Scribd
Upload
41 views

Msba MKTG Analytics Day2

This document provides an introduction to a course on causality and marketing science. It discusses some of the challenges of estimating causal effects from observational marketing data, including endogeneity and selection bias. It outlines several approaches that can be used to estimate causal effects, such as randomized experiments, instrumental variables, differences-in-differences models, regression discontinuity designs, and exploiting natural experiments. The document also provides a high-level overview of key concepts and applications in marketing analytics, including different types of analytics and how marketing analytics can be implemented.

Uploaded by

jshen5
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
41 views

Msba MKTG Analytics Day2

This document provides an introduction to a course on causality and marketing science. It discusses some of the challenges of estimating causal effects from observational marketing data, including endogeneity and selection bias. It outlines several approaches that can be used to estimate causal effects, such as randomized experiments, instrumental variables, differences-in-differences models, regression discontinuity designs, and exploiting natural experiments. The document also provides a high-level overview of key concepts and applications in marketing analytics, including different types of analytics and how marketing analytics can be implemented.

Uploaded by

jshen5
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 21

Intro to Marketing Analytics:

Causality and Marketing Science

Professor Kenneth C. Wilbur

University of California, San Diego


Rady School of Management

1
Let’s discuss: Causal Inference in Economics and Marketing

We often want to estimate the effects of interventions, e.g.


I y is salary
I x is MSBA attainment
I y=xb+e :: what is e?
I bols = cov (x, y )/var (x) = b + cov (x, e)/var (x)
I How should you interpret bols ? What does it mean? What does it not mean?
I descriptive vs prescriptive
Endogeneity is when x and e are likely to be correlated. x is endogenous
I Endogeneity is not testable. e is unobserved. Can you correlate x and e?
I Endogeneity is widely misunderstood

2
OK so what can we do?

Live with the ambiguity, make decisions with your gut


I This approach is easiest and most common. Is it good?
Assume the correlation is causal
I Cheap, easy, fast, quantitative, might be better than gut
I Confounds selection with treatment
I Risky. Hope boss doesn’t understand causality; it’s not your money on the line;
and no one goes back to check your prediction quality
Try to estimate a “causal effect” or “treatment effect”
I Increasingly possible and common in digital environments

3
Basic identity of causal inference

Assume a discrete treatment, such as MSBA attainment


I Observed outcome = Outcome for treated - Outcome for untreated
=(Outcome for treated - Outcome for treated if untreated)
+ (Outcome for treated if untreated - Outcome for untreated)
=Impact of treatment on treated + selection bias
This proves the observed outcome is not the treatment effect. I.e. treatment data
alone do not distinguish selection bias (e) from effect of treatment (b)
I Usually in marketing, treatment is nonrandom and depends on unobservables
I Usually in marketing, Impact of treatment on treated < selection bias
We have to measure or predict what would have happened to treated if they had been
untreated

4
5 ways to get causal effects / prescriptive analytics

Run an experiment / RCT : exogeneity by design


I Random assignment eliminates confounding factors. But, not always possible
Instrumental variables
I If you have variables z that (i) predict x and (ii) do not correlate with e, then
biv = cov (z,y )
cov (z,x) =
cov (z,x)+cov (z,e)
cov (z,x) b=b
I i is testable; ii is not
Diff-in-diff (including synthetic control): Estimate how treated and untreated units correlate
pre-treatment, then use untreated units to predict outcome for treated if untreated.
Depends on untestable “parallel trends” assumption
Regression discontinuity: measure local effects of x when treatment depends on a
known, discontinuous threshold. Esp. helpful when algorithms assign treatments
“Natural experiment”: use known, exogenous variation in treatment. This is rare
5
Let’s discuss
Hands-on Guide to Google Data
I Google Trends
I Google Consumer Surveys
Marketing Analytics by France & Ghose (2018)
Preview: Day3, Team Assignment, Day4

6
Marketing Analytics: 1,000’ View

Marketing data show how competitors, collaborators, customers and consumers learn,
feel, behave and use products and services
Marketing analytics is the use of data to improve marketing decisionmaking
Data is taking an ever more central role
Analytics have enabled new marketing techniques, e.g. geofencing, retargeting,
recommendations and seeded word-of-mouth, and improved many traditional marketing
functions

7
Types of Marketing Analytics

Descriptive analytics
I Summarize data, produce facts
Diagnostic analytics
I Test theories, identify drivers
Predictive analytics
I Forecasting: what’s going to happen tomorrow? Helps to plan
I Value depends on forecasting error and similarity(today,tomorrow)
Prescriptive analytics
I Algorithmic decisionmaking e.g. bandits
I “Marketing engineering,” “expert systems”
I Theory-driven structural models of behavior
What are the main differences between these?
8
Key applications of Marketing Analytics

Customer relationship management: Acquire, develop, retain and “fire” customers


Marketing mix: Improve product offerings, prices, promotions and distribution
I E-commerce
Understand customer heterogeneity for customization, personalization,
recommendations, product development, etc.
Privacy and security, e.g. misuse, theft, regulatory compliance

9
How do we “do” Marketing Analytics?

Collect relevant data


Manage data
Analyze data
Communicate analyses and recommendations
Make decisions
Implement data-driven decisions

10
History of Marketing Science

Emerged in the 1940s and 1950s


Branched off from management science and economics
I Quickly assimilated psychology
I Relies on tools from statistics, econometrics, computer science, etc.
Expanded rapidly in business schools, recently in policy schools
Differentiators
I merges experiments on humans with market data and business-relevant phenomena
I research is often driven by real-world phenomena (e.g., “How to...’)0
I beliefs about human behavior come more from observation than assumption
I inherently interdisciplinary
I lacks a unifying theory

11
Milestones in Marketing Data and Research Techniques

1910s: Opinion polls became popular


1920s: Questionnaire survey research
1920s: Copy research based on AIDA model
1920s: Retail product sales measurements (Nielsen)
1930s: Field experiments, telephone surveys and panel data
1930s: Radio audiences (Arbitron)
1940s: Panel data on consumer purchases
1950s: TV audiences (Nielsen)
1960s: Companies’ own customer data, RFM metrics
1960s: Inventory tracking data

12
Milestones in Marketing Data and Research Techniques

1970s: Census geodemographic data and credit agencies


1970s: UPC codes and digital registers enabled scanner panel data
1980s: Retail loyalty cards, TV set-top boxes
1990s: IRI in-home barcode scanning panel
1990s: CRM software, metrics and personalization
1990s: Email, Online display ads, WWW clickstream data
1990s: Search engine optimization and marketing
2000s: Social network data, Online marketplace data
2000s: User-generated content e.g. reviews, blogs, video
2010s: Smartphone-enabled data collection, especially geolocation
All of these are still in active use; more will be invented
13
Major Marketing Science Modeling Frameworks
Multidimensional scaling (Coombs 1950)

Decision calculus to optimize marketing spend (Dorfman and Steiner 1954, Little and Lodish 1969)

New product diffusion (Bass 1969)

Stochastic models of buyer behavior (Massy et al. 1970)

Market share mdoels for store level data (Nakanishi and Cooper 1974)

Choice models (McFadden 1974, Guadagni and Little 1983)

Conjoint analysis (Green and Srinivasan 1978)

Models of forward-looking behavior (Rust 1987, Erdem and Keane 1996)

Choice models with latent heterogeneity (Kamakura and Russell 1989, Rossi et al. 1996)

Marketing mix models (Dekimpe and Hanssens 1995)

Multi-touch attribution ?

Machine learning and fast, flexible models (e.g. Breiman 2001)


14
Main challenges to implementing marketing analytics

Executives
I Some prefer hunches
I “In God we trust; others must provide data”
I Some are territorial, or incentivized to be
I Some worry that analytics will constrain or replace them
I Some think data = magic
I Some misunderstand uncertainty
Analysts
I Expensive
I Hard to find
I Not always current
I Not always interested in business

15
Main challenges to implementing marketing analytics

Organizational/Cultural
I Are analytics used to make decisions or justify them?
I What is the tolerance for uncertainty?
I Is it a high- or low-trust environment? (Do messengers get shot?)
I Are data available and integrated?
I Do departments work together?
I Simple data descriptives are often displayed on “dashboards”
I Facilitates human anomaly detection and decisionmaking, but underuses data

16
Signs of a great analytics org

I C-level champion (e.g., CEO, CMO, CFO, CAO or CSO)


I Centralized team provides integrated data, architecture and tools; avoids silos
I Decentralized analysts collaborate with execs
I Efficient planning: only do things that contribute to the bottom line.
Avoid unproductive efforts
I Analytics career tracks are well established, including individual contributor roles
I Data are used to drive innovation, not just improve operations
I Careful thought about retaining/outsourcing analytic functions
I My favorite examples: Google, Netflix

17
How to evaluate a commercial data set
1. Utility
I What can you do with it? How much money will that make?
2. Frequency
I One-time snapshot, regular installments, or a real-time stream?
I Movies teach you more than pictures
3. Reliability and validity
I Is it what it purports to be? Can you verify it?
I Who else is using it? How long have they used it?
4. Size
I How much is there? Do you need a sample or a population? Can you handle it?
5. Cost
I Can you test-drive it? Is cost commensurate with utility?
I What usage restrictions, contract terms (e.g., exclusivity from competitors)?
I Is it bundled with tools or other data sets? Does it carry liability (e.g. privacy risk)?
I What else can you do with the money?
Pro tips: “Modeled data” are often worthless. Beware B2B salespeople 18
Popular languages and software for marketing analytics

FORTRAN, C++
R, Python
SQL
SAS
Matlab, Octave, Julia
S-Plus, STATA, SPSS
Hadoop, Dremel, MapReduce, ...
Focus on the languages that analysts in your niche use; preferably free and extensible
To generalize, learn multiple languages, e.g. SQL, R and Python
Ultimately, the language is not the thing

19
Marketing analytics generalities

Always start with univariate descriptions and bivariate visualizations to test your
understanding of the data
I Assume the data description is incorrect and incomplete; it usually is
Beware outliers, especially in internet data. Always try to understand why they occurred
Model-free evidence helps establish credibility before building models
Bayesian and Frequentist approaches usually offer similar insights
I despite what the cults believe
Simple models often perform similarly to complicated models in small and medium data
With big enough data, most flexible models will perform similarly

20
Marketing analytics generalities

Advertising agencies do not tell the truth about data


Marketing and analytics carry more weight in B2C than B2B
Selection effects are usually large; treatment effects are usually small
Demographics predict surprisingly little behavior
To speed estimation of complicated models, consider stochastic gradient descent or
variational bayes
Usually, you only get a few strikes
I Make sure nobody broke your reporting
If it’s written in LaTeX, it’s probably correct

21

You might also like

pFad - Phonifier reborn

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy