Analysis of Variance and

Design
g of Experiments
Experiments-
p -I
MODULE – II
LECTURE - 10

GENERAL LINEAR HYPOTHESIS

AND ANALYSIS OF VARIANCE
Dr. Shalabh
Department of Mathematics and Statistics
Indian Institute of Technology Kanpur
 2

3. Duncan’s multiple comparison test

The test procedure in Duncan’s multiple comparison test is the same as in the Student-Newman-Keuls test except the
observed ranges are compared with Duncan’s 100α % critical range

s2
D p = qα p , p ,γ
*

where α p = 1 − (1 − α ) p −1 , qα* p , p,γ denotes the upper 100α % points of the Studentized range based on Duncan’s range.

Tables for Duncan’s range are available.

Duncan feels that this test is better than the Student-Newman-Keuls test for comparing the differences between any
two ranked means. Duncan regarded that the Student-Newman-Keuls method is too stringent in the sense that the
true differences between the means will tend to be missed too often. Duncan notes that in testing the equality of a
subset k , (2 ≤ k ≤ p) means through null hypothesis, we are in fact testing whether (p - 1) orthogonal contrasts between
the β ' s differ from zero or not. If these contrasts were tested in separate independent experiments, each at level α ,
th probability
the b bilit off incorrectly
i tl rejecting
j ti theth nullll hypothesis
h th i would b ⎡⎣1 − (1 − α ) pp−1 ⎤⎦ . So
ld be S Duncan
D proposed
d to
t use
⎡⎣1 − (1 − α ) p −1 ⎤⎦ in place of α in the Student-Newman-Keuls test.
[Reference: Contributions to order statistics, Wiley 1962, Chapter 9 (Multiple decision and multiple comparisons, H.A.
David, pages 147-148)].
.
 3

Case of unequal sample sizes

When sample means are not based on the same number of observations, the procedures based on Studentized range,
Student-Newman-Keuls test and Duncan’s test are not applicable. Kramer proposed that in Duncan’s method, if a set

of p means is to be tested for equality, then replace

s 1 ⎛ 1 1 ⎞
qα* * , p ,γ *
by qα * , p ,γ s ⎜ + ⎟
p
n p
2 ⎝ nU nL ⎠

where nU and nL are the number of observations corresponding to the largest and smallest means in the data
data. This
procedure is only an approximate procedure but will tend to be conservative, since means based on a small number of
observations will tend to be overrepresented in the extreme groups of means.

p
Another option is to replace n by the harmonic mean of n1 , n 2 ,..., n p , i.e., .
p
⎛1 ⎞
∑ ⎜
i =1 ⎝ ni
⎟
⎠
 4

The “Least Significant Difference” (LSD)

In the usual testing of H 0 : β i = β k against H1 : βi ≠ β k , the t-statistic

yio − yko
t=
Vm ( yio − yko )
Var
is used which follows a t-distribution, say with degrees of freedom ‘df’. Thus H 0 is rejected whenever

t >t α
df , 1−
2

and it is concluded that β1 and β2 are significantly different. The inequality t > t α
can be
d f , 1−
2
equivalently written as

yio − yko > t m ( y − y ).

Var
V
α io ko )
df , 1−
2

If every pair of sample for which yio − yko exceeds t m(y − y )

Var
α io ko
df , 1−
2

then this will indicate that the difference between βi and βk is significantly different . So according to this
this, the

quantity t m ( y − y ) would be the least difference of y and y for which it will be declared that the
Var
α io ko io ko
df , 1−
2
difference between βi and βk is significant. Based on this idea, we use the pooled variance of the two sample

Var ( yio − y ko ) as s2 and the Least Significant Difference (LSD) is defined as

⎛1 1⎞
LSD = t α s 2 ⎜ + ⎟.
df , 1−
2 ⎝ ni nk ⎠
 5

If n1 = n2 = n, then

2s 2
LSD = t α .
df , 1− n
2

p ( p − 1)
Now all pairs of yio and yko , (i ≠ k = 1, 2,..., p) are compared with LSD. Use of LSD criterion may not lead to
2
good results if it is used for comparisons suggested by the data (largest/smallest sample mean) or if all pairwise

comparisons are done without correction of the test level. If LSD is used for all the pairwise comparisons then these

tests are not independent. Such correction for test levels was incorporated in Duncan’s test.
 6

Tukey’s “Honestly Significant Difference” (HSD)

In this procedure , the Studentized rank values q α , n , γ are used in place of t-quantiles and the standard error of the
difference of pooled mean is used in place of standard error of mean in common critical difference for testing H 0 : β i = β k
against H 0 : β i ≠ β k and Tukey’s Honestly Significant Difference is computed as

MSerror
HSD = q α
1− , p ,γ n
2

p ( p − 1)
assuming all samples are of the same size n. All pairs yio − yko are compared with HSD. If yio − y ko > HSD
2
then β i and β k are significantly different.
 7

We notice that all the multiple comparison test procedure discussed up to now are based on the testing of hypothesis.
Th
There i one-to-one
is t relationship
l ti hi bbetween
t th
the ttesting
ti off h
hypothesis
th i and
d th
the confidence
fid interval
i t l estimation.
ti ti So
S the
th
confidence interval can also be used for such comparisons. Since H 0 : β i = β k is same as H 0 : βi − β k = 0, so first we
establish the relationship and then describe the Tukey’s and Scheffe’s procedures for multiple comparison test which
are based on the confidence interval. We need the following concepts.

Contrast
p
A linear parametric function L = l ' β = ∑ A i β i where β = ( β1 , β 2 ,..., β P ) and A = (A 1 , A 2 ,..., A p ) are the p × 1 vectors of
i =1
p
parameters and constants respectively is said to be a contrast when ∑A
i=1
i = 0.

For example β1 − β 2 = 0, β1 + β 2 − β 3 − β1 = 0, β1 + 2β 2 − 3β 3 = 0

etc are contrast whereas β1 + β 2 = 0, β1 + β 2 + β3 + β 4 = 0, β1 − 2β 2 − 3β 3 = 0 etc.

etc. etc are not contrasts.
contrasts

Orthogonal contrast
p p p
If L1 = A ' β = ∑ A i β i and L2 = m ' β = ∑ mi β i are contrasts such that
i =1
A ′m = 0 or ∑A m
i =1
i i =0 then L1 and L2
i =1
are called orthogonal contrasts.

For example, L1 = β1 + β 2 − β3 − β 4 and L2 = β1 − β 2 + β 3 − β 4 are contrasts. They are also the orthogonal contrasts.

p p
The condition ∑A m
i =1
i i = 0 ensures that L1 and L2 are independent in the sense that Cov ( L1 , L2 ) = σ 2 ∑ A i mi = 0.
i =1
 8

Mutually orthogonal contrasts

If there
h are more than
h two contrasts then
h they
h are said
id to be
b mutually
ll orthogonal,
h l if they
h are pair-wise
i i orthogonal.
h l
It may be noted that the number of mutually orthogonal contrasts is the number of degrees of freedom.

Coming back to the multiple comparison test, if the null hypothesis of equality of all effects is rejected then it is
reasonable to look for the contrasts which are responsible for the rejection. In terms of contrasts, it is desirable to have
a procedure
i. that permits the selection of the contrasts after the data is available.
ii
ii. with which a known level of significance is associated.
associated

Such procedures are Tukey’s and Scheffe’s procedures. Before discussing these procedure, let us consider the
following example which illustrates the relationship between the testing of hypothesis and confidence intervals.
.
Example Consider the test of hypothesis for
H 0 : βi = β j (i ≠ j = 1, 2,..., p)
or H 0 : βi − β j = 0

or H 0 : contrast = 0
or H 0 : L = 0.

The test statistic for H 0 : β i = β j is

( βˆi − βˆ j ) − ( βi − β j ) Lˆ − L
t= =
m(y − y )
Var m ( Lˆ )
Var
io ko
 9

where βˆ denotes the maximum likelihood (or least squares) estimator of β and t follows a t-distribution with df
degrees
g of freedom. This statistic,, infact,, can be extended to anyy linear contrast,, sayy e.g.,
g,

L = β1 + β 2 − β 3 − β 4 , Lˆ = βˆ1 + βˆ2 − βˆ3 − βˆ4 .

The decision rule is

reject H 0 : L = 0 against H1 : L ≠ 0 if m ( Lˆ ).
Lˆ > tdf V
Var )

The 100(1 − α )% confidence interval of L is obtained as

⎡ Lˆ − L ⎤
P ⎢ −tdf ≤ ≤ tdf ⎥ = 1 − α
⎢ m ( Lˆ ) ⎥
⎣ Var ⎦

or P ⎡ Lˆ − tdf Var m ( Lˆ ) ⎤ = 1 − α
m ( Lˆ ) ≤ L ≤ Lˆ + t Var
⎢⎣ df
⎥⎦

so that the 100(1 − α ) % confidence interval of L is

⎡ Lˆ − t Var m ( Lˆ ) ⎤
m ( Lˆ ), Lˆ + t Var
⎢⎣ df df
⎥⎦

and

m ( Lˆ ) ≤ L ≤ Lˆ + t Var
Lˆ − t df Var m ( Lˆ ).
df
 10

pp confidence limits,, then H0 : L = 0 is accepted.

If this interval includes L= 0 between lower and upper p Our objective
j is
to know if the confidence interval contains zero or not.

Suppose for some given data the confidence intervals for β1 − β 2 and β1 − β 3 are obtained as

−3 ≤ β1 − β 2 ≤ 2 and 2 ≤ β1 − β3 ≤ 4.

Thus we find that the interval of β1 − β2 includes zero which implies that H 0 : β1 − β 2 = 0 is accepted. Thus β1 = β2 . On
the other hand interval of β1 − β3 does not include zero and so H 0 : β1 − β3 = 0 is not accepted. Thus β1 ≠ β3 .

If the interval of β1 − β3 is −1 ≤ β1 − β 3 ≤ 1 then H 0 : β1 = β3 is accepted. If both H 0 : β1 = β 2 and H 0 : β1 = β3 are

accepted then we can conclude that β1 = β 2 = β3 .