MULTIPLE COMPARISON PROCEDURES IN FACTORIAL DESIGNS USING THE ALIGNED RANK TRANSFORMATION by MARISELA ABUNDIS, B.S. A THESIS IN STATISTICS Submitted to die Graduate Faculty of Texas Tech University in Partial Fulfillment of the Requirements for the Degree of MASTER OF SCIENCE Approved May, 2001
63
Embed
MULTIPLE COMPARISON PROCEDURES IN FACTORIAL DESIGNS USING ...
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
MULTIPLE COMPARISON PROCEDURES IN FACTORIAL
DESIGNS USING THE ALIGNED RANK TRANSFORMATION
by
MARISELA ABUNDIS, B.S.
A THESIS
IN
STATISTICS
Submitted to die Graduate Faculty of Texas Tech University in
Partial Fulfillment of the Requirements for
the Degree of
MASTER OF SCIENCE
Approved
May, 2001
ACKNOWLEDGMENTS
First of all, I would like to thank my advisor, Dr. Mansouri for his help, support
and patience throughout my coursework and thesis. Thank you so much. Also. I
greatly appreciate the help I received from Dr. Westfall with some of my programming
required for the thesis. Dr. Duran, I thank you for your suggestions on my thesis and
answers to my questions. Thanks again to these outstanding professors.
Para mi familia, especialmente mis padres, agradezco mucho su apollo y comprension
durante todos los afios escolares de mi vida. Ustedes son la razon por la cual sigo
adelante. Los quiero mucho!
I would also like to thank my fiance, Jeffrey Martinez. Even though as undergrads
we went through very stressful and hard times, we managed to brighten each other
up and enjoy our college years. I love you!
Also, I greatly appreciate the friendship and love I have received from Jefirey's
parents. Thank you! To my best friend, Monet Alvarez, I thank you for always being
there for me since we were kids. I thank my other two closest friends, John Gomez
and Ina Aguirre. You two always keep my laughing!
I thank all the professors I have had at Texas Tech University, who have allowed
me to gain more mathematical and statistical skills. Thanks to Dr. Bennett, who
has always been a wonderful person to me, and without him I would not be here.
I would also like to thank my closest college friend, Ruby Martinez. I have always
enjoyed your company and will continue to do so as we proceed with our lives. Thanks
for j)utting up with me. I thank Richard Campos, Norma Aguirre, Bernard Omolo,
.Armando Arciniega, Carrie Mahood, and the custodians of the mathematics and
statistics department for being great friends to me.
11
CONTENTS
ACKNOWLEDGMENTS h
LIST OF TABLES iv
LIST OF FIGURES viii
I INTRODUCTION 1
II ANALYSIS OF VARIANCE AND MULTIPLE COMPARISON PROCE
DURES FOR A TWO-FACTOR FACTORIAL DESIGN 3
2.1 Analysis of Variance 3
2.2 Multiple Comparison Procedures 5
III ALIGNED RANK TRANSFORM TECHNIQUE 7
I\^ APPLICATIONS 11
4.1 A Balanced Two-Factor Factorial Design 11
4.2 Residual Analysis 15
4.3 An Unbalanced Two-Factor Factorial Design 25
V SIMULATIONS AND POWER STUDY 40
5.1 Simulations 41
5.2 Results 44
\T CONCLUSION 51
BIBLIOGRAPHY 52
111
LIST OF TABLES
2.1 Analysis of Varianc(> (ANOVA) for a Fixed Effects Two-Factor Facto
rial Design 4
4.1 Data for the Industrial Waste Experiment U
4.2 Analysis of \ariance for Industrial Waste Experiment based on LS.
Dependent Variable: waste 13
4.3 Analysis of Variance for Industrial Waste Experiment on testing inter
action, based on ART. Dependent Variable: W 13
4.4 Analysis of Variance for Industrial Waste Experiment on testing for
main effect of Temperature based on ART. Dependent Variable: i?" 14
4.5 Analysis of Variance for Industrial Waste Experiment on testing for
main eflfect of Environment based on ART. Dependent Variable: R^ 14
4.6 Tukey's Studentized Range (HSD) Test for Temperature based on
LS. Alpha=0.05 DF=15 MSE=1.174783, Critical Value of Studentized
Range=3.67338, Minimum Significant Diflference= 1.2591, Means with
the same letter are not significantly diflferent 15
4.7 Tukey's Studentized Range (HSD) Test for Temperature based on
ART. Alpha=0.05 DF=15 MSE=58.9, Critical Value of Studentized
Range=3.67338, Minimum Significant Diflference=8.915, Means with
the same letter are not significantly diflferent 15
4.8 Scheffe's Test for Temperature based on LS. Alpha=0.05 DF=15 MSE=1.174783,
Critical Value of F= 3.68232, Minimum Significant Diflference=1.3154,
Means with the same letter are not significantly diflferent. NOTE: This
test controls the Type I experimentwise error rate 16
IV
4.9 Scheflfe's Test for Temperature based on .ART. .Alpha=0.05 DF=15
MSE=58.9. Critical \ alue of F=3.68232, Minimum Significant Diflfer-
ence=9.3143. Means with the same letter are not significantly different.
XOTE: This test controls the Type I experimentwise error rate. . . . 16
4.10 Tukey's Studentized Range (HSD) Test for Environment based on
LS. Alpha=0.05 DF-=15 MSE=1.174783, Critical \ alue of Studentized
Range=4.36699, Minimum Significant Difference=1.9323, Means with
the same letter are not significantly diflferent 17
4.11 Tukey's Studentized Range (HSD) Test for Environment based on
ART. Alpha=0.05 DF=15 MSE=61.83333. Critical \ alue of Studen
tized Range=4.36699, Minimum Significant Diflference=14.019. Means
with the same letter are not significantly diflferent 17
4.12 Scheflfe's Test for Environment based on LS. Alpha=0.05 DF=15 MSE=1.174783,
Critical \'alue of F=3.05557, Minimum Significant Difference=2.1877,
Means with the same letter are not significantly different. NOTE: This
test controls the Type I experimentwise error rate IS
4.13 Scheffes Test for Environment based on ART. Alpha=0.05 DF=15
MSE=61.83333, Critical \'alue of F=3.05557, Minimum Significant
Difference=15.872. Means with the same letter are not significantly
different. NOTE: This test controls the Type I experimentwise error
rate 18
4.14 Testing hypothesis at a = 0.05 significance level for both least squares
and ART methods/Industrial Waste Experiment 19
4.15 Multiple Comparison Procedures for both least squares and ART meth
ods/Industrial Waste Experiment 19
4.16 Analysis of Variance for testing equality of variances. Dependent \'ari-
able: Y 24
4.17 Drug Study Data 26
4.18 Drug Study Data (cell means for zj) 27
4.19 Analysis of Variance for Drug Study based on LS. Dependent X'ariable:
Y 28
4.20 Analysis of Variance for Drug Study on testing interaction, based on
ART. Dependent Variable: W 28
4.21 Analysis of Variance for Drug Study on testing for the main eflfect of
Drug, based on ART 29
4.22 Analysis of Variance for Drug Study on testing for the main eflfect of
Disease, based on ART. Dependent Variable: R^ 29
4.23 Least Squares Means for Eflfect drug Pr > \t\ for HO: LSMean(i)=LSMean(j)
Dependent Variable: Y based on LS, Adjustment for Multiple Com
parisons: Tukey 30
4.24 Least Squares Means for Eflfect drug Pr > \t\ for HO: LSMean(i)=LSMean(j)
Dependent Variable: R^ based on ART, Adjustment for Multiple Com
parisons: Tukey 30
4.25 Least Squares Means for Eflfect drug Pr > \t\ for HO: LSMean(i)=LSMean(j)
Dependent Variable: Y based on LS, Adjustment for Multiple Com
parisons: Scheflfe 31
4.26 Least Squares Means for Eflfect drug Pr > \t\ for HO: LSMean(i)=LSMean(j)
Dependent Variable: R^ based on ART, Adjustment for Multiple Com
parisons: Scheflfe 31
VI
4.27 Least Squares Means for Effect drug Pr > \t\ for HO: LSMean(i)=LSMean(j)
Dependent Variable: Y based on LS, Adjustment for Multiple Com
parisons: Tukey 32
4.28 Least Squares Means for eflfect Disease Pr > |t| for HO: LSMean(i)=LSMean(j)
Dependent Variable: R^ based on ART, Adjustment for Multiple Com
parisons: Tukey 32
4.29 Least Squares Means for eflfect Disease Pr > \t\ for HO: LSMean(i)=LSMean(j)
Dependent Variable: Y based on LS, Adjustment for Multiple Com
parisons: Scheflfe 33
4.30 Least Squares Means for effect Disease Pr > \t\ for HO: LSMean(i)=LSMean(j)
Dependent Variable: R^ based on ART, Adjustment for Multiple Com
parisons: Scheflfe 33
4.31 Testing hypothesis at a = 0.05 significance level for both least squares
and ART methods/Drug Study 34
4.32 Multiple Comparison Procedures for both least squares and ART meth
ods/Drug Study 34
5.1 Design 1 43
5.2 Design 2 43
5.3 Design 3 43
5.4 Design 4 43
5.5 Design 5 43
5.6 Power simulation for all pairwise comparisons of Design 1 based on the
least squares and aligned rank transform methods 47
5.7 Power simulation for all pairwise comparisons of Design 2 based on the
least squares and aligned rank transform methods 47
Vll
5.8 Power simulation for all pairwise comparisons of Design 3 based on the
least squares and aligned rank transform methods 48
5.9 Power simulation for all pairwise comparisons of Design 4 based on the
least squares and aligned rank transform methods 48
5.10 Power simulation for all pairwise comparisons of Design 5 based on the
least squares and aligned rank transform methods 49
5.11 Power simulation for all pairwise comparisons of Table 4.1, Data for
Industrial Waste Output based on the least squares and aligned rank
transform methods 49
5.12 Power simulation for all pairwise comparisons of Table 4.5, Drug Study
Data based on the least squares and aligned rank transform methods 50
viu
LIST OF FIGURES
4.1 Residuals versus YEAR 20
4.2 Residuals versus Temperature 21
4.3 Residuals versus Environment 22
4.4 Normal Probability Plot 23
4.5 Residuals versus YEAR 35
4.6 Residuals versus Drug 35
4.7 Residuals versus Disease 37
4.8 Normal Probability Plot 38
IX
CHAPTER I
INTRODUCTION
A factorial design is used for an experiment that involves the study of two or more
factors simultaneously, with each factor having two or more levels. The importance
of factorial designs is that they permit simultaneous examination of the effects of
individual factors and their interactions. All possible combinations of the levels of
the factors are investigated in each replication of an experiment. .A main effect is
defined as the change in response produced by a change in the level of one factor
while keeping the remaining factors at a fixed level. Interaction exists between two
factors if the difference in response between the levels of one factor is not the same
at all levels of the other factors. Thus, the preliminary focus of analysis is testing
hypotheses about interaction, equality of row treatment effects, and column treatment
eflfects. If interaction exists, main effects exists as well. If interaction does not exist.
then testing main effects should proceed. In Chapter II, we present the analysis of
variance for a two-factor factorial fixed effects factorial design.
The hypotheses do not provide detailed information about the difference in inter
actions and main eflfects. Multiple comparison procedures are capable of responding
to specific questions about more meaningful comparisons on any of the above eflfects.
These procedures allow the comparisons between groups or pairs of treatment means.
The comparisons are made in terms of treatment totals or treatment averages. It
must be noted that multiple comparison techniques are not dependent on the re
jection of null hypothesis. The testing of interaction between two factors and main
eflfects can also be performed by multiple comparison methods. Common multiple
comparison procedures that will be implemented are Tukey's Studentized Range T( st
and Scheflfe's Method. These multiple comparison procedures are discussed in Chap
ter II. Additionally, we will employ a macro called %SimPower in order to perform
multiple comparisons. This macro was developed by Tobias (see Westfall et al.. 1999).
%SimPower simulates power for multiple comparisons. It uses complete, mini
mal, and proportional power definitions. Complete power is defined as the probability
of rejecting all false null hypotheses. Minimal power is the probability that at least
one false null hypothesis is rejected implying a significant result. Proportional power
is the proportion of false null hypotheses detected to all false null hypotheses, that is
false nulls expected to be detected. Further discussion is provided in Chapter \ ' .
One of the main purposes of this investigation is to carry out multiple comparisons
for tw^o-factor factorial designs based on the aligned rank transformation. The aligned
rank transform procedure provides a robust and powerful alternative method of data
analysis to the classical least squares method. In this investigation, we will study the
validity and power of the aligned rank transform technique for multiple comparisons.
In Chapter III, we define the classical least squares F-statistic and the aligned rank
transform technique.
Analysis of two applications based on the least squares and aligned rank transform
methods will be examined in Chapter IV. The final conclusion will be provided in
Chapter VI. We will utilize the SAS programming language to perform the classical
least squares method and the aligned rank transform technique. P R O C GLM,
PROG REG, and P R O C R A N K will execute the two techniques. These ideas
will be explained in more detail later on.
CHAPTER II
ANALYSIS OF VARIANCE AND MULTIPLE COMPARISON PROCEDURES
FOR A TWO-FACTOR FACTORIAL DESIGN
In this chapter, we present the analysis of variance and multiple comparison pro-
cefures for a fixed effects two-factor factorial design.
2.1 Analysis of Variance
Suppose that an experiment involves the simultaneous study of two factors A and
B. Let yijk denote the k-th replication in response under the i-th level of factor A and
j-th level of factor B. It is assumed that yijk follows the linear model
Vijk = Id + ai -\- Pj -\- {al3)ij + Cijk,
i=l,...,a
j=l,...,b
k=l,...,n
(2.1)
where,
a is the number of levels of factor A,
b is the number of levels of factor B,
n is the number of replications,
// is the overall mean,
a, is the main eflfect of the (ith) level of factor A,
Pj is the main eflfect of the (jth) level of factor B,
{aP)ij is the interaction between the i-th level of factor A and j-th level of factor B,
and Cijk are the random error components.
The linear model has the following restrictions:
EU ^^ = 0
Table 2.1: Analysis of Variance (ANOVA) for a Fixed Effects Two-Factor Factorial Design
Source of Variation
Sum of Squares
Degrees of Freedom
Mean Squares Fa
A
B
AB
Error
Total
SSA
SS B
SS AB
SSt
SSi
a- 1
6 - 1
( a - l ) ( 6 - l )
ab(n — 1)
abn — 1
SSA MSA = ^
^ a—I
MSB = I^
A f c _ SSAB IMbAB - ( a - l ) ( 6 - l )
MSE - ^^^ ab{n—l)
MSA
MSE
MSB
MSE
-U5.45
MSE
It is assumed that eij are independent and indentically distributed random variables
with ^"(0,^^) distribution, where N{/j,,a'^) denotes a normal distribution with mean
Ij, and variance cr .
The quantities in the table are presented as:
SSA - TZ Jli^i y y:.. abn
abn
hj. yl abn
abn
SSB - j5 E ; = i ytj. -
SS.B = i EU EU Vl - £ - SSA - SS
SSs = SSr-i; EU EU y
SST = zJi=l l^j = l Z A:=1 Vijk
where,
E a \~^^ v"^" i = l 2 . ^ j= l Z^k^l Vij^
i*"" abn
E n k-i y^ji^
y^J- ~ n
E a ^r-^n i= l 2^k=l yijii
y.j.
B
y.j. an
4
(2.2)
(2.3)
(2.4)
yi.. — /^j=i z^k=\ y^ji^
yi.. fjj^
for i := 1,2, • • -,0 j = 1,2, • • -,6.
From the ANOVA Table 2.1, several hypotheses can be tested by using the ap
propriate F-statistic. The following tests of hypotheses are of interest:
Test the interaction between Factor .A and Factor B Effects:
Ho: (a/3),, = 0
Hi : (oi.p)ij ^ 0 for some ij:
Test the equality of Factor A main Eflfects:
HQ : Off = 0
Hi : tti 7 0 for some i\
Test the equality of Factor B main Effects:
Ho: ft = 0
Hi : Pj ^ 0 for some j .
The corresponding test statistic and rejection region for each hypothesis will be in
troduced in Chapter HI.
2.2 Multiple Comparison Procedures
In this section, we consider a brief discussion of multiple comparison procedures.
As previously mentioned in Chapter I, testing the overall hypothesis does not give
sufficient information about the difference in interactions and main effects. Multiple
comparison procedures allow the experimenter to gain detailed information on which
treatments or treatment combinations actually diflfer.
Suppose that comparing all pairs of treatment means are of interest and wv would
like to test
Ho,ij : iit = Pj, \fiy^j = l,....a (2.0)
Hi,ij : Hi / Pj.
0
For our purposes, Tukey's studentized range statistic is used tu test the above
hypotheses. On experimentwise premise, Tukey's test has a family-wise Type I error
rate of a for all pairwise comparisons (see Montgomery, 1997). It strongly controls
the family-wise Type I error rate and is the most powerful for pairwise comparison.
Rejection of 7/o,ij results if |^j. — yj.\ > Ta = ^0(0, f)Sy^,, where qa{a, f) is the upper
a percentage point of the studentized range for groups of means, Sy^ = v/^^^^ is the
standard error of each average, a is the number of means in a group, / is the degrees
of freedom for error.
If the comparison involves more than two means, then Scheffe's test is the best
method for such comparisons. Scheflfe's test compares any and all contrasts between
means. .A contrast is a linear combination of treatment totals. That is, C = Yll^=i ^iVf
with the constraint J2^=i^i — 0- - 0 is rejected if \Cu\ > Sa,u' where C^ is any
possible linear combination of interest and Sa,u = Sc^ \J{a — l)Fa^a-i,iv-a< where
Scy, = SJMSE miLi(Qu^/^i)- The Type I error is at most a for all possible compar
isons. Scheffe's test could also be applied to pairwise comparisons of means, though
it is not the most sensitive technique.
Tukey's and Scheffe's test are applied to the applications in Chapter I\', while in
simulation studies of Chapter V, we only apply Tukey's test.
6
CHAPTER III
ALIGNED RANK TRANSFORM TECHNIQUE
In this chapter, we define the classical least squares (LS) F-statistic and present
the aligned rank transform (ART) technique. From this point on, we will make
comparisons between the least squares and aligned rank transform techniques on hy
potheses testing, multiple comparison procedures, and power.
The ratio FQ = '^^^Tr.a^tment, obtained from Table 2.1 is distributed as Snedecor's
F with a - I and ab{n - 1) degrees of freedom. We use this statistic to test the
hypothesis of interest, (2.2), (2.3), and (2.4).
An alternative method to the usual least squares technique is the aligned rank
transform (ART) technique (see Mansouri, 1998). The ART is a robust and powerful
technique. This method is not sensitive to outliers and violations of assumptions on
error distribution. The aligned rank transform statistic is obtained the same way as
the least squares statistic, with the only exception that it replaces the observations
with ranks of the residuals (aligned observations) of the reduced model. Thus, hy
pothesis testing and multiple comparison procedures can be performed for any linear
model based on the ART technique.
For the aligned rank transform method, the statistical linear model and assump
tions (without the normality of the random error component) of a particular design is
the same as model (2.1). Alternatively, the latter is presented as the full rank linear
model:
y = 1/i + X^a -f Xp^ + X^j + € (3.1)
where,
Y is an TV X 1 vector of observations {N = abn)
7
1 is an A X 1 vector of ones
p is the y-intercept
cx= (ai, ...,aa-i)'
^=(Pu...jb-iy
7 = (7ll, •••,7l,6-l,-",7a-l,l'---'7a-l,6-l)'
e is an A X 1 vector of independent random variables with a continuous, unknown
distribution function
A'Q, Xp, X-y are full rank matrices with -1,0,1 entries, according to the specific design.
A reduced model is a linear model that includes all the unknown parameters except
the parameter of interest which is held constant. That is when the null hypothesis
(HQ) holds, then the reduced model is obtained.
The ART technique is developed in the following manner:
1. Fit the reduced model and obtain the corresponding residuals.
2. Rank the residuals.
3. Fit the reduced and the full models to the ranks.
The ART test statistic is:
J, ^[SSEreMR)]-SSEfMa{R)] ^"^ qMSEfuiila(R)] ^ ^^
where,
SSEreMR)] = «'(^)(^ - Hi)a(R)
SSEf^uHR)] = a'(R){I - H)a(R)
MSEf^uHR)] = {N- P)-'SSEfuu[a{R)]
P is the rank of H
q is the rank of H — Hi
[a{R)] is the vector of rank scores
Hi=Xi{X[Xi)-'^X[
H = X{X'X)-'^X'.
Hi and H are the respective hat matrices for the reduced and full models.
8
We now specify the corresponding test statistics and rejection regions for hypothe
ses (2.2), (2.3) and (2.4) based on both the classical least squares technique and the
aligned rank transform. Let T.S and R.R denote the test statistic and rejection re
gion for the classical least squares method, and T.SART and R.RART denote the
test statistic and rejection region for the ART method.
Referring to null and alternative hypotheses mentioned previously, we test the
following at a particular significance level (a):
Interaction between factor A and B effects:
^ c . P — MSA*B
T Q A R T • r ^ — MSRC[a{R-^)] l .O / \ rLl . r ^ ^ — MSElaiR-^)]
R . R : Fo > FaXa-\){b-\),ab{n-l)
R . R A R T : F ^ ^ > Fa^i^a-\){b-\),ab{n-\)
Main effect of factor A:
T Q • ;r — ^SA
T Q A R T • /?« - ^SR[a{R'^) i . D / \ n i . Fj^j^ — f^sE[a{R^)]
R . R : FQ > Fa,{a-l),ab{n-l)
R . R A R T : F J ^ > FaXa-\),ab{n-l)
Main effect of factor B:
T c; . F — ^SB
T Q A R T • F ^ - ^SC[a{R^)
(3.3)
(3.4)
(3 .5 )
R . R : FQ > Fa,{b-l),ab{n-l)
R . R A R T : F ^ ^ > Fa,(b-\),ab{n-l)
The following SAS program for the analysis of a two-way layout based on the .ART
method is given in Mansouri (1998):
data new; /*create a data set*/ do A = ai to Gr] /*A has r levels, B has c levels and rep has n replications*/ do B = bi to br] do rep = 1 to n; input Y@@;
9
output; end; end; end; datalines; data two way; set new;
/*define A^V if ,4 = "ai" then X^i = 1; if ^ = "a^" then X^i = - 1 ; else A' i = 0:...if -4 = "ar-i" then Xar-i = f; if ^ = "flr" then A^r-i = — 1; else X^r-i — 0;
/*define A>*/ if B = "6i" then Xpi = l;\iB = "6c" then A>i = - 1 ; else X^i = 0;...if B = "'6^-1'' then A/3C-1 = 1; ii B = ''be" then A^c-i = —^] else Xpc-i = 0;
/*define A / / A'-yii = A ' Q I A / 3 i ; . . . X ^ r - l , c - l = - ^ a r - l ^ / S c - l J
/*ART test statistic for testing i/o"*/ proc reg data=twoway; model Y = Xpi - A^c-1^711 ~ -^7r-i,c-i; output out=aligna r = K"; proc rank data=aligna out==rankya; ranks R°'; var V"; proc glm data=rankya; classes .A B; model R^=A B A*B; Ismeans A/lsd tukey scheffe;
/*ART test statistic for testing HQ^*/ proc reg data=twoway; model Y = A^ i ~ ^Qr-1^711 ~ ^7r-i,c-i; output out=alignb r = Y^\ proc rank data=alignb out=rankyb; ranks R^\ var Y^] proc glm data=rankyb; classes A B; model R^=A B A*B; Ismeans B/lsd tukey scheflfe;
/*ART test statistic for testing i/o^*/ proc reg data=twoway; model Y = AQI - Xar-iXpi — Xpc-i] output out=aligng r = Y^; proc rank data=aligng out=rankyg; ranks R^; var Y'^; proc glm data=rankyg; classes A B; model R'^=A B A*B;
10
CHAPTER IV
APPLICATIONS
In this chapter, the least squares and the ART method are applied to data sets
resulting from factorial experiments. The main objective is to show that the ART
procedure provides valid results for multiple comparisons and should be considered
as a feasible alternative to the procedures based on the least squares. We present
two types of designs, a balanced and an unbalanced two-factor factorial design. All
procedures are performed on SAS and the results are analyzed.
4.1 A Balanced Two-Factor Factorial Design
Industrial Waste Experiment: To remain competitive, businesses have adopted
a philosophy of continuous improvement of their manufacturing and service processes.
An important element of this philosophy is experimentation, to better understand
systems and to optimize performance. The data in the following study are from
an experiment designed to study the eflfect of temperature (at low. medium, and
high levels) and environment (five different enivornments) on waste ouput in a man
ufacturing plant (see Table 4.1). Two replicate measurements are taken at each
temperature/environment combination (Westfall et al., 1999, pp. 176 e 178).
Table 4.1: Data for the Industrial Waste Experiment
Temperature Low (1)
Medium (2)
High (3)
1 7.09 5.90 7.01 5.82 7.78 7.73
Environment 2
7.94 9.15 6.18 7.19 10.39 8.78
3 9.23 9.85 7.86 6.33 9.27 8.90
4 5.43 7.73 8.49 8.67 12.17 10.95
5 9.43 6.90 9.62 9.07 13.07 9.76
11
The questions of interest are:
a. Overall, are there significant differences among the temperature levels?
b. Overall, are there significant differences among the environment levels?
c. Do the differences between temperature levels depend on the environment (or vice
versa)?
The statistical linear model that is used to investigate this problem is the same as
(2.1):
yijk = P +(^1 +Pj-^ {c^(^)ij-^Ujk 2 = 1,2,3 7 = 1,2,3,4,5 k = 1.2
where,
y^jk - the (kth) waste output of (ith) temperature, (jth) environment
p - overall mean of waste outputs
ai - main effect of the (ith) temperature
13j - main effect of the (jth) environment
(Q^P)ij - level of interaction between the (ith) temperature and (jth) environment
Eijk - random error component,
with the following conditions:
E l l « . = 0
EU ft = 0
It is assumed that e,, ~ NID{0,a^).
In order to answer questions (a) - (c), hypotheses testing are performed at a
0.05 significance level. Testing interaction is performed first. Then main effects is
tested if there is no interaction. The classical least squares technique and the aligned
rank transform procedure explained in Chapter II are applied to the Industrial Waste
Experiment. Utilizing P R O C GLM , P R O C R E G , and P R O C R A N K in SAS
allows us to obtain the ANOVA tables for the least squares and aligned rank transform
techniques, along with multiple comparison techniques such as tukey's studentized
range test and scheffe's test. We can make inferences from these results.
12
Table 4.2: Analysis of Variance for Industrial Waste Experiment based on LS. Dependent Variable: waste
F Value Pr > F Source Model Error Corrected Total
Source envir*temp temp
DF 14 15 29
DF 8 2
Sum of Squares 78.28974667 17.62175000 95.91149667
Type I SS 22.91156000 30.69280667
Mean Square 5.59212476 1.17478333
Mean Square 2.86394500 15.34640333
4.76 0.0024
envir
F Value Pr > F 2.44 0.0651 13.06 0.0005
4 24.68538000 6.17134500 5.25 0.0075
Table 4.3: Analysis of Variance for Industrial Waste Experiment on testing interaction, based on ART. Dependent Variable: R^
Source Model Error Corrected Total
Source Temp Envir Temp*Envir
DF 14 15 29
DF 2 4 8
Sum of Squares 1295.000000 952.500000 2247.500000
Type I SS 12.800000 13.000000 1269.200000
Mean Square 92.500000 63.500000
Mean Square 6.400000 3.250000 158.650000
F Value 1.46
F Value 0.10 0.05 2.50
Pr > F 0.2391
Pr > F 0.9047 0.9945 0.0602
The classical least squares and aligned rank transform ANOVA tables for the
industrial waste experiment are given in Tables 4.2 - 4.5.
The least squares (LS) and aligned rank transform (ART) multiple comparison
procedures for temperature and environment are given in Tables 4.6-4.13.
13
Table 4.4: Analysis of Variance for Industrial Waste Experiment on testing for main eflfect of Temperature based on ART. Dependent \'ariable: R^
Source Model Error Corrected Total
Source Temp Envir Temp* Envir
DF 14 15 29
DF 2 4 8
Sum of Squares 1364.000000 883.500000 2247.500000
Type I SS 1349.600000 4.000000 10.400000
Mean Square 97.428571 58.900000
Mean Square 674.800000 1.000000 1.300000
F Value 1.65
F Value 11.46 0.02 0.02
Pr > F 0.1723
Pr > F 0.0010 0.9994 1.0000
Table 4.5: Analysis of Variance for Industrial Waste Experiment on testing for main effect of Environment based on ART. Dependent Variable: R^^
Source Model Error Corrected Total
Source Temp Envir Temp*Envir
DF 14 15 29
DF 2 4 8
Sum of Squares 1320.000000 927.500000 2247.500000
Type I SS 15.000000 1285.333333 19.666667
Mean Square 94.285714 61.833333
Mean Square 7.500000 321.333333 2.458333
F Value 1.52
F Value 0.12 5.20 0.04
Pr > F 0.2135
Pr > F 0.8866 0.0079 1.0000
Table 4.14 and Table 4.15 summarize the results from the conducted analysis. Let
us define the following abbreviations as:
LS is the the value of the least squares test statistic,
ART is the value of the aligned rank transform test statistic,
CV is the critical value used to determine if HQ is rejected or not,
P-VAL-LS is the smallest level of significance at which HQ can be rejected for the
usual least squares method,
P-VAL-ART is the smallest level of significance at which HQ can be rejected for the
aligned rank transform method,
{ti,tj) denotes temperature i as significantly different from temperature j .
14
Table 4.6: Tukey's Studentized Range (HSD) Test for Temperature based on LS. Alpha=0.05 DF=15 MSE=1.174783, Critical Value of Studentized Range=3.67338, Minimum Significant Difference=1.2591, Means with the same letter are not significantly different.
Tukey Grouping A B B B
Mean 9.8800 7.8650
7.6240
N 10 10
10
Temp 3 1
2
Table 4.7: Tukey's Studentized Range (HSD) Test for Temperature based on ART. Alpha=0.05 DF=15 MSE=58.9, Critical Value of Studentized Range=3.67338, Minimum Significant Difference=8.915, Means with the same letter are not significantly different.
Tukey Grouping A B B B
Mean 24.900 11.900
9.700
N 10 10
10
Temp 3 1
2
(ej,ej) denotes environment i as significantly different from environment j .
We make conclusions about the analysis after performing residual analysis, which
will be explained in section 4.2.
4.2 Residual Analysis
Before we formulate a formidable conclusion about the analysis of variance for
the industrial waste experiment, we must check the validity of the underlying con
ditions for linear model (2.1) based on the least squares technique. To verify these
assumptions, we perform residual analysis such that Cij are normally, identically and
independently distributed with p = 0 and constant variance cr .
Hence, we examine the residuals, which are defined as
eijk = yijk - yijk^ (44)
15
Table 4.8: Scheflfe's Test for Temperature based on LS. Alpha=0.05 DF=15 MSE=1.174783, Critical Value of F= 3.68232. Minimum Significant Difference=1.3154, Means with the same letter are not significantly different. NOTE: This test controls the Type I experimentwise error rate.
Scheffe Grouping Mean N Temp ~7. 9.8800 To 3
B 7.8650 10 1 B B 7.6240 10 2
Table 4.9: Scheflfe's Test for Temperature based on ART. Alpha=0.05 DF=15 MSE=58.9, Critical Value of F=3.68232, Minimum Significant Diflference=9.3143, Means with the same letter are not significantly different. NOTE: This test controls the Type I experimentwise error rate.
Scheflfe Grouping Mean N Temp ~A~ 24.900 To 3
B 11.900 10 1 B B 9.700 10 2
where
yiji^ is the observation
y^jk is the predicted value or estimate of the corresponding observation
Note that yijk = yij., the average of the observations in the ijth. cell implies that
^ijk ^^ yijk ~ yij.-
We provide the plots for residuals versus y (Figure 4.1), temperature (Figure 4.2),
environment (Figure 4.3) and the normal probability plot of residuals (Figure 4.4) for
the industrial waste experiment. A normal probability plot plots each residual versus
its expected value under normality.
16
A A A A A A A
9.6417
8.9067
8.5733
8.2717
6.8883
6
6
6
6
6
Table 4.10: TUKEY'S STUDENTIZED RANGE (HSD) TEST for Environment based on LS. Alpha=0.05 DF=15 MSE=1.174783, Critical \ alue of Studentized Range=4.36699, Minimum Significant Difference=1.9323, Means with the same letter are not significantly diflferent.
Tukey Grouping Mean N Envir 5
4
B A 8.5733 6 3 B B A 8.2717 6 2 B B 6.8883 6 1
Table 4.11: Tukey's Studentized Range (HSD) Test for Environment based on ART. Alpha=0.05 DF=15 MSE=61.83333, Critical Value of Studentized Range=4.36699, Minimum Significant Difference=14.019, Means with the same letter are not significantly diflferent.
Tukey Grouping Mean N Envir T
4 A
B A 17.000 6 3 B B B B
A A A A A A A
23.000
19.667
17.000
14.000
3.833
6
6
6
6
6
Table 4.12: Scheffe's Test for Environment based on LS. Alpha=0.05 DF=15 MSE=1.174783, Critical Value of F=3.05557, Minimum Significant Difference=2.1877, Means with the same letter are not significantly diflferent. NOTE: This test controls the Type I experimentwdse error rate.
Scheflfe Grouping Mean N Envir A 9.6417 6 5 A
B A 8.9067 6 4 B A B A 8.5733 6 3 B A B A 8.2717 6 2 B B 6.8883 6 1
Table 4.13: Scheflfe's Test for Environment based on ART. Alpha=0.05 DF=15 MSE=61.83333, Critical Value of F=3.05557, Minimum Significant Diflference=15.872, Means with the same letter are not significantly different. NOTE: This test controls the Type I experimentwise error rate.
Scheflfe Grouping Mean N Envir
5
4
3
2
A A A A A A A
23.000
19.667
17.000
14.000
3.833
6
6
6
6
6
B B B B B B B 3.833 6 1
18
Table 4.14: Testing hypothesis at a = 0.05 significance level for both least squares and ART methods/Industrial Waste Experiment
EFFECTS Interaction
Temp. Environ.
LS 2.44 13.05 5.25
ART 2.50 11.46 5.20
CV 2.64 3.68 3.06
P-VAL-LS 0.0651 0.0005 0.0075
P-VAL-ART 0.0602 0.0010 0.0079
Table 4.15: Multiple Comparison Procedures for both least squares and ART methods/Industrial Waste Experiment
EFFECTS Temp. (1,2,3)
Environ. (1,2,3,4,5)
Tukey-LS ( ^ 3 , ^ l )
(^3,^2)
( 6 5 , 6 1 )
( 6 4 , 6 1 )
Tukey-ART (^3,^1)
(^3,^2)
( 6 5 , 6 1 )
( 6 4 , 6 1 )
Scheflfe-LS (^3,^1)
(^3,^2)
( 6 5 , 6 1 )
Scheflfe-ART (^3 ,^1)
(^3 ,^2)
( 6 5 , 6 1 )
19
ybar
Figure 4.1: Residuals versus YBAR
20
2 . 5
2 . 0 -
O
1 . 5
O
o
Temperature
Figure 4.2: Residuals versus Temperature
21
Environment
Figure 4.3: Residuals versus Environment
22
i 1 /
/ /
/ /
/ /
9 /
/ /
G /
/
o
o /
/ o^
o
/ / o
o GJ'OO /
O / o/
/
e
- 2 i / /
/
/ /
/ 1/ - 3 —I 1 1 1 1 r
3 - 2 - 1 0 1 2 3
N o r m a l Q u a n t i l e s
N o r m a l L i n e : — Mu=0, S i g m a = l
Figure 4.4: Normal Probability Plot
23
Table 4.16: .Analysis of Variance for testing equality of variances. Dependent Variable: Y
Sum of Source DF Squares Mean Square F \'alue Pr > F
Model 4 3.68815333 0.92203833 0.70 0.5986 Error 25 32.88431667 1.31537267 Corrected Total 29 36.57247000
Source DF Type I SS Mean Square F \'alue Pr > F environment 4 3.68815333 0.92203833 0.70 0.5986
From Figures 4.1 and 4.2, we observe that the plots do not contain outliers. \\V
note that Figure 4.3 clearly shows a trumpet-like pattern, implying that the error
variance is not constant for different environments. We can further investigate or
reexamine the data from the study of regression analysis. Let us perform the modified
Levene's test. Levene's test conducts a test for equality of variances in all treatments
(environment) at a 0.05 significance level. The modified Levene test (see Montgomery,
1997) calculates the absolute deviation of observations yijk in each treatment from the
treatment median {yjk)- Similarly the previous is denoted as dijk = \yijk-yjk\^ where
2 = 1,2,3 j = 1,2,3,4,5 k = 1,2. The test statistic is the classical F-statistic,
used in Table 4.14. Let us provide the reader with the corresponding ANOVA Table
4.16.
From the ANOVA table, we obtain a p-value=0.5986 greater than the 0.05 signif
icance level, thus implying that we do not reject HQ. That is, variances are constant,
hence satisfying the assumption. We observe from the normal probability plot (Fig
ure 4.4) that a straight line is represented by the residuals, implying that the errors
are normally distributed.
We can now answer questions (a) - (c). From Table 4.14, we observe that the least
squares and aligned rank transform techniques show no interaction to be present be
tween temperature and environment at the 0.05 level of significance. Since interaction
24
does not exist, we test for main effects. There is suflficient evidence to support the
claims that there are differences in temperature levels and diflferences in environment
levels. Tukey's studentized range test and Scheflfe's test reveal that mean waste output
for temperatures 1 and 3 and temperatures 2 and 3 are significantly different based on
both the LS and ART methods. The same two multiple comparison procedures under
both techniques show that environments 1 and 5 are significantly different, along with
a significant diflference of environments 1 and 4 demonstrated by Tukey's studentized
range test, but not Scheflfe's test. Note that Scheffe's test considers comparisons of
all possible contrast between means, thus implying that not all significances will be
determined. Scheffe's test is not sensitive to pairwise comparisons.
Both the usual least squares technique and the aligned rank transform method
provide similar results and conclusions. This is an example that supports the use of
the ART technique as an alternative to the least squares technique.
4.3 An Unbalanced Two-Factor Factorial Design
In practice, an experimenter may be presented with missing data or different
number of observations within each treatment, such that a design is incomplete or
unbalanced. We must know how to handle such an instance. Let us consider and
examine an unbalanced design. In order to perform the analysis on an unbalanced
design, we calculate cell means and proceed with this analysis.
Outcomes of Alternative Drugs on Different diseases: Developing phar
maceutical drugs is a long and laborious process, often taking 10 years from discovery
to marketing. During the process the drug must be tested, first on animals, and later
on humans, for evidence of safety and efficacy. Human testing alone requires four
phases of clinical trials (labeled Phase I through Phase IV). Development can be
stopped at any point in the process should the candidate drug be determined unsafe
and/or ineflfective. Multiplicity adjustment is important because it is costlv for thv
drug company to pursue false leads. However, the tests should be as powerful as
possible to avoid halting development of a promising drug.
Phase I trials tend to be small and exploratory. The data in Table 4.17, shows
how patients with different diseases respond to alternative drug formulations. While
there were plans for six patients at each drug/disease combination, in many cases
there were fewer than that number (Westfall et al., 1999, pp. 187 & 188).
26
Table 4.18: Drug Study Data (cell means for ij)
Disease Drug 1 2 3 A (1) 29.33 18.83 17.00 B (2) 23.33 22.33 18.17 C (3) 8.17 3.67 5.67 D (4) 11.33 12.83 11.83
The questions of interest are:
a. Are there differences in effect between drugs?
b. Do the differences between drugs levels depend on the disease (or vice versa)
The statistical model that is used to investigate this problem is the same as (2.1):
yijk = l^-^Oi^-\-|3j-\-{a|3)ij-^e^Jk i = 1,2,3,4 j = l ,2 ,3 A: = 1 6
where,
yijk - the (kth) patient reaction of (ith) drug, (jth) disease,
fi - overall mean of patient reaction,
ttj - main effect of the (ith) drug,
Pj - main eflfect of the (jth) disease,
(oi(5)ij - level of interaction between the (ith) drug and (jth) disease ,
Cijk - random error component,
with the following conditions:
E-=i »i = 0
E •=! ft = 0 E t i M ) . ; = 0 Vi
E ' = i M ) . j = o Vi
It is assumed that e^j ~ NIDlO^a"^).
As mentioned previously, we will transform the unbalanced design to a balanced
one by calculating cell means and performing hypotheses tests on them. Thus, the
transformed data set becomes the condensed cell means data set in Table 1.18.
Analysis on this data can now be performed. Let us assume that the standard
27
Table 4.19: Analysis of Variance for Drug Study based on LS. Dependent Variable: Y
Source Model Error Corrected Total
Source drug disease drug*disease
DF 11 60 71
DF 3 2 6
Sum of Squares 4445.353612 1249.488580 564.84219
Type I SS 3315.994238 448.878110 680.481264
Mean Square 404.123056 20.824810
Mean Square 1105.331413 224.439055 113.413544
F Value 19.41
F Value 53.08 10.78 5.45
Pr > F <.00Ol
Pr > F <.00Ol 0.0001 0.0002
Table 4.20: Analysis of Variance for Drug Study on testing interaction, based on ART. Dependent Variable: R'^
Source Model Error Corrected Total
Source drug disease drug* disease
DF 11 60 71
DF 3 2 6
Sum of Squares 9527.66667 21570.33333 31098.00000
Type I SS 215.777778 43.750000 9268.138889
Mean Square 866.15152 359.50556
Mean Square 71.925926 21.875000 1544.689815
F Value 2.41
F Value 0.20 0.06 4.30
Pr > F 0.0149
Pr > F 0.8959 0.9410 0.0011
deviation is set to equal 5. Again, we compare the classical least squares and aligned
rank transform techniques with respect to this problem. Inferences may be made
based on the results obtained from the analysis. The corresponding ANO\A tables
are presented in Tables 4.19 - 4.22.
The Classical Least Squares and Aligned Rank Transform multiple comparison
techniques for drug and disease are given in Tables 4.23 - 4.30.
28
Table 4.21: Analysis of Variance for Drug Study on testing the main Drug, based on ART. Dependent Variable: R^
Sum of
Source Model Error Corrected Total
Source drug disease drug*disease
DF 11 60 71
DF 3 2 6
Squares 22452.00000 8646.00000 31098.00000
Type I SS 22431.22222 9.75000 11.02778
Mean Square 2041.09091 144.10000
Mean Square 7477.07407 4.87500 1.83796
F Value 14.16
F Value 51.89 0.03 0.01
Pr > F <.0001
Pr > F <.0001 0.9668 1.0000
Table 4.22: Analysis of Variance for Drug Study on testing the main effect of Disease, based on .ART. Dependent Variable: i?^
Sum of Source DF Squares Mean Square F Value Pr> F Model 11 7961.00000 723.72727 1.88 0.0608 Error 60 23137.00000 385.61667 Corrected Total 71 31098.00000
Source drug disease drug*disease
DF 3 2 6
Type I SS 20.111111 7896.083333 44.805556
Mean Square 6.703704 3948.041667 7.467593
F Value 0.02 10.24 0.02
Pr > F 0.9968 0.0001 1.0000
Table 4.31 and Table 4.32 contain results from the the analysis conducted for
the study on outcomes of alternative drugs on different diseases. Let us define the
following:
(dg^.dgj) denotes drug i as significantly diflferent from drug j ,
[diSi.diSj) denotes disease i as significantly different from disease j .
Again, we make conclusions about the analysis after performing residual analysis.
Explanation for checking the model's assumptions was given in section 4.2.
Residual and normal probability plots for the outcomes of alternative drugs on
different diseases experiment are presented as Figures 4.5. 4.6, 4.7 and 4.8.
29
Table 4.23: Least Squares Means for eflfect Drug Pr > \t\ for HO: LSMean(i)=LSMean(j) Dependent Variable: Y based on LS, Adjustment for Multiple Comparisons: Tukey
i/j 1 2 3 4
drug 1 2 3 4
1
0.8920 <.0001 <.0001
Y LSMEAN 22.8606358 21.7777742 6.8677643 11.4566408
2 0.8920
<.0001 <.0001
LSMEAN Number
1 2 3 4
3 <.0001 <.0001
0.0191
4 <.0001 <.0001 0.0191
Table 4.24: Least Squares Means for eflfect Drug Pr > \t\ for HO: LSMean(i)=LSMean(j) Dependent Variable: R^ based on ART, Adjustment for Multiple Comparisons: Tukey
Table 4.25: Least Squares Means for effect Drug Pr > 11 for HO: LSMean(i)=LSMean(j) Dependent Variable: Y based on LS, .Adjustment for Multiple Comparisons: Scheflfe
i/J 1 2 3 4
drug 1 2 3 4
1
0.9170 <.0001 <.0001
Y LSMEAN 22.8606358 21.7777742 6.8677643 11.4566408
2 0.9170
<.0001 <.0001
LSMEAN Number
1 2 3 4
3 <.0001 <.0001
0.0360
4 <.0001 <.0001 0.0360
Table 4.26: Least Squares Means for effect Drug Pr > \t\ for HO: LSMean(i)=LSMean(j) Dependent Variable: R^ based on ART, Adjustment for Multiple Comparisons: Scheflfe
Table 4.27: Least Squares Means for eflfect Disease Pr > 11 for HO: LSMean(i)=LSMean(j) Dependent Variable: Y based on LS, .Adjustment for Multiple Comparisons: Tukey
disease Y LSMEAN LSMEAN Number
1 2 3
i/J 1 2 3
Table 4.28:
18.9894027 15.3147257 12.9179829
1
0.0191 <.0001
CO
to
I—'
2 0.0191
0.1720
3 <.0001 0.1720
Least Squares Means for effect Disease Pr > \t\ for HO: LSMean(i)=LSMean(j) Dependent \'ariable: R^ based on ART, Adjustment for Multiple Comparisons: Tukey
disease R^ LSMEAN LSMEAN Number
i/j
1 2 3
1 2 3
50.2083333 34.5000000 24.7916667
1
0.0200 <.0001
1 2 3
2 0.0200
0.2089
3 <.0001 0.2089
32
Table 4.29: Least Squares Means for eflfect Disease Pr > \t\ for HO: LSMean(i)-LSMean(j) Dependent Variable: Y based on LS, Adjustment for Multiple Comparisons: Scheffe
Table 4.30: Least Squares Means for eflfect Disease Pr > \t\ for HO: LSMean(i)=LSMean(j) Dependent Variable: R^ based on ART, Adjustment for Multiple Comparisons: Scheflfe ^___^_
Table 5.10: Power simulation for all pairwise comparisons of Design 5 based on both the least squares and aligned rank transform methods
tech
LS
ART
distn NORM
EXP LNORM
GAM CAU
NORM EXP
LNORM GAM CAU
COM 1.000 1.000 1.000 1.000
0.32600 1.000 1.000 1.000 1.000
0.8050
MIN 1.000 1.000 1.000 1.000
0.57200 1.000 1.000 1.000 1.000
0.9910
PROP 1.000 1.000 1.000 1.000
0.43725 1.000 1.000 1.000 1.000
0.9355
FWE 0.032 0.022 0.033 0.036
0.00800 0.059 0.005 0.047 0.002 0.5470
DIR 0.032 0.022 0.033 0.036
0.00800 0.059 0.005 0.047 0.002 0.5610
Table 5.11: Power simulation for all pairwise comparisons of Table 4.1, Data for Industrial Waste Output based on both the least squares and aligned rank transform methods
tech
LS
ART
distn NORM
EXP LNORM
GAM CAU
NORM EXP
LNORM GAM CAU
COM 0.08600 0.08100 0.02300 0.03400 0.00000 0.118 0.181
The applications provided in Chapter IV demonstrated that the ART technique
performs just as well as the least squares technique in testing hypotheses of interest
specifically for factorial designs, in our case. Both methods provided similar statistical
inferences based on the Industrial Waste Output and Drug Study data sets. The
results from our simulation study have shown that in most cases considered, the
aligned rank transform technique is equally as powerful or more powerful than the
classical least squares technique for multiple comparisons.
An experimenter may substitute the usual Icctst squares method by the aligned
rank transform technique when performing data analysis. We have demonstrated
that it fares just as well or better than the least squares procedure in data analysis.
From the simulation power study we observed that for a heavy tailed distribution,
such as Cauchy, the Type I family-wise error rate was higher than expected under the
aligned rank transform. Thus, future work should consider the evolution of procedures
that improve the empirical Type I family-wise error rate for heavy tailed distributions.
51
6
7
8
9
BIBLIOGRAPHY
Conover, W.J., (1977). Practical Nonparametric Statistics. Wiley . - Sons. Inc. New York.
Dey, A., Mukerjee, R., (1999). Fractional Factorial Plans. Wiley k Sons. Inc. New York.
Duran, B.S., 12/00, Thesis suggestions, Mathematics and Statistics Department of Texas Tech University, Mathematics and Statistics.
John, J.S., Quenoville, M.H., (1977). Experiments: Design & Analysis. Griffin c ' Company Limited, London.
Kendall M.G., Buckland W.R., (1971). A Dictionary of Statistical Terms. Hafner Publishing Company, Inc., New York.
Kurtz, A.K., Edgerton, H.A., (1939). Statistical Dictionary. Wiley k Sons, Inc., New York
Mansouri, H., (1998). Aligned rank transform tests in linear models. Elsevier, New York.
Mansouri, H., (1998). Multifactor analysis of variance based on the aligned rank transform technique. Elsevier, New York.
Montgomery, D., (1997). Design and Analysis of Experiments. Third edition. Wiley & Sons, Inc., New York.
10] Myers J.L., Well A.D., (1991). Research Design & Statistical Analysis. Harper Collins Publishers Inc., New York.
11] Neter J., Kutner M.H., Nachtsheim C.J., Wasserman, W., (1996). Applied Linear Regression Model. Third edition. McGraw-Hill Companies, Inc., New York.
12] Westfall, P.H., Tobias, R.D., Rom, D., Wolfinger, R.D., and Hochberg, A'., (1999). Multiple comparison and Multiple Tests, using the SAS System. S.AS Institute Inc., Gary, NC.
52
[13] Westfall, P.H., 02/01, %SimPower macro discussion, Business School Department of Texas Tech University, Business school.
53
PERMISSION TO COPY
In presentmg this thesis m partial fulfillment of the requuements for a master's
degree at Texas Tech University or Texas Tech University Health Sciences Center, I
agree that the Library and my major department shall make it freely avaUable for
research purposes. Permission to copy this thesis for scholarly purposes may be
granted by the Director of the Library or my major professor. It is understood that
any copying or publication of this thesis for financial gain shall not be allowed
without my further written permission and that any user may be liable for copyright