Chapter 8

Inferences About More Than

Two Population Central Values
 The procedures in this chapter generalize the test of the equality of means of
two independent populations. This generalization is often called the one-way
layout. While this design has somewhat limited value in practice, the material
in this chapter is fundamental for further generalizations. The key ideas that
are first developed in the one-way analysis of variance are: the generalization
of the t-test, the expected mean square calculation (which is described in
Chapter 14 and is crucial for power calculations), and the introduction to
multiple testing of hypotheses in Chapter 9.
The Model of Observations in a Completely
Randomized Design

The usual “effects” model is 𝑌𝑖𝑗 = 𝜇 + 𝛼𝑖 + 𝜎1𝑤 𝑍𝑖𝑗 , for 𝑖 = 1, … 𝐼 (where 𝐼 is the
number of treatment settings), 𝑗 = 1, … , 𝐽𝑖 , and σ𝐼𝑖=1 𝐽𝑖 𝛼𝑖 = 0. The use of 𝑍𝑖𝑗 in
this model is the assumption that the dependent variable data is normally
distributed and independent. The use of the multiplier 𝜎1𝑤 is the assumption that
the variances within groups are homogeneous. The important assumption is
independence of the error terms. This is guaranteed when there is a random
assignment of experimental unit to treatments. Sometimes researchers apply these
techniques to data not generated by a randomized experiment. In that event,
checking the assumption of independence is crucial. The {𝛼𝑖 } parameters are
called the treatment effects. Under the effects model, 𝐸 𝑌𝑖𝑗 = 𝜇 + 𝛼𝑖 , and the
2
distribution of 𝑌𝑖𝑗 is 𝑁𝐼𝐷(𝜇 + 𝛼𝑖 , 𝜎1𝑤 ).
OLS Estimates

A model that is equivalent to the “effects” model is called the “means” model
and is 𝑌𝑖𝑗 = 𝜇𝑖 + 𝜎1𝑤 𝑍𝑖𝑗 , where 𝜇𝑖 = 𝜇 + 𝛼𝑖 . The sum of squares function is
𝐽𝑖
then 𝑆𝑆 𝑚1 , … , 𝑚𝐼 = σ𝐼𝑖=1 σ𝑗=1 (𝑦𝑖𝑗 − 𝑚𝑖 )2 . We seek values of the arguments
that make the SS function as small as possible. As before, we take the partial
derivatives and solve the normal equations.
7
Sum of Squared Errors

As in Chapters 11 and 12, the minimized value of the SS function is the sum of
squared error and is crucial for our analysis. Now,
Fisher’s decomposition of the total sum of squares
Now shift the discussion from a realized experiment to a planned experiment.
That is, we will use the random variable notation.
Analysis of Variance Table
Analysis of Variance Table
Complete Randomized Experiment

Source DF Sum of Squares Mean Square F

𝐼
𝐼−1 𝑆𝑆𝑡𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡 𝑀𝑆𝑡𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡
Treatment ෍ 𝐽𝑖 (𝑌𝑖⦁ − 𝑌⦁⦁ )2
𝐼−1 𝑀𝑆𝐸
𝑖=1

𝐼 𝐽𝑖 𝐼
𝑆𝑆𝐸
Error 𝑛−𝐼 ෍ ෍(𝑌𝑖𝑗 − 𝑌𝑖⦁ )2 = ෍(𝐽𝑖 −1)𝑆𝑖2
(𝑛 − 𝐼)
𝑖=1 𝑗=1 𝑖=1

𝐼 𝐽𝑖
Total 𝑛−1 ෍ ෍(𝑌𝑖𝑗 − 𝑌⦁⦁ )2
𝑖=1 𝑗=1

As in Chapters 11 and 12, the statistical estimate of the variance parameter in the
2
model is the mean squared error. Then 𝜎ො1𝑤 = 𝑀𝑆𝐸.
Tests of Hypotheses

The most common issue is whether the expected value of the outcome variable is
the same for each setting of the treatment. The usual null hypothesis is then
𝐻0 : 𝜇1 = 𝜇2 = ⋯ = 𝜇𝐼 . The alternative hypothesis is 𝐻1 : 𝜇𝑖 ≠ 𝜇𝑖 ′ , 𝑖 ≠ 𝑖′; that is,
there is at least one pair of treatment settings with unequal means. An equivalent
statement using the effects model is that 𝐻0 : 𝛼1 = 𝛼2 = ⋯ = 𝛼𝐼 = 0. The
equivalent alternative hypothesis is 𝐻1 : 𝛼𝑖 ≠ 𝛼𝑖 ′ , 𝑖 ≠ 𝑖′ for at least one pair of
settings. The test statistic for this null hypothesis is 𝐹 = 𝑀𝑆𝑡𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡ൗ𝑀𝑆𝐸. The
distribution of the test statistic under the null hypothesis is a central F
distribution with 𝐼 − 1 numerator and 𝑛 − 𝐼 denominator degrees of freedom.
We call this test the “overall F-test” or “global F-test.”

When the null hypothesis is true, the test statistic should be one, modulo
statistical variability. When the alternative hypothesis is true, the test statistic
should be greater than one modulo statistical variability. That is, the test of the
null hypothesis is a right sided test.
When the global null hypothesis is rejected, researchers want to know which
settings of the treatment variable are associated with larger expected values and
which are associated with smaller expected values. Such questions lead to the
issues of multiple comparisons, which is covered more deeply in Chapter Nine. A
relatively simple approach is to use Fisher’s protected t confidence intervals
(which is also called Fisher’s Least Significant Difference). Fisher’s protected
confidence intervals are calculated only when the global null hypothesis is
rejected. Then one calculates a confidence interval for 𝐸 𝑌𝑖𝑗 − 𝐸 𝑌𝑖 ′ 𝑗 = 𝜇𝑖 −
𝜇𝑖 ′ for each pair of treatment settings using a procedure analogous to the
procedures in Chapter 6. These comparisons are called post hoc comparisons.
For example, the 99% confidence interval for 𝜇1 − 𝜇2 would be
1 1
𝑦1⦁ − 𝑦2⦁ ± 𝑡0.005,𝑛−𝐼 𝑀𝑆𝐸( + )
𝐽1 𝐽2

When the experiment is underpowered, it might well happen that the global null
hypothesis is rejected with no protected confidence interval excluding zero.
Balanced One-way Layouts

Typically, the questions that you will see on examinations used a balanced one-
way layout. That is, 𝐽1 = 𝐽2 = ⋯ = 𝐽𝐼 = 𝐽 . This simplifies the calculations and
permits natural and more complex issues discussed in Chapter Nine. The
interpretation of the global F-test is more clear for balanced one-way layouts.

That is, under the global null hypothesis, 𝐸0 𝑀𝑆𝑡𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡 = 𝐸(𝑀𝑆𝐸). The
global F-test is then the ratio of two quadratic forms with the same expectation
under the global null hypothesis. When the global null hypothesis does not
hold, 𝐸0 𝑀𝑆𝑡𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡 > 𝐸(𝑀𝑆𝐸). One should reject the global null
hypothesis when the global F-test is larger than one.
Example question:
A research team wishes to specify a manufacturing process so that Y, the area in a
product affected by surface flaws is as small as possible. They have four levels of
concentration of a chemical used to wash the product before the final
manufacturing step and want to determine whether the concentration level causes a
change in E(Y). They run a balanced one-way layout with 6 observations for each
concentration with level 1 set at 10%, level 2 set at 15%, level 3 set at 20%, and
level 4 set at 25%. They observe that 𝑦1⦁ = 264.5, 𝑦2⦁ = 255.9, 𝑦3⦁ = 216.2,
𝑦4⦁ = 263.8, where 𝑦𝑖⦁ is the average of the observations taken on the ith level.
They also observe that 𝑠12 = 411.9, 𝑠22 = 522.2, 𝑠32 = 631.8, 𝑠42 = 521.9, where
𝑠𝑖2 is the unbiased estimate of the variance for the observations taken on the ith
level.
a. Complete the analysis of variance table for these results; that is, be sure to
specify the degrees of freedom, sum of squares, mean square, and F-test.
b. What is your conclusion? Use significance levels set to 0.10, 0.05, and 0.01.
Make sure that you discuss the optimal setting of the concentration level and
how you could document it.
c. In Chapter 9, questions will be added asking for the decomposition into linear,
quadratic, and cubic components.
Checking on the AOV Conditions

The most important assumption is that of independence. This is guaranteed in

a randomized experiment in which the experimental units are randomly
assigned to treatment. When the data do not come from a randomized
experiment, this assumption should be checked carefully. Common problems
occur when time series data (for example, an exchange rate on successive days
as the dependent variable) is used. Also data describing a geographical area
such as a census tract have spatial autocorrelation. Data on students from the
same class will be correlated because of the common instruction.

The analysis procedures for balanced analyses of variances are not sensitive to
violations of the normality assumption and the homogeneity of variance
assumption. The residuals in a one-way AOV are 𝑟𝑖𝑗 = 𝑦𝑖𝑗 − 𝑦𝑖⦁ . Residual
analysis is simple for this model. One can and should generate a probability
plot of the residuals. Closeness of the plot to a straight line suggests that the
assumption of normality appears to be true.
 2
𝑆𝑚𝑎𝑥
Hartley’s 𝐹𝑚𝑎𝑥 = 2 is sensitive to normality. The Brown-Forsythe-Levene
𝑆𝑚𝑖𝑛
test for homogeneity of variance is more robust to violations of the assumption
of normality. Many statistical packages will calculate this test, and you should
use it routinely. There is another more robust test of this null hypothesis that
corrects for the estimated kurtosis of the sampled random variables. Some
statistical packages report this test as well or instead of Levene’s test. If the
hypothesis of constant variance is rejected, there are two common next steps.
One is to use weighted least squares (with weights reflecting the difference in
variance of observations), and the other approach is to transform the data to
lessen the differences in variance. These transformations are called variance
stabilizing transformations and are commonly used. They are helpful,
especially when predictions of future values are to be made.
An Alternative Analysis: Transformation of the Data

Delta Method

The objective is to calculate the approximate mean and variance of a

random variable 𝑊 = 𝑓(𝑌). The random variable Y has expected value
𝜇𝑌 and variance 𝜎𝑌2 , and the function f has finite derivatives. The delta
method approximates the value of W using the first term of the Taylor
series: 𝑊 ≅ 𝑓 𝜇𝑌 + 𝑓′(𝜇𝑌 )(𝑌 − 𝜇𝑌 ). Using this approximation, and
E(𝑊) ≅ 𝑓 𝜇𝑌 , and 𝑣𝑎𝑟 𝑊 ≅ 𝑓 ′ 𝜇𝑌 2 𝑣𝑎𝑟(𝑌).
Example question:
The random variable Y has the Poisson distribution with expected value 𝜇. Find the
approximate mean and variance of 𝑌.

Comment

Since 𝑣𝑎𝑟( 𝑌) ≅ 0.25 independently of the expected value of the Poisson

random variable, the square root transformation is said to be a variance stabilizing
transformation for a Poisson random variable. Another transformation of a
Poisson random variable is 𝑊 = 𝑌 + 𝑌 + 1 has approximate expected value
𝐸 𝑊 ≅ 𝜇 + 𝜇 + 1 and approximate variance 𝑣𝑎𝑟(𝑊) ≅ 1. This
transformation is an example of a Freeman-Tukey deviate.
Example question:
The random variable Y, 𝑌 > 0, has 𝐸 𝑌 = 𝜃 and 𝑣𝑎𝑟 𝑌 = 𝜃 3 , 𝜃 > 0. Find the
approximate mean and variance of 𝑊 = ln(𝑌).
Exploratory Data Analysis Tool to Identify Variance Stabilizing
Transformation

When the dependent variable in an analysis of variance is always positive,

𝑛𝑖 (𝑦 −𝑦 )2
σ𝑗=1 𝑖𝑗 𝑖⦁
calculate 𝑦𝑖⦁ and 𝑠𝑖2 = , where i indexes the treatments in the
𝑛𝑖 −1
analysis of variance. Plot log(𝑠𝑖 ) against log(𝑦𝑖⦁ ) and fit a straight line to the
data.

Call the slope m.

1−𝑚
When 𝑚 ≠ 0, then analyze the transformed values 𝑡𝑖𝑗 = 𝑦𝑖𝑗 , with 𝑚 ≠ 1.
When 𝑚 = 1, use 𝑡𝑖𝑗 = log(𝑦𝑖𝑗 ).
When 𝑚 ≅ 0, no transformation is necessary.

There is a related set of techniques called the Box-Cox transformations that is

also helpful. We will deal with this when we study multiple comparisons.
Example question:
The random variable Y, 𝑌 > 0, has 𝐸 𝑌 = 𝜃 and 𝑣𝑎𝑟 𝑌 = 𝜃 2.5 , 𝜃 > 0.
Find transformation W that makes the variance of W approximately constant.
What are the approximate mean and variance of W?