Chapter 6:
Comparison of Several Multivariate Means
Paired Comparisons
Measurements are recorded under different sets of conditions
See if the responses differ significantly over these sets
Two or more treatments can be administered to the same or similar
experimental units
Compare responses to assess the effects of the treatments
7/14/2017 22
Example 6.1:
Effluent Data from Two Labs
7/14/2017 3
Single Response (Univariate) Case
D j X j1 X j 2 , j 1, 2, , n
D j : N ( , )
2
d
D
t : t n 1
sd / n
Reject H 0 : 0 in favor of H 1 : 0 if t t n 1 ( / 2)
100(1 )% confidence interval for
sd sd
d t n 1 ( / 2) d t n 1 ( / 2)
n n
7/14/2017 4
Multivariate Extension: Notations
X 1 j1 variable 1 under treatment 1
X 1 j 2 variable 2 under treatment 1
X 1 jp variable p under treatment 1
--------------------------------
X 2 j1 variable 1 under treatment 2
X 2 j 2 variable 2 under treatment 2
X 2 jp variable p under treatment 2
7/14/2017 5
Result 6.1
D j1 X 1 j1 X 2 j1
D j2 X1 j2 X 2 j2
D jp X 1 jp X 2 jp
D j D j1 , D j 2 , , D jp '
D j : N p (, d ), j 1, 2, , n
(n 1) p
T nD ' S D :
2 1
F p ,n p
(n p)
d
1 n
D D j , Sd
1 n
D j D D j D '
n j 1 n 1 j 1
7/14/2017 6
Test of Hypotheses and Confidence Regions
d 'j d j1 , d j 2 , , d jp : observed differences
Reject H 0 : 0 in favor of H 1 : 0 if
(n 1) p
1
T nd ' S d
2
F p ,n p ( )
n p
d
(n 1) p
Confidence regions : d ' S d
1
F p ,n p ( )
n p
d
s d2i di
2
(n 1) p s
i : di F p ,n p ( ) , i : d i t n 1
n p n 2p n
7/14/2017 7
Example 6.1: Check Measurements from Two
Labs
d1 9.36 199.26 88.38
d , Sd
2
d 13.27 88.38 418.61
0.0055 0.0012 9.36
T 11 9.36 13.27
2
13.27
0.0012 0.0026
2 10
13.6 F2,9 (0.05) 9.47
9
Reject H 0 : 0
1 : 9.36 9.47 199.26 / 11 or 22.46, 3.74
2 : 13.27 9.47 418.61 / 11 or - 5.71, 32.25
Both includes zero
7/14/2017 8
Experiment Design for
Paired Comparisons
1 2 3 n
...
...
Treatments Treatments Treatments Treatments
1 and 2 1 and 2 1 and 2 1 and 2
assigned at assigned at assigned at assigned at
random random random random
7/14/2017 9
Alternative View
x' x , x , , x , x
11 12 1p 21 , x 22 , , x 2 p
S11 S12
S
S 21 S 22
1 0 0 | 1 0 0
0 1 0 | 0 1 0
C
( p 2 p ) |
0 0 1 | 0 0 1
d j Cx j , d Cx, S d CSC' , T 2
1
nx' C' CSC' Cx
7/14/2017 10
Repeated Measures Design for Comparing
Measurements
q treatments are compared with respect to a single response variable
Each subject or experimental unit receives each treatment once over
successive periods of time
7/14/2017 11
Example 6.2: Treatments in an Anesthetics
Experiment
19 dogs were initially given the drug pentobarbitol followed by four
treatments
Present 3 4
Halothane
Absent 2 1
Low High
CO2 pressure
7/14/2017 12
Example 6.2: Sleeping-Dog Data
7/14/2017 13
Contrast Matrix
X j1
X
Xj j2
, j 1, 2, , n E ( X j )
X jq
1 2 1 1 0 0 1
1 0 1 0 2
C
1 3
1 q 1 0 0 1 q
7/14/2017 14
Test for Equality of Treatments in a Repeated
Measures Design
X : N q (, ), C : contrast matrix
Test of H 0 : C 0 vs. H 1 : C 0
Reject H 0 if
(n 1)(q 1)
T nCx ' CSC'
1
2
Cx Fq 1,n q 1 ( )
(n q 1)
7/14/2017 15
Example 6.2: Contrast Matrix
3 4 1 2 Halothane contrast
1 3 2 4 CO 2 contrast
1 4 2 3 H CO 2 interaction
1 1 1 1
C 1 1 1 1
1 1 1 1
7/14/2017 16
Example 6.2: Test of Hypotheses
368.21 2819.29
404.63 3568.42 7963.14
x , S
479.26 2943.49 5303.98 6851.32
502.89 2295.35 4065.44 4499.63 4878.99
209.31 9432.32 1098.92 927.62
Cx 60.05, CSC' 1098.92 5195.84 914.54
12.79 927.62 914.54 7557.44
T 2 nCx ' CSC' Cx 116
1
(n 1)(q 1)
Fq 1,n q 1 0.05 10.94
(n q 1)
Reject H 0 : C 0
7/14/2017 17
Example 6.2: Simultaneous Confidence Intervals
Contrast of halothane influence
x3 x4 x1 x2 18(3)
F3,16 (0.05) 209.31 73.70
16
CO 2 pressure influence
5195.84
60.05 10.94 60.5 54.70
19
H - CO 2 " interaction"
7557.44
12.79 10.94 12.79 65.97
19
7/14/2017 18
Assumptions Concerning the Structure of Data
X11 , X12 , , X1n1 : random sample from p variate
population with mean vector 1 and covariance 1
X 21 , X 22 , , X 2 n2 : random sample from p variate
population with mean vector 2 and covariance 2
X11 , X12 , , X1n1 are independent of X 21 , X 22 , , X 2 n2
Further assumptions when n1 and n2 small :
- -Both populations are multivariate normal
1 2
7/14/2017 19
Pooled Estimate of
Population Covariance Matrix
7/14/2017 20
Result 6.2
7/14/2017 21
Proof of Result 6.2
7/14/2017 22
Test of Hypothesis
Reject H 0 : 1 2 0 in favor of H 1 : 1 2 0
1
1
if T x1 x 2 0 ' x1 x 2 0
1
2
S pooled
n1 n2
n1 n2 2 p
F p ,n1 n2 p 1 ( )
n1 n2 p 1
Note E ( X1 X 2 ) 1 2
Cov( X1 X 2 )
Cov( X1 ) Cov( X1 , X 2 ) Cov( X 2 , X1 ) Cov( X 2 )
1 1
Cov( X1 ) Cov( X 2 )
7/14/2017
n1 n2 23
7/14/2017 24
Example 6.3: Comparison of Soaps Manufactured
n n 50
in Two Ways
1 2
8.3 2 1 10.2 2 1
x1 , S 1 , x2 , S2
4.1 1 6 3.9 1 4
49 49 2 1 1.9
S pooled S1 S 2 , x1 x 2
98 98 1 5 0 .2
Eigenvalues and eigenvectors of S pooled :
1 5.303, e1 0.290 0.957'
2 1.697, e1 0.957 0.290'
1 1 n1 n2 2 p
F p ,n1 n2 p 1 (0.05) 0.25
n1 n2 n1 n2 p 1
1 0.25 1.15, 2 0.25 0.65
7/14/2017 25
Example 6.3
7/14/2017 26
Result 6.3: Simultaneous Confidence Intervals
n1 n2 2 p
c
2
F ( )
p , n1 n2 p 1
n1 n2 p 1
1 1
a' ( X1 X 2 ) c a' S pooled a
n1 n2
will cover a' ( 1 2 ) for all a
In particular, 1i 2i will be covered by
1 1
( X 1i X 2i ) c sii , pooled
n1 n2
7/14/2017 27
7/14/2017 28
Example 6.4: Electrical Usage of Homeowners with
and without ACs
204.4 13825.3 23823.4
x1 , S1 , n1 45
556.6 23823.4 73107.4
130.0 8632.0 19616.7
x2 , S2 , n1 55
355.0 19616.7 55964.5
n1 1 n2 1
S pooled S1 S2
n1 n2 2 n1 n2 2
10963.7 21505.5
21505 .5 63661 .3
98(2)
c2 F2,97 (0.05) 6.26
97
7/14/2017 29
Example 6.4: Electrical Usage of Homeowners with
and without ACs
95% simultaneous confidence intervals
1 1
11 21 : (204.4 130.0) 6.26 10963.7
45 55
or 21.7 11 21 127.1
1 1
12 22 : (556.6 355.0) 6.26 63661.3
45 55
or 74.7 12 22 328.5
7/14/2017 30
Example 6.4:
95% Confidence Ellipse
7/14/2017 31
Bonferroni Simultaneous Confidence Intervals
1 1
1i 2i : x1 x 2 t n1 n2 2 sii , pooled
2p n1 n2
7/14/2017 32
Result 6.4
n1 p and n2 p are large
100% confidence ellipsoid for 1 2 :
1
1
x1 x 2 1 2 ' S1 S 2
1
n1 n2
x1 x 2 1 2 p2 ( )
Simultaneous confidence intervals for a' 1 2 :
1
a' x1 x 2 ( ) a' S1 S 2 a
2 1
p
n1 n2
7/14/2017 33
Proof of Result 6.4
7/14/2017 34
Remark
If n1 n2 n
n 1 1
nn2 2
S1 S 2 S1 S 2
1 1 1
n1 n2 n
n 1S1 n 1S 2 1
1 1 1
S pooled
nn2 n n n n
7/14/2017 35
Example 6.5
Example 6.4 Data
1 1 464.17 886.08
S1 S 2
n1 n2 886.08 2642.15
11 21 : 74.4 5.99 464.17 or 21.7, 127.1
12 22 : 201.6 5.99 2642.15 or 75.8, 327.4
H 0 : 1 2 0
1
1
T x1 x 2 ' S1 S 2 x1 x 2 15.66 22 (0.05) 5.99
2 1
n1 n2
1
1 0.041
1
Critical linear combination : S1 S 2 x1 x 2
n1 n2 0.063
7/14/2017 36
Example 6.9: Nursing Home Data
Nursing homes can be classified by the owners: private (271), non-
profit (138), government (107)
Costs: nursing labor, dietary labor, plant operation and maintenance
labor, housekeeping and laundry labor
To investigate the effects of ownership on costs
7/14/2017 37
One-Way MANOVA
Population 1 : X11 , X12 , , X1n1
Population 2 : X 21 , X 22 , , X 2 n2
Population g : X g1 , X g 2 , , X gng
MANOVA (Multivariate ANalysis Of VAriance)
is used to investigate whether the population mean
vectors are the same, and, if not, which mean
components differ significantly
7/14/2017 38
Assumptions about the Data
X 1 , X 2 , , X n : random sample from a population
with mean , 1, 2, , g
Random sample from different populations are
independent
All populations have a common covariance
matrix
Each population is multivariate normal
7/14/2017 39
Univariate ANOVA
X 1 , X 2 , , X n : random sample from N ( , 2 )
1, 2, , g
Null hypothesis H 0 : 1 2 g
Reparameterization
H 0 :1 2 g 0
g
X j ej , ej : N (0, 2 ), n
1
0
xj x x x xj x
7/14/2017 40
Univariate ANOVA
x j x x x xj x 2 x x xj x
2 2 2
x x 0
n
j
j 1
x x n x x xj x
n n
2 2 2
j
j 1 j 1
x x n x x xj x
g n g g n
2 2 2
j
1 j 1 1 1 j 1
(SS cor ) (SS tr ) (SS res )
n x x xj x
g n g g n
x 2
j 1 2
n n n x 2 2 2
1 j 1 1 1 j 1
(SS obs ) (SS mean ) (SS tr ) (SS res )
7/14/2017 41
Univariate ANOVA
7/14/2017 42
Univariate ANOVA
Reject H 0 : 1 2 g 0 at level if
SS tr / g 1)
F Fg 1, n g ( )
g
SS res / n g
1
1 SS res
1 SS tr / SS res SS res SS tr
7/14/2017 43
Examples 6.6 & 6.7
9 6 9 4 4 4 4 4 4 1 2 1
0 2 4 4 3 3 1 1
3 1 2 4 4 4 2 2 2 1 1 0
SS obs 216, SS mean 128
SS tr 78, d.f. 3 1 2
SS res 10, d.f. (3 2 3) - 3 5
SS tr /( g 1) 78 / 2
F 19.5 F2,5 (0.01) 13.27
SS res / n g 10 / 5
H 0 : 1 2 3 0 is rejected at the 1% level
7/14/2017 44
MANOVA
X j e j ; j 1, 2, , n ; 1, 2, , g
e j : N p (0, ), : overall mean (level)
g
: th treatment effect, n 0
1
x j x x x x j x e j
x x x j x ' n x x x x '
g n g
j
1 j 1 1
x j x x j x ' B W
g n
1 j 1
7/14/2017 45
MANOVA
7/14/2017 46
MANOVA
W x j x x j x '
g n
1 j 1
n1 1S1 n2 1S 2 n g 1S g
Reject H 0 : 1 2 g 0 if Wilk' s lambda
x x x j x '
g n
j
W 1 j 1
*
B W
x x x j x '
g n
j
1 j 1
is too small
7/14/2017 47
Distribution of Wilks Lambda
7/14/2017 48
Test of Hypothesis for Large Size
If H 0 is true and n n is large,
pg
- n 1 ln *
: 2
p ( g 1)
2
Reject H 0 at significance level if
p g W
p2 ( g 1) ( )
- n 1 ln
2 BW
7/14/2017 49
Popular MANOVA Statistics Used in Statistical
Packages
W
Wilk' s lambda *
BW
Lawley - Hotelling trace tr BW 1
Pillai trace tr B(B W ) 1
Roy' s largest root
1
maximum eigenvalue of W (B W )
7/14/2017 50
Which do you choose?
For the most part stick with Wilks lambda. Its
the most widely used
Use Hotellings Trace if
Manipulated (experimental) variables
Very clean design with no internal validity
problems
Pillais trace is the most conservative, but if
your design has many problems (e.g.
unbalanced, assumption violation, etc) pillais is
supposed to be robust to these problems
7/14/2017 51
Assessing Individual variable
If multivariate test is significant
Run multiple univariate F-tests (one per
variable) in order to see on which
variable there are group differences, this
assumes uncorrelated among variables.
7/14/2017 52
Assessing Individual variable
If variables are correlated than individual F-tests are problematic
but usually this is ignored and univariate Fs interpreted anyway
7/14/2017 53
Example 6.8
9 6 9
2 7
3
0 2 8 1 4 4
0 x1 , x 2 , x 3 , x
4 4 2 5 5
3 1 2
9 7
8
9 6 9
0 2 , SS obs SS mean SS tr SS res 128 78 10 216
3 1 2
3 2 7
4 0 , SS obs SS mean SS tr SS res 200 48 24 272
8 9 7
7/14/2017
54
Example 6.8
9 6 9 4 4 4 4 4 4 1 2 1
0 2 4 4 3 3 1 1
3 1 2 4 4 4 2 2 2 1 1 0
3 2 7 5 5 5 1 1 1 1 2 3
4 0 5 5 3 3 2 2
8 9 7 5 5 5 3 3 0 1 1
3
Cross products
Mean : 8 4 5 160
Treatment : 3 4 (-1) 2 (-3) (-3) 3 (-2) 3 -12
Residual : 1 (-1) (-2) (-2) 1 3 0 (-1) 1
Total : 9 3 6 2 9 7 2 7 149
7/14/2017 55
Example 6.8
7/14/2017 56
Example 6.8
10 1
W 1 24
*
0.0385
BW 88 11
11 72
1 * n g 1 1 0.0385 8 3 1
g 1 0.0385 3 1
*
8.19 F2 ( g 1), 2 n g 1 (0.01) F4,8 (0.01) 7.01
Reject H 0
7/14/2017 57
Example 6.9: Nursing Home Data
Nursing homes can be classified by the owners: private (271), non-
profit (138), government (107)
Costs: nursing labor, dietary labor, plant operation and maintenance
labor, housekeeping and laundry labor
To investigate the effects of ownership on costs
7/14/2017 58
Example 6.9
7/14/2017 59
Example 6.9
W (n1 1)S1 (n2 1)S 2 (n3 1)S 3
182.962
4.408 8.200
1.695 0.633 1.484
9.581 2.428 0.394 6.538
n1 x1 n2 x 2 n3 x 3
x 2.136 0.519 0.102 0.380'
n1 n2 n3
3.475
g 1.111 1.225
B n x x x x '
1 0.821 0.453 0.235
0.584 0.610 0.230 0.304
7/14/2017 60
Example 6.9
W
*
0.7714
B W
n p 2 1 *
17.67
p
*
F24, 2510 (0.01) 82 (0.01) / 8 2.51
or, approximate analysis
W
(n 1 ( p g ) / 2) ln 132.76
B W
p2 ( g 1) (0.01) 82 (0.01) 20.09
7/14/2017 Reject H 0 by both analyses 61
Bonferroni Intervals for
Treatment Effects
ki x ki xi , ki i x ki xi
1 1
Varki i Var x ki xi ii
n k n
W n1 1S1 n2 1S 2 n g 1S g
(n g )S pooled (n g )
1 1 wii
Varki i
n k n ( n g )
m pg ( g 1) / 2
7/14/2017 62
Result 6.5: Bonferroni Intervals for
Treatment Effects
With confidence at least (1 )
ki i belongs to
wii 1 1
x ki xi t n g
pg ( g 1) n g nk n
7/14/2017 63
Example 6.10: Example 6.9 Data
1 x1 x 0.070 0.039 0.020 0.020'
3 x 3 x 0.137 0.002 0.023 0.003'
13 33 0.20 0.023 0.043, n 516
1 1 w33 1 1 1.484
0.00614
n1 n3 n g 271 107 516 3
t 513 (0.05 / 4 3 2) 2.87
95% simultaneous confidence interval for 13 33
- 0.043 2.87 0.00614 or 0.061,0.025
95% simultaneous confidence intervals for
13 23 and 23 33 : 0.058,0.026, 0.021,0.019
7/14/2017 64
Two-Way ANOVA
X kr k k ekr
1, 2, , g ; k 1, 2, , b; r 1, 2, , n
g b g b
k k k
1
0, e
k 1
kr : N
1
( 0, 2
)
k 1
E X kr k k
xkr x x x xk x xk x xk x xkr xk
g b n g b
xkr x bn x x gn xk x
2 2 2
1 k 1 r 1 1 k 1
g b g b n
n xk x xk x xkr xk
2 2
1 k 1 1 k 1 r 1
SS cor SS fac1 SS fac 2 SS int SS res
7/14/2017 65
Two-Way ANOVA
7/14/2017 66
Two-Way ANOVA
F ratio tests
SS fac1 /( g 1)
: for effects of factor 1
SS res / gb(n 1)
SS fac 2 /(b 1)
: for effects of factor 2
SS res / gb(n 1)
SS int /( g 1)(b 1)
: for effects of
SS res / gb(n 1)
factor 1 - factor 2 intercation
7/14/2017 67
Example 6.11: Plastic Film Data
7/14/2017 68
Two-Way MANOVA
X kr k k e kr
1, 2, , g ; k 1, 2, , b; r 1, 2, , n
g b g b
1
k k k 0, e kr : N p (0, )
k 1 1 k 1
x kr x x x x k x x k x x k x x kr x k
g b n
x
1 k 1 r 1
kr x x kr x '
g b
bnx
1
x x x ' gnx k x x k x '
k 1
g b
nx k x x k x x k x x k x '
1 k 1
g b n
x kr x k x kr x k '
7/14/2017
1 k 1 r 1 69
Effect of Interactions It would be useful to add an
interaction term to the model if you
want to test the hypothesis that the
expected responses of Y at different
levels of factor 1 are different at
different levels of factor 2.
7/14/2017 70
Two-Way MANOVA
7/14/2017 71
Two-Way MANOVA
Test for interaction
For large samples, reject H 0 : 11 12 gb 0 if
p 1 ( g 1)(b 1)
gb(n 1) ln *
( g 1)( b 1) ( )
2
2
SSPres
Wilk' s lambda *
SSPint SSPres
If interaction effects exist, the factor effects do not have a
clear interpretation
7/14/2017 72
Two-Way MANOVA
Test for factor 1 effect
For large samples, reject H 0 : 1 2 g 0 if
p 1 ( g 1)(b 1)
gb(n 1) ln *
( g 1) p ( )
2
2
SSPres
Wilk' s lambda *
SSP fac1 SSPres
Test for factor 2 effect
For large samples, reject H 0 : 1 2 b 0 if
p 1 ( g 1)(b 1)
gb(n 1) ln *
( b 1) p ( )
2
2
SSPres
Wilk' s lambda *
SSP fac 2 SSPres
7/14/2017 73
Bonferroni Confidence Intervals
With negligible interactions,
the simultaneus confidence intervals are
Eii 2
x i x mi t p for i mi
pg ( g 1) bn
and
Eii 2
xki xqi t p for ki qi
pb(b 1) gn
gb(n 1), E SSPres
7/14/2017 74
Example 6.11: MANOVA Table
7/14/2017 75
Example 6.11: Interaction
SSPres
*
0.7771
SSPint SSPres
( g 1)(b 1) 1
1 * gb(n 1) p 1 / 2
F *
( g 1)(b 1) p 1 / 2
: F 1 , 2
1 ( g 1)(b 1) p 1 3
2 gb(n 1) p 1 14
F 1.34 F3,14 (0.05) 3.34
H 0 : 11 12 21 22 0 (no interaction) is not rejected
7/14/2017 76
Example 6.11: Effects of
Factors 1 & 2
SSPres
*
1 0.3819
SSP fac1 SSPres
SSPres
*
2 0.5230
SSP fac 2 SSPres
1 *1 2 / 2
F1 * 7.55, 1 g 1 p 1 3
1 1 / 2
1 *2 2 / 2
F2 * 4.26, 1 b 1 p 1 3
2 1 / 2
2 gb(n 1) p 1 14
F1 F3,14 (0.05) 3.34, reject H 0 : 1 2 0
F2 F3,14 (0.05) 3.34, reject H 0 : 1 2 0
7/14/2017 77
Profile Analysis
We begin with a definition of profile analysis (a.k.a. growth curves or repeated
measures).
For each subject we have a vector of observations:
We assume that the observations are measured in the same units.
Typically, these analyses are done with data that has been collected over
time and we are interested in seeing whether there is a general linear trend
(an ordering is not necessary though).
7/14/2017 78
Profile Analysis
A battery of p treatments (tests, questions, etc.) are administered to
two or more group of subjects
The question of equality of mean vectors is divided into several
specific possibilities
Are the profiles parallel?
Are the profiles coincident?
Are the profiles level?
7/14/2017 79
Previously we only used a test to see if we believed that the
means were all equal, well now with two vectors we can test
three things.
1. Are the two profiles parallel?
1a. Is H01 : 1i 1i1 = 2i 2i1 (i = 2, 3, . . . , p)?
2. If they are parallel, are they coincident?
2a. Is H02 : 1i = 2i (i = 1, 2, 3, . . . , p)?
3. If they are coincident, are they level?
3a. Is H03 : 1i = 12 = = 1p = 2i = 22 = = 2p ?
7/14/2017 80
Example 6.12:
Love and Marriage Data
To demonstrate profile analysis, we will use the example data set from your
text-- Example 6.12: Love and Marriage Data:
As part of a larger study of love and marriage, E. Hatfield, a sociologist,
survey adults with respect to their marriage contributions and outcomes
and their levels of passionate and companionate love. Recently married
males and females were asked to respond to four questions using an 8-point
scale (8 representing extremely positive for the first two questions) and using
the 5-point scale (5 representing extremely positive for the last two
questions).
7/14/2017 81
Example 6.12:
Love and Marriage Data
7/14/2017 82
7/14/2017 83
Part 1: Parallel Profiles?
We begin by asking whether the two population mean vectors are parallel.
If they are parallel, it could mean that:
One group performed uniformly better than the other group.
Both groups are equal.
To test this we need to think what it would mean that they are parallel.
It would mean that the difference between any two means in group one is
the same as the difference between the same two means in group two.
7/14/2017 84
7/14/2017 85
7/14/2017 86
7/14/2017 87
7/14/2017 88
7/14/2017 89
7/14/2017 90
Population Profile
7/14/2017 91
Profile Analysis
Assume two populations
Are the profiles parallel?
H 01 : 1i 1i 1 2i 2i 1 , i 2, 3, , p
Are the profiles coincident?
H 02 : 1i 2i , i 1, 2, , p
Are the profiles level?
H 03 : 11 12 1 p 21 22 2 p
7/14/2017 92
Test for Parallel Profiles
1 1 0 0 0 0
0 1 1 0 0 0
C
( p 1) p
0 0 0 0 1 1
CX1 j : N p 1 (C 1 , C ' ), CX 2 j : N p 1 (C 2 , C ' )
Reject H 01 : C 1 C 2 at level if
1
1 1
T x1 x 2 ' C' CS pooled C' Cx1 x 2 c 2
2
n1 n2
(n1 n2 2)( p 1)
c
2
F p 1,n1 n2 p ( )
n1 n2 p
7/14/2017 93
Test for Coincident Profiles
Given parallel profiles
Reject H 02 : 1' 1 1' 2 at level if
1
1 1
T 1' x1 x 2 1' S pooled 1 1' x1 x 2
2
n1 n2
2
1' x1 x 2
t n1 n2 2 ( ) F1,n1 n2 2 ( )
2
1 1 2
1' S pooled 1
n1 n2
7/14/2017 94
Test for Level Profiles
Given coincident profiles
Reject H 03 : C 0 at level if
(n1 n2 ) x' C' CSC' Cx c
1 2
(n1 n2 1)( p 1)
c
2
F p 1,n1 n2 p 1 ( )
n1 n2 p 1
n1 n2
x x1 x2
n1 n2 n1 n2
7/14/2017 95
Example 6.14
6.833 6.633
7.033 7.000
x1 , x 2
3.967 4.000
4.700 4.533
0.606 0.262 0.066 0.161
0.262 0.637 0.173
0.143
S pooled
0.066 0.173 0.810 0.029
0.161 0.143 0.029 0.306
7/14/2017 96
Example 6.14:
Test for Parallel Profiles
1 0 0
1 1 0 0 1 1 0
CS pooled C' 0 1 1 0S pooled
0 1 1
0 0 1 1
0 0 1
0.200
1 1 0 0 0.167
Cx1 x 2 0 0.033
1 1 0 0.066
0.033
0 0 1 1 0.200
0.167
(30 30 2)(4 1)
T 1.005
2
F3,56 (0.05) 8.7
30 30 4
7/14/2017 97
Example 6.14: Sample Profiles
7/14/2017 98
Example 6.14:
Test for Coincident Profiles
1' x1 x 2 0.367
1' S pooled 1 4.207
2
0.367
T
2
0.501 F1,58 (0.05) 4.0
1 1 4.207
30 30
7/14/2017 99
Example 6.15: Ulna Data,
Control Group
7/14/2017 100
Example 6.15: Ulna Data,
Treatment Group
7/14/2017 101
Comparison of Growth Curves
X j : vector of p measurements on subject j in group
j 1, 2, , n ; 1, 2, , g
X j : Multivariate normal with covariance
Putthoff - Roy model
0 1t1 q t1q 1 t1 t1q 0
q q
0 1t 2 q t 2 1 t 2 t 2 1
E ( X j )
q q
0 1t p q t p 1 t p t p q
B
7/14/2017 102
Comparison of Growth Curves
Maximum likelihood estimators of :
B' S pooled
1
B
1
B' S pooled
1
X
S pooled
1
(n1 1S1 n g 1S g ) W
Ng Ng
g
k ^
N n , Cov( )
1
B' S pooled
1
B
1 n
k ( N g )( N g 1) /( N g p q )( N g p q 1)
g n
W
Wq
X j B X j B ' ,
*
1 j 1 Wq
Reject the null hypothesis that the polynomial is adequate if
N ( p q g ) / 2 ln * (2p q 1) g ( )
7/14/2017 103
Example 6.15
Use quadratic growth model
73.0701 70.1387
1 2 3.6444 4.0900
2.0274 1.8534
Control Group : 73.07 3.64t 2.03t 2
Treatment Group : 70.14 4.09t 1.85t 2
* 0.7627
N ( p q g ) / 2 ln * 7.86 (24 21) 2 (0.01) 9.21
7/14/2017 104
Example 6.16: Comparing Multivariate and
Univariate Tests
7/14/2017 105
Example 6.16: Comparing Multivariate and
Univariate Tests
Univariate test on x1 : F 2.46 F1,18 (0.10) 3.01
Univariate test on x 2 : F 2.68 F1,18 (0.10) 3.01
Accept 1 2
Hotelling' s test :
18 2
T 17.29 c
2 2
F2,17 (0.01) 12.94
17
Reject 1 2
7/14/2017 106
Strategy for Multivariate Comparison of
Treatments
Try to identify outliers
Perform calculations with and without the outliers
Perform a multivariate test of hypothesis
Calculate the Bonferroni simultaneous confidence intervals
For all pairs of groups or treatments, and all characteristics
7/14/2017 107
Importance of Experimental Design
Differences could appear in only one of the many
characteristics or a few treatment combinations
Differences may become lost among all the inactive
ones
Best preventative is a good experimental design
Do not include too many other variables that are not
expected to show differences
7/14/2017 108
Wishart Distribution
e
( n p 2) / 2 tr A 1 / 2
A
wn 1 ( A | )
1
p
n i
( n 1) / 2
2 p ( n 1) / 2
p ( p 1) / 4
i 1 2
A : positive definite
Properties :
A 1 : Wm1 ( A 1 | ), A 2 : Wm2 ( A 2 | )
A 1 A 2 : Wm1 m2 ( A 1 A 2 | )
A : Wm ( A | ) CAC': Wm (CAC' | C ' )
7/14/2017 109
Concept of Degrees of Freedom
y ' x11 , , x1n1 , x 21 , , x 2 n2 , , x gng : d.f. n n1 n2 n g
Treatment vector
1 0 0
1 0 0
0
1
0
x1 x x2 x x g x
0 1 0
0 0 1
0 0 1
x1 x u1 x 2 x u 2 x g x u g
7/14/2017 110
Concept of Degrees of Freedom
1 1, , 1 u1 u 2 u g
Treatment vector and 1 are all on the hyperplane
spanned by u1 , u 2 , , u g : d.f. g
1 is perpendicular to the treatment vector
mean vector x1 : d.f. g 1
Residual vector
e y x1 x1 x u1 x 2 x u 2 x g x u g
perpendicular to the hyperplane spanned by u1 , u 2 , , u g
d.f. of e : n g
7/14/2017 111