DOCUMENT RESUME ED 064 343 Draper, John For Jr. A Monte ... · DOCUMENT RESUME ED 064 343 TM 001...
Transcript of DOCUMENT RESUME ED 064 343 Draper, John For Jr. A Monte ... · DOCUMENT RESUME ED 064 343 TM 001...
DOCUMENT RESUME
ED 064 343 TM 001 508
AUTHOR Draper, John For Jr.TITLE A Monte Carlo Investigation of the Analysis of
Variance Applied to Non-Independfunt BernoulliVariates.
PUB DATE Apr 72NOTE 40p.; Paper presented at the annual meeting of the
American Educational Research Association (Chicago,Illinois, April 1972)
EDRS PRICE MF-$0.65 HC-$3.29DESCRIPTORS *Analysis of Variance; *Factor Analysis;
*Mathematical Models; Research Methodology;*Statistical studies; *Tables (Data)
ABSTRACTThe applicability of the Analysis of Variance, ANOVA,
procedures to the analysis of dichotomous repeated measure data isdescribed. The design models for which data were simulated in thisinvestigation were chosen to represent simple cases of twoexperimental situations; situation one, in which subjects' responsesto a single randomly selected set of three stimuli or items wereevaluated dichotomously on four successive occasions, and situationtwo, in which subjects' responses ar3 evaluated dichotomously on foursuccessive occasions with respect to four separate randomly selectedset of three stimuli or items. The two experimental situations arerepresented by "three way factorial" designs with random factors forsubjects ani items and a fixed factor for occasions; the first withsubjects, items and occasions completely crossed and the second withitems and occasions crossed with subjects, but with items nestedwithin occasions. Results of the study are presented in 11 tables and7 figures. (Author/DB)
U.S. DEPARTMENT OF HEALTH.EDUCATION & WELFAREOFFICE OF EDUCATION
THIS DOCUMEAT HAS BEEN REPRO-DUCED EXACTLY AS RECEIVED FROMTHE PERSON OR ORGANIZATION ORIG-INATING IT POINTS OF VIEW OR OPIN-IONS STATED DO NOT NECESSARILYREPRESENT OFFICIAL OFFICE OF EDU-CATION POSITION OR POLICY.
A MONTE CARLO INVESTIGATION OF THE ANALYSIS OFVARIANCE APPLIED TO NON-INDEPENDENT BERNOULLI VARIATES
By
John F. Draper, Jr.CTB/McGraw-Hi11.
Paper presented at the annualmeeting of the American Educational Research Association
April, 1972
The problem with which this paper will be concerned is the applicability
of the Analysis of Variance, ANOVA, procedures to the analysis of dichotomous
repeated measure data. Recall, that in the employment of ANOVA procedures it
is assumed that the data can be reasonably modeled by a vector variate with
multi-variate normal density and diagonal or patterned "compound symmetric"
covariance matrix, that is with equal variances and equal covariances in
diagonal blocks for subjects luld zeros elsewhere, (Greenhouse dad Geisser,
1959). Can these familiar procedures still be employed when the data can
more reasonably be modeled by a vector variate with non-independent Bernoulli
variates as elements? Other investigators, Mandeville (1970) and Seegar"and
Gabrielson (1968) have employed Monte Carlo methods in order to achieve a
partial answer to the above question. Their results tend toward a cautious
but affirmative answer with respect to the design models and model parameter
configurations with which their simulations were concerned. The results re-
ported in this paper are less sanguine. That is, however, not.ta say that
the results reported in this paper are in disagreement with those of previous
investigatOrs, since the design models and model parameter configurations
with which this paper is concerned are somewhat different. Bradley (1968)
has indicated that the robustness of a parametric statistical test is idio-
syncratic rather than general. For example, the F-test has bean shown to be-
robust to heteroscedasiticity for balanced designs (equal cell n's), but
generally not rdbust to heteroscedasiticity for unbalanced Llesigns.
The design models for which datA were simulated in this investigation
were dhosen to represent simple cases of two experimental situations, situation
one, in which subjects responses to a single randomly selected set of three
stimuli or items were evaluated dichotomously on four successive occasions
and situation two, in which subjects' responses are evaluated dichotomously
on four successive occasions with respect to four separate randomly selected
sets of three stimuli or items. Table 1 contrasts these two experimental
situations.
Table 1. Contrast of experimental Situation 1 with Experimental Situation 2.
Situation 1
Occasion 01
02 03 04
Item 11 1213
111213 11 12 1 3
11 1213
Situation 2
Occasion 01
02 03 04
Item 111213 1415 16 17 18 19 1
10111
112
In terms of experimental design the two experimental situations can be represented
by "three way factorial" designs with random factors for subjects and items1
and a fixed factor for occasions; the first with subjects, items an4 occasions
completely crossed and the second with items and pccasions crossed with subjects,
but with items nested within occasions.
IMFIN.1Items were considered to represent a random source of variation, sinceexperimenters often wish to generalize their results to some domain ofitems or stimuli, rather than to restrict their findings to only thoseitems they actually employed.
3
These two experimental situations were considered to be of interest because
experimenters often form occasion measures for subjects as the sum of dichotomous
response evaluations to a series of items and then ignore items as a factor in
their experimental design. And because it may be shown, (Draper, 1971), that
for such experimental situations, there is a source of variation associated
with items2 which if nonrnull will be confounded-with the source of variation
for occasions. A situation which holds whether or not, items are included as
a factor in the experimental design and an ANOVA consistent with the design
is employed for analysis and which also holds whether the dependent measures,
or item scores, are dichotomous or have distributions which may be reasonably
approximated by a normal probability density.
For situation one an examination of expected mean squares, suggested
possible variance ratio tests for sources of variation due to subjects, occasions,
items, subjects by occasions interaction, and items by occasions interaction.
Further, since the items by occasions interaction could be nonnull and would
thus be confounded with occasions, a QuasiF variance ratio test for occasions
(Satterthwaite, 1941) suggested itself. Similarly for situation two possible
variance ratio tests were suggested for the sources of variation, subjects,
occasions, items, and subjects.by occasions interaction, where both a "regular"
variance ratio test and a "QuasiF" test were suggested to test for occasions
effects. Calculations with respect to all the suggested tests were made on
simulated data, but of major interest were those tests for occasions and the
subjects by occasions interaction.
2An item by occasion interaction effect in situation one or a main effectdue to items in situation VATO.
It was considered of interest to vary the following model parameters over
examples of the experimental situations to be investigated.
1) The base probability of one3, on item difficulty, in the data,
2) The number of subjects,
3) The degree of relative heterogenity of the subjects4,
4) The effect of items4,
5) The effect of occasions4,
and
6) The effect of subject by occasion interaction4.
Three different values of the base probability of a one, .5, .2, and .1, suggested
by results due to Lunney (1969), were investigated as were five values of the
number of subjects 4, 6, 8, 10, and 12. The four levels of relative subject
heterogenity, Levels 1, 2, 3, and 4, which were employed, correspond to four
ratios of subject to error variance, 2:8, 3:7, 4:6 and 5:5 respectively. Com-
binations of these parameter values in combination with the values for effects
due items, occasions and subject by occasion interaction resulted in 720 different
model parameter configurations for data were simulated. The data for
the model parameter configurations were gimulated as pseudo-random numbers
with distributions which could reasonably be approximated by a normal proba-
bility density function and all ANOVA statistics calculated, then the data
were dichotomized and the ANOVA calculations repeated. Determinations were
then made as to which statistics fell to various a rejection regions and counters
which had been initialized at zero were incremented by one where appropriate.
3The base probability of a one will refer to the general probability of aone in the absence of possible main and interaction effects.
4Whether null or non-null.
5
This process was repeated 999 times for each parameter configuration for a
total of 1,000 simulations.
As an example of the type of analysis which was done on the frequency
data obtained from the simulation runs consider the following. Table 2 contains
frequencies in a = .05, .025 and .01 rejection regions of variance ratio tests
of occasion effects under conditions in which the occasion effects were null,
items were crossed with occasions and there was no interaction between subjects
and occasions. Table 2 is laid out as a six-dimensional array with three left
margins and three upper margins. The left-most margin indicates the number of
subjects employed with respect to a given simulation of data. Proceeding from
left to right the next margin indicates the nature of the items, whether null
in effect or non-null, and the right-most indicates the level of subject
heterogeneity. The upper margins from top to bottom indicate: (1) the probability
of a one with respect to a given simulation df data, (2) one thousand times
nominal a or the expected frequency given a true F-variate, and (3) an indication
of whether the 1,000 variance ratios with respect to a given cell of the table
were calculated from the dependent variables when they were variates with
approximate normal density (N) or from the subsequently dichotomized "normals"
(D).
In order to test for trends in the data reported in tables, the margins
were considered as fixed sources of variaeon and a multivariate analysis of
variance, MANOVA, was performed on the frequency data three-tuples5 within the
table, employing the highest order interaction mean products as an estimate of
error under the assumption that the highest order interaction is truly null6.
5For a = .05, .025, and .01.
6An assumption checked by X2 tests separately on each interaction varianceestimate.
An initial analysis of the data in each of the tables indicated that the
frequencies for tests based on the dichotomous variates were in each case
significantly different from the frequencies for tests based on the normal
variates.
A more detailed account of the results may be found in the hand out. However.
without extensive reference to the results Figures 1. and 2. tell most of the
story with respect to the "regular" variance ratio test of occasions under
null occasion effect conditions. Note the interaction of the effects due to
the number of subjects and the probability of a one in both figures. Note also
the strong effzet due to nm-null items in Figure 2. The effect of the non-
null items condition confirms the previously indicated confounding of a non-
null items effect with occasions.
An examination of Tables 4. and 5. indicate something about the behavior
of the Quasi-F test under null occasions effects. Note that its behavior is not
particularly good, even when based on the "normal" data.. Tables 6. through 9.
contain the most startling results. Note that the Quasi-F tests based on
normal variates responded in at, appropriate manner to non-null occasion effects.
However, the Quasi-F tests based on dichotomous data had significantly fewer
frequencies in the a rejection regions when the effect being tested was non-null
than they did when the effect being tested was null! Thus it would appear
that Quasi-F tests based on dichotomous data, such as that simulated in this
study, are biased tests!
Tables 10. and 11. and Figures 6. and 7. axe with respect to.the variance
ratio tests of subject by occasion interactions. .Although the results concerning
the Quasi-F were the most startling of the results of this study, the most
disappointing results were those concerning the tests of subject by occasion
interactions. Examine in Figure 6. how the test becomes increasingly liberal
with increases in the number of subjects under conditions where the probability
of a one is not equal to .5.
In order to briefly examine the implications of this study, consider what
the results might suggest to an experimenter who would like to do a repeated
measures type of experiment and analyze his results with hypothesis testing in
mind. First an experimenter who is considering doing a repeated measures type
of experiment should consider the nature of the items or response evokeis that
will be employed in hir experiment, examine them for possible confounding,
and then employ a design and analysis which includes the items as a factor in it.
If an experimenter must have confounding in the analysis consi.stent with
the design of his repeated measures experiment, he is in a somewhat difficult
position with regard to analysis of variance testing of the source of variation
with which confounding is present. For even if he can expect his dependent
variables to have an approximately normal distribution, the results of this
study suggest that the Quasi-F test will not have particularly good properties.
If, on the other hand, the experimenter must evaluate responses in such a manner
that his dependent variables are dichotomous, the Quasi-F test appears to be
completely unacceptable.
On the other hand, if an experimenter finds that he can expect to have no
confounding such as that mentioned above, but must have dichotomous dependent
variables the results suggest, he can expect to appropriately employ the
ordinary analysis of variance, variance ratio test for the occasion effect, if
he employs enough subjects. The results also suggest that an experimenter
should expect the power of analysis of variance tests based on dichotomous
data to be between one third and one half the power he could expect if his
dependent variables could be considered normal in -distribution. The practical
suggestion implied by the results of this study is, that if the above experimenter
must have dichotomous dependent variables he should employ a still larger number
of subjects in order to obtain a reasonable power situation.
The results of this study with respect to the test of the subject by.
occasion interaction when based on dichotomous data, suggest that even a very
large variance ratio statistic may not indicate that the null hypothesis is
false if there are a large number of subjects and the probability of a one is
not close to .5 (see again Figure 6). For a probability of a one close to .5,
however, it would appear that the probability of a Type I error is only
approximately 1.2 times the nominal a level. Since the power of the test of
the subject by occasion interaction based on dichotomous data with a .5
probability of a one was greater than half the power of the test based on
normal data, the results suggest that when the probability of a one is close
to .5 the test may be appropriately employed. If the probability of a one
is not close to .5, however, the results suggest the above test may not be
appropriately employed. As a practical suggestion, an experimenter who could
expect possible subject by occasion interaction and who must have dichotomous
dependent variables should endeavor to employ items or stimuli which will give
him in general a .5 probability of a one in this data.
BIBLIOGRAPHY
Bradley, J. V. Distribution-Free Statistical Tests. New York:Prentice-Hall, 1968.
Donaldson, T. S. Power of the F-test for non-rcrmal distributionsand unequal error variances. Rand Corporation research memorandumRH-5072-PR, September, 1966.
Draper, J. F. A Monte Carlo investigation of the analysis of varianceapplied to non-independent Bernoulli variates. Unpublisheddissertation, Michigan State University, 1971.
Greenberger, M. Notes on a new pseudo-random number generator."Journal Assoc. Comp. Mach.," 8, 163-67.
Greenhouse, S. W. and Geisser, S. On methods in the analysis ofprofile data. "Psychometrika," 1959, 24, 95-112.
Hammersley, J. M. and Handscomb, D. C. Monte Carlo Method. New York:John Wiley, 1964.
Hsu, T. C. and Feldt, L. S. The effect of limitations on the number ofcriterion score values on the significance level of the F-test."American Educational Research Journal," 1969, 6, No. 4, 515-28.
Hudson, M. D. and Krutchkoff, R. G. A Monte Carlo investigation ofthe size and power of tests employing Satterthwaite's syntheticmean squares. "Biometrika," 1968, 55, 431-33.
Lunney, G. H. A Monte Carlo investigation of basic analysis ofvariance models when the dependent variable is a Bernoullivariable. Unpublished dissertation, Univ. of Minnesota, 1968.
Mandeville, G. K. An empirical investigation of repeated measuresanalysis of variance for binary data. Paper tnesented at theannual meeting of the American Educational Rz:earch Association,1970.
Marsaglia, G. and Bray, T. A. One line random number generators andtheir use in combinations. "Communications of ACM," 1968, 11,757-59.
Satterthwaite, F. E. Synthesis of variance. "Psychometrica," 1941,6 309-16.
Seeger, P. and Gabrielsson. Applicability of the Cochran Q test andthe F test for statistical analysis of dichotomous data fordwendent samples. "Psychol. Bull.," 1968, 69, 269-77.
10
TABLES AND FIGURES TO ACCOMPANY THE PAPER ENTITLED"A MONTE CARLO INVESTIGATION OF THE ANALYSIS OF
VARIANCE APPLIED TO NON-INDEPENDENT BERNOULLI VARIATES"
By
John F. Draper, Jr.CTB/McGraw-Hill
Paper presented at the annualmeeting of the American Educational Research Association
April, 1972
11
INDEX
TABLES PAGE
1. Contrast of Experimental Situation 1 with Experimental Situation 2. 1
2. The Empirical Frequencies, in a.1000.Rejection Regions, of theVariance Ratio Test for Occasions, Based on Normal and DichotomousData, Items Crossed with Occasions, Interaction and Occasion EffectsBoth Null.
3. The Empirical Frequencies, in a.1000 Rajection Regions, of theVariance Ratio Test for Occasions, Based on the Dichotomous Data,Items Nested within Occasions Interaction and Occasion Effects BothNull.
4. The Empirical Frequencies, in a.1000 Rejection Regions, of the Quasi-F for Occasions, Based on Normal and Dichotomous Data, Items Crossedwith Occasions, Interaction and Occasion Effects Both Null.
5. The Empirical Frequencies, in a.1000 Rejection Regions, of the Quasi-F for Occasions, Based on Dichotomous Data, Items Nested withinOccasions, Interaction and Occasion Effects Both Null.
6. The Empirical Frequencies, in a.1000 Rejection Region, of the Quasi-F for Occasions, Based on Normal and Dichotomous Data, Items Crossedwith Occasions, Occasion Effects Non-null, Interaction Null.
7. The Empirical Frequencies, in a.1000 Rejection Regions, of the Quasi-F for Occasions, Based on Dichotomous Data, Items Nested withinOccasions, Occasion Effects Non-null, Interaction Null.
8. The Empirical Frequencies, in a.1000 Rejection Regions, of the Vari-ance Ratio Test and Quasi-F for Occasions, Based on Normal andDichotomous Data, Items Crossed with Occasions, Twelve Subjects.
9. The Empirical Frequencies, in a.1000 Rejection Regions,ance Ratio Test and Quasi-F for Occasions, Based on theDichotomous Data, Items Nested within Occasions, Twelve
of the Vari-Normal andSubjects.
10. The Empirical Frequencies, in a.1000 Rejection Regions, of the Vari-ance Ratio Test for the Subjects by Occasions Interaction, Based on theNormal and Dichotomous Data, Items Crossed with Occasions, Interactionand Occasions Effects Both Null.
11. The Empirical Fi:equencies, in a.1000 Rejection Regions, of the Vari-ance Ratio Test for the Subjects by Occasions Interaction, Based on theNormal and Dichotomous Data, Items Nested within Occasions, Interactionand Occasions Effects Both Null.
3
4
10
11
13
14
19
20
21
22
FIGURES PAGE
1. The Interaction of the Probability of a One and the Number of Subjects,with Respect to the Data in Table 2., Data for the "Regular" OccasionsTest.
2. The Interaction of the Probability of a One, the Number of Subjects;and Items Null vs. Non-null, with Respect to the Data in Table 3., Data
for the "Regular" Occasions Test.
3. The Interactionwith Respect toOccasions.
4. The Interactionwith Respect toOccasions.
of the Probability of a One and the Number of Subjects,the Data in Table 4., Data for the Quasi-F Test of
of the Probability of a One and Items Null vs. Non-null,the Data in Table 4., Data for the Quasi-F Test of
5. The Interaction of the Probability of a One and the Number of Subjects,with Respect to the Data in Table 5., Data on the Quasi-F Test ofOccasions.
6. The Interaction of the Probability of a One and the Number of Subjects,with Respect to the Data in Table 10., Data on the Test of Subjects byOccasions Interaction.
7. The Interaction of the Probability of a One and Items Null vs. Non-null,with Respect to the Data in Table 10., Data on the Test of Subjects byOccasions Interaction.
BIBLIOGRAPHY
APPENDIX
ii
5
7
15
16
17
23
Table 1. Contrast of Experimental Situation 1 with Experimental Situation 2.
Situation 1
Occasion 01 02
03 04
Item 111213 111213 111213 111213
Situation 2
Occasion 01
02
03
04
Item 111213 141516 171819 110111112
"Re ular" Variance Ratio Test of Occasion Results
With respect only to the frequencies based on dichotomous variates
in Table 2, it was concluded that there were significant main effects due
to the probability of a one, Pr (1), and the number of subjects as well as
an interaction between the two significant main effects. The significant
interaction is represented in Figure 1. In the figure the two horizontal
lines represent .95 probability limits1 for mean frequencies such as those
graphed, given an expected value of 50. Observing Figure 1, it appears
that a favorable comparison of nominal and empirical probabilities of a
Type I error occurred when the probability of a one was .5 and there were
six or more subjects. Also a favorable camparison occurred when the
probability of a one was .2 and there were ten or more subjects. However,
when the probability of a one was .1 no favorable comparisons occurred,
although the graph suggests that a favorable comparison might occur given
more subjects.
Table 3 differs from Table 2 in only two aspects: (1) the values in
the table were obtained by simulating data which could have arisen from
situation 2 rather than situation 1, and (2) no frequencies appear with
respect to tests based on normal variates.
From the analysis of the frequencies in Table 3, it was concluded that
there were significant main effects due to: (1) the probability of a one,
(2) the number of subjects, and (3) items null in effect versus items non-
null in effect. There were also significant interactions between
1See Appendix A
Table 2.
The Empirical Frequencies, in 0'4.1000 Rejection Regions, of the
Variance Ratio Test for Occasions, Based on Normal and Dichotomous Data,
Items Crossed with Occasions, Interaction and Occasion Effects Both Null
piob
abili
ty o
f a
One
.2.1
Of
1000
5025
1050
2510
5025
10
Dic
hoto
mou
s or
Rom
a11
BD
IID
BD
110
8D
IIla
A13
1111
NN
o. o
fSu
b.*.
II
Item
sI
Lev
el o
fSu
b. B
et.
151
4426
207
841
4121
2412
1141
6715
427
15nu
ll2
4755
2827
99
2233
917
76
1053
525
414
339
4618
274
1614
566
332
1512
1.: C
211
27
44
4545
2223
010
1255
637
324
949
515
51
11
4750
2825
311
2756
930
119
2644
1420
712
non-
null
240
5220
332
1429
4*7
154
1015
44ls
:15
*9
327
5018
296
1127
5111
253
1413
561
141
64
3845
1530
717
2040
624
017
1249
419
09
151
4025
238
850
3216
157
642
:a0
1317
97
null
266
5031
226
630
3219
146
637
246
126
33
5251
3428
1014
3534
2012
105
2229
314
i)2
64
6157
3827
1412
3932
2115
28
lb25
612
05
148
5522
285
1241
2523
1415
522
2314
117
3no
n-nu
ll2
4649
2127
610
3029
1216
57
1529
912
74
353
3426
149
930
3117
196
1028
240
113
44
5138
1513
56
3328
2013
35
3222
114
01
148
3630
1911
834
3616
257
431
2919
141
4nu
ll2
4739
2227
89
4241
1831
104
3045
2230
C4
359
4729
2614
1335
4117
3212
640
388
300
0e
461
4638
2314
1129
4413
266
816
480
340
S1
5352
2630
1011
5234
1524
74
3945
1931
97
non-
null
247
6216
355
932
4316
283
216
5416
3d4
03
6651
4037
199
5532
727
33
2654
937
24
457
3830
2917
239
4422
336
728
45b
27i
S
146
5530
299
1536
4921
265
162*
*922
2513
10nu
ll2
4246
1621
510
4559
1928
413
4257
le30
64
355
6332
285
1644
6515
363
1425
513
230
1110
453
5216
287
1364
5316
294
1423
651
401
131
5160
2140
1222
3246
1922
19
4156
3335
12Id
non-
null
257
5530
2912
1747
6220
358
16 '
3643
1225
715
353
5529
348
1759
6423
276
1323
553
200
94
5259
2937
1221
3059
1840
514
2343
035
012
146
5328
2711
1344
6223
352
860
5319
364
9nu
ll2
4563
2539
1516
7049
2829
74
3947
1635
24
365
5523
278
837
5015
264
1136
4814
368
124
6353
2433
1012
5846
2129
95
1750
434
.4
91
5957
2933
7lb
4756
3537
1314
3145
2526
106
non-
null
246
SO18
224
671
6550
3925
1039
4126
29le
33
53B
O28
4312
1860
5433
3422
1845
3125
2211
5'4
4963
3340
148
484e
2130
107
6450
1930
44
.4 A
N, n
40 {
.4.,
1.44
,JA
kdat
-.
V11
1,11
4;M
:7,
4.1
Table 3.
The Empirical Frequencies, in e0,1000 Rejection Regions, of the
Variance Ratio Test for Occasions, Based on the Dichotomous Data, Items Nested
within Occasions, Interaction and Occasion Effects Both Null.
Prob
abili
ty o
f a
One
.5.2
A
or10
0058
2510
5025
1050
2510
No:
of
Lev
el o
fub
item
sSu
b. B
et.
151
26?
4121
1241
157
247
289
227
105
4
nul
l3
39re
414
612
122
2
44
4522
512
63
95
51
7338
1411
561
2244
176
256
3615
7247
1419
50
non-null
374
2314
5926
654
328
462
265
.69
289
2910
0
151
258
5016
742
139
266
316
3019
637
,6
6nul 1
352
3410
3520
1021
30
64
6138
1439
242
le6
01
100
6837
120
6326
6936
21.
299
5620
119
41ta
8751
'6non-null
312
567
2411
948
2410
456
454
134
6419
119
6331
7834
19
-
148
3011
3416
731
191
247
22a
4218
1030
220
null
359
2914
3517
124Q
80
84
6/38
1429
136
160
01
127
6339
131
8638
133
6433
212
570
3015
310
564
82 .
2311
non-null
312
276
3617
111
556
9047
244
115
5631
156
101
5719
571
21
146
309
3821
.5
1822
132
4216
545
194
it218
8nul 1
355
325
4415
325
30
204
5316
764
164
201
1
114
583
4215
511
077
200
122
602
172
106
5820
413
569
153
9649
non-null
316
289
4620
714
167
111
5030
417
310
449
211
148
9410
042
24
146
.20
1 1
4423
a60
194
nu 1
12
4525
1570
287
3916
23
6523
a37
154
3614
a12
463
2410
5621
917
44
120
614
380
250
160
101
177
129
50non-null
218
812
661
253
185
123
213
108
543
202
148
8120
615
912
117
497
484
174
120
8426
816
410
716
090
55
60
50
40
20
Number of Subjects
12 =
RI = 08 =6 7-1 04 .= *
1
0.5 0.2 0.1
Probability of a One
Figure 1. The interaction of the probability of a one and thenumber of subjects, with respect to the data in Table 2., datafor the "regular" occasions test..
'11
the probability of a one and items null vs. non-null and between the
number of subjects and items null vs. non-null. Both of these interactions
were represented on one graph, Figure 2. The horizontal lines in Figure 2
are .95 probability limits about 50, for means such as those graphed. An
inspection of Figure 2 indicates that a favorable comparison of nominal a
and empirical probabilities of a Type I error occurred when items were null
in effect and the probability of a one was .5. Also a favorable camparison
occurred when the items were null, the probability of a one was .2 and there
were 10 or more subjects. In the absence of the above conditions, however,
only unfavorable comparisons resulted.
The marginal mean vectors2 for items null in effect was [40, 18, 6] and
for items non-null it was [131, 78, 42]. The large mean frequencies for items
under non-null conditions confirm the contention that non-null item effects
are confounded with occasions and will cause the regular variance ratio test
for occasions to have too many Type I errors.
In order to examine the behavior of the "regular" variance ratio test
for occasions effects under non-null occasion conditions, an index of
relative pawer was formed. The relative power of the test statistic3 based
on dichamous data was defined to be, the frequency of the test statistic
based on normal non-null data which fell into an a mjection region divided
into the frequency in the same rejection region of that same statistic
based on the same data after it had been subsequently dichotomized. Relative
power values were obtained for the non-null occasion effect analogs of all
20rdered for frequencies in a = [.05, .025, .01] rejection regions.
3For each design model and parameter configuration.
240
200
yr-4lifte
fatio
01111/ord
do,
Oleo
de
or
o,or 144,dror
Oe
mawOOP
fr odogoo
nos or 41°.
Dr Or or OW MO
ow dr 41 olio
4144,111%
elldsallr IMP 4M. dal 0 1111.
4Ildb%el
4144,
or or dor goo or goo dr 44,46. ***.444
.11.1.11111.
Number of Subjects
12 = 6 = 0
10 =0 4=*8 = 0
S.
doe
%0
am? Non-Null
em"-"1 Null
0.5 0.2Probability of a One
Figure 2. The interaction of the probability of a one, the numberof subjects, and items null vs. nonvinull, mlth respect to the datain Table 3., data for the "regular" occasions test.
0.1
of the parameter configurations indicated in Table 2. These data were put
to an arc-sin transformation and then analyzed by means of a MANOVA. This
analysis indicated three significant main effects and no interactions. The
significant effects were (1) the probability of a one, Pr (1), (2) the number
of subjects, and (3) the level of subject heterogeneity. The overall mean
vector of relative power variables was [.45, .38, .32], which indicates that
the relative power of the variance ratio test for occasions effects decreased
as the nominal a level decreased from .05, to .025, to .01. An ANOVA revealed
no interaction between nominal a level and other sources of variation.
For probabilities of a one equal to .5, .2, and .1 the mean relative
powers for a = .05 ware .58, .49, and .28 respectively. For numbers of
subjects 4, 6, 8, 10, and 12, the means were .39, .41, .43, .50 and .54.
For the four levels of subject heterogeneity 1, 2, 3, and 4 the means were
.52, .46, .43, and .41 respectively.
The differences in mean relative power show clear trends. As the prob-
ability of a one becames smaller so does the relative power, which is the
opposite of the trend which might have been expected, since the degree of
non-null effect in the simulated data was selected to counter the effect
of decreased variance corresponding to a decreased pr6bability of a one.
Thus the power of tests based on the dichotamous variates should not have
changed across levels of a probability of a one, whereas the power of the
test based on normals should have and did decrease across levels of a
probability of a one (and decreasing non-null effects). The results
indicate however, that the power of tests based on the dichotomous variates
fell off more rapidly across levels of the probability of a one than did
the power of tests based on the normals. The explanation of the above may
be that as the probability of a one becomes smaller the point of dichoto-
mization is such that more of the "information" carried in the normals is
lost. Another clear trend is that as the number of subjects increases, the
relative power does so as well. The trend with respect to the number of
subjects is most likely a function of the effect of the central limit theorem.
The third trend indicates a loss in relative power with an increase in
subject heterogeneity, a trend for which this investigator has at present
no explanation.
Quasi-F Test of Occasion Results
The most startling result of this study concerned the empirical testing
the application of the Quasi-F test of occasion on dichotomous data.
Tables 4 and 5 contain the frequencies in a = .05, .025, and .01
rejection regions of the Quasi-F ratio t=.-st statistics for tests of
occasion effects when the data were simulated under null occasion effect
conditions and null subject by occasion interaction effect conditions.
That is Tables 4 and 5 are the Quasi-F analogs to Tables 2 and 3.
Again the data in which were with respect to normal variates were
significantly different from the data in Table 4 which were with respect
to dichotomous variates.
The analysis of the data with respect to the dichotomous variates in-
dicated significant effects due to: (1) the probability of a one, Pr (I)
(2) the number of subjects, the interaction of (1) and (2)0 and the inter-
action of (1) and items null vs nm-null respectively. The above two
significant interactions are reptesented in Figures 3 and 4 respectively.
Zs:
°Col
.!d!
.d
Table 4.
The Empirical Frequencies, in 04.1000 Rejection Regions,
of the
Quasi-F for Occasions, Based onNormal and Dichotomous Data, Items Crossed
with Occasions, Interaction and Occasion
Effects Both Null.
Probability of a One
.5
.2
.1
a.1.000
50
25
10
50
25
10
50
25
10
4iehotosous or dorsal
DODNDN
DVIDND
D
.;
No. of
Sub.s
11..svel
of
Items
Sub. Het
175
4537
2515
1245
5923
299
1520
606
332
22nu
l I2
6347
3326
1912
6142
3917
156
1654
42
92
143
6847
3630
128
3066
2027
FS14
646
32
00
d4
451
6426
3210
1727
5610
336
Id6
613
330
141
6857
4325
179
6171
2833
911
355
023
014
non-
null1
255
5728
2715
1234
4726
2313
IC8
595
332
173
7639
2512
128
3152
2322
1310
2252
1127
912
457
4426
2512
1223
6212
383
191
b50
1423
413
178
3945
1818
874
3536
1721
342
5519
2711
5nu
ll2
8134
4713
245
8337
5017
295
4338
1 b
213
3
373
4546
2518
971
. 32
3119
165
2652
82
91
46
487
5645
2619
749
4825
3014
512
487
210,
4i
7739
4120
137
6643
2319
147
4341
721
02
non-
nul li
260
4431
2113
240
4633
2716
650
4522
164
1
383
4349
2126
660
4944
2613
738
3715
130
04
6750
452?
276
5548
3019
123
2439
1610
20
110
224
5712
295
5717
4011
150
4814
306
110
null
281
3052
1730
771
1629
913
041
2621
149
03
7338
4419
148
7225
4815
140
3027
245
I 5
38
467
2347
1723
7dC
2551
1617
037
2520
136
01
6232
3018
16a
51la
348
190
-82
2753
942
0no
n-nu
ll2
7236
4321
2610
5419
385
160
9633
be1
235
03
7727
4210
172
5813
343
90
4634
3815
120
464
1440
516
060
20.
389
190
5321
305
00
198
3759
1925
1094
2976
631
643
2414
37
2nu
ll2
9327
6310
337
9436
538
328
110
3881
053
33
9238
5315
270
104
4251
1229
958
2129
612
310
480
3646
1423
478
3441
531
553
3935
118
41
9234
6212
267
7329
394
52
6231
3310
187
non-
nul I
289
2752
1125
775
3632
715
472
2442
610
33
8429
4412
246
9827
632
392
92:2
051
015
44
823?
5013
IT9
112
2476
953
581
4143
1221
7
113
225
6312
357
9025
59IC
271
137
3354
1641
5
null
283
4653
2027
574
2451
1026
289
2763
222
72
394
2050
923
291
2544
1023
292
2135
1417
64
8028
596
350
7913
397
110
7431
5319
2 7
5
193
3956
1934
366
3537
2426
412
133
7415
400
non-
null
2$4
2241
1024
167
3549
2019
110
117
4910
303
378
4533
2321
48
?37
4721
267
9519
5414
335
471
2147
1027
296
3457
1243
368
3438
1522
4.
11...
-1
Table 5.
The Empirical Frequencies,
0(41000 Rejection Regions, of the
Quasi. -F for Occasions, Based on Dichotomous Data, Items Nestedwithin Oc-
casions, Interaction and Occasion Effects Both Null.
Prob
abilL
q of
a O
ne.5
.2.1
1000
5025
10SO
2510
5025
10
No.
of
Sub.
sIt
ems
Lev
el o
fSu
b. H
et.
4 6 8
10 12
null
non
nti
nu 1
1
non-
nun
huII
bah.
hull
hon-
m41
/
non-
wa
If
1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 1 3 4 1 2 3 4
61 59 62 47 59 68 64 58 77 62 68 85 108 86 84 72 92 73 66 67 102 92 140
119 82 100 87 es 106
102
113
87 101 '78
77 97 123
122
128
174
37 35 29 25 38 40 37 32 41 39 39 43 61 55 30 46 52 44 38 41 65 55 101
85 49 51 48 53 59 65 70 46 67 48 49 69 77 87 09 120
14 17 11 4 14 19 13 17 17 22 15 16 20 28 13 16 23 18 14 16 33 34 51 54 22 23 25 21 32 48 28 20 39 19 27 25 40 68 58 04
27 50 25 24 45 50 42 35 61 68 70 52 67 61 61 64 54 51 65 83 74 0511
3 86 90 78 BO 73 96
134
114
113
102 92 82 85 05 72 81 115
15 32 12 9 25 28 13 14 36 38 31 21 44 35 42 37 19 29 35 47 45 49 62 46 57 49 50 50 63 82 85 75 54 53 47 4$ 63 45 53 60
9 13 6 516 13
5 9 13 20 23 11 21 23 23 21 18 914 24 17 33 38 27 35 28 31 35 52 52 50 44 29 28 26 12 29 .28 37 31
a 6 4 618 27 32 24 51 47 35 16 46 35 34 47 70 54 53 37 44 46 85 64 71 93 66 67 50 56 61 45 94 96 10
5 7710
321
310
210
1
1 4 4 3 4 19 10 17 24 20 13 10 24 17 12 25 25 20 30 10 23 32 36 38 27 66 29 38 21 34 41 34 46 59 58 52 02 108 68 43
a 0 3 0 4 711
4 3 10 5 2 11 a 9 11 11 15 221 a 11 25
.20 13 41 12 11 12 24 24 10 31 33 30 16 41 54 44 31
Both figures indicate that the empirical probability of a Type I error is
not in general close to .05. Figure 3 indicates that although the frequency
in the rejection region is less affected by the probability of a one as the
number of subjects increases, the tests became rather liberal. Figure 4 is
self explanatory.
The analysis of the data in Table 4 with respect to the normal variates
was interesting in that it tends to contradict earlier findings. The analysis
indicated a strong significant effect due to the number of subjects and
indicates a general downward trend in the empirical probability of a Type
error with an increase in number of subjects.
The data in Table 5 were analyzed by means of a multivariate analysis
of variance and significant effects were found with respect to: (1) the
probability of a one, (2) the number of subjects, (3) items null vs. non-
null, and the interaction of (1) and (2). The significant interaction is
represented in Figure 5 which is somewhat similar to Figure 3 and which in
general lends itself to the same interpretation.
Tables 6 and 7 contain frequencies in a = .05, .025, and .01 rejection
regions of the Quasi-F test statistics for tests of occasion effects when
the data were simulated under non-null occasion effect conditions. The
results in Tables 6 and 7 are most startling for it is apparent that although
the Quasi-F tests based on normal variates responded in an appropriate manner
to non-null effects that the Quasi-F tests based on dichotomous variates
did not, The Quasi-F test based on dichotomous variates had significantly
fewer frequencies in rejection regions when the effect it was testing was
non-null Chan it did when the effect it was testing was null, and in all of
the cases investigated the empirical power of the Quasi-F test was consistently
less than the nominal a level:
-12-
25
Table 6.
The Empirical Frequencies, in the *1.1000 Rejection Region, of
the Quasi-F for Occasions, Based on Normal and Dichotomous Data, Items
Crossed mith Occasions, Occasion Effects Noo-null, Interaction Null.
irobabillty of
a O
ne.
..
0410
0050
2510
SO
25
10
50
25
10
Die
boto
sow
s or
Nor
mal
DNDN
DN
D.NDNDN
DNDNDM
1
No.
of
JI
Lev
el o
fMet
128
298
1318
96
107
2010
616
917
4329
111
1767
1234
null
220
328
922
53
129
2620
816
126
365
810
98
666
303
1138
47
258
415
510
244
1513
38
627
140
668
235
44
649
02
350
121
615
209
718
43
825
151
588
046
124
288
1020
5S
122
3318
319
101
1056
1411
34
551
32no
n-nu
ll2
1734
211
237
312
514
193
210
91
5512
118
361
232
312
346
423
21
126
2524
411
141
578
1313
69
675
2)4.
1346
35
345
121
611
280
119
40
945
151
110
01
47
114
412
627
52
150
2223
211
168
461
4314
225
6510
30nu
ll2
547
42
343
122
223
266
810
35
9345
166
1179
344
311
545
640
63
277
2828
97
202
712
325
177
1091
451
64
665
63
516
235
413
386
626
42
149
1421
99
114
164
19
403
328
60
182
2223
615
155
073
5112
719
6410
39no
n-nu
ll2
843
94
309
320
218
257
818
97
8448
142
2067
047
37
544
441
52
276
2926
524
197
1998
3017
810
0910
484
76C
21
472
131
923
401
1930
011
193
3422
127
126
1469
110
544
932
94
233
2130
810
219
811
246
218
1911
31
48nu
ll2
562
91
497
029
932
342
1724
16
145
4220
922
109
1358
34
697
256
70
393
2042
09
275
917
330
253
1217
65
928
42
792
169
01
529
2251
514
376
722
831
295
2523
020
131
15
582
444
31
273
3431
523
218
1411
841
207
2411
216
53no
n-nu
ll2
1161
45
487
331
418
353
424
13
147
5723
334
133
2476
315
672
957
97
399
2443
012
2 90
019
457
245
2CP
1 51
2091
48
790
265
30
512
1652
014
388
023
737
304
1422
30
109
19
696
655
94
371
937
95
291
220
756
234
1616
3.1
-677
null
25
732
458
24
436
1844
110
307
622
939
249
2118
914
107
32
915
270
21
525
1352
56
377
226
97
294
120
3a
122
104
291
00
804
(`65
811
620
247
50
299
2034
37
2 31
514
11
368
91
534
134
724
356
1828
72
197
4122
318
160
1182
non-
null
25'
760
362
00
440
IC44
910
314
123
043
266
301
5613
110
32
827
072
50
532
2152
612
357
425
530
330
1422
50
142
42
911
278
40
644
2261
15
465
031
123
343
132
315
151
-1
373
4.0
622
044
4.10
432
631
33
173
4635
332
1 19
2453
null
2I
902
167
90
49'3
'12
462
633
13
208
3825
425
1 39
2282
30
880
077
70
625
456
94
404
227
440
319
3122
110
115
124
093
70
878
0 na
066
60
519
036
747
330
3120
717
141
10
514
038
70
278
940
24
308
016
352
261
3813
121
63no
n-nu
ll2
152
20
420
027
913
494
636
00
221
4525
722
146
671
31
591
144
70
331
654
85
391
024
633
318
V2
397
110
42
603
247
92
364
661
62
497
234
535
329
352
7124
16,)
Table 7.
The Empirical Frequencies, in 0(41000 Rejection Regions, of the
Quasi-F for Occasions, Based on Dichotomous Data, Items Nested within Oc-
casions, Occasion Effects Non-null, Interaction Null.
Prob
abili
of a
One
.1
cc 1
000
5025
1050
2510
5025
10
No.
of
Lev
el o
fSu
b.s
item
sSu
b. H
et.
3.25
14a
1?a
415
10nu
lIl
2 312 a
4 42 2
28 1613 13
2 55 13
3 1
0 o4
4 14 30
2 17o 8
a38
4 25o 12
? 167 3
0 1
non-
,iuIS
2 336 30
13 126 5
47 2426 9
1? 915 21
9 a5 5
416
95
2313
623
73
113
64
2613
445
328
non
210
41
256
239
310
315
53
2916
328
236
410
33
114
218
121
127
177
4425
1240
2313
213
42
3831
1025
04
n O
n- h
ull
313
62
4433
1636
24a
414
63
2622
1150
204
19
72
178
359
207
23
1o
329
349
1711
34
42
226
133
114
84
21
o32
106
3?24
101
2414
260
4332
7234
12
non-
non
2 328 17
12 144 4
39 4534 29
16 1654 53
30 2010 15
422
1?9
4637
2479
2615
15
55
189
547
2114
noa
24
33
95
232
219
33
20
1?11
120
60
104
00
010
30
111
1412
1914
538
2914
5132
11
noil-
2 39 4
8 23 1
52 4239 31
23 1349 35
33 2622 17
46
31
1410
67
?3
13
o0
e3
33?
3211
nu2
21
06
33
6229
43
11
06
32
5531
612
4o
00
0o
043
2417
18
41
4120
1151
3927
non
-114
411
2 3a
105 10
3 638 33
19 2012 5
91 8556 57
40 374
108
469
400
9942
24
100
90
an
50
ariSUBJECTS
(pp 40
12=
a 30 10= 0i4P4
20
10
8 =
6 =0
4 =*
00.5 0.2Probability of a One
Figure 3. The interaction of the probability of a one and thenumber of subjects, with respect to the data in Table 4., datafor the quasi-F test of occasions..
0.1
90
"14..
1416 111,
ms,
Non-Null
11111110
Nu
l
0
0.5
0.2
Probability
of a One
Figure
4.
The
interaction
of
the
probability
of a one
and
items
null
vs.
non-null,
with
respect
to
the
data
in
Table
4.,
data
for
the
quasi-F
test
of
occasions.
114,
411.011114.ftimm
eivassio
0.1
II
110
100
90
80
70
60
0 50
40
30
Pei
20
Number of Subjecis
12 m
10 =0
8 =In
6 = 0
4 m*
0.5 0.2Probability of a One'
Figure 5. The interaction of the probability of a one and thenumber of subjects, with respect to the data in Table 5., dataon the quasi-F test of occasions.
0.1
4.3
7.3
The data in Table 6 with respect to normal variates has an overall mean
vector of [407, 300, 195] which is significantly less than the mean vector of
analogous power data based on the regular variate ratio test.
Table 8 is a rearrangement of data which has been previously presented.
The arrangement of data in Table 8 was established to allow easy contrast of
data with respect to regular F and Quasi-F tests based on normal and dichoto-
mous variates under null and non-null repeated measures conditions.
Table 9 is laid out in the same manner as Mlle 8 and includes some new
data, that with respect to normal variates under situation 2.
ect
Tables 10 and 11 present frequencies in a = .05, .025, and .01 rejection
regions for variance ratio tests of subject by occasion interaction effects,
when the data were simulated under null occasion and null subject by occasion
interaction effect conditions. The frequency data in Table 10 are with respect
to situatior. 1 and the frequency data in Table 11 with respect to situation 2.
Again MANOVA indicated a significant difference between the data in
Table 10 based on normal variates and the data in Table 10 based on dichotomous
data.
The multivariate analysis of the data in Table 10 based on dichotomous
variates indicated significant effects due to: (1) the probability of a one,
(2) the number of subjects, (3) items fixed vs. random, (4) the level of
subject heterogeneity the interaction (1) and (2) and the interaction of (1)
and (3). The two interactions are displayed graphically in Figures 6 and 7.
Inspection of Figure 6 indicates that a favorable comparison of empirical
probability of a Type I error to nominal occurred only for 6 subjects and a
Table 8.
The Empirical Frequencies, in ot.1.000 Rejection Regions, of the
Variance Ratio Test and Quasi-F for Occasions, Based on Normal and Dicho-
tomous Data, Items Crossed with Occasions, Twelve Subjects.
0g 1
009
5025
10SO
2520
5025
Nor
mal
or
Die
hotO
noua
D1
8V
8II
8I
811
811
0If
8II
00
FI
1cel
oft.
Sub.
fle
I1
4653
2027
1113
4462
2335
28
6053
1534
49
ol
nqn
245
6825
3915
1670
4929
297
439
4716
352
43
6555
2327
00
3750
1526
411
3648
1436
312
&ri
ft*
463
5324
3310
1258
4621
299
517
504
344
91
5957
2933
718
4756
3537
1314
3145
2520
106
naa-
nall
246
SO13
224
671
6550
3925
1339
4128
29Id
33
5380
2843
12la
6054
3334
22.1
845
3125
2211
54
4963
3340
140
4840
2130
107
6450
1930
44
Cen
tral
110
225
6312
357
9025
5910
271
8730
5416
415
nun
283
46SO
2027
574
2451
1026
269
2763
2227
2gu
llet-
394
2050
923
291
2544
1023
292
2135
1417
64
8028
596
350
7918
39T
110
7431
5319
273
190
3956
1934
366
3537
2428
412
1a3
7415
406
tios_
rolf
204
2241
1024
167
3549
2019
110
117
4910
303
378
4533
2321
487
3747
2126
795
1954
1433
54
7121
4710
272
9634
5712
403
6834
3815
224
161
589
649
482
035
169
936
560
926
350
717
434
915
834
385
244
4514
2no
II2
623
938
563
878
356
746
343
653
250
556
133
407
125
379
9425
244
158
366
197
753
292
939
705
239
677
126
965
619
151
214
746
475
357
3622
7P°
711.
475
198
863
696
944
393
645
485
432
075
418
661
417
648
910
240
554
291
161
191
547
283
533
371
936
555
525
546
215
734
915
035
011
725
554
.
171
-om
it2
563
931
495
874
349
766
376
664
267
558
185
409
146
380
115
294
6515
73
666
966
538
93C
387
858
405
752
296
642
176
435
157
407
119
34a
4122
04
696
992
582
97C
420
941
402
839
308
712
201
615
160
491
9243
545
214
Nor
:-ce
ntra
l1
373
40
622
044
410
432
631
33
173
4635
332
119
2453
non
21
002
267
90
497
1246
26
331
320
8 -
3025
425
139
.22
132
30
830
077
73
525
456
94
404
227
440
319
3122
118
115
Qua
si.-
40
937
087
83
720
066
60
519
036
747
330
3126
717
141
.1
051
40
387
027
39
402
430
50
163
5226
13d
131
2163
atal
-sd'
21
528
042
00
279
1349
48
360
022
145
247
2214
66
913
159
11
447
033
16
548
539
10
246
3331
89
239
711
04
260
31
479
236
46
616
249
72
345
3532
935
271
2415
0
!14
.11,
314,
4
.,4.
Table 9.
The Empirical Frequencies, in ota1000 Rejection Regions, of the
Variance Ratio Test and Quasi-F for Occasions, Based on the Normal and
Dichotomous Data, Items Nested within Occasions, Twelve Subjects.
,Probability of a On
.5
r
1311000
50
25
10
SO
25
10
SO
25
10
Normal or Diehotosbous
1:1
11D
alD
11D
110
11E
l11
011
011
D2
Oecautins
FItems
Lev
el o
fSub. Het. ....,
146
53
28
27
11
13
44
62
23
35
28
60
53
19
36
49
hitt/
245
63
25
39
15
16
70
49
28
29
74
39
47
16
35
24
365
55
23
27
aa
37
50
15
26
411
36
48
14
36
a12
00-
4 163
206
53
236
24
143
33
207
10
8Z
12
127
58,
258
46
445
21
160
29
339
9101
5
266
17
177
50
486
4129
34
369
450
9
291
608-00 0
2 ,188
202
3C3
334
126
143
217
245
61
81
154
181
253
206
484
534
185
159
410
476
123
121
313
352
213
114
515
546
10o
97
44 5
459
5446
353
360
4174
393
128
304
64
204
268
571
164
495
107
413
160
652
90
550
55
45
Central
1101
40
67
14
39
8102
34
54
14
29
494
46
46
20
31
9
pull
278
55
48
27
19
13
92
39
53
14
28
496
41
59
17
33
12
(Nast-
3 '4
77
29
49
727
382
38
47
13
26
4105
40
58
14
30
9
97
40
69
11
25
285
31
48
912
377
46
52
20
16
10
1123
54
77
44
40
20
85
104
63
89
29
51
103
93
82
71
41
SO
non-nut/
2122
75
87
52
68
17
72
113
45
82
26
48
213
119
166
1SO
54
66
3126
69
89
66
58
28
81
109
53
77
37
47
102
127
66
107
44
66
4174
95
128
73
84
36
115
109
60
88
31
49
ICI
144
43
101
31
59
.....,
1615
898
494
820
351
699
365
609
263
507
174
349
158
343
85
244
45
142
PIM
2623
933
500
878
356
746
363
653
250
556
133
407
125
379
94
292
44
150
3661
977
632
929
397
852
396
771
265
656
191
512
147
464
75
357
36
227
Parkt
4751
988
636
969
448
936
452
654
320
754
206
614
176
489
102
465
54
292
1655
835
550
827
447
744
492
764
404
631
293
593
311
650
265
564
199
400
2686
909
590
868
479
795
459
818
404
720
317
604
272
707
214
593
182
479
II0IF411111
3734
928
646
931
508
349
516
536
422
792
313
705
317
774
214
699.
170
574
Hoa-
4750
933
667
907
541
876
515
907
445
871
333
791
2.44
834
221
749
136
676
Central
13
791
06:19
0541
8499
3360
3250
37
259
32
206
11
99
hUn
22
649
1745
C600
6525
3413
3273
62
262
29
175
491
31
904
1828
0708
8612
3513
2341
55
317
31
254
i153
40
948
0914
3842
,0
698
0591
0466
43
366
24
306
17
197
Qua
si-
1.8
553
4446
1319
41
245
20
173
11
136
51
167
3/
134
27
113
MM441
23
577
5469
3354
30
278
12
189
12
137
91
179
56
124
44
66
310
593
10
524
6391
33
321
20
217
5144
'
85
209
57
153
37
111
410
632
8523
4416
69
321
48
229
0143
99
224
42
167
24
1)2
J
k'
Table 10.
The Empirical Frequencies, in 4;4)1000 Rejection Regions, of the
Variance Ratio Test for the Subjects by Occasions Interaction, Based on the
Normal and Dichotomous Data, Items Crossed with Occasions, Interaction and
Occasions Effects Both Null.
Piro
ba61
11of
a O
no.
..
.2.
ot .1
000
5025
1050
25.1
050
2510
Dic
hoto
mou
s or
dor
eal
01
01
08
0II
111
08
D11
011
718
No.
of
sub_
5462
3130
1610
4750
2428
1113
3040
2527
1714
null
254
4026
257
1062
6537
3324
2%.,
4657
3020
2713
371
4540
2225
771
3556
1631
1364
5647
1928
94
454
4126
2C12
865
3550
143E
,10
4735
3213
217
a55
6835
3514
1253
4530
2115
951
5433
1816
10no
n-nu
ll2
4962
2929
2213
4543
2417
138
5746
3619
2912
362
713?
3626
2144
4525
1813
961
4031
1827
114
.75
5146
3226
1847
5224
2521
1564
5337
2732
18
158
4329
2010
746
4653
2:12
556
6631
3912
18nu
ll2
6353
3020
185
6549
3?22
1211
101
5170
2757
143
6534
4119
173
8162
5729
238
116
55d3
2752
106
468
5739
2818
1587
5259
3123
1211
360
7234
3516
139
3723
1510
1061
4836
2425
852
4533
2917
20no
n-nu
ll2
5443
2?14
146
4858
3134
1921
8549
5127
2812
363
4543
2615
951
5141
3134
1181
5649
3337
204
6950
4427
117
4351
'37
3222
1475
3947
2337
13
152
5229
249
553
3639
20,
157
7833
3518
184
null
258
5831
3114
.11
6640
4116
287
111
437/
2767
73
4445
1525
512
6555
5222
2211
13J
29la
1951
2a
473
4736
1922
1090
5349
3028
1312
243
8322
497
155
3836
2111
1160
3944
2325
667
3146
2229
10no
n-nu
ll2
4942
2318
1010
7C33
4519
1911
7438
7025
364
256
5231
3314
1283
4245
2833
1375
3347
1426
34
6049
3522
1713
0438
'38
2625
876
3649
).9
154
168
6041
2915
1564
4945
1615
1593
5154
2625
22nu
ll1
5641
2317
1310
9349
4721
2219
115
3567
il51
93
6545
3123
1311
112
5257
2038
1714
247
109
2563
2010
467
4235
18la
1010
362
6724
4318
150
5511
627
8 /
201
4841
la21
a11
6661
2817
2014
112
6047
2930
26no
n-nu
ll2
4746
2215
98
5958
2424
1819
7755
4226
3918
349
4921
1810
1291
052
5926
2421
130
5264
2337
214
5343
2214
128
9940
6030
5226
124
5464
2151
17
145
5120
286
984
4247
2515
1113
046
9230
5012
null
240
5223
2811
1611
052
7235
3819
121
5892
3555
203
7456
3428
2011
6849
4530
2012
7951
7222
5711
466
4642
3112
710
850
6033
249
153
5(.
1d:9
3491
161
5445
3229
1012
5655
3038
1716
64.;
5636
2027
12no
n-nu
ll2
6259
2740
1519
6144
5134
268
104
5370
2650
103
4648
1936
1116
6247
3232
415
9240
5529
35-
104
6546
2924
1412
121
478
Z26
4214
100
4783
2445
15
Table 11.
The Empirical Frequencies, in cc+1000 Rejection Regions, of the
Variance Ratio Test for the Subjects by Occasions Interaction, Based on
the Normal and Dichotomous Data, Items Nested within Occasions, Interaction
and
Occ
asio
nsEffects Both Null.
Probability of a One
.5
.2
.1
Of
1000
SO
25
MW
25
20
SO
25
10
Oichotomous or Normal
OVANDODIMON
DNONEIN
No.
of
Sub
.s.1
/Will
iLe
vel o
fS
Ub.
Het
.1
5957
3222
1610
5053
2920
912
4636
3126
2514
,..
null
255
4427
168
765
4528
2014
5251
412
189
364
5234
2418
642
4718
247
7146
5526
3012
44
5951
3321
1510
4343
1132
663
2939
17ld
b.
155
4926
2811
1450
4213
199
854
4942
2223
10no
n-nu
ll2
6563
3226
910
5842
2018
B11
6834
4425
3211
354
5528
299
1376
5239
1914
461
395L
2621
144.
5760
3731
1717
5247
4215
256
7029
5121
)25
7
155
4924
2312
861
4344
1814
a65
6536
3519
18
null
269
5342
2822
870
4741
2321
1699
51o4
2957
lo3
'66
4539
2221
973
5259
2326
1110
057
7)38
63.L
76
485
6442
3221
1781
4869
3036
1710
654
ol31
5814
164
4430
2713
1484
3843
1921
1155
5237
2417
1-no
n-nu
ll2
6947
3925
1610
8044
5618
3510
7342
5123
519
379
5044
2818
1161
4453
2740
1283
505d
3748
174
7257
3733
2514
102
5273
3452
1586
5260
3339
13
161
4936
3012
851
4431
1410
255
5240
ld21
L
null
252
5535
3117
670
4438
1520
611
462
66lo
5610
357
4526
279
1360
5851
2126
512
348
0517
56d
84
8549
5528
314
106
5364
2429
312
651
7120
509
161
4330
229
769
4337
22lti
777
6844
1735
b
non-
null
261
5331
1713
781
4741
1525
3o3
55bl
16'id
23
5159
3623
182
7640
4614
224
9262
5113
350
462
4330
2117
7-
9549
5617
295
8143
526
430
169
5238
3012
1470
5655
2821
1288
5529
3229
lbnu
ll2
6750
2425
712
8549
4230
2815
112
4598
1759
7
368
6032
2513
1311
959
7037
3616
15d
5511
J31
7617
104
6946
3623
1813
104
5667
3448
1714
669
109
3673
ld1
6354
3426
1417
6465
4545
1915
7176
7137
2612
non-null
2 376 67
56 4651 36
33 2322
614 10
8810
062 67
61 5234 32
35 3415 17
106
128
73 6271 16
4.)
3545 61
lo 134
9954
6225
2413
9565
7737
2819
141
61llo
3892
Id
150
5321
2810
1482
4749
2819
1711
940
t-.7
3341
17
null
2 3
44 8459 72
19 4728 35
IC 1715 12
105 90
49 5074 61
39 3754 21
24 1810
8 9347 48
19 7.;
37 2656 41
28 1412
479
5245
2313
1210
639
7434
4114
143
5211
733
8621
162
7942
4225
1691
3941
3417
207
fjeo
l21
3646
132
4452
2534
1314
7646
4637
1620
103
3564
2651
12non-null
346
5033
3412
e5t
.56
2.;
403
2311
842
1C,;
g5
6..
144
6047
3529
1918
7628
4116
1914
101
33o6
2%;
347
,
120
1 1 0
Number of Subjects
1 00 12 =
10 = 0
90 8 =
6 = 0
80 4 = *
40
0.5 0.2 0.1
Probability or a One
Figure 6. The interaction of the probability of a one and tilenumber of subjects, with respect to the data in Table 10.1, 4ataon the test of subjects by occasions interaction.
probability of a one equal to .5, and for 4 subjects and a probability of
a one equal to .2 or .1. For all other conditions the test is too liberal.
Figure 7 is self explanatory.
A multivariate analysis of dhe data in Table 11 indicated the same trends
and significant effects that were found in Table 10.
Relative power variables were formed for both Table 10 and Table 11
and separate multivariate analyses of variance were performed on the relative
power variables for both tables. In both analyses significant effects were
found due to; (1) the probability of a one, (2) the number of subjects, the
interaction of (1) and (2), the interaction of (1) and items fixed vs.
random, and the interaction of (1) and the level of subject heterogeneity.
In addition the analysis of the relative power variables from Table 11
disclosed a significant main effect due to items fixed vs. random.
In general it may be observed that higher relative powers correspond to
greater "liberalness" in the test of a true null hypothesis. Thus the
general increases in relative power observed across levels of a probability
of a one .5 to .2 to .1 are in same sense specious.4
I 00
90
700 /1:3
60eel
a)
50k
40
0
"6" Non-Nullaserrore Null
0.5 0.2Probability of a One
Figure 7. The interaction of the probability of a one and itemsnull vs. non-null, with respect to the data in Tabl 10., dataon the test of subjects by occasions interaction.
0.1
BIBLIOGRAPHY
Bradley, J. V. Distribution-Free Statistical Tests. New York:
Prentice-Hall, 1968.
Donaldson, T. S. Power of the F-test for non-normal distributionsand unequal error variances. Rand Corporation research memorandumRM-5072-PR, September, 1966.
Draper, J. F. A Monte Carlo investigation of the analysis of vuriance
applied to non-independent Bernoulli variates. Unpublisheddissertation, Michigan State University, 1971.
Greenberger, M. Notes on a new pseudo-random number generator."Journal Assoc. Comp. Mach.," 8, 163-67.
Greenhouse, S. W. and Geisser, S. On methods in the analysis ofprofile data. "Psychometrika," 1959, 24, 95-112.
Hammersleyt J. M. and Handscomb, D. C. Monte Carlo Method. New York:
John Wiley, 1964.
Hsu, T. C. and Feldt, L. S. The effect of limitations on the number ofcriterion score values on the significance level of the F-test."American Educational Research Journal," 1969, 6, No. 4, 515-28.
Hudson, M. D. and Krutchkoff, R. G. A Monte Carlo investigation of
the size and power of tests employing Satterthwaite's syntheticmean squares. "Biometrika," 1968, 55, 431-33.
Lunney, G. H. A Monte Carlo investigation of basic analysis ofvariance models when the dependent variable is a Bernoullivariable. Unpublished dissertation, Univ. of Minnesota, 1968.
Mandeville, G. K. An empirical investigation of repeated measuresanalysis of variance for binary data. Paper presented at the
annual meeting of the American Educational Research Association,1970.
Marsaglia, G. and Bray, T. A. One line random number generators andtheir use in combinations. "Cammunications of ACM," 1968, 11,757-59.
Satterthwaite, F. E. Synthesis of variance. "Psychometrica," 1941,
6, 309-16.
Seeger, P. and Gabrielsson. Applicability of the Cochran Q test and
the F test for statistical analysis of dichotamous data fordependent samples. "Psychol. Bull.," 1968, 69, 269-77.
APPENDIX A
If 1,000 samples are simulated so that a null hypothesis is true
and so that the assumptions of an ANOVA are met, the number of F-statistics
testing the above hypothesis which have values which exceed the Fl_a
quantile of a corresponding F-distribution4 will be approximately 1,000a.
If 1,000 samples of data are simulated so that a null hypothesis is not
true and so that the assumptions of an ANOVA are met, the number of F-
statistics testing the above hypothesis which have values which exceed the
F1-a
quantile of a corresponding F-distribution will be approximately 1,000
times the nominal power (1-$) for the situation simulated.
Let X. be defined according to the rule
1, F FXi
= {
0, otherwise
where Fi
is the F-variate calculated on the ith
sample, i = 1, 2, ...,
1,000, and F1_01 is the 1-a quantile of the corresponding F-distribution.
The variate Xiis then an indicator variable, which takes on the value
one 'when F.is in the a rejection region and which takes on the value zero
when F does not occur in the rejection region. Note that the frequency
of Fi's which fall in the a rejection region,
1000f, may be represented by the expression f = E X.
i=1
Observe that f is a binominal variate with parameters p = a and n = 1000.
The variance of f is then np(1-p) or for n = 1000 and a = .05, the var
(f) = 47.5. Therefore by employing the normal approximation a, .95
probability interval may be formed about the expected value of f, E(f) =
1000 a, which for a = .05 is Pr(36.5 < f < 63.5) = .95.
4An F-distribution with the same degrees of freedom as the variance ratiofor the test.
40