INE340 Advanced Statistics

Dr. Rola ZINTOUT

Part 3

Department of Industrial & Mechanical Engineering

1
Week Course outline Relevant Chapters
1 Introduction 1
2 Simple Comparative Experiments 2
3 Simple Comparative Experiments 2
4 The Analysis of Variance 3
5 The Analysis of Variance 3
6 Randomized Blocks, Latin Squares and Related Designs 4
7 Randomized Blocks, Latin Squares and Related Designs 4
8 Midterm
9 Introduction to Factorial Design 5
10 The 2𝑘 Factorial Design 6
11 Blocking and Confounding in the 2𝑘 Factorial Design 7
12 Blocking and Confounding in the 2𝑘 Factorial Design 7
13 Two-Level Fractional Factorial Designs 8
14 Fitting Regression Models 10
15 Exercises
2
Chapter 3:

Experiments with a Single Factor

The Analysis of Variance

3
Chapter 3 overview:

- One way ANOVA

- Least Significant Difference (LSD)
- Estimation of the Model Parameters
- Residuals
- Statistical Tests for Equality of Variance. (Bartlett’s test, Modified Levene test)
- Contrasts
- Scheffé’s Method for Comparing All Contrasts

4
5
Recall from chapter 1
Experimental Design
• a plan and a structure to test hypotheses in which
the researcher controls or manipulates one or more variables.

4
 Introduction to Design of Experiments (Recall from chapter 1)

Independent Variable
• Treatment variable - one that the experimenter controls or modifies in the
experiment.
• Classification variable - a characteristic of the experimental subjects that was
present prior to the experiment, and is not a result of the experimenter’s
manipulations or control.
• Levels or Classifications - the subcategories of the independent variable used
by the researcher in the experimental design.
• Independent variables are also referred to as factors.
• Dependent Variable is also called the response to the different levels of the
independent variables.
7
 Three Types of Experimental Designs

• Completely Randomized Design – subjects are assigned randomly to

treatments; single independent variable.
• Randomized Block Design – includes a blocking variable; single
independent variable.
• Factorial Experiments – two or more independent variables are
explored at the same time; every level of each factor are studied under
every level of all other factors.

8
1) Completely Randomized Design

• The completely randomized design contains only one independent variable with
two or more treatment levels
• If two treatment levels of the independent variable are present, the design is the
same used to test the difference in means of two independent populations which
used the t test to analyze the data.
• A technique has been developed that analyzes all the sample means at one time
and precludes the buildup of error rate: ANOVA
• A completely randomized design is analyzed by one way analysis of variance.

9
Introduction to Design of Experiments ANOVA

• Analysis of Variance (ANOVA) – a group of statistical techniques used to

analyze experimental designs.
• ANOVA begins with notion that individual items being studied are all the
same

10
What If There Are More Than Two Factor Levels?

• The t-test does not directly apply

• There are lots of practical situations where there are either more than two
levels of interest, or there are several factors of simultaneous interest

• The analysis of variance (ANOVA) is the appropriate analysis “engine” for

these types of experiments

• The ANOVA was developed by Fisher in the early 1920s, and initially
applied to agricultural experiments

• Used extensively today for industrial experiments

11
Example 1:
An engineer is interested in investigating the relationship between the RF power
setting and the etch rate for this tool. The objective of an experiment like this is to
model the relationship between etch rate and RF power, and to specify the power
setting that will give a desired target etch rate.

• The response variable is etch rate.

• She is interested in a particular gas (C2F6) and gap (0.80 cm), and wants to test four
levels of RF power: 160W, 180W, 200W, and 220W. She decided to test five wafers at
each level of RF power.

• The experimenter chooses 4 levels of RF power 160W, 180W, 200W, and 220W

• The experiment is replicated 5 times → 20 runs made in random order.

• This is an example of a single-factor experiment with a=4 levels of the factor and
n=5 replicates. 12
• Does changing the power change the mean etch rate?

• Is there an optimum level for power ?

13
Let’s examine experimental data graphically.
The first figure presents box plots for etch rate at each level of RF power
The second figure represents a scatter diagram of etch rate versus RF power.

14
▪ Both graphs indicate that etch rate increases as the power setting increases.

▪ There is no strong evidence to suggest that the variability in etch rate around the
average depends on the power setting.

▪ On the basis of this simple graphical analysis, we strongly suspect that (1) RF power
setting affects the etch rate and (2) higher power settings result in increased etch rate.

Test: differences between the mean etch rates at a = 4 levels of RF power

• t-test for all six possible pairs of means: inflates the type I error

• The appropriate procedure for testing the equality of several means is the analysis of
variance. It is probably the most useful technique in the field of statistical inference.

15
The appropriate procedure for testing the equality of several means is the analysis
of variance.

o There will be, in general, n observations under the ith treatment.

o a treatments of a single factor
o 𝑦𝑖𝑗 : the j th observation taken under taken under factor level or treatment i.

16
Models for the Data.

One way to write this model is using this formula called the means model:

where 𝑦𝑖𝑗 is the ijth observation, µ𝑖 is the mean of the ith factor level or treatment, and
ε𝑖𝑗 is a random error component.

17
• The name “analysis of variance” stems from a partitioning of the total
variability in the response variable into components that are consistent with a
model for the experiment

• Both means model and the effect models are linear statistical models

• one-way or single-factor analysis of variance model

• the effect model is more widely encountered in the experimental design

literature

• object:
- test hypotheses about the treatment means
- estimate model parameters: 𝜇, 𝜏𝑖 , 𝜎 2

18
 Analysis of Variance

• The null hypothesis states that the population means for all treatment levels
are equal
• Even if one of the population means is different from the other, the null
hypothesis is rejected
• Testing the hypothesis is done by portioning the total variance of data into the
following two variances
• Variance resulting from the treatment (columns)
• Error variance or that portion of the total variance unexplained by the
treatment
The Null hypothesis and the alternative hypothesis are:

Ho: 1 =  2 =  3 = =  k
Ha: At least one of the means is different from the others

The test value is : MSC

F=
MSE

If F > F , reject H .
c
o

The decision rule: If F  F , do not reject H .

c
o
20
The ANOVA table is as follows:

21
total sum of squares = error sum of squares + between sum of squares
SST = SSC + SSE

 (Xij− X ) =  n j X j − X
nj C

i =1 j= 1
2 C

j =1
( )
2 nj
+ 
C

i =1 j =1
(X ij − X j )
2

where : i = particular member of a treatment level

j = a treatment level
C = number of treatment levels
n j
= number of observations in a given treatment level
X = grand mean
X j
= mean of a treatment group or level

X ij
= individual value

22
One-Way ANOVA: Computational Formulas

ANOVA is used to determine statistically whether the variance between

the treatment level means is greater than the variances within levels
(error variance)

Assumptions underlie ANOVA

- Normally distributed populations.
- Observations represent random samples from the population.
- Variances of the population are equal.

23
 𝟐
SSC = σ𝒄𝒋=𝟏 𝒏𝒋 ഥ𝒋 − 𝑿
𝑿 ഥ with 𝒅𝒇𝑪 = C - 1

𝒏𝒋 𝟐
SSE = σ𝒊=𝟏 σ𝒄𝒋=𝟏 𝒏𝒋 ഥ𝒋
𝑿𝒊𝒋 − 𝑿 with 𝒅𝒇𝑬 = N – C

𝒏𝒋 𝟐
SST = σ𝒊=𝟏 σ𝒄𝒊=𝟏 𝒏𝒋 ഥ
𝑿𝒊𝒋 − 𝑿 with 𝒅𝒇𝑻 = N – 1

SSC
MSC =
𝒅𝒇𝑪

SS𝑬 Where:
MSE=
𝒅𝒇𝑬 i: a particular member of the treatment
j: a treatment level
MSC C: Number of treatment levels
F= MSE
𝑛𝑖 : number of observations in each treatment level
ത grand mean
𝑋:
𝑋ത𝑗 : Column mean
𝑋𝑖𝑗 : Individual value 24
Example 2:
In a company we have 3 divisions: Division1, division 2, and division 3. We want to
study if there is any difference in the mean age between employees between the 3
groups. Here are the given data:

25
Let’s construct the ANOVA table for the given data:

Source of Sum of Squares Degrees of Mean Squares F-ratio

Variance Freedom (df) (SS/df)
Treatment SST= 129.733 k-1= 3-1= 2 MST= 64.8665 64.8665/1.6333=39.715
Error SSE = 19.6 N-k = 15-3 = 12 MSE= 1.6333
Total SStotal= 149.333 N-1= 15-1 =14

We have: N=15, K=3 , 𝑋ത1 = 28.2 ; 𝑋ത2 = 32 ; 𝑋ത3 =24.8 and 𝑋തG = 28.333.

SSE= (29−28.2) 2 +(27−28.2) 2 +…+ (32−32) 2 +(33−32) 2 + ⋯ + (25−24.8) 2 + ⋯ +

(26−24.8) 2
= 6.8 + 10 + 2.8 = 19.6

SStotal= (29−28.333) 2 + ⋯ + (26−28.333) 2 =149.333

SSt = SStotal – SSE = 149.333 – 19.6 = 129.733. 26

F= 39.715
To find the critical value from an F distribution you must know the numerator
(MSTR) and denominator (MSE) degrees of freedom, along with the
significance level. F critical has df1 and df2 degrees of freedom, where:
- df1 is the numerator degrees of freedom equal to c-1, here, df1 = k-1= 3-1= 2

- df2 is the denominator degrees of freedom equal to N-c. Here df2 = N-k =
15-3 = 12

If α = 10% then:
So F2,12 = 2.81.

Decision Rule: We reject the null hypothesis if: F (observed value) > F critical
(critical value).
At least one mean age is different from others.

Since F= 39.715 > 2.81, we reject the null hypothesis

27
28
Example 3:
Suppose the National Transportation Safety Board (NTSB) wants to examine the
safety of compact cars, midsize cars, and full-size cars. It collects a sample of three
for each of the treatments (cars types). Using the hypothetical data provided below,
test whether the mean pressure applied to the driver’s head during a crash test is
equal for each types of car. Use α = 5%.

Compact cars Midsize cars Full-size cars

643 469 484

655 427 456
702 525 402

29
a- Find the mean and the standard deviation for each category of cars.

X1 = 666.67 X 2 = 473.67 X 3 = 447.33

S1 = 31.18 S2 = 49.17 S3 = 41.68

b- Find the general mean and the general standard deviation.

X G = 529.22 and SG = 109.7176

c- State the null and alternative hypotheses.

The null hypothesis for an ANOVA always assumes the population

means are equal.
H0 : µ1 = µ 2 = µ 3 - The mean head pressure is statistically equal
across the three types of cars.
The alternative hypothesis is: 30

Ha : At least one mean pressure is not statistically equal.

d- Using F table, find the critical value and deduce the condition for accepting
H0 .

To find the critical value from an F distribution you must know the numerator
(MSTR) and denominator (MSE) degrees of freedom, along with the significance
level. F critical has df1 and df2 degrees of freedom, where:
- df1 is the numerator degrees of freedom equal to c-1, here, df1 = 3 - 1 = 2
- df2 is the denominator degrees of freedom equal to N-c. Here df2 = 9 - 3 = 6.
So F2,6 = 5.14.

Decision Rule: We reject the null hypothesis if: F (observed value) > F critical
(critical value).

31
32
 The ANOVA table is:

Source of Sum of Squares Degrees of Mean Squares F-ratio

Variance Freedom (df) (SS/df)
Treatment SST= 86049.55 k-1= 3-1= 2 MST= 43024.775/1709 =25.175
43024.775
Error SSE = 10254 N-k = 9-3 = 6 MSE= 1709
Total SStotal= 96303.556 N-1= 9-1 =8

33
d- Calculate the appropriate test statistic by constructing the ANOVA table.

We have: N=9, K=3 , X1 = 666.67 ; X 2 = 473.67; X 3 = 447.33 and 𝑋തG = 529.22

SSE= (643−666.67) 2 +(655−666.67) 2 +(702−666.67) 2 + (469−473.67) 2 +

(427−473.67) 2 +(525−473.67) 2 +(484−447.33) 2 +(456−447.33) 2 +(402−447.33) 2
= 10254

SStotal=(643−529.22) 2 + ⋯ + (402−529.22) 2 = 96303.556

SSTR = [3x (666.67−529.22) 2 ] + [3x (473.67−529.22) 2 ] + [3x (447.33−529.22) 2 ]

= 86049.55

Or simply: SSt = SStotal – SSE = 96303.556 – 10254 = 86049.55

34
e- According to the ANOVA table, shall we reject the null hypothesis? Interpret
your result.

Since the F value 25.17 > 5.14 (Fcritical), so we reject the null hypothesis.

Interpretation: Since we rejected the null hypothesis, we are 95% confident (1-α )
that the mean head pressure is not statistically equal for compact, midsize, and full
size cars.

However, since only one mean must be different to reject the null, we do not yet
know which mean(s) is/are different.

In short, an ANOVA test will test us that at least one mean is different, but an
additional test must be conducted to determine which mean(s) is/are different.

35
f- Determine which mean(s) are different .

If you fail to reject the null hypothesis in an ANOVA then you are done. You know,
with some level of confidence, that the treatment means are statistically equal.

However, if you reject the null then you can conduct a separate test to determine which
mean(s) is/are different. There are several techniques for testing the differences
between means, but the most common test is the Least Significant Difference Test.

Least Significant Difference (LSD) for a balanced sample:

2 x MSE x F1,N−c
LSD =
n

where: MSE is the mean square error and r is the number of rows in each treatment.

2 x 1709 x 5.99
In the example above, LSD = = 82.61 36
3
Thus, if the absolute value of the difference between any two treatment means is greater
than 82.61, we may conclude that they are not statistically equal.

1) Compact cars vs. Midsize cars: |666.67 − 473.67|= 193. Since 193 > 82.61this means
head pressure is statistically different between compact and midsize cars.

2) Midsize cars vs. Full-size cars: |473.67 − 447.33| = 26.34. Since 26.34 < 82.61 this
means head pressure is statistically equal between midsize and full-size cars.

3) Compact vs. Full-size cars: | 666.67 − 447.33| = 219.34> 82.61this means head
pressure is statistically different between compact and Full-size cars.

Since 1&2 are different; 2&3 are equal and 1& 3 are different, we can conclude that 1
is different from 2 and 3.

Conclusion: Compact cars head pressure statistically different from Midsize cars and
Full-size cars. 37
Least Significant Difference (LSD) for a balanced sample:

2
LSD= 𝒕α,𝑵−𝒂 𝑀𝑆𝐸
𝟐 𝑛

Where

𝒕α,𝑵−𝒂 is the critical value from the t-distribution with N-a degrees of freedom ( same
𝟐
df as SSE)
MSE is the mean square error from the ANOVA table
n is the number of observations in each group (since the sample size is balanced).

38
Example 3 (revisited):
We have α = 5%, N-a = 9-3 = 6, MSE= 1709 and 𝒕α,𝑵−𝒂 = 2.447
𝟐

2
LSD = 𝒕α,𝑵−𝒂 𝑀𝑆𝐸 =
𝟐 𝑛
2
= 2.447 x 1709
3
= 82.596

39
1) Compact cars vs. Midsize cars: |666.67 − 473.67|= 193. Since 193 > 82.6 this means
head pressure is statistically different between compact and midsize cars.

2) Midsize cars vs. Full-size cars: |473.67 − 447.33| = 26.34. Since 26.34 < 82.6 this
means head pressure is statistically equal between midsize and full-size cars.

3) Compact vs. Full-size cars: | 666.67 − 447.33| = 219.34> 82.6 this means head
pressure is statistically different between compact and Full-size cars.

Since 1&2 are different; 2&3 are equal and 1& 3 are different, we can conclude that 1
is different from 2 and 3.

Conclusion: Compact cars head pressure statistically different from Midsize cars and
Full-size cars.
40
Example:
Suppose we conduct an experiment with three different fertilizers (A, B, C) on plant
growth and want to see if there's a significant difference between the mean plant heights
for each fertilizer. After running an ANOVA, we find that the overall test is significant,
meaning there is a difference somewhere between the groups. Now, we perform the LSD
test to determine which fertilizers differ significantly from each other. Use α = 5%.
Let’s assume:
MSE= 2.5 (from the ANOVA table)
The sample size for each group is n=10
Degrees of freedom for error = N-a = 30 – 3 = 27
And 𝒕α,𝑵−𝒂 = 2.052
𝟐
Using the following formula:
2 2
LSD= 𝒕α,𝑵−𝒂 𝑀𝑆𝐸 = 2.052 x 2.5 = 1.45
𝟐 𝑛 10

41
The least significant difference is 1.45 units. If the difference between any two fertilizer
means exceeds 1.45, we would conclude that the difference is statistically significant.

For example, if the mean heights for Fertilizer A, B, and C are 15, 14, and 12
respectively:

•The difference between A and B is 15−14=115 - 14 = 115−14=1, which is not

significant (less than 1.45).

•The difference between A and C is 15−12=315 - 12 = 315−12=3, which is significant

(greater than 1.45).

•The difference between B and C is 14−12=214 - 12 = 214−12=2, which is significant

(greater than 1.45).

42
Example 4:

Three different traffic routes are tested for mean driving time. The entries in the table
are the driving times in minutes on the three different routes.
a- Construct the one-way ANOVA table.
b- State the null and alternative hypothesis and deduce the optimal route using 5% level
of significance.

Route 1 Route 2 Route 3

30 27 16
32 29 41
27 28 22
35 36 31

43
44
45
Analysis of the Fixed Effects Model

Example 1: (revisited)
An engineer is interested in investigating the relationship between the RF power setting
and the etch rate for this tool. The objective of an experiment like this is to model the
relationship between etch rate and RF power, and to specify the power setting that will
give a desired target etch rate.

46
Note that:
▪ The RF power or between-treatment mean square (22290.18) is many times larger than
the within-treatment or error mean square (333.70). This indicates that it is unlikely that
the treatment means are equal.
▪ Suppose that the experimenter has selected α =0.05. So critical F0.05,3,16 = 3.24.
Because 66.80 > 3.24, we reject H0 and conclude that the treatment means differ; that is,
the RF power setting significantly affects the mean etch rate.
47
Graphical interpretation:

The P-value is very small:

F0.05,3,16 = 3.24 and F0.01,3,16 = 5.29 and because the test value F> 5.29,
we can conclude that an upper bound for the P-value is 0.01; that is, P <0.01
48
(the exact P-value is P =2.88 x10−9 ).
49
Estimation of the Model Parameters

The mean of the ith treatment is:

µ𝑖 = µ + 𝜏

Where µ is the population mean and 𝜏 is the estimation error.

We will also estimate

50
Estimation of the Model Parameters
if we assume that the errors ε𝑖𝑗 are normally distributed, each treatment average is
distribute NID(µ𝑖 , σ2 /n) .
A confidence interval estimate of the ith treatment mean may be easily determined
using the least squares estimation method.

51
Estimation of the Model Parameters

Where MSE: mean squared error

Df=N-k

n: number of variables in this particular group

52
Example 4: (revisited)
Route 1 Route 2 Route 3
Let’s find the CI estimate of the first treatment mean:
30 27 16
We have :𝑌ത1 = 31
MSE= 49; n=4, t=2.262; df= N-k=12-3 =9 32 29 41
27 28 22
A 95% confidence interval for the mean of route 1: 35 36 31
49 49
[31 – 2.262 ; 31 + 2.262 ] = [ 23.083 ; 38.917 ]
4 4

53
ν is the degree of
freedom
1- α: is CI

54
 Example 1: (revisited)
The general mean

This confidence interval expressions given is called one-at-a-

time confidence intervals. That is, the confidence level (1 – α)
applies to only one particular estimate.
55
Unbalanced Data

In some single-factor experiments, the number of observations taken within each

treatment may be different. We then say that the design is unbalanced.

The analysis of variance described above may still be used, but slight modifications
must be made in the sum of squares formulas.

56
Model Adequacy Checking

The use of the partitioning to test formally for no differences in treatment means
requires that certain assumptions be satisfied.
These assumptions are that the observations are adequately described by the model

𝑦𝑖𝑗 = µ + 𝜏𝑖 + ε𝑖𝑗

and that the errors are normally and independently distributed with mean zero and
constant but unknown variance σ2 .

If these assumptions are valid, the analysis of variance procedure is an exact test of
the hypothesis of no difference in treatment means. However, these assumptions
will usually not hold exactly.

57
It is usually unwise to rely on the analysis of variance until the validity of these
assumptions has been checked.
Violations of the basic assumptions and model adequacy can be easily investigated
by the examination of residuals.
We define the residual for observation j in treatment i as :
𝑒𝑖𝑗 = 𝑦𝑖𝑗 - 𝑦ො𝑖𝑗

Where 𝑦ො𝑖𝑗 is an estimate of the corresponding observation 𝑦𝑖𝑗 obtained as follows:

𝑦ො𝑖𝑗 = µො + 𝜏Ƹ = 𝑦ഥ.. + (𝑦𝑖. - 𝑦ഥ.. ) = 𝑦𝑖.

This means that the estimate of any observation in the ith treatment is just the
corresponding treatment mean.
Examination of the residuals should be an automatic part of any analysis of variance.

58
If the model is adequate, the residuals should be structureless; that is, they should
contain no obvious patterns.
Through analysis of residuals, many types of model inadequacies and violations of the
underlying assumptions can be discovered.
We show how model diagnostic checking can be done easily by graphical analysis of
residuals and how to deal with several commonly occurring abnormalities:

- The normal probability plot of the residuals is an extremely useful procedure

𝑒𝑖𝑗
The standardized residuals are: 𝑑𝑖𝑗 = with 𝑒𝑖𝑗 = 𝑦𝑖𝑗 - 𝑦ො𝑖𝑗
𝑀𝑆𝐸

59
The table shows the original data and the residuals for the etch rate data in Example1.
Using 𝑒𝑖𝑗 = 𝑦𝑖𝑗 - 𝑦ො𝑖𝑗 for example: 575 – 551.2 = 23.8 and 565-587.4 = -22.4
60
The normal probability plot .
i X 𝑓𝑖 100 x Z score
𝑓𝑖
1 -25.4 0.025 2.5 -1.96
2 -22.4 0.075 7.5 -0.93
3 -22 0.125 12.5 -1.15
4 -21.2
5 -15.4
6 -12.2
7 -9.2
8 -8.4
9 -7
10 2.6
11 3
12 3.6
13 5.6
14 8
15 11.6
16 18
17 18.8
18 22.6
19 23.8
20 25.6 0.975 97.5 1.96
61
The normal probability plot .

The normal probability plot is a graphical technique to identify substantive

departures from normality.
This includes identifying outliers, skewness, kurtosis.
Normal probability plots are made of raw data, residuals from model fits, and
estimated parameters.

Step 1:Arrange your x-values in ascending order.

(𝒊 −𝒂)
Step 2: Calculate 𝒇𝒊 = , where i is the position of the data value in the
(𝒏+𝟏 −𝟐𝒂)
ordered list and n is the number of observations.
Step 3: Find the z-score for each 𝒇𝒊
Step 4: Plot your x-values on the horizontal axis and the corresponding z-score on
the y-axis.
(𝒊 −𝟎.𝟓)
Note that: a= 3/8 if n ≤ 10, otherwise a= 0.5 and 𝒇𝒊 =
𝒏
62
 i X 𝑓𝑖 100 x 𝑓𝑖 Z score
1 -25.4 0.025 2.5 -1.96
2 -22.4 0.075 7.5 -0.93
Most normal probability plots present 3 -22 0.125 12.5 -1.15
4 -21.2
100(i - 0.5)/n on the left vertical scale. 5 -15.4
6 -12.2
An outlier is a residual that is very 7 -9.2
8 -8.4
much larger than any of the others. It is 9 -7
a very common defect. 10 2.6
11 3
12 3.6
13 5.6
14 8
15 11.6
16 18
17 18.8
18 22.6
19 23.8
20 25.6 0.975 97.5 1.96
63
Note that:
- the error distribution is approximately normal.
- The tendency of the normal probability plot to bend down slightly on the left side
and upward slightly on the right side implies that the tails of the error distribution
are somewhat thinner than would be anticipated in a normal distribution; that is, the
largest residuals are not quite as large (in absolute value) as expected.
- This plot is not grossly non-normal.
- In general, moderate departures from normality are of little concern in the fixed
effects analysis of variance.
- An error distribution that has considerably thicker or thinner tails than the normal is
of more concern than a skewed distribution. Because the F test is only slightly
affected, we say that the analysis of variance is robust to the normality assumption.

64
Checking for outliers may be made by examining the standardized residuals :
𝒆𝒊𝒋
𝒅𝒊𝒋 =
𝑴𝑺𝑬

Where 𝑒𝑖𝑗 = 𝑦𝑖𝑗 - 𝑦ො𝑖𝑗

If the errors ε𝑖𝑗 are N(0, σ2 ), the standardized residuals should be approximately
normal with mean zero and unit variance.

Thus, about 68 percent of the standardized residuals should fall within the limits ±1,
about 95 percent of them should fall within ±2, and virtually all of them should fall
within ±3.
A residual bigger than 3 or 4 standard deviations from zero is a potential outlier.
𝑒1 651−625.4 25.6 25.6
For example for the 14th value (651): 𝑑1 = = = = = 1.4
𝑀𝑆𝐸 333.27 333.27 18.27
65

It is the largest standardized residual and it should cause no concern.

Statistical Tests for Equality of Variance.

Bartlett’s test:

𝐻0 : σ1 2 = σ2 2 = … = σ𝐾 2
𝐻1 : Not true for at least σ𝑖 2

The procedure involves computing a statistic whose sampling distribution is closely

approximated by the chi-square distribution with a - 1 degrees of freedom when the
a random samples are from independent normal populations. (a is the number of
categories)

66
The test statistic is χ 𝟐 = 2.3026 q/c

Where 𝑙𝑜𝑔10 𝑒 = 2.3026 with 𝑒 =2.7182 (This means 𝑒 2.3026 = 10)

and 𝑆𝑖2 is the sample variance of the ith population.

67
The quantity q is :
- large when the sample variances 𝑆𝑖2 differ greatly
- is equal to zero when all are equal.

We reject 𝐻0 when:
χ02 > χα2 ,𝑎−1

Note that: Bartlett’s test is very sensitive to the normality assumption.

Consequently, when the validity of this assumption is doubtful, Bartlett’s test
should not be used.

68
Example 4: (revisited)

In the plasma etch experiment, the normality assumption is not in question, so we can
apply Bartlett’s test to the etch rate data.

69
Using: χ 2 = 2.3026 q/c and

70
71
Example 5: Bartlett’s Test
Suppose a professor wants to know if three different studying techniques lead to different
average exam scores.
She randomly assigns 10 students to use each technique for one week, then makes each student
take an exam of equal difficulty.
The exam scores of the 30 students are shown below.
Conduct Bartlett’s Test to verify that the three groups have equal variances.

72
Brief solution:

•Test statistic B: 3.30244 (with df=2)

•P-value: 0.19182
Since the p-value is not less than 0.05, the professor will fail to reject the null
hypothesis. In other words, she doesn’t have sufficient evidence to say that the three
groups have different variances.

73
Statistical Tests for Equality of Variance (Modified Levene test)
This test :
➢ Is robust to departures from normality.
➢ It uses the absolute deviation of the observations 𝑦𝑖𝑗 in each treatment from
the treatment median 𝑦ത𝑖 in order to test the hypothesis of equal variances in all
treatments.
➢ These deviations are:

Note that 𝑦෤𝑖 is the median of row i.

74
The modified Levene test then:

➢ It evaluates whether or not the means of these deviations are equal for all
treatments.

➢ It turns out that if the mean deviations are equal, the variances of the
observations in all treatments will be the same.

➢ The test statistic for Levene’s test is simply the usual ANOVA F statistic
for testing equality of means applied to the absolute deviations.

➢ The Levene test rejects the null hypothesis if 𝐹𝑠𝑡𝑎𝑡𝑖𝑠𝑡𝑖𝑐 > 𝐹α,𝐾−1 ,𝑁−𝑘

75
Example 6:
A civil engineer is interested in determining whether four different methods of
estimating flood flow frequency produce equivalent estimates of peak discharge
when applied to the same watershed.

Each procedure is used six times on the watershed, and the resulting discharge
data (in cubic feet per second) are shown in the next Table.

The analysis of variance for the data, implies that there is a difference in mean
peak discharge estimates given by the four procedures.

The plot of residuals versus fitted values, is disturbing because the outward-
opening funnel shape indicates that the constant variance assumption is not
satisfied.

76
We will apply the modified Levene test to the peak discharge data.

The upper panel of Table 3.7 contains the treatment medians and the lower panel contains
the deviations dij around the medians.

Levene’s test consists of conducting a standard analysis of variance on the dij.

The F test statistic that results from this is F0 = 4.55, for which the P-value is P = 0.0137.

Therefore, Levene’s test rejects the null hypothesis of equal variances, essentially
confirming the diagnosis we made from visual examination of Figure 3.7.

The peak discharge data are a good candidate for data transformation.

77
78
𝑒𝑖𝑗 = 𝑦𝑖𝑗 - 𝑦ො𝑖 𝑑𝑖𝑗 = |𝑦𝑖𝑗 - 𝑦෤𝑖 |

𝑒11 = 0.34 – 0.71= - 0.37 d11 = |0.34 – 0.52| = 0.18

𝑒12 = 0.12 – 0.71= - 0.59 d12= |0.12 – 0.52| = 0.4 …

𝑒21 = 0.91 – 2.63= - 1.72 d21= |0.91 – 2.61| = 1.7 ..

𝑒23 = 2.14 – 2.63= - 0.49

Note that 𝑦෤𝑖 is the median of row i.

79
 Here:
𝑦ො𝑖𝑗 =0.71; 2.63 ;
7.93; 14. 72

Plot of residuals 𝑒𝑖𝑗 versus 𝑦ො𝑖𝑗

80
Practical Interpretation of Results:

1- A Regression Model

The experimenter is interested in determining the differences, if any, between the levels of
the factors.
In fact, the analysis of variance treat the design factor as if it were qualitative or categorical.

A linear model of the data: 𝑦ො = β0 + β1 X + ε

The method often used to estimate the parameters in a model such as this is the method of
least squares.

81
In the previous example: 𝑦ො = 137.62 + 2.527X

We can do an improvement by using a quadratic model fit :

𝑦ො = 1147.77 + 8.2555 X + 0.028375 𝑋 2

82
The quadratic model appears to be superior to the linear model because it provides a
better fit at the higher power settings.

In general, we would like to fit the lowest order polynomial that adequately describes
the system or process.

In this example, the quadratic polynomial seems to fit better than the linear model, so
the extra complexity of the quadratic model is justified.

Selecting the order of the approximating polynomial is not always easy, however, and
it is relatively easy to overfit, that is, to add high-order polynomial terms that do not
really improve the fit but increase the complexity of the model and often damage its
usefulness as a predictor or interpolation equation

83
Contrasts

Many multiple comparison methods use the idea of a contrast.

Consider the plasma etching experiment.

Because the null hypothesis was rejected, we know that some power settings produce
different etch rates than others, but which ones actually cause this difference? We might
suspect at the outset of the experiment that 200 W and 220 W produce the same etch rate,
implying that we would like to test the hypothesis:

H0 : µ 3 = µ 4
Ha : µ 3 ≠ µ 4

It is equivalent to:
H0 : µ 3 - µ 4 = 0
Ha : µ 3 - µ 4 ≠ 0 84
If we had suspected at the start of the experiment that the average of the lowest levels of
power did not differ from the average of the highest levels of power, then the hypothesis
would have been:
H0 : µ1 + µ2 = µ3 + µ4
Ha : µ1 + µ2 ≠ µ3 + µ4
Which is equivalent to:
H0 : µ1 + µ2 − µ3 − µ4 = 0
Ha : µ1 + µ2 - µ3 − µ4 ≠ 0

A contrast is a linear combination of parameters of the form:

Г = σ𝑎𝑖=1 𝑐𝑖 µ1

where the contrast constants 𝑐𝑖 sum to zero

σ𝑎𝑖=1 𝑐𝑖 = 0
85
We can express the hypotheses in terms of contrasts:

H0 : σ𝑎𝑖=1 𝑐𝑖 µi = 0
Ha : σ𝑎𝑖=1 𝑐𝑖 µi ≠ 0

Because H0 : µ3 = µ4
Ha : µ 3 ≠ µ 4

The contrast constants for the hypotheses in our previous example are:
c1 = c2 = 0,
c3 = 1 and c4 = -1,

Or they are
c1 = c2 = 1,
c3 = c4 = -1,
86
Two basic ways for testing hypotheses involving contrasts:

The first approach uses a t-test.

- Write the contrast of interest in terms of the treatment averages:

C= σ𝒂𝒊=𝟏 𝒄𝒊 𝒚
ഥ𝒊
σ2
With variance V( C ) = σ𝑎𝑖=1 𝑐𝑖 2
𝑛

𝐶 σ𝒂 ഥ𝒊
𝒊=𝟏 𝒄𝒊 𝒚
The test statistic is: = and it follows a normal distribution N(0,1) if
𝑉 (𝐶) σ2
σ𝑎
𝑖=1 𝑐𝑖
2
𝑛
H0 is true.

σ𝒂 ഥ𝒊
𝒊=𝟏 𝒄𝒊 𝒚
The following test statistic will be used: 𝑡0 =
𝑀𝑆𝐸
σ𝑎
𝑖=1 𝑐𝑖
2
𝑛
The null hypothesis would be rejected if |𝑡0 | > 𝑡α;𝑁−𝑎 . 87

2
The second approach uses an F test.

Now the square of a t random variable with ν degrees of freedom is an F random variable
with 1 numerator and ν denominator degrees of freedom.

𝟐 (σ𝒂 ഥ 𝒊 )𝟐
𝐢=𝟏 𝒄𝒊 𝒚
𝑭𝟎 = 𝒕𝟎 = 𝑴𝑺𝑬
σ𝒂
𝐢=𝟏 𝒄𝒊
𝟐
𝒏

The null hypothesis would be rejected if 𝐹0 > 𝐹α ;1;𝑁−𝑎

𝑀𝑆𝐶 𝑆𝑆𝐶/1
We can write the test statistic 𝐹0 = =
𝑀𝑆𝐸 𝑀𝑆𝐸

where the single degree of freedom contrast sum of squares is:

(σ𝒂 ഥ 𝒊 )𝟐
𝐢=𝟏 𝒄𝒊 𝒚
SSc = 𝟏
88

σ𝒂
𝐢=𝟏 𝒄𝒊
𝟐
𝒏
The confidence interval of the contrast C= σ𝑎𝑖=1 𝑐𝑖 µ𝑖 is:

The mean and the variance are:

89
When more than one contrast is of interest, it is often useful to evaluate them on
the same scale. One way to do this is to standardize the contrast so that it has
variance σ2 .

A useful special case is that of orthogonal contrasts. Two contrasts with

coefficients {ci } and {di } are orthogonal if

σ𝑎𝑖=1 𝑐𝑖 𝑑𝑖 =0

or, for an unbalanced design, if

σ𝑎𝑖=1 𝑛𝑖 𝑐𝑖 𝑑𝑖 = 0

90
➢ There are many ways to choose the orthogonal contrast coefficients for a set of
treatments.
➢ Usually, something in the nature of the experiment should suggest which
comparisons will be of interest.
➢ For example, if there are a = 3 treatments, with treatment 1 a control and treatments
2 and 3 actual levels of the factor of interest to the experimenter, appropriate
orthogonal contrasts might be as follows:

𝒄𝒊 𝒅𝒊

Note that contrast 1 with 𝑐𝑖 = -2, 1, 1 compares the average effect of the factor with
the control, whereas contrast 2 with 𝑑𝑖 = 0, -1, 1 compares the two levels of the 91

factor of interest.
Example 1: (revisited)

Consider the plasma etching experiment. There are four treatment means and three
degrees of freedom between these treatments.
Suppose that prior to running the experiment the following set of comparisons among
the treatment means (and their associated contrasts) were specified:

Note that the contrast coefficients are orthogonal.

92
 (σ𝒂 ഥ 𝒊 )𝟐
𝐢=𝟏 𝒄𝒊 𝒚
Using the data from this Table and the following formulas: SSc = 𝟏 ,we find
σ𝒂
𝐢=𝟏 𝒄𝒊
𝟐
𝒏
from the numerical values of the contrasts and the sums of squares to be as follows:
k
Since
C =  ci Yi.
i =1
(−36.2)2
𝐶1 = 𝑦ത1 - 𝑦ത2 = 1(551.2) - 1(587.4) = - 36.2 and 𝑆𝑆𝑐1 = 1 = 3276.10
(2)
5
𝐶2 = 𝑦ത1 + 𝑦ത2 - 𝑦ത3 - 𝑦ത4 = 1(551.2) + 1(587.4) -1(625.4) – 1(707) = - 193.8
93
(−193.8)2
and 𝑆𝑆𝑐2 = 1 = 46948.05
5
(4)
 𝑐3 = 𝑦ത1 - 𝑦ത2 = 1(625.4) - 1(707.6) = - 81.6

(−81.6)2
and 𝑆𝑆𝑐3 = 1 = 16646.40
(2)
5

These contrast sums of squares completely partition the treatment sum of squares.

Where SS(RF power)= σ 𝑆𝑆𝑐𝑖 = 3276.10 + 46948.05 + 16646.40 = 66870.55 94

2
(σ 𝑦𝑖𝑗 )
SSTotal = σ(𝑦𝑖𝑗 )2 - = 7704511 - (12355)2 /20 = 72209.75
20
The tests on such orthogonal contrasts are usually incorporated in the ANOVA, as shown
in the following table.

𝑀𝑆𝐶 𝑆𝑆𝐶/1
Since 𝐹0 = = we have:
𝑀𝑆𝐸 𝑀𝑆𝐸

22290.18/333.70 = 66.80 95
3276.10/333.70 = 9.82
We conclude from the P-values that there are significant differences in mean etch rates
between levels 1 and 2 and between levels 3 and 4 of the power settings, and that the
average of levels 1 and 2 does differ significantly from the average of levels 3 and 4 at
the 0.05 level.

F critical is 𝐹α ;1;𝑁−𝑎

96
 More about contrasts

➢ For a One-way ANOVA, a contrast is a specific comparison of Treatment group means.

➢ Contrast constants are composed to test a specific hypothesis related to treatment means
based upon some prior information about the Treatment groups.
➢ For k treatment groups, contrast constants are a sequence of numbers c1 , c2, ......, ck
such that
k

c
i =1
i =0

97
Contrasts and Hypothesis testing
A given contrast will test a specific set of hypotheses:
k
H 0 :  ci i = 0
i =1

and k
H a :  ci i  0
i =1

k
using C =  ci Yi.
i =1

to create an F-statistic with one numerator df.

98
Example: Control and two equivalent treatments
Suppose we have two treatments which are supposed to be equivalent.

For example, each of two drugs is supposed to work by binding to the receptor for
adrenalin. Propanolol is such a drug sometimes used for hypertension or anxiety.

We may think that:

• the two drugs are equivalent, and
• they are different from Control

The Layout of the experiment:

99
The two contrasts:

Control Drug A Drug B

Contrast 1 -1 ½ ½
Contrast 2 0 -1 +1

Contrast 1 tests whether or not the Control group differs from the groups
which block the adrenalin receptors.

Contrast 2 tests whether or not the two drugs differ in their effect.

100
Orthogonal Contrasts

• The contrasts in the last example were orthogonal.

• Two contrasts are orthogonal if the pairwise products of the terms sum to zero.
• The formal definition is that two contrasts
' ' '
c1, c2 ,..., ck and c1, c2 ,..., ck

Contrasts are orthogonal if:

k

c c
i =1
'
i i =0

101
Orthogonal Contrasts allow the Trt. Sums of Squares to be decomposed

The Treatment Sums of Squares can be written as a sum of two Statistically

independent terms:
SSTrt = SSC1 + SSC2

Which can be used to test the hypotheses in the example. The a priori structure in
the Treatments can be tested for significance in a more powerful way.

102
Why?

If all of the differences in the means are described by one of the contrasts, say the first
contrast, then

F = SSC1 MSE

is more likely to be significant than

F = SSTrt MSE

Since the signal in the numerator is not combined with “noise”.

103
Example 7:
Suppose we are testing three treatments, T1, T2 and T3 (control) with treatment
means m1, m2, and m3 (two d.f.).

The null hypothesis for the ANOVA is

H0 : µ1 = µ2 = µ3

Since there are two degrees of freedom for treatments, there are in principle two
independent comparisons that can be made.

104
For example, one could in principle test the two hypotheses that µ1 and µ2 are not
significantly different from the control: µ1 = µ3 and µ2 = µ3 .

We represent the means µi with their estimates 𝑌ത𝑖 .

1) 1 = 3 can be rewritten as 11 + 02 -13= 0

the coefficients of this contrast are: c1 = 1, c2 = 0, c3 = -1

2) 2 = 3 can be rewritten as 01 + 12 -13= 0

the coefficients of this contrast are: d1 = 0, d2 = 1, d3= -1
k
These linear combinations of means are contrast because c
i =1
i =0

However these contrasts are not orthogonal because we should have σ𝑎𝑖=1 𝑐𝑖 𝑑𝑖 =0

But here σ𝑎𝑖=1 𝑐𝑖 𝑑𝑖 = 1 105

So, not every pair of hypotheses can be tested using this approach.

In addition to summing to 0, the 𝑐𝑖 coefficients are almost always taken to be

integers, a constraint which severely restricts their possible values.

For t = 3 such a set of values are

1) c1 = 1, c2 = 1, c3 = -2;
2) d1 = 1, d2 = -1, d3 = 0.

These are contrasts and are orthogonal. The hypotheses they define are:

1) The average of the two treatments is equal to the control (i.e. is there a
significant average treatment effect?); and
2) The two treatments are equal to one another (i.e. is one treatment significantly
different from the other?). 106
There are two general kinds of linear combinations:
▪ Class comparisons
▪ Trend comparisons.

Class comparisons (or group)

The first category of hypothesis we can pose using orthogonal contrasts is class (or
group) comparisons.
Such contrasts compare specific treatment means or combinations of treatment
means, grouped is some meaningful way.

107
Extra example :
To illustrate orthogonal contrasts in class comparisons we will use the following
data represented in the table showing results (mg shoot dry weight) of an
experiment (CRD) to determine the effect of seed treatment by  acids on the
early growth of rice seedlings.

108
The analysis of variance for this experiment is given the following table:

The treatment structure of this experiment suggests that the investigator had
several specific questions in mind from the beginning:

1) Do acid treatments affect seedling growth?

2) Is the effect of same organic acids different from that of inorganic acids?
3) Is there a difference in the effects of the two different organic acids?

109
In the following table, coefficients are shown that translate these three questions
into contrasts.
The table shows orthogonal coefficients for partitioning the SST into 3
independent tests.

Control HC1 Propionic Butyric

Control vs. acid +3 -1 -1 -1
Inorganic vs. 0 -2 +1 +1
organic
Between organics 0 0 +1 -1
Totals 20.95 19.34 18.64 18.2
Comparisons Means 4.19 3.87 3.73 3.64

110
The 1st contrast (first row) compares the control group to the average of the three
acid-treated groups, as can be seen from the following manipulations:
• 3μCont - 1μHCl - 1μProp - 1μBut = 0
• μCont = (1/3)*(1μHCl + 1μProp + 1μBut)
• Mean of the control group = Mean of all acid-treated groups

The H0 for this 1st contrast is that there is no average effect of acid treatment on
seedling growth. Since this Ho involves only two group means, it costs 1 df.

111
The 2nd contrast (second row) compares the inorganic acid group to the average of the
two organic acid groups:

0μCont + 2μHCl - 1μProp - 1μBut = 0

μHCl = (1/2)*(1μProp + 1μBut)
Mean of the HCl group = Mean of all organic acid-treated groups

The H0 for this second contrast is that the effect of the inorganic acid treatment on
seedling growth is no different from the average effect of organic acid treatment. Since
this null hypothesis involves only two group means (means than before), it also costs 1
df.

112
Finally, the third contrast (third row of coefficients) compares the two organic acid
groups to each other:

• 0μCont + 0μHCl + 1μProp - 1μBut = 0

• 1μProp = 1μBut

The H0 for this third contrast is that the effect of the propionic acid treatment on
seedling growth is no different from the effect of butyric acid treatment.
Since this null hypothesis involves only two group means (different means than
before), it also costs 1 df.
At this point, we have spent all our available degrees of freedom (dftrt = t – 1 = 4 –
1 = 3).

113
Because each of these questions are contrasts (each row of coefficients sums to
zero) and because the set of three questions is orthogonal, these three question
perfectly partition SST into three components, each with 1 df.

The SS associated with each of these contrasts serve as the numerators for three
separate F tests, one for each comparison. The critical F values for these single df
tests are based on 1 df in the numerator and dfError in the denominator.
All of this can be seen in the expanded ANOVA table below representing orthogonal
partitioning of SST via contrasts.
Source df SS MS F
Total 19 1.0113
Treatment 3 0.8738 0.2912 33.87
1. Control vs. acid 1 0.7415 0.7415 86.22
2. Inorg. vs. Org. 1 0.1129 0.1129 13.13
3. Between Org. 1 0.0194 0.0194 2.26
Error 16 0.1376 0.0086 114
 Control HC1 Propionic Butyric
Control vs. acid +3 -1 -1 -1
Inorganic vs. 0 -2 +1 +1
organic
Between organics 0 0 +1 -1
Totals 20.95 19.34 18.64 18.2
Comparisons Means 4.19 3.87 3.73 3.64
(σ𝒂 ഥ𝒊 )𝟐
𝐢=𝟏 𝒄𝒊 𝒚
Using the following formula: SSc = 𝟏 we get
σ𝒂
𝐢=𝟏 𝒄𝒊
𝟐
𝒏
SS1 (control vs. acid) = [3(4.19) – 3.87 – 3.73 – 3.64]2 / [(12)/5] = 0.74
SS2 (inorg. vs. org.) = [– 2(3.87) + 3.64 + 3.73]2 / [(6)/5] = 0.11
SS3 (between org.) = [ 3.73 -3.64]2 / [(2)/5] = 0.02

We have:
SStreatment= SS1 +SS2 +SS3 =0.87125
SStotal= 298.46 - (77.13)2/20 = 1.008155
Sserror = 0.138155 115
Conclusion:
From this analysis, we conclude that in this experiment acids significantly reduce
seedling growth (F = 86.22, p < 0.01), that the organic acids cause significantly
more reduction than the inorganic acid (F = 13.13, p < 0.01), and that the
difference between the organic acids is not significant (F = 2.26, p > 0.05).

116
Scheffé’s Method for Comparing All Contrasts
In many situations, experimenters may not know in advance which contrasts they
wish to compare, or they may be interested in more than a -1 possible comparisons.

In many exploratory experiments, the comparisons of interest are discovered only

after preliminary examination of the data.

Scheffé (1953) has proposed a method for comparing any and all possible contrasts
between treatment means.

In the Scheffé method, the type I error is at most for any of the possible
comparisons.

117
118
Example 1: (revisited)
Suppose that the contrasts of interests are:
Г1 = µ1 + µ2 - µ3 - µ4 and Г2 = µ1 - µ4

The numerical values of these contrasts are:

𝐶1 = 𝑦ത1. + 𝑦ത2. - 𝑦ത3. - 𝑦ത4. = 551.2 + 587.4 - 625.4 - 707.0 = -193.8 (this is the tested value for Г1 )

and 𝐶2 = 𝑦ത1. - 𝑦ത4. = 551.2 - 707.0 = - 155.8 (this is the tested value for Г2 )

The standard error of the contrast are:

𝑆𝐶𝑢 = 𝑀𝑆𝐸 σ𝑎𝑖=1(𝐶𝑖𝑢 2 /𝑛𝑖 ) where 𝑛𝑖 is the number of observations in the ith treatment.

119
The 1 percent critical values are

120
121
Conclusion:

Because |C1|> 𝑆0.01,1 (193.8 > 65.09) we conclude that the contrast Г1 = µ1 + µ2 -
µ3 - µ4 does not equal zero; that is, we conclude that the mean etch rates of power
settings 1 and 2 as a group differ from the means of power settings 3 and 4 as a
group.

Because |C2|> 𝑆0.01,2 (155.8 > 45.97) we conclude that the contrast Г2 = µ1 - µ4
does not equal zero; that is, the mean etch rates of treatments 1 and 4 differ
significantly.

122
Exercise 1: (problem 3.25)
Four chemists are asked to determine the percentage of methyl alcohol in a certain
chemical compound. Each chemist makes three determinations, and the results are
the following:

Chemist Percentage of Methyl Alcohol

1 84.99 84.04 84.38
2 85.15 85.13 84.88
3 84.72 84.48 85.16
4 84.20 84.10 84.55

(a) Do chemists differ significantly? Use α = 0.1.

(b) Analyze the residuals from this experiment.
(c) If chemist 2 is a new employee, construct a meaningful set of orthogonal
contrasts that might have been useful at the start of the experiment.
123
a)

124
 125

α = 0.1 this means F=2.92

b) The residuals :

126
c)

127
Exercise 2: (problem 3.26)
Three brands of batteries are under study. It is suspected that the lives (in weeks) of the
three brands are different. Five batteries of each brand are tested with the following
results:
Week of life
Brand 1 Brand 2 Brand 3
100 76 108
96 80 100
92 75 96
96 84 98
92 82 100

128
(a) Are the lives of these brands of batteries different?

The Model F-value of 38.34 > 6.93 implies the model is significant. There is only
a 0.01% chance that a "Model F-Value" this large could occur due to noise. Yes, at
least one of the brands is different.

129
(b) Analyze the residuals from this experiment.

Using the following formula 𝑒𝑖𝑗 = 𝑦𝑖𝑗 - 𝑦ො𝑖𝑗 and data from this table

Residuals are:

Week of life
Brand 1 Brand 2 Brand 3
100- 95.2 = 4.8 76- 79.4 = -3.40 7.60
96-95.2 = 0.8 0.60 -.40
92-95.2 = -3.2 -4.40 -4.40 130
96-95.2 = 0.8 4.60 -2.40
92-95.2 =-3.2 2.60 -0.40
131
132
(c) Construct a 95% interval estimate on the mean life of battery brand 2.
Construct also a 99% interval estimate on the mean difference between the lives
of battery brands 2 and 3.

133
(d) Which brand would you select for use? If the manufacturer will replace without
charge any battery that fails in less than 85 weeks, what percentage would the
company expect to replace?

Chose brand 3 for longest life. Mean life of this brand in 100.4 weeks, and the
variance of life is estimated by 15.60 (MSE).

Assuming normality, then the probability of failure before 85 weeks is:

85 −100.4
Φ( ) = Φ (-3.9) = 0.00005
15.6

That is, about 5 out of 100,000 batteries will fail before 85 week.

134
(e) Use the modified Levene test to determine if the assumption of equal variances is
satisfied. Use α = 0.05. Did you reach the same conclusion regarding the equality of
variances by examining the residual plots?

Solution:
The absolute value of Battery Life – brand median is (we find deviations): 𝑑𝑖𝑗 = |𝑦𝑖𝑗 - 𝑦෤𝑖 |

Brand 1 Brand 2 Brand 3

4 4 8
0 0 0
4 5 4
0 4 2
4 2 0

135
The analysis of variance indicates that there is not a difference between the different
brands and therefore the assumption of equal variances is satisfied.

136
Comparing Pairs of Treatment Means
In many practical situations, we will wish to compare only pairs of means.

We are interested in contrasts of the form Г = µ𝑖 - µ𝑗 for all i ≠ j.

The Scheffé method could be easily applied to this problem, it is not the most
sensitive procedure for such comparisons.

Suppose that we are interested in comparing all pairs of a treatment means and that
the null hypotheses that we wish to test are H0 : µ𝑖 - µ𝑗 for all i ≠ j.

Two popular methods for making such comparisons:

1) Tukey’s Test
2) The Fisher Least Significant Difference (LSD) Method.
137
Tukey’s Test.
Suppose that, following an ANOVA in which we have rejected the null hypothesis of
equal treatment means, we wish to test all pairwise mean comparisons:
H0 : µi = µj
H1 : µi ≠ µj
for all i ≠ j.

➢ The Tukey procedure controls the experimentwise or “family” error rate at the
selected level .
➢ Tukey’s procedure makes use of the distribution of the studentized range statistic.

ഥ𝒎𝒂𝒙 −ഥ
𝒚 𝒚𝒎𝒊𝒏
q=
𝑴𝑺𝑬/𝒏
where 𝑦ത𝑚𝑎𝑥 𝑎𝑛𝑑 𝑦ത𝑚𝑖𝑛 are the largest and smallest sample means.
138
For equal sample sizes, Tukey’s test declares two means significantly different if the
absolute value of their sample differences exceeds:

𝑻𝜶 = 𝒒𝜶 (a,f) 𝑴𝑺𝑬/𝒏

Values of q ( a, f ), the upper percentage points of q, where f is the number of degrees of

freedom associated with the MSE (f=N-a), are in the table on the next 2 slides.

When sample sizes are not equal, we use:

(a,f) 𝟏
𝑻𝜶 =
𝒒𝜶
𝟐
𝑴𝑺𝑬 (
𝒏𝟏
+ 𝒏𝟏 )
𝟐

139
140
a

141
We can construct a set of 100(1 - 𝛼) percent confidence intervals for all pairs of
means as follows:

When sample sizes are not equal, we use:

142
Note that: The unequal sample size version is sometimes called the Tukey–Kramer
procedure.

143
Example 1 (revisited):

144
The Fisher Least Significant Difference (LSD) Method.
The Fisher method for comparing all pairs of means controls the error rate for each
individual pairwise comparison but does not control the experimentwise or family error
rate.

This procedure uses the t statistic for testing: H0 : µi = µj

ഥ𝒊. −ഥ
𝒚 𝒚𝒋.
𝒕𝟎 = 𝟏 𝟏
𝑴𝑺𝑬 (𝒏 +𝒏 )
𝟏 𝟐

1 1
The quantity called the least significant difference is: LSD= 𝒕α,𝑵−𝒂 𝑀𝑆𝐸 ( + )
𝟐 𝑛1 𝑛2

145
Example 1 (revisited):

146
Problems 3.7 – 3.8 – 3.9 - 3.22 - 3.25 – 3.26 – 3.49 from the book

147
Comparing Treatment Means with a Control (Dunnett )

In many experiments, one of the treatments is a control, and the analyst is interested in
comparing each of the other a -1 treatment means with the control.

Thus, only a -1 comparisons are to be made.

Suppose that treatment a is the control and we wish to test the hypotheses:
H0 : µ i = µ a
H1 : µi ≠ µa
for all a=1,2,…, a-1.

148
Dunnett’s procedure is a modification of the usual t-test.

For each hypothesis, we compute the observed differences in the sample means | 𝒚ഥ𝒊. − 𝒚ഥ𝒂.|.

The null hypothesis is rejected using a type I error α rate if:

𝟏 𝟏
ഥ𝒊. − 𝒚
|𝒚 ഥ𝒂. | > 𝒅𝜶 (a-1 , f) 𝑴𝑺𝑬 ( + )
𝒏𝒊 𝒏𝒋
where the constant 𝑑𝛼 (a-1 , f) is given in the next table.

149
Example 1: (revisited)
To illustrate Dunnett’s test, consider the experiment with etching rate with treatment 4
considered as the control. In this example, a = 4, f = n-a =16, and n = 5. At the 5 percent
level, we find from Appendix Table VIII that
𝒅𝜶 (a-1 , f) = 𝒅𝟎.𝟎𝟓 (3 , 16) =2.59.
𝟏 𝟏 𝟏 𝟏
The critical difference becomes 𝒅𝜶 (a-1 , f) 𝑴𝑺𝑬 (𝒏 + 𝒏 ) = 2.59 𝟑𝟑𝟑. 𝟕 (𝟓 + 𝟓 ) =29.92
𝒊 𝒂

150
Any treatment mean that differs in absolute value from the control by more than 29.92 would be
declared significantly different.
The observed differences are:
1 vs. 4: 𝑦ത1 - 𝑦ത4 = 551.2 - 707.0 = -155.8
2 vs. 4: 𝑦ത2 - 𝑦ത4 = 587.4 - 707.0 = - 119.6
3 vs. 4: 𝑦ത3 - 𝑦ത4 = 625.4 - 707.0 = -81.6

Note that all differences are significant. Thus, we would conclude that all power settings are
different from the control.

151
152
153
Example 1 (revisited):

154
Tables

155
 The cumulative standard normal
distribution table.

156
ν is the degree of
freedom
1- α: is CI

157
158
159
160
161
162
 Formulas

𝟐
Treatment: SSC = σ𝒄𝒋=𝟏 ഥ𝒋 − 𝑿
𝑿 ഥ with 𝒅𝒇𝑪 = C - 1

𝒏𝒋 𝟐
Error: SSE = σ𝒊=𝟏 σ𝒄𝒋=𝟏 ഥ𝒋
𝑿𝒊𝒋 − 𝑿 with 𝒅𝒇𝑬 = N – C

𝒏𝒋 𝟐
Total: SST = σ𝒊=𝟏 σ𝒄𝒊=𝟏 ഥ
𝑿𝒊𝒋 − 𝑿 with 𝒅𝒇𝑻 = N – 1

SSC
MSC =
𝒅𝒇𝑪

SS𝑬
MSE=
𝒅𝒇𝑬

MSC
F= MSE
163
Formulas

𝟐 𝐱 𝐌𝐒𝐄 𝐱 𝑭𝟏,𝐍−𝒄
LSD =
𝒓

χ 𝟐 = 2.3026
q/c

𝑑𝑖𝑗 = |𝑦𝑖𝑗 - 𝑦෤𝑖 |

164
Contrasts
C= σ𝑎𝑖=1 𝑐𝑖 𝑦ത𝑖
𝜎2
V( C ) = σ𝑎𝑖=1 𝑐𝑖 2
𝑛
σ𝒂 ഥ𝒊
𝒊=𝟏 𝒄𝒊 𝒚
𝑡0 =
𝑀𝑆𝐸
σ𝑎𝑖=1 𝑐𝑖
2
𝑛
(σ𝒂 ഥ 𝒊 )𝟐
𝐢=𝟏 𝒄𝒊 𝒚
SSc = 𝟏
σ𝒂
𝐢=𝟏 𝒄 𝒊
𝟐
𝒏

2
(σ 𝑦𝑖𝑗 )
SSTotal = σ(𝑦𝑖𝑗 )2 -
𝑁

𝑀𝑆𝐶 𝑆𝑆𝐶/1
𝐹0 = =
𝑀𝑆𝐸 𝑀𝑆𝐸

165
Г𝑗 = σ 𝑐𝑖 µ𝑖

𝐶𝑗 =σ 𝑐𝑖 𝑦ത𝑖.

𝑆𝐶𝑢 = 𝑀𝑆𝐸 σ𝑎𝑖=1(𝐶𝑖𝑢 2 /𝑛𝑖 )

𝑆𝛼,𝑢 = 𝑆𝐶𝑢 (𝑎 − 1)𝐹𝛼,𝑎−1,𝑁−𝑎

𝑦ത𝑚𝑎𝑥 −𝑦ത 𝑚𝑖𝑛

q=
𝑀𝑆𝐸/𝑛

𝑇𝛼 = 𝑞𝛼 (a,f) 𝑀𝑆𝐸/𝑛

2 1 1
LSD = 𝒕 α
,𝑵−𝒂 𝑀𝑆𝐸 𝑑𝛼 (a-1 , f) 𝑀𝑆𝐸 ( + )
𝟐 𝑛 𝑛𝑖 𝑛𝑗 166