluntary 25
Santa Clara Valley Water Voluntary 25
San Jose Water Works Voluntary 25
Great Oaks Water Co. No program Mild 25 Prior use
Napa Mild 20 Prior use
Vallejo Mild 10 Prior use
Source: Bruvold 1979; Hoffman et al. 1979.
with widely disparate conservation programs. Respondents also had negative reac-
tions to what they perceived as double messages from their utilities, urging them to
conserve and seemingly at the same time encouraging greater water use. We wonder if
this resulted from utility concern about falling revenues resulting from conservation
choices by consumers. Household water bills fell because of conservation in all
districts except San Rafael. According to Hoffman and colleagues, these water agencies
placed much more emphasis on equity and public perception of conservation
programs than on revenue considerations and administrative convenience. Several
water-rich districts admitted that they adopted voluntary rationing programs to
support other districts experiencing shortages. Surprisingly, the drought reduced
these utilities’ interest in using water rates to encourage conservation in the future. It
also increased their desire to maintain long-term conservation programs.
Corpus Christi, Texas, 1984.
• Specific program elements
• Nonsymetric response to restrictions
Background and program. Corpus Christi depends on reservoir storage for its
municipal water supply. When reservoir storage fell because of severe drought
conditions, the city implemented voluntary, and then mandatory, conservation
programs with the following conditions:
• Condition 1: Voluntary—customers were asked to limit watering to every
other day. Municipal operations were ordered to conserve water.
• Condition 2
— Lawn watering was permitted only every 10 days, rotated by street
address. Watering was allowed only with a hand-held hose, bucket, drip
irrigation, or an attended sprinkler system.
— Other vegetation watering was permitted only on Tuesday, Thursday,
Saturday, and Sunday between 6 a.m. to 9 a.m. and 6 p.m. to 9 p.m.
— Washing of vehicles, boats, and buildings was restricted.
• Condition 3
— Total ban on lawn watering
— Monthly limits on residential water use, ranging from 6,000 gal per
month for houses with 1 or 2 occupants to 12,000 for 11 or more
occupants. Equivalent restrictions were placed on apartments and
mobile home parks. Commercial and Industrial customers were limited
to 75 percent of their normal monthly use. For residential customers,
use above the limit was surcharged $3/1,000 gallons month for the first
1,000, increasing to $25/1,000 for each 1,000 gallons per month
exceeding the limit by 3,000 gal up to an absolute maximum. Use above
the absolute maximum resulted in termination of service, fines, and
very likely fines for violation of outdoor watering restrictions. As
reservoir levels dropped, the city provided extensive publicity about the
problem and the conservation program using all available means of
communication (Shaw and Maidment 1988).
Results. The statistical analysis showed no effect from the voluntary watering
restrictions. When Condition 2 restrictions were imposed, however, water use was
31 percent less than what it would have been with no shortages. Under Condition 3
restrictions, water use was 39 percent less than its expected normal level. When
restrictions were removed, water use remained about 20 percent less than its
predrought level. Corpus Christi base water use (low irrigation season) is about
86 percent of total water deliveries because of the heavy concentration of petrochem-
ical plants. Many large commercial operations and industrial plants made extensive
capital investments in water-saving technology to meet the 25 percent mandatory
Condition 3 restrictions. Residential customers may have also contributed to this
decrease by adjusting their water-use practices.
Iowa, 1977.
• Penalty water rates for excessive use
• Coordination of programs among utilities
Background and program. Iowa experienced a drought in which some commu-
nities had their wells go dry, yet others experienced only modest supply reductions.
The populations of the communities in this research ranged from 700 to 29,000, and
some of the small communities were within commuting distance of metropolitan
areas. Two communities imposed mandatory restrictions of 30 and 40 gpcd with
severe penalty rates of 5–10 times normal marginal price for excess use. One district
set restrictions based on prior year winter use, also with a high penalty for excess use
and a $25 fine for violating the outdoor water-use ban. The remaining five
communities used only voluntary restrictions. Lee and Warren (1981) measured
conservation effectiveness by constructing an econometric model of what water use
would have been in the summer of 1997 without the drought and associated
conservation programs. Conservation scores were then constructed with the following
formula, which shows the number of standard deviations by which water use
changed. The advantage of this measure is that it incorporates the estimated change in
water use and its typical variability in a single measure.
Dj
CS j = ----- (Eq 11-2)
S
where:
CSj = the conservation score of the jth observation
Dj = actual minus predicted normal consumption
S = estimated standard error of predicted value of water use.
Results. Mandatory restrictions on a per capita basis with severe penalty rates
were highly effective, reducing water use by 34 to 59 percent with a conservation score
of 2.9 to 5.4 compared to the previous nondrought summer. These reported
percentage drops in use are raw numbers; the scores are based on estimated usage
taking into account differences in the weather and other factors. The differences
between these raw percentages and the normalized score values reflect the need for
more than raw numbers when analyzing conservation programs.
Mandatory restrictions based on prior year use resulted in a 1 to 14 percent
drop in use with a score of 1.9. The score means that water use in this community was
almost two standard deviations less than its normal level. One community with a
voluntary program that was close to a community where wells went dry reduced water
use by 21 to 27 percent with a conservation score of 2.2. In contrast, three
communities with voluntary programs did not achieve any significant drop in usage.
Athens, Greece, 1990.

• Effectiveness of rate increase
Background and program. By the summer of 1990, a combination of a prolonged
dry season, inadequate investment in infrastructure, and low water prices resulted in
a projected shortage of water. Overall water rates were increased substantially, and the
rate structure was changed to increasing block rates. A publicity campaign, which
included numerous news items, displayed frequent pictures of nearly empty reser-
voirs. Consumers were given plastic bricks for toilet tank displacement, and leaking
water pipes were fixed.
Results. The result was a drop in monthly water sales of about 25 percent, most
of which is attributed to the increased water rates and the shift to increasing block
rates (Briassoulis 1994).
Santa Barbara and Goleta, California, 1988–1992.
• Effectiveness of rate increase
• Price elasticity of demand by income level of customers
Background and program. Many parts of California and the West experienced
drought conditions of varying intensity between 1988 and 1992. In August 1988,
Santa Barbara offered free showerheads and toilet rebates. In June 1989, they changed
from a uniform marginal price to increasing block rates. In 1990, they made the
schedule much steeper, and banned outdoor water use except for drip irrigation and
hand-held hoses. Water police enforced outdoor rules.
Adjacent Goleta started offering toilet rebates in 1986, and in 1987, increased
the rebates and added showerheads. As the drought became increasingly severe, they
imposed an allocation program, limiting use to 132 Ccf per year plus 55 percent of
average base year use more than 132 Ccf. In July 1990, water rates were raised, and a
uniform marginal price replaced increasing block rates. Renwick and Archibald
(1998) developed a three-equation two-stage regression model of water use to
estimate the impacts of these programs.
Results. Goleta’s allocation program was estimated to reduce water use by
4.58 Ccf (out of 16.26 Ccf) per month, or 28 percent. Santa Barbara’s ban on most
exterior water use lowered water use by 4.37 Ccf (out of 27.5) per month, or 16
percent. For the entire sample, installation of a new toilet reduced household use by
10 percent and a new showerhead reduced it by 8 percent. Raising prices also reduced
demand. Among high-income households, the estimated price elasticity of demand
was –0.11, almost doubled to –0.21 for middle income, and jumped to –0.53 for low-
income homes.
Austin, Texas, 1984–1985.
• Trigger points to impose restrictions
• Declining effectiveness in second year
Background and program. Austin failed to expand their water treatment,
distribution, and wastewater systems to keep up with growth, resulting in water
shortages. In response to impending water shortages in the 1980s, the following
programs were initiated.
• Stage 1: Voluntary compliance with Stage 2 restrictions (May 1 to September
30).
• Stage 2: Mandatory compliance (imposed if total water use was more than
150 mgd for three consecutive days or 157 mgd for one day):
— Hand-held hoses, buckets, or drip irrigation allowed at any time.

Irrigation with sprinklers allowed once every five days between 7 p.m.
and noon on a rotating schedule.
— Noncommercial car washing and swimming pool filling allowed only on
same days and hours.
— Restrictions were imposed on ornamental fountains, golf course water-
ing, use of water from fire hydrants, and spraying of foundations.
• Stage 3 (as revised): Mandatory compliance (water use greater than 157 mgd
for three consecutive days or 165 mgd for one day). Same as Stage 2 except
for the following:
— Individual sprinklers or sprinkler systems were allowed on designated
days only from 6 a.m. to 10 a.m. and 7 p.m. to 10 p.m.
— Hand-held hoses, buckets, and drip irrigation were allowed any day but
only during the hours of 6 a.m. to 10 a.m. and 7 p.m. to 10 p.m.
Stage 3 was terminated when the triggering conditions ceased to exist for at
least five consecutive days.
Results. Residents were given three days to prepare for Stage 2 restrictions. There
is some evidence that residents increased lawn watering during the three-day period
between when restrictions were announced and when they went into effect. The
rotating schedule was modified after the first year to even out water use. Voluntary
restrictions reduced water use by 2.77 mgd the first time they were used, and by
1.27 mgd the second time. Mandatory restrictions in 1984 reduced peak water use by
13.5 mgd (out of 170 mgd) and by 5.5 mgd in 1985. It is unclear if the reduced impact
of these restrictions was caused by higher temperatures in 1985 (which were included
in the model) or by falling compliance rates. Shaw and Maidment (1987) did not
mention enforcement activity.
Honolulu, Oahu, 1977–1979, 1984–1985.
• Price increase versus restrictions
Background and program. The first drought episode (1977–1979) was met with
voluntary conservation programs, with heavy emphasis on education and publicity.
During the second episode (1984–1985), mandatory restrictions were added. Moncur
(1987, 1989) developed simple regression models of water demand during these time
periods using a sample of single-family residential water customers. Unfortunately, he
was unable to differentiate the characteristics of the various conservation programs.
Results. The regression models showed a water-use reduction in response to
conservation programs, but the analysis presented focuses primarily on the potential
role of price increases to reduce demand. Real (inflation-adjusted) prices fell steadily
from 1972 through 1977, and increased significantly during 1978 and 1979. By 1983,
although nominal prices had increased between 1978 and 1982 by 127 percent, prices
had drifted back downward to their 1972 level in real terms. These price variations
allowed Moncur to estimate a price elasticity of demand of –0.345, implying that a

10 percent reduction in water use could be achieved with a price increase of about
29 percent. Based on the results of voluntary conservation programs during the
drought, the same reduction could be achieved with these programs and a smaller
price increase of 25 percent (Moncur, 1987). Thus the nonprice conservation
programs had a rather limited impact.
Utah, 1976–1977.
• Ineffectiveness of voluntary program, time of day restrictions
• Effectiveness of penalty rates
Background and program. Communities in Utah experienced a drought during

1976–1977, and because they depended heavily on surface storage, undertook a
variety of conservation programs, ranging from higher prices to restrictions limiting
times or amounts of irrigation to aggregate use limits. Narayanan, Larson, and
Hughes (1985) estimated a linear regression model using data from 33 communities.
Results. A general price increase of 1 percent reduced water use by only
0.1 percent, an elasticity of –0.1. The author suggests that the price elasticity may be
lower during a drought. Price levels in these communities were generally quite low,
however, and even with the increases, water bills may not have been sufficiently large
to attract much attention. They found that requests for voluntary restraint had little
or no effect, and time of day restrictions on watering had small effects. However, one
community that imposed a charge of $10/1,000 gal for use above a mandatory limit
experienced a 50 percent drop in water use.
Blacksburg, Virginia, 1973.

• Nonsymetric price response
Background and program. A new townhouse development changed water

suppliers two years after opening, with the result that the marginal cost of water
increased from $1.728/1,000 gal per month to $2.205. For many owners, this made
water bills larger than heating bills. Townhouse owners received extensive publicity
about the rate change ahead of the change. After 10 months at the higher rates,
marginal price was reduced to $1.26/1,000 gal.
Results. Water use fell significantly with the rate increase. Hogarty and MacKay
(1975) computed the price elasticity for each owner for each available period for the
rate increase and later decrease. The average price elasticity for the periods of increase
was –0.71. During the following year, when prices were lower, there was no rebound
in water use. The authors attributed the drop in water use to leak repairs, although
this seems strange since the units were only a few years old. Residents may not have
given any attention to their water-use habits before the rate increase and found that
they could cut back with no significant loss of satisfaction.
SUMMARY
This chapter provides a comprehensive review of short term conservation programs,
and the impact on water use and utility revenue that can be anticipated from various
program components. The wide dispersion of realized conservation results signals the
importance of a careful evaluation of local conditions and preferences as well as
continual monitoring of programs once they are implemented. Perhaps the most
significant conclusion of this research is that short-term conservation goals can be
achieved if local officials remain flexible by adjusting their programs as they gain
experience and as the emergency situation unfolds.
REFERENCES
Briassoulis, H. 1994. Effectiveness of Water Conservation Measures in Greater Athens Area. Journal of
Bruvold, W.H. 1979. Residential Response to Urban Drought in Central California. Water Resources
Research 15(6): 1297–1304.
Chestnutt, T.W., J.A. Beecher, P.C. Mann, D.M Clark, W.M. Hanemann, G.A. Raftelis, C.N. McSpadden,
et al. 1997. Designing, Evaluating, and Implementing Conservation Rate Structures. Sacramento:
California Urban Water Conservation Council.
The Cross Section 52(8) August 2006, p. 1.
Hoffman, M., R. Blickstein, and S. Linoff. 1979. Urban Drought in the San Francisco Bay Area: A Study
of Institutional and Social Resiliency. Journal AWWA 71(7): 356–363.
Hogarty, T.F., and R.J. MacKay. 1975. The Impact of Large Temporary Rate Changes on Residential Water
Use. Water Resources Research 11(6): 791–794.
Kenney, D.S., M.P. Clark, and R.A. Klein. 2004. Use and Effectiveness of Municipal Water Restrictions
During Drought in Colorado. Journal of American Water Resources Association 40(Feb.): 77–87.
Lee, M.Y., and R.D. Warren. 1981. Use of a Predictive Model in Evaluating Water Consumption and
Conservation. Water Resources Bulletin 17(6): 948–955.
Maddaus, W.O., and L.A. Maddaus. 2006. Water Conservation Programs—A Planning Manual. AWWA
M52. Denver, Colo.: AWWA.
Mayer, P.W., W.B. DeOreo, E.M. Opitz, J.C. Kiefer, W.Y Davis, B. Dziegielewski, and J.O. Nelson. 1999.
Residential End Uses of Water. Denver, Colo.: Awwa Research Foundation (AwwaRF).
Moncur, J.E.T. 1987. Urban Water Pricing and Drought Management. Water Resources Research 23(3):
393–398.
———. 1989. Drought Episodes Management: The Role of Price. Water Resources Bulletin 25(3): 499–
505.
Narayanan, R., D.T. Larson, and T.C. Hughes. 1985. Effectiveness of Drought Policies for Municipal
Water Management. Water Resources Bulletin 21(3): 407–416.
Pint, E.M. 1999. Household Response to Increasing Water Rates During the California Drought. Land
Economics 75(2): 246–266.
Renwick, M.E., and S.O. Archibald. 1998. Demand Side Management Policies for Residential Water Use:
Shaw, D.T., R.T. Henderson, and M.E. Cardona. 1992. Urban Drought Response in Southern California,
1990–1991. Journal AWWA 84(10): 34–41.
Shaw, D.T., and D.R. Maidment. 1987. Intervention Analysis of Water Use Restrictions: Austin, Texas.
Water Resources Bulletin 23(6): 1037–1046.
———. 1988. Effects of Conservation on Daily Water Use. Journal AWWA 80(9): 71–77.
Tabors Caramanis and Associates. 1994. Long-Term Water Conservation and Shortage Management
Practices: Planning that Includes Demand Hardening. Sacramento: California Urban Water
Agencies.
TCEQ (Texas Commission on Environmental Quality). 2006. Handbook for Drought Contingency
Planning for Retail Public Water Suppliers. Austin: TCEQ.
———. 2006. Handbook for Drought Contingency Planning for Wholesale Public Water Suppliers. Austin:
TCEQ.
———. 2006. Handbook for Drought Contingency Planning for Irrigation Districts. Austin: TCEQ.
———. 2006. Model Drought Contingency Plan for the Investor Owned Utility. Austin: TCEQ.
———. 2006. Model Drought Contingency Plan for the Water Supply Corporation. Austin: TCEQ.
CHAPTER 12
Forecasting With Regression
Unit use coefficient models are an easy-to-apply forecasting technique, but utility
decision makers frequently can benefit from more detailed approaches. This chapter
discusses key steps in developing regression models that can be used in specialized
studies and in forecasting water use.
These models are based on common-sense theories of cause and effect about
the relation between water sales and various underlying factors or drivers of water use.
They are sometimes referred to as “causal/structural” models. They offer the
opportunity to explore the effect of alternative assumptions and changing local
conditions on the water use or demand forecast. They give management a valuable
perspective on how future circumstances, including changing business conditions,
climatic shifts, or new government policies, may affect the water system under various
sets of assumptions.
Regression models describe the essential aspects of how the world works. Such
models depend on the ability of the analyst to determine the relationships among
variables and also on being able to find or estimate future values of such independent
variables as personal income, the real price of water, population, employment by
major industry group, and the use of various water-related fixtures.
Additionally, obtaining meaningful results requires substantial variation in the
values of each variable. For example, if water rates have not changed during a 10-year
study period, the price elasticity of demand cannot be estimated except insofar as
consumers respond to the erosion of the real cost of water by overall inflation.
The first major section in this chapter focuses on general regression modeling.
Topics include (1) the dependent variable, explanatory variables, and functional form
of the regression; (2) the interpretation of regression coefficients; (3) the error term of
the regression; (4) diagnostic statistics, including standard errors and t statistics for
each coefficient, the F statistic for the significance of the regression as a whole, and R2
253
and standard error of the estimate describing goodness of fit, and (5) nonlinear
regression models. The discussion is intended as a quick introduction or refresher.
The second major section presents a case study of a cross-section regression
estimated with household data. The discussion follows the procedures outlined in the
first section, adding materials on sample data and using regression equations in
forecasting. The discussion also considers exploratory data analysis through inspec-
tion of means, standard deviations, and correlations of the explanatory variables with
each other (“correlation matrix”).
The data in this causal/structural regression model are a cross-sectional sample
of observations on the water use and socioeconomic characteristics of individual
residential customers, derived as a random sample from utility billing records. These
data, available on the accompanying CD, sometimes are called microdata, distinguish-
ing them from aggregate water-use data by customer categories.
The final major section focuses on general issues relating to the validity of
regression models, and their relevance to selecting the set of explanatory variables to
be used in the regression. This section briefly describes some problems that can arise
in ordinary least squares regressions, including serial correlation, heteroscedasticity,
and multicollinearity. This section also considers “antidotes” to these “pathologies” of
classical linear regression.
This chapter expands on the discussion in chapter 6, where the estimation of
weather influences within causal/structural water-demand models is considered.
REGRESSION MODELING—GENERAL CONCEPTS

The outset of a regression analysis involves three basic questions:
1. What is the dependent variable?
2. What are the explanatory variables?
3. What is the functional form of the regression?
After answering these questions, the regression is estimated, producing esti-
mates of the values of the coefficients of the explanatory variables and other
information valuable in diagnostic testing. The goodness of fit of the regression to the
data is a primary number used in evaluating the regression. Others diagnostic tests
relate to the statistical significance of the individual coefficients. And tests of adequacy
relate to the residuals of the regression.
A regression relationship for monthly water use of the average residential
customer as the dependent variable can be written as:
Q = β0 + β1 X1 + β1 X1 + β2 X2 + β3 X3 + β4 X4 + β5 X5 + E (Eq 12-1)
where:
Q = water use each month
β 0 = the constant or intercept term of the model
FORECASTING WITH REGRESSION 255
β1, β2, β3, β4, β5 = estimated slope coefficients

X1, X2, X3, X4, X5 are explanatory variables (e.g., price of water)
E = the residual error of the estimate.
1
The functional form of this estimating equation is strictly linear. The variable
on the left-hand side of the equation is the target variable we are attempting to explain
and is referred to as the dependent variable. The variables on the right-hand side of the
equation are called the independent, explanatory, exogenous, or regressor variables.
Taken together, these explanatory variables are called the specification of the model.
The literature also refers to specifying variables in a regression, which means that
these variables are the explanatory variables.
The computer program used for solving the linear regression problem
determines the β values so that the sum of squared residual errors of the estimation is
minimized. This equation says that the true value of water use is equal to β0 plus a
weighted linear sum of the values of X1, X2, X3, X4, and X5, plus a random error or
disturbance term.
Interpretation of Regression Coefficients. The coefficient for each vari-

able shows the change in the dependent variable for a one-unit change in that
independent variable, holding all other independent variables constant. Each regres-
sion coefficient can be thought of as the partial derivative of the dependent variable
with respect to that specific explanatory variable. For example, a coefficient of –3.0 on
the independent variable X1, when X1 represents price in dollars, means that a $1.00
increase in price reduces water use by three units per month. The size of the
coefficient depends on the way the variable is defined. Changing the unit of
measurement of X1 from dollars to cents simply reduces the price coefficient to –0.03,
but does not change the result, since this means that a 1 cent increase in price reduces
water use by 0.03 units per month.
The Error Term of the Regression. It is probably not widely recognized

outside of econometrics, but it is the assumptions regarding the error term of the
regression that provide the basis for the standard statistical interpretation and
diagnostics in regression analysis.
The basic concept is extremely simple. In explaining phenomena in the real
world, it is seldom possible to account for all the influences. In the dry language of
regression analysis, “The independent variables of a regression equation are seldom
able to include all influences on the dependent variable.” Thus, the effects of
numerous possible explanatory variables, each of which is too minor to include in the
specification of the regression, or for which there are no usable data, create a residual
or error term, denoted as E in Equation 12-1. Possible errors in measuring the
variables of the model also contribute to this error term E.
1. Note that if we take the logarithm of both sides of the exponential equation Q = β0X1β1 X2β2 X3β3
X4 X5β5 E, we transform a multiplicative relationship into a linear form like Equation 12-1.
β4
The next step is to assume that this error process (the values of E for each
observation) is generated by or follows a normal probability distribution—the
famous bell-shaped curve. The basic motivation is the Central Limit Theorem of
statistics, which asserts that the mean value of a series of random variables tends
toward a normal distribution as the number of variables in this sum increases. This is
true, given certain broad conditions, even when the individual random variables are
not themselves normally distributed. Therefore, the combined effect of innumerable
small influences omitted from the regression specification will tend to be described by
a normal probability distribution.
Diagnostics Statistics. Typical diagnostic tests for regressions include:
• Standard errors and t statistics for each coefficient
• F statistic for the significance of the regression as a whole
2
• R and standard error of the estimate describing goodness of fit
Standard errors and t-statistics. The standard error for an estimated coefficient
provides an indication of the likely difference between the estimated β value and the
unknown true value. For a normal distribution and a data set of more than 200
observations, there is a 32 percent chance that the true value of the coefficient is more
than ± one standard error from its estimated value. For a two-standard error interval,
the chance drops to only 4.6 percent, and for a three standard error interval, to less
than 0.5 percent.
The t statistic is the coefficient divided by its standard error. Thus, if the
standard error for the coefficient of X1 in Equation 12-1 is 1.3 when the estimated
coefficient is –3.0, the t statistic is –3.0/1.3 = –2.30. The estimated coefficient and its t
statistic always have the same sign. As a rule of thumb, a coefficient is likely to be
accepted as significantly different from 0 if the absolute value of the t statistic is 2 or
larger, or equivalently, if the estimated coefficient is at least twice as large as its
standard error.
The significance level for regressions using less than about 200 observations
depends on both the computed t statistic and the degrees of freedom of the estimated
regression. Degrees of freedom is equal to the number of observations in the data set
minus the number of estimated coefficients, including the constant. Using the degrees
of freedom and the computed t statistic, an exact statement of significance can be
obtained from a table of values for the t distribution.
The example of Equation 12-1 has six estimated β coefficients. Assuming 48
observations in the data set, this leaves 48 – 6 = 42 degrees of freedom. From the table
of values for the t distribution, found in the back of most statistics books, the critical
value for 95 percent significance for a two-tailed test with 40 degrees of freedom is
2.021. At the 98 percent level, the critical value of the t statistic is 2.423. Thus, we
could reject the null hypothesis that this coefficient is actually equal to 0 if we accept
a 5 percent chance of being wrong, since it has an estimated t of –2.3. The null
hypothesis, however, cannot be rejected if we demand a higher standard of only a 2
percent chance.
A two-tailed test is used to test the null hypothesis that the true but unknown
coefficient may equal to 0 against the alternative hypothesis that the coefficient is not
0 but could be either positive or negative. Statistical tests of significance require the
statement of a null hypothesis and an alternative hypothesis. A statistic is then
computed that results in either being unable to reject the null hypothesis (in effect
accepting it as true) or rejecting the null hypothesis (in effect accepting the alternative
as true).
For the coefficient of X1 in Equation 12-1, however, the null hypothesis is that
the coefficient is 0 or larger. The alternative hypothesis is that the coefficient is less
than 0, that is, has a negative value, showing that higher prices reduce water use.
Therefore, a one-tailed test is used, making the critical value for a 95 percent level (all
in one tail of the t distribution) equal to 1.684. In this example, therefore, the critical
t from the table of 2.021 represents the 0.975 level—a 1 in 25 chance of rejecting a
true null hypothesis. Variables with insignificant coefficients are often dropped from
the model.
Confidence interval. A confidence interval is an interval around a coefficient
estimate that, at a given probability, can be said to contain the true value of the
parameter. Assume a 95 percent confidence interval is to be constructed for an
estimated coefficient β with a standard error of S. The 95 percent confidence interval
is said to contain the true value of the estimated coefficient with a 19 in 20 odds ratio,
i.e., 95 percent of the time. Then, if enough observations (more than 200) have been
used to estimate the regression, a table of probabilities for the t distribution indicates
a critical value of the t statistic of 1.96 for a 95 percent confidence interval. Therefore,
the confidence interval will be:
β – ( 1.96S ) ≤ ( trueβ 2 ) ≤ β + ( 1.96S ) (Eq 12-2)
For example, if the estimated coefficient of X2 in Equation 12-1 is 0.071, with

a standard error of 0.024, and using the critical t statistic of 1.96, the confidence
interval is:
0.071 – (1.96) × (0.024) ≤ true β2 ≤ 0.071 + (1.96) × (0.024)
0.024 ≤ true β2 ≤0.118
This assumes a large sample. If the sample size is smaller, so that degrees of
freedom is 42, the critical value from the t table is 2.021 and the confidence interval
becomes 0.0225 ≤ true β2 ≤ 0.1195.
2
R and the F statistic. Is the equation itself significant? The F statistic shows
whether or not the entire equation is significantly better than the mean value of the
dependent variable in explaining water use. An equation would generally be accepted
if it is significant at the 0.05 level or better, meaning that there is a 5 percent chance
that the equation does not explain variations in the dependent variable better than its
mean value. Since the F statistic is adjusted for degrees of freedom, dropping variables
with low t statistics from the equation may improve this statistic and make an
otherwise unacceptable equation usable.
2
The R shows the proportion of actual minus mean water use explained by the
2
model. Generally, a higher R indicates a more effective model, with a value of 1,
which is generally unobtainable, indicating a perfect relationship. For a time series
2
model, such as one using monthly or annual water-use data, R values between 0.7
and 0.9 are considered to be quite good. For a cross-section model, however, such as
500 individual water customers’ water use during a single summer, the R2 is typically
between 0.20 and 0.50, because there are many factors that cannot be measured
influencing individual behavior.
Nonlinear Relationships. The forecast equation need not be the simple
linear form of Equation 12-1. To apply standard regression procedures, however, it
must be possible to transform the forecast equation into a linear form. There are
several ways to do this. Many water-demand models, including the often-cited Howe
and Linaweaver (1967) study, present demand or water use as the product of a
number of influences, as expressed in the equation
Qi = eβ0 Rainβ1 Temperatureβ 2 Priceβ3 Incomeβ 4 Hhsizeβ5 (Eq 12-3)
where:
Qi = quantity of water used
Rain = warm weather precipitation
Temperature = average temperature
Price = price per unit of water
Income = household income
Hhsize = number of persons in the household
e = the base of the natural logarithms
βi are estimated exponents.
Written this way, as a multiplicative form, the model is nonlinear. It can,

however, be transformed so that the parameters (the betas) become linear by taking
the natural logarithms of both sides, and noting that ln(e) = 1, where ln stands for the
natural logarithm. In this transformation, Equation 12-3 becomes
ln(Qi) = β0 + β1 ln(Rain) + β2 ln(Temperature) + β3 ln(Price) +

β4 ln(Income) + β5 ln(Hhsize) (Eq 12-4)
Equation 12-4 is just like Equation 12-3, except that the data on the explanatory
and dependent variables are subject to a logarithmic transformation. Equation 12-4 is
linear in the parameters βi and can be estimated with ordinary regression.
The payoff in transforming to an estimating equation that is linear in the
parameters is that the well-known formulas and procedures of regression analysis can
be directly applied. By contrast, when the forecast relationship is nonlinear in the
parameters, methods of successive approximation must be used to estimate the values
of parameters. Such approximation methods are considerably more complex from a

mathematical standpoint. Many econometric computer programs include options for
estimating nonlinear relationships.
A second nonlinear form has been used in some water-demand forecasting
models such as that of Foster and Beattie (1979). A salient feature of this form is that
elasticities computed from these estimated coefficients vary in size with the average
values of the variables themselves. For example, the price elasticity is the estimated
coefficient on price times price, so higher prices imply greater elasticity. In these
models, the initial equation form and its logarithmic transformation are
β0 +β1Rain+ β2Temperature+ β3Price+ β4Income+ β5Hhsize

Q= e (Eq 12-5)
LnQ = β0 + β1Rain + β2Temperature + β3Price + β4Income + β5Hhsize
ESTIMATING A REGRESSION MODEL WITH HOUSEHOLD DATA

Developing a causal/structural model of water use per household allows a utility to
answer such questions as: What if the price of water falls 15 percent, in real terms,
because of inflation? What if the proportion of households with water-conserving
fixtures rises by 7 percent?
Answers to such questions are available by substituting the new values of the
specified independent variables into the model and recalculating the result. Hypothet-
ical situations can be analyzed quickly. If the changes imposed on the model are not
too large, the analyst can be relatively confident of these what-if results.
Sample Size. Suppose two urban water utilities plan a joint effort to collect
data from a random sample of single-family households to develop a demand
forecasting model. The required sample size depends to an extent on the unknown
characteristics of the population and on the objective of the research. This means that,
initially, sample size estimation must begin with a guess.
If the utilities wish to simply estimate and forecast average water use, the
required sample size is obtained from the formula:
2 2 2
n = (t × σ ) / d (Eq 12-6)
where:
t = the t statistic showing the level of accuracy desired
σ = the population standard deviation
d = the margin of error the utility is willing to accept.
For accuracy of ± 5 percent and a mean water use of 180,000 gal/year,

d = 5 percent of 180,000, or 9,000 gal. If the utilities are willing to take a 1 in 20
chance of a larger error, a t of 2 can be chosen. Utility records of water consumption
along with informed judgment suggest that the population standard deviation is
2
90,000 gal. Then, the sample size required to achieve a tolerance of ± 5 percent is (2
2 2
× 90,000 )/9,000 = 400.
Half of the sample would be randomly selected from the service area of each
utility. For more accuracy, a larger sample would be used.
If the utilities planned to estimate other variables such as the proportion of the
population with low-flow fixtures, a larger sample might be required. The required
sample size is given by (t2 × P × Q)/d2, where P is the proportion of the population
with the fixtures, Q is the proportion without, and d is the margin of error. The
largest sample size is required if P and Q are each 50 percent. Given an acceptable
error of 5 percent, the required sample size is: 22 × 0.5× 0.5 ÷ 0.052 = 400. If the
results of the survey were to be disaggregated based on a variable obtained from the
survey, such as income level, the required sample size is much larger to obtain
significant results for each subgroup.
Based on these calculations, a sample size of 400 single-family residents is
assumed to be randomly drawn from utility billing records. Residences with no water
use in any month are excluded, since they are presumably vacant. This accounts for
the absence of any water use values close to 0. Note that this sample size does not
ensure the regression model will produce accurate estimates—only that the cases
sampled from utility records should present a representative picture of the water use
of the population, i.e., the residential customers of these utilities.
The initial sample size would be increased to offset customers who did not
respond to the survey. If a 35 percent response rate was anticipated, the sample size
would be set at 1,143. Appendix A, Survey Methods for Direct Data Collection,
presents more extensive information on survey techniques.
Model Specification—Choice of Variables. It is imperative to have a
sound conceptual basis for the variables in the model. The variables in this example
were chosen because they have been used in numerous published water-demand
models. Because the data is cross-sectional, weather factors are relevant only insofar as
there is microclimatic variation in temperature and precipitation in the area sampled
for the analysis. Weather stations are usually too sparsely located to permit collection
of such data within an urban area.
Water rates are much higher in one district than the other, allowing for an
assessment of the effect of price on water use. This makes the resulting relationship a
full-fledged demand function, as opposed to merely a water-use relationship. Annual
household water use is the dependent variable.
For this example, a database representing two geographically contiguous water
districts is synthesized from empirical data available to the authors. Residential
customers in these districts have similar socioeconomic profiles.
The variables in the sample database are
• Quantity = annual water use per household, in 1,000-gal units
• Price = marginal price of water per 1,000 gal
• Billfixed = fixed service charge for water for the year
• Income = household income per year, in thousands of dollars

• Hhsize = number of people in the household
• Pool = swimming pool (1 if present, 0 otherwise)
• Conservation = indoor water-conserving fixtures (1 if present, 0 otherwise)
• Lot size = lot size, in thousands of square feet
The water-rate schedule is similar to that in Figure 9-2 with a fixed monthly
service charge of $3.00 and a constant marginal price regardless of the quantity used.
The marginal price in District 1 is $2.00/1,000 gal, while the marginal price of water
in District 2 is $2.35/1,000 gal. The marginal price of water represents the cost to the
consumer of purchasing additional quantities of water. Given this type of rate
schedule, the price is the same for all customers in each district.
Water, like other goods and services, follows the law of demand—less is
purchased at higher prices—and thus the estimated coefficient on Price is expected to
be negative.
Because the monthly service charge is constant, it has the same effect as a small
decrease in household income. Thus the monthly service charge (Billfixed) is
subtracted from income (Income) and the resulting variable Income✻ = (income – $36)
is used in the model. Because the differences between Income and Income✻ are so
minor from a numerical standpoint, we make no additional distinction in discussion
for this adjustment. So refer to the adjusted value is simply “Income.”
Water use per household is known to increase with household size, so the
2
expected sign on Hhsize is positive. Water use is likely to increase with lot size due to
summer sprinkling demand.
Water use is also likely to increase with the presence of a swimming pool
because of both evaporation and backwashing requirements. If a swimming pool
replaces irrigated vegetation, however, its effect on water use may be quite small.
Microlevel studies of actual water use with real-time logging, such as that conducted
by Mayer and colleagues (1999) under the auspices of the Awwa Research Foundation
(AwwaRF), might be able to answer this question of whether pools or turf use more
water.
To illustrate the effect of conservation measures on water use, a composite
dummy variable representing indoor water-conserving fixtures is specified in the
model. Because this composite variable equals 1 when indoor water-conserving
fixtures are present, the sign on its estimated coefficient is expected to be negative. In
a more complete model, a variable might be included for each type of fixture
individually.
2. There are efficiencies in water use per resident on a household basis. This is, in part, because lot
size and outdoor residential water use usually increase less than proportionally to the number of
household residents. Thus, a single three-member household generally uses less water than three one-
member households. There are also economies of scale in clothes and dishwashing.
Exploratory Data Analysis. Before developing the regression model, the

analyst must carefully examine the data collected from the sample survey and from
utility records. Means, standard deviations, and other descriptive statistics of the
dependent and explanatory variables (listed in Table 12-1) should be computed.
Table 12–1 Descriptive Statistics
Std.
Mean Median Maximum Minimum Dev.
Quantity 205.94 191.51 590.27 67.16 83.55
Price 2.19 2.35 2.35 2.00 0.17
Income 23.27 20.69 83.88 4.75 11.94
Hhsize* 1.92 2.00 5.00 1.00 0.82
Pool 0.08 0.00 1.00 0.00 0.27
Conserve 0.36 0.00 1.00 0.00 0.48
Lot size 9.61 8.71 32.23 4.01 4.79
*Number of persons in household.
Figure 12-1 shows the frequency distribution of the sample observations on

water use. This curve is skewed, or left-steep, with a long tail of observations at high
water-use levels characteristic of a lognormal distribution. Skewness is also shown by
the mean of 205.94 exceeding the median of 191.51, and the maximum value of 590
being much further from the mean than the minimum value of 67. The presence of
skewness suggests a multiplicative form for the water-demand model.
Correlation matrixes in Table 12-2 show the sign and size of the simple linear
relationship between each pair of variables. The top set of numbers are correlations of
the raw data, and the bottom panel of numbers are the correlations of the log-
transformed data. The diagonal entries in each of these triangular arrays equals 1
because any variable is perfectly correlated with itself.
Although several of the correlation coefficients change somewhat with the log
transformation, all of their signs remain the same. The highest simple correlation is
between lot size and income, as shown in the third column of the table. Both income
and lot size are also strongly correlated with quantity. This suggests the possibility of
multicollinearity between household income and lot size if both are included in the
regression specification. Multicollinearity occurs when two or more independent
variables in a regression are highly correlated (vary together). (For further discussion
of multicollinearity, see chapter 13.) Household income has the higher simple
correlation with quantity, so it is the first choice for the regression specification. Thus,
lot size is omitted from the model.
Regression Equation. Exploratory data analysis suggests that the effects of
the explanatory variables operate in a multiplicative relationship to the dependent
variable—annual water use per household. Thus, the regression model is estimated by
taking natural logs of the dependent variable and each of the explanatory variables.
The dummy variables Pool and Conserve assume values of 1 or the base of the natural
140
120
Number of Customers
100
80
60
40
20
0
0– 50– 100– 150– 200– 250– 300– 350– 400– 450 500 550– 600 +
50 100 150 200 250 300 350 400 450 –500 –550 600
Water Use (1,000 gal/year)
Figure 12–1 Frequency Distribution of Water Use
Table 12–2 Correlation Matrix for Residential Microdata—Raw and Log-Transformed Versions
Raw Data
Quantity Price Income Hhsize* Pool Conserve Lot Size
Quantity 1.00
Price –0.07 1.00
Income 0.52 0.04 1.00
Hhsize 0.17 0.07 –0.04 1.00
Pool 0.34 0.04 0.71 –0.03 1.00
Conserve –0.05 0.05 -0.01 0.02 –0.08 1.00
Lot Size 0.46 0.03 0.93 –0.05 0.67 0.00 1.00
Log-Transformed Data
Quantity Price Income Hhsize Pool Conserve Lot Size
Quantity 1.00
Price –0.08 1.00
Income 0.56 0.01 1.00
Hhsize 0.22 0.05 –0.02 1.00
Pool 0.31 0.04 0.53 –0.01 1.00
Conserve –0.07 0.05 -0.01 0.03 –0.08 1.00
Lot Size 0.47 0.00 0.87 –0.03 0.54 0.01 1.00
*Number of persons in the household.

logarithms, e = 2.71828, rather than 0 and 1 in the original exponential model.

Therefore, after the logarithmic transformation, the values of these dummy variables
will be 0 or 1, respectively.
The results of the regression can be expressed in equation form as
ln(Quantity) = 4.2 – 0.43ln(Price) + .043ln(Income) + 0.21ln(Hhsize)

+ 0.02ln(Pool) – 0.05ln(Conserve) (Eq 12-7)
where:
ln = the natural logarithm of each respective variable
Other variables are as defined earlier.
This relationship is estimated with annual data for 400 households. Table 12-3
lists the regression results and key diagnostic statistics.
Table 12–3 Multiple Regression Results

Multiple R 0.61
R2 0.37
Adjusted R 2 0.37
Standard Error 0.31
Observations 400
ANOVA
Regression 5.00 22.28 4.46 47.22 0.00
Residual 394.00 37.18 0.09
Total 399.00 59.47
Standard Upper
Coefficients Error t Stat P-value Lower 95% 95%
Intercept 4.20 0.19 22.53 0.00 3.83 4.56
Price 0.43 0.19 –2.24 0.03 –0.81 –0.05
Income 0.43 0.04 11.75 0.00 0.35 0.50
Hhsize 0.21 0.04 5.83 0.00 0.14 0.28
Pool 0.02 0.07 0.36 0.72 –0.11 0.16
Conserve 0.05 0.03 –1.66 0.10 –0.12 0.01
NOTES: ANOVA = analysis of variance; DF = degrees of freedom; SS = sum of squares, used to compute R2;
MS = mean squares, used to compute F statistic; F statistic = test for the statistical significance of the
overall equation; Hhsize = number of persons in household.
The adjusted R2 of this regression indicates that the model explains 37 percent
2
of the variation of annual water use. This is often as high an R as can be expected
from cross-section microdata, because of the noise typically generated by individual
patterns of behavior. As a whole, the equation is highly significant with an F statistic
of 47.22.
Significance Levels of the Estimated Coefficients. The estimated coef-

ficients all have the expected signs. Furthermore, the coefficients on Price, Income, and
Hhsize are highly significant, as indicated by their standard errors and t statistics.
When Lot size is retained in the model, its coefficient has an unexpected negative sign.
This is a clear indication that it is indeed multicollinear with income, and the decision
3
to drop it from the model was correct.
The estimated coefficients on Pool and Conserve, unfortunately, are not
significantly different than 0 at the 2.5 percent level, as indicated by the inclusion of 0
in their 95 percent confidence intervals. The appropriate test of significance uses only
one tail of the t distribution, because we are testing the hypothesis that the presence
of a pool increases water use and conservation devices reduce it. If the hypothesis were
simply that water use would change, a two-tailed test would be appropriate. With a t
statistic of –1.66, Conserve is significant at the 0.05 level, and Pool is significant only at
0.40. One option is to believe that the sign of the estimated coefficient is correct, but
to accept the higher odds that you are in fact being misled. Thus, there is a 1 in 20
chance that Conserve has a positive coefficient, but more than a 1 in 3 chance that Pool
decreases water use. The analyst might in this instance reconsider the initial belief that
a pool with its usually paved surrounding deck uses more water than the turf it may
have replaced.
Swimming pools were, however, retained in the model. The other option is to
omit one or both of the relatively insignificant variables and re-estimate the equation
without them. Because the sign on the effects of these variables is confirmed by the
correlation matrix, leaving them in the specification is appropriate. The model can be
re-estimated dropping one or both of these variables using the data CD.
Strategies for improving the statistical reliability of estimates of conservation
effects are considered in chapter 10.
Coefficient Confidence Intervals. Because Equation 12-2 is in log-log
form, the coefficients are elasticities. This makes impact analysis and what-ifs
particularly easy. In addition, the standard errors of Price, Income, and Hhsize are
small enough (and the t statistics large enough) to provide informative confidence
intervals for these elasticity or impact estimates.
As an illustration, let us derive confidence bounds for the effect of a 15 percent
rate increase. Assume both utilities want to increase water rates to meet increases in
operating costs. Management is interested in exploring the probable impact on
residential water use of a 15 percent rate increase.
The price elasticity of demand is estimated at –0.43, which is relatively inelastic,
as expected. Its t statistic of –2.24 is significant at the 0.025 level, which means that
there is less than a 1 in 40 chance that the true value is equal to or larger than 0.
Other things being equal, this price elasticity suggests that a 15 percent increase
in Price would reduce residential water sales by (–0.43) × (0.15) = –6.45 percent, while
increasing utility revenue from sales to residences by 8.55 percent. Here, a handy rule
3. The regression including lot size is included on the data CD.

is that the price elasticity of revenue equals 1 plus the price elasticity of demand (see
chapter 9).
What are the confidence bounds on this impact estimate? From Table 12-3, the
standard error of the coefficient of Price is 0.19, so the 95 percent confidence interval
is computed as: –0.43 + (1.96 × 0.19) ≥ true price elasticity ≥ –0.43 – (1.96 × 0.19).
Thus the odds are only 1 in 20 that the residential price elasticity will not be
4
between –0.058 and –0.80.
The regression analysis, therefore, suggests that a 15 percent rate increase will
reduce water demand by between 1 and 12 percent while increasing revenues by
between 14 and 3 percent. The highest probability in this forecast is for the mean
value, and the chance of experiencing any particular value in this range decreases as
the value moves away from the mean.
The Income coefficient of 0.43 means that an increase in real (inflation-
adjusted) income per household of 10 percent can be expected to increase water sales
by 4.3 percent, presumably because higher income households invest in more water-
using features such as decorative fountains or Jacuzzis. An interesting exercise is to
re-estimate the model including Lot size instead of Income in the specification. The
reader could also explore the effect of either smaller or larger lot sizes in new
construction.
Single Household Forecast. The regression in Equation 12-1 provides
valuable structural information, but its forecasts of water use appear to be located in
an extremely wide band. Careful analysis, however, shows that although the
confidence intervals for forecasts of water use of specific residences—or a residence
having specific values for the explanatory variables—are plus or minus nearly
100 percent, confidence intervals for a forecast of aggregate residential water use
become tighter and tighter as more residences are included in the forecast. Although
we examine the forecast for a single household to better explain this process, the water
utility will seldom if ever be interested in what any one residential customer does. The
variability of forecasted values from a regression depend on:
• The variance of the regression or its goodness of fit to the data
• The degree to which projected values of explanatory variables deviate from
their mean values in the sample data
Thus, the most reliable forecast occurs when it is assumed that the explanatory
variables will take on values equal to their sample means. This is a “business as usual”
prediction, in which the only source of variability is the fit of the regression to the
sample data. For the regression of Equation 12-2, we can explore this steady-state
forecast of water use of an average residence as follows:
1. Determine the estimate of variance of the regression, which is usually
provided along with other regression results.
4. These values differ slightly from the confidence interval in Table 12-3 because of rounding error.
2. Compute the square root of the variance, to obtain the standard deviation,
which is also typically given on the printout of results.
3. Multiply the estimated regression coefficients by their associated explana-
tory variable values, which are equal to the mean values of the sample.
4. Build a 95 percent confidence interval around the resulting regression
estimate for the average sample values.
The standard approximation to the variance of the regression estimate is the
residual sum of squared errors divided by the regression degrees of freedom. In Table
12-3, this number is in the middle (analysis of variance [ANOVA]) panel in the row
titled “Residual” and column titled “MS.” The square root of the variance is the
standard deviation of the regression error process. Thus, √0.094372 is 0.3072, which
is shown as the standard error in the top panel.
This standard deviation (error) refers to the log-transformed data, so the
following discussion refers to averages of the log-transformed values of explanatory
variables. The sample average of the log-transformed water-use numbers in the
dataset is 5.328 = ln(205.94). Substituting the average values of the log-transformed
explanatory variables in the regression Equation 12-2 produces this same number
except for rounding errors.
Suppose that next year we anticipate that the average log-transformed values for
Price, Income, Hhsize, Pool, and Conserve will precisely equal their averages in these
sample data. Because there is considerable unexplained variation in the sample data,
it is possible that the average of the log-transformed water sales figures for next year
will be quite different than 5.328.
Given 400 observations, a 95 percent confidence interval is associated with a
t value of 1.96. Thus, the confidence bounds are 5.328 – (1.96 × 0.3072) = 4.726 =
ln(112.85) and 5.328 + (1.96 × 0.3072) = 5.930 = ln(376.16). Converting these
numbers into thousands of gallons, we ascertain that a 95 percent confidence interval
for the regression estimate of the average log-transformed water use in the sample is
between 113 and 376 thous gal. These 95 percentage confidence bounds are –55 to
+83 percent of the predicted point estimate of 206 thous gal. Thus, although the
regression rather precisely estimates effects for certain of the explanatory variables, it
does not forecast water use of an individual household having average values of the
independent variables with much accuracy at all.
Some analysts prefer linear rather than double logarithmic demand models.
When this same model is estimated in strictly linear form, the results, including the
confidence intervals, are similar to those found from the double logarithmic model.
The linear model can be estimated using the data on the CD.
Total Population Forecast. The confidence interval for a forecast for the
combined water use of all of the utility’s customers is dramatically smaller. For
computational simplicity, this example is developed for a strictly linear regression. For
a population of 100,000, the forecast of aggregate water use when each of the
independent variables of the regression equals its mean value from the sample is
100,000 × 205.94 thous gal = 20,594 mil gal. The variance of this forecast is
5
estimated by
2
2 S2 2 2 68.9 2
Var = K ----- + KS = 100, 000 × ------------ + 100, 000 × 68.9 (Eq 12-8)
N 400
where:
K = the forecast population = 100,000
S = the standard error of the regression = 68.9 thous gal, which is taken
to be the best estimate of σ, the population standard error
N = the number of observations used in the regression = 400.
Thus, the standard error of the forecasts, which is equal to the square root of the
forecast variance, is 345.188 mil gal. This yields a forecast confidence interval of
20,594 ± 345.2 × 1.96, or 20,594 ± 677 mil gal. Therefore, the predicted water use falls
between 19,917 and 21,271 mil gal, for a forecast range of ± 3.3 percent.
If one of the independent variables takes on a value different than its mean
value in the sample, the forecast variance includes an additional term that can result
in a wider interval. For example, suppose marginal water rates increase by 15 percent
from their mean value (P of $2.19 to P ✻ of $2.52). Using the price elasticity from the
log model of –0.43 yields a drop in water sales of –0.43 × 15 percent = 6.45 percent.
The new forecast of water sales per household is 192.66 thous gal, so aggregate use of
100,000 customers is 19,266 mil gal. The forecast variance now includes the following
additional term:
K 2 2 2
S 2 68.9
∑ ( P∗ – P ) × -------------------------------- = 33, 000 × ------------
N 11.53
(Eq 12-9)
i=1 2
∑ ( Pi – P )
i=1
2
Adding this to the previous forecast variance yields the new forecast variance, S
= 56,752,545 × 104, and the standard error of the forecast S = 753 mil gal. The
95 percent forecast confidence interval now becomes 19,266 ± 753 × 1.96, or 19,266 ±
1,477 mil gal, which is a forecast interval of 18,513 to 20,019 mil gal, or a range of
±3.9 percent.
If more than one independent variable takes on a value that differs from its
mean, the computation of the forecast variance becomes much more complex, unless
the correlation among the independent variables is 0. It is also difficult to estimate the
confidence interval for an aggregate forecast from a model with a nonlinear
5. This formula is derived and explained in detail in the first edition of this book. The derivation is
available from the authors on request.
dependent variable, such as the double log model, although the same principle
operates.
ESTIMATING A REGRESSION MODEL WITH TIME-SERIES DATA

Many utility analysts use time-series rather than cross-section data for developing
forecasting models. Equation 12-10 is an example of a demand model to be estimated
with time-series data that could be used for forecasting purposes.
Q = β0 + β1Price + β2Rain + β3Temp + β4Income + β5Hhsize + E (Eq 12-10)
where:
Q = average household water use for each month
β0 = the constant or intercept term of the model
β1, i = 1, 2, 3, 4, 5 = estimated slope coefficients
Price = marginal price per unit of water
Rain = monthly rainfall
Temp = average monthly high temperature
Income = average monthly income per household
Hhsize = number of persons in household
E = the residual error of the estimate.
This model is formulated to estimate the water use of a typical household.

Thus, a forecast of total residential water use requires multiplying the output of this
model by a projection of the number of households. Because the data used to estimate
this model consist of average values per household, the accuracy of any forecast using
these data depends on this model having a very low standard error of the estimate, or
2
somewhat equivalently, a very high R . The variance for a forecast of average
household water use is estimated by:
2
2 ( X∗ – X ) -
1 + --------------------------------
Var = S 1 + --- (Eq 12-11)
N N 2
∑ ( Xi – X )
i=1
where:
S2 = the variance of the regression estimate
N = the number of observations used in the estimating equation
X = the mean value of X over the data set
X✻ = the forecasted value of that independent variable
Xi, i = 1, 2, 3, . . . = observed values of an independent variable.
The standard error, S, of the forecast used in developing the confidence interval
is the square root of this variance.
When all the independent variables of the regression are forecasted to assume
their average values, the third term in the equation is 0. For example, suppose in the
2
regression S = 0.57 thous gal, N = 72, and Q = 13.42 thous gal/month. The forecast
2
variance = (0.57) × (1 + 1/72) = 0.3294 and the standard error of the forecast is =
0.574.
For a 95 percent confidence interval, the t statistic for a regression with 72
observations and 6 estimated coefficients (66 degrees of freedom) is 2.00. Thus the
confidence interval is 13.42 ± (0.574 × 2.00), or 13.42 ±1.15, giving a forecast interval
from 12.27 to 14.57, or ±8.6 percent. Since the forecast for the entire residential sector
is this forecast multiplied by the total population, it would have the same relative
confidence interval of ±8.6 percent, assuming a perfect population forecast. If the
population forecast also was associated with a confidence interval, the two would
need to be combined to obtain the total forecast confidence interval.
If one of the independent variables was forecasted with a value different than its
mean value, the forecast interval would be wider. Equation 12-11 shows that the
forecast variance increases as the predicted value of an independent variable moves
away from its mean value within the sample. When more than one independent
variable is forecasted with a value that differs from its mean value, the computation of
the forecast interval is best carried out with matrix algebra, since it involves both
variances and covariances of the independent variables.
Improving Forecast Accuracy With Disaggregation. An excellent strategy
for improving forecast precision is disaggregating utility customers by categories such
as single-family residential, commercial, and industrial. For a large utility, it may be
worthwhile to disaggregate customers into numerous categories to develop a
forecasting model for each. Geographic disaggregation is also a possibility when
distinctly different areas are served by the same utility. Chapter 5 illustrated this
concept for sectoral forecasts.
The primary benefit of disaggregation is that accuracy can be improved by
developing a different model specification for each category to reflect its individual
characteristics. Disaggregation also permits the analyst to use entirely different
forecast methods on various categories. For example, water use by one category of
customers could be forecast with simple extrapolation, another with water-use
coefficients, a third with a causal/structural model, and a fourth category with a
6
Delphi process. Having separate forecast models for each sector allows the analyst to
focus attention on whichever category forecast most needs improvement.
Another benefit of disaggregation is that customer response to changes in some
of the independent variables included in a regression model may differ markedly
among various groups. For example, Agthe and Billings (1987) found that in Tucson,
Ariz., the marginal price elasticity of demand varied systematically by income group,
6. See appendix A for a brief introduction to Delphi methods.

from –0.40 to –0.57, with the lowest elasticity among the highest income group.
Ignoring differences among groups of customers such as these results in a larger
standard error, lower R2, and less precise forecasts. Differences such as this can be
accommodated by running individual regressions for each group. An often-used
alternative is to include intercept and/or slope dummy variables in the model to
capture these effects.
Slope and Intercept Dummy Variables. An intercept dummy variable

such as Pool or Conserve in the preceding model takes on values of 0 or 1. A change in
the dummy variable from 0 to 1 causes a parallel shift in the regression line by the
amount of the estimated coefficient. This is equivalent to adding the coefficient
estimated for the dummy variable to the equation constant when the dummy variable
equals 1. This type of variable is often used to represent the presence or absence of a
condition or status, such as the presence of a low-flush toilet.
In contrast, a slope dummy variable is used to allow the slope or estimated
coefficient of an ordinary variable to differ between two groups of customers or two
situations. For example, the effect of precipitation on water use could be divided
between months with a low average temperature and those with a high monthly
temperature. This would require using precipitation (R) as one independent variable,
and a second variable (HR) that would be equal to R during periods of high
temperature and equal to 0 otherwise. Thus, the estimated effect of precipitation
during periods of high temperature is equal to the sum of the two estimated
coefficients. An equivalent result would be obtained by defining R as precipitation
during cool months and 0 otherwise, and HR as precipitation during hot months and
0 otherwise. Variables such as this can be designed to meet the peculiarities of the
utility’s customers. An alternative in this instance, which has often been used, is to
estimate separate models for the irrigation and nonirrigation seasons, or winter and
summer. Fall and spring are often omitted from such models.
CROSS-SECTION, TIME-SERIES MODELS

The ideal causal/structural model probably is one using pooled cross-section and
time-series data. Weber (1989) described a model using this approach for Oakland,
Calif. Using dummy variable techniques, customers with widely varying income
levels, geographic locations, water-use habits, and so on can be combined into a single
regression. The benefit of this tactic is a greatly increased sample size and an enhanced
ability to generate precise parameter estimates. The increase in precision arises from
the greater variation in the values of the independent variables and the larger sample
size. A data set containing wide variation in water price leads to a much more accurate
estimate of the price elasticity of demand than a data set containing only a narrow
range of prices. At the extreme, when there is no variation in the price of water within
the data set, the price coefficient cannot be estimated.
In Equation 12-2, the largest component of the forecast variance shrinks in
proportion to sample size. For a pooled cross-section, time series model, forecasting
variance shrinks both as sample size increases and as the variance of the independent
variable increases. Thus, even if a combined cross-section, time-series model has a
higher standard error, these other effects of combined samples are likely to dominate,
leading to more precise forecasts (i.e., narrower forecast intervals).
Forecast Evaluation. Forecasts should be tested against reality. The ultimate
test comes when the actual values of the variables in question are available.
Differences between the actual and forecast results do not necessarily mean the
forecast is poor. Large forecast errors can reflect inherent randomness and volatility in
the dependent variable. The existence of such unpredictability is signaled by a wide
forecast confidence interval.
There also are surprises. A forecast based on a structural regression equation
attempts to integrate and project the best current guess about the future values of the
explanatory variables. The demand for urban water in an area, for example, may be
linked to the completion of a major infrastructure project, perhaps a port, an airport,
or a highway.
The emphasis in forecast evaluation, therefore, should be on accounting for the
forecast errors. Would the forecast model have predicted with greater accuracy if an
emergent development or surprise did not happen? Did the error arise from the
estimated model itself, from a misspecification of the model caused by the omission
of some variable that turned out to be important, or from errors in the forecast values
of the independent variables? Identifying the source of the forecast error allows the
analyst to evaluate the potential for forecast improvement, and to determine how
efforts toward revision should be directed. This analysis will also reveal the extent to
which the analyst might have reduced the forecast error by a more thorough initial
effort.
Instead of waiting for time to pass before evaluating the forecast, the analyst can
develop a backcast. This requires forecasting the variables for the time period before
the beginning of the data series used in developing the model or for a “holdout”
sample. The backcast results then can be compared with the actually known but
unused values of the dependent variable. This comparison provides a reality check on
model accuracy.
THE VALIDITY OF REGRESSION MODELS

One desirable feature of a regression is that its parameter estimates are “unbiased.”
Establishing this property in the context of statistical theory involves the following
procedure. Take a number of equal-sized subsamples from the population. Estimate
the regression in question over each subsample, and examine the resulting spread or
distribution of coefficient values. If each of the sets of coefficient estimates are evenly
distributed around the “true” or “underlying” value of the coefficient(s), the
regression model is said to be unbiased.
Basically, this means there is as good a chance of underestimating the true value
of a regression coefficient as overestimating it, and, furthermore, the larger the
sample, the closer the parameter estimate is likely to be to the true or underlying
value. Consistency of the parameter estimates is a weaker, but related and perhaps
more realistic, condition. A parameter estimate is said to be consistent when its
expected value becomes closer and closer to the true value as the number of
observations in the regression increases.
Several conditions can produce unbiased and consistent regression models. The
most fundamental condition is that the error or disturbance components of the
regression are normally distributed, independent random variables, with zero mean
and constant variance. In addition, for individual parameter estimates to be unbiased,
the set of regressors must be orthogonal, not mutually correlated.
These are actually very strong conditions, so it is important to be able to
diagnose and “treat” the “pathologies” that arise when these conditions are not met.
Some of the most common of so-called pathologies are serial correlation or
autocorrelation, heteroscedasticity, and multicollinearity. Also, when an analyst learns to
throw these terms about in the conference room, just about everybody thinks things
have gotten really deep.
Serial Correlation of Error Terms. Regressions with time-series data such

as monthly water use often suffer from serial correlation. This is a tendency for large
positive residuals (errors) in one time period to be followed by large positive residuals
in the next time period (or large negative residuals followed by more large negative
residuals). Serial correlation, or autocorrelation, often indicates that some important
variable has been omitted from the model specification.
The primary effect of serial correlation of the error terms is on the estimates of
the standard errors of the coefficients. The resulting distortion of the size of the
standard errors or t statistics may generate results from the computer program
showing that some explanatory variables are not statistically significant when they in
fact are, or vice versa. Possible autocorrelation of the residuals can often be detected
from an examination of a plot of the residuals. Most statistical computer packages also
provide the Durbin-Watson test statistic for first order autocorrelation. The ideal
value for this statistic is 2, indicating no autocorrelation. As a rule of thumb, values of
this statistic between 1.5 and 2.5 are acceptable; values outside this range suggest
remedial action. Exact critical values can be obtained from the table of Durbin-
Watson statistics in any econometrics book.
The best solution for autocorrelation of the regression equation residuals is to
“discover” the missing variable and add it to the model specification. Another tactic is
to estimate the change or first difference of the dependent variable and selected
independent variables instead of using the actual level. For example, instead of
estimating an equation explaining total water use, compute the first difference of
water use (water use in period t minus water use in period t – 1) and use it as the new
dependent variable. The first differences of each of the independent variables would
also be used in the model instead of the original data. The analyst using such variables
should expect a much lower R2 because a small error in total water use becomes a
much larger error in the change in water use.
When these tactics are not successful, perhaps because the analyst cannot find
data for the needed variable or a first difference would not make any sense,
econometrics and forecasting computer programs offer several mechanical fixes.
Cochrane-Orcutt and Hildreth-Lu are two frequently used procedures.
Heteroscedasticity. Heteroscedasticity is the situation in which the variance

of the error process is not constant, but is instead related to the size of the dependent
variable or to some explanatory variable(s) such as income. This occurs most often in
the analysis of a cross-section sample of individual utilities or businesses of different
sizes. As with serial correlation, the major effect of this condition is on the estimates
of the standard errors of coefficients. Heteroscedasticity can often be eliminated by
converting the dependent and/or independent variable(s) from an aggregate form to
a per capita, per employee, or per square foot basis. For a cross-section study of water
utility districts, for example, using a dependent variable in the form of water use per
capita rather than total water use might alleviate problems of heteroscedasticity in a
regression relationship. Water use per connection might still suffer from this problem,
because some customers use far more water than others, especially certain commercial
or industrial firms.
Specification. The best forecasts are produced when the major explanatory
or exogenous variables influencing the target or dependent variable are in the
regression equation and insignificant variables are omitted. Some explanatory
variables almost always must be omitted from the regression specification, either
because the concept they represent cannot be measured or because their effect is not
very significant. Leaving an explanatory variable out of the regression can bias
estimated coefficients if the omitted variable is correlated with an explanatory variable
in the regression specification. Because the included variable is correlated with an
omitted variable, its estimated coefficient picks up the influence that rightfully should
be attributed to the excluded variable. The amount of bias is related to (1) the
correlation between the omitted and included variable(s) and (2) the magnitude of
the true coefficient of the omitted variable. If the included and excluded independent
variables maintain the same relationship to each other in the future as in the historical
record, excluding the omitted variable has little effect on forecast accuracy.
Weather variables often are not correlated with many of the other significant
influences on water use, such as income. This suggests that it may be possible to
evaluate the contribution of weather to water demand without developing a full
specification of socioeconomic variables in the regression equation.
Multicollinearity. Multicollinearity occurs when two or more explanatory

variables in the sample are highly correlated. As a result, regression procedures may
not be able to distinguish the separate contributions of these variables to the
dependent variable, and the estimated regression coefficients may be unreliable. The
extreme case is that of pure or perfect collinearity. If one explanatory variable is three
times the other in every case or observation in the dataset, the variables are perfectly
collinear. The variables are also perfectly collinear if the sum of several variables is
exactly equal to another variable. Then it becomes impossible to compute the

regression. For example, if a model contains a dummy variable (0 or 1) for each
month of the year, and a constant term, one of the months (or the constant) must be
omitted since all 12 months together add up to exactly the constant. The simple
solution to perfect collinearity is to omit one of the offending variables. An initial
screening for possible multicollinearity can be obtained from computing the simple
correlations among the variables. Pairs of variables with high simple correlations may
cause multicollinearity problems when regression models are developed.
Ordinary multicollinearity is the situation in which there is a close, but not
perfect, linear relationship between some of the explanatory variables in the sample
data. Multicollinearity is usually considered to be a data or sample problem. Thus, the
standard recommendation is to collect more data in hopes that additional observa-
tions will dispel the high correlations. Other tactics include reducing the dimension-
ality of the data with principal components. Such techniques may reduce forecast
errors but also introduce bias to the parameter estimates.
One interesting example is when strong correlations exist between variables
such as income and wealth. In this situation, regression forecasts can be relatively
unaffected by the distortions introduced by multicollinearity if the relationship
between income and wealth does not change much over the forecast period; one
estimated coefficient will pick up the effect missed by the other, and their combined
effects will give a good representation of the economic well-being of these households.
The principle of parsimony (using the simpler model when greater complexity
does not provide significant benefits), however, suggests that when two or more
variables are highly correlated, one of them should be omitted from the model
specification. Thus, most models include either personal income or home values but
not both.
When two of the independent variables are multicollinear, dropping one of
them from the estimated model can be expected to cause large changes in the
estimated coefficient and t statistic of the remaining variable. The appearance of
unexpected signs on estimated coefficients and insignificant t statistics often signals
multicollinearity.
Stepwise Regression. Model building often involves trying various combi-

nations of the available independent variables to determine which combination
provides the best fit to the data. Many statistical programs have an automated process
for variable selection in which the regression equation is automatically estimated
several times, with each run bringing in a single new variable. The program first
includes the independent variable that makes the greatest contribution to explaining
variations in the dependent variable. Then, in the second estimation, a second
variable is added, based on the potential contribution of each of the remaining
variables. This process continues until some given statistical criterion is met. This
process is known as forward stepwise inclusion. The program can also be set to start
with all of the variables and eliminate them one at a time until the statistical criterion
for exclusion is met. This is referred to as backward stepwise regression. There is no
reason to expect that forward and backward processes will necessarily result in exactly
the same set of selected variables. For an example of the use of these techniques, see
Brekke (2002). Although using stepwise regression programs can speed up developing
a forecasting model, it is no substitute for the analyst’s careful selection of the
independent variables and careful evaluation of the regression model results.
Dropping insignificant variables. Often, in the process of developing a model, the
initial equation includes several variables, some of which turn out to be statistically
insignificant and add very little to the explanatory power of the model. When
comparing equations with different numbers of explanatory variables, the adjusted R2
2
is a more sophisticated measure of effectiveness than the simple R statistic. The
2
change in the adjusted R and F statistics when an independent variable is added to or
dropped from the model is one measure of the importance of that variable to the
model specification.
Although the simple R2 generally increases when additional independent
2
variables are added to the equation, the adjusted R increases only when the statistical
benefit of the new independent variable is greater than the statistical cost in degrees of
freedom lost. This typically limits the number of variables accepted in a final model
2
to less than 10 and often to less than 5. If the F and R statistics are not satisfactory,
some important influence on water use may have been omitted from the model, or
the amount of data used in the analysis may be inadequate. It is also possible that
there is no significant relationship among the selected variables or that the actual
relationship changed midway through the period used for analysis. Graphs of the
data, including both the independent and dependent variables, may suggest possible
changes in the model.
Standard error of the estimate. Is the equation sufficiently accurate? The standard
error of the estimate provides a measure of the typical difference between the true value
of the dependent variable (water use) and the value estimated by the regression. If this
value is unacceptably large, the equation needs further work.
2
There are pitfalls in a straightforward “maximum R ” criterion, especially when
models are built with time-series data. The classic pitfall of a casual/structural model
built with time-series data is that variables that increase with time can show up with
spurious correlations. Regression models built with time-series data, as a result, often
2
show much higher R than regression models built with cross-sectional data.
This apparently greater explanatory power can be illusory, however, and suggest
incorrect forecasts if hidden factors relating to the growth of the explanatory variable,
but not related to the growth of the dependent variable, change. One precaution that
may be taken is to replace variables exhibiting trends with their first differences. It is
less likely that the time differences in two variables will show strong correlation when
there is not true causal or functional interdependence than it is for their raw series to
show correlation.
Another proviso pertains to univariate time-series models or to elaborate time-
related adjustments to fit the residuals in multiple regression models. Such models
can contain mathematical features that can “overfit” to the sample data. Univariate
time-series models that follow every contour of the sample data may not generalize or
2
forecast as well as other models with lower R .
SUMMARY
Regression is a very powerful curve-fitting technique because it can incorporate the
effects of many independent or driver variables in explaining the variations in a
dependent or target variable. At the same time, mastery of regression procedures and
modeling involves excursions into statistical theory and the tactics of model building.
The best approach is to focus on collecting good data and estimating relatively
simple models. This chapter gives information about diagnostic tests for adequacy of
these models and some remedies for problems associated with violation of the
assumptions of the classical linear regression model. The acid test, however, is
whether the regression model is successful in forecasting. One possible measure of
success is if the errors in the forecast are no larger than the randomness in water
demand caused by fluctuations in the weather.
The case study in this chapter focuses on a regression model estimated with
household microdata from a random sample of utility billing records and customer
surveys. The discussion considers sample size issues and exploratory data analysis.
The regression is estimated in log-log form, so that the estimated coefficients can be
read as elasticities—suggesting the percentage change in water use that will result
from a given percentage change in each explanatory variable. We computed a
confidence interval for the price elasticity and used it to estimate a confidence interval
for the change in revenue and water sales based on a 15 percent increase in water
rates.
A forecast confidence interval is computed for a hypothetical time-series
forecasting model using average household water sales as the dependent variable. Even
though this model has a high R2, compared with the cross-section model R2 of 0.37,
the forecast interval for the time-series model is more than twice as large.
This example, along with examination of the forecast interval formulas,
suggests that forecasts based on cross-section models of individual customer water use
may be more precise than those developed with more aggregated time-series data. The
ideal forecasting model, however, is likely to be a combined time-series, cross-section
model, because the data in such a model incorporate increased variability of the
independent variables, leading to more precise parameter estimates, along with a
larger sample size.
Frequently, regressions fit to microdata on the behavior of individuals,
households, or individual business firms have a relatively low explanatory power as
measured by R2. These regressions, however, can discriminate the effects of key
explanatory variables with relatively high precision. Thus, regressions based on
microdata tend to be looked to for unbiased estimates of response elasticities for use
in policy analysis and as adjustments to be used in combination with simpler
forecasting models.
In addition, regressions based on microdata often can forecast aggregate water

use for relevant classes or types of customers with relatively tight confidence intervals,
even though the confidence intervals for forecasts of the water use of a specific
customer can be quite large.
REFERENCES
Agthe, D.E., and R.B. Billings. 1987. Increasing Block Rates for Water: Equity, Price Elasticity and
Household Income. The American Journal of Economics and Sociology 46(3): 273–286.
Brekke, L. 2002. Suburban Water-Demand Modeling Using Stepwise Regression. Journal AWWA 94(10):
65–75.
Foster, H.S., and B.R. Beattie. 1979. Urban Residential Water Demand for Water in the United States.
Land Economics 55:43–58.
Howe, C.W., and F.P. Linaweaver Jr. 1967. The Impact of Residential Water Demand and Its Relation to
System Design and Price Structure. Water Resources Research 3(1): 13–32.
Jones, C.V., and J.R. Morris. 1984. Instrumental Price Estimates and Residential Water Demand, Water
Weber, J.A. 1989. Forecasting Demand and Measuring Price Elasticity. Journal AWWA 81(5): 57–65.

CHAPTER 13
Advanced Topics
and Methods
This chapter considers advanced topics and methods important to a deeper

understanding of modeling issues and options in water-demand forecasting. Major
topics include time-series analysis, neural networks, the econometric analysis of
consumer responses to water rates, and risk simulation.
TIME-SERIES ANALYSIS
This section focuses on time-series analysis and begins with a basic discussion of
trend extrapolations. The discussion also includes a quick, conceptual-level introduc-
tion to autoregressive integrated moving average (ARIMA) models and exponential
smoothing. These more advanced time-series methods are sometimes used in
identifying underlying patterns in hourly, daily, or monthly water-use data.
Time Trends. A time series is a time-labeled series or sequence of numbers.

In abstract notation, a time series can be represented as a set of points
(x1, x2, …, xt, ..), where t is the time index for the discrete observations on the value
of the variable x.
Trend Models. The simplest trend model takes the form
Qt=a+bT (Eq 13-1)
281
where:
Qt = water usage in time period t
t = the time index
T = a variable in the equation which takes the value of the time index t
a and b = parameters to be estimated
Equation 13-1 is the equation of a straight line, which can be charted on a

graph with coordinates measuring quantity of water usage and time, respectively.
Generally, this sort of time trend equation is “fit” to historic data by means of a least-
squares algorithm, such as that embodied in statistical regression.
The time index customarily begins with 1 for the first entry in the history of a
time series, and thereafter increments by whole numbers, as in 1, 2, 3, 4, ... .
Any consecutive series of whole numbers, however, such as years, differing by a
constant whole number amount, can be used.
Thus, Equation 13-1 specifies or determines the values of a time series (Q1, Q2,
..., Qt, ...), given the values of the constants a and b.
A time trend can take several functional forms. Equation 13-1 is a linear
equation, but there also are exponential, logarithmic, power, and polynomial time-
trend equations.
In Microsoft Excel, for example, after creating the graph, the cursor can be
placed on the curve that is charted, and with a left click, a window pops up listing
“Add Trendline.” Selecting this option brings up another Windows box with tabs for
the type of trend line and “options.” This procedure supports creation of linear,
logarithmic, polynomial, power, exponential, and moving-average trend lines.
Excel also calculates the best fit parameters for the chosen trend line. If you click
on the “options” tab and select the box for “display equation on chart” and “display
R2 value on chart,” these items of information are added to the chart when the trend
line is drawn.
As an example, Figure 13-1 shows three lines. The jagged line charting values
between 2000 and 2008 represents the historical data. The other two lines are trends
estimated and extrapolated from these historical data. One is a linear trend, generated
by the equation:
Q = 115.25 + 3.917 T (Eq 13-2)
where:
the time index T = year – 1999, so the historical data are effectively
renumbered from 1 to 9 and the projections go to year 17. The other curve
trend line is a logarithmic trend curve, expressed with the equation
Q = 110.31 + 17.24 ln(T) (Eq 13-3)
where ln(.) is the natural logarithm.

ADVANCED TOPICS AND METHODS 283
The extrapolations of this historical series generated by these two trend lines
differ more and more, the further one goes toward 2018. Which is right? Which is the
best trend to use for projection purposes?
The goodness of fit of these two trend lines to these historical data provides
some clues. The R2 for the linear trend is 0.5865, and the R2 or coefficient of
determination for the logarithmic trend is 0.7838. Because closer to 1 is better, the
logarithmic trend may be chosen.
200
180
160
Population (thousands)
140
120
100
80
60
40
20
0
00
02
04
06
08
10
12
14
16
18
20
20
20
20
20
20
20
20
20
20
Historical Data Linear Trend Fitted Logarithmic Trend Line

to Historic Data
Figure 13–1 Historical Population and Trend Extrapolations
In many applied contexts, this is about as far as a trend analysis goes. At the
same time, more complex time-series analysis—including ARIMA models and
exponential smoothing models—has many proponents.
ARIMA Models. The time-trend equations just described are a very simple
form of a time-series model.1 More generally, the two major classes of time-series
models are autoregressive and moving average. Autoregressive time-series models
essentially predict future values of a time-series xt based on weighted values of past
1. Note that a linear time-trend equation is in fact a first order autoregressive process, because we
have Xt = a + bt →Xt + 1 = a + bt + 1 →Xt + 1 – Xt = b or Xt + 1=Xt + b.
values of the time series. A first-order autoregressive process is expressed by the

formula
xt = a1xt–1 + et (Eq 13-4)
where:
et is a normally distributed error process and a1 is a parameter. Note that
if a1>1, the autoregressive process can result in ever-increasing values of x,
depending on the effect of the error process. If a1=1, the autoregressive
process is said to have a unit root.2 A second order autoregressive process
is described by the equation:
xt = a2xt–2 + et (Eq 13-5)
In general, we might write an autoregressive process with k lags as
xt = a0 + a1xt–1 + a2xt–2 + … + akxt–k + et
Time series generated by autoregressive processes have a characteristic signature

that can be unmasked by inspecting the autocorrelation function.
A moving-average time-series model (not to be confused with ordinary moving
averages presented in chapter 5), on the other hand, expresses the current value of a
time series in terms of a moving average of some number of previous estimation
errors.
A famous theorem, proved by Wold in the 1930s, asserts that autoregressive and
moving-average processes have transforms into each other. For example, any
autoregressive process can be expressed as a sum of sufficiently many and possibly
infinitely many moving-average processes and vice versa.
In the 1960s and 1970s, the statisticians Box and Jenkins developed an approach
to time-series modeling of autoregressive and moving average processes. The capstone
of this modeling is the autoregressive integrated moving average or ARIMA model.
These highly structured time-series models have seen relatively little application
in water-demand studies, except in a short-term or very-short-term context. They are
appropriate for modeling the seasonal pattern in monthly water usage, as well as fixed
patterns that can occur for daily water usage.
Exponential Smoothing Models. Exponential smoothing models have an
expression in terms of the classic time-series concepts expressed previously, but
originated in an applied setting and are more intuitive than ARIMA models. The first
2. Unit root processes have come to play an important role in modern econometrics and macro-
economics. A first order autoregressive process with a unit root is called a random walk with drift. Many
important economic and business time series have been shown to be approximately a random walk
with drift—although whether these are truly unit root processes or whether there may be some slight
convergence over time is a point of considerable and continuing controversy.
exponential smoothing models were developed in the late 1950s in connection with
sales forecasting and inventory control. The idea is extremely simple.
Suppose you are delegated to update a forecast produced by upper manage-
ment, relating to the level of sales orders. The original forecast indicates that sales will
run 1 million units per period for the next several periods. In fact, sales this current
period S rose to 1.1 million units. Your job is to revise the forecast for the next period.
An adaptive forecasting method would be to determine a positive fraction, θ ,
that would be multiplied by the forecast error S – Sˆ to produce an adjustment to the
forecast. So the new forecast becomes Sˆˆ = Ŝ + θ (S- Ŝ ).
This simple method has roots in the theory of stochastic processes. So, if sales
in this context tend to oscillate around a level, there is a way to pick θ , the smoothing
constant, so that future forecast errors will be minimized.
In general, exponential smoothing models work with the level of a time series as
well as its trend and seasonal variation. Exponential smoothing models that focus
only on the level of a series are usually called simple exponential smoothing. Holt or
two-parameter exponential smoothing models deal with time series exhibiting
random movements around a level, plus a trend component. Winters or three-
parameter exponential smoothing allows for three components or influences—a level,
a trend, and a seasonal component that can be either additive or multiplicative.
This approach to forecasting is widely employed in business forecasting and is
supported by off-the-shelf software. Forecast Pro, for example, originated as a
program developed for the Electric Power Research Institute, and has evolved into
software that successfully applies artificial intelligence to parameterizing forecast
models. Forecast Pro supports exponential smoothing, Box-Jenkins, or ARIMA
models, as well as several more specialized forecasting applications.
Finally, it is interesting to note that exponential smoothing performed well in
the international forecasting competitions organized and held in the late 1970s and
early 1980s by Spyros Makridakis—the famous M-competitions (Ashley 1988;
Makridakis 1986).
NEURAL NETWORK MODELS

This section provides a brief introduction to artificial neural networks, especially
multilayer neural networks trained by backpropagation of errors.
Artificial neural networks are fascinatingly complex but also boast real
achievements. Voice recognition and other working, day-to-day pattern-recognition
software generally depend on one or another type of artificial neural network. In
water demand, neural networks are reported to perform well in short-term
forecasting applications.
“Artificial neural networks” suggest an analogy to brain cells or neurons. Brain
neurons transmit electrical impulses and process information. They usually have two
types of ends or synapses—an axon for outward going electrical impulses and
dendrites extending from the cell body with possible connections to axons of other
nerve cells.
Similarly, artificial neural networks can be described as a “system of intercon-
nected computational units or nodes” (Salas et al. 2000, p. 25). Generally, there are
places for information inputs, internal processing areas, and places where outputs are
assembled.
Figure 13-2 shows a simple diagram of a multilayer neural network (MNN), a
major type of artificial neural networks. This MNN has two inputs, coming in at the
left side of the schema, and a hidden layer with two nodes. Both input variables are
connected to both nodes of the hidden layer, and the nodes of the hidden layer, in
turn, are connected to another node, which generates the final value of the output
variable. These “connections” are weights that multiply into the values of the attached
input variables or nodes.
Hidden L ayer
Input 1
Output
Input 2
Figure 13–2 A Simple Diagram of a Multilayer Neural Network
The “hidden layer” is so-called because the analyst usually never “sees” or
“tracks” the value of the nodes in the hidden layer. These values are simply in process
to obtaining the value of the output, based on the observable values of the input
variables.
The nodes take the sum of the weighted values of the input variables and apply
an activation function, usually a hyperbolic tangent or sigmoid function. Essentially,
linear sums of weighted values of inputs or nodes pass through these activation
functions. This produces the end value of the node—rather like neural messages are
transmitted through neurons in the brain—and, mathematically, allows the neural
network to capture nonlinearities in the data.
The point of neural network estimation is to determine the values of the
weights to minimize an error criterion—usually measured as the squared differences
of the predicted and actual output. This is an iterative process, often guided by
something known as the backpropagation of errors. The backpropagation of errors can

be interpreted as a gradient descent process familiar from numerical optimization.
The bottom line is that a multilayer neural network can capture and predict
intricately nonlinear relationships between input variables and output variables.
Apart from the error criterion, there are few “diagnostic statistics” in the sense
familiar from regression analysis. Rather, the emphasis in neural network estimation
is on keeping some of the sample data out of the estimation process as a “holdout” or
test sample. At the outset of the analysis, the sample is divided into a two parts: a
training or estimation sample and a usually smaller holdout or test sample. The
neural network is “trained” on the estimation sample, and then tested on the holdout
sample.
This is important because the number of parameters in a neural network grows
rapidly with the number of nodes and hidden layer(s). The danger is overparameter-
ization. Thus, it is possible to define a neural network architecture that achieves a
perfect fit to a body of data, but results in a neural network that completely fails to
generalize.
Spreadsheet add-in programs can define, estimate, and test MNNs for purposes
of forecasting. In terms of the data structures, the process is similar to estimating a
regression relationship in a spreadsheet. The input variables, designed to “explain” the
behavior of the output variable, are grouped together—like the explanatory variables
for a regression.
An important step is designating the architecture of the neural network to be
estimated with the data. How many hidden layers are to be estimated? How many
nodes will there be in each hidden layer?
The program also typically asks the user to specify values or levels of learning
and other parameters. Learning parameters govern the rapidity of movement along
the gradient to the minimum of the sum of squared errors.
The process of developing a working neural network model would seem
inordinately elaborate were it not for an interesting fact—MNNs and their cousins,
recurrent neural networks, often put in superior performances in forecasting contexts.
An abstract from the 15th International Conference on Pattern Recognition reports
typical results, as follows:
In this paper, the relatively new technique of Artificial Neural Networks
(ANNs) has been investigated for use in forecasting short-term water
demand. Other methods investigated for comparison purposes include
regression and time-series analysis. The data employed in this study
consist of weekly water demand at the Indian Institute of Technology
(IIT) Kanpur campus, and rainfall and maximum temperature from the
City of Kanpur, India. The ANN models consistently outperformed the
regression and time-series models developed in this study. (Jain et al.
2000, p. 439)
Multilayer neural networks have been applied in short-term forecasting in
water-demand studies, and the possibilities for applications are broader than this. In
work conducted for the Awwa Research Foundation (AwwaRF), for example, Baxter
(2001) applied neural networks to water treatment plant optimization.
STATISTICAL ANALYSIS OF CONSUMER RESPONSE

TO RATE SCHEDULES
This section considers developments in econometric analysis of consumer responses to
water rates. This is a central topic in the literature relating to urban water demand.
Many studies, over the past two decades, focus on how to correctly model consumer
responses to sale of a commodity at a schedule of rates and charges, as compared with
a uniform price. This section highlights conceptual issues that have generated this
attention. The section also points out the importance of robust approximations in this
context.
The Problem. Chapter 3 notes strong evidence customers react to water sales
at a rate schedule, as opposed to a flat fee per billing period. Chapter 9 surveys wide-
ranging statistical research indicating that higher water rates—and here we mean
“inflation-adjusted rates”—influence water customers to reduce water consumption,
as long as other things, such as income and housing type, remain the same over the
period of comparison. These facts are generally manifestations of the Law of Demand
in the water context (see Jones 1987).
There are problems, however, in specifying the water price or relevant water
rates in statistical water-demand functions, and in developing methods of estimating
these price or rate effects that converge to true values in large samples.
The first issue is: what is the relevant price, feature, or set of features of the rate
schedule influencing customer water use? Currently, suggested answers include (a) the
average price, (b) the marginal price, (c) a two-part price specification involving the
marginal price and a second variable reflecting differences between the marginal rate
and prior rates and overall charges, and (d) the entire rate schedule.
A second issue is how to deal with the simultaneity between water rates or price
and the quantity of water sold in options (a) through (c).
Adjustments for General Inflation. Prior to developing this discussion, let
us emphasize the importance of adjusting water rate and fee data for inflation, in any
analysis involving more than two years of data. It is standard practice to adjust prices
by the consumer price index (CPI), published by the US Bureau of Labor Statistics. If
possible, the researcher should have recourse to regional or local CPI, developed for
his or her specific locale.
Average Price. The simplest way to model price is to compute the average
price for each observation in the data set. The customer’s water bill in this approach
includes all charges for water and sewer service, such as a fixed monthly fee, a
commodity charge, and a sewerage usage fee. This total water bill for each observation
is divided by total water use for that observation to get average price. This
computation can be carried out for individual customers or for some more aggregated
data, such as total billed revenue and total water deliveries for each billing period or
for each of a large number of utility districts in a cross-section model. Estimation
results using average price are usually statistically robust, and often produce
regression statistics that are better than those obtained from more complex models.
The use of average price, however, raises several concerns.
Consumer-choice theory. Economists have developed a theory of consumer
choice that suggests consumers react to the marginal price of goods and services when
deciding how to use their resources. Marginal price should be understood to include
all of the extra money, time, and effort required to obtain a particular good or service.
For most consumers of urban water in high-income countries, the marginal price is
well represented by information on periodic water bills. When reliable piped water is
not available, the effective marginal price may include the value of time and other
resources spent obtaining water.
Often, wide discrepancies exist between average and marginal price when
analysts use individual customer data. With any rate schedule including a periodic
service charge and one or more commodity rates, average price will be above marginal
price when water use is low, as shown in Figures 9-2, 9-3, and 9-4. When analyzing
customers buying larger quantities, this may not be a significant issue. It is also not a
significant issue for aggregate data, because for the average customer in each use class,
average and marginal price will generally be similar.
Simultaneity bias. The difficulty is that average price and quantity are not
independently determined when the rate arrangement includes a periodic service fee
and commodity charges. The estimating model attempts to determine the effect of
price on quantity purchased, yet the price depends on the quantity. Most consumer
goods are not priced this way. When you buy oranges at the store, the price does not
generally depend on how many oranges you purchase.
The extent to which this problem has caused biased (inaccurate) estimates of
the effect of price on water purchases is unclear. It is generally thought that the price
estimates in some reported research may be inflated because of simultaneity bias,
making the reported price response larger than its true value. This contention is
supported by the observation that when a researcher reports models estimated with
both marginal price and average price, the marginal price elasticity is generally closer
to 0.
One approach to removing this bias in studies of water is to parameterize each
rate schedule so that average price used in the statistical model changes only with
inflation adjustments and when explicit rates change. For a very simple example, add
up total utility revenue for all periods when a rate schedule was in effect and divide by
the total water delivered during those periods. Thus the average price used in the
model changes only when rates are changed or adjusted for inflation.
Marginal Price. Marginal price is the price to the customer of the last unit of
water used. For research using individual customers, this can be obtained directly
from billing records or indirectly from the quantity of water used and the applicable
rate schedule. This variable is recommended by the economic theory of consumer

choice, especially when microdata on individual customers are available.
The marginal price has often been challenged because most consumers are
unaware of the value of the marginal price of water and are only vaguely aware of the
size of their water bills unless the bills are exceptionally large. Agthe and colleagues
(1988), for example, report that even after extensive controversy about water and
water rates, only 20 percent of Tucson, Ariz., homeowners were aware that water had
been billed using an increasing block rate schedule for more than 10 years.
When examining the effect of rate changes, both marginal and average price
tend to rise together, although not necessarily by the same proportion, especially for
individual users.
Simultaneity bias. The problem of simultaneity bias occurs under many rate
schedules incorporating a nonuniform commodity charge, such as increasing or
decreasing block rates, step rates, and lifeline rates. The problem is that the marginal
price of water for the individual customer depends on the quantity purchased. This
can result in issues relating to the consistency or large-sample properties of the
estimation of price effects.
One remedy is to develop an instrumental variable for price from the
observations in the data, such that the instrument is independent of the quantity
actually chosen. A simple way to do this is to use the marginal price of the median
consumer over the entire time period during which each rate schedule is in effect as
the price for all consumers. Thus the price is independent of choices of how much
water to use, yet provides a good measure of changes in prices over time or among
utility districts. Simple approaches such as this work well as long as the basic structure
of water rates remains unchanged. Other methods are needed when rate structures
change, such as a shift from declining to increasing block rates.
Other bill components. If the customer’s payment for water is represented by the
marginal price alone, other components of the bill, such as the periodic meter change
and sewerage usage fees, are ignored. To the extent that consumers react to variations
in these charges by modifying their water purchases, the omission of the other bill
components will bias any resulting estimates. If sewerage-use fees are based on
monthly water use, they should be included as part of the marginal cost of water.
Many utilities, however, base sewerage-use fees on water use during the two or three
lowest use months, which are typically December, January, and February in the
Northern Hemisphere. In this case, a totally informed consumer would be aware that
water use during these months will incur charges throughout the year, and include
this implicit marginal cost in deciding how much water to use during these base
period months. Billings and Agthe (1980) found, however, that consumers in Tucson,
Ariz., did not respond this way. Other researchers have found similar results.
Marginal Price and Difference. This approach to modeling water demand

requires two variables derived from billing information:
1. Marginal price = the price per unit of the last unit of water purchased
2. Difference = customer’s water (and sewer) bill minus marginal price times
quantity of water
This formulation is in full accord with the economic theory of consumer choice,
which suggests that although marginal price conveys the primary effect of water rates,
difference measures a secondary effect. The theoretical effect of difference on the
consumer is a pure income effect, and should be the same as a corresponding change
in disposable (after-tax) income. Surprisingly, most of the research using difference
has not obtained the expected results on this variable. One way this has been
addressed is by adding difference to income to create a single adjusted income
variable, often referred to as virtual income. An advantage of this approach is
parsimony.
Simultaneity bias. This two-part rate specification does not eliminate potential
simultaneity bias. This has been dealt with by using models that include an equation
in which quantity is the dependent variable and additional equations in which
marginal price and difference are the dependent variables. These multiequation
models can be estimated by ordinary regression by first estimating the two secondary
equations and then substituting the estimated values of difference and marginal price
into the demand equation. This makes difference and price independent of quantity
in the demand model, thus satisfying the requirements of unbiased estimation. These
multiequation models can also be estimated by other, more complex methods that are
beyond the scope of this work (see, for example, Agthe et al. 1986 and Billings 1982).
Another approach is to parameterize marginal price and difference from each rate
schedule, and then use these “sanitized” values in the demand model.
Discrete-Continuous Choice (DCC) Model for Block Rate Sched-

ules. Another approach for estimating demand models when block rates are
present involves taking into account the entire rate schedule, as it were. Estimating
consumer rate responses in this approach involves considering both the block in
which consumption takes place and the specific amount of water chosen within that
block. This is implemented with a statistical construct called a “likelihood function,”
which is maximized using nonlinear estimating techniques.
The benefit of this approach is that it explicitly accounts for all aspects of the
rate schedule, and should therefore provide better estimates of the effects of changes
in marginal price than earlier models. The price elasticity estimates from these
models, however, are not consistently different from those derived from simpler
formulations. Recent examples include work by Cavanagh and colleagues (2004) and
Olmstead and colleagues (2006).3 The disadvantage of the model is its complexity.
It is interesting to consider an example application of the DCC approach. Thus,
Pint (1999) reports on a study of water use in Alameda County, Calif., comparing
3. Both of these working papers are available on the Web site of Michael Hanemann (see are.ber-
keley.edu/~hanemann/).
demand models estimated with ordinary regression, fixed effects models, and three
versions of the DCC model.
The three models are as follows:
1. Heterogeneous Preferences model, which assumes estimation errors exist

because some individual household characteristics cannot be observed,
and these are by necessity omitted from the model.
2. Error Perception model, in which estimation errors occur because there is

a discrepancy between the amount of water a household intends to use,
given the rate schedule and other influences, and the amount that is
actually used. These may arise from causes such as equipment malfunction
(a leaking toilet) or unintended use by guests or children.
3. Both 1 and 2.
The research uses detailed information on 599 single-family households,

including variables such as lot and house size. The study covers five predrought years
and five drought years. Increasing block rates were imposed starting in July 1991.
Water rates increased in real terms by about 41 percent over the years of uniform
marginal prices. After the increasing block rates, the percentage of customers in each
of the higher use blocks falls dramatically, as shown in Table 13-1.
Table 13–1 Distribution of Customers by Water Use Under Uniform and Increasing Block Rates
With Increasing Block Rates
% Increase Over Uniform Rate

Uniform Rates
Use Block % Users July 1991 July 1992 % of Users
First 60 9 11 78
Second 18.3 100 35 14
Third 10 200 35 5
Fourth 11.7 300 71 3
Source: Pint 1999.
The first and third DCC models produced acceptable results, with the first
model showing the greatest strength. Pint (1999), however, thinks the elasticity
estimates from the combined model are more realistic. Table 13-2 gives the estimated
price elasticities from the first and third models.
The estimated models forecast both residential water sales and revenues for the
entire utility. The forecast error for quantities ranged from 3.7 percent more than to
9.3 percent less than actual values. Revenue forecast errors vary from 2 to 10 percent
less than realized levels. This provides evidence that DCC models can estimate
residential demand and successfully generate forecasts, but much work remains to be
done before these models can be recommended for use.
Table 13–2 Price Elasticity Estimates for Discrete-Continuous Choice (DCC) Model
Low Price ($) Elasticity High Price ($) Elasticity

Heterogeneous Characteristics Model
Summer 0.60 –0.04 3.75 –0.14
Winter 0.60 –0.07 3.75 –0.29
Combined Two Error Model
Summer 0.60 –0.33 2.00 –0.47
Winter 0.60 –0.33 2.20 –1.24
Source: Pint 1999.
Additional Perspective and Information. The research on water consumer

rate responses and their estimation is complex and often draws on advanced
econometric procedures. An interesting question is whether, in sum, the findings
result in significant differences in water-demand forecasts.
Three points can be made quickly in regards to this question of practical
consequence
1. For a typical system price elasticity of – 0.3, water sales would drop by
6 percent for a 20 percent rate increase. Such projections relate to short-
term effects, and assume other explanatory factors, such as weather, do not
change. If mistakes in the price or rate specification are associated, as some
research suggests, with forecast errors on the order of 10 to 40 percent,
errors in predictions of water quantities would vary from less than
1 percent to as much as 2.4 percent.
2. Computations suggest that errors in revenue projections resulting from

mistakes in the price or rate specification can be somewhat larger than the
error in the projection of physical quantities. Thus, given a system price
elasticity of – 0.3 and a 20 percent rate increase, revenue from water sales
would rise by 14 percent in the short run. Mistakes in the price or rate
specification might translate into errors in this revenue forecast of between
1 and 6 percent.
3. Here, an important issue concerns the nature of the rate change. Is it, for
example, a proportional increase in all rates and charges, or a rate
restructuring, whereby some rates and fees increase, while others decrease?
In specific revenue targets—such as “revenue neutrality”—a higher degree of
statistical calibration of water-demand forecasting models is recommended.
For those wishing to pursue the questions raised here, Renzetti (2002) offers a
good overview. Arbues and colleagues (2003) give a detailed exposition of the
development of estimating methods, with extensive references to the literature.
RISK SIMULATION
The final section discusses risk simulation of water-demand forecasts.
A major challenge for all forecasts is assessing their accuracy. Risk simulation
offers important tools for estimating the most likely intervals within which future
water sales and revenues will occur. Many forecasts give only the most likely value for
each time period of the forecast. The chance that realized values will be exactly as
forecast is, however, essentially zero. Risk simulation is a valuable tool for assessing
the likely variability of future events.
Risk simulation, sometimes called Monte Carlo simulation, is a technique for
assessing the potential variability of an output variable or variables, given probabilistic
assumptions about the input variables that determine this output variable.
In the urban water-demand field, the idea is to consider the variation in future
projections of water demand when, for example, per capita water use varies within its
historically observed range, and/or population growth is lower or higher than
anticipated. In essence, this leads to a series of plausible scenarios. Summary statistics
and descriptions of how this variability may look over the forecast horizon can help
stakeholders assess and respond to risks in the planning situation.
The procedure is to establish the probable ranges and variability of the key
input variables and then to sample their values, based on randomized techniques,
tabulating the resulting value(s) of the output variables(s). An off-the-shelf add-in to
Microsoft Excel called @RISK provides a convenient platform for these risk
simulations.
In the per capita water-demand forecast example in chapter 5, for example, the
average gpcd is 258, and the population growth rate is 2 percent per year. Of course,
the per capita water-use rate in any specific year of the forecast horizon is no more
likely to be 258 than it was in any year of the historical record. The standard deviation
of the per capita water-use rate from 1993 to 2006 is approximately 5.
Over this same historic period, the annual average population growth rate was
2.4 percent, varying, in a coordinated way, between 0.5 and 5 percent. The population
growth series shows evidence of first order autocorrelation—a year of higher than
average population growth tends to be followed by another year of higher than
average population growth.
Accordingly, we develop a method for simulating future population growth that
takes both the variability and autocorrelation of the historical population growth
series into account.
Figure 13-3 shows 5 runs of a 30-year simulation of the population for this
community based on this generating equation. In simulation 2, the population
growth rate dropped early in the forecast period, returning to approximately 2 percent
thereafter.
One difference between risk simulation and simply considering a few scenarios
is that risk simulation involves a large number, say 10,000, of runs or separate
samplings of the relevant risk factors. Thus, the population growth rates and per
250,000
200,000
150,000
Population
100,000
50,000
0
07
09
11
13
15
17
19
21
23
25
27
29
31
33
35
37
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
Year
Simulation Simulation Simulation Simulation Simulation

Run 1 Run 2 Run 3 Run 4 Run 5
Figure 13–3 Five Simulations of Population Growth Over the Forecast Horizon
capita water-demand values are resampled from the applicable random distributions
and period-by-period correlations.
The results of these resamplings of the population growth can be summarized
with a risk distribution, such as that shown in Figure 13-4 for the population in the
year 2037 for this example utility. Figure 13-4 is skewed, so the average 2037
population is around 175,000 persons, but the modal, or most likely value, is 155,000
persons.
Moving from this risk distribution of population to water demand means
adding randomly sampled values of the per capita water-use rate for each year of the
planning or forecast horizon. To accomplish this, we could link the per capita use with
variation in underlying annual temperatures and precipitation. However, here we
adopt a simpler approach, assuming that the annual gpcd for any year in the forecast
horizon is a normally distributed random variable with mean 282.7 gpcd and
standard deviation of 10—allowing for somewhat more variability in the forecast than
in the historical sample.
Combining these two risk factors—the annual per capita water use and the
population growth rate—we develop 10,000 runs or scenarios based on these
6
Probability
0
56 76 96 116 135 155 175 195 215 234 254 274 294 314 333 353 373 393 413 432 452 472
2037 Population in 100,000 Persons
Figure 13–4 Frequency Distribution of Simulated Population
0.03
0.025
0.02
Probability
0.015
0.01
0.005
0
0
7
4
1
7
4
1
8
5
2
52
17
24
31
38
45
52
58
65
72
79
86
93
10
10
11
12
12
13
14
14
15
16
2037 System Water Demand (mgd)

Figure 13–5 Simulation Results for 2037 System Water Demand
probabilistic input parameters. The results are summarized in the risk distribution of
water demand at various points in the planning horizon, shown in Figure 13-5.
Risk simulation of water-demand forecasts can involve a several-step procedure,
panels of experts, and groups of stakeholders. The first step is to identify the major
risk factors in the situation. For per capita water-demand forecasts, the problem is
relatively simple. Essentially, the object is to develop the subjective probability

distributions for key risk factors. Several techniques for developing the probability
distributions of risk factors exist; Delphi methods, for example, can be used to elicit
parameters of risks from panels of experts or groups of stakeholders in the planning
process. Historical data also can be and probably should be collected to see what light
these data cast on the potential variability of risk factors. (For a more extensive
discussion in the context of water investment, see Jones 1992).
Generally, identifying the potential variability of risk factors is fairly straightfor-
ward if historical data series are available. It is somewhat more difficult to develop
realistic correlations between risk factors. From data, this generally requires more
sophisticated time-series analysis such as identifying the order of autocorrelation in
sample series. In addition, assuming that the future will be like the past is not always
valid. In essence, this means that risk factors must be carefully chosen to reflect the
probable realities in the future.
The per capita water-demand forecast and risk simulation illustrates several key
points. First, if the risk factors have “most probable” and centrally located average
values, the resulting distributions of the output variables are likely also to be peaked
distributions. The distributions of the output variables, however, may range consider-
ably more widely than one might expect.
In this example, total system water use in the simulation ranges between 17 and
168 mgd, with an average of 51.2 mgd. This paints a much different picture than
suggesting the system water demand in 2037 will be 51.2 mgd.
The issue becomes how to summarize the results of the risk analysis. One mode
of expression that community stakeholders seem to favor is the odds ratio. Thus,
based on this risk simulation, it is possible to assert that the odds of total system water
demand not exceeding 60 mgd in 2037 are approximately 3:1. In three out of four
possible universes generated according to these stochastic assumptions, then, total
system water demand in 2037 will be less than or equal to 60 mgd. An interesting
feature of the risk distribution is its skewed characteristic: its average value is much
greater than its mode.
This risk situation can be parameterized in numerous ways. For example, it
might be plausible to link the variability of the per capita water-use rate in any year
with the total population served. The idea is that if population growth ramped up
significantly, a kind of risk pooling could occur. Thus, suppose the per capita water
use of individual households can be expressed as GPCDHH = 282 + e, where e is a
normally distributed error process with zero mean and a variance of 64. Then,
consider the variability of total annual water usage of 100,000 such households:
100,000 × 282 + E, where E has a variance of 100,000 × 64. Note that the standard
deviation of the total error process, however, has increased by only the square root of
100,000, or less than proportionally to the total population served. In short, variances
add, but the standard deviation is the key parameter for the normal distribution. The
standard deviation increases more slowly than the variance, in this sense.
SUMMARY
This chapter covers advanced material relating to water-demand forecasting, includ-
ing time-series analysis and neural networks, recent developments in the econometric
analysis of consumer responses to water rates, and risk simulations of water-demand
forecasts.
REFERENCES
Agthe, D.E., R.B. Billings, J.L. Dobra, and K. Raffiee. 1986. A Simultaneous Equation Demand Model for
Block Rates. Water Resources Research 22(1): 1–4.
Agthe, D.E., R.B. Billings, and J.M. Dworkin. 1988. Effects of Rate Structure Knowledge on Household
Water Use. Water Resources Bulletin 24(3): 627–630.
Arbues, F.M., A. Garcia-Valinas, and R. Martinez-Expineira. 2003. Estimation of Residential Water
Demand: A State of the Art Review. Journal of Socio-economics 32(1): 81–102.
Ashley, R. 1988. On the Relative Worth of Recent Macroeconomic Forecasts. International Journal of
Forecasting: 4(3): 363–376.
Baxter, C.W., R. Tupas, Q. Zhang, R. Shariff, S.J. Stanley, B.M. Coffey, and K.G. Graff. 2001. Artificial
Intelligence Systems for Water Plant Optimization. Denver, Colo.: Awwa Research Foundation
(AwwaRF).
Billings, R.B. 1982. Specification of Block Rate Price Variables in Demand Models. Land Economics 58(3):
386–394.
Economics 56(1): 73–84.
Cavanaugh, S.M., W.M. Hanemann, and R.N. Stavins. 2006. Do Consumers React to the Shape of
Supply? Water Demand Under Heterogenous Price Structures. Working paper 05-10. Washington,
D.C.: AEI-Brookings Joint Center. Available through Social Science Research Network Web site at
www.ssrn.com/.
Jain, A., U.C. Joshi, and A.K. Varsheny. 2000. Short-Term Water-Demand Forecasting Using Artificial
Neural Networks: IIT Kanpur Experience. Pages 439–462 in Proceedings of 15th International
Conference on Pattern Recognition. Washington, D.C.: IEEE Computer Society.
Jones, C.V. 1987. Nonlinear Pricing and the Law of Demand. Economics Letters 23:125–128.
———. 1992. Financial Risk Analysis of Infrastructure Debt: The Case of Water and Power Investments.
Greenwood, Conn.: Quorum Books.
International Journal of Forecasting 2(1): 15–39.
Olmstead, S.M., W.M. Hanemann, and R.N. Stavins. 2004. Muffled Price Signals and Household Water
Demand Under Increasing Block Prices. Available through Social Science Research Network Web
site at www.ssrn.com/.
Pint, E.M. 1999. Household Response to Increasing Water Rates During the California Drought. Land
Economics 75(2): 246–266.
Renzetti, S. 2002. The Economics of Water Demands. Boston, Mass.: Kluwer.
Salas, J.D., M. Markus, and A.S. Tokar. 2000. Streamflow Forecasting Based on Artificial Neural
Networks. Pages 23–51 in Artificial Neural Networks in Hydrology. Edited by R.S. Govindaraju and
A. Ramachandra Rao. Dordrecht, Netherlands: Kluwer.

CHAPTER 14
Forecast Uses, Evaluation,

and Improvement
The forecasting process consists of nine steps, as follows:

1. Choice of the forecast approach
2. Collection and analysis of data
3. Identification of forecast model(s)
4. Estimation of the forecast model(s)
5. Diagnosis of the statistical adequacy of the model(s)
6. Production of the forecast, including confidence intervals
7. Evaluation of the forecast
8. Use of the forecast by decision makers
9. Expost facto analysis of forecast error
This book has dealt generally with this process, and this chapter focuses on the
initial and final stages—the choice of the forecast approach, evaluations of forecasts,
and their improvement.
A strategic or pivotal feature of forecasts is their time horizon. There are short-
term, medium-term, and long-term forecasts. Broadly speaking, short-term forecasts
are for months. Medium-term forecasts range from one to several years. In the water
field, long-term forecasts can extend up to several decades. These time spans can
301
overlap, but the classification is useful since it relates to the types of problems
forecasts are called upon to solve, as well as data and forecast models.
The next section pulls together observations on forecast horizon, forecast use,
and type of model. The discussion then turns to issues related to forecast confidence
intervals and forecast evaluation. We present several of the most widely used measures
of forecast accuracy. Suggestions are offered on how to improve forecast accuracy. The
chapter closes with a discussion of the benefits of developing formal forecasts.
FORECAST HORIZON AND APPROACH

Short-Term Forecasts. Short-term water-demand forecasts have applica-
tions in the management of utility operations. These forecasts also can inform
maintenance and operating schedules for pumps, wells, reservoirs, and mains. Short-
term forecasts are also relevant to decisions about water levels and drawdown in
reservoirs. Almost all utilities rely on forecasts for these purposes, even if these are as
simple and intuitive as, “We do our pump maintenance work in January when
demand is low.”
Short-term fluctuations in water use tend to be dominated by the weather.
Although a forecaster can adjust for normal or average seasonal variation, specific
weather forecasts for horizons of more than a few days are not reliable, at the current
state of the art.
Because of weather forecasting limitations, short-term water-demand models
frequently are simple ratios or proportions developed from the historical record by
averaging. Otherwise, regression performs well in this context, where statistical
interaction between temperature and precipitation is finessed by expressing weather
variables in terms of differences from their long-term norms. At the system level, such
regression models usually incorporate weather variables, seasonal factors, and
underlying trends in water use. Chapter 5 illustrates deployment of regression to the
explanation of monthly and seasonal water demands. Good performance also is
reported with neural network models of short-term water demand.
When such key factors as price or conservation programs vary, forecasts must
be adjusted. Often, this adjustment is based on judgment and knowledge of the
probable effect of the changed conditions.
For some situations, sample survey data of utility customers might produce or
enhance a short-term forecast. Such a survey might ask respondents to describe their
water-using activities, any recent changes in those activities, and plans for the future.
So if numerous customers are planning to adopt low water-use landscaping (or
replace grass with gravel), the water utility can anticipate a corresponding change in
water demand.
Surveys are especially useful for industrial customers, who may have plans to
expand or modify production activities that would affect their future water use. A
utility with a few large industrial water users could benefit from regular, periodic
FORECAST USES, EVALUATION, AND IMPROVEMENT 303
surveys of these customers. Utilities should also collect information on the probable
effect of changes in government regulations of water use or wastewater disposal.
Long-Term Forecasts. Water utilities face a formidable long-term forecast-

ing problem. The challenge is to build sufficient capacity to meet customer demands,
while avoiding costly excess capacity. The record suggests that overforecasting long-
term population growth is more common than underforecasting. Variation in water
use over the long term tends to be dominated by the number and type of customers,
as well as factors such as household size and change in the industrial mix. Sensitivity
or risk analysis is often helpful in evaluating the potential range of required supply
capability. Analysts sometimes produce long-term forecasts using intuitive methods
such as Delphi (see appendix A).
Qualitative issues also can be important over the longer term, especially those
pertaining to directions in the local or regional economy. Depending on local
conditions, key questions may be:
• Are recent declines in use per capita likely to persist?
• Will recent rapid growth continue?
• When will vigorous economic growth be reinstituted?
Per capita and unit-use or sectoral water-demand forecasts are the primary
long-range water-demand forecasting techniques. Water-use forecasts by customer
sector often help sort out trends applying to components of water use that are masked
in simpler per-capita forecasts.
Recently, modeling of water conservation effects with end-use estimates has
come to be incorporated in medium and long-term water-demand forecasts—often
based on sophisticated benefit-cost analysis.
Medium-Term Forecasts. Medium-term forecasts with horizons from one

to five years support utility operations, minor system improvements, revenue
forecasts, and rate setting. As such, medium-term forecasts share features of long-
term and short-term forecasts, but can be have horizons long enough for slower-
acting influences to play themselves out.
Thus, business and weather cycles are important influences on water demand
over multiyear periods—both are subject to recurring fluctuations of uncertain and
somewhat random severity and duration. It is well-known, for instance, that wet and
dry years tend to run in multiyear cycles. Additionally, there is a tendency for long
droughts (or wet periods) to continue. These multiyear cycles, however, can and do
change abruptly. The business cycle influences the level of existing customer demand
and the number of customers over multiyear periods. In the recovery phase of the
cycle, construction activity picks up and some residential customers may opt to
expand their homes, adding bathrooms or even swimming areas. On the other hand,
a slowdown in local, national, or export markets may influence industrial customers
to significantly reduce water use.
The development of alternative forecast scenarios or development of risk

simulations can be useful in this context.
Many university research departments produce state and local economic
forecasts, combining models of the local economy with macroeconomic forecasts
from national organizations. The water-demand forecaster can use these economic
forecasts as independent variables in a causal/structural model.
Causal/structural models incorporating explanatory variables based on surveys
of individual customers can inform estimates of the effect of changes in price, income,
and other key explanatory variables in a formal water-demand forecasting system.
Often such models have relatively low R2 but strong t-statistics (low standard errors)
for individual regression coefficients. Such regressions can often produce relatively
accurate estimates of the effect of selected explanatory variables as well as forecasts of
overall sectoral water use.
Reliable forecasts of the future values of multiple independent variables may,
however, be difficult to locate or produce. One approach to resolving this dilemma is
to develop alternative scenarios with values of the independent variables consistent
with high, medium, and low future water use. This not only produces a range of
possible values for the forecast but also focuses attention on the characteristics of the
local economy that would produce each result.
FORECAST INTERVAL ESTIMATES

Forecasts generally gain credibility by offering a range within which the actual
outcome is likely to occur, rather than a single value. Several approaches are available
to generate forecast intervals, including statistically based confidence intervals,
probable ranges indicated by risk or Monte Carlo simulation, and the construction of
alternative scenarios.
Regression-Generated Confidence Intervals. Regression analysis and
other parametric statistical approaches to forecasting provide methods for computing
the confidence intervals of forecasts. For a regression-based forecast, the general result
is that the distance from the lower to the upper boundary on a 95 percent confidence
interval is related to:
• The standard error of the regression
• The difference(s) between the mean value(s) of the explanatory variable(s)
and the projected value(s) of the explanatory variable(s)
If one explanatory variable such as population is projected to grow or decrease
or, in other words, exhibit a trend, the confidence intervals will grow increasingly
1
larger over time. Obtaining an overall confidence interval for a forecast requires
taking into account the forecast errors for the explanatory variables.
Risk or Monte Carlo Simulation. As noted in chapter 13, often the best
way to account for the possible variation in a forecast is to conduct what is called a
Monte Carlo or risk simulation, or risk analysis. Thus, in principle, if we know

realistic ranges for all the input variables that go into determining a forecast, we
should be able to utilize this information with other statistical measures to come up
with a likely distribution of the water-demand forecast in future time periods.
As illustrated in chapter 13, this type of analysis can be carried out within
standard business spreadsheets.
Alternative Scenarios. Analysts also can develop alternative values of the

independent variables of the forecasting model from a set of scenarios of the future.
Thus, a medium- to long-term demand forecast might include a low-growth scenario,
in which the economy and population of the utility area grow slowly, if at all; a high-
growth scenario driven by new industries and a revitalized infrastructure; and a
medium-growth scenario. Each scenario can be used to develop a consistent set of
predictions for values of the independent variables. The scenario discussions would
elucidate for the forecast user the underlying forecast assumptions, thereby making
the forecast more valuable.
FORECAST EVALUATION
Evaluating a forecast starts with computing the forecast error, which is the difference
between the actual values of the dependent variable and the forecasted values. Here
we present several of the most widely used measures of forecast accuracy.2
Mean Squared Error. The differences between the predicted and actual
values are squared and added together, creating a sum of squared errors that would be
0 if the forecasts were perfectly accurate. The sum of squared errors accords the same
weight to positive and negative forecast errors, but weighs one large error much more
heavily than several small errors adding up to the same total. The sum of squared
1. Suppose y = βx + E expresses a simple regression of y = total system water use with x = population
as the independent variable, where both x and y are expressed as deviations from their respective
means. Then, the variance of the forecast error E, when x ✻ is a projected population for a given year and
S is the standard error of the regression, is
⎛ ⎞
⎜ 2 ⎟
( x∗ ) -⎟
Var ( y∗ ) = S ⎜ 1 + -----------------
2
⎜ 2⎟
⎜
⎝ ∑ xt ⎟
⎠
sample
where:
xt is summed over the sample of historical data on population and water use. Thus, the larger
x✻ is, or, in other words, the further the forecast value of population is from the average pop-
ulation in the sample or historical data, the larger the variance of the forecast and the larger
the confidence interval.
2. For further discussion, see Hanke and Wichern 2005.
errors is typically divided by N, the number of observations, yielding the mean

squared error (MSE), which can be compared among samples of different sizes. Using
monthly data for the example, the MSE is:
N
1 ˆ )2
MSE = ---
N ∑ ( Qi – Q i (Eq 14-1)
i= 1
where:
Qi =the observed water use during the ith month
ˆ
Q = the estimated water use during the ith month
i
i =the sequence number of each individual month
N =the total number of months in the sample.
The square root of the MSE is the standard error of the estimate, which is one
measure of the “typical” size of the forecast error.
Mean Absolute Deviation. The mean absolute deviation (MAD) is an
alternative measure used when the analyst wants to weigh one large error the same as
several small errors adding up to the same total. The MAD considers positive and
negative errors equally by taking their absolute values. Using the same variables as
Equation 14-1, MAD is:
N
1
∑
MAD = ---- ˆ )
( Qi – Q (Eq 14-2)
i
N
i=1
The MAD is a measure of the typical or average size of the error or difference
between the actual data and the forecast. The most important difference between the
MSE and the MAD is that the MSE squares the individual errors, thus putting heavy
weight on outliers; the MAD does not.
Mean Absolute Percent Error. The forecast error is often presented in
percentage terms by dividing each measured error by the actual value of its
corresponding observation. This produces an error measure that is unit free, so that
the analyst can compare the accuracy of forecasts for different-sized utilities or market
segments. The mean absolute percent error (MAPE) is:
N
Qi – Q ˆ
1
∑
i
MAPE = ---- -------------------- (Eq 14-3)
N Qi
i=1
Bias. The mean bias gives the amount and direction by which the average
estimated value strays from the true average.
N
1 ˆ
Mean bias = ----
N ∑ ( Qi – Qi ) (Eq 14-4)
i=1
The mean bias can also be put in percentage terms, giving the mean percent
bias, or the average percentage and direction by which the forecast strays from the
actual value. Its value is
N
( Qi – Q ˆ )
1
∑
i
Mean percent bias = ---- ---------------------- (Eq 14-5)
N Qi
i=1
ERROR MEASURES AND LOSS FUNCTIONS

The analyst will probably compute several measures of forecast errors when
evaluating a forecast. The most important error measure depends on the losses
associated with various types of error. If large errors are more important than several
combined smaller errors, a measure such as the MSE, which uses a sum of squared
errors, is most appropriate. For example, the political cost to water utility officials
may increase more than proportionately with the size of a water shortage.
If instead, the loss imposed by a forecast error is simply proportional to its size,
a measure such as MAD is appropriate. In some situations, the loss may be more
closely associated with the relative error size, so that one of the measures using a
percentage would be appropriate.
There is no particular reason to believe that the loss imposed by forecast errors
is symmetrical. For example, the political or economic cost of a 20 percent supply
shortage may be much more severe than that of an equal-size surplus or excess
capacity. In this case, positive errors (actual greater than forecast) would be given a
much greater weight than negative errors (actual less than forecast). For a revenue
forecast, the opposite may be true, so that a one-million dollar deficit is much more
serious than an equal surplus.
STRATEGIES TO IMPROVE ACCURACY

The most important source of improved forecast accuracy accrues from carefully
evaluating past forecast results. Evaluation will reveal the size and pattern of forecast
errors and often suggest ways to improve the next set of forecasts. The greatest
obstacle to improvement is a failure to compare prior forecasts with actual events.
In a regression-based forecast, the analyst can disaggregate the total forecast
error between errors in the values of the independent variables and error resulting
from the model specification. An expost facto forecast using the actual values of the
independent variables will sort out these effects. The realized value of each
independent variable can be compared to its forecasted value to determine errors in

forecasting the independent variable values. The forecast may have been off primarily
because the weather was abnormal. If so, no adjustment in the forecast model is
warranted. Conversely, if the error arose because of the omission of some important
variable from the model specification or from a change in the utility system, the
analyst would need to reformulate the model. Forecast error could also have arisen
from errors in the historical data used to estimate the model parameters. Improving
the available data can be highly beneficial.
When uncertainty about the general characteristics of the future is great,
forecast accuracy can often be improved by averaging alternative forecasts of water
use. Thus, instead of putting all of their effort into a single forecast model, analysts
could develop three or four models with the same total effort, using a wide range of
techniques. The results of these models would then be averaged to obtain a “best
guess” forecast. At least one national forecasting service follows this strategy by
obtaining forecasts of inflation, interest rates, economic growth, and so on from
several dozen individual forecasting organizations, and then takes an average to
obtain a “consensus” forecast.
Before committing additional resources to improve forecast accuracy, the
forecaster and the forecast user need to jointly consider the costs and benefits of
forecast improvement. This requires evaluating the expected incremental benefit to
the organization from forecast improvement compared to the incremental cost.
Additional resources would be devoted to the forecasting effort only when expected
benefits exceed expected cost. Generally, the greater the potential loss resulting from
forecast errors, the more the resources put into increasing forecast accuracy.
WHY FORECAST FORMALLY?

Taking all the potential sources of variation in a forecast of water use for an urban
area into account may seem to lead to indeterminate answers. This is not necessarily
true, however, because some risks tend to balance out. Indeed, as Spiros Makridakis
states, it may be in the medium-term forecasting period where risks are amplified to
their greatest extent:
In the long term, economic laws and S-shaped curves always prevail,
making it possible to identify and measure long term trends and average
relationships using least squares methods … In the short term, seasonality
rarely changes. In addition, the momentum or inertia of most phenomena
makes extrapolation routine and allows the use of simple methods (such as
exponential smoothing and decomposition) which isolate and project
accurately the most recent pattern. The combination of seasonality and
inertia/momentum is indeed being used to a great extent by most statistical
methods to make their predictions which on average are more accurate
than … judgment. However, forecasts must be made for additional
horizons including the medium term, which is the hardest to deal with
because the depth and length of each cycle is unique. … Only through a
combination of judgment and systematic analysis of past fluctuations can
we arrive at reasonable conclusions about medium term cyclical fluctua-
tions which can help us more accurately monitor and more successfully
predict future cycles and the resulting uncertainty. (1986, p. 71)
The question is not “Can we develop a perfect forecast?” Rather, can the utility
demand and revenue forecast be improved by developing a formal forecasting
program? Even without much data, and using only simple models, the utility can
develop forecasts of water demand that will be more accurate than the implicit
forecasts necessarily used in the absence of a formal effort.
Forecasting is a continuing project. The initial forecasting model should evolve
and become more accurate over time. Accumulating experience, including careful
comparisons of forecasts and actualizations, can suggest refinements. When condi-
tions change slowly, and water rates incrementally increase each year, there may be
little apparent need for a comprehensive forecasting model. For contingency
planning, however, incorporating variables representing weather, population, price,
and conservation effects is necessary to obtain accurate revenue predictions under
various unusual circumstances, and therefore aid in being prepared to handle them
ahead of time.
There is no guarantee that the future can be accurately predicted. Nevertheless,
it is usually feasible to analyze water-use patterns, obtaining relatively accurate
forecasting models for next month's water use or for periods of up to several months.
It is also possible to inform long-term planning with regression models based on
explanatory variables such as population, income, and price. These really are
hypothetical models in which the random element that is so apparent in day-to-day
demand is factored out over the long term. Beyond that, making credible statements
about water demand in an urban system over a period of several years requires careful
study of the available data and general knowledge about population, climate,
socioeconomic, and other relevant processes.
To develop forecasts that are as accurate as possible, the forecasting function
should be independent of the utility’s planning and operating activities. The forecast
must not be unduly influenced by what management wants to hear. Of course, the
forecast will take utility operating policies and programs into account, but the
forecaster should have the independence to report results that disagree with
management expectations.
REFERENCES
Armstrong, S.J. 1978. Long-Range Forecasting: From Crystal Ball to Computer. New York: John Wiley &
Sons.
Hanke, J.E., and W. Wichern. 2005. Business Forecasting. 8th ed. Upper Saddle River, N.J.: Pearson
Prentice Hall.
International Journal of Forecasting 2(1): 15-39.
APPENDIX A
Survey Methods for

Direct Data Collection
Water utilities often want to gather information about their customers beyond the
basic water-use data generated by the accounting system. Some demographic
information is available from outside sources, such as the decennial census of
population and local governmental agencies (discussed in chapter 2). Specific detailed
information can be obtained only from the utility’s individual customers. Desired
information might include customer response to publicity about water conservation,
the rate at which they accept water-saving devices, their willingness to spend for water
system improvements, and the extent to which they might be willing to pay now to
avoid some probability of water rationing in the future.
This appendix touches on some considerations that must be taken into account
when designing and conducting surveys. Topics discussed include methods for
customer feedback; development of alternative scenarios, questionnaires, and ques-
tions; determination of the sample size; selection of a sample; random sampling; and
mail and telephone surveys.
CUSTOMER FEEDBACK
Day-to-day customer contact with the utility is one low-cost source of information.
Although most customers simply pay their bills, a small percentage will spend time
discussing water-related problems with utility employees. This feedback does not
provide a statistical basis for inferring how the vast majority of customers feel, but
carefully tabulating these comments and complaints results in continuous feedback
311
on how well the utility is serving customer needs. Additional low-cost information
can be obtained from such sources as letters to local newspaper editors and the results
of water-related bond elections.
Opinions about the future can also be collected from current and prospective
water customers. For example, major commercial and industrial water users could be
asked to share plans for changes in their technology and output. This would allow the
utility to incorporate the impact of these changes into demand forecasts. Similar
information could be obtained from governmental regulators, who may be developing
new regulations on water use or wastewater disposal, either of which would affect
water demand over the next few years. Survey results could be directly compiled into
a forecast of future water demands, or, more likely, the survey results would be
incorporated into the forecast either as a judgmental adjustment or as values for an
independent variable in a causal/structural model.
EXPERT OPINION
One approach used to examine possible future trends involves the development of
alternative scenarios. For the long term of a decade or more, scenario analysis seems
advisable as a complement to a complex, multivariable, causal/structural model of
demand. This is because each of its several explanatory variables must themselves be
forecasted. Qualitative issues may be more important in the longer run, especially
those pertaining to directions in the local or regional economy. For many areas, the
key questions are whether industrial decline will be reversed or vigorous economic
growth reinstituted. For other areas, the problem is whether the current soundness of
the economy can be sustained in the face of challenges from global competitors.
Scenario analysis is useful in this regard. This approach involves brainstorming about
how future trends will play themselves out for a particular urban area under several
alternative assumptions.
Delphi and Other Expert-Opinion Techniques. Intuitive forecasts take a
number of specific forms, of which the Delphi methods are the best known. The
advantage of such forecasts is the incorporation of expert opinion—without the time
and cost of formal model building. The process is not necessarily cheap, however.
Considerable time and effort are required to obtain the desired responses, and it may
be necessary to pay the experts.
The Delphi method for forecasting water demand uses a group of experts
selected from a wide range of specialties from disciplines such as civil engineering,
sociology, economics, economic development, land development, urban agriculture,
and political science. Each expert is asked a series of questions about the long-term
future, such as
• What will happen to exterior water use by each major customer category?
• What will happen to indoor water use by each major customer category?
• How much will the size of each customer class change?
APPENDIX A 313
• How would the customers respond to an increase in water rates of X

percent?
Each expert writes out a set of responses and submits it to the facilitator, who
prepares anonymous summaries of the various opinions and sends them back to the
panel members. Each panel member then has an opportunity to modify his or her
forecast in light of the opinions of other experts. This process goes through several
iterations, with an end result that will either lead toward consensus or to a
clarification of why different experts hold varying opinions on the most likely
outcome.
Expert panels with personal interaction can also be used. In this method, the
experts assemble and share their viewpoints with each other. This has the advantage
of allowing direct interaction, which may lead to quicker clarification of differences or
movement toward a consensus. The primary disadvantage is that the more vocal or
forceful members of the panel may dominate the discussion, leading to less accurate
forecasts. This process also leaves little time for reflection, which may short-circuit the
development of new insights.
Regardless of the process used, the experts making up the panel should be
diverse in training and experience. The panel should also include some members with
no stake in the results of the forecast. A wide diversity among the panel members
helps to minimize the bias that tends to develop among any group in the same
geographic or professional area.
Various attempts have been made to assess the accuracy of judgmental forecasts.
Armstrong (1985) concluded that for long-range forecasts, causal/structural models
are superior to judgmental models.
SAMPLE SURVEY OF CUSTOMERS

Although valuable information can be obtained from a variety of sources, only a well-
designed effort using statistically based random sampling models is capable of
providing data from which valid inferences about the entire population can be drawn.
When a utility decides that information about customers is useful and worth the
required resources, several steps are necessary for carrying out scientifically valid
research in the form of a survey (see Cochran 1963 for an example).
Steps in Designing a Survey. A professional statistician can help with some
of these steps:
1. Determine the question or problem requiring a solution. There is no point
in spending resources to collect information unless that information can
be used to help solve a problem. Customer information might help to
address issues such as the need to conserve water or to plan for capacity
expansion. In addition, this information can be used to determine
customer preferences between current water-rate financing and bond
financing of utility capital expenditures. Surveys might also be used to
determine customer preferences regarding water treatment, especially if

removal of dissolved minerals is an option.
2. Determine what information will help to solve the problem identified in
Step 1 and how the information will be used once it is collected. There is
no value in collecting information unless there is a plan showing how it
will be used. Carefully thinking about how the new information will be
used is essential in designing questions that will yield the desired
information.
3. Determine the population about which more information is needed. For a
water utility, this may be all customers, only residential or commercial
customers, only customers in certain parts of the service area, or only
customers with certain equipment at their homes or workplaces. Addition-
ally, the population of interest may be only those customers whose water
use exceeds a certain level, either continually or once in a while.
4. Determine how much precision is needed. A substantial part of the cost of
a survey results from developing appropriate questions and testing them,
but another significant part of the total cost results from the number of
customers in the sample. Increased precision carries a higher price tag.
In this context, precision measures the closeness of results of repeated samples
of the same size, and can be estimated before collecting the data based on statistical
theory. In other words, a measure of precision indicates the likelihood that collecting
more data will not significantly change the results. For example, sample results are
often said to be ± 5 percent except for a 1 in 20 chance. This simply means that if we
collected a large number of samples from the population, the mean would vary by no
more than 5 percent from the true value for 95 percent of the repeated samples.
The utility can opt for a low-budget effort with correspondingly low precision,
or for high precision at a much higher cost. If statistically valid results for various
subgroups of the population are required, the total sample must be commensurately
larger. Sample size will be discussed in detail later in this appendix.
One important limitation to the accuracy of the survey results is that
respondents can systematically misunderstand or misrepresent their situation. It is
well-known, for example, that survey respondents consistently overstate the extent of
their recycling activities and understate the amount of their consumption of tobacco
products and alcoholic beverages. Similarly, customers may self report more extensive
water-conservation activities than they are actually undertaking. This source of error
can be reduced—but not eliminated—by careful question design. It can also be
compensated for by using other sources of data.
Questionnaire Development. Developing an appropriate set of questions is
one of the critical phases of obtaining information from customers. Each question
needs to be clearly stated so that every respondent will understand what is being
asked. The questions must avoid even subtle bias or the impression of a preferred or
“socially correct” response by the customer. This is sometimes difficult, but it is well-
APPENDIX A 315
known that respondents unconsciously want to please the interviewer by giving the
“correct” answer. As a result, people tend to claim to recycle their beverage containers
and newspapers much more often than they actually do, and will profess greater
interest in water conservation than they actually have. It is important, then, that
questions be specific and nonjudgmental. Often this can be more effectively
accomplished by using questions with multiple-choice responses. For example, to
determine a customer’s knowledge about water rates, this question could be asked:
“Do water rates per unit of water vary during the year?” Answers might include (1)
“Yes, they are higher in the summer”; (2) “Yes, they are higher in the winter”; (3) “No,
they are the same all year”; or (4) “I don’t know.”
The questions should be sequenced to avoid having questions near the
beginning that might bias the answers near the end. The questions also need to
progress from easy to difficult, to build the respondent’s confidence in his or her
ability to answer the questions. A progression from general to personal is also helpful,
as the respondent’s commitment to completing the interview hopefully increases. The
questionnaire should also not be overly long, or potential respondents will lose
interest and skip some questions or not send in the responses at all.
Reviewing Questions. Subjecting the draft questions to rigorous review is
extremely important. The first review should be by water utility officials and other
professionals, followed by a test sample of water customers. Pretesting often reveals
unexpected problems with the questions and provides an opportunity to avoid serious
error.
Regardless of what medium is used for the survey, obtaining voluntary
participation from the selected sample requires careful planning. The interest of the
potential respondent needs to be quickly piqued. The introduction to the interview
process should demonstrate the interviewer’s credibility and clearly explain the
importance of the requested information. The interviewee must be confident that this
is not a sales pitch for some product or service. The attention and cooperation of the
interviewee can be obtained with simple questions that can be readily answered before
asking difficult questions or questions eliciting personal information such as the
respondent’s income group.
Target Population. Because the population of interest will usually be repre-
sented by either the size of the water meters or type of water customers, it is relatively
simple for the utility to specify a target population and to determine the individuals
in that population, along with their mailing addresses.
Even these lists, however, are likely to have errors of classification and
omissions. Some small number of water users may have connected to the system
without going through proper procedures, thereby receiving free water. Others may be
misclassified. For a sample of apartment complexes, for example, some properties
may have changed use without a change in billing record from apartment to personal
residence or commercial use. An enumeration of all water meters assigned to this
category will likely include some properties that have been misclassified, either
because of recording error or change in use. It is also likely that an enumeration will
miss some apartments that were incorrectly classified under other headings. These
random errors usually will not significantly alter the results. Analyzing misclassifica-
tion errors does provide a check on the accuracy of utility billing records that might
be useful in developing corrective measures.
Adding to the challenge, some apartment complexes and larger businesses have
several meters. In some complexes, individual apartments are metered, either directly
by the utility or indirectly by the complex management. Common areas frequently
have water service separate from the residential uses of the tenants.
The problem of identifying the target population becomes substantially more
challenging when the utility is interested in customers with certain characteristics not
observable within the utility accounting system. Such characteristics may include the
ownership of a swimming pool or the use of water filters for household drinking
water. The presence of permanent swimming pools could be determined from aerial
photography or from property assessor’s records; use of water filters could be
determined only by asking individuals within the context of a more general survey.
Nonresponse Bias. An important limitation to the precision of a survey’s

results is nonresponse by some customers to the survey efforts. Nonrespondent
characteristics may be quite different from those of the group answering the
questions.
The cost of a survey rises sharply as efforts are made to increase the response
rate. To the extent that the characteristics being investigated are correlated with water
use per customer, the bias in the results may be less important than for other surveys
because the utility knows the water use of each respondent and nonrespondent. Note
that bias in the results caused by nonresponse would still be a problem even if the
utility attempted to obtain information from the entire population. Increasing the
sample size has no effect on this source of error. Methods of increasing the response
rate for each type of survey are discussed later in this appendix.
Selecting the sample. The simplest sample selection method is a random sample
of the entire population to determine some characteristic of that population, such as
the proportion of homes having low-flow showerheads. In this case, the entire
population is listed (perhaps as a list of all single-family residential water customers in
the utility computer), and a table of random numbers is used to select a sample of the
desired size. Typically, the computer will have a routine to generate these random
numbers and make the sample selection.
Systematic sampling. Frequently, the task of drawing a sample can be simplified
by substituting a systematic sample for a purely random sample. In this case, we
might select every 200th water customer from the utility list of customers. The
selection would be started by drawing a random number between 1 and 200. This
eliminates placing all of the customers in a numeric sequence and drawing random
numbers to determine which ones are in the sample.
A systematic sample will approximate the results of a truly random sample as
long as the individual items are randomly distributed within the overall file with
respect to the variables of interest. This would generally be valid for customers
APPENDIX A 317
arranged alphabetically by name or by arbitrarily assigned billing numbers. Selecting

every tenth house along each street, however, could result in systematically including
too many corner lots or in excluding most corner lots.
Sample size for a proportion. The sample size selected for a survey depends on
the precision desired—not, surprisingly, on the size of the population, unless the
population is relatively small (under 10,000). Sample size is also based on estimates of
the standard deviation of the variable being studied, such as water use. The basic
relationship is that the variability of the sample mean equals σ/N, where σ is the
standard deviation of water use per customer in the population of the customer class
being considered. Furthermore, the sample mean is an unbiased estimator of the
population mean (regardless of the size of the population), and the sampling
distribution tends toward a normal distribution as the sample sizes become larger. All
this means that it is possible to derive simple formulas that link the desired tolerance
of the estimate to be derived from the sample with the variability of the phenomena
under consideration.
As an example, if the proportion of households with low-flow showerheads is to
be determined with an error of ±5 percent, and we are willing to accept a 1 in 20
chance of being outside this range, the sample size would be no more than 400.
Required sample size for a proportion is
t 2 1.96 2
N = PQ -- = ( 0.3 ) ( 0.7 ) --------- = 323 (Eq 15-1)
d 0.05
where:
P =the proportion of the population expected to have low-flow fixtures
Q =1 – P
t =the statistic from the normal distribution indicating the chance of an error
greater than selected (1 in 20 in this example)
d =the acceptable proportion of error in decimal form (5 percent in this
example).
Thus, if 30 percent of the population is expected to have low-flow showerheads,

the required N computed in Equation A-1 is 323. If a response rate of 40 percent is
expected, 808 questionnaires would be sent out. The sample size could be reduced if
the total population being sampled comprised fewer than 10,000. The formula for
reducing sample size for small populations, known as the finite sample correction
(FSC), is:
population – sample size-

---------------------------------------------------
FSC = (Eq 15-2)
population – 1
which reduces the estimate of N slightly. For example, if the total population were
4,000, the finite sample size correction factor is 0.96, which would reduce the sample
size from 323 to 310.
If separate results are required for various groups within the total population,
several options exist. If it can be determined which customers are in each group
before doing the survey, a stratified random sample can be used, and the sample size
within each group can be determined in the same way as for a single population.
Thus, if the utility wishes to obtain results by customer group, and these groups can
be identified through the utility’s computerized records, the task is straightforward.
For example, single-family residential customers might be divided into low-,
medium-, and high-volume water users. Each customer would be placed into the
appropriate category using utility billing records, and a random sample would be
drawn from each category. As an alternative, the sample could be stratified by census
block or tract (perhaps based on census-reported income levels), by postal zip code,
or by political district.
When customers will be placed into subcategories based on their survey
responses, it is more difficult to determine the needed sample size. In this case, the
analyst needs to make some initial guess or estimate of the proportion of the
population in each subcategory, and select a large enough sample of the total
population to make the results for each subcategory statistically valid. For example,
suppose that the analyst plans to divide customers into those with one bathroom, two
bathrooms, or three or more bathrooms. (Note that this will require a definition of
what fixtures are necessary to constitute a bathroom.)
An estimate could probably be obtained for this distribution from real-estate
agents, home builders, or the property tax assessor. Suppose they estimate that
20 percent of the homes have one bathroom, 68 percent have two, and 12 percent have
three or more. To obtain a large enough random sample to ensure statistical validity
for the group with three or more bathrooms, 323 responses are required from this
group. This would require a simple random sample of about 2,700 responses and a
mailing of 6,750, assuming a 40 percent response rate.
If the sample can be stratified ahead of time, perhaps by using property
assessor’s data or census-block data (because home value and number of bathrooms
are highly correlated), the necessity of this large sample could be avoided. Prestratifi-
cation of the sample would be done, however, only if it is less expensive than
increasing the sample size sufficiently to obtain the desired accuracy for the smallest
strata.
MAIL SURVEY
It is relatively inexpensive to do a one-shot mail survey, perhaps sent as an insert in a
random sample of monthly water bills that meet some specified criteria. A more
intense program could reduce nonresponse, but each added step increases cost. For
example, a survey program might include several or all of the following steps
(Dillman 1978):
APPENDIX A 319
1. Mail a postcard or short letter explaining why survey information is

important to the city. It may be helpful to have the message sent by a
recognizable figure such as the mayor.
2. A few days later, mail the survey form.
3. After about 10 days, mail a reminder postcard to nonrespondents.
4. After another 10 days, mail a second questionnaire to those who still have
not responded.
5. Mail a second reminder postcard after another 10 days.
6. Mail a certified letter with another copy of the questionnaire to a sample of
remaining nonrespondents after another 2 weeks.
7. Send a trained interviewer to visit a sample of remaining nonrespondents
after another 2 weeks.
The reason for intensively pursuing information from only a sample of
nonrespondents is to hold down the total cost while still determining the extent to
which previous nonrespondents differ from those who initially responded.
Few survey efforts would go to the extent described, unless there was reason to
believe that nonrespondents were quite likely to be different with respect to the
activities of interest than respondents. Any survey that includes follow-up requires
keeping track of who submitted each completed questionnaire and extending
appropriate assurances of confidentiality to those answering the questions. Tracking is
also essential if the survey responses are to be matched with the utility’s water-use
records. Survey information can also be enhanced by matching it with property
assessor’s records of the physical characteristics of each property. If such follow-up or
matching is to be used, it is necessary to directly tell the potential survey respondent
and give appropriate assurances of the confidentiality of the responses. This requires
follow-through to make sure the responses remain confidential, are used only for
statistical analysis, and do not single out individual customers.
If the utility plans to single out individual customers based on their responses,
such as for an audit of outdoor water uses, this must be clearly stated in the
introduction to the survey.
TELEPHONE SURVEY
Although obtaining a sample of water customers and their mailing and/or service
addresses from utility billing records is relatively simple, obtaining telephone numbers
to conduct a telephone survey may be difficult.
For surveys of business customers, phone numbers are generally readily
available; the yellow pages of the telephone directory provide a ready source of
possible customers by category. For example, a utility might wish to gather
information from plumbing companies about the water-use rates of faucets,
showerheads, and toilets that they install. The population of plumbing companies is
approximated by the yellow page listings, and a random sample could readily be
developed. As mentioned previously, the cost will escalate as efforts to obtain
responses from nonrespondents increase. Numerous callbacks may be required to
obtain information from some business owners or managers, if they are willing to
respond to questions at all.
An additional cost of telephone or in-person surveys is training and monitoring
the interviewers to maintain accuracy and consistency. Respondents have the
opportunity to ask for clarification of various questions, and their answers may be
open to interpretation by the interviewer. Utilities or others wishing to conduct
telephone surveys frequently hire commercial survey firms for this reason. The survey
of Tucson, Ariz., apartment managers described in chapter 111 was done by a
professional telephone survey company. A notable advantage of this was that data
were entered into a computerized database as the interviewers collected the
information, avoiding the later task of data entry that is encountered in a mail survey.
Keep in mind that there is a considerable setup cost for entering the questionnaire
into the computer system and testing it for errors. An added advantage of the
computerized interview system was that the program allowed irrelevant questions to
be automatically skipped over in real time. If a respondent reported no grass, then the
series of questions on grass irrigation was automatically skipped, taking the
interviewer to the next relevant question. Skipping requires a little more attention
from the respondent when done in a mail survey.
An Example Telephone Survey. To develop the sample for the Tucson
apartment survey, a list of all apartment complexes served by the water utility was
obtained from utility billing records, and then a random sample was drawn from the
list. Obtaining telephone numbers was unexpectedly difficult. The telephone book,
city directory, apartment locator guides, and other sources were used. Finding
telephone numbers for the larger complexes proved to be easy, but those for small
complexes were not always available. Some owners of small complexes do not list a
telephone number in any of the published sources and do not have listed personal
phone numbers. As a result, it was impossible to conduct telephone interviews with
apartment owners or managers whose phone numbers were unknown.
A large number of nonresponses also resulted from no answer or from
answering-machine responses. Virtually all managers or owners who were actually
reached by telephone agreed to participate in the survey. Carrying out the survey
under sponsorship of a major state university may have increased the credibility of the
effort and enhanced the response rate. One unexpected quirk in the survey was that
some individuals managed several apartments in the sample and were interviewed
more than once concerning different properties. It would have been interesting and
useful to keep track of these multiple responses to see if the manager’s answers
differed across their properties. Essentially identical answers would raise the question
1. Some results from this study are provided on the data CD.
APPENDIX A 321
of whether they were able to supply answers for each managed property or were
combining their experiences at several properties. If the project budget had been
larger, an interviewer could have gone to a sample of apartments with no listed phone
number or where no response was obtained.
An Alternative Approach. Another approach to telephone survey sampling
is to take a random sample of all telephone numbers in the survey area (excluding
emergency numbers, hospital numbers, and the like), and “qualify” the person
answering the phone. Qualify means to determine if the telephone is associated with
the target population, such as a single-family residential customer. This requires a
much larger number of total phone calls, but with automated dialing equipment, this
may not be a problem. Having a real person on the phone, however, is undoubtedly
more effective in obtaining responses. Few people will listen to a machine unless they
really want to make or receive a specific phone call.
SUMMARY
When a water utility decides that it needs information about its customers beyond
that generated in the normal course of business, direct data collection is a viable
option. Information about customer behavior and attitudes can be obtained directly
from customers by using sampling techniques. Such data can then be used to improve
utility decision making about conservation programs, pricing initiatives, and other
policies.
Long-range planning can be aided by using intuitive information gathering and
analysis. Popular approaches include Delphi techniques, scenario building, and
customer surveys. Gathering information from as wide a range of sources as possible
will aid the utility decision makers in making the best possible assessment of future
long-term trends. Such data can help avoid costly policy blunders that could result in
(1) large excess capacity and its accompanying cost, (2) insufficient capacity that
would deprive customers of their accustomed level of service, or (3) unacceptable
changes in water quality.
REFERENCES
Armstrong, J.S. 1985. Long Run Forecasting, From Crystal Ball to Computer. New York: Wiley Interscience.
Cochran, W.G. 1963. Sampling Techniques. New York: John Wiley & Sons.
Dillman, D.A. 1978. Mail and Telephone Surveys—The Total Design Method. New York: Wiley
Interscience.
APPENDIX B
AWWA Water Demand

Survey Findings
This appendix contains findings of the AWWA Water Demand Survey (WDS), which
was conducted via the Internet during a 73-day period from Monday, September 11,
2006, to Wednesday, November 22, 2006. During this time, 662 completed responses
were received.
To begin the WDS, AWWA provided a master list of approximately 3,269 North
American water utilities, predominantly from the United States. After defunct e-mails
were eliminated, the master sample constituted 2,815 utilities. These included several
score of the largest water supply systems in North America, more than 400 midsize
systems serving 100,000 to 500,000 customers, and many smaller systems. The WDS
achieved a representative mix of utilities by size and geographic area for the United
States, plus responses from Canadian utilities.
The size of utilities participating in the WDS ranged from less than 10,000
persons to cities of one million persons or more. The breakout included
• Four large metropolitan utilities serving one million persons or more
• Five utilities serving populations of between 1 to 1.5 million persons
• 70 utilities serving between 100,000 to 500,000 persons
• 55 systems in the 50,000 to 100,000 range
• 217 utilities serving between 10,000 to 50,000 persons
• 311 utilities serving populations of 10,000 persons or fewer
Table B-1 gives the geographic distribution of utilities participating in the WDS.
323
Table B–1 WDS Participants by US State or Canadian Province
AK 6 NB* 6
AL 18 NC 12
AR 3 ND 4
AZ 14 NE 10
BC* 6 NH 5
CA 74 NJ 7
CO 27 NL 1
CT 4 NM 5
DE 1 NV 5
FL 20 NY 19
GA 10 OH 17
HI 1 OK 1
IA 14 ON* 12
ID 4 OR 24
IL 28 PA 21
IN 24 RI 1
KS 4 SC 7
KY 9 SD 4
LA 6 SK* 1
MA 13 TN 21
MB* 1 TX 41
MD 1 UT 12
ME 6 VA 17
MI 19 VT 1
MN 17 WA 27
MO 13 WI 14
MS 1 WV 1
MT 4 WY 3
* Canadian Province
In total, participating utilities served a population of 39,383,758 persons.

The average value for gallons per capita per day in the sample is 173.52. The
peak-day to average-day demand ratio is 1.91.
The following tables, taken directly from the survey results, tabulate results to
individual questions.
APPENDIX B 325
1. Which types of forecasts are used for planning supply and treatment
capacity in your water system? Please indicate Yes (Y) or No (N).
(Percentages)
Yes No
Annual per capita water-demand 65.6 34.4
forecasts
Annual water-demand forecasts by
major customer class (residential, 58.1 41.9
industrial, commercial, etc.)
Peak-day forecasts 74.0 26.0
Monthly system water-demand 65.7 34.3
forecasts
Daily water-demand forecasts 66.0 34.0
Revenue forecasts linked with 57.9 42.1
water-demand forecasts
1.1 Annual per capita water-demand forecasts.
Response Count Percent

Yes 375 65.6
No 197 34.4
1.2 Annual water-demand forecasts by major customer class (residential,

industrial, commercial, etc.).

Yes 319 58.1
No 230 41.9
1.3 Peak day forecasts.

Yes 436 74.0
No 153 26.0
1.4 Monthly system water-demand forecasts:

Yes 366 65.7
No 191 34.3
1.5 Daily water-demand forecasts.

Yes 367 66.0
No 189 34.0
1.6 Revenue forecasts linked with water-demand forecasts.

Yes 314 57.9
No 228 42.1
2. Basic Water System Facts. Please check True (T) or False (F).
(Percentages)
True False Not Sure
Our system has universal metering; 90.9 8.9 0.2
all customers are on meters.
Our water system promotes water
conservation through education 53.9 43.6 2.5
and other programs, incentives, or
rebates.
Our water system has experienced a
supply emergency in the past three 19.7 79.7 0.6
years.
Our water system sells or delivers
water to wholesale water customers 33.4 65.9 0.6
outside our primary utility service
area.
Our number of customers has 84.1 15.1 0.8
grown in the past five years.
New water sources for our water 80.3 11.6 8.1
system will be more expensive.
We anticipate total system water
production to increase by at least 70.6 18.6 10.8
10 percent over the next 10 years.
APPENDIX B 327
2.1 Our system has universal metering; all customers are on meters.

True 591 90.9
False 58 8.9
Not Sure 1 0.2
2.2 Our water system promotes water conservation through education and
other programs, incentives, or rebates.

True 349 53.9
False 282 43.6
Not Sure 16 2.5
2.3 Our water system has experienced a supply emergency in the past three
years.

True 127 19.7
False 515 79.7
Not Sure 4 .6
2.4 Our water system sells or delivers water to wholesale water customers
outside our primary utility service area.

True 216 33.4
False 426 65.9
Not Sure 4 0.6
2.5 Our number of customers has grown in the past five years.

True 545 84.1
False 98 15.1
Not Sure 5 0.8
2.6 New water sources for our water system will be more expensive.

True 518 80.3
False 75 11.6
Not Sure 52 8.1
2.7 We anticipate total system water production to increase by at least

10 percent over the next 10 years.

True 456 70.6
False 120 18.6
Not Sure 70 10.8
3. In recent years, per capita water demand in your water system has been

Increasing 218 31.4
Level 209 32.7
Decreasing 153 23.9
Without a clear trend 9 .2
4. In the past 10–20 years, the ratio of peak-day water use to average-day
water use in your water system has been been

Increasing 217 35.1
Level 187 30.2
Decreasing 121 19.5
Without a clear trend 4 5.2
APPENDIX B 329
5. Which (if any) of the following water-conservation programs are being

implemented in your water system? Please mark all that apply.

Public education programs 365 55.3
Rebates for installation of low-flow or 67 10.2
ultra low-flow toilets
Free distribution of shower flow restric- 117 17.7
tors and toilet dams
Rebates for front-loading washing ma- 49 .4
chines
Rebates for other water-efficiency appli- 28 4.2
ances, hardware, or fixtures
Conservation rates, such as increasing
block rates, seasonal rates, water bud- 201 30.5
gets
Programs to affect landscaping 147 22.3
Emergency water restrictions 209 31.7
None of the above 176 26.7
6. If you imposed emergency water restrictions, please check all years that
apply.

2004 63 9.5
2005 61 9.2
2006 66 10.0

GLOSSARY
WATER MEASUREMENT
AF = acre foot = 325,851 gal = 43,560 ft3 = 1,234 m3
CcF = hundred ft3 = 748.1 gal = 2.832 m3
ft3 = cubic foot = 7.481 gal = 28.32L
gal = gallon = 3.785L = 0.003785 m3
gpcd = gallons per capita per day
gpd = gallons per day
L = liter = 0.2642 gal = 0.03531 ft3
mgd = million gallons per day
m3 = cubic meter = 264.2 gal = 35.32 ft3
TERMS AND ABBREVIATIONS

aggregation. The combining of individual observations to obtain a new set of data
representing the total or average values of the original set.
ARIMA. Mixed autoregressive integrated moving average model used for forecasting a
single time series using only data from the history of that time series itself.
autocorrelation. The correlation or relationship between successive observations in a
time series. The correlation of the series with itself lags by one period.
average price. Applied to water, the customer’s total water bill divided by the quantity of
water purchased.
b or β-coefficient. In a multiple regression model, the estimated relationship between
the dependent variable and an individual independent variable; the slope of the
function between two variables.
backcast. An estimate of the values of a data series for a time period prior to that used
in estimating the model. Backcasts are sometimes used to test the validity of a
model that will be used for forecasting.
block rate. A rate structure in which the marginal price of water changes based on the
quantity purchased. The first units of water have an initial price, and water use
within each successive block has a different price.
331
Box-Jenkins model. See ARIMA.

business cycles. Recurring multiyear fluctuations in the economy, consisting of an
economic expansion eventually followed by a recession or contraction. Cycles vary
in length from less than two years to more than eight years.
causal/structural model. A model in which variations in the variable of interest (the
dependent variable) are explained by variations in one or more independent
variables.
climate. The long-term tendency of temperature, precipitation, and wind speed, among
others, that characterizes a particular region.
coefficient model. A model in which the water-use rate per capita (the coefficient) is
multiplied by total population to forecast total future water use.
conditional forecast. Using projected values for the period being forecasted; for
example, these values might include expected future prices for water and personal
income.
confidence interval. Used to illustrate the precision of an estimate or forecast. The
confidence interval might include the range in which the value of a variable is
expected to occur 95 times out of 100. Thus a 95 percent confidence interval leaves
a 5 percent chance that the true value is outside the interval shown.
correlation. The degree to which two data series are related. A correlation coefficient of
1 means that the two series are perfectly related (both change together in exact
unison). A coefficient of 0 indicates no relationship. A coefficient of –1 shows an
exact inverse relationship.
cross-section data. Data on a number of different units such as households, census
tracts, or water utilities for the same time interval, such as a year; generally used to
explain differences among these units.
decomposition. The strategy of breaking a problem into subproblems, solving them, and
then recombining the individual solutions to get an overall result.
decreasing block rates. A rate structure in which the price per unit falls as more is
purchased. The first block of water used has the highest marginal price, and each
successive block has a lower price.
degrees of freedom. The total number of observations (such as months of data) minus
the number of coefficients estimated by some statistical procedure such as multiple
regression.
Delphi. A method of obtaining forecasts from a panel of experts.
demand. As used in economics, the quantity of a product that will be purchased at each
possible price, given that all other prices are held constant and considering
consumer’s incomes, stocks of durable equipment, habits, and preferences. Also
referred to as the demand curve.
dependent variable. The variable of primary interest to the analyst. The variable that is
being forecasted. The variable on the left-hand side of a regression equation.
GLOSSARY 333
dummy variable. An variable that only takes on values of 1 or 0, representing the

presence or absence of a particular condition. For example, a drought dummy
variable would have a value of 1 during a drought and a value of 0 during normal
years.
Durbin-Watson statistic. A measure of the correlation of successive error terms from an
estimated time-series equation (between errors in time, t, and time t-1). A value of
2 shows that no autocorrelation is present. Values smaller than 1.5 or larger than 2.5
indicate a problem that should not be ignored.
elasticity. An expression of the relationship between two variables showing the
percentage change in the dependent variable for a 1 percent change in the
independent variable. For example, the price elasticity of demand shows the
percentage of change in water use for a 1 percent change in water price.
end-use model. A forecasting model developed by adding up the end uses for water, such
as toilet flushes, showers, lawn irrigation, and so forth.
endogenous variable. See dependent variable.
error term. The difference between the actual value of a variable and its estimated value
from a statistical model.
evapotranspiration. The water loss from leaf surfaces and soil; used to measure
irrigation requirements.
exogenous variable. See independent variable.
explanatory variable. See independent variable.
extrapolation model. A forecasting model in which past variations in the dependent
variable are used to estimate future values.
exponential smoothing. A univariate forecasting model that removes the irregularities
from a historical time series to create a forecast. This is a special case of the more
general ARIMA models. Used for short-term forecasts.
forecasting. Estimating future values of a dependent variable in “unknown” situations.
F-statistic. In regression, used to describe the statistical significance of the entire model.
Also used generally in statistics to test a hypothesis.
forecasting model. A model used to obtain forecasts.
goodness of fit. How well the values estimated by a model fit the actual known data
values. A number of different statistical measures are used to assess goodness of fit.
heteroscedasticity. A condition in which the general size of the error term from an
estimated model varies over the range of data in a systematic manner. The error
may be small when the variables in the model are small and large when they are
large.
homoscedasticity. A condition in which the variability of the error term is constant over
the entire range of data.
hypothesis. The statement of a relationship that is to be tested statistically.
increasing block rates. A term describing a situation in which the price per unit
increases as the quantity purchased increases. The first units purchased carry a low
price, and each successive block of use is increasingly expensive.
independent variable. A variable used to help explain variations in a dependent variable.
A change in an independent variable is said to cause a corresponding change in the
dependent variable. A causal variable.
inflation. A reduction in the value of money, measured as the percentage increase in the
weighted market prices of a basket of consumer goods.
interaction. A situation where the relationship between X and Y depends on the level of
X. For example, the price elasticity of demand for water may increase as the price of
water increases.
intercept dummy variable. See dummy variable.
logistic curve. A growth curve where the percentage of growth gets smaller over time;
the S curve. It can be computed from 1/Y = 1/a + cXt, where Y is the dependent
variable, a is the asymptote, c is a coefficient to be estimated, X is the growth factor,
and t represents time.
log-normal variable. A variable that has a normally distributed logarithm.
long-term forecast. A forecast for time periods generally 10 or more years ahead.
MAD. Mean absolute deviation; the average error, ignoring the sign.
MAPE. Mean absolute percent error; a relative error measure that weighs errors
according to their absolute size.
marginal price. The price of the last unit purchased. It may be lower than the average
price when a water-rate schedule includes decreasing block rates but higher when it
includes increasing block rates.
mean. Arithmetic average.
mean absolute deviation. See MAD.
mean absolute percent error. See MAPE.
mean squared error. The average of the squared errors of a forecast or estimate.
Squaring puts more weight on larger errors than the otherwise similar mean
absolute deviation, and counts positive and negative errors equally.
measurement error. The extent to which errors result from shortcomings in the way that
numbers are assigned to a concept. Faulty water meters or erroneous meter readings
would result in errors in water-use data.
measurement model. A model used to estimate relationships among variables. These
relationships can then be used in subsequent forecasting models.
median. The observation in the center of a distribution; half the values in the series are
larger and half are smaller.
medium-range forecast. Generally includes a time period ranging from a few years up to
10 years.
GLOSSARY 335
microdata. Data on individual customers; the opposite of aggregated data.

mode. The category with the most observations.
model. A representation of the most important characteristics of the real world for the
purpose at hand.
model specification. The choice of the set of independent variables and their form to be
used in a model.
MSE. See mean squared error.
multicollinearity. The extent to which two or more independent variables are correlated
or vary together. High multicollinearity among independent variables makes it
difficult or impossible to separate the individual effects of the variables.
Normal distribution. The bell-shaped frequency distribution for a variable. This
distribution is symmetric around its mean value, which is also the median in this
case.
observation. Measurements for a given unit (individual customer or water utility) for a
given time period such as a day or month.
outlier. An observation on a variable that is unexpectedly large or small compared to the
remaining observations and the analyst’s expectations.
parameter. The estimated relationship or descriptive statistic from a computation.
peak demand. The greatest amount of water delivered for a particular time interval. For
example, the peak hour for a system might be from 5 p.m. to 6 p.m. on July 23.
price elasticity of demand. The relative responsiveness of demanded quantity to a
change in price. The percentage change in quantity caused by a 1 percent change in
price.
price elasticity of revenue. The relative responsiveness of revenue to a change in price.
The percentage of change in revenue resulting from a 1 percent change in price.
Equal to 1 plus the price elasticity of demand.
principle of composition. The idea that forecast accuracy is improved by disaggregation
with separate forecast models for each category.
principle of noise. The idea that it is easier to predict an average characteristic of a group
than the specific characteristic of one selected individual. Also called the law of large
numbers.
R2. The coefficient of determination of a regression showing the proportion of variation
of the dependent variable accounted for by variations in the independent variables.
regression. A statistical method of estimating the relationships among several indepen-
dent variables and a dependent variable that we are attempting to explain.
reliability. The extent to which a repetition of the measurement will yield the same
result.
seasonality. The recurring variation in temperature, rainfall, and water use (or any other
variable of interest) over the year.
scenario. A cohesive story about future events to be used in developing a specific forecast
of the variable of interest.
sensitivity analysis. A type of analysis in which the values of one or more independent
variables are changed to allow observation of the effect on the dependent variable.
serial correlation. The correlation of a data series with itself lagged by one or more
periods; measures the extent to which larger (smaller) than average values are
followed by additional larger (smaller) than average values.
simultaneous model. A model in which the dependent variable in one equation appears
as an independent variable in another equation. For example, under block rate
pricing, the marginal price of water depends on the quantity taken and the quantity
also depends on the price.
short-term forecast. A forecast for a period ranging from a few hours or days to a few
years.
slope dummy variable. An ordinary continuous independent variable multiplied by a

dummy variable, resulting in a variable that is either 0 or its normal value.
specification of the model. The list of independent variables to be used to explain

variations in the dependent variable.
specification error. The omission of one or more important independent variables from
a model.
standard deviation. A statistical measure of the dispersion of the values of a variable.
standard error of an estimated coefficient. A statistical measure showing the significance

of an estimated coefficient.
standard error of an estimated equation. A statistical measure of the “typical” difference

between the true value and the estimated value of a data series.
statistical significance. The degree to which a result is likely to have occurred by chance.
Thus a variable that is significantly greater than 0 at the 0.10 level incurs a 10
percent chance that the true value is 0 or negative in value.
std. dev. See standard deviation.
step rates. A rate structure in which marginal price changes with quantity in many small
steps but all water purchased by each customer is invoiced at the customer’s final
step.
t-statistic. A statistic showing the significance of the coefficient for the relationship
between two variables. The t-statistic in regression is equal to an estimated
coefficient divided by its standard error. As a rule of thumb, the t-statistic should be
greater than 2 for a useful relationship.
R. Bruce Billings
R. Bruce Billings is professor emeritus of economics, having
taught at the University of Arizona for 35 years. His teaching
and research interests include economic and business forecast-
ing, water resources, public finance, and managerial economics.
His research has been published in refereed journals including
Land Economics, Water Resources Research, Journal AWWA, the
National Tax Journal, the Journal of Regional Science, the Journal
of Water Resources, Planning and Management, and the American
Journal of Economics and Sociology. He has also reviewed many articles submitted for
possible publication to these journals.
Bruce has served on project review committees for Tucson Water, the
Universities Council on Water Resources, the US Geological Survey, the Water
Resources Research Institute, and the State of Arizona. His consulting clients have
included the states of Arizona and Hawaii, Tucson Water, the Bureau of Reclamation,
and the Idaho National Engineering Laboratory. His concern for water conservation
has led him to install a 2,000-gal rainwater storage tank for irrigation water in his back
yard.
Clive V. Jones
Clive Vaughan Jones is Managing Director of Economic Data
Resources LLC. As an economist and forecast analyst, he has 25
years of professional experience for public and municipal
agencies and private business. Clients include City of Calgary
Water Works, the Chicago Water Department, RW Beck Engi-
neers, New Century Energies, Tri-State Generation and Trans-
mission Company, Denver Water, Wyoming Water Development
Commission, Alcatel, Agilent Technologies, Hewlett Packard,
US Environmental Protection Agency, and the US Department of the Interior.
Clive has served on project panels for the Awwa Research Foundation
(“Effectiveness of Residential Water Conservation Price and Nonprice Programs,”
Michelson et al. AwwaRF, 1998) and on the AWWA Rates Committee. He has written
extensively on the topic of water-demand forecasting and explaining water use with
demographic, socioeconomic, and technological factors.
337

Index
NOTE: f. indicates figure; n. indicates footnote; t. dishwashers, 120

indicates table. evaporative coolers, 121
garbage disposals, 120–121
reverse-osmosis water filters, 121
A water softening, 121
ARIMA models, 34, 86, 283–284, 331
Accuracy, 39–40 @RISK, 294
and complexity, 123 Autocorrelation, 132
disaggregation for improving, 270–271 defined, 331
as an evolving goal, 307–308 Average (mean), 53–54
improving, 306 Average price, 159, 288–289
with regard to population and employment, defined, 331
122 simultaneity bias, 289
Aggregation, 331 AWWA. See American Water Works Association
American Meteorological Society, 127, 133–134 AwwaRF, 24
American Statistical Association, 123
American Water Works Association
data on forecasting methods and models in B
use, 31–32, 32t.
research on Energy Policy Act of 1992, 28 ß-coefficient, defined, 331
Water Demand Survey, 2, 2t., 2n., 19 Backcast, 331
Water Demand Survey results, 323–330 Bias, 305
AMS. See American Meteorological Society Billing records, 45, 46
Analysis Boulding, Kenneth, poem on water, 15
adequacy of, 79 Box-Jenkins models, 11, 34, 284
consistency of assumptions, 79 Business cycles, defined, 332
by customer class, 2 Butterfly effect, 133
risk (Monte Carlo simulation), 79–80
sensitivity, 79–80, 301
simple, 51–57 C
spreadsheet formats, 51, 52t.
Appliances California, global warming impact database, 3
clothes washers, 120, 195–196 California Urban Water Conservation Council, 28
conservation manuals, 239
339
Carbon dioxide, 140–141 steps in designing, 313–314

Causal/structural models systematic sampling, 316–317
defined, 332 target population, 315–316
See also Regression modeling by telephone, 319–321
Chaotic process, 127
CIEW. See Commercial and Institutional End
Uses of Water D
Climate
and carbon dioxide, 140–141 Data, 45, 58
change, 140–141 average (mean), 53–54
climatographs, 135–140, 136f.–139f. climate, 50
defined, 332 collection, 45–47
distinguished from weather, 127 commercial and industrial production,
global models, 140, 141–142 50–51
and global warming, 3, 127–128, 140–141 composition, 48, 49
heat-island effects, 140 confidence interval, 54–55
implications of change for demand fore- cross-section, 48, 58
casts and water utility planning, 142, 143 customer records, 46–47
medium-term predictions, 140 disaggregation, 47–49
and water demand, 9 employment, 50–51, 77
Coefficient models, 332 flow variables, 47
Commercial and Institutional End Uses of Water, formats for spreadsheet analysis, 51, 52t.
25, 26t. frequency distributions, 51, 53f.
Composition (data), 48, 49 graphs, 51, 53f.
Conditional forecasts, 332 irregular distributions, 56
Confidence intervals maximum and minimum values, 52–53
for data, 54–55 median, 54
defined, 332 missing values, 49
in regression modeling, 104, 104n., 257, mode (modal value), 54
302–303, 303n. noise, 48–49
and weather, 134 organizing, 47–48
Correlation, defined, 332 outliers, 49
Cross-section data, 48, 58 peak-to-average use ratio, 56, 57f., 57t.
cross-section, time-series regression personal income, 50
models, 271–272 pooled cross-section and time-series, 48
defined, 332 population, 50
graph, 53f. simple analysis, 51–57
pooled with time-series data, 48 simple correlations, 55, 56t.
spreadsheet format, 51, 52t. standard deviation, 54, 63, 336
Customer records, 46–47 standardized values, 62, 63t.
Customer surveys, 300–301, 311, 321 status variables, 47
customer feedback, 311–312 system deliveries, 46
by mail, 318–319 time-series, 45, 47–48, 52t., 58
nonresponse bias, 316–318 variance, 54
questionnaire development, 314–315 weather, 50
random sampling, 316 Data loggers, 3, 25
reviewing questions, 315 DCC model. See Discrete-continuous choice
sample size for a proportion, 317–318 model
INDEX 341
Decomposition, 332 Austin (Texas) case study, 248–249

Degrees of freedom, 332 Blacksburg (Virginia) case study, 250
Delphi method, 301, 312–313 central California case study, 242, 242t.
defined, 332 Colorado (Front Range) case study, 241–
Demand, defined, 332 242
Demand forecasting contingency plan elements, 239–240
analysis by customer class, 2 Corpus Christi (Texas) case study, 245–246
and capital investments, 5 evaluating program effectiveness, 241
and climate change, 142, 143 goal achievement process, 239, 240f.
and conservation, 6 Hawaii case study, 249–250
and data loggers, 3, 25 Iowa case study, 246–247
defined, 1 model contingency plans, 239
end-use models, 11 San Francisco (California) case study, 244–
forecast evaluation, 304–306 245, 245t.
hybrid models, 11 Santa Barbara and Goleta (California) case
incorporating water conservation estimates, study, 248
211–214 Southern California case study, 242–244,
interval estimates, 302–303 243t.
long-term, 6–7, 7t., 301, 307, 334 Utah case study, 250
medium-term, 7, 7t., 301–302, 307, 334 See also Emergency conservation
per capita use per customer class, 10–11 DSS model, 11, 32, 33
purposes, 1, 3–4, 5–6, 37 Dummy variables, 333
regression models, 11, 39, 41 Durbin-Watson statistics, 333
short-term, 7–8, 7t., 300–301, 307, 336
simple method (per capita use times
population), 2, 10, 31, 42 E
and socioeconomic factors, 11
steps, 299 Economic forecasts, 109, 116–117
and system optimization, 6 checklist, 122–123
time-series models, 11 effect of inflation, 117–118
types reported in Water Demand Survey, 2t. and housing composition, 118
very-short-term, 7t., 8 income trends, 118
and weather variables, 142–143 and medium-term forecasting, 301–302
See also Forecasting methods and models; and personal income, 117
Per capita water-demand forecasts; Survey of Professional Forecasters, 116
Sectoral water-demand forecasts El Niño/Southern Oscillation, 134
Demand hardening, 232–233, 233t. Elasticity, 333
Demand hypothesis, 4–5 Emergency conservation, 228, 251
Demand management, 1 AWWA Water Demand Survey results, 330
Demand monitoring, 8 changing to increasing block rates, 238–
Dependent variables, 332 239
Disaggregation, 37–38, 47–49, 218–222, 219f., choice of program, 232
270–271 and demand hardening, 232–233, 233t.
Discrete-continuous choice model, 291–293, drought funds, 229
293t. education and publicity, 233–234, 242–
Double logarithmic functional form, 171 244, 243t.
Drought planning, 239 effectiveness, 228, 229t.
Athens (Greece) case study, 247–248 effects on revenue, 229–230
and faucet aerators, 234 Expert opinion, 312–313

and flow-reducing devices, 234 Exponential smoothing models, 11, 284–285,
hydro-illogic cycle, 227f. 333
incentives, 230–231 Extrapolation models, 333
increased water and sewer rates, 236–238,
242, 247–248, 249–250
leak detection and repair, 234–235 F
and low-flow showerheads, 234
mandatory limits on outdoor water use,
235, 241–247, 243t., 245t., 249–250 F-statistics, 257–258, 276, 333
metropolitan area coordination, 232, 244– Federal Reserve Bank of Philadelphia, 116, 123
245, 245t. Flow variables, 47
price penalties for excessive use, 238, 246– Flow-reducing devices, 234
247, 250 Forecast Pro, 32
and rate increases, 230 Forecasting, defined, 333
and synergy and competition among pro- Forecasting methods and models, 33–34, 41–
grams, 233 42
techniques, 230–232 accuracy, 39–40, 306
trigger points for restrictions, 248–249 and automatic meter-reading systems, 41
and utility funding sources, 229–230 AWWA Water Demand Survey results, 325–
voluntary programs, 235, 241–244, 243t., 326
246–247, 249–250 complexity and accuracy, 123
water budgets for customers, 236 comprehensive disaggregated causal/struc-
See also Drought planning tural model for water conservation effects,
218–222, 219f.
and customer disaggregation, 37–38
E and data availability, 38–39
and data frequency, 39
Employment, 109 data on current use, 31–33, 32t.
accuracy issues, 122 defined, 333
data, 50–51, 77 estimating, 40–41
forecasts, 114–116, 115t. extrapolation, 34–35
offshoring and outsourcing, 115–116 and forecasting purpose, 37
regional and local forecasts, 116 fuzzy logic, 35, 86
scenarios, 123–124, 301–302 for long-term forecasts, 301
and water demand, 9 for medium-term forecasts, 301–302
End-use models, 333 multivariate regression, 35, 39, 41
Energy Policy Act of 1992, 28 neural networks, 35, 86
ENSO. See El Niño/Southern Oscillation nonparametric methods, 35
Environment Canada, 128 sectoral models, 37–38
Error terms, 333 selecting, 35–36, 36n., 41, 299–302
Evaluation methods, 304 for short-term forecasts, 300–301
bias, 305 software, 32–33
mean absolute deviation, 304–305 subjective methods, 34
mean absolute percent error, 305 See also DSS model; Forecast Pro; IWR-
mean squared error, 304 Main; Per capita water-demand
selecting, 305–306 forecasts; Regression modeling; Sectoral
Evapotranspiration, 333 water-demand forecasts
INDEX 343
Frequency distributions, 51, 53f. for selected cities, 188, 189t.

Fuzzy logic, 35, 86 typical single-family residential indoor water
uses, 189, 190f.
water use by turf and xeriscape, 198, 199f.
G Inflation, 334
Interaction, 334
Global climate models (GCMs), 140, 141–142 Interval estimates, 302
Global warming, 127–128, 140–141 and alternative scenarios, 303
database, 3 regression-generated confidence intervals,
Goodness of fit, 93, 254, 333 302–303, 303n.
Graphs, 51 and risk simulation, 303
cross-section, 53f. Irregular distributions, 56
residential and CI water-use rates, 75, 76f. IWR-MAIN, 11, 32–33
time series, 53f.
Graywater reuse (on-site), 200–201
L
H La Niña, 134
Landscape conversion, 197–199, 199f.
Household regression model, 259 Law of Demand, 4
choice of variables, 260–262 Law of large numbers, 89, 89n.
coefficient confidence intervals, 265–266 Leakage reduction, 188–190
comparison of descriptive statistics for in emergency conservation, 234–235
variables, 262, 262t. Log-normal variable, defined, 334
correlation matrix for residential microdata, Logistic curve, defined, 334
262, 263t. Lorenz, Edward, 133
equation and results, 262–264, 264t. Lotus, 51
forecast confidence interval for total Low-flow fixtures
population, 267–269 for emergency conservation, 234
frequency distribution of water use, 263, faucets, 196, 197f.
263f. and offsetting behaviors, 196–197
preliminary data analysis, 262, 262t., showers, 190, 194–195, 195t.
263f., 263t. toilets, 191–193, 191f., 193t., 194t.
sample size, 259–260 urinals, 193–194
significance levels of estimated coefficients,
265
single household forecast, 266–267 M
Hypothesis, defined, 333
MAD. See Mean absolute deviation
Makridakis, Spyros, 123, 307
I MAPE. See Mean absolute percent error
Marginal price, 159, 289–290
Independent variables, 334 defined, 334
Indoor/outdoor use, 24–25, 25t., 29, 83, 90, plus difference, 159–160, 290–291
90f., 124 simultaneity bias, 290, 291
changing indoor habits, 204 Maximum and minimum values, 52–53
mandatory limits on outdoor water use M-Competition, 123
(emergency conservation), 235 Mean, defined, 334
Mean absolute deviation, 304–305 time window, 91

defined, 334 Multicollinearity, 335
Mean absolute percent error, 305 Multivariate regression, 35, 39, 41
defined, 334
Mean squared error, 304
defined, 334 N
Measurement error, 334
Measurement model, 334 National Bureau of Economic Research, 123
Median, 54, 334 National Climate Data Center, 128
Meters and metering National Drought Mitigation Center, 135
automatic meter-reading systems, 41 National Oceanic and Atmospheric
estimating impact on water use, 169–170, Administration, 128
170t. National Research Council, 140
meter-reading records, 47 National Weather Service, 128
and rates, 159–160 NCDC. See National Climate Data Center
and sectoral water-demand forecasts, 61 Neural network models, 35, 86, 285–287, 286f.
and water conservation, 205 backpropagation of errors, 286
and water demand, 27–28, 87 hidden layer, 286
Methane, 140 NOAA. See National Oceanic and Atmospheric
Microdata, 335 Administration
Microsoft Excel, 51, 282, 294 Noise, 48–49
Missing values, 49 Normal distribution, 335
Mitchell, J. Murray, 133 NRC. See National Research Council
Mode (modal value), 54 NWS. See National Weather Service
defined, 335
Model, defined, 335
Model specification O
defined, 335
See also Specification of the model Observation, 335
Models and modeling. See ARIMA models; Box- Outdoor use. See Indoor/outdoor use
Jenkins models; Discrete-continuous choice Outliers, 49
model; DSS model; Exponential smoothing defined, 335
models; Global climate models; Forecasting
methods and models; Neural network mod-
els; Regression modeling; Software P
Monte Carlo simulation. See Risk simulation
Monthly proportions of annual water usage, 83, Palmer Drought Severity Index, 131, 132f.,
84, 85f., 87–89, 88t., 105 131n.
data arrays, 87, 88t. Parameters, 335
data requirements, 89 Peak demand, 21, 22f., 29, 83
Monthly water production, 84, 85f. defined, 335
regression modeling, 93–102 Peak-to-average day ratios, 84, 84f., 86–87, 105
Moving averages, 83, 84, 90–91, 90f., 91t., AWWA Water Demand Survey results, 329
93f., 105 Peak-to-average use ratio, 56, 57f., 57t.
calculating, 91, 92t. Per capita water-demand forecasts, 2, 10–11,
centered, 93f., 91, 92t. 31, 42, 59–60, 80
extrapolating trends, 92 adequacy of analysis, 79
periods, 91 and climate data, 62n.
INDEX 345
comparison with sectoral forecasts, 77–78, and ratio of water bills to income, 164
78t. relatively elastic, 162, 163
conservation and rate impacts, 64–65 relatively inelastic, 161, 163
consistency of assumptions, 79 for residential categories, 172–173, 173t.,
example, 61–62, 61t. 174t.
fundamental equation, 59 of revenue, 161, 335
identifying year of normal weather, 62–64, selecting, 167
63t. and single-family housing, 172, 173t.
long-term, 301 and time, 163–164
long-term weather index, 64 translating revenue objective into rate
monitoring population, new customers, and increase, 166–167
water demand, 80 unitary elastic, 162
reality checks, 60, 79–80 Price. See Rates
risk analysis (Monte Carlo simulation), 79–80 Principle of composition, 335
sensitivity analysis, 79–80 Principle of noise, 335
trends, 64 Public water systems, 17, 17n., 28
See also Sectoral water-demand forecasts water use, 17, 18–20, 18f., 19t., 28
Personal income Pumpage, 59–60
data, 50
and water use, 117, 123
Population, 109, 123 R
accuracy issues, 122
data, 50 Rainwater harvesting, 200–201
forecast checklist, 121–122 Rates, 147–148, 176
forecasts, 110, 110n. adjusting analysis data for inflation, 288
growth (U.S.), 110, 111f. average price, 159, 288–289
monitoring, 80 bills displaying marginal prices, 205–206,
national trends (U.S.), 110–111 206f., 207f.
regional and local forecasts, 113 block, 331
regional trends (U.S.), 112–113, 112t. conservation effects of proportional
scenarios, 123–124 increases in all rates, 206–208, 208t.
and water demand, 9 and consumer-choice theory, 289
Potential evapotranspiration, 130–131, 130n. decreasing block, 150–151, 151f., 152t.,
Precipitation, formula for standardized value, 156t., 332
62–63 discrete-continuous choice model for block
Price elasticity, 177 rate schedules, 291–293, 292t.
cautions, 168–169, 169t. and econometric studies, 170–172, 288–
and commercial users, 174, 174t. 293
computation, 162–163, 166–168, 169t. effect on revenue on 10 percent drop in
defined, 160–161, 335 sales, 155, 156t.
and demand curves, 161, 163 estimating impact of metering and rate
in demand forecasts, 166–169 changes on water use, 169–170, 170t.
estimating quantity impacts and revenue flat, 150
target, 167–168 as function of supply constraints, 236–237
and industrial users, 174–176, 175t., 176t. impact on per capita water-demand
and multifamily housing, 172–173, 173t. forecasts, 64–65
numeric rules of thumb, 164–166, 165t. increase for desired decrease in water use,
perfectly inelastic, 161–162 237–238
increasing block and conservation effects, 302–303, 303n.

183, 183t., 209–211, 210t., 211t., cross-section, time-series models, 271–
238–239 272
increasing block (inverted), 152–153, daily demand, 103, 103t.
152f., 152t., 154–155, 156t., 334 daily water-demand regressions and peak-
increasing marginal price while decreasing day demand, 102
monthly service charge, 208–209 dependent variables, 93, 94, 94n., 254,
and inflation, 237 255
lifeline, 152–153 dropping insignificant variables, 276
marginal cost and rebate, 156–157, 158t. dummy monthly variables, 96–97
marginal price, 159, 289–290, 334 error bands, 104, 104n.
marginal price plus difference, 159–160, error term, 255–256
290–291 estimating, 41, 93
metered service with commodity charges, explanatory variables, 93, 94, 254, 255
159–160 F-statistics, 257–258, 276, 333
monthly service fee and constant volume forward stepwise inclusion, 275–276
charge, 150, 151f., 156t. frequency and disaggregation, 39
penalties for excessive use, 238 functional forms, 93, 97–98
price-related conservation strategies, 184, general concepts, 254–259
185 goodness of fit, 93, 254, 333
and revenue stability, 154–158 heteroscedasticity, 274, 333
schedules, 149–154 in medium-term forecasts, 302
and sectoral water-demand forecasts, 74– monthly residential water use equation,
75 254–255, 255n.
step, 153–155, 154f., 155t., 336 monthly water use, 93–102
total bill, 160 multicollinearity, 95, 274–275
translating revenue objective into rate nonlinear relationships, 258–259
increase, 166–167 normal probability distribution, 104
unmetered service (flat rate), 158 null hypothesis, 256–257
and water demand, 9 omitting selected explanatory variables
See also Price elasticity from regression specification, 274
Reclaimed water, 202–204 pathologies, 273–275
Regression, defined, 335 and policy analysis, 83–84
Regression model with time-series data, 269– preliminary data analysis, 262, 262t.,
270 263f., 263t.
disaggregation for improved accuracy, 270– R2, 258, 276–277, 335
271 regression coefficients, 255
dummy variables for slope and intercept, residuals, 93, 93n.
271 results, 98–101, 100t., 101t.
Regression modeling, 11, 85–86, 105–106, sample database, 96, 96t.
253–254, 277–278 scaling, 94
acid test, 93 serial correlation (autocorrelation), 273–
applications, 101–102 274
applications to forecasting and risk in short-term forecasts, 300
evaluation, 104–105 sign test, 105
backward stepwise regression, 276 specification, 255
basic questions, 92–93 stability of coefficients, 101
confidence intervals, 104, 104n., 257, standard errors, 256, 276–277
INDEX 347
statistical significance, 254 reference (unconstrained) forecast, 66,

statistical significance of the estimated 66f., 69, 71t., 77–78, 78t.
coefficients, 105 risk analysis (Monte Carlo simulation), 79–
stepwise regression, 275–277 80
t-statistics, 256–257, 336 sensitivity analysis, 79–80
trend variables, 97–102 stages of, 66–67, 66f.
unbiased, 272–273 trending commercial/industrial use rates,
weather variables, 94–95 75–76, 76f.
See also Household regression model; trending commercial/industrial water use
Multivariate regression; Regression based on employment forecasts, 77
model with time-series data trending residential water-use rates, 75,
Reliability, defined, 335 76f.
Residential End Uses of Water, 24 two-step computation method, 68–69
Risk simulation, 79–80, 293–297, 295f. typical equation, 60
and interval estimates, 303 use rate times number of customers, 65
and long-term forecasts, 301 and water loss control, 74
and medium-term forecasts, 301–302 and water losses, 65–66, 66n.
risk distribution, 294–295, 296f. See also Per capita water-demand forecasts
Sensitivity analysis, 79–80, 301
defined, 336
S in per capita water-demand forecasts, 79–
80
Scenarios, 336 in sectoral water-demand forecasts, 79–80
Seasonal forecasting, 83 Serial correlation, 132
Seasonality, 335 defined, 336
Sectoral water-demand forecasts, 37–38, 59– Simple correlations, 55, 56t.
60, 80–81 Simultaneous models, 336
adequacy of analysis, 79 Slope dummy variables, 336
breaking out an dseparating large CI Software. See @RISK; DSS model; Forecast Pro;
customer water use, 76–77 IWR-Main; Lotus; Microsoft Excel;
comparison with per capita forecast, 77– Stochastic weather generators
78, 78t. Specification errors, 336
consistency of assumptions, 79 Specification of the model
estimating savings from water conservation, defined, 336
70–74, 73t. See also Model specification
example data, 67–68, 68t. Spreadsheet analysis, 51, 52t.
final or adjusted sectoral forecast, 67f., 67 Standard deviation, 54, 63
69–78, 72t., 77–78, 78t. defined, 336
forecasting household size changes, 77 Standard error of an estimated coefficient, 336
fundamental equation, 65 Standard error of an estimated equation, 336
and inflation, 68 Standardized values, 62, 63t.
long-term, 301 formulas for precipitation and temperature,
and metering, 61 62–63
monitoring population, new customers, and Statistical significance, 336
water demand, 80 Status variables, 47
and outside sales, 65 Stochastic weather generators, 134–135
and rate impacts, 74–75 Summer/winter water use, 83, 90
reality checks, 60, 79–80 Survey of Professional Forecasters, 116
Surveys. See Customer surveys; Survey of Water Conservation Program—A Planning

Professional Forecasters; Water Demand Manual, 239
Survey (AWWA) Water conservation, 181–182, 222–223
aggregate before-and-after analysis, 215–
216
T analyzing results, 185
anticipated savings by device or program,
Technological changes, 9, 109, 119, 123 181, 182t.
appliances, 120–121 automatic timers on watering systems,
clothes washers, 120 199–200
dishwashers, 120 AWWA Water Demand Survey results, 329
evaporative coolers, 121 behavioral programs, 204–211
garbage disposals, 120–121 benefit-cost comparisons, 187–188, 301
and industrial water use, 119–120 benefits axiom, 187n.
reverse-osmosis water filters, 121 bills displaying marginal prices, 205–206,
water softening, 121 206f., 207f.
Temperature, formula for standardized value, 62–63 changing indoor and outdoor habits, 204
Texas Commission on Environmental Quality, 239 characteristics of successful programs,
Texas Water Development Commission, 3 214–215
clothes washers, 195–196
Time index, 282
commercial, industrial, and institutional
Time series, 281
programs, 201–202, 202f.
ARIMA models, 283–284
commercial, industrial, and institutional
cross-section, time-series regression
users and ultra low-flush toilets, 193,
models, 271–272
193t., 194t.
data, 45, 47–48, 58
competition or antagonism between
defined, 281
programs, 186–187, 233
exponential smoothing models, 284–285
comprehensive residential model, 218–
graph, 53f.
222, 219f.
models, 11
control group comparisons, 216–218
pooled with cross-section data, 48
data, 46
regression model with time-series data,
and demand forecasting, 6
269–271
effect of metering, 205
spreadsheet data formats, 51, 52t.
end-use estimates of water-savings
spreadsheet displays, 282
potential, 212
and state-dependent effects, 130n.
estimating effects with local studies, 215
trend models, 281–283, 283f.
estimating savings in sectoral water-
t-statistics, 256–257, 336
demand forecast, 70–74, 73t.
faucets, 196, 197f.
goals as determinants of programs, 213
V graywater reuse (on-site), 200–201
impact on per capita water-demand
Variance, 54 forecasts, 64–65
implementation amoung U.S. utilities, 183,
183t.
W incorporating into forecasting, 211–214
and increasing marginal price while decreasing
Water audits, 188–190 monthly service charge, 208–209
INDEX 349
indoor/outdoor use for selected cities, 188, and employment, 9

189t. equation, 4–5
landscape conversion, 197–199 indoor use, 24, 25t., 29, 124
Las Vegas landscape conversion program, indoor/outdoor use for selected cities, 188,
198–199, 199f. 189t.
and long-term forecasts, 301 and metering, 27–28
low-flow showers, 190, 194–195, 195t. monitoring, 80
manuals, 239 outdoor use, 24–25, 29, 124
mediated engineering, 184 and peak demand, 21, 22f., 29, 83, 335
motivating customers, 183–184 per capita water use (U.S.), 18–20, 18f.,
offsetting behavior, 196–197 19t.
price-related strategies, 184, 185 and population, 9
program evaluation, 183–188 and price, 9
program management, 214–222 residential, 22f., 23–24, 23f., 25, 29
and proportional increases in all rates, seasonal patterns, 21, 21f., 29, 83
206–208, 208t. stabilizing trend in, 18–21, 19t.
and public education and publicity, 183, statistical profiles, 20–21, 20f., 21f., 29
183t., 184–185, 204, 233–234 and technological change, 9
pure engineering, 184 typical single-family residential indoor water
rainwater harvesting, 200–201 uses, 189, 190f.
reclaimed water use, 202–204 variation among residential customers, 23–
regulations, 184 24, 23f., 24n.
structural programs, 188–204 and water loss, 4, 4n., 10, 26–27, 27f.,
synergism between programs, 186, 233 27n.
test statistic, 216, 217 and weather and climate, 9
toilet flush volumes, 191–192, 191f. Water Demand Survey (AWWA), 2, 2t., 19
toilet rebate programs, 191–193 data on forecasting methods and models in
toilet repairs, 190 use, 31–32, 32t.
typical single-family residential indoor water emergency water restrictions, 330
uses, 189, 190f. participants by US state or Canadian
urinals, 193–194 province, 323, 324t.
water audits and leakage reduction, 188–190 ratio of peak-day to average-day use, 329
and water demand, 9–10, 28 results, 323–330
water use by turf and xeriscape, 198, 199f. sampling technique, 2n.
water use in selected cities, 185, 186f. types of forecasts used, 325–326
and wholesale demand forecast, 213–214, water conservation programs, 329
214t. water demand increase/decrease, 329
See also Drought planning; Emergency water system descriptions and facts, 327–
conservation 328
Water demand Water loss
AWWA Water Demand Survey results, 329 and sectoral water-demand forecasts, 65–
commercial and industrial (CI), 25–26, 26t. 66, 66n., 74
components of, 21–27, 29 and water demand, 4, 4n., 10, 26–27, 27f.,
curves, 4, 4n., 5f., 148–149, 148n., 149f. 27n.
daily pattern of residential use, 25 Water measurement units, 331
defined, 4 Water production 59–60
and efficiency and conservation programs, Water shortages. See Emergency conservation
9–10, 28, 29 Water system parameters, 8
Water-delivery records, 46 identifying year of normal weather, 62–64,

WDS. See Water Demand Survey (AWWA) 63t.
Weather interannual forecasts, 134
autocorrelation (serial correlation) in data limits of predictability, 127, 133–134, 142–
analysis, 132 143
average temperature, 128–130, 129f. long-term index, 64
and chaotic process, 127, 133 Palmer Drought Severity Index, 131, 131f.,
confidence intervals, 134
131n.
cycles, 132–133, 301
regression modeling variables, 94–95
data, 50
distinguished from climate, 127 simulations, 134–135
drought cycles, 133 special variables, 130
evapotranspiration (potential or measured), statistical analysis of data, 132
130–131, 130n. stochastic weather generators, 134–135
extreme events, 131, 131f. total precipitation, 128–130, 129f.
forecasting, 133–134 variables, 128–131
heat waves, 132 and water demand, 9

Forecasting Urban Water Demand

Uploaded by

Copyright:

Available Formats

Forecasting Urban Water Demand

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Forecasting Urban Water Demand

Uploaded by

Copyright:

Available Formats

Forecasting

Forecasting Urban Water Demand, Second Edition

20410 cover 1.indd 1 2/6/2008 8:31:10 AM

Science and Technology

Forecasting Urban Water Demand, Second Edition

AWWA Publications Manager: Gay Porter De Nileon

Library of Congress Cataloging-in-Publication Data

6666 West Quincy Avenue

LIST OF TABLES xiii

Summary and Conclusions 142

AUTHOR BIOGRAPHIES 335

CD-ROM OF SELECTED DATA TABLES (INSIDE BACK COVER)

4-1 Time-Series of Water Use by Month..........................................................53

5-1 Simplified Flow of Work Diagram for Sectoral Water-Demand

6-1 Historical Peak-to-Average Day Demand ..................................................84

8-1 Average Daily Temperatures, Atlanta, Ga. ..............................................129

9-1 Typical Linear and Exponential Demand Curves....................................149

9-3 Decreasing Block Rates and Average Price With a $7.00

10-1 Water Use in Cities .................................................................................. 186

11-1 Flow Diagram of Goal Achievement Process ......................................... 240

12-1 Frequency Distribution of Water Use ..................................................... 263

13-1 Historical Population and Trend Extrapolations .................................... 283

3-1 Methods for Developing Water-Demand Forecasts Reported

4-1 Time-Series Data Format ...........................................................................52

5-1 Simple per Capita Forecast Approach ......................................................61

6-1 Data Array for Calculating Monthly Proportions of

7-1 Most Rapidly Growing Metropolitan Areas, 1990–2000 ..................... 112

9-1 Water Rate Schedules ............................................................................ 152

9-6 Detailed Computations for a 200 Percent Price Increase.................... 169

10-1 Anticipated Water Savings by Device or Program ................................. 182

11-1 Projected Savings by Program ................................................................ 229

12-1 Descriptive Statistics............................................................................... 262

13-1 Distribution of Customers by Water Use Under Uniform

B-1 WDS Participants by US State or Canadian Province............................ 322

This page intentionally blank.

This page intentionally blank.

Forecasting Urban Water Demand

Urban water-demand forecasting is the process of making predictions about future

2 FORECASTING URBAN WATER DEMAND

65.4 Annual per capita water-demand forecasts

Many water utilities rely on simple forecasting techniques, such as multiplying

Numerous research projects sponsored by the AWWA Research Foundation

WHAT IS WATER DEMAND?

4 FORECASTING URBAN WATER DEMAND

forecasting is to make forward-looking information available to public water suppliers

The existence of this function for a consumer or group of consumers—sometimes

Figure 1–1 Schematic of Demand Curve

6 FORECASTING URBAN WATER DEMAND

FORECAST HORIZON AND FORECAST APPLICATIONS

Long-Term Water-Demand Forecasts. Long-term water-demand forecasts

4. Forecasts evaluated ex-post can be compared with actual numbers.

Table 1–2 Types of Water-Demand Forecasts and Major Applications

Forecast Type Forecast Horizon Applications

Long-Term Decades, 10–50 years Sizing system capacity, raw

Medium-Term Years to a decade, 7–10 years Sizing, staging treatment

Short-Term Years, 1–2 years Budgeting, program tracking

Very-Short-Term Hours, days, weeks, Optimizing, managing system

8 FORECASTING URBAN WATER DEMAND

Very-Short-Term Water-Demand Forecasts. These forecast models, focus-

Related Applications. Special studies, sponsored by utilities or conducted