STAT 530 Experimental Design and Analysis

3.2. The Analysis of Variance.

Example. A textile company weaves a certain fabric on three looms. It

is suspected that different looms produce fabric of different strength. To test
the claim, a quality control engineer randomly chooses four samples of fabrics
for every loom. The observed data are textile strength in pounds per square
inch.
Looms Strength
1 88 93 90 89
2 91 89 92 94
3 97 96 94 93
Definition. In a one-way analysis of variance (one-way ANOVA), a single
factor (treatment) with several levels is considered.
In our example, the treatments are the looms. There are three treatments
(three levels of the factor).
In a one-way problem, the data are presented by the following table

Treatments Observations Totals Averages

1 y11 y12 . . . y1n y1. ȳ1.
2 y21 y22 . . . y2n y2. ȳ2.
. . . ... . . .
. . . ... . . .
. . . ... . . .
a ya1 ya2 . . . yan ya. ȳa.
y.. ȳ..

Definition. The statistical model for the observations is yij = µ + τi + εij ,

where yij is the jth observation for the ith treatment (the ijth observation),
i = 1, . . . , a, j = 1, . . . , n. The parameter µ is the overall mean, τi is the ith
treatment effect, and εij is a random error. It is assumed that εij ’s are i.i.d.
N (0, σ 2 ). This is a completely additive model (there is no interaction).
ANOVA hypotheses. We are interested in testing for zero treatment ef-
fect, that is, we would like to test H0 : τ1 = τ2 = · · · = τa = 0 versus
H1 : τi ̸= 0 for some i.

Definition. Alternatively, the statistical model is given by yij = µi + εij ,

where µi is the mean of the ith factor level (µi = µ + τi ). The ANOVA
hypotheses for this model are H0 : µ1 = µ2 = · · · = µa versus H1 : µi ̸= µj
for some i and j. ∑ ∑
Notice the relations µ = ai=1 µi /a and ai=1 τi = 0.

1
Derivation of the test statistic.

Definition. A fixed effects model is the model in which levels of the fac-
tor are chosen by the experimenter. In our example, the three looms are the
only looms at the factory, so it is a fixed effects model.
Definition. A random effects model is the model in which the a treatments
are a random sample from a larger population of treatments. For example, a
textile factory has one hundred looms. A quality control engineer randomly
chooses three looms, and obtains four random samples of fabrics for every
loom. This is an example of a random effects model.

We derive the test statistic for the fixed effects model.

Notation.
∑
n
yi.
yi. = yij , ȳi. = , i = 1, . . . , a,
j=1
n

∑
a ∑
n
y..
y.. = yij , ȳ.. = , N = na.
i=1 j=1
N

3.3.1. Decomposition of the Total Sum of Squares.

∑ ∑
Definition. The total corrected sum of squares is SST = ai=1 nj=1 (yij −
ȳ.. )2 . It represents the total variability in the data.
It can be partitioned into component parts (whence, the name “Analysis of
Variance”). It can be written as
∑
a ∑
n ∑
a ∑
n
SST = (yij − ȳ.. ) =
2
[(ȳi. − ȳ.. ) + (yij − ȳi. )]2
i=1 j=1 i=1 j=1

∑
a ∑
a ∑
n ∑
a ∑
n
=n (ȳi. − ȳ.. ) +
2
(yij − ȳi. ) + 2
2
(ȳi. − ȳ.. )(yij − ȳi. ).
i=1 i=1 j=1 i=1 j=1

The cross product is equal to zero:

∑
n
(yij − ȳi. ) = yi. − ȳi. = yi. − n(yi. /n) = 0.
j=1
∑
Thus, SST = SStr +SSE where SStr = n ai=1 (ȳi. −ȳ.. )∑
2
is the
∑sum of squares
due to treatments (between treatments), and SSE = i=1 j=1 (yij − ȳi. )2
a n

is the sum of squares due to error (within treatments).

3.3.2. Fixed Effects Model: Statistical Analysis.

∑ ∑
Recall that SST = SStr + SSE where SST = ai=1 nj=1 (yij − ȳ.. )2 , SStr =
∑ ∑ ∑
n ai=1 (ȳi. − ȳ.. )2 , SSE = ai=1 nj=1 (yij − ȳi. )2 .

2
 SStr SSE
The quantities M Str = a−1
and M SE = N −a
are called mean squares. It can
be shown that
∑a
n τi2
E(M Str ) = σ +2 i=1
and E(M SE ) = σ 2 .
a−1
Under H0 , both M SE and M Str are estimates of σ 2 . Moreover, it can be
shown (Cochran’s Theorem, page 69) that, under H0 , SSE /σ 2 and SStr /σ 2
are independent chi-square random variables with N − a and a − 1 degrees
of freedom. Therefore, if the null hypothesis is true, the ratio

SStr /(a − 1) M Str

F = =
SSE /(N − a) M SE

has an F distribution with a − 1 and N − a degrees of freedom. Thus, one

would reject the null, if F > Fα, a−1, N −a .

Computational Formulas.
∑
a ∑
n
y..2
SST = yij2 − ,
i=1 j=1
N

∑
a
y2 y..2
SStr = i.
− ,
i=1
n N
SSE = SST − SStr .
The one-way ANOVA table looks like this

Source of Sum of Degrees of Mean

Variation Squares Freedom Square F
Between treatments SStr a−1 M Str F =M Str
M SE
Error(within
treatments) SSE N −a M SE
Total SST N −1

3
Example. In the example with looms, test whether the looms produce fab-
ric of different strength. Use α = 0.05.
Solution: We want to test H0 : τ1 = τ2 = τ3 = 0 against H1 : τi ̸= 0 for some
i = 1, 2, 3.

Looms Strength Totals

1 88 93 90 89 360
2 91 89 92 94 366
3 97 96 94 93 380
1106
∑ ∑ 2
a = 3, n = 4, N = 12, SST = ai=1 nj=1 yij2 − yN.. = 102026 − 101936.33 =
∑a yi.2 2

i=1 n − N = 101989 − 101936.33 = 52.67,

y..
89.67, SStr = SSE =
SST − SStr = 37, M Str = a−1 = 2 = 26.34, M SE = N −a = 37
SStr 52.67 SSE
9
=
M Str 26.34
4.11, F = M SE = 4.11 = 6.41. The ANOVA table is

Source SS df MS F
Looms 52.67 2 26.34 6.41
Error 37 9 4.11
Total 89.67 11

F0.05, 2, 9 = 4.26 ⇒ reject H0 and conclude that the fabrics differ in strength.

3.3.3. Fixed Effects Model: Estimation of Model Parameters.

Proposition. If Xi ∼ N (µ, σ 2 ) where i = 1, . . . , n, then the MLE of µ

is µ̂ = X̄ = (X1 + · · · + Xn )/n.

The observations in a one-factor model satisfy yij = µi + εij ∼ N (µi , σ 2 )

for all j = 1, . . . , n and some fixed i = 1, . . . , a. Therefore, the MLE ∑ of µi
is µ̂i = (yi1 + · · ·∑+ yin )/n = ȳi. . Also, since the overall mean µ = µi /a,
its MLE is µ̂ = µ̂i /a = (ȳ1. + · · · + ȳa. )/a = (y1. /n + · · · + ya. /n)/a =
(y1. + · · · + ya. )/N = ȳ.. .
Writing the model as yij = µ + τi + εij , we see that the MLEs of the model
parameters are µ̂ = ȳ.. and τ̂i = ȳi. − ȳ.. .

In the example with looms, µ̂ = 1106/12 = 92.17, µ̂1 = 360/4 = 90, µ̂2 =
366/4 = 91.5, µ̂3 = 380/4 = 95, τ̂1 = 90−92.17 = −2.17, τ̂2 = 91.5−92.17 =
−0.67, τ̂3 = 95 − 92.17 = 2.83.

Confidence Intervals.

The distribution of ȳi. is N (µi , σ 2 /n). Use M SE as an estimate for σ 2 .

It can be shown that √ȳi. −µi ∼ tN − a . Therefore, a 100(1 − α)% confi-
M SE /n
√
dence interval for the ith treatment mean µi is ȳi. ± tα/2, N −a M SE /n. A
100(1 − α)% confidence interval for the difference in two treatment means

4
 √
µi − µj is ȳi. − ȳj. ± tα/2, N −a 2M SE /n.

In our example, the 95% confidence intervals for the treatment means and
the difference in treatment means are computed as follows:

ȳ1. = 90, ȳ2. = 91.5, ȳ3. = 95, M SE = 4.11, N = 12, a = 3, t.025, 9 = 2.262,
n = 4, 87.71 ≤ µ1 ≤ 92.29, 89.21 ≤ µ2 ≤ 93.79, 92.71 ≤ µ3 ≤ 97.29,
−1.74 ≤ µ2 − µ1 ≤ 4.74, 1.76 ≤ µ3 − µ1 ≤ 8.24, 0.26 ≤ µ3 − µ2 ≤ 6.74.
Notice that the confidence interval for the difference between µ2 and µ1 in-
cludes zero, but the other two intervals don’t. It means that there is no
statistical difference between looms one and two, but there is difference be-
tween looms one and three, and two and three.

Example. A manufacturer of television sets is interested in the effect on

tube conductivity of four different types of coating for color picture tubes.
The following conductivity data are obtained:

Coating Type Conductivity Total

1 143 141 150 146 580
2 152 149 137 143 581
3 134 136 132 129 531
4 129 127 132 129 517
2209
(a) Write down the math model for this experiment.
iid
Solution: yij = µi + εij , i = 1, . . . , 4, j = 1, . . . , 4, ε ∼ N (0, σ 2 ), µ̂1 =
145, µ̂2 = 145.25, µ̂3 = 132.75, µ̂4 = 129.25, √ SST = 1040.94, SStr =
822.69, SSE = 218.25, M SE = 18.19, σ̂ = M SE = 4.26.
iid
Or, alternatively, yij = µ + τi + εij , i = 1, . . . , 4, j = 1, . . . , 4, ε ∼
N (0, σ 2 ), σ̂ = 4.26, µ̂ = 138.06, τ̂1 = 6.94, τ̂2 = 7.19, τ̂3 = −5.31, τ̂4 =
−8.81.

(b) Construct the 99% confidence intervals for the treatment means and
for the differences in treatment means.
Solution: t0.005, 12 = 3.055, 138.49 ≤ µ1 ≤ 151.51, 138.74 ≤ µ2 ≤ 151.76, 126.24 ≤
µ3 ≤ 139.26, 122.74 ≤ µ4 ≤ 135.76, −8.98 ≤ µ2 − µ1 ≤ 9.48, 3.02 ≤
µ1 − µ3 ≤ 21.48, −5.73 ≤ µ3 − µ4 ≤ 12.73.

3.5.7. Comparing Pairs of Treatment Means.

To compare all pairs( )of treatment means in a one-way ANOVA model, we

test simultaneously a2 pairs of hypotheses H0 : µi = µj for all i ̸= j against
H1 : µi ̸= µj for some i ̸= j.

We study three tests: Tukey’s test, the Fisher Least Significant Difference

5
(LSD) method, and Duncan’s Multiple Range test.

Tukey’s Honestly Significant Difference (HSD) Test.

The test declares two means significantly different if the absolute value of
the difference of the√ respective sample treatment means exceeds
Tα = qα (a, N − a) M SE /n where the critical value qα (a, N − a) comes from
table VIII on pages 656-657.
Example. In the example with looms, we would like to test simultaneously
three pairs of hypotheses H0 : µ1 = µ2 against H1 : µ1 ̸= µ2 , H0 : µ2 = µ3
against H1 : µ2 ̸= µ3 , and H0 : µ1 = µ3 against H1 : µ1 ̸= µ3 .
The three sample treatment means are ȳ1. = 90, ȳ2. = 91.5, ȳ3. = 95. The
differences in the means are
ȳ1. − ȳ2. = −1.5
ȳ2. − ȳ3. = −3.5
ȳ1. − ȳ3. = −5.
The absolute values √ of these quantities
√ should be compared to
T0.05 = q0.05 (3, 9) M SE /n = 3.95 4.11/4 = 4.00. The conclusion is that
there is a significant difference between µ1 and µ3 but the differences between
µ1 and µ2 , and µ2 and µ3 are insignificant. This can be presented graphically
as

ȳ1. ȳ2. ȳ3.

The Fisher Least Significant Difference (LSD) Method.

√
The means µi and µj are significantly different if |ȳi. −ȳj. | > tα/2, N −a 2M SE /n.
√ √
Example. In our example, t0.025, 9 2M SE /n = 2.262 4.11/2 = 3.24. The
conclusion is that µ1 and µ2 are not different, but µ1 and µ3 and µ2 and µ3
are different.

ȳ1. ȳ2. ȳ3.

Duncan’s Multiple Range Test.

Step 1. Arrange the sample means in increasing order.√

Step 2. Compute a − 1 quantities Rp = rα (p, N − a) M SE /n, p = 2, . . . a,
where the significant ranges rα (p, N − a) can be found in table VII on page
655.
Step 3. Compare the difference between the largest and the smallest means
to Ra , between the largest and the second-smallest to Ra−1 , and so on, until
all means are compared to the largest. Then, the difference between the
second-largest
(a ) and the smallest means are compared to Ra−1 , etc. Until all
2
pairs of means has been considered.

6
 √ √
Example. R2 = √ r0.05 (2, 9) M S
√ E /n = 3.2 4.11/4 = 3.24,
R3 = r0.05 (3, 9) M SE /n = 3.34 4.11/4 = 3.39.
ȳ3. − ȳ1. = 5 > 3.39(R3 )
ȳ3. − ȳ2. = 3.5 > 3.24(R2 )
ȳ2. − ȳ1. = 1.5 < 3.24(R2 )
The conclusion is that µ1 and µ3 , as well as µ2 and µ3 differ significantly, but
µ1 and µ2 do not.

ȳ1. ȳ2. ȳ3.

3.5.4. Contrasts.

Definition. A ∑contrast is a linear combination of the treatment means of

a
the
∑a form Γ = i=1 ci µi where the coefficients c1 , . . . , ca sum to zero, that is,
i=1 ci = 0. ∑ ∑
The goal is to test H0 : ai=1 ci µi = 0 against H1 : ai=1 ci µi ̸= 0.
∑a
We∑estimate the ∑ contrast by Γ̂ = i=1 ci ȳi. . The distribution of Γ̂ is
N ( ai=1 ci µi , ai=1 c2i σ 2 /n). Therefore, under the null hypothesis, the vari-
able ∑a
ci ȳi.
√∑ai=1
2 2
i=1 ci σ /n

has a N (0, 1) distribution. Thus, the test statistic

∑a
i=1 ci yi.
t= √ ∑ .
nM SE ai=1 c2i

has the t-distribution with N − a degrees of freedom. The null hypothesis

should be rejected if |t| > tα/2,
∑ N− a
.
( a ci yi. )2
Or, equivalently, F = t = nM SE a c2 has F distribution with 1 and N − a
2 i=1∑
i=1 i
degrees of freedom. The null should be rejected if F > Fα, 1, N − a .

Example. In the example with looms, suppose we would like to test H0 :

µ2 = (µ1 + µ3 )/2 against H1 : µ2 ̸= (µ1 + µ3 )/2. The contrast Γ =
µ1 − 2µ2 + µ3 .
Notice that, for example, testing H0 : µ3 = 2µ1 cannot be expressed in terms
of a contrast since Γ = 2µ1 − µ3 is not a contrast.
The test statistic is
360 − (2)(366) + 380
t= √ = 0.81.
(4)(4.11)(1 + 4 + 1)

The critical value t0.025, 9 = 2.262. We accept the null and conclude that the
mean strength for loom 2 equals the average of the means for loom 1 and 3.

7
3.5.5. Orthogonal Contrasts.

Definition.
∑a Two contrasts with coefficients {ci } and {di } are orthogonal
if i=1 ci di = 0.
Example. Γ1 = µ1 − 2µ2 + µ3 and Γ2 = µ1 − µ3 are orthogonal since
(1)(1) + (−2)(0) + (1)(−1) = 0.

For a treatments the set of a − 1 orthogonal contrasts partition the sum

of squares due to treatments into a − 1 independent single-degree-of-freedom
components.
∑a Denote by C a contrast in terms of the treatment totals, that
is, C = i=1 ci yi. . Then the contrast sum of squares is
∑
( ai=1 ci yi. )2
SSC = ∑ .
n ai=1 c2i
∑
( a c y )2 SSC /1
It can be shown by noticing that F = t2 = nM Si=1∑ia i. 2 = M SC = .
E i=1 ci M SE M SE
Example. In the example with looms, suppose we would like to test the
hypotheses H0 : µ2 = (µ1 + µ3 )/2 and H0 : µ1 = µ3 . The sums of squares of
the corresponding orthogonal contrasts

C1 = y1. − 2y2. + y3. = 360 − (2)(366) + 380 = 8, and C2 = y1. − y3. = −20

are
(8)2 (−20)2
SSC1 = = 2.67, and SSC2 = = 50.
(4)(6) (4)(2)
Notice that SStr = 52.67 = SSC1 + SSC2 .
The ANOVA table is
Source SS df MS F
Looms 52.67 2 26.34 6.41
C1 : 2µ2 = µ1 + µ3 (2.67) 1 2.67 0.65
C2 : µ1 = µ3 (50) 1 50 12.17
Error 37 9 4.11
Total 89.67 11
Since F0.05, 1, 9 = 5.12, we reject H0 : µ1 = µ3 and conclude that the means
for loom and one and loom three differ significantly, whereas we accept
H0 : 2µ2 = µ1 + µ3 and conclude that the mean for loom 2 equals the
average of the means for looms 1 and 3.

3.4. Model Adequacy Checking.

In a one-way ANOVA the observations are described by the model yij =

µi + εij . The hypotheses testing of no difference in treatment means is valid
iid
only if the following assumptions on the model are satisfied: ε ∼ N (0, σ 2 ).
There are, in fact, three assumptions: (i) identical Normal distribution of
the errors, (ii) independence of the random errors, and (iii) constant vari-
ance (“homoscedasticity”).

8
Prior to performing the analysis of variance, it is wise to check the validity of
the assumptions (“model diagnostic checking”). It can be done by examining
the residuals.
Definition. The residual for ijth observation is eij = yij − ŷij where ŷij is
the fitted value, the estimate of the ijth observation, ŷij = µ̂i = ȳi. . Thus,
eij = yij − ȳi. .

3.4.1. The Normality Assumption.

(a) Plot a histogram of the residuals. It should display a bell-shaped figure
centered at zero.
(b) Construct a Normal probability plot (QQ-plot) of the residuals. It should
display a linear pattern.
(c) Perform normality tests (see SAS output).

3.4.2. Plot of Residuals in Time Sequence.

To verify the independence assumption, plot residuals in time order (if possi-
ble). If a pattern is detected, then there is a correlation between the residuals.
The pattern might be a run of positive residuals, then a run of negatives;
a megaphone a linear pattern. If the plot is structureless, the residuals are
independent.

3.4.3. Plot of Residuals Versus Fitted Values.

(a) To verify for equality of variances, plot residuals against the fitted values.
Again, the plot should show no structure.
(b) To test the homogeneity of variance, perform the Bartlett test (see pages
84-85 and SAS output). The hypotheses of interest are H0 : σ12 = σ22 =
· · · = σa2 and H1 : H0 is false for at least one σi2 .

4.1. The Randomized Complete Block Design.

Example. The quality control engineer suspects that the difference in

fabric strength for the three looms is caused by difference in work experience
and skills of the three operators. He decides to set up an experiment with
operators being the blocking factor, a nuisance source of variability. Within
a block, the order in which the three looms are tested is randomly deter-
mined. He gets the following results:

Operators
Looms 1 2 3
1 88 93 90
2 90 92 92
3 91 96 94
Notice that each loom is tested for each operator, that is, each block contains
all treatments. This defines a randomized complete block design.

9
4-1.1. Statistical Analysis of the RCBD.

Definition. The statistical model for this design is yij = µ + τi + βj +

εij , i = 1, . . . , a, j = 1, . . . , b, where µ is the overall mean, τi is the ith
treatment effect, βj is the jth block effect, and εij is a random error. It is
assumed that εij ’s are i.i.d. N (0, σ 2 ). Since τi ’s and βj ’s are defined
∑aas devia-
tions from the overall mean, the assumptions are imposed such as i=1 τi = 0
∑
and bj=1 βj = 0. Notice that this is an additive model.
We are interested in testing for zero treatment effect, that is, H0 : τ1 = τ2 =
· · · = τa = 0 versus H1 : τi ̸= 0 for some i.
Note. We may also be interested in testing for zero blocking effect since, if
H0 : β1 = β2 = · · · = βb = 0 is accepted, blocking may not be necessary in
future experiments.
Derivation of the test statistic.

Notation.
∑
b
yi.
yi. = yij , ȳi. = , i = 1, . . . , a,
j=1
b
∑
a
y.j
y.j = yij , ȳ.j = , j = 1, . . . , b,
i=1
a
∑
a ∑
b ∑
a ∑
b
y..
y.. = yij = yi. = y.j , ȳ.. = , N = ab.
i=1 j=1 i=1 j=1
N
We express the total corrected sum of squares as
∑
a ∑
b ∑
a ∑
b
SST = (yij − ȳ.. ) =
2
[(ȳi. − ȳ.. )+(ȳ.j − ȳ.. )+(yij − ȳi. − ȳ.j + ȳ.. )]2
i=1 j=1 i=1 j=1

∑
a ∑
b ∑
a ∑
b
=b (ȳi. − ȳ.. ) + a
2
(ȳ.j − ȳ.. ) +
2
(yij − ȳi. − ȳ.j + ȳi. )2
i=1 j=1 i=1 j=1

∑
a ∑
b ∑
a ∑
b
+2 (ȳi. − ȳ.. )(ȳ.j − ȳ.. ) + 2 (ȳi. − ȳ.. )(yij − ȳi. − ȳ.j + ȳi. )
i=1 j=1 i=1 j=1

∑
a ∑
b
+2 (ȳ.j − ȳ.. )(yij − ȳi. − ȳ.j + ȳi. )
i=1 j=1

= {the last three cross products are zero!} = SStr + SSbl + SSE
where SStr is the sum of squares due to treatments, SSbl is the sum of squares
due to blocks, and SSE is the sum of squares due to error. Define the mean
squares by M Str = SS tr
a−1
, M Sbl = SS
b−1
bl SSE
, and M SE = (b−1)(a−1) . It can be
shown that ∑
b ai=1 τi2
E(M Str ) = σ +
2
a−1

10
 ∑b
a j=1 βj2
E(M Sbl ) = σ 2 +
b−1
E(M SE ) = σ . 2

To test H0 : τi = 0, we would use the test statistic F = M Str

M SE
which is
distributed as Fa−1,(a−1)(b−1) if the null hypothesis is true. Thus, we would
reject the H0 , if F > Fα, a−1, (a−1)(b−1) .
Remark. It seems that the hypothesis H0 : βj = 0 may be tested by com-
M Sbl
paring the statistic F = M SE
to Fα, b−1, (a−1)(b−1) . However, the randomization
has been applied only to treatments within blocks, making the F distribu-
M Sbl
tion questionable. It is advisable, however, to examine the ratio M SE
. If it
is large, the blocking factor has a large effect.

The ANOVA table looks like this

Source of Variation SS DF MS F
∑ 2
yi. 2
Treatments b
− y..
ab
a−1 SStr
a−1
M Str
M SE
∑ 2
y.j 2
Blocks a
− y..
ab
b−1 SSbl
b−1
Error SST − SStr − SSbl (a − 1)(b − 1) SSE
∑∑ 2 2
(a−1)(b−1)
Total yij − yab.. ab − 1

Example. In our example,

Operators
Looms 1 2 3 Totals
1 88 93 90 271
2 90 92 92 274
3 91 96 94 281
Totals 269 281 276 826
∑∑ 2 2 ∑ yi.2
a = b = 3, SST = yij − yab.. = 75854 − 682276
9
≈ 45.56, SStr = b
−
2 ∑ y 2 2
y..
ab
= 227478
3
− 682276
9
≈ 75826 − 75808.44 = 17.56, SSbl = .j
a
− yab.. ≈
227498
3
−75808.44 = 24.23, SSE = SST −SStr −SSbl = 45.56−17.56−24.23 =
3.77, M Str = SS tr
a−1
= 17.56
2
= 8.78, M Sbl = SS bl
b−1
= 24.23
2
≈ 12.12, M SE =
SSE
(a−1)(b−1)
= 4 ≈ 0.94, F = M SE = 0.94 ≈ 9.34. The ANOVA table is
3.77 M Str 8.78

Source of Variation SS DF MS F
Looms 17.56 2 8.78 9.34
Operators 24.23 2 12.12
Error 3.77 4 0.94
Total 45.56 8

Fα, a−1, (a−1)(b−1) = F0.05, 2, 4 = 6.94. Therefore, we reject H0 : τi = 0 and

conclude that the looms produce fabric of different strength. Also, operators
seem to differ significantly, since the mean square for blocks is large relative

11
to error.

4.2. The Latin Square Design.

Example. The effect of five different formulations of an explosive mixture

on the explosive force is studied. Each formulation is mixed from a batch of
raw material that is only large enough for five formulations to be tested. Fur-
thermore, the formulations are prepared by several operators, and there may
be substantial difference in the skills and experience of the operators. Thus,
there are two blocking factors: batches of raw material and operators. The
appropriate design for this problem consists of testing each formulation ex-
actly once in each batch and for each formulation to be prepared exactly once
by each of five operators. The resulting design is called a Latin square design.
The design is a square arrangement and the five treatments (formulations)
are denoted by the Latin letters A, B, C, D, and E (hence, the name “Latin
square”). The coded data are

Operators yi..
Batches 1 2 3 4 5
1 A=-1 B=-5 C=-6 D=-1 E=-1 -14
2 B=-8 C=-1 D=5 E=2 A=11 9
3 C=-7 D=13 E=1 A=2 B=-4 5
4 D=1 E=6 A=1 B=-2 C=-3 3
5 E=-3 A=5 B=-5 C=4 D=6 7
y..k -18 18 -4 5 9 10=y...
Definition. The statistical model for this design is yijk = µ + αi + τj + βk +
εijk , i, j, k = 1, . . . , p, where yijk is the observation in the ith row and kth
column for the jth treatment, µ is the overall mean, αi is the ith row effect,
τj is the jth treatment effect, βk is the kth column effect, and εijk is a random
error. It is assumed that εijk ’s are i.i.d. N (0, σ 2 ). Notice that only two out of
the three subscripts are necessary to identify an observation.
∑p This
∑p model is
∑p The assumptions on the effect terms are i=1 αi = 0,
additive. j=1 τj = 0,
and k=1 βk = 0.
We are interested in testing for zero treatment effect, that is, H0 : τ1 = τ2 =
· · · = τp = 0 versus H1 : τj ̸= 0 for some j.
The ANOVA table for the Latin square design is

Source SS DF MS F
∑p y2 2
Treatments SStr = j=1 p.j. − yp...2 p−1 SStr M Str
∑ y2 2
p−1 M SE
Rows SSrows = pi=1 pi.. − yp...2 p−1 SSrows
∑ y2 2
p−1
Columns SScol = pk=1 ..k
p
− yp...2 p−1 SScol
p−1
Error SSE (by subtraction) (p − 2)(p − 1) SSE
∑ ∑ ∑ 2 2
(p−2)(p−1)
Total SST = i j k yijk − y...
p2
p2 − 1
If the test statistic F > Fα, p−1, (p−2)(p−1) , the null is rejected. In addition,
we might want to test for no row effect and no column effect by forming the

12
ratios of M Srow or M Scol to M SE . However, since the rows and columns
represent restrictions on randomization, these tests may not be appropriate.
2
Example. In our example, p = 5, SST = 680 − 10 25
= 676, SSbatches =
(−14)2 +92 +52 +32 +72 102 (−18)2 +182 +(−4)2 +52 +92 2
5
− 25 = 68, SSoper = 5
− 10
25
= 150.
Formulation Total
A y.1. = 18
B y.2. = −24
C y.3. = −13
D y.4. = 24
E y.5. = 5
182 +(−24)2 +132 +242 +52 2
SSf orm = 5
− 10
25
= 330, SSE = 128. The ANOVA table is

Source SS DF MS F
Formulations 330 4 82.5 7.73
Batches 68 4 17
Operators 150 4 37.5
Error 128 12 10.67
Total 676 24
F = 7.73 > F.05, 4, 12 = 3.26, therefore, we reject H0 and conclude that there
is a significant difference between the formulations. Also, MMSSoper
E
= 3.51, in-
M Sbatches
dicating that there are differences between operators, while M SE = 1.59
implies no difference between batches. So, blocking on operators is reason-
able, whereas blocking on batches is unnecessary (though advisable).

4.3. The Græco-Latin Square Design.

The Græco-Latin square design is used to block in three directions. It can be

viewed as superposition of two Latin square designs in which treatments are
denoted by Latin letters, and the rows, columns, and Greek letters represent
blocks. Each Greek letter appears exactly once with each Latin letter.
Example. Suppose that in the explosive force example, test assemblies is
an additional blocking factor, so we run a Græco-Latin square design. The
design is
Operators yi...
Batches 1 2 3 4 5
1 Aα=-1 Bγ=-5 Cε=-6 Dβ=-1 Eδ=-1 -14
2 Bβ=-8 Cδ=-1 Dα=5 Eγ=2 Aε=11 9
3 Cγ=-7 Dε=13 Eβ=1 Aδ=2 Bα=-4 5
4 Dδ=1 Eα=6 Aγ=1 Bε=-2 Cβ=-3 3
5 Eε=-3 Aβ=5 Bδ=-5 Cα=4 Dγ=6 7
y...k -18 18 -4 5 9 10=y....
Definition. The statistical model for this design is yijkl = µ + θi + τj +
ωk + Ψl + εijkl , i, j, k, l = 1, . . . , p, where yijkl is the observation in the

13
ith row, Latin letter j, Greek letter k, and lth column, µ is the overall
mean, θi is the ith row effect, τj is the effect of Latin letter j treatment,
ωk is the effect of Greek letter k blocking factor, Ψl is the lth column ef-
fect, and εijkl ’s are i.i.d. N (0, σ 2 ) errors. Two subscripts suffice to iden-
tify an observations. Græco-Latin square designs exist for all p ≥ 3 except
p∑= 6. The ∑ model is additive.
∑p The assumptions
∑p on the effect terms are
p p
θ
i=1 i = 0, τ
j=1 j = 0, k=1 kω = 0, and l=1 l = 0.
Ψ
We are interested in testing for zero treatment effect, that is, H0 : τ1 = τ2 =
· · · = τp = 0 versus H1 : τj ̸= 0 for some j.
The ANOVA table for the Græco-Latin square design is

Source SS DF MS F
∑p y2 2
Treatments SSL = j=1 .j.. − yp....
2 p−1 M SL
∑p y..k.
p
2 2
M SE
Greek letter SSG = k=1 p − p2 y....
p−1
∑ y2 2
Rows SSrows = pi=1 i... − yp....
2 p−1
∑p y...l
2
p
2
Columns SScol = l=1 p − p2 y....
p−1
Error SSE (by subtraction) (p − 3)(p − 1)
∑ ∑ ∑ ∑ 2 2
Total SST = i j k l yijkl − y....
p2
p2 − 1
If the test statistic F > Fα, p−1, (p−3)(p−1) , the null is rejected. In addition, we
might want to test for no row effect, or no column effect, or no Greek let-
ter factor effect by forming the ratios of M Srows or M Scol or M SG to M SE .
However, since the blocks represent restrictions on randomization, these tests
may not be appropriate.

Example. In our example, p = 5, SST = 676, SSbatches = 68, SSoper =

150, SSf orm = 330.
The total for the test assemblies are
Greek letter Total
α 10
β -6
γ -3
δ -4
ε 13

The sum of squares due to assemblies is SSassem = 330/5 − 100/25 = 62.

The error sum of squares is SSE = 66. The ANOVA table is

Source SS DF MS F
Formulations 330 4 82.5 10.0
Batches 68 4 17
Operators 150 4 37.5
Assemblies 62 4 15.5
Error 66 8 8.25
Total 676 24

14
F = 10 > F.05, 4, 8 = 3.84, therefore, we reject H0 and conclude that there
is a significant difference between the formulations. Also, MMSSoper
E
= 4.55, in-
M Sbatches
dicating that there are differences between operators, while M SE = 2.06
and MM Sassem
SE
= 1.88 imply no difference between batches or test assemblies.
So, blocking on operators is reasonable, whereas blocking on batches and
assemblies is unnecessary (though advisable).

4.4. The Balanced Incomplete Block Design.

Example. A chemical engineer suspects that the time of reaction for a chem-
ical process is a function of the type of catalyst employed. Four catalysts
are investigated. The experiment consists of taking a batch of raw material,
loading the pilot plant, applying each catalyst separately, and observing the
reaction time. The batches of the raw material are used as blocks. The
order in which the catalysts are run in each block is randomized. However,
each batch is only large enough for three catalysts to be run. Therefore, an
incomplete block design must be used.
Definition. An incomplete block design is a randomized block design, in
which every treatment is not present in every block.
Definition. A balanced incomplete block design is an incomplete block de-
sign in which any two treatments appear together an equal number of times.
Example (continued...). The balanced incomplete block design for this
experiment is

Batch of Raw Material

Catalyst 1 2 3 4 Totals
1 73 74 – 71 218
2 – 75 67 72 214
3 73 75 68 – 216
4 75 – 72 75 222
Totals 221 224 207 218 870

4-4.1. Statistical Analysis of the BIBD.

Notation. We assume that there are a treatments and b blocks. In addition,

we assume that each block contains k treatments, that each treatment occurs
r times in the design, and that there are N = ar = bk total observations.
Proposition. The number of times each pair of treatments appears in the
same block is λ = r(k−1)
a−1
.
Proof: Consider, for example, treatment 1. It appears in r blocks and there
are k − 1 other treatments in each of those blocks. Therefore, there are
r(k − 1) additional observations in the design, which must represent the re-
maining a − 1 treatments λ times. Thus, (a − 1)λ = r(k − 1). 2
Definition. The statistical model for this design is yij = µ + τi + βj + εij ,
where µ is the overall mean, τi is the ith treatment effect, βj is the jth
block effect, and εij is a random error. It is assumed that εij ’s are i.i.d.

15
N (0,∑σ 2 ). This is an∑
additive model. The assumptions on the effect terms
are ai=1 τi = 0 and bj=1 βj = 0.
We are interested in testing for zero treatment effect, that is, H0 : τ1 = τ2 =
· · · = τa = 0 versus H1 : τi ̸= 0 for some i.
Derivation of the test statistic.
∑∑ y..2
SST = yij2 − = SStr(adj) + SSbl + SSE
i j
N

where the sum of squares for treatments is adjusted to separate the treatment
and the block effects. This adjustment is necessary because each treatment
is represented in a different set of r blocks.
∑ y2 2
The block sum of squares SSbl = bj=1 k.j − yN.. has b − 1 degrees of freedom.
The adjusted treatment sum of squares is
∑
k ai=1 Q2i
SStr(adj) =
λa
∑
where Qi = yi. − k1 b
j=1 nij y.j , i = 1, . . . , a with nij = 1 if treatment i ap-
pears in block j and nij = 0 otherwise. Qi ’s are called the adjusted treatment
totals and always sum up to zero.
SStr(adj) has a − 1 degrees of freedom. The error sum of squares is computed
by subtraction; it has N − a − b + 1 degree of freedom. The test statistic is
MS
F = Mtr(adj)
SE
. The null hypothesis is rejected if F > Fα, a−1, N −a−b+1 .

The ANOVA table looks like this

Source of Variation SS
∑
DF MS F
k Q2i SStr(adj) M Str(adj)
Treatments (adj) λa
a−1 a−1 M SE
∑ 2
y.j 2
Blocks k
− y..
N
b−1
Error SST − SStr(adj) − SSbl N −a−b+1 SSE
N −a−b+1
∑∑ 2 2
Total yij − yN.. N −1

Remark. We might be interested in testing H0 : βj = 0 ∀j. To do so, com-

∑ y2 2
pute SStr = ai=1 ri. − yN.. and SSbl(adj) = SST − SStr − SSE . If the ratio
SSbl(adj) /(b−1)
SSE /(N −a−b+1)
is large, we reject the null.
Example. In our example, a = 4, b = 4, r = 3, k = 3, λ = 3(3−1) =
∑ ∑ 2 4−1y..2
2, N = (4)(3) = 12. The sum of squares are SST = i j yij − 12 =
2 ∑ y 2 2 2 2 2 +(218)2 2
63, 156 − (870)
12
= 81, SSbl = j 3.j − y12.. = (221) +(207) +(224)
3
− (870)
12
=
55. The adjusted treatment totals are Q1 = 218 − 3 (221 + 224 + 218) =
1

−9/3, Q2 = 214− 13 (224+207+218) = −7/3, Q3 = 216− 13 (221+224+207) =

−4/3, Q4 = 222− 13 (221+207+218)
∑
= 20/3. The adjusted sum of squares for
k Q2 2 2 2 2
treatments is SStr(adj) = λia i = (3)((−9/3) +(−7/3) +(−4/3) +(20/3) )
(2)(4)
= 22.75.
The error sum of squares is SSE = SST − SSbl − SStr(adj) = 81 − 55 − 22.75 =
3.25. The ANOVA table is

16
 Source of Variation SS DF MS F
Catalysts 22.75 3 7.58 11.66
Batches 55.00 3
Error 3.25 5 0.65
Total 81.00 11
The test statistic F = 11.66 > F0.05, 3, 5 = 5.41, therefore, the null hypothesis
of no treatment effect is rejected, and we conclude that the catalyst employed
has a significant effect on the time of reaction. Now, to test whether there is
a blocking effect, compute

SStr = 11.67, SSE = 3.25, SSbl(adj) = 66.08

SSbl(adj) /(b − 1)
=⇒ = 33.89.
SSE /(N − a − b + 1)
We conclude that it is reasonable to use batches as the blocking factor.

5.3. The Two-Factor Factorial Design.

Suppose there are two treatments (factors) A and B. An experiment is

run for all possible combinations of the levels of the factors, and there are,
in general, more than one replicate in each cell.
Example. It is suspected that temperature and concentration might affect
the yield of a chemical process. The data are

Concentration
Temperature 40 60 80
50 17 16 24
20 21 22

75 12 18 17
9 13 12

100 16 18 25
12 21 23

5.3.2. Statistical Analysis of the Fixed Effects Model.

Definition. The statistical model is yijk = µ + τi + βj + (τ β)ij + εijk , i =

1, . . . , a, j = 1, . . . , b, k = 1, . . . , n, where µ is the overall mean, τi is the
effect of the ith level of treatment A, βj is the effect of the jth level of treat-
ment B (main effects), (τ β)ij is the effect of the interaction between τi and
βj , and εijk ∼ N (0, σ 2 ) is a random error. Notice that this model is not
additive. ∑ ∑b
The assumptions on the effect terms are ai=1 τi = 0, j=1 βj = 0,

17
∑a ∑b
i=1 (τ β)ij = 0, j = 1, . . . , b, and j=1 (τ β)ij = 0, i = 1, . . . , a.
The hypotheses of interest are H0 : τi = 0 ∀i, H0 : βj = 0 ∀j, or
H0 : (τ β)ij = 0 ∀i, j. The ANOVA table for the analysis is
Source SS DF F
∑a2
yi.. 2

i=1 bn − abn a−1

y... M SA
A M SE
∑b y.j.2 2

j=1 an − abn b−1

y... M SB
B M SE
∑a ∑b yij.2 2

j=1 n − abn − SSA − SSB (a − 1)(b − 1)

y... M SAB
AB i=1 M SE
Error SS − SS − SS − SSAB ab(n − 1)
∑aT ∑b A ∑n B 2 2

k=1 yijk − abn abn − 1

y...
Total i=1 j=1

Example (continued...). The ANOVA table is

Source SS DF MS F
Temperature 150.11 2 75.06 10.72
Concentration 114.78 2 57.39 8.20
Interaction 40.55 4 10.14 1.45
Error 63 9 7
Total 368.44 17
F0.05,2,9 = 4.26, F0.05,4,9 = 3.63, therefore, we conclude that there are main
effects, but no interaction present.

The Main Effect Plots and the Interaction Plot.

Proposition. The estimated effect of the ith level of factor A is ȳi.. , that of
the jth level of factor B is ȳ.j. , and that of their interaction is ȳij. .
Proof: The model parameter estimates for a two-factor factorial design are
(see Section 5-3.4) µ̂ = ȳ... , τ̂i = ȳi.. − ȳ... , β̂j = ȳ.j. − ȳ... , and (τˆβ)ij =
ȳij. − ȳi.. − ȳ.j. + ȳ... .
The estimated main effect of the ith level of factor A is µ̂ + τ̂i = ȳ... +
ȳi.. − ȳ... = ȳi.. . Likewise, the effect of the jth level of factor B is esti-
mated by ȳ.j. . The estimate of the observation in the ijth cell is given by
ŷijk = µ̂ + τ̂i + β̂j + (τˆβ)ij = ȳij. .

Definition. A main effect plot displays the estimated values for different
levels of a factor.
In our example, the estimates for factor A are ȳ1.. = 20, ȳ2.. = 13.5, ȳ3.. =
19.17, and the estimates for factor B are ȳ.1. = 14.33, ȳ.2. = 17.83, ȳ.3. = 20.5
Definition. An interaction plot displays the estimated values for different
treatment-treatment combinations.
In our example, these estimated values are ȳ11. = 18.5, ȳ12. = 18.5, ȳ13. =
23.0, ȳ21. = 10.5, ȳ22. = 15.5, ȳ23. = 14.5, ȳ31. = 14.0, ȳ32. = 19.5, ȳ33. =
24.0.

6.2. The 22 Factorial Design.

18
Definition. The 22 factorial design is a design with two factors A and B,
each at two levels. The levels of the factors may be arbitrarily called “low”
and “high.”
Example. In a chemical reaction, the reactant concentration is factor A run
at two levels, 15% and 25%, and the catalyst is factor B, with two levels,
one bag used or two bags used. The experiment is replicated three times,
and the data are
Replicate
Treatments I II III Total
A low, B low 28 25 27 (1) = 80
A high, B low 36 32 32 a = 100
A low, B high 18 19 23 b = 60
A high, B high 31 30 29 ab = 90

The effect of treatment A at the low level of B is [a − (1)]/n and the effect of
A at the high level of B is [ab − b]/n. Therefore, the main effect of treatment
A is the average of this two quantities
1 1 [contrast]
A= [ab − b] + [a − (1)] = [a + ab − (1) − b] = .
2n 2n 2n
Likewise, the main effect of treatment B is
1 1
B= [ab − a] + [b − (1)] = [b + ab − (1) − a].
2n 2n
The interaction between A and B is the average difference between the effect
of A at the high level of B and the effect of A at the low level of B. Thus,
1 1
AB = [ab − b] − [a − (1)] = [(1) + ab − a − b].
2n 2n
Remark. Alternatively, AB can be defined as the average difference be-
tween the effect of B at the high level of A and the effect of B at the low
level of A. Both definitions lead to the same equation.

Example (continued...) A = 2(3) 1

(100 + 90 − 80 − 60) = 8.33, B =
1
2(3)
(60 + 90 − 80 − 100) = −5.00, AB = 2(3)
1
(80 + 90 − 100 − 60) = 1.67.

To test for main and interaction effects, compute the sum of squares
[contrast]2 ∑∑∑ y2
SS = , SST = 2
yijk − ... ,
4n 4n
SSE = SST − SSA − SSB − SSAB .
Example (continued...) The ANOVA table is
Source SS DF MS F
A 208.33 1 208.33 53.15
B 75.00 1 75.00 19.13
AB 8.33 1 8.33 2.13
Error 31.34 8 3.92
Total 323.00 11

19
F0.05, 1, 8 = 5.32. There are main effects of A and B but no interaction.
Remark. To coefficients for estimating the effects are

Effect (1) a b ab
A -1 +1 -1 +1
B -1 -1 +1 +1
AB +1 -1 -1 +1

The coefficients for estimating AB is the product of the corresponding coef-

ficients for the two main effects.
Example. The following data were obtained by running four replicates of a
22 design. Analyze the data and draw conclusions.

Replicate
A B I II III IV
- - 18.2 18.9 12.9 14.4
+ - 27.2 24.0 22.4 22.5
- + 15.9 14.5 15.1 14.2
+ + 41.0 43.0 36.3 39.9
6.3. The 23 Factorial Design.

Definition. The 23 factorial design is a design with three factors, A, B,

and C, each at two levels.
Algebraic signs for calculating the effects in the model are:

Treatment Factorial Effect

Combination I A B AB C AC BC ABC
(1) + - - + - + + -
a + + - - - - + +
b + - + - - + - +
ab + + + + - - - -
c + - - + + - - +
ac + + - - + + - -
bc + - + - + - + -
abc + + + + + + + +

For example,

1 [contrast]
A= [a + ab + ac + abc − (1) − b − c − bc] = .
4n 4n
To test the main and interaction effects, compute

[contrast]2 ∑∑∑∑ y2
SS = , SST = 2
yijkl − .... , SSE by subtraction.
8n 8n
Example. The following data were obtained by running two replicates of a
23 design. Analyze the data and draw conclusions.

20
 Treatment Replicate
Combination I II
(1) 18.2 18.9
a 12.9 14.4
b 27.2 24.0
ab 22.4 22.5
c 15.9 14.5
ac 15.1 14.2
bc 41.0 43.0
abc 36.3 39.9

5.6. Blocking in a Factorial Design.

Example. It is suspected that temperature and concentration might affect

the yield of a chemical process. There is enough time to run one replicate of
the experiment in one day, so days represent blocks. The order in which the
experiment is run within each block is randomized. The data are

Day 1 Day 2
Concentration Concentration
Temperature 40 60 80 40 60 80
50 17 16 24 20 21 22
75 12 18 17 9 13 12
100 16 18 25 12 21 23

Statistical Analysis of a Two-Factor Factorial Design with Blocking.

Suppose n replicates of a two-factor factorial design are run in n blocks each
containing a single replicate. The model for this experiment is yijk = µ + τi +
βj + (τ β)ij + δk + εijk , i = 1, . . . , a, j = 1, . . . , b, k = 1, . . . , n where δk is the
effect of the kth block. Notice that the interaction between the block and
the treatments is assumed negligibly small, that is, the terms (τ δ)ik , (βδ)jk ,
and (τ βδ)ijk are assumed equal to zero. The assumptions on the effect terms
∑a ∑b ∑a ∑b
are
∑n i=1 τi = 0, j=1 β j = 0, i=1 (τ β) ij = 0, j=1 (τ β)ij = 0, and
k=1 δk = 0.
The ANOVA table is
Source SS DF F
∑a
2
yi.. 2

i=1 bn − abn a−1

y... M SA
A M SE
∑b y.j.2 2

j=1 an − abn b−1

y... M SB
B M SE
∑a ∑b yij.2 2

j=1 n − abn − SSA − SSB (a − 1)(b − 1)

y... M SAB
AB i=1
∑n y..k2 2 ( MM SE
)
k=1 ab − abn n−1
y... Sblocks
Blocks M SE
Error subtraction (ab − 1)(n − 1)
∑a ∑b ∑n 2

k=1 yijk − abn − 1

2 y...
Total i=1 j=1 abn

21
Example (continued...). y..1 = 163, y..2 = 153, SSblocks = (163)2 /9 +
(153)2 /9 − (316)2 /18 = 5.56.The ANOVA table is

Source SS DF MS F
Temperature 150.11 2 75.06 10.45
Concentration 114.78 2 57.39 7.99
Interaction 40.55 4 10.14 1.41
Days 5.56 1 5.56 (0.77)
Error 57.44 8 7.18
Total 368.44 17

F0.05,2,8 = 4.46, F0.05,4,8 = 3.84, therefore, we conclude that there are main
effects, but no interaction present.

7.2. Blocking a Replicated 2k Design.

Example. In a chemical reaction, the reactant concentration is factor A

run at two levels, and the catalyst is factor B, with two levels. Only four
experiments can be made from a single batch of raw material. Therefore,
three batches (blocks) are used to run three replicates. The data are

Batch
Treatments 1 2 3
(1) 28 25 27
a 36 32 32
b 18 19 23
ab 31 30 29
Total 113 106 111

The sum of squares for the main effects and the interactions are computed
as in the 2k factorial design. The sum of squares for blocks is SSblocks =
∑n Bi2 2

i=1 2k − n2k . The sum of squares for error is calculated by subtraction.

y...

In our example, SSblocks = 6.5. The ANOVA table is

Source SS DF MS F
A 208.33 1 208.33 50.32
B 75.00 1 75.00 18.12
AB 8.33 1 8.33 2.01
Blocks 6.5 2 6.5 (1.57)
Error 24.84 6 4.14
Total 323.00 11

F0.05, 1, 6 = 5.99. There are main effects of A and B but no interaction.

7.3. Confounding in the 2k Factorial Design.

Definition. Sometimes a complete replicate of a 2k factorial design cannot

be run within one block. Then the design has to be run in 2, 4, 8, or, in

22
general, 2p incomplete blocks (p < k). In this case certain treatment effects
(usually high-order interactions) cannot be distinguished from blocks or are
confounded with blocks.

7.4. Confounding in the 2k Factorial Design in Two Blocks.

Example. In a chemical reaction, the reactant concentration is factor A

run at two levels, and the catalyst is factor B, with two levels. Three repli-
cates of the 22 design are to be run. Only two experiments can be made from
a single batch of raw material. Therefore, six batches (incomplete blocks)
are used to run the replicates. The data are

Replicate I Replicate II Replicate III

Batch 1 Batch 2 Batch 3 Batch 4 Batch 5 Batch 6
(Block 1) (Block 2) (Block 1) (Block 2) (Block 1) (Block 2)
(1) = 28 a = 36 (1) = 25 a = 32 (1) = 27 a = 32
ab = 31 b = 18 ab = 30 b = 19 ab = 29 b = 23
∑ Bi2 ȳ...
2
The sum of squares for blocks is computed by SSblocks = 2
− 12
where Bi
1
is the total for the ith block. Consider the main effects A = 2n [a+ab−(1)−b]
1
and B = 2n [b+ab−(1)−a]. The treatment combinations a and b, and (1) and
ab have opposite signs. Since they come from different blocks, the main effects
are unaffected by the blocking. The interaction AB, however, is confounded
with the blocks since in AB = 2n 1
[(1) + ab − a − b], a and b, and (1) and ab
have the same signs. This means that SST = SSblocks + SSA + SSB + SSerror .
The ANOVA table is
Source SS DF MS F
A 208.33 1 208.33 36.76
B 75.00 1 75.00 13.24
Blocks 17.00 5 3.4 (0.60)
AB 8.33 1

replicates 2

Error for ABC (replicates×blocks) 2

Error 22.67 4 5.67

replicates×effects

Total 323.00 11

23
How to assign treatment combinations to the two blocks in the 2k
design with a certain combination being confounded?

There are a number of methods to assign treatments to blocks. One method

is to look at the table of plus and minus signs for the 2k design and assign
the treatment combinations that are plus on the confounded combination to
block 1, for example, and those that are minus on that combination to block
2.
Example. An experimenter wants to run four replicates of 23 design in two
blocks with ABC confounded. He should assign treatment combinations
a, b, c, and abc to block 1, and (1), ab, ac, and bc to block 2.

Another method is to consider a defining contrast L = α1 x1 + α2 x2 + · · · +

αk xk where αi = 0 or 1 depending on whether the ith treatment is present
in the confounded treatment combination. In our example, the defining con-
trast corresponding to ABC is L = x1 + x2 + x3 .
The rule for placing treatment combinations into the two blocks is the follow-
ing. The treatment combinations that produce the same value of L(mod 2)
are placed in the same block. In our example,
(1) → x1 = 0, x2 = 0, x3 = 0 → L = 0 → Block 1
a → x1 = 1, x2 = 0, x3 = 0 → L = 1 → Block 2
b → x1 = 0, x2 = 1, x3 = 0 → L = 1 → Block 2
c → x1 = 0, x2 = 0, x3 = 1 → L = 1 → Block 2
ab → x1 = 1, x2 = 1, x3 = 0 → L = 0 → Block 1
ac → x1 = 1, x2 = 0, x3 = 1 → L = 0 → Block 1
bc → x1 = 0, x2 = 1, x3 = 1 → L = 0 → Block 1
abc → x1 = 1, x2 = 1, x3 = 1 → L = 1 → Block 2

Example. Set up a 24 design in two blocks with ABCD confounded.

Solution: The defining contrast L = x1 + x2 + x3 + x4 . Therefore, treatment
combinations (1), ab, ac, bc, ad, bd, cd, and abcd are assigned to block 1, and
the others to block 2.

7.6. Confounding in the 2k Factorial Design in Four Blocks.

The 2k factorial designs (k ≥ 3) can be run in four blocks with 2k−2 observa-
tions each. In this case one has to specify two effects to be confounded with
blocks.
Example. Construct a 24 factorial design in four blocks with ABC and
BCD confounded.
Solution: The defining contrasts are L1 = x1 + x2 + x3 and L2 = x2 + x3 + x4 .
L1 = 0, L2 = 0 for (1), bc, abd, acd
L1 = 1, L2 = 0 for b, c, ad, abcd
L1 = 0, L2 = 1 for d, ab, ac, bcd
L1 = 1, L2 = 1 for a, bd, cd, abc

Remark. Another effect must be confounded. Since four blocks have three

24
degrees of freedom, three effects should be confounded with blocks. To find
the third confounded effect compute the generalized interaction which is de-
fined as the product of the two effects modulus 2. Since the confounded
effects are constant in a block (equal either 1 or -1), their product is a con-
stant in a block (either 1 or -1), and, thus, is confounded with the block.
Example. Also (ABC)(BCD) = AD must be confounded.

7.7. Confounding in the 2k Factorial Design in 2p Blocks.

In general, the 2k factorial design can be run in 2p (p < k) blocks with

2k−p observations each. Then p effects are confounded with the blocks. The
assignment of the treatments to the blocks can be made by writing the p
defining contrasts L1 , L2 , . . . , Lp associated with the confounded effects. In
addition, 2p − p − 1 other effects will be confounded with blocks. These are
the generalized interactions of the p initially specified effects.

7.8. Partial Confounding.

A confounded effect cannot be estimated, therefore, a partially confounded

design is used. In a partially confounded design, different effects are con-
founded for different replicates of the design. This way all the effects can be
estimated.
Example. For the data considered early, confound A in replicate I, B in
replicate II, and AB in replicate III. The data are

Replicate I Replicate II Replicate III

Batch 1 Batch 2 Batch 3 Batch 4 Batch 5 Batch 6
(Block 1) (Block 2) (Block 1) (Block 2) (Block 1) (Block 2)
(1) = 28 a = 36 (1) = 25 b = 19 (1) = 27 a = 32
b = 18 ab = 31 a = 32 ab = 30 ab = 29 b = 23

The ANOVA table is

Source SS DF
A (from replicates II and III) 105.125 1
B (from replicates I and III) 60.5 1
AB (from replicates I and II) 10.125 1
Blocks 17.00 5
Replicates 2 ...
A (rep.I)+B(rep.II)+AB(rep.III) 3

Error 130.25 3
Total 323.00 11

8.2. The One-Half Fraction of the 2k Design.

Sometimes it is very expensive to run a whole replicate of the 2k design.

25
In this case the fractional factorial design can be run, which is a fraction of
the complete design.
Example. Suppose that experimenters cannot afford all eight runs of a
23 design, but can afford four runs. Therefore, a one-half fraction of the 23
design is run. It is called 23−1 design.
Suppose treatments a, b, c, and abc are chosen for the experiment. From
the table of plus and minus signs, these treatments correspond to the plus
signs in the ABC column. The combination ABC is called the generator
of this fraction. Because the identity I always has the positive sign, the
defining relation for this fraction is I = ABC.

Treatment Factorial Effect

Combination I A B AB C AC BC ABC
a + + - - - - + +
b + - + - - + - +
c + - - + + - - +
abc + + + + + + + +
(1) + - - + - + + -
ab + + + + - - - -
ac + + - - + + - -
bc + - + - + - + -

From the table, the linear combinations used to estimate the main effects of
A, B, and C are
a − b − c + abc −a + b − c + abc
lA = = lBC , lB = = lAC ,
2 2
and
−a − b + c + abc
lC = = lAB .
2
Thus, the effects of A and BC, B and AC, and C and AB are indistin-
guishable. These pairs are called aliases. When estimating A, we are really
estimating A + BC, etc.
To find the alias to a treatment combination, multiply (modulus 2) the defin-
ing relation by this treatment. For example, A · I = A · ABC = BC.
A fractional factorial designs are set up in such a way that by running addi-
tional fractional designs, one gets a complete replicate of a factorial design.
Example. Terms (1), ab, ac, and bc could have been chosen for the one-half
fractional design. The defining relation of this design is I = −ABC. The
aliases are AB · I = −AB · ABC = −C, AC · I = −B, and BC · I = −A.
Thus, by estimating A, for example, we really estimate A − BC.
Notice that combining the two described one-half fractions, we get a com-
plete replicate of the 23 design.
Example. A 23−1 design with defining relation I = ABC produced the

26
following data. Find the important effects.

Treatment Observation
a 3
b 4
c 7
abc 5

(a − b − c + abc)2
SSA+BC = = 9/4 = 2.25, SSB+AC = 1/4 = 0.25,
4
SSC+AB = 25/4 = 6.25.
The main effects of C and possibly A are important.

Example. Set up a 24−1 design with the defining relation I = −ABCD.

Find the aliases for the main effects.
Solution: The treatments used for the design are a, b, c, d, abc, abd, acd, and
bcd since they have minus signs in the ABCD column.

13.1. The Random Effects One-Way ANOVA Model.

Definition. A random effects one-way ANOVA model is the model in which

the a treatments are a random sample from a larger population of treatments.
The population size is assumed infinite.
The statistical model is yij = µ + τi + εij , where εij ’s are i.i.d. N (0, σ 2 ), τi ’s
are i.i.d. N (0, στ2 ), and τi and εij are independent.
The variances σ 2 and στ2 are called the variance components. We are in-
terested in testing for zero treatment effect, that is, we would like to test
H0 : στ2 = 0 against H1 : στ2 > 0.
It can be shown that E(M Str ) = σ 2 + nστ2 , and E(M SE ) = σ 2 . Thus, un-
der the null, M Str and M SE estimate the same thing, and the test statistic
F = M Str /M SE has the F distribution with a − 1 and N − a degrees of
freedom.
The estimators of the variance components are σ̂ 2 = M SE and σ̂τ2 = (M Str −
M SE )/n.
Example. A manufacturer suspects that the batches of raw material fur-
nished by his supplier differ significantly in calcium content. There is a large
number of batches currently in the warehouse. Five of these are randomly
selected for study. Three observations are taken on each batch. The com-
puted sums of squares are SStr = 5.16 and SSE = 1.13.
(a) Is there significant variation in calcium content from batch to batch?
F = 11.42, F0.05, 4, 10 = 3.48
(b) Estimate the variance components.
σ̂ 2 = 0.113, σ̂τ2 = 0.39

13.2. The Two-Factor Factorial with Random Effects.

27
In this model, a levels of factor A and b levels of factor B are randomly
chosen. The model is yij = µ + τi + βj + (τ β)ij + εij , where τi , βj , (τ β)ij , and
εij are independent normal random variables with mean zero and variances,
respectively, στ2 , σβ2 , στ2β , and σ 2 . We are interested in testing H0 : στ2 = 0
against H1 : στ2 > 0; H0 : σβ2 = 0 against H1 : σβ2 > 0; and H0 : στ2β = 0
against H1 : στ2β > 0.
It can be shown that E(M SA ) = σ 2 + nστ2β + bnστ2 , E(M SB ) = σ 2 + nστ2β +
anσβ2 , E(M SAB ) = σ 2 + nστ2β , and E(M SE ) = σ 2 .
Thus, the test statistic for H0 : στ2β = 0 is F = M SAB /M SE which has the
F distribution with (a − 1)(b − 1) and ab(n − 1) degrees of freedom.
The statistic for H0 : στ2 = 0 is F = M SA /M SAB which has the F distribu-
tion with a − 1 and (a − 1)(b − 1) degrees of freedom.
For testing H0 : σβ2 = 0 the test statistic is F = M SB /M SAB which has the
F distribution with b − 1 and (a − 1)(b − 1) degrees of freedom.
The estimators of the variance components are σ̂ 2 = M SE , σ̂τ2β = (M SAB −
M SE )/n, σ̂τ2 = (M SA − M SAB )/bn, and σ̂β2 = (M SB − M SAB )/an.
Example. It is suspected that the strength of a fabric depends on looms
(factor A) and operators’ experience (factor B). Five looms and four opera-
tors are randomly chosen from a large number of looms and operators in a
factory. Every operator works on every loom producing three samples of fab-
ric. The data yield SSA = 23.6, SSB = 17.8, SSAB = 5.6, and SSE = 12.0.
(a) Analyze the data.
M SAB /M SE = 1.56, M SA /M SAB = 12.64, M SB /M SAB = 12.71, a =
5, b = 4, n = 3, F0.05, 12, 40 = 2.00, F0.05, 4, 12 = 3.26, F0.05, 3, 12 = 3.49
(b) Estimate the variance components.
σ̂ 2 = 0.3, σ̂τ2 = 0.45, σ̂β2 = 0.36, σ̂τ2β = 0.06

Sometimes the estimate of a variance is negative. There are two ways to

deal with a negative estimate of a variance component. The first one is to
assume that the variance is equal to zero and leave the other estimates un-
changed. The second method is to assume the variance is zero and recompute
the other estimates using the reduced model.
Example. Suppose σ̂τ2β = −0.14. The reduced model is yijk = µ + τi +
βj + εijk , E(M SA ) = σ 2 + bnστ2 , E(M SB ) = σ 2 + anσβ2 , σ̂τ2 = (M SA −
M SE )/bn, σ̂β2 = (M SB − M SE )/an.

13.3. The Two-Factor Mixed-Effects Model.

Definition. A two-factor mixed-effects model has factor A fixed and fac-

tor B random. The interaction term is, thus, also random.
The statistical model ∑ is yijk = µ + τi + βj + (τ β)ij + εijk where
(1) τi ’s are fixed and ai=1 τi = 0;
(2) βj ’s are i.i.d. N (0, σβ2 );
∑
(3) (τ β)ij ’s are N (0, a−1 a
στ2β ) and ai=1 (τ β)ij = 0;
(4) Cov((τ β)ij , (τ β)i′ j ) = − a1 στ2β , i ̸= i′ and Cov((τ β)ij , (τ β)ij ′ ) = 0, j ̸= j ′ ;
(5) βj ’s, (τ β)ij ’s and εijk ’s are independent;

28
(6) εijk ’s are N (0, σ 2 ).
∑a
bn τ2
It can be shown that E(M SA ) = σ 2 + nστ2β + a−1 i=1 i
, E(M SB ) = σ 2 + anσβ2 ,
E(M SAB ) = σ + nστ β , and E(M SE ) = σ .
2 2 2

The statistic for H0 : τi = 0 for all i is F = M SA /M SAB which has the F

distribution with a − 1 and (a − 1)(b − 1) degrees of freedom.
For testing H0 : σβ2 = 0 the test statistic is F = M SB /M SE which has the
F distribution with b − 1 and ab(n − 1) degrees of freedom.
Thus, the test statistic for H0 : στ2β = 0 is F = M SAB /M SE which has the
F distribution with (a − 1)(b − 1) and ab(n − 1) degrees of freedom.
The estimators of the fixed factor effects are µ̂ = ȳ... and τ̂i = ȳi.. − ȳ... . The
variance components are estimated by σ̂ 2 = M SE , σ̂β2 = (M SB − M SE )/an
and σ̂τ2β = (M SAB − M SE )/n.
Example. Experimenters wish to know whether there is a difference in per-
formance of a certain heart valve by center and by size. Three centers are
randomly chosen from a long list of sites where the valve is implanted. The
five valve sizes are 19mm, 21mm, 23mm, 25mm, and 27mm. Two patients for
each size are implanted at each center and their survival times are recorded.
The data are summarized as SSA = 2.13, SSB = 4.12, SSAB = 9.87, SSE =
7.45, y... = 294, y1.. = 56, y2.. = 88, y3.. = 72, y4.. = 34, y5.. = 44.
(a) Analyze the data.
M SA = 0.53, M SB = 2.06, M SAB = 1.23, M SE = 0.5, M SA /M SAB =
0.43, M SB /M SE = 4.12, M SAB /M SE = 2.46, a = 5, b = 3, n =
2, F0.05, 4, 8 = 3.84, F0.05, 2, 15 = 3.68, F0.05, 8, 15 = 2.64
(b) Estimate the parameters of the model.
µ̂ = 9.8, τ̂1 = −0.47, τ̂2 = 4.87, τ̂3 = 2.2, τ̂4 = −4.13, τ̂5 = −2.47, σ̂ 2 =
0.5, σ̂β2 = 0.156, σ̂τ2β = 0.365

14.1. The Two-Stage Nested Design.

Example. A study is performed to compare the reliability of several models

of cars. Two different car models from each of three domestic car manufac-
turers are tested. Two different cars of each make and model are subjected
to a reliability test. The scores are recorded. The data are

Make 1 Make 2 Make 3

Model 1 Model 2 Model 1 Model 2 Model 1 Model 2
62 77 72 58 94 85
67 73 69 57 90 88

This is a nested design.

Definition. A two-stage nested design (or hierarchical design) has two fac-
tors A and B, with the levels of factor B being similar but not identical for
different levels of factor A. Factor B is said to be nested under the levels of
factor A.
In our example, the factor “Model” is nested under the factor “Make.” The
two Models are not the same for different Makes, and could be potentially

29
numbered 1 through 6.

Statistical Analysis.

The model is yijk = µ+τi +βj(i) +ε(ij)k , i = 1, . . . , a, j = 1, . . . , b, k = 1, . . . , n

where βj(i) is the effect of the jth level of B nested under the ith level of A.
The ANOVA table is
Source SS DF
∑a 2

i=1 yi.. − abn a−1

1 2 y...
A
∑a ∑
bn ∑a
j=1 yij. − bn a(b − 1)
1 b 2 1 2
B within A n i=1 i=1 yi..
Error by subtraction ab(n − 1)
Total SST abn − 1

Here are the expected values of the mean squares depending on whether
factors are fixed or random.
A Fixed A Fixed A Random
E(M S) B Fixed
∑
B Random∑ B Random
bn τi2 bn τ 2
E(M SA ) 2
σ + ∑a−1
∑
σ + nσβ2 + a−1 i
2
σ + nσβ2 + bnστ2
2

n 2
βj(i)
E(M SB(A) ) σ2 + a(b−1)
σ 2 + nσβ2σ 2 + nσβ2
E(M SE ) σ 2
σ2 σ2
∑a
Case 1. A fixed, B fixed. The model constraints are i=1 τi = 0 and
∑b
j=1 βj(i) = 0, i = 1, . . . , a. The hypothesis H0 : τi = 0 is tested by
M SA /M SE , and H0 : βj(i) = 0 is tested by M SB(A) /M SE . ∑
Case 2. A fixed, B random. The model assumptions are ai=1 τi = 0 and
βj(i) are i.i.d. N (0, σβ2 ). Test statistic for H0 : τi = 0 is M SA /M SB(A) and
that for H0 : σβ2 = 0 is M SB(A) /M SE .
Case 3. A random, B random. Model assumptions are τi are i.i.d. N (0, στ2 )
and βj(i) are i.i.d. N (0, σβ2 ). Test statistic for H0 : στ2 = 0 is M SA /M SB(A)
and that for H0 : σβ2 = 0 is M SB(A) /M SE .

Example. In our example, both A and B are fixed. The ANOVA table
is
Source SS DF MS F
Make 1401.17 2 700.58 110.68
Model (within Make) 309.5 3 103.17 16.29
Error 38 6 6.33
Total 1748.67 11

15.3. The Analysis of Covariance (ANCOVA).

Definition. ANCOVA is an experimental design in the presence of an inde-

pendent variable (called covariate).

30
Statistical Model The model is yij = µ + τi + β(xij − x̄.. ) + εij , where
iid ∑
i = 1, . . . , a, j = 1, . . . , n, εij ∼ N (0, σ 2 ), and ai=1 τi = 0. To obtain the
least-squares estimates of the parameters µ, β, and τi , minimize the sum of
squares
∑ a ∑ n
( )2
yij − µ̂ − τ̂i − β̂(xij − x̄.. )
i=1 j=1
∑
under the assumption that ai=1 τ̂i = 0. Solving the system of normal equa-
tions for µ̂ and τ̂i , arrive at the estimates

µ̂ = ȳ.. , and τ̂i = ȳi. − ȳ.. − β̂(x̄i. − x̄.. ),

Substituting τ̂i into the normal equation for β̂ yields

∑
a ∑
n
[ ] ∑
a ∑
n
[ ]
β̂ (xij −x̄.. ) − (x̄i. −x̄.. )(xij −x̄.. )
2
= yij (xij −x̄.. ) − ȳi. (xij −x̄.. )
i=1 j=1 i=1 j=1

or
Sxy − Txy Exy
β̂ = =
Sxx − Txx Exx
where
∑
a ∑
n
Sxx = (xij − x̄.. )2 ,
i=1 j=1

∑
a ∑
n ∑
a ∑
n
Sxy = (xij − x̄.. )(yij − ȳ.. ) = yij (xij − x̄.. ),
i=1 j=1 i=1 j=1

∑
a ∑
n ∑
a
Txx = (x̄i. − x̄.. )(xij − x̄.. ) = n (x̄i. − x̄.. )2 ,
i=1 j=1 i=1

∑
a ∑
n ∑
a ∑
n ∑
a
Txy = (x̄i. −x̄.. )(yij −ȳ.. ) = ȳi. (xij −x̄.. ) = n (x̄i. −x̄.. )(ȳi. −ȳ.. ).
i=1 j=1 i=1 j=1 i=1

31
The ANCOVA table.

∑
a ∑
n
y..2 ∑a
yi.2 y2
SST = yij2 − , SStr = − .. ,
i=1 j=1
an i=1
n an
SSE = SST − SStr .
The sums of squares adjusted to the presence of the covariate are
2
Sxy
SST (adj) = SST − , dfT (adj) = an − 2,
Sxx
2
Exy
SSE(adj) = SSE − , dfE(adj) = a(n − 1) − 1,
Exx
SStr(adj) = SST (adj) − SSE(adj) , dftr(adj) = a − 1.
The test statistic for testing H0 : τ1 = τ2 = · · · = τa = 0 againts H1 : τi ̸= 0
for some i is
SStr(adj) /(a − 1) ( )
F = ∼ F a − 1, a(n − 1) − 1) .
SSE(adj) /(a(n − 1) − 1)

To test the hypotheses H0 : β = 0 against H1 : β ̸= 0, compute

2
Exy
SSslope = , dfslope = 1,
Exx
SSslope /1
F = .
SSE(adj) /(a(n − 1) − 1)
The estimator of σ 2 is σ̂ 2 = M SE(adj) .

Example. Experimental units = 12 steel brackets, treatments = 3 shops

where steel brackets are zinc plated, response y= thickness of zinc plating,
covariate x= thickness of steel bracket before plating. The data are

Shop
1 2 3
x y x y x y
110 40 60 25 62 27
75 38 75 32 90 24
93 30 38 13 45 20
98 47 140 35 59 13

The model is yij = µ + τi + β(xij − x̄.. ) + εij , where i = 1, . . . , 3, j =

iid ∑a
1, . . . , 4, εij ∼ N (0, σ 2 ), and i=1 τi = 0. Compute x̄1. = 94, ȳ1. =
38.75, x̄2. = 78.25, ȳ2. = 26.25, x̄3. = 64, ȳ3. = 21, x̄.. = 78.75, ȳ.. =
28.6667 = µ̂.

32
Sxx = 9278.25, Sxy = 2351, Txx = 1801.5, Txy = 1072.25, Exy = 1278.75, Exx =
7476.75, β̂ = 0.17, τ̂1 = 7.48, τ̂2 = −2.33, τ̂3 = −5.14. SST = 1208.67, SStr =
665.17, SSE = 543.5, SST (adj) = 612.95, SSE(adj) = 324.80, SStr(adj) =
288.15, SSslope = 218.70, σ̂ 2 = 40.60.
The ANCOVA table is
Source SS DF MS F Fcr
Treatment(adj) 288.15 2 144.08 3.55 < 4.46
(Slope) (218.70) (1) (218.70) (5.39) > 5.32
Error(adj) 324.80 8 40.60
Total(adj) 612.95 10

The conclusion of the ANCOVA analysis is that there is no difference be-

tween the shops but the presence of the covariate in the model is significant.

If we ignore the covariate, then the ANOVA table is

Source SS DF MS F Fcr
Treatment 665.17 2 332.58 5.507 > 4.26
Error 543.5 9 60.39
Total 1208.67 11

The conclusion of the ANOVA analysis is that the shops differ significantly.

The Cross-Over Design.

Definition. Cross-over design is a design in which several treatments

are applied sequentially to the same experimental unit. The advantage of
this design is that treatments are compared on the same unit, and therefore,
there is no between-unit variability. The disadvantage is that there might be
a carry-over (or residual) effect: a treatment may have an effect beyond the
period during which it is applied.
As an example, we will consider the simplest case of a cross-over design.

SIMPLE REVERSAL (AB/BA) CROSS-OVER DESIGN

In this design, half of the experimental units receive treatment A, then, after
a wash-out period, treatment B; the other half of the units receive treatment
B first, and then, after the wash-out period, treatment A.

The collected data are of the form

33
 Period Treatment Group 1 Total
1 A y111 y121 . . . y1n1 1 y1·1
2 B y112 y122 . . . y1n1 2 y1·2
Difference D11 D12 . . . D1n1 G1

Period Treatment Group 2 Total

1 B y211 y221 . . . y2n2 1 y2·1
2 A y212 y222 . . . y2n2 2 y2·2
Difference D21 D22 . . . D2n2 G2
Here yijk is an observation on the j-th subject from the i-th group, at period
k, i = 1 or 2, j = 1, . . . , ni , and k = 1 or 2. The difference Dij = yij 1 − yij 2 ,
and the grand differences G1 = y1·1 − y1·2 and G2 = y2·1 − y2·2 .

The model is more conveniently written in the form:

yijkl = µ + αi + βij + γk + τl + εijkl ,

where µ is the overall mean, αi is the i-th group effect, i = 1 or 2, βij is the
iid
effect of the j-th subject from group i, j = 1, . . . , ni , βij ∼ N (0, σβ2 ), γk is
the k-th period effect, k = 1 or 2, τl is the l-th treatment effect, l = 1 or 2
iid
depending on i and k, and εijkl ∼ N (0, σ 2 ). The parameters of this model
are µ, αi , σβ2 , γk , τl and σ 2 . We are interested in testing H0 : τ1 = τ2 = 0
against H1 : τ1 ̸= 0 and τ2 ̸= 0.

The analysis of the data is as follows:

(n2 G1 − n1 G2 )2
SStr = , dftr = 1,
2n1 n2 (n1 + n2 )
1( ∑ ∑ )
n
1 2n
1 1
SSE = 2
D1j + 2
D2j − G21 − G22 , dfE = n1 + n2 − 2,
2 j=1 j=1
2n 1 2n 2

M Str
F = ∼ F (1, n1 + n2 − 2).
M SE
When n1 = n2 = n, the formulas simplify to become

(G1 − G2 )2
SStr = , dftr = 1,
4n
and
1 ∑∑ 2 1 ( 2 )
2 n
SSE = Dij − G1 + G22 , dfE = 2(n − 1).
2 i=1 j=1 2n

Example. Two drugs A or B are applied to subjects and the reaction time
(in seconds) to certain stimulus is measured. There are inherent differences

34
in reaction time among individuals, therefore, a cross-over design is used with
a one-week wash-out period. The data are

Group 1 Subject
Period Drug 1 2 3 4 Total
1 A 30 57 52 66 205
2 B 28 50 38 49 165
Difference 2 7 14 17 40

Group 2 Subject
Period Drug 1 2 3 4 Total
1 B 34 32 18 28 112
2 A 41 21 27 36 125
Difference -7 11 -9 -8 -13

The ANOVA table is

Source SS DF MS F F0.05,1,6
Treatment 175.5625 1 175.5625 5.129 < 5.987
Error 205.375 6 34.229

The conclusion is that there is no statistically significant difference between

the two drugs.

35
STAT 530 SAS CODES AND OUTPUTS

3.3.2. Fixed Effects Model: Statistical Analysis.

SAS Code

title 'The Difference in Fabric Strength for the Three Looms';

data textile;
input loom strength;
cards;
1 88
1 93
1 90
1 89
2 91
2 89
2 92
2 94
3 97
3 96
3 94
3 93
;

proc anova data=textile;

class loom;
model strength = loom;
run;

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

SAS Output
The Difference in Fabric Strength for the Three Looms

The ANOVA Procedure

Dependent Variable: strength

Sum of
Source DF Squares Mean Square F Value Pr > F

Model 2 52.66666667 26.33333333 6.41 0.0186

Error 9 37.00000000 4.11111111

Corrected Total 11 89.66666667

1
 3.5.7. Comparing Pairs of Treatment Means.
SAS Code
title 'The Difference in Fabric Strength for the Three Looms';
data textile;
input loom strength;
cards;
1 88
1 93
1 90
1 89
2 91
2 89
2 92
2 94
3 97
3 96
3 94
3 93
;

proc anova data=textile;

class loom;
model strength = loom;
means loom / lsd clm cldiff;
run;

proc anova data=textile;

class loom;
model strength = loom;
means loom / tukey lsd duncan;
run;
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

SAS Output

95% Confidence
loom N Mean Limits

3 4 95.000 92.707 97.293

2 4 91.500 89.207 93.793
1 4 90.000 87.707 92.293

2
 The Difference in Fabric Strength for the Three Looms

Difference
loom Between 95% Confidence
Comparison Means Limits

3 - 2 3.500 0.257 6.743 ***

3 - 1 5.000 1.757 8.243 ***
2 - 3 -3.500 -6.743 -0.257 ***
2 - 1 1.500 -1.743 4.743
1 - 3 -5.000 -8.243 -1.757 ***
1 - 2 -1.500 -4.743 1.743

Tukey's Studentized Range (HSD) Test for strength

Alpha 0.05
Error Degrees of Freedom 9
Error Mean Square 4.111111
Critical Value of Studentized Range 3.94850
Minimum Significant Difference 4.003

Means with the same letter are not significantly different.

Tukey Grouping Mean N loom

A 95.000 4 3
A
B A 91.500 4 2
B
B 90.000 4 1

t Tests (LSD) for strength

Alpha 0.05
Error Degrees of Freedom 9
Error Mean Square 4.111111
Critical Value of t 2.26216
Least Significant Difference 3.2433

Means with the same letter are not significantly different.

t Grouping Mean N loom

A 95.000 4 3

B 91.500 4 2
B
B 90.000 4 1

3
 Duncan's Multiple Range Test for strength

Alpha 0.05
Error Degrees of Freedom 9
Error Mean Square 4.111111

Number of Means 2 3
Critical Range 3.243 3.385

Means with the same letter are not significantly different.

Duncan Grouping Mean N loom

A 95.000 4 3

B 91.500 4 2
B
B 90.000 4 1

4
 3.5.5. Orthogonal Contrasts.
SAS Code

title 'Orthogonal contrasts';

data textile;
input loom strength @@;
datalines;
1 88 1 93 1 90 1 89
2 91 2 89 2 92 2 94
3 97 3 96 3 94 3 93
;

proc glm data=textile;

class loom;
model strength=loom;
means loom;
contrast ‘Contrast 1’ loom 1 0 -1;
contrast ‘Contrast 2’ loom 1 -2 1;
run;

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

SAS Output

Sum of
Source DF Squares Mean Square F Value Pr > F

Model 2 52.66666667 26.33333333 6.41 0.0186

Error 9 37.00000000 4.11111111

Corrected Total 11 89.66666667

Contrast 1 1 50.00000000 50.00000000 12.16 0.0069

Contrast 2 1 2.66666667 2.66666667 0.65 0.4414

5
 3.4. Model Adequacy Checking.
SAS Code

title 'The Difference in Fabric Strength for the Three Looms';

data textile;
input loom strength;
cards;
1 88
1 93
1 90
1 89
2 91
2 89
2 92
2 94
3 97
3 96
3 94
3 93
;

proc glm data=textile;

class loom;
model strength=loom;
means loom/ hovtest=bartlett;
output out=new r=residuals p=pstrength;
run;

proc capability normaltest data=new;

var residuals;
histogram residuals/normal;
qqplot residuals/normal (mu=est sigma=est);
run;

proc gplot data=new;

plot residuals*pstrength/vref=0;
run;

SAS Output

6
7
 Goodness-of-Fit Tests for Normal Distribution
Test Statistic DF p Value
Kolmogorov-Smirnov D 0.12387579 Pr > D >0.150
Cramer-von Mises W-Sq 0.03377973 Pr > W-Sq >0.250
Anderson-Darling A-Sq 0.24883072 Pr > A-Sq >0.250
Chi-Square Chi-Sq 1.73172167 2 Pr > Chi-Sq 0.421

8
 Bartlett's Test for Homogeneity
of strength Variance
Source DF Chi-Square Pr > ChiSq
loom 2 0.0792 0.9612

9
 4.1. The Randomized Complete Block Design.
SAS Code

data RCBD;
input loom operator strength @@;
datalines;
1 1 88 1 2 93 1 3 90
2 1 90 2 2 92 2 3 92
3 1 91 3 2 96 3 3 94
;

proc anova;
class loom operator;
model strength = loom operator;
run;

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

SAS Output

Source DF Squares Mean Square F Value Pr > F

Model 4 41.77777778 10.44444444 11.06 0.0195

Error 4 3.77777778 0.94444444

Corrected Total 8 45.55555556

R-Square Coeff Var Root MSE strength Mean

0.917073 1.058890 0.971825 91.77778

Source DF Anova SS Mean Square F Value Pr > F

loom 2 17.55555556 8.77777778 9.29 0.0314

operator 2 24.22222222 12.11111111 12.82 0.0182

10
 4.2. The Latin Square Design.
SAS Code
data LATIN_SQUARE;
input batch operator formulation $ force @@;
datalines;
1 1 A -1 1 2 B -5 1 3 C -6 1 4 D -1 1 5 E -1
2 1 B -8 2 2 C -1 2 3 D 5 2 4 E 2 2 5 A 11
3 1 C -7 3 2 D 13 3 3 E 1 3 4 A 2 3 5 B -4
4 1 D 1 4 2 E 6 4 3 A 1 4 4 B -2 4 5 C -3
5 1 E -3 5 2 A 5 5 3 B -5 5 4 C 4 5 5 D 6
;

proc anova;
class batch operator formulation;
model force = batch operator formulation;
run;

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

SAS Output
Source DF Squares Mean Square F Value Pr > F

Model 12 548.0000000 45.6666667 4.28 0.0089

Error 12 128.0000000 10.6666667

Corrected Total 24 676.0000000

R-Square Coeff Var Root MSE force Mean

0.810651 816.4966 3.265986 0.400000

Source DF Anova SS Mean Square F Value Pr > F

batch 4 68.0000000 17.0000000 1.59 0.2391

operator 4 150.0000000 37.5000000 3.52 0.0404
formulation 4 330.0000000 82.5000000 7.73 0.0025

11
 4.3. The Græco-Latin Square Design.
SAS Code
data GRAECO_LATIN;
input batch operator formulation $ assembly $ force @@;
datalines;
1 1 A a -1 1 2 B c -5 1 3 C e -6 1 4 D b -1 1 5 E d -1
2 1 B b -8 2 2 C d -1 2 3 D a 5 2 4 E c 2 2 5 A e 11
3 1 C c -7 3 2 D e 13 3 3 E b 1 3 4 A d 2 3 5 B a -4
4 1 D d 1 4 2 E a 6 4 3 A c 1 4 4 B e -2 4 5 C b -3
5 1 E e -3 5 2 A b 5 5 3 B d -5 5 4 C a 4 5 5 D c 6
;

proc anova;
class batch operator formulation assembly;
model force = batch operator formulation assembly;
run;

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

SAS Output
Source DF Squares Mean Square F Value Pr > F

Model 16 610.0000000 38.1250000 4.62 0.0171

Error 8 66.0000000 8.2500000

Corrected Total 24 676.0000000

R-Square Coeff Var Root MSE force Mean

0.902367 718.0703 2.872281 0.400000

Source DF Anova SS Mean Square F Value Pr > F

batch 4 68.0000000 17.0000000 2.06 0.1783

operator 4 150.0000000 37.5000000 4.55 0.0329
formulation 4 330.0000000 82.5000000 10.00 0.0033
assembly 4 62.0000000 15.5000000 1.88 0.2076

12
 4.4. The Balanced Incomplete Block Design.
SAS Code
data BIBD;
input catalyst batch time @@;
datalines;
1 1 73 1 2 74 1 4 71
2 2 75 2 3 67 2 4 72
3 1 73 3 2 75 3 3 68
4 1 75 4 3 72 4 4 75
;

proc glm;
class catalyst batch;
model time=batch catalyst;
run;

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

SAS Output
Source DF Squares Mean Square F Value Pr > F

Model 6 77.75000000 12.95833333 19.94 0.0024

Error 5 3.25000000 0.65000000

Corrected Total 11 81.00000000

R-Square Coeff Var Root MSE time Mean

0.959877 1.112036 0.806226 72.50000

Source DF Type I SS Mean Square F Value Pr > F

batch 3 55.00000000 18.33333333 28.21 0.0015

catalyst 3 22.75000000 7.58333333 11.67 0.0107

Source DF Type III SS Mean Square F Value Pr > F

batch 3 66.08333333 22.02777778 33.89 0.0010

catalyst 3 22.75000000 7.58333333 11.67 0.0107

13
 5.3.2. The Two-Factor Factorial Design: Fixed-Effects Model.
SAS Code

data factorial_design;
input temp conc yield @@;
datalines;
1 1 17 1 1 20 1 2 16 1 2 21 1 3 24 1 3 22
2 1 12 2 1 9 2 2 18 2 2 13 2 3 17 2 3 12
3 1 16 3 1 12 3 2 18 3 2 21 3 3 25 3 3 23
;

proc glm;
class temp conc;
model yield=temp conc temp*conc;
lsmeans temp conc/out=maineff;
lsmeans temp*conc/out=intereff;
run;

symbol interpol=join value=x;

proc gplot data=maineff;

plot lsmean*temp;
plot lsmean*conc;
run;

proc gplot data=intereff;

plot lsmean*conc=temp;
run;

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

SAS Output
Error 9 63.000 63.000 7.000
Total 17 368.444

Source DF Squares Mean Square F Value Pr > F

Model 8 305.4444444 38.1805556 5.45 0.0101

Error 9 63.0000000 7.0000000

Corrected Total 17 368.4444444

R-Square Coeff Var Root MSE yield Mean

0.829011 15.07074 2.645751 17.55556

Source DF Type I SS Mean Square F Value Pr > F

temp 2 150.1111111 75.0555556 10.72 0.0042

conc 2 114.7777778 57.3888889 8.20 0.0094
temp*conc 4 40.5555556 10.1388889 1.45 0.2951

14
LSMEAN
21

20

19

18

17

16

15

14

13

1 2 3

t emp

LSMEAN
21

20

19

18

17

16

15

14

13

1 2 3

conc

15
LSMEAN
24

23

22

21

20

19

18

17

16

15

14

13

12

11

10

1 2 3

conc

t emp 1 2 3

16
 6.3. The 23 Factorial Design.
SAS Code

data two_cubed;
input temp conc catalyst yield @@;
datalines;
1 1 1 18.2 1 1 1 18.9 2 1 1 12.9 2 1 1 14.4
1 2 1 27.2 1 2 1 24.0 2 2 1 22.4 2 2 1 22.5
1 1 2 15.9 1 1 2 14.5 2 1 2 15.1 2 1 2 14.2
1 2 2 41.0 1 2 2 43.0 2 2 2 36.3 2 2 2 39.9
;

proc glm;
class temp conc catalyst;
model yield=temp | conc | catalyst;
run;
_______________________________________________________________________

SAS Output
Sum of
Source DF Squares Mean Square F Value Pr >
F

Model 7 1658.110000 236.872857 115.83

<.0001

Error 8 16.360000 2.045000

Corrected Total 15 1674.470000

R-Square Coeff Var Root MSE yield Mean

0.990230 6.014868 1.430035 23.77500

Source DF Type I SS Mean Square F Value Pr > F

temp 1 39.062500 39.062500 19.10 0.0024

conc 1 1092.302500 1092.302500 534.13 <.0001
temp*conc 1 0.640000 0.640000 0.31 0.5912
catalyst 1 220.522500 220.522500 107.83 <.0001
temp*catalyst 1 3.240000 3.240000 1.58 0.2436
conc*catalyst 1 295.840000 295.840000 144.67 <.0001
temp*conc*catalyst 1 6.502500 6.502500 3.18 0.1124

17
 5.6. Blocking in a Factorial Design.
SAS Code

data one;
input temp day conc yield @@;
datalines;
1 1 1 17 1 1 2 16 1 1 3 24 1 2 1 20 1 2 2 21 1 2 3 22
2 1 1 12 2 1 2 18 2 1 3 17 2 2 1 9 2 2 2 13 2 2 3 12
3 1 1 16 3 1 2 18 3 1 3 25 3 2 1 12 3 2 2 21 3 2 3 23
;

proc glm;
class temp day conc;
model yield=temp conc temp*conc day;
run;
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

SAS Output

Sum of
Source DF Squares Mean Square F Value Pr > F

Model 9 311.0000000 34.5555556 4.81 0.0187

Error 8 57.4444444 7.1805556

Corrected Total 17 368.4444444

R-Square Coeff Var Root MSE yield Mean

0.844089 15.26386 2.679656 17.55556

Source DF Type I SS Mean Square F Value Pr > F

temp 2 150.1111111 75.0555556 10.45 0.0059

conc 2 114.7777778 57.3888889 7.99 0.0124

temp*conc 4 40.5555556 10.1388889 1.41 0.3135

day 1 5.5555556 5.5555556 0.77 0.4047

18
 7.4. Confounding in the 2k Design in Two Blocks.
SAS Code

data one;
input temp conc batch yield @@;
datalines;
1 1 1 28 1 1 3 25 1 1 5 27
2 1 2 36 2 1 4 32 2 1 6 32
2 2 1 31 2 2 3 30 2 2 5 29
1 2 2 18 1 2 4 19 1 2 6 23
;

proc glm;
class temp conc batch;
model yield = temp conc batch;
run;

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

SAS Output

Sum of
Source DF Squares Mean Square F Value Pr > F

Model 7 300.3333333 42.9047619 7.57 0.0344

Error 4 22.6666667 5.6666667

Corrected Total 11 323.0000000

R-Square Coeff Var Root MSE yield Mean

0.929825 8.656277 2.380476 27.50000

Source DF Type I SS Mean Square F Value Pr > F

temp 1 208.3333333 208.3333333 36.76 0.0037

conc 1 75.0000000 75.0000000 13.24 0.0220
batch 5 17.0000000 3.4000000 0.60 0.7080

19
 13.2. The Two-Way ANOVA Design with Random Effects.
SAS Code

data random_effects;
input temp conc yield @@;
datalines;
1 1 17 1 1 20 1 2 16 1 2 21 1 3 24 1 3 22
2 1 12 2 1 9 2 2 18 2 2 13 2 3 17 2 3 12
3 1 16 3 1 12 3 2 18 3 2 21 3 3 25 3 3 23
;

proc glm;
class temp conc;
model yield=temp conc temp*conc;
random temp conc temp*conc/test;
run;

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

SAS Output

The GLM Procedure

Tests of Hypotheses for Random Model Analysis of Variance

Dependent Variable: yield

Source DF Type III SS Mean Square F Value Pr >

F

temp 2 150.111111 75.055556 7.40

0.0452
conc 2 114.777778 57.388889 5.66
0.0682

Error: MS(temp*conc) 4 40.555556 10.138889

Source DF Type III SS Mean Square F Value Pr >

F

temp*conc 4 40.555556 10.138889 1.45

0.2951

Error: MS(Error) 9 63.000000 7.000000

20
 13.3. The Two-Way ANOVA Design with Mixed Effects.
SAS Code

data mixed_effects;
input temp conc yield @@;
datalines;
1 1 17 1 1 20 1 2 16 1 2 21 1 3 24 1 3 22
2 1 12 2 1 9 2 2 18 2 2 13 2 3 17 2 3 12
3 1 16 3 1 12 3 2 18 3 2 21 3 3 25 3 3 23
;

proc glm;
class temp conc;
model yield=temp conc temp*conc;
random conc temp*conc / test;

run;

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

SAS Output

The GLM Procedure

Tests of Hypotheses for Mixed Model Analysis of Variance

Dependent Variable: yield

Source DF Type III SS Mean Square F Value Pr > F

temp 2 150.111111 75.055556 7.40 0.0452

conc 2 114.777778 57.388889 5.66 0.0682

Error: MS(temp*conc) 4 40.555556 10.138889

Source DF Type III SS Mean Square F Value Pr > F

temp*conc 4 40.555556 10.138889 1.45 0.2951

Error: MS(Error) 9 63.000000 7.000000

21
 14.1. The Two-Stage Nested Design.
SAS Code
data NESTED;
input make model score @@;
datalines;
1 1 62 1 1 67 1 2 77 1 2 73 2 1 72 2 1 69
2 2 58 2 2 57 3 1 94 3 1 90 3 2 85 3 2 88
;
proc sort;
by make model;

proc nested;
class make model;
var score;
run;

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

SAS Output

The NESTED Procedure

Coefficients of Expected Mean Squares

Source make model Error

make 4 2 1
model 0 2 1
Error 0 0 1

Nested Random Effects Analysis of Variance for Variable score

Variance Sum of Error Variance

Percent
Source DF Squares F Value Pr > F Term Mean Square Component of
Total

Total 11 1748.666667 158.969697 204.104167

100.0000
make 2 1401.166667 6.79 0.0770 model 700.583333 149.354167
73.1755
model 3 309.500000 16.29 0.0027 Error 103.166667 48.416667
23.7215
Error 6 38.000000 6.333333 6.333333
3.1030

score Mean 74.33333333

Standard Error of score Mean 7.64080784

22
 15.3. Analysis of Covariance (ANCOVA).
SAS Code

data ancova;
input shop x y;
xcentered=x-78.75;
cards;
1 110 40
1 75 38
1 93 30
1 98 47
2 60 25
2 75 32
2 38 13
2 140 35
3 62 27
3 90 24
3 45 20
3 59 13
;

proc glm;
class shop;
model y= shop xcentered;
run;
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

SAS Output

Sum of
Source DF Squares Mean Square F Value Pr > F

Model 3 883.871527 294.623842 7.26 0.0114

Error 8 324.795140 40.599392

Corrected Total 11 1208.666667

R-Square Coeff Var Root MSE y Mean

0.731278 22.22709 6.371765 28.66667

Source DF Type I SS Mean Square F Value Pr > F

shop 2 665.1666667 332.5833333 8.19 0.0116

xcentered 1 218.7048601 218.7048601 5.39 0.0488

Source DF Type III SS Mean Square F Value Pr > F

shop 2 288.1556320 144.0778160 3.55 0.0788

xcentered 1 218.7048601 218.7048601 5.39 0.0488

23
 The Cross-Over Design.
SAS Code
data reaction_time;
input group subject period drug $ time;
cards;
1 1 1 A 30
1 2 1 A 57
1 3 1 A 52
1 4 1 A 66
1 1 2 B 28
1 2 2 B 50
1 3 2 B 38
1 4 2 B 49
2 1 1 B 34
2 2 1 B 32
2 3 1 B 18
2 4 1 B 28
2 1 2 A 41
2 2 2 A 21
2 3 2 A 27
2 4 2 A 36
;

proc glm data=reaction_time;

class group subject period drug;
model time=group subject(group) period drug/SS3; /* SS3 = Type III SS
*/
random subject(group);
run;
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
- - -
SAS Output
The GLM Procedure

Sum of
Source DF Squares Mean Square F Value Pr >
F

Model 9 2539.562500 282.173611 8.24

0.0092

Error 6 205.375000 34.229167

Corrected Total 15 2744.937500

R-Square Coeff Var Root MSE time Mean

0.925180 15.42160 5.850570 37.93750

Source DF Type III SS Mean Square F Value Pr > F

group 1 1105.562500 1105.562500 32.30 0.0013

subject(group) 6 1212.875000 202.145833 5.91 0.0242
period 1 45.562500 45.562500 1.33 0.2925
drug 1 175.562500 175.562500 5.13 0.0641

24