A study of the power of the rank transform test in a 2 3 factorial...
Transcript of A study of the power of the rank transform test in a 2 3 factorial...
This article was downloaded by: [Northeastern University]On: 30 November 2014, At: 07:58Publisher: Taylor & FrancisInforma Ltd Registered in England and Wales Registered Number: 1072954 Registered office:Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK
Communications in Statistics - Simulation andComputationPublication details, including instructions for authors and subscriptioninformation:http://www.tandfonline.com/loi/lssp20
A study of the power of the rank transformtest in a 23 factorial experimentChoi Young Hun aa Department of Statistics , Hanshin University , Osan , 447-791 , KoreaPublished online: 23 Dec 2010.
To cite this article: Choi Young Hun (1998) A study of the power of the rank transform test in a 23
factorial experiment, Communications in Statistics - Simulation and Computation, 27:1, 259-274, DOI:10.1080/03610919808813478
To link to this article: http://dx.doi.org/10.1080/03610919808813478
PLEASE SCROLL DOWN FOR ARTICLE
Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”)contained in the publications on our platform. However, Taylor & Francis, our agents, and ourlicensors make no representations or warranties whatsoever as to the accuracy, completeness, orsuitability for any purpose of the Content. Any opinions and views expressed in this publicationare the opinions and views of the authors, and are not the views of or endorsed by Taylor &Francis. The accuracy of the Content should not be relied upon and should be independentlyverified with primary sources of information. Taylor and Francis shall not be liable for any losses,actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoeveror howsoever caused arising directly or indirectly in connection with, in relation to or arising out ofthe use of the Content.
This article may be used for research, teaching, and private study purposes. Any substantialor systematic reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, ordistribution in any form to anyone is expressly forbidden. Terms & Conditions of access and usecan be found at http://www.tandfonline.com/page/terms-and-conditions
COMMUN. STATIST.-SIMULA., 27(1), 259-274 (1998)
A STUDY O F T H E POWER O F T H E RANK TRANSFORM TEST IN A 23 FACTORIAL EXPERIMENT
Young Hun Choi
Department of Statistics Hanshin University
Osan, 447-791, Korea
Key Words: Rank transform; power; factorial experiment; ANOVA test; Monte CQrlo simulation.
ABSTRACT
By means of a simulation study it is shown that the rank
transform test has for many cases in a z3 factorial design superior
power over the parametric analysis of variance test. However the
rank transform test should be cautiously used for some situations in
which the model contains both three main effects and interaction
effects.
1. INTRODUCTION
The rank transform procedure is to allow use of existing
parametric tests based on data replaced with their corresponding
ranks. In short the usual parametric procedure is performed on the
ranks instead of on the data themselves.
Copyright O 1998 by Marcel Dekker, Inc
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
07:
58 3
0 N
ovem
ber
2014
CHOI
Recently much research has been done by statisticians in
several fields, especially in the field of experimental designs, to
make the rank transform procedure invaluable. However it is now
well understood why the rank transform procedure will work in
balanced cases and in cases where there is no interaction. Some
explanation will help the reader and will also prevent the practitioner
from applying the rank transform procedure in other cases hoping it
will still work.
Warnings about the limitations of the rank transform procedure
came early from its main architects [Hora and Conover (1984)l and
were repeated by several others [e.g. Brunner and Neumann (1984,
1986)l. The paper by Akritas (1990) considered for the first time
the rank transform procedure in unbalanced designs, as well as for
testing main effects in the presence of interactions, and uncovered
the reasons why the rank transform procedure does not work in
such cases. In particular, Akritas (1990) showed that ranked data
are heteroscedastic even if the original data are homoscedastic. He
also demonstrated that the ranked data, having undergone a
monotone but nonlinear transformation, will not conform to a
parametric null hypothesis even if the original data do and, as a
consequence, they cannot be used to test these parametric
hypotheses. He also demonstrated that in the balanced case, the
ANOVA procedure for testing for no main effects when interactions
are not there is valid even for heteroscedastic data. This explained why the rank transform method works in this balanced case in spite
of the fact that the rank transformed data are heteroscedastic.
Thompson (1991, 1993) studied the asymptotic properties of the
rank transform statistic for testing for interactions in a two-way
layout linear model with interactions. This author found that if and only if there are exactly two levels of each main effect, the
asymptotic distribution of the rank transform test for testing for
interactions is chi-squared. Akritas and Arnold (1994) considered
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
07:
58 3
0 N
ovem
ber
2014
multivariate repeated measures designs which are supposed to be
arbitrarily heteroscedastic and proceeded to define the stronger
hypotheses called nonparametric hypotheses for no main effects and
no interactions that are invariant under monotone transformations
and showed that the rank transform statistics test these stronger
hypotheses. Further Akritas, Arnold and Brunner (1997) extended
all possible nonparametric hypotheses for three- and higher-way
layouts, both balanced and unbalanced, to factorial designs with
independent errors.
Meanwhile we will examine the articles that employed computer
generated simulation methods in order to compare the Type I error
and power properties of the rank transform test to those of the
usual analysis of variance F test.
Although many simulation papers applied in the context of a
two-way layout have shown that the rank transform approach is
fairly proper for many given circumstances [e.g. Conover and Iman
(1976), Pirie and Rauch (19841, Pavur and Nath (198611, it seems to
be controversial under some situations [Blair, Sawilowsky and
Higgins (1987, 1989)l. Further until recently, the rank transform
procedure has not provided useful solutions for problems involving
more complicated linear models, particularly factorial designs with
high-order interactions. Namely due to complexity of theoretical
development for the rank transform method, the theory for tests
based on ranks over a three-way layout with interactions has not
concretely been considered. Moreover the simulation stud)- of the
rank transform test in a three-way layout has not thoroughly been
developed.
Thus the primary concern of this paper is comparing the power
of the usual analysis of variance F test and the rank transform test
in a 2"actorial experiment. The comparison is made by means of
a Monte Carlo simulation study. Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
07:
58 3
0 N
ovem
ber
2014
CHOI
2. DESCRIPTION OF THE PROCEDURES
The model used in our simulation study is the three-way
crossed classification model as follows:
Xijkn = /1 + a , + Pj + Y k + &ij + aYik + P Y , ~ + ahi jk + eijkn, for i , j , k = 1 , 2 , n = 1 , 2 ,..., N ,
where /1, Q i , B j , Yk , @ij, ayik, P Y , ~ , a h i j k are the overall mean, main effects and interaction effects. The error terms, e i j h , are assumed to be independently and identically distributed with a
continuous distribution. Four underlying distributions namely the
standard normal with mean of zero and variance equal to 1, exponential with scale parameter of 1, double exponential with
location parameter of zero and scale parameter of 1 and uniform
distributions on ( 0 , 1 ) are chosen for the error terms. The process
is carried out for N = 2 , 5 , 10, 20, 30, 50 observations per cell. However only results for N= 2 , 10 are presented here since results
for N= 5 , 20, 30 , 50 are consistent with those for N = 2 , 10 and
increases in sample size beyond N = 10 contribute to fast power
inflations near to 1 in almost all situations so that we can not sharply contrast the difference of power for tests under the given
conditions.
In this article to test for the null hypothesis of no B main
effects Ho : Pj= 0 for all j and the null hypothesis of no AB
interaction effects Ho : aDij=0 for all i and j , the symbols F and FR will denote the usual ANOVA test statistic and F statistic on
the ranks of the data respectively. The A, B and C main effects
are formed by setting al = - a2 = cr, P1 = - P2 = B and
Y , = - yz = y respectively, where a, P and y are either 0 or
take values from 0.25 to 1.25 with increments of 0 . 2 5 , or are
0.5 or 1.5 times the latter values. Meanwhile the AB interactions
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
07:
58 3
0 N
ovem
ber
2014
are created by setting = aBn = - aP12 = - aPzl = aP, with
a@ taking on the values just mentioned above. The AC, BC and ABC interactions are formed in a similar manner. Further all ABC
interactions, f f f i ~ ~ , ~ ' ~ , are taken to be zero in order to clarify the simulation results focused on the main effects and the two-factor
interactions.
The first step in our simulation study is to generate pseudo
random deviates using a C program, after which the specified main
and interaction effects are added to obtain the X-values. Next the classical ANOVA F statistic is computed. After comparing the
critical values corresponding to 8(N- 1) degrees of freedom for
denominator and the significance level of 0.05, we calculate the proportion of rejections under the given effects for the power. In
addition the procedure is repeated after replacing original
observations with their respective ranks. For each case 10,000 repetitions are performed.
For standard normal random variates we use Box and Muller
(1958)'s transformation. For exponential random variates we
generate uniform random variates and use the inverse transform
method to convert to exponential random variates [Marsaglia (196111. For double exponential random variates we generate exponential
random variates and use the composition method to yield double
exponential random variates [Law and Kelton (1991 11. For uniform
random variates we use the linear congruential method drectly.
3. SIMULATION RESULTS
The simulation study is conducted over a variety of experimental situations. We wish to investigate the effects of allowing the
magnitude of the nuisance parameters to vary. We also consider
several fashions in which main effects and interaction terms are
constructed.
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
07:
58 3
0 N
ovem
ber
2014
264 CHOI
The results reported in TABLE I through TABLE IV are
representative of a variety of configurations considered. Namely the
power comparison of the parametric analysis of variance test and
the rank transform test is presented. In these tables the second
column labeled "c" indicates the value of constant employed in
generating the given effects. The "Statistic" column exhibits which
of the two statistical tests is being considered; F represents the
usual ANOVA statistic and FR represents the rank transform
statistic.
TABLE I and TABLE II examine the power of tests for main
effects (i.e. B effect for our case) when sampling is from N= 2 and
N= 10 observations per cell respectively. For the power analysis, the following nine different situations are considered: one main
effect is nonnull; P = c , a = r = aP = ay = By = aBy = 0, O two
main effects are nonnull; a = B = c , y = aP = cry = By = aBy = 0 , 3' three main effects are nonnull; a = P = y = c , aB = ay = By
= aBy = 0 , O' another construction in which three main effects are nonnull; a = 1.5c, P = c , y = 0.5c, aP= ay= B y = aBy= 0 , @ meanwhile one main effect and one interaction are nonnull;
1 P=aB= C , a = y=ay=By=aPy=O, O three main effects and one interaction are nonnull; a = P = r = ay = c , UP = By =
aBy = 0 , O2 another construction in which three main effects and one interaction are nonnull; a = P = r = aB = c , ay = By = aBy
= 0 , 63' three main effects and two interactions are nonnull; a = B = y = aB = ay = c , By = aBy = 0 , @ 2 another construction
in which three main effects and two interactions are nonnull;
a = 1 .5c , B= c, r = 0 . 5 c , aB=0.5c, ay=1 .5c , By= aBy=O.
The primary points of focus in this study are (a) number of
nonnull effects in the model, (bi effect size, ( c ) number of
observations per cell, ( d l types of population distributions, in order to
provide a more rigorous examination of power of the rank transform
test for a factorial design.
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
07:
58 3
0 N
ovem
ber
2014
2' FACTORIAL EXPERIMENT
TABLE I
Power of tests for B main effect at the significance level of 0.05 when sampling is from N= 2 and for
all cases of F
B = c , a = y = aB= ay= By= &y=O a = B = c , y = aB= ay= By= aBy=O
O' a = B = y= c , aB= ay=By= aBy=O
O2 a = 1 . 5 c , B = c , y= 0.5c, aB= ay- By= aBy= 0 @ B = a B = c , a = y=uy=By=aBy=O 6)' a = b= y= ay= c , aB= By= aBy= 0 0' a = B = y= aB= c , ay=By= & = O 6)' a = B = y = a B = a y = c , By=aBy=O 8' a = 1 . 5 c , B = c , y=0.5c, aB=0.5c , ay=1.5c, By=aBy=O.
Statistic Population c @F aFR SFR ~ ' F R O'FR @FR O'FR O'FR O'FR ~ ' F R
Normal 0.25 0.145 0.145 0.143 0.142 0.142 0.141 0.143 0.142 0.134 0.137
0.50 0.426 0.420 0.405 0.399 0.392 0.406 0.402 0.369 0.305 0.3%
0.75 0.756 0.742 0 717 0.707 0.706 0.716 0.703 0.630 0.481 0.618
1.00 0.942 0.935 0 914 0.903 0.911 0.914 0.901 0.824 0.643 0.822
1.25 0.993 0.991 0.985 0.976 0.982 0.986 0.978 0.931 0.780 0 934
Exponential 0.25 0.177 0.256 0.234 0.224 0.215 0.233 0.225 0.194 0.179 0.193
0.50 0.498 0.597 0.550 0.526 0.523 0.553 0.528 0.451 0.369 0.455
0.75 0.776 0.822 0 785 0.758 0.771 0.785 0.762 0.684 0.550 0.684
1.00 0.917 0.927 0.907 0.889 0.908 0.906 0.888 0.824 0.680 0.836
1.25 0.972 0.972 0.964 0.952 0.962 0.964 0.954 0.910 0.772 0.920
Double 0.25 0.106 0.126 0.121 0.121 0.117 0.119 0.116 0.118 0.111 0.111
Exponentia10.50 0.279 0.323 0304 0.294 0.288 0.304 0.299 0.272 0.233 0.265
0.75 0 514 0.563 0 527 0.511 0.506 0.534 0.516 0.464 0.372 0.4%
1.00 0.727 0.762 0.727 0.709 0.709 0.730 0.708 0.637 0.496 0.632
1.25 0.872 0.891 0.863 0.840 0.850 0.865 0.845 0.766 0.608 0.771
Uniform 0.25 0.869 0.821 0.803 0.789 0.798 0.803 0.781 0.702 0.528 0.701
0.50 1.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000 0.938 0.999
0.75 1.000 1.000 1000 1.000 1.000 1.000 1.000 1.000 0.938 1000
1.00 1.000 1.000 1000 1.000 1.000 1.000 1.000 1.000 0.938 1.000
1.25 1.000 1.000 1000 1.000 1.000 1.000 1.000 1.000 0.938 1.000 Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
07:
58 3
0 N
ovem
ber
2014
TABLE I .
CHOI
Power of tests for B main effect at the significance level of 0.05 when sampling is from N= 10 and for the same cases as TABLE I .
Statistic Population c @)F O F R Q F R D'FR O'FR @FR ~ ' F R ~ ' F R O'FR @'FR
Normal 0.25 0 602 0.581 0.579 0.580 0.579 0.581 0.578 0.576 0.547 0.565
0.50 0.992 0.990 0.989 0.988 0.988 0.990 0.987 0.985 0.962 0.979
0.75 1.000 1,000 1,000 1000 1.000 1.000 1.000 1.000 0.989 1.000
1.00 1.000 1.000 1.000 1000 1.000 1.000 1.000 1.000 1,000 1.000
1.25 1.000 1.000 1.000 1000 1000 1,000 1.000 1000 1,000 1.000
Exponential 0.25 0.619 0.876 0.839 0.815 0.801 0.845 0.810 0.780 0.720 0.744
0.50 0.988 1.000 1.000 0 998 0.998 0.999 0.998 0.997 0.983 0.993
0.75 1.000 1.000 1.000 1000 1000 1.000 1.000 1.000 0.999 1,000
1.00 1.000 1.000 1.000 1.090 1.000 1.000 1.000 1.000 1.000 1.000
1.25 1.000 1.000 1.000 1.000 1.000 1.000 1.000 1000 1.000 1.000
Double 0.25 0.364 0.480 0.469 0.462 0.454 0.468 0.455 0.449 0.422 0.433
Exponential 0.50 0.880 0.953 0.937 0 928 0.923 0.940 0.925 0.912 0.863 0.891
0 75 0.994 0.999 0.998 0 997 0.998 0.998 0.997 0.996 0.983 0.993
1.00 1.000 1.000 1.000 1000 1.000 1.000 1.000 1,000 0.998 1,000
1.25 1.000 1.000 1.000 1000 1.000 1.000 1.000 1,000 1.000 1,000
Uniform 0.25 1.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000
0.50 1.000 1.000 1.000 1000 1.000 1.000 1.000 1.000 1.000 1,000
0.75 1.000 1.000 1.000 1000 1.000 1.000 1.000 1.000 1.000 1,000
100 1.000 1.000 1.000 1000 1.000 1.000 1.000 1.000 1.000 1,000
1.25 1.000 1.000 1.000 1000 1.000 1.000 1000 1.000 1,000 1,000
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
07:
58 3
0 N
ovem
ber
2014
TABLE IJI
Power of tests for AB interaction effect at the significance level of 0.05 when sampling is from N= 2 and for
62 all cases of F Q a,!?= c , a = ,8= y= ay= By= @y=O @ aB= ay= c , a = ,9= y= By= & = 0 (3 B = aB= c , a = y= a y = By= a,8y=0 @' B = y = aB= c , a = ay= By= u&?y=O
a = B = a@= c , y= ay= By= aBy=O a' y = & = c , a y = B y = a B y = O B2 a = c , B = 0 . 5 c , y = 1 . 5 c , a B = c , ay=By=aBy=O @' a = B = y = a B = a y = c , By=aBy=O Q2 a = c , B = 0 . 5 c , y = 1 . 5 c , a B = c , a r = 0 . 5 c . BY= aBy=O.
Statistic Population c F DFR OFR 3 F R @'FR 02FR B'FR B2FR @'FR D'FR
Normal 0.25 0.148 0.149 0.150 0.151 0.149 0.146 0.145 0.144 0.134 0.140
0.50 0.427 0.419 0.409 0.409 0.399 0.351 0.375 0.391 0.304 0.365
0.75 0.752 0 739 0.714 0.714 0.704 0.504 0.629 0.675 0.484 0.636
1.00 0.938 0,930 0.910 0.914 0.898 0.550 0.817 0.862 0.646 0.830
1.25 0.993 0.991 0.983 0.985 0.974 0.558 0.931 0.941 0.772 0.934
Exponential 0.25 0.177 0.254 0.232 0.234 0.223 0.185 0.192 0.205 0.180 0.201
0.50 0.495 0.595 0.545 0.547 0.519 0.361 0.451 0.499 0.374 0.469
0.75 0.773 0.823 0.784 0.782 0.761 0.471 0.679 0.730 0.546 0.702
1.00 0,917 0.931 0.912 0.911 0.895 0.523 0.832 0.872 0.675 0.847
1.25 0.971 0.974 0.963 0.963 0.953 0.543 0.909 0.934 0.768 0.923
Double 0.25 0.106 0.126 0.122 0.125 0.121 0.122 0.121 0.120 0.114 0.118
Exponential 0.50 0.274 0.313 0.300 0.296 0.290 0.252 0.268 0.281 0.228 0.266
0.75 0.503 0.552 0.520 0.527 0.509 0.385 0.457 0.484 0.364 0.455
1.00 0.724 0.755 0.722 0.721 0.698 0.473 0.628 0.670 0.490 0.632
1.25 0.870 0.887 0.866 0.861 0.840 0.516 0.762 0.804 0.604 0.771
Uniform 0.25 0.866 0.816 0.793 0.793 0.785 0.554 0.693 0.782 0.522 0.723
0.50 1,000 1.000 1.000 1.000 1.000 0.554 1,000 0.984 0.937 0.992
0.75 1.000 1.000 1.000 1.000 1.000 0.554 1.000 1.000 0937 1.000
1.00 1000 1000 1.000 1.000 1.000 0.554 1.000 1.000 0937 1.000
1.25 1.000 1.000 1.000 1.000 1,000 0.554 1.000 1.000 0937 1.000 Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
07:
58 3
0 N
ovem
ber
2014
TABLE N
CHOI
Power of tests for AB interaction effect at the significance level of 0.05 when sampling is from N= 10 and for the same cases as TABLE IU.
S t a t ~ s t ~ c Population L $ F I)FR 3FR JIFR ~ ' F R Z'FR ~ ' F R ~ ' F R E'FR s2FR
Normal 0.25 0.602 0.581 0.580 0.583 0.582 0.577 0.575 0.576 0 547 0 565
0.50 0.994 0.991 0.991 0.991 0.990 0.988 0.989 0.990 0 965 0 984
0 7 5 1,000 1 0 0 0 1.000 1.000 1,000 1.000 1.000 1 0 0 0 0.999 1 0 0 0
1.00 1,000 1,000 1.000 1.000 1,000 1.000 1.000 1,000 1 000 1 000
1 2 5 1,000 1 0 0 0 1.000 1,000 1,000 1.000 1,000 1,000 1 0 0 0 1 0 0 0
Double 0 25 0.364 0 479 0.464 0.467 0.458 0.455 0.447 0.446 0 418 0 435
Exponential 0 50 0.878 0.956 0.942 0.942 0.932 0 912 0.914 0 917 0 866 0 902
0 75 0 995 0.999 0.999 0.999 0.998 0 992 0.996 0.997 0 984 0 996
1.00 1 0 0 0 1.000 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1.000 1 0 0 0 1 0 0 0
1.25 1 0 0 0 1.000 1.000 1.000 1.000 1 0 0 0 1.000 1.000 1 0 0 0 1 0 0 0
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
07:
58 3
0 N
ovem
ber
2014
The results of TABLE I show that in all cases the power of
the parametric F statistic is the same under all scenarios because
the data are balanced. Further the parametric test has good power
and seems to be favorable because the parametric test does not
depend on other effects in the model.
However the simulation results also indicate that the power of
the rank transform FR statistic is superior in certain respects to
that of the parametric F statistic. Namely first of all in the absence
of interactions (cases a, 0, @', 0') there are many instances in
which the rank transform test appears to have greater power than
the parametric test, especially for small alternative values to the
null. Likewise in the presence of interactions if the number of main
effects and interactions is small (case a), the power of the rank
transform test is greater than or nearly equal to that of the
parametric test. Further if three main effects and one two-factor
interaction without a testing factor (i.e. B factor for our case) given
under the null hypothesis are simultaneously present (case a'), the
rank transform test still maintains the power advantage. However if
three main effects and interactions including a testing factor are
simultaneously present (cases a', @', B2), the rank transform test
seems to lose the power.
In general increases in the number of effects are associated with
decreases in the power of the rank transform test. Besides
increases in the value of effect size and sample size show manifest
increases in the power of both the normal theory (parametric) test
and the rank transform test. Note that in the nonnormal populations
such as exponential and double exponential ones, the rank transform
test indicates pronounced power advantage over the normal theory
test, especially for the small effect size. Note also that as compared
with other distributions, the power of light-tailed dstributions like
the uniform dstribution increases rapidly. Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
07:
58 3
0 N
ovem
ber
2014
270 CHOI
The results of TABLE II agree with those seen in TABLE I . Furthermore it is interesting to note that in all instances the rank
transform test produces reasonable power nearly equal to the normal
theory test. Particularly the gain of powers of the rank transform test becomes high in heavy-tailed distributions such as exponential
and double exponential ones. In addition the performance of the
rank transform test is remarkably improved with increase in sample
size.
Meanwhile TABLE III and TABLE IV report the power of tests
for interaction effects (i.e. AB effect for our case) when sampling is
from N= 2 and N= 10 observations per cell respectively. Likewise the following nine different situations are considered for the power
analysis: Q one interaction effect is nonnull; aB= c , a = 8 = r = ay = By = aBy = 0 , @I two interaction effects are nonnull;
aB=ay= c , a = B = y=Py=aBy=O, 13 one main effect and
one interaction effect are nonnull; B = aB = c , a = y = cry = By
= a h = 0 , @ ' two main effects and one interaction effect are
nonnull; B = y = aB = c , a = ay = By = aBy = 0 , @I2 another construction in which two main effects and one interaction effect are
nonnull; a = B = a B = c , r = a y = B y = a B y = O , Q' three main effects and one interaction effect are nonnull; a = B = y = aB = c ,
cry = By = aBy = 0 , 0' another construction in which three main effects and one interaction effect are nonnull; cr = c , B = 0.5 c ,
r = 1 . 5 c , a B = c , a r = B y = a B r = O , 8' three main effects and two interaction effects are nonnull; a = B = y = aB = ay = c ,
BY = aBy = 0 , a2 another construction in which three main effects and two interaction effects are nonnull; a = c , B = 0.5 c , y = 1.5 c ,
UP= c , cry=0.5c, BY= aBy=O.
The results of TABLE III and TABLE IV show that when there
are only interaction effects and there esist interaction effects with
one main effect (cases a, 0, a), the power of the rank transform
test is excellent or very similar with that of the normal theory test.
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
07:
58 3
0 N
ovem
ber
2014
Specifically it is shown that the rank transform procedure is
superior to the parametric procedure under the heavy-tailed
distributions like the exponential and double exponential distributions
of the error terms.
On the other hand when there exist interactions with two or
three main effects related to testing factors (i.e. A or B factor for
our case) simultaneously (cases a'-B2 of TABLE m), the power
of the rank transform test tends to be decreased somewhat
(especially for the cases of @I2 and @'I. [Side comments seem to
be appropriate here. First of all if there exist interactions with one
main effect unrelated to testing factors simultaneously, the rank
transform test obviously maintains much of its substantial power
even though the results are not presented here. Secondly it is
notable that when there exists only one interaction with two main
effects related to testing factors at the same time (case @I2 of
TABLE a), in spite of increase of effect size the power for only
N=2 seems to be bounded around 0.56 regardless of the type of distributions. However when we increase the number of
observations per cell beyond N=2 for the same case of 'a2, the
power rapidly approaches to 1 with increase of effect size. Thirdly when both three main effects and interactions related to testing
factors exist with great effect size, the power for N= 2 of the
uniform distribution seems to be bounded by a number of 0.94 less
than but very close to 1 (case a' of TABLE I and TABLE Ill).] This result closely agrees with that of Thompson (1991)'s paper
which may be applicable to a three-way layout. However note that
when the dstribution of the error terms is heavy-tailed with small
effect size or large sample size, the power of the rank transform
test is still prominent over the parametric test.
In overall these results of testing for interaction effects quite
agree with those of testing for main effects mentioned above. That
is, the power of testing for main effects and interaction effects
behaves in a similar fashion for most instances.
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
07:
58 3
0 N
ovem
ber
2014
4. CONCLUSIONS
CHOI
The results of this simulation study exhibit that the rank transform test in a z3 factorial design has considerable power advantages over the parametric test in many cases. Generally if the effect size or the number of effects is small, the sample size is large and the distribution of error terms is heavy-tailed, the rank transform procedure appears to maintain much of its substantial power and is more desirable. Further when the dstribution of the error terms is heavy tailed with small effect size or large sample size, the power of the rank transform test is superior to that of the parametric test.
However the rank transform test should be especially approached with caution when three main effects and interactions related to testing factors exist at the same time.
BIBLIOGRAPHY
Akritas, M.G. (1990). "The rank transform method in some two- factor designs," Journal of the American Statistical Association, 85, 73-78.
Akritas, M.G. and Arnold, S.F. (1994). "Fully nonparametric hypotheses for factorial designs," Journal of the Ameriwn Statistical Association, 89, 336-343.
Akritas, M.G., Arnold, S.F. and Brunner, E. (1997). "Nonparametric hypotheses and rank statistics for unbalanced factorial designs," To appear : Journal of the American Statistical Association
Blair, R. C., Sawilowsky, S. S. and Higgins, J. J. (1987). "Limitations of the rank transform statistic in tests for interactions," Communications in Statistics, Part B, Simulation and Computation, 16, 1133- 1145.
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
07:
58 3
0 N
ovem
ber
2014
Blair, R. C., Sawilowsky, S. S. and Higgins, J. J. (1989). "An investigation of the type I error and power properties of the rank transform procedure in factorial ANOVA," Journal of Eduaztional Statistics, 14, 255-267.
Box, G. E. P. and Mulier, M. E. (1958). "A note on the generation of random normal deviates," The Annals of Mathematiazl Statistics, 29, 610-611.
Brunner, E. and Neumann, N. (1984). "Rank tests for the 2 x 2 split-plot design," Metrika, 31, 233-243.
Brunner, E. and Neumann, N. (1986). "Rank tests in 2 x 2 designs," Statistiaz Neerlandiaz, 40, 251-271,
Conover, W. J. and Iman, R. L. (1976). "On some alternative procedure using ranks for the analysis of experimental designs," Communiaztions in Statistics, Part A, Theory and Methods, 5, 1349-1368.
Hora, S. C. and Conover, W. J. (1984). "The F statistic in the two-way layout with rank score transformed data," Journal of the American Statistical Association, 79, 668-673.
Law, A. M. and Kelton, W. D. (1991). Simulation Modeling and Analysis, McGraw-Hill, Inc..
Marsaglia, G. (1961). "Generating exponential random variables," Annals of Mathematical Statistics, 32, 899-902.
Pavur, R. and Nath, R. (1986). "Parametric versus rank transform procedures in the two-way factorial experiment: A comparative study ," Journal of Statistiazl Computation Simulation, 23, 231 - 240.
Pirie, W. R. and Rauch, H. L. (1984). "Simulated efficiencies of tests and estimators from general linear models analysis based on D
ownl
oade
d by
[N
orth
east
ern
Uni
vers
ity]
at 0
7:58
30
Nov
embe
r 20
14
274 CHOI
ranks: the two-way layout with interaction," Journal of Statistical Computation Simulation, 20, 197-204.
Thompson, G. L. (1991). "A note on the rank transform for interactions," Biometrika, 78, 697-701.
Thompson, G. L. (1993). "A correction note on the rank transform for interactions," Biometrika, 80, 711.
Received March, 1997; Rev ised J u l y , 1997.
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
07:
58 3
0 N
ovem
ber
2014