Chapter 11A Multi-Factor Repeated Measures ANOVA Repeated Measures On Both Factors

Download as pdf or txt
Download as pdf or txt
You are on page 1of 41

Chapter 11A

Multi-Factor Repeated Measures ANOVA


Repeated Measures on Both Factors

Page
1. Introduction 11-2
2. Structural model, SS partitioning, and the ANOVA table 11-3
3. Two-Factor Repeated Measures ANOVA in SPSS 11-8
4. Contrasts and Effect Sizes 11-12
5. An example 11-27

11A-1 2007 A. Karpinski


Repeated Measures ANOVA
Two-Factor Repeated Measures

1. Introduction

Participants take part in a training program to help them prepare for a


standardized test. Before the training, they take the test and scores are
recorded for all three sub-scales of the test. After the 12-week training
program, participants retake the test.

Pre-training Post-training
Participant Subscale1 Subscale2 Subscale3 Subscale1 Subscale2 Subscale3
1 42 42 48 48 60 78
2 42 48 48 36 48 60
3 48 48 54 66 78 78
4 42 54 54 48 78 90
5 54 66 54 48 66 72
6 36 42 36 36 48 54
7 48 48 60 54 72 84
8 48 60 66 54 72 90
9 54 60 54 48 72 78
10 48 42 54 54 66 78
46.2 51.0 52.8 49.2 66.0 76.2

Test Performance Test Performance

80 80
Test Score
Test Score

70 Scale1 70
Pre
60 Scale2 60
Post
50 Scale3 50
40 40
Pre Post Scale1 Scale2 Scale3
Time Time

With this design, several questions come to mind:


o Overall, does the training improve test scores?
Does training improve test scores for subscale 1?
Does training improve test scores for subscale 2?
Does training improve test scores for subscale 3?
o Overall, is there a difference in performance on the three sub-scales?

11A-2 2007 A. Karpinski


We have two repeated measures factors:
o Pre-test and post-test scores
o The three subscales of the test

We can classify this design as a 2*3 repeated measures design, with


repeated measures on both factors.
Subscale of test (Factor A)
Time (Factor B) Subscale 1 Subscale 2 Subscale 3
Pre-test X .11 = 46.2 X .21 = 51.0 X .31 = 52.8 X ..1 = 50.0
Post-test X .12 = 49.2 X .22 = 66.0 X .32 = 76.2 X ..2 = 63.8
n = 10 X .1. = 47.7 X .2. = 58.5 X .3. = 64.5

Everything we learned about interpreting two-way between-subjects designs


applies here. The only difference will be the assumptions of the test, and the
construction of the error term.

2. Structural model, SS partitioning, and the ANOVA table


We will only consider the case where the factors are fixed variables.

Here is the structural model for a two-factor repeated measures design:


Yijk = + j + k + + ( ) jk + ( ) + ( ) + ( )

o Factor A ( j ) and Factor B ( k ) and the A*B interaction ( jk ) are fixed


effects
o The Subject effect ( i ) is a random effect. Thus, all interaction terms
involving the subject effect are also random effects

o Because we have one observation per participant, we do not have enough


information to estimate both the ( ) interaction and the within cell
residuals ( ijk ) .
In the randomized block design, we omitted the interaction term and
retained the estimate of error
For factorial within-subjects resigns, we will omit the error term, and
consider the information to be an estimate of the A*B*Subject
interaction term.
This difference is a difference of terminology, not a sustentative
difference.

11A-3 2007 A. Karpinski


o We can compute estimates for the fixed terms in the model, just as we
have for factorial designs:

The overall mean of the scores

j The effect of being in level j of Factor A


a
j = . j . ...
j =1
j =0

k The effect of being in level k of Factor B


b
k = ..k ...
k =1
k =0

( ) jk The effect of being in level j of Factor A and level k of Factor B


(the interaction of level j of Factor A and level k of Factor B)
( ) jk = . jk . j . ..k + ...
a

( )
j =1
jk = 0 for each level of j
b

( )
k =1
jk = 0 for each level of k

o The remaining terms are random effects.

11A-4 2007 A. Karpinski


o What do we do with all the random effect parameters? Lets take a closer
look at the ( ) parameter. To look at the Factor B * Subject effect, we
need to collapse across Factor A

Participant Difference
Pre- Post-
Training Training
1 44 62 18
2 46 48 2
3 50 74 24
4 50 72 22
5 58 62 4
6 38 46 8
7 52 70 18
8 58 72 14
9 56 66 10
10 48 66 18
50.0 63.8 13.8

The B*Subject interaction examines if the effect of B (Pre vs. Post-


training) is the same across all participants.

In other words, the B*Subject interaction is a measure of the


variability in the B effect or how much error we have in the
measurement of the B effect
(And so intuitively it makes sense that we can use the B*Subject
term as an error term when we test the B effect)

This logic extends across each of the fixed effects

The A*Subject interaction measures the variability in the A effect

The B*Subject interaction measures the variability in the B effect

The A*B*Subject interaction measures the variability in the A*B


interaction

11A-5 2007 A. Karpinski


o For a two-factor repeated measures design, we have the following SS
decomposition.

SS Total
(SS Corrected Total)

SS Model SS Error

SS Main SS 2-Way SS SS SS SS
Effects Interactions A*S B*S A*B*S Subject

SS SS SS
A B A*B

11A-6 2007 A. Karpinski


The ANOVA table for a two-factor repeated measures design:

o Remember that to construct a valid F-test for an effect, we need:


The numerator to contain exactly one more term than the
denominator
The extra term must correspond to the effect being tested

o When these conditions hold:


The F-ratio will equal 1 when the null hypothesis is true (because the
numerator and denominator will be estimating the same effects)
The F-ratio will be greater than 1 when the null hypothesis is false

Source SS df MS E(MS) F
Factor A SSA a-1 SSA nb 2j MSA
a 1 + b +
2 2
MS ( A * S )
a 1
A*S SS (a-1)(n-1) SS ( A * S ) + b
2 2

(Factor A Error) (A*S) (a 1)(n 1)

Factor B SSB b-1 SSB na k2 MSB


b 1 + a +
2 2
MS ( B * S )
b 1
B*S SS (b-1)(n-1) SS ( B * S ) + a
2 2

(Factor B Error) (B*S) (b 1)(n 1)

A*B SSAB (a-1)(b-1) SSAB n jk2 MSAB


(a 1)(b 1) + +
2 2

(a 1)(b 1) MS ( A * B * S )

A*B*S SS (a-1)(b-1) SS ( A * B * S ) 2 +
2

(A*B Error) (A*B*S) *(n-1) (a 1)(b 1)(n 1)

Subjects (S) SSS (n-1) SSS 2 + ab 2


n 1
Total SST N-1

11A-7 2007 A. Karpinski


o For example, lets consider the test for Factor A

H 0 : .1. = .2. = ... = .a .


H 0 : 1 = 2 = ... = a = 0
nb 2j
+ b +
2 2
MSA a 1
FA [(a 1), (a 1)(n 1)] = =
MS ( A * S ) + b
2 2

If H0 is true: 2
j =0
2 + b
2
Then FA = 2 =1
+ b
2

If H0 is false: 2
j >0
nb 2j
2 + b
2
+
Then FA = a 1 >1
2 + b
2

o Note that unlike the one-way within-subjects design, it is not possible to


construct an F-test for the effect of subjects.

3. Two-Factor Repeated Measures ANOVA in SPSS

Lets see how the ANOVA looks in SPSS.


o We need to enter the within subjects factors correctly. First, we enter the
name and number of levels of each repeated factor.

11A-8 2007 A. Karpinski


o Next, we need to identify which variables go with which factors:

(1,1) means time 1 and scale 1 pre1


(1,3) means time 1 and scale 3 pre3
(2,3) means time 2 and scale 3 post3

If you do not identify the factors properly, you will misinterpret your
results!

11A-9 2007 A. Karpinski


o Or you can enter the following syntax:
GLM pre1 pre2 pre3 post1 post2 post3
/WSFACTOR = time 2 scale 3
/PRINT = DESC.

Time is the first repeated factor with 2 levels


Scale is the second repeated factor with 3 levels
The order of the variables needs to be
Time 1, Scale 1 pre1
Time 1, Scale 2 pre2
Time 1, Scale 3 pre3
Time 2, Scale 1 post1
Time 2, Scale 2 post2
Time 2, Scale 3 post3

o If we switched the order of the factors, we would need to also switch the
order of the variables:
GLM pre1 post1 pre2 post2 pre3 post3
/WSFACTOR = scale 3 time 2
/PRINT = DESC.

This syntax will give us exactly the same output as the syntax above

o Now, we can check the sphericity assumption (presumably, we already


checked the normality assumption before starting to run the ANOVA)
Mauchly's Test of Sphericity

Measure: MEASURE_1

Epsilon
Greenhous
Within Subjects Effect e-Geisser Huynh-Feldt Lower-bound
TIME 1.000 1.000 1.000
SCALE .962 1.000 .500
TIME * SCALE .904 1.000 .500
Tests the null hypothesis that the error covariance matrix of the
orthonormalized transformed dependent variables is proportional to
an identity matrix.

We get an epsilon for each effect (main effect and interactions)


We can use our same rules of thumb for determining if we have
compound symmetry. In this case, we are actually OK!

11A-10 2007 A. Karpinski


o Here is the SPSS ANOVA table with the epsilon-adjusted tests removed:

Tests of Within-Subjects Effects

Measure: MEASURE_1
Type III Sum
Source of Squares df Mean Square F Sig.
TIME Sphericity Assumed 2856.600 1 2856.600 33.766 .000
Error(TIME) Sphericity Assumed 761.400 9 84.600
SCALE Sphericity Assumed 2899.200 2 1449.600 40.719 .000
Error(SCALE) Sphericity Assumed 640.800 18 35.600
TIME * SCALE Sphericity Assumed 1051.200 2 525.600 45.310 .000
Error(TIME*SCALE) Sphericity Assumed 208.800 18 11.600

Each within-subjects factor is immediately followed by its appropriate


error term

Main effect of time: F(1,9) = 33.77, p < .001


Compares X ..1 = 50.0 vs. X ..2 = 63.8

Main effect of scale: F(2,18) = 40.72, p < .001


Compares X .1. = 47.7 vs. X .2. = 58.5 vs. X .3. = 64.5

Time by scale: F(2,18) = 45.31, p < .001


Examines if the time effect is the same for each scale
OR Examines if the scale effect is the same at each time

Subscale of test (Factor A)


Time (Factor B) Subscale 1 Subscale 2 Subscale 3
Pre-test X .11 = 46.2 X .21 = 51.0 X .31 = 52.8 X ..1 = 50.0
Post-test X .12 = 49.2 X .22 = 66.0 X .32 = 76.2 X ..2 = 63.8
n = 10 X .1. = 47.7 X .2. = 58.5 X .3. = 64.5

Test Performance

80
Test Score

70
Pre
60
Post
50
40
Scale1 Scale2 Scale3
Time

11A-11 2007 A. Karpinski


However, the main effect for scale and the time*scale interaction are
omnibus tests. We previously stated that we wanted to avoid omnibus
tests at all costs for repeated-measures designs

Technically, in this case we are OK because we have spherical data


but it is good practice to avoid omnibus tests for these designs.

4. Contrasts and Effect Sizes


The formulae for tests of contrasts are the same formulae we used for one-
factor within-subjects designs.

t observed =

=
c j X.j
standard error ( ) c 2j
MSE
n

2 SS
SS = F (1, df ) =
c 2j MSE
n

o The strongly recommended (and the SPSS) approach


MSE will be the contrast-specific error term (with df = n-1).
o The alternative, use at your own risk approach relies on the data being
spherical. If the data are spherical, then we can use the appropriate
omnibus error term:
For contrasts on the marginal Factor A means, use the omnibus Factor
A error term, MSE = MSA *S (with df = (a-1)(n-1)).
For contrasts on the marginal Factor B means, use the omnibus Factor
B error term, MSE = MSB *S (with df = (b-1)(n-1)).
For contrasts on the A*B cell means, use the omnibus A*B interaction
error term, MSE = MSA *B *S (with df = (a-1)(b-1)(n-1)).
I recommend that you always use the contrast-specific error term.

11A-12 2007 A. Karpinski


o Just as for one-factor within-subjects designs, we have a number of
options for effect sizes

Partial eta-squared is a measure of percentage of the variance


accounted for (in the sample) that can be used for omnibus tests or
contrasts:
SSeffect
(Effect
2
) =
SSeffect + SSErrorTermForEffect
SS A SS B SS S *B
A2 = B2 = A2*B =
SS A + SS A*S SS B + SS B*S SS A*B + SS A*B*S
SS Contrast
Contrast
2
=
SS Contrast + SS ErrorTermForContrast

Tests of Within-Subjects Effects

Measure: MEASURE_1
Type III Sum
Source of Squares df Mean Square F Sig.
TIME Sphericity Assumed 2856.600 1 2856.600 33.766 .000
Error(TIME) Sphericity Assumed 761.400 9 84.600
SCALE Sphericity Assumed 2899.200 2 1449.600 40.719 .000
Error(SCALE) Sphericity Assumed 640.800 18 35.600
TIME * SCALE Sphericity Assumed 1051.200 2 525.600 45.310 .000
Error(TIME*SCALE) Sphericity Assumed 208.800 18 11.600

2856.6 2899.2
Time
2
= = .80 Scale
2
= = .82
2856.6 + 761.4 2899.2 + 640.8
1051.2
Time
2
*Scale = = .83
1051.2 + 208.8

This formula can be used for omnibus tests and for contrasts.

For contrasts (except maybe polynomial trends), we can also compute a d as


a measure of the effect size, just as we did for the paired t-test.

d =

but if and only if c i

Where: is the average value of the contrast of interest


is the standard deviation of the contrast values

For all contrasts, we can also compute an r as a measure of the effect size.
2
t Contrast FContrast
r = =
2
t Contrast + df contrast FContrast + df contrast

11A-13 2007 A. Karpinski


There are four methods we can use in SPSS to test contrasts:
o Create a new variable reflecting the value of the contrast and conduct a
one-sample t-test on this new variable
o Selecting only the groups of interest and running a contrast or paired t-
test on those groups
o SPSSs brand-name contrasts
o SPSSs special subcommand

Method 1: Compute a new variable for each contrast, and test if the value of
the contrast differs from zero.
o Lets start by testing three of our hypotheses

i. Does training improve test scores for subscale 1?


compute diff1 = post1-pre1.
T-TEST /TESTVAL=0
/VARIABLES=diff1.
One-Sample Statistics

Std. Error
N Mean Std. Deviation Mean
DIFF1 10 3.0000 7.61577 2.40832

One-Sample Test

Test Value = 0
95% Confidence
Interval of the
Mean Difference
t df Sig. (2-tailed) Difference Lower Upper
DIFF1 1.246 9 .244 3.0000 -2.4480 8.4480

3 .0
d = = = 0.39
7.61577

Test Performance

80
No. The scores on sub-scale 1
do not change significantly
Test Score

70
60
Pre between pre- and post-test, t(9)
Post = 1.25, p = .24, d = .39
50
40
Scale1 Scale2 Scale3
Time

11A-14 2007 A. Karpinski


ii. Does training improve test scores for subscale 2?
compute diff2 = post2-pre2.
T-TEST /TESTVAL=0
/VARIABLES=diff2.

One-Sample Statistics

Std. Error
N Mean Std. Deviation Mean
diff2 10 15.0000 10.67708 3.37639

One-Sample Test

Test Value = 0
95% Confidence
Interval of the
Mean Difference
t df Sig. (2-tailed) Difference Lower Upper
diff2 4.443 9 .002 15.00000 7.3621 22.6379

15
d = = = 1.40
10.67708

Test Performance
Yes. The scores on sub-scale
80
2 significantly improve
Test Score

70
Pre between pre- and post-test,
60 t(9) = 4.44, p < .01, d = 1.40
Post
50
40
Scale1 Scale2 Scale3
Time

11A-15 2007 A. Karpinski


iii. Does training improve test scores for subscale 3?
compute diff3 = post3-pre3.
T-TEST /TESTVAL=0
/VARIABLES=diff3.

One-Sample Statistics

Std. Error
N Mean Std. Deviation Mean
diff3 10 23.4000 6.60303 2.08806

One-Sample Test

Test Value = 0
95% Confidence
Interval of the
Mean Difference
t df Sig. (2-tailed) Difference Lower Upper
diff3 11.207 9 .000 23.40000 18.6765 28.1235

22.4
d = = = 3.54
6.60303

Test Performance

80
Yes. The scores on sub-scale 3
significantly improve between
Test Score

70
60
Pre pre- and post-test, t(9) = 11.21, p
Post < .01, d = 3.54
50
40
Scale1 Scale2 Scale3
Time

11A-16 2007 A. Karpinski


Method 2: Selecting the groups of interest and running paired-test on those
groups
o Note that this approach would not be recommended for between subjects
designs. However, for within-subjects designs, the default is to use a
contrast-specific error term. That is, we only use information from the
groups that are involved in the contrast to construct the error term. Thus,
for within-subjects designs, it is acceptable to select the groups of interest
and run a test only on those groups.

T-TEST PAIRS = pre1 pre2 pre3 WITH post1 post2 post3 (PAIRED).
Paired Samples Test

Paired Differences
95% Confidence
Interval of the
Std. Error Difference
Mean Std. Deviation Mean Lower Upper t df Sig. (2-tailed)
Pair 1 pre1 - post1 -3.00000 7.61577 2.40832 -8.44800 2.44800 -1.246 9 .244
Pair 2 pre2 - post2 -15.00000 10.67708 3.37639 -22.63792 -7.36208 -4.443 9 .002
Pair 3 pre3 - post3 -23.40000 6.60303 2.08806 -28.12352 -18.67648 -11.207 9 .000

Scale 1, pre vs post: t(9) = 1.25, p = .24, d = .39


Scale 2, pre vs post: t(9) = 4.44, p < .01, d = 1.40
Scale 3, pre vs post: t(9) = 11.21, p < .01, d = 3.54

These analyses are identical to the previously conducted analyses on


the difference scores we computed.

11A-17 2007 A. Karpinski


Method 3: SPSSs brand-name contrasts. SPSS conducts contrasts on the
marginal main effect means of the repeated measures factor, using contrast
specific error estimates
o Contrasts can only be specified on the marginal means. Tests on the cell
means are obtained by multiplying together main effect contrasts.
o So far, we have examined the effect of training for each subscale. Now,
we would like to test whether:
The effect of training on subscale 1 is the same as the effect of
training on subscale 2
The effect of training on subscale 1 is the same as the effect of
training on subscale 3
The effect of training on subscale 2 is the same as the effect of
training on subscale 3

70

65

60
Pre-test
55
Post-test
50

45

40
Subscale 1 Subscale 2

80 80
75 75
70 70
65 65
Pre-test Pre-test
60 60
Post-test Post-test
55 55
50 50
45 45
40 40
Subscale 1 Subscale 3 Subscale 2 Subscale 3

11A-18 2007 A. Karpinski


GLM pre1 pre2 pre3 post1 post2 post3
/WSFACTOR = time 2 Simple (1) scale 3 simple (1)
/PRINT = DESC.
Tests of Within-Subjects Contrasts

Measure: MEASURE_1
Type III Sum
Source time scale of Squares df Mean Square F Sig.
time Level 2 vs. Level 1 1904.400 1 1904.400 33.766 .000
Error(time) Level 2 vs. Level 1 507.600 9 56.400
scale Level 2 vs. Level 1 1166.400 1 1166.400 38.368 .000
Level 3 vs. Level 1 2822.400 1 2822.400 66.566 .000
Error(scale) Level 2 vs. Level 1 273.600 9 30.400
Level 3 vs. Level 1 381.600 9 42.400
time * scale Level 2 vs. Level 1 Level 2 vs. Level 1 1440.000 1 1440.000 45.000 .000
Level 3 vs. Level 1 4161.600 1 4161.600 83.903 .000
Error(time*scale) Level 2 vs. Level 1 Level 2 vs. Level 1 288.000 9 32.000
Level 3 vs. Level 1 446.400 9 49.600

o There are only two groups in the time factor, so no matter what we ask
SPSS to do, it will give us a pairwise contrast
Pre-test Post test
X ..1 = 50.0 X ..2 = 63.8
Coefficients -1 +1

SS Contrast 1904.4
Contrast
2
= = = .79
SS Contrast + SS ErrorTermForContrast 1904.4 + 507.6

F (1,9) = 33.77, p < .01, 2 = .79

o For the scale effect there are three groups, so we can ask SPSS to conduct
two main effect contrasts:
Subscale 1 Subscale 2 Subscale 3
X .1. = 47.7 X .2. = 58.5 X .3. = 64.5
Level 2 vs Level 1 -1 1 0
Level 3 vs Level 1 -1 0 1

1166.4 2822.4
22vs1 = = .81 32vs1 = = .88
1166.4 + 273.6 2822.4 + 381.6

Level 2 vs Level 1: F (1,9) = 38.37, p < .01, 2 = .81


Level 3 vs Level 1: F (1,9) = 66.57, p < .01, 2 = .88

11A-19 2007 A. Karpinski


o For the time by scale interaction there are 2 dfs, so SPSS will provide
two follow-up tests. SPSS multiplies each of the main effect contrasts
together to obtain interaction contrasts.
These are the tests we are interested in!

Level 2 vs Level 1* Subscale 1 Subscale 2 Subscale 3


Level 2 vs Level 1
Pre-test -1
Post-test +1
-1 1 0

Level 2 vs Level 1* Subscale 1 Subscale 2 Subscale 3


Level 2 vs Level 1
Pre-test 1 -1 0
Post-test -1 1 0

This contrast tests whether the effect of training on Subscale 1 is the


same as the effect of training for Subscale 2 (an interaction!)
H 0 : Pr e1 Post1 = Pr e 2 Post 2
H 1 : Pr e1 Post1 Pr e 2 Post 2

70

65

60 1440
Pre-test Contrast
2
= = .83
55
Post-test
1440 + 288
50

45 F (1,9) = 45.00, p < .01, 2 = .98


40
Subscale 1 Subscale 2

11A-20 2007 A. Karpinski


Level 2 vs Level 1* Subscale 1 Subscale 2 Subscale 3
Level 3 vs Level 1
Pre-test -1
Post-test +1
-1 0 1

Level 2 vs Level 1* Subscale 1 Subscale 2 Subscale 3


Level 3 vs Level 1
Pre-test -1 0 1
Post-test 1 0 -1

This contrast tests whether the effect of training on Subscale 1 is the


same as the effect of training for Subscale 3
H 0 : Pr e1 Post1 = Pr e 3 Post 3
H 1 : Pr e1 Post1 Pr e 3 Post 3

80
75
70
65 4161.6
Pre-test Contrast
2
= = .90
60
Post-test
4161.6 + 446
55
50
F (1,9) = 83.90, p < .01, 2 = .90
45
40
Subscale 1 Subscale 3

11A-21 2007 A. Karpinski


o To test whether the effect of training on Subscale 2 is the same as the
effect of training for Subscale 3, we need to run a new command.
GLM pre1 pre2 pre3 post1 post2 post3
/WSFACTOR = time 2 Simple (1) scale 3 simple (2)
/PRINT = DESC.
Tests of Within-Subjects Contrasts

Measure: MEASURE_1
Type III Sum
Source time scale of Squares df Mean Square F Sig.
time Level 2 vs. Level 1 1904.400 1 1904.400 33.766 .000
Error(time) Level 2 vs. Level 1 507.600 9 56.400
scale Level 1 vs. Level 2 1166.400 1 1166.400 38.368 .000
Level 3 vs. Level 2 360.000 1 360.000 10.588 .010
Error(scale) Level 1 vs. Level 2 273.600 9 30.400
Level 3 vs. Level 2 306.000 9 34.000
time * scale Level 2 vs. Level 1 Level 1 vs. Level 2 1440.000 1 1440.000 45.000 .000
Level 3 vs. Level 2 705.600 1 705.600 12.250 .007
Error(time*scale) Level 2 vs. Level 1 Level 1 vs. Level 2 288.000 9 32.000
Level 3 vs. Level 2 518.400 9 57.600

Level 2 vs Level 1* Subscale 1 Subscale 2 Subscale 3


Level 3 vs Level 2
Pre-test -1
Post-test +1
0 -1 1

Level 2 vs Level 1* Subscale 1 Subscale 2 Subscale 3


Level 2 vs Level 1
Pre-test 0 1 -1
Post-test 0 -1 1
H 0 : Pr e1 Post1 = Pr e 2 Post 2
H 1 : Pr e1 Post1 Pr e 2 Post 2

80
75
70
65 705.6
Pre-test Contrast
2
= = .58
60
Post-test
705.6 + 518.4
55
50
45
F (1,9) = 12.25, p < .01, 2 = .58
40
Subscale 2 Subscale 3

11A-22 2007 A. Karpinski


o Most cell mean contrasts that you would find of interest can be obtained
by multiplying two main effect contrasts (with the exception of simple
effect contrasts).

Method 4: SPSSs special subcommand.

o Lets return to our original questions:


i. Overall, does the training improve test scores?
ii. Does training improve test scores for subscale 1?
iii. Does training improve test scores for subscale 2?
iv. Does training improve test scores for subscale 3?
v. Overall, is there a difference in performance on the three sub-scales?
This final hypothesis is an omnibus hypothesis, so we will only
consider the first four

o To use the special subcommand:


Treat your design as a one-factor repeated-measures design
Enter the appropriate coefficients in the special command

GLM pre1 pre2 pre3 post1 post2 post3


/WSFACTOR = factor 6 special ( 1 1 1 1 1 1
-1 -1 -1 1 1 1
-1 0 0 1 0 0
0 -1 0 0 1 0
0 0 -1 0 0 1
-1 -1 2 1 1 -2).

Note that contrast 6 is of no interest to us, but we must enter five


contrasts after the row of ones.
Tests of Within-Subjects Contrasts

Measure: MEASURE_1
Type III Sum
Source FACTOR of Squares df Mean Square F Sig.
FACTOR L1 17139.600 1 17139.600 33.766 .000
L2 90.000 1 90.000 1.552 .244
L3 2250.000 1 2250.000 19.737 .002
L4 5475.600 1 5475.600 125.587 .000
L5 8294.400 1 8294.400 45.474 .000
Error(FACTOR) L1 4568.400 9 507.600
L2 522.000 9 58.000
L3 1026.000 9 114.000
L4 392.400 9 43.600
L5 1641.600 9 182.400

11A-23 2007 A. Karpinski


Tests of Within-Subjects Contrasts

Measure: MEASURE_1
Type III Sum
Source FACTOR of Squares df Mean Square F Sig.
FACTOR L1 17139.600 1 17139.600 33.766 .000
L2 90.000 1 90.000 1.552 .244
L3 2250.000 1 2250.000 19.737 .002
L4 5475.600 1 5475.600 125.587 .000
L5 8294.400 1 8294.400 45.474 .000
Error(FACTOR) L1 4568.400 9 507.600
L2 522.000 9 58.000
L3 1026.000 9 114.000
L4 392.400 9 43.600
L5 1641.600 9 182.400

i. Overall, does the training improve test scores? (L1)


Yes. Averaging across the subscales, post-test scores are higher
than pre-test scores, F(1,9) = 33.77, p < .01

ii. Does training improve test scores for subscale 1? (L2)


No. The scores on sub-scale 1 do not change significantly
between pre- and post-test, F(1,9) = 1.55, p = .24

iii. Does training improve test scores for subscale 2? (L3)


Yes. The scores on sub-scale 2 are higher at post-test than at
pre-test, F(1,9) = 19.74, p < .01

iv. Does training improve test scores for subscale 3? (L4)


Yes. The scores on sub-scale 3 are higher at post-test than at
pre-test, F(1,9) = 125.59, p < .01

Test Performance

80
Test Score

70
Pre
60
Post
50
40
Scale1 Scale2 Scale3
Time

11A-24 2007 A. Karpinski


Depending on how these contrasts are conducted, you may need to adjust
their p-values

o The recommended approach is to forgo the omnibus tests, and to conduct


three of fewer planned contrasts. If you take this approach, no correction
is necessary
If you have a large number of planned tests, you may need to apply a
Bonferroni correction.

o The alternative is to conduct the tests for main effects and interactions,
and then conduct the contrasts as follow-up tests. Now, the contrasts are
post-hoc tests. If they are pairwise, then you need to use the Tukey
procedure; if they are complex, you need to use the Scheff procedure to
adjust the p-values.

To use Tukeys HSD, compute q(1-,a,)


Where = Familywise error rate
a = Number of repeated-measures in the family
= df(error)

For single-df tests, df(error) should be (n -1), the df associated with


the contrast-specific error estimate.

To determine significance at the (1-) level,

qcrit (qcrit )2
Compare tobserved to or Fobserved to
2 2

To use the Scheff correction, compute FCrit = (r ) F =.05;r ,


Where = Familywise error rate
r = Degrees of freedom associated with the family
= df(error), (n -1) for a contrast-specific error estimate.

Compare Fobserved to Fcrit

11A-25 2007 A. Karpinski


o As an example, lets consider the following question as a post-hoc test
Does training improve test scores for subscale 2?

When we tested this contrast, we found the test statistic to be:


F(1,9) = 19.74

We need to discard the p-value and compute a Tukey adjusted critical


value.
q(1-,a,) with = .05 a =6 =9
q(.95,6,9) = 5.02 Fcrit =
(5.02)2 = 12.60
2

And so we can report the test to be significant at the = .05 level:


F(1,9) = 19.74, p < .05

Test Score Improvement

80

70
Test Score

Pre-test
60
Post-test
50

40
Scale 1 Scale 2 Scale 3
Scale

o We should have a separate variance estimate of the error bars for each
cell
s 2jk
StdError =
n
o SPSS and EXCEL are not good for plotting separate error bars for each
cell. The best you can do is to compute a common standard error based
on the error term for the highest order interaction. This error bar is
misleading (because you did not actually use it in your analyses). If you
plan to publish using repeated measures data, get better graphical
software.

11A-26 2007 A. Karpinski


5. An Example

Consider an experiment on facial perception. Faces vary on two


dimensions: orientation (upright, 90 rotation, and 180 rotation) and
distortion (none, eyes & mouth upside down, eyes whitened and teeth
blackened). Participants rate each of the six resulting faces on how bizarre
each face looks on a 7-point scale, with higher numbers indicating more
bizarreness. The following data were obtained:

No Distortion Upside-Down Whitened and Blackened


Subject 0 90 180 0 90 180 0 90 180
1 1.18 2.40 2.48 4.76 4.93 3.13 5.56 4.93 5.21
2 1.14 1.55 1.25 4.81 4.73 3.89 4.85 5.43 4.89
3 1.02 1.25 1.30 4.98 3.85 3.05 4.28 5.64 6.49
4 1.05 1.63 1.84 4.91 5.21 2.95 5.13 5.52 5.69
5 1.81 1.65 1.01 5.01 4.18 3.51 4.90 5.18 5.52
6 1.69 1.67 1.04 5.65 4.56 3.94 4.12 5.76 4.99

Distortion 0 90 180
None 1.32 1.69 1.49 1.50
Upside-Down 5.02 4.58 3.41 4.34
Whitened/Blackened 4.81 5.41 5.47 5.23
3.72 3.89 3.46 3.69

None
3 Upside-Down
Whitened/Blackened

0
0 90 180
Rotation

11A-27 2007 A. Karpinski


We can solve for the (fixed) model parameters
Yijk = + j + k + + ( ) jk + ( ) + ( ) + ( )

The overall mean of the scores


= 3.69

j The effect of being in level j of Orientation


j = Y . j . Y ...
1 = 3.72 3.69 = 0.03
2 = 3.89 3.69 = 0.20
3 = 3.46 3.69 = 0.23

k The effect of being in level k of Distortion


k = Y ..k Y ...
1 = 1.50 3.69 = 2.19
2 = 4.34 3.69 = 0.65
3 = 5.23 3.69 = 1.54

( ) jk The effect of being in level j of Orientation and


level k of Distortion
( ) jk = Y . jk Y . j . Y ..k + Y ...

( )11 = 1.32 3.72 1.50 + 3.69 = 0.21


( )12 = 5.02 3.72 4.34 + 3.69 = 0.65
( )13 = 4.81 3.72 5.23 + 3.69 = 0.45
.
.
( )33 = 5.47 3.46 5.23+ 3.69 = 0.47

First, we need to check assumptions


o This design is a two-factor repeated measures design
o Participants must be independent and randomly selected from the
population
o Normality/ symmetry of difference scores (but in practice normality
within each condition)

11A-28 2007 A. Karpinski


EXAMINE
VARIABLES=nod_zer nod_90 nod_180 usd_zer usd_90 usd_180 wb_zer wb_90 wb_180
/PLOT BOXPLOT STEMLEAF NPPLOT
/COMPARE VARIABLES.

Tests of Normality
6
Shapiro-Wilk 6

Statistic df Sig. 5

NOD_ZER .807 6 .068


4
NOD_90 .831 6 .109
NOD_180 .846 6 .147 3

USD_ZER .771 6 .032 1

2
USD_90 .980 6 .950
USD_180 .872 6 .235 1
3

WB_ZER .954 6 .776


0
WB_90 .955 6 .779 N= 6 6 6 6 6 6 6 6 6

NOD_ZER NOD_180 USD_90 WB_ZER WB_180


WB_180 .909 6 .432 NOD_90 USD_ZER USD_180 WB_90

o Sphericity

Mauchly's Test of Sphericity

Measure: MEASURE_1

Epsilon
Greenhous
Within Subjects Effect e-Geisser Huynh-Feldt Lower-bound
ORIENTAT .963 1.000 .500
DISTORT .932 1.000 .500
ORIENTAT * DISTORT .462 .720 .250

Sphericity is not satisfied.


We must either conduct only non-parametric tests or contrasts

Contrasts of interest (all post-hoc):


o Are there linear (and quadratic) trends in the marginal orientation means?
o Are there linear (and quadratic) trends in the orientation means within
each level of distortion?
o Are the linear (and quadratic) trends in the orientation means within each
level of distortion different from each other?

11A-29 2007 A. Karpinski


First, lets test for linear and quadratic trends in orientation.

Distortion 0 90 180
None
Upside-Down
Whitened/Blackened
Linear -1 0 +1
Quadratic +1 -2 +1
Marginal Orientation Means

4
Rating of Bizarreness

3.75

3.5

3.25

3
0 90 180
Orientation

o Method 1: Use SPSSs built-in contrasts to test main effect contrasts


GLM nod_zer usd_zer wb_zer nod_90 usd_90 wb_90
nod_180 usd_180 wb_180
/WSFACTOR = orientat 3 polynomial distort 3
/PRINT = DESCRIPTIVE.
Tests of Within-Subjects Contrasts

Measure: MEASURE_1
Type III Sum
Source orientat distort of Squares df Mean Square F Sig.
orientat Linear .606 1 .606 5.772 .061
Quadratic 1.143 1 1.143 13.537 .014
Error(orientat) Linear .525 5 .105
Quadratic .422 5 .084

SS Linear .606
Linear
2
= = = .54
SS Linear + SS ErrorTermForLinear .606 + .525
1.143
Quadratic
2
= = .73
1.143 + .422

These are complex, post-hoc tests, so a Scheff correction is required.


FCrit = 2 * F (.05,2,5) = 2 * 5.78 = 11.57
FCrit = 2 * F (.10,2,5) = 2 * 3.78 = 7.56

Linear trend: F (1,5) = 5.78, ns, 2 = .54


Quadratic trend: F (1,5) = 13.54, p < .05, 2 = .73

11A-30 2007 A. Karpinski


o Method 2: Compute and test the contrasts manually
compute lin_ori = -1*nod_zer + 0*nod_90 +1*nod_180
- 1*usd_zer + 0*usd_90+ 1* usd_180
- 1*wb_zer + 0*wb_90+1* wb_180.
compute quad_ori = 1*nod_zer -2*nod_90 +1*nod_180
+ 1*usd_zer -2*usd_90+1* usd_180
+ 1*wb_zer -2*wb_90+1* wb_180.

T-TEST /TESTVAL=0
/VARIABLES=lin_ori quad_ori.

One-Sample Statistics

Std. Error
N Mean Std. Deviation Mean
LIN_ORI 6 -.7783 .79356 .32397
QUAD_ORI 6 -1.8517 1.23274 .50327

One-Sample Test

Test Value = 0
95% Confidence
Interval of the
Mean Difference
t df Sig. (2-tailed) Difference Lower Upper
LIN_ORI -2.402 5 .061 -.7783 -1.6111 .0545
QUAD_ORI -3.679 5 .014 -1.8517 -3.1454 -.5580

These results exactly match the results obtained from using SPSSs
built-in main effect contrasts.

2
t Contrast 2.402 2 3.679 2
rlinear = = = .73 rquad = = .85
2
t Contrast + df contrast 2.404 2 + 5 3.679 2 + 5

Linear trend: F (1,5) = 5.78, ns, r = .73


Quadratic trend: F (1,5) = 13.54, p < .05, r = .85

11A-31 2007 A. Karpinski


Second, lets test for linear and quadratic trends in orientation within each
level of distortion.

Distortion 0 90 180
None Linear -1 0 +1
Quadratic +1 -2 +1
Upside-Down
Whitened/Blackened

4
Bizarreness

None
3 Upside-Down
Whitened/Blackened

0
0 90 180
Rotation

11A-32 2007 A. Karpinski


Distortion 0 90 180
None
Upside-Down Linear -1 0 +1
Quadratic +1 -2 +1
Whitened/Blackened

4
Bizarreness

None
3 Upside-Down
Whitened/Blackened

0
0 90 180
Rotation

Distortion 0 90 180
None
Upside-Down
Whitened/Blackened Linear -1 0 +1
Quadratic +1 -2 +1

4
Bizarreness

None
3 Upside-Down
Whitened/Blackened

0
0 90 180
Rotation

11A-33 2007 A. Karpinski


o Method 1: Compute and test the contrasts manually
Compute lin_nod = -nod_zer + 0*nod_90 + nod_180.
Compute quad_nod = nod_zer - 2*nod_90 + nod_180.

Compute lin_usd = -usd_zer + 0*usd_90 + usd_180.


Compute quad_usd = usd_zer - 2*usd_90 + usd_180.

Compute lin_wb = -wb_zer + 0*wb_90 + wb_180.


Compute quad_wb = wb_zer - 2*wb_90 + wb_180.

T-TEST /TESTVAL=0
/VARIABLES=lin_nod quad_nod lin_usd quad_usd lin_wb quad_wb .
One-Sample Statistics

Std. Error
N Mean Std. Deviation Mean
LIN_NOD 6 .1717 .81121 .33117
QUAD_NOD 6 -.5817 .33030 .13484
LIN_USD 6 -1.6083 .38039 .15529
QUAD_USD 6 -.7217 1.28395 .52417
LIN_WB 6 .6583 .87894 .35883
QUAD_WB 6 -.5483 1.13125 .46183

One-Sample Test

Test Value = 0
95% Confidence
Interval of the
Mean Difference
t df Sig. (2-tailed) Difference Lower Upper
LIN_NOD .518 5 .626 .1717 -.6796 1.0230
QUAD_NOD -4.314 5 .008 -.5817 -.9283 -.2350
LIN_USD -10.357 5 .000 -1.6083 -2.0075 -1.2091
QUAD_USD -1.377 5 .227 -.7217 -2.0691 .6258
LIN_WB 1.835 5 .126 .6583 -.2641 1.5807
QUAD_WB -1.187 5 .288 -.5483 -1.7355 .6388

2
t Contrast 0.518 2 4.314 2
rLinearNoDistort = = = .23 rQuadNoDistort = = .89
2
t Contrast + df contrast 0.518 2 + 5 4.314 2 + 5
10.357 2 1.377 2
rLinearUpsideDown = = .98 rQuadUpsideDown = = .28
10.357 2 + 5 1.377 2 + 5
1.835 2 1.187 2
rLinearWhitenedBlacked = = .63 rQuadWhitenedBlackened = = .47
1.835 2 + 5 1.187 2 + 5

o These are complex, post-hoc tests, so a Scheff correction is required.


FCrit = 4 * F (.05,4,5) = 4 * 5.19 = 20.76 FCrit = 4 * F (.10,4,5) = 4 * 3.52 = 14.08
t crit = Fcrit = 20.76 = 4.56 t crit = Fcrit = 14.08 = 3.75

11A-34 2007 A. Karpinski


o For faces that were not distorted: There is a marginally significant
quadratic trend such that sideways faces are rated to be most bizarre and
deviations from 90 are less bizarre, t (5) = 4.31, p < .10, r = .89 .
o For faces with upside-down mouths and faces: There is a linear trend in
ratings of bizarreness such that as orientation increases, bizarreness
decreases, t (5) = 10.36, p < .05, r = .98
o For faces with whitened eyes and blacked teeth: Ratings of bizarreness
are unaffected by orientation, rs .63 .

4
Bizarreness

None
3 Upside-Down
Whitened/Blackened

0
0 90 180
Rotation

11A-35 2007 A. Karpinski


o Method 2: Selecting only the groups of interest and running a contrast on
those groups

No Distortion: Linear and quadratic trends


GLM nod_zer nod_90 nod_180
/WSFACTOR = orientat 3 Polynomial.
Tests of Within-Subjects Contrasts

Measure: MEASURE_1
Type III Sum
Source orientat of Squares df Mean Square F Sig.
orientat Linear .088 1 .088 .269 .626
Quadratic .338 1 .338 18.608 .008
Error(orientat) Linear 1.645 5 .329
Quadratic .091 5 .018

.08841 .338
LinearNoDi
2
stortion = = .05 QuadraticN
2
oDistortion = = .79
.08841 + 1.645 .338 + .09091

Upside-down eyes and mouths: Linear and quadratic trends


GLM usd_zer usd_90 usd_180
/WSFACTOR = orientat 3 Polynomial.
Tests of Within-Subjects Contrasts

Measure: MEASURE_1
Type III Sum
Source orientat of Squares df Mean Square F Sig.
orientat Linear 7.760 1 7.760 107.262 .000
Quadratic .521 1 .521 1.896 .227
Error(orientat) Linear .362 5 .072
Quadratic 1.374 5 .275

7.76 .521
LinearUpsi
2
deDown = = .96 QuadraticU
2
psideDown = = .27
7.76 + .362 .521 + 1.374

Whitened eyes and blackened mouths: Linear and quadratic trends


GLM wb_zer wb_90 wb_180
/WSFACTOR = orientat 3 Polynomial.
Tests of Within-Subjects Contrasts

Measure: MEASURE_1
Type III Sum
Source orientat of Squares df Mean Square F Sig.
orientat Linear 1.300 1 1.300 3.366 .126
Quadratic .301 1 .301 1.410 .288
Error(orientat) Linear 1.931 5 .386
Quadratic 1.066 5 .213

1.300 .301
LinearWhit
2
enedBlackened = = .40 QuadraticW
2
hitenedBlackened = = .22
1.300 + 1.931 .301 + 1.066

11A-36 2007 A. Karpinski


Finally, wed like to test for differences between these trends
o For example, does the linear trend for no distortion differ from the linear
trend for upside-down eyes and mouth?

o Method 1: Compute and test the contrasts manually


Linear (No distortion) vs. Linear (Up-side down)
Distortion 0 90 180
None Linear -1 0 +1
Upside-Down Opposite Linear +1 0 -1
Whitened/Blackened

Quadratic (No distortion) vs. Quadratic (Up-side down)


Distortion 0 90 180
None Quadratic +1 -2 +1
Upside-Down Opposite Quadratic -1 +2 -1
Whitened/Blackened

The syntax and output for this method is not included here.

4
Bizarreness

None
3 Upside-Down
Whitened/Blackened

0
0 90 180
Rotation

11A-37 2007 A. Karpinski


o Approach 2: Use SPSSs built-in contrasts
Notice that differences in trends can be obtained by examining the
interaction between polynomial contrasts on orientation and simple
contrasts on distortion

Linear (orientation) by None vs.Upside-down (distortion)


Distortion 0 90 180
None -1 0 +1 +1
Upside-Down +1 0 -1 -1
Whitened/Blackened
-1 0 +1

Quadratic (orientation) by None vs.Upside-down (distortion)


Distortion 0 90 180
None +1 -2 +1 +1
Upside-Down -1 +2 -1 -1
Whitened/Blackened
+1 -2 +1

GLM nod_zer usd_zer wb_zer nod_90 usd_90 wb_90 nod_180 usd_180 wb_180
/WSFACTOR = orientat 3 Polynomial distort 3 Simple(1)
/PRINT = DESCRIPTIVE.
Tests of Within-Subjects Contrasts

Measure: MEASURE_1
Type III Sum
Source ORIENTAT DISTORT of Squares df Mean Square F Sig.
ORIENTAT Linear .202 1 .202 5.772 .061
Quadratic .381 1 .381 13.537 .014
Error(ORIENTAT) Linear .175 5 3.499E-02
Quadratic .141 5 2.814E-02
DISTORT Level 2 vs. Level 1 145.010 1 145.010 382.927 .000
Level 3 vs. Level 1 250.358 1 250.358 621.185 .000
Error(DISTORT) Level 2 vs. Level 1 1.893 5 .379
Level 3 vs. Level 1 2.015 5 .403
ORIENTAT * DISTORT Linear Level 2 vs. Level 1 9.505 1 9.505 20.243 .006
Level 3 vs. Level 1 .711 1 .711 .755 .425
Quadratic Level 2 vs. Level 1 1.960E-02 1 1.960E-02 .082 .787
Level 3 vs. Level 1 1.111E-03 1 1.111E-03 .004 .951
Error(ORIENTAT*DIST Linear Level 2 vs. Level 1 2.348 5 .470
ORT) Level 3 vs. Level 1 4.705 5 .941
Quadratic Level 2 vs. Level 1 1.201 5 .240
Level 3 vs. Level 1
1.338 5 .268

FCrit = 4 * F (.05,4,5) = 4 * 5.19 = 20.76 FCrit = 4 * F (.10,4,5) = 4 * 3.52 = 14.08


t crit = Fcrit = 20.76 = 4.56 t crit = Fcrit = 14.08 = 3.72

11A-38 2007 A. Karpinski


Difference in trends between no distortion and upside-down eyes and
mouth:
9.505 .0196
LinearDiff
2
= = .80 QuadraticD
2
iff = = .02
9.505 + 2.348 .0196 + 1.201
Difference in linear trends: F (1,5) = 20.24, p < .10, 2 = .80
Difference in quadratic trends: F (1,5) = 0.08, ns, 2 = .02

Difference in trends between no distortion and whitened eyes and


black teeth:
.711 .00111
LinearDiff
2
= = .13 QuadraticD
2
iff = = .0008
.711 + 4.705 .00111 + 1.338
Difference in linear trends: F (1,5) = 0.77, ns, 2 = .13
Difference in quadratic trends: F (1,5) = 0.01, ns, 2 < .01
6
6

5
5

4
4
Bizarreness
Bizarreness

None
None
3 Upside-Down
3 Upside-Down
Whitened/Blackened
Whitened/Blackened

2
2

1 1

0 0
0 90 180 0 90 180
Rotation Rotation

To obtain differences in trends between upside-down eyes and mouth


and whitened eyes and black teeth, we need to run another analysis:
GLM nod_zer usd_zer wb_zer nod_90 usd_90 wb_90 nod_180 usd_180 wb_180
/WSFACTOR = orientat 3 Polynomial distort 3 Simple(2)
/PRINT = DESCRIPTIVE.

Difference in linear trends: F (1,5) = 24.18, p < .05, 2 = .83


Difference in quadratic trends: F (1,5) = 0.04, ns, 2 < .01
6

4
Bizarreness

None
3 Upside-Down
Whitened/Blackened

0
0 90 180
Rotation

11A-39 2007 A. Karpinski


All of the previous analysis examined the effect of orientation or the effect
of orientation within each level of distortion.
Alternatively, we may be interested in the effect of distortion or the effect of
distortion within each level of orientation. The following analysis are a few
examples of these types of contrasts

Distortion 0 90 180
None 1.32 1.69 1.49 1.50
Upside-Down 5.02 4.58 3.41 4.34
Whitened/Blackened 4.81 5.41 5.47 5.23
3.72 3.89 3.46 3.69

o Within each level of orientation, lets compare the distorted faces to the
non-distorted control.
compute comp1 = usd_zer - nod_zer.
compute comp2 = wb_zer - nod_zer.

compute comp3 = usd_90 - nod_90.


compute comp4 = wb_90 - nod_90.

compute comp5 = usd_180 - nod_180.


compute comp6 = wb_180 - nod_180.

T-TEST /TESTVAL=0
/VARIABLES=comp1 to comp6.
One-Sample Statistics

Std. Error
N Mean Std. Deviation Mean
COMP1 6 3.7050 .29187 .11916
COMP2 6 3.4917 .71065 .29012
COMP3 6 2.8850 .42505 .17353
COMP4 6 3.7183 .64691 .26410
COMP5 6 1.9250 .90697 .37027
COMP6 6 3.9783 .82956 .33867

One-Sample Test

Test Value = 0
95% Confidence
Interval of the
Mean Difference
t df Sig. (2-tailed) Difference Lower Upper
COMP1 31.093 5 .000 3.7050 3.3987 4.0113
COMP2 12.035 5 .000 3.4917 2.7459 4.2374
COMP3 16.626 5 .000 2.8850 2.4389 3.3311
COMP4 14.079 5 .000 3.7183 3.0394 4.3972
COMP5 5.199 5 .003 1.9250 .9732 2.8768
COMP6 11.747 5 .000 3.9783 3.1078 4.8489

11A-40 2007 A. Karpinski


These are pair-wise posthoc comparisons, so a Tukey HSD correction
is required.
qcrit (.05,9,5) 6.80
tcrit = = = 4.81
2 2
Within each level of orientation, all distorted faces are rated as more
bizarre than the control, non-distorted faces, all ps <.05, ds > 2.12.
o We decide to follow these tests up with pair wise comparisons between
the two distorted faces at each level of orientation.
compute comp7 = usd_zer - wb_zer.
compute comp8 = usd_90 - wb_90.
compute comp9 = usd_180 - wb_180.
T-TEST /TESTVAL=0
/VARIABLES=comp7 to comp9.
One-Sample Statistics

Std. Error
N Mean Std. Deviation Mean
COMP7 6 .2133 .80746 .32964
COMP8 6 -.8333 .64242 .26227
COMP9 6 -2.0533 .95007 .38786

One-Sample Test

Test Value = 0
95% Confidence
Interval of the
Mean Difference
t df Sig. (2-tailed) Difference Lower Upper
COMP7 .647 5 .546 .2133 -.6340 1.0607
COMP8 -3.177 5 .025 -.8333 -1.5075 -.1592
COMP9 -5.294 5 .003 -2.0533 -3.0504 -1.0563

Again, a Tukey HSD correction is required.


qcrit (.05,9,5) 6.80
tcrit = = = 4.81
2 2

When faces are presented upside-down, then the faces with eyes
whitened and teeth blacked are rated as more bizarre than faces with
up-side down eyes and mouth, t (5) = 5.29, p < .05, d = 2.16
In other orientations (upright and 90), there are no significant
differences in ratings of the two distorted faces, ds < 1.29.
Distortion 0 90 180
None 1.32 1.69 1.49 1.50
Upside-Down 5.02 4.58 3.41 4.34
Whitened/Blackened 4.81 5.41 5.47 5.23
3.72 3.89 3.46 3.69

11A-41 2007 A. Karpinski

You might also like