r

AD-AO14 715
TREATMENT CONTRASTS IN PAIRED COMPARISONS. I.
BASIC PROCEDURES WITH APPLICATION TO FACTORIALS
A. Bradley,SRalph
et al
Florida State University

Prepared for:

(Office of Naval Research

Department of the Army
Department of the Air Force

July 1975

DISTRIBUTED BY:

National Technical Information Service

U. S. DEPARTMENT OF COMMERCE
 266all

-~ Th .'lrida.IE
I. bAI
CNAStSW
R~lUE
In iv~feD
Arsity-S
7thAPiTINOFCORL

ON
ecnclReort N 9

Ta11a a-4. B g,,

ka 1
 TREATMINT CONTRASTS IN PAIRED COMIARISONS
I. BASIC PROCEDURES WIITH APPLICATION TO FACTORIALS 1

BY
Ralph A. Bradley and Abdalla T. El-I1elbawy

FSU Technical Report lo. H1348

ONR Technical Report No. 99

July, 1975
The Florida State University
Department of Statistics
Tallahassee, Florida 32306 '

1
Research supported at the Florida State University by the Army, Navy
and Air Force through ONR Contract N00014-67-A-0235-0006. Reproduction in
whole or in part is permitted for any purpose of the United States Government.

F IA~ od for p.. bLic r*•...

Dthtlibutiou Unjr i •if
A
 TREATZW1NT CONTRASTS IN PAIRED C(I1PARISONS
I. BASIC PROCEDURES WITH APPLICATION TO FACTORIALS 1

Ralph A. Bradley Abdalla T. I1-Ilelbawy

Florida State University University of Cairo
'Tallaliassoe, Florida, U.S.A. Cairo, United Republic of Egypt

A simple procedure for consideration of specified treatment contrasts or

* sets of contrasts in a paired comparisons experiment is developed. General

likelihood estimation and likelihood ratio tests are given. Specified treatment

comparisons as appropriate in a particular experiment may be made. The procedure

may be used for consideration of factor effects and interactions when the treat-

ments in paired conparisons are factorial treatment combinations. An example

is given of a taste preference experiment on coffee with factors, brew strength,

roast color and brand, each at two levels. Results are summarized in an

analysis of chi-square table very analogous to the typical analysis of variance

Sndmar oC.
Research supported at the Florida State University by the Army, Navy
nNR through
and Air Force Contract 1100014-57-A-0235-0006. Reproduction in
wholc or in part is p~ermitted for any purpose of the United State6 Covernment. .
i
1. INTRODUCTIOII

Bradley and Terry (1952) presented a model and a method of analysis for

paired comparisons generalized slightly, by Dykstra (1960). In the basic experi-

ment, t treatments, T1 , ... , Tt, are compared with nij comparisons of Ti

and Tj, ni 2 00 i < j, ijj = 1, ... , t. Some of the comparison sizes n

may be zero but linkage of comparisons is required in the sense that there must

not be any subset of the treatments for which no treatment is compared with any

treatment of the complementary subset. The model postulates the existence of

treatment parameters, -f. for T., . >: 0,

'1 1 1

2:"'=[ "i 1 ,(1,1)

such that the probability of selection of Ti from the pair (Ti,T ) j 0 1,

is

PCTi > T) u ii/(7 i + 1T. • (1.2)

Likelihood methods were used. On the assumption of independence of selec-

tions, the likelihood function is

a. n..
LC!) = It (,.T"/ + w) ,(1.3s)

I3<j i
i

where ai is the total number of selections of Ti in the entire experiment,

i • It ... , t, I a i nij, and -' a (I1' "' ." hen L(Q) is maximized

subject to (1.1) as a constraint and if pi is the likelihood estimator of

i (and p of !), the likelihood equations are

pi ni

0 i1 , (1.5)
i
 -2-

where •' represents a sum with j • i. Bradley (1975) has summarized the
j
various bases for the nodel and results associated with it, giving an extensive

bibliography.
In applications, situations arise in which special comparisons or contrasts

among treatments are of interest. Abclson and Bradley (1954) considered the
2 x 2 factorial and El-llelbawy (1974) the 2i factorial. This paper considers
orthogonal treatment contrasts more generally and the results may be applied to

any set of factorial treatment combinations. The factorial models attempted

earlier have been modified and sinplified although numzerical results in applica-

tions are closely sinilar.

David (1963, Section 4.1) noted that the basic vie'hod is unchanged if (1.1)

is replaced by other scale-determining constraints and suggested

i -e 7r.- 0 .(1.6)

Bradley (19S3), in comparing the method with that of Thurstone (1927), suggested
that Cn i. plays the role of a location parameter for Ti. Treatment contrasts
will be specified as linear contrasts among

Yi = Zn ri1 i = is ... , t .(1.7)

2. LIKELIHOOD ESTIMATION

Let B-m be an m x t matrix, 0 < m < (t - 1), with zero-sum, orthonormal

rows. Let y(: and _(p) be the column vectors with i-th elements Yi in
(1.7) and tn pi, i 1, ... , t; let
1 r be an m-element column vector with

elements r ,1"
k = 1, . m, to be used as Lagrange nultipliers; and let
m., 0
and 1i be respectively column vectors of m zero elements and t unit elements.

The general estimation problem is to find estimators pi of n. i u 1, ... ,

or p of r, maximizing L(i_) in (I.3) subject to the constraints,

it) - 2,1÷4l (2.1)

incorporating (1.6). Thus, using the method of Lagrange multipliers, we maximize

Q
Q!) UL L(! ,+ .'(Z [• ] (2.2)

subject to (2.1). Throughout this work we assume that p > 0, i , ... ,t;

Ford (1957), in show-ing convergence of an iterative solution to (1.4) and (1.5)

assured this by the following Ford Condition: In every possible partition of

the t treatments into two non-empty subsets, some treatment in the second

subset has been preferred at least once to some treatment in the first subset.

Let

-i ""ipi = ,. t (2.3)
(
3 Pi * p.'j =~..

and lot 1(p) be the t-clcemnt -;lumn vector with elements in (2.3). The

equations necessary for maximization of Q(Z) reduce to

YO B'r a 20t (2.4)4

and

Bny-. 0 (2.5)
 -4-

together with the scale-determining equation,

.en 1), a 0 (2.6)

1

or (1.5) as desired; it i5 easy to show that r0 0. M1ultiplication on the

left of (2.4) by B shows that r_ = -Bf•(p_) and that (2.4) reduces to

(It - BIB)E(p) = O , (2.7)

-1-0
-5t It

wihere It is the t x t identity matrix. Let

_., - (2.8)

with typical eler.mcnt Dij. Then (2.7) may be rewritten

a._
1 - i() = 0 i = 1 t
, ... , , (2.9)
Pi

where

n . U- -U
D.. (2.10)
j Pi Pj ID

Note that D.. > 0, m = 1, ... , t - 2. '.hon m (t - 1), a deterministic

solution follows from (2.1) alone - each yi u0, ni 1 or, with scale adjust-

nont, each 1 * it./ !'Ihen m = 0, B is non-existent and (2.4) oT (2.7)

reduces to (1.4) or we may use (2.9) with D.. = 0, i j j, D.. - 1,1

i,j 1
1, ... , t, in (2.10). An iterative solution to (2.9), (2.5), and (2.6)
is considered now.
 e ~-5-1

3. ITERATIVE SOLUTION OF LIKELIHOOD EQUATIONS

Let p(0) and (r) satisfying (2.5) and (2.6) be initial and r-th

approximations to •, the solution of (2.9), (2.5), and (2.6). That

exists is clear because p(o) a I meets the requirements. The means of

obtaining p(r*l) from (r) are describod.

Think of r = t(J - 1) * j - 1, 3 " 1,2,3, ... , j o Iv ... , t; J indi-

cates a cycle of iterations and j. the step in the J-th cycle. To initiate the

calculation of p(r~l) from P(r), calculate ýj(r from (2.10): Thb

iteration is done by calculation of

(r) )k, r) .
i (k) " exp(L P' *Di (D] - , = i1 ... t , (3.1)

for successive values of k, k - 0,1,2, ... , where

fi[a /P (r) *-((r) ) if 4j ( (3.2)

jI otherwise.

When a/ip(r) * * (r)) p l * (r); if this occurs for t successive

values of r, p r) a/c )he our))( and

Lp(r+l)(k)]for k= 0,1,2, ... until a value k* of k is found for which

1 )k*)]
L[E(r+C > L(C)M). Then
prl p(r+l) (kt).

El-Helbawy Lnd Bradley (1975) show that k* exists and examine the convergence

of this iterative process. The procedure is easy to program on a computer.

Note that pCr+l) satisfies conditions (2.5) and (2.6) if a(r) does.
4 through use of (3.1),
"this is demons i rated easily for p(r )(k) for all k
 -6-

the orthonorriality of '-he rows of Bn, , and thc fact that Bm has zero-sum rows.
Nlote also that if n -O0 Di = 0, i 0j, D.i M1, i'j= 1, ... , t' k= 0, and

then 0 jr) > 0 and

~r~l) a)c (r+l) (3r) *

where

[~~yA~y~ -/t
1

(r c (r ()
is such that Rfl rl 1. Then the iterative procedure is equivalent to the
iI
one suggested by Bradley and Terry (1952) and shown to converge by Ford (1957).

4. INFERENCE XMD FACTORIALS

Likelihood estimlation and likelihood ratio tests were proposed by Bradley

and Terry in cons ideration-of the basic miodel for paired comparisons used here.

In extending those results, w:o can formulate a single test procedure that may
be used to test particular treatment contrasts or for the analysis of factorial

treatment combinations.

Consider two matrices B-m and B...+n i where 05 m < mn+ n :5 (t - 1),

BMdefined as before and BM ~ having zero-sun~, orthouiormal row.%AlWe assume that

BmyQ) 0ý and test the hypothesis,

110:Bn
IT (4.1)
 -7-

against the alternative,

1-1a: PejQ) ~0 n (4.2)

Let wt maximize L() given B_,.ny(r) a 0 and (1.6) and let r p

maximize L(D) given L(y)=)
- 0 and (1.6); both p0 and • may be

obtained through use of the likelihood estimation procedure given in Sections

2 and 3. The likulihood ratio statistic for testing 110 versus Ha i.ven

S- 26 is

-2 tn A - -2[tki L(pO) - tn L(p)] (4.3)

which, for large values of the nij, has the central chi-square distribution
with n degrees of freedom as its limiting distribution under H When
0'
m w 0 and n = (t - 1), the original chi-square test with (t - 1) degrees of

freedom for treatment equality is obtained. The test based on (4.3) may be used
to partition that chi-squaro with (t - 1) degrees of freedom into several
independent and additive chi-squares for properly chosen• and sequenced ortho-

normal sets of contrast hypothesesand designated assuned null contrasts.

An important and natural application of the technique developed is to

experirments wherein the treatments constitute a full or fractional set of

factorial treatment combinations. Tests of factorial effects, main effects

and interactions, may be made directly based on (4.3) and suitably chosen rows

for DB and B-+n" Alternatively, the paired comparisons model may be reparam-
eterized to introduce factorial parameters.
Consider a q-factor mixed factorial for which factor s has bs levels,

bs Z 2, s = l, ... , q and let a = (at ... ' a q) be a vector designating the

 -0Q-

levels of the q factors for a particular treatmlent combination. Now replace

q
Ti by T and ir.by r C; if a full factorial is considered, t = R bS
a -- s=l
Factorial parameters are introduced multiplicatively; take

ITli2 7 .T r7 1 .. q (4.4)
1 aiI <i 2 ai1 i2 1 <...<i' ai I OL a, .. q

for all a whore 1 :5i I < . . < i r ý5q, r = 1 ... p q, and r,

represents the -r-factor interaction (main effect if r = 1) of factors

i1, ... ,Vir at levels ail ""..'Ccr The new factorial parameters are
q riq q
H (I + bs -1 in number and must be subject to [ H (1 +b) - IN bs]
S=1 s=l S=l
functionally independent constraints. But (4.4) is linear in the logarithms of
•arameters. The system is exactly analogous to that of the analysis of variance

(anova) of, say, a randomized complete block design with the same factorial

treatment combinations. The anova constraints on anova factorial parameters

il...i il'''1
apply to the t i r If r is the likelihood estimator of
11 1 r

1 '"r 1l' ir is the same linear function of the

an . OL , '• -(A .en pIaa as is
i I "r I Ir

the anova estimator of the corresponding anova interaction parameter of the

anova treatment efEcets estimators. Inotm alporithms for the analysis of

factorial experiments may be applied in the analysis of factorial treatment

combinations in paired comparisons. NA hypothesis of no r-factor interaction

among factors il, ... , ir is formulated by the specification that the usual
r il *.oi r
TI (bi - 1) interaction contrasts anong the Za 7T arc zero, a
s=l s r1

a
 -9--

i ir
specification equivalent to the requirement that each n a. 0..1. = 0 in
1 1r
1
view of the parameter constraints or that Tr .. r - 1 for a *is 1, ... b is

S = 1, ... , TTr 1

It has been seen that likelihood estimators i of parameters wT under

constraints (2.1) may be found for various choices of B . In each such situation,

nultivariate normal distributions may be specified for the large-sample distri-

butions of 47 Q - L or of VIT - .y.()] with means zero and determined

variance-covariance matrices where n NX 0. ij as N - nij

. . Large-
sample variances and covariances for specified treatment contrasts may be found

and power functions for local alternatives for tests based on (4.3) may be speci-

fied. While these results are needed in some aspects of data analysis for

paired comparisons, they will be given by the authors in a subsequent paper.

S. EXAMIPLE - A 2 FACTORIAL

Brew strength, roast color, and cofee brand were tho three factors, each

at two levels, in an example of the use of paired comparisons with factorial

treatment combinations in consumer preference testing. Twenty-six preference

judgments were obtained on each of the 28 possible treatment comparisons.. The

data are summarized in Table 1. Mhe three-element vectors a and B designate

the treatmest combinations.

 -10-

1
TABLE 1 - Preference Data in Coffee Testing

Preferred Treatment Not Preferred Total

Treatment T Ta Preferences

2: 000 001 010 011 100 101 110 111 a_.

000 0 15 15 16 19 14 19 16 114

001 11 0 10 15 15 14' 15 12 92

010 11 16 0 15 15 14 10 15 104

011 10 11 11 0 14 11 15 13 85

100 7 11 11 12 0 9 14 13 77

101 12 12 12 15 17 0 16 18 102

110 7 11 8 11 12 10 0 12 71

111 10 14 11 13 13 8 14 0 83

1 Data provided through the courtesy of Ilavis B. Carroll and John C. leimlich
and of the General Foods Corporation.

Consider the matrix,

1 1 1 1 -1 -1 -1 -1
1 1 -1 -1 1 1 -1 -1
1 -1 1 -1 1 -1 1 -1

R71 1 -1 -1 -1 -1 1 1 (5.1)
1 -1 1 -1 -1 1 -1 1
1 -1 -1 1 1 -1 -1 1]
1 -1 -1 1 -1 1 1 -
 -11-

The overall test of treatment preference equality, equivalent to that of Bradley

and Terry (19S2), is obtained from (4.3) writh B in (4.1) equal' to 1 7 in
-n
(5.1) and m - 0. Under the null hypothesis (4.1), each ir 1 and L(!) in
(4.3) is L(la) - the scale constraint (1.6) is used instatd of (1.1). In the
analysis summary of Table 2, -2 tn L(l ) is given in the first rowl. Under
the alternative hypothesis (4.2), m - 0' and does not exist; L(Q) is
maximized subject only to (1.6) with the resultant values of p8 and -2 tA L(ka)
given in row 2 of Table 2. The chi-square statistic of (4.3) is, by subtraction,

29.58 with 7 degrees of froedom. 7his value is shoim in the totals rotqs of

analysis of chi-square Tables 3 and 4.

The total chi-square may be partitioned in various ways, two of which are

showm in Tables 3 and 4. ,'!e give details on Table 3 and leave the reader to

check Table 4. Hypothesize no three-factor interaction. Then _ -(7), the

last row in (5.1), m n 0, B does not exist and B Bl(7).
B Values of
p and -2 tn L(p) under 110 are given in row 3 of Table 2. The chi-square
statistic of (4.3), now with 1 degree of freedom, is the difference of entries
-2 tn L(p) in rows 2 and 3 of Table 2, the value being 0.63. Now assume no
three-factor interaction and hypothesize no two-factor interactions; B _

B - 1(4,5,6), B•+n (4,5,6,7), the arguments of B indicating the rows

chosen from (5.1). The required chi-squared statistic from (4.3) has 3 degrees

of freedom and the value 15.34 from rows 3 and 4 of Table 2. This chi-square
may be partitioned in a number of ways mid we choose only one of them. Chi-
square statistics, each with 1 degree of freedom, are calculated for the following

hypotheses and assumptions: (i) No brand, roast color interaction; no three-

lactor interaction assumed, (ii) No brand, brew strength irteraction, no three-

factor interaction and no brand, roast color interaction assumed, (iii) no roast
 In InCto 0 n vo

Oh) 0 In n, In an C, LN mn V) C3 In @
40D,m as
O as0m 0) 0) 0r Ci- ~0
Mh 4h0
-4- -4

0 CO C% 9.4 m .4 qt In on 0 V-4 C% if'

- C I0
* tl %0 0- In 1N %0 0% a P-4 %D
0 07 to 90i'- v* N i- CO 0i 0-4 04 Oh -4 4m
+. 0. 0 , 0
-4C5 04 C5 0 0 - .-4 C, 4

Q coO -4 1-4 In In V n -q C'M ; W) t

%0 t- %0
an t- co t- caG a3
a-4 Q- V-4 n i m W D In an n t

C- C4 P4 4: l m o v C4 1 -4

C) cIa an an 14n m0 '0 an q C

~~ "-4 0)
In
anaC)
01
0
'ci
~
ICn
-
%Q
£
Go 01
) V
CA
0
a0
I-
co
a
V
P4 CA
0) DoI
I cl C) Co 03 to r m l Cm
p 1114 0 4 aI,4 *I .-

-4 0 r) I/ Ch n an ý M4 wt
0a .0 'Si 0' -e
0- ý C30 o CD CD V-4 COý Ca to C) Go 0)
$4 . 4 0 0 0- Cl P4 : 0 0; 0; 0 0
oa '-4 an
"I "'
PI '0 %0 vh 'C ", 4-n P

0
CII CO
n an %D - w V an an an Ch
a .4 . 4 .) -4a r. %1 wt fn (4
- 0 0 co %0 CD'I 0 "0 %0i m to

P C
0p l Cl
ClI i
-InC
~ n- aIn
-
Q0.0
CD %t
V-4
a-
to
I,,
w-I
"C
In
0%
@
"4
N
%D
h.

-m CD V4 In 1n aa C o %0 t- l -
Ch 0 Oh
"C co
o 0D N an
4J
N 0 0N 4
914 a-, -. N- N- (A -4 I

U) on q* Q c 4 v, V e t oI

4 r- 'a LI Olt t-~ w0 anr an

C5 " n an
an 2 In In In Nn
C ~ -4
C14 le ( C an c an i

a'- a-i 4- 0
1- 5L-
NO - 0N
0 n In an anD an N n - n N
WN 0

$4e * . wtI
S a* C11a *
4P)
gel
 -13-

color, brew strength interaction; no three-factor interaction, no brand, roast

color interaction, and no brand, brew strength interaction assumed. The matrices

D , B and 13 respectively are (i) E1(7), _B(6). L(6,7), (ii) 12(6,7)8

_ n -~~ 1 6) 3(,)5..(,)
Bjj(S), C(5,6,7), (iii) 1(5,6,7), B1 (4), R4(4,5,6,7). Pertinent row pairs

in Table 2 are (i) 3,5, (ii) 5,6, (iii) 6,4 and the chi-square values are

0.22, 14.96, 0.15. T1hree remaining degrees of freedom arc for tests on main

effects. The following hypotheses and assumptions arc ised: (iv) No brand

effect, no interactions assumed; (v) No roast color effect; no brand effect

and no interactions assumed; (vi) No brew strength effect; no roast color

effect, no brand effect and no interactions assumed. The matrices B , B and

-M n
B n+n, the row pairs of Table 2 and the corresponding values of chi-square,

each with I degree of freedom, are as follows: (iv) R4(4,5,6,7), BI(3),

.. (3,4,5,6,7), 4, 7, 0.04; (v) !6(3,4,5,6,7), B (2), B (2,3,4,5,6,7), 7, 8,

4.29; (vi) 4 (2,3,4,5,6,7), B ( 1). &7, 8, 1, 9.28. T1ise results are given in

Table 3. The alternative analysis of Table 4 is based on rows 1, 2, 9-14 of

Table 2.

The analyses of Tables 3 and 4 are remar!ably similar. i•hile our purpose

has been to illustrate a technique, it is clear from both analyses that brew

strength and roast color have main effects while brand interacts with brew

strength even though it has no main effect.

Factorial parameters may be estimated. Ue illustrate using the full factorial

model and p from row 2 of Table 2. Because of the constraints on the factorial

parameters and the one degree of froedomn available for the estimation of each

main effect or interaction, each main effect or interaction may be obtained from

the vector I Thus

 -14-

TABLE 3 - An Analysis of Chi,,Square for the Coffee Data

Hypothesis Conditioas Degrees of Chi

Tested1 Assumed Freedom Square

No F1 effect No F2 . F 3 , N•o interactions 1 9.28

No F2 effect ITo F3 , No interactions 1 4.29

No F3 effect iHo interactions 1 0.04

No F1 F 2 , F 1 F 3 , F 2 F 3 Io FI 1F2 F 3 interaction 3 15.34

interactions

14o F2 F3 interaction iTo FIF 2 F3 interaction 1 0.22

No F 1 F 3 interaction 14o F 2 F5, F 1 F 2 F3 interactions 1 14.96

No F 1 F 2 interaction F1 F3 , F2 F 3 , F1 F 2 F3
Fo 1 0.15
interactions

1 F 2 F3 interaction
F1o None 0.63

No treatment effects Hone 7 29.58

F is brew strength, F2 is roast color, F is brand.

12
 -15-

TABLE 4 - An Alternative Analysis of Chi Square for the Coffee Data

Hypothesis Conditions Degrees of Chi

Tested1 Assumed Freedom Square

14o F 1 effect None 1 9.47

No F 2 effect Ro F1 effect 1 4.33

No F3 effect No F1 , F2 effects 1 0.04

No F1 F 2 , FlF3 , F2 F3 No main effects 3 15.12

interactions

14o F1F2 interaction Ilo main effects 1 0.16

W4o F 1 F3 interaction No main effects, No F1 F 2 14.73

interaction

14o F •3 interaction Hie main effects, lie F 1 F 2 , 1 0.24

FI1F 3 intera~ctions

N4o F I F2 F3 interaction Ilo main effects, Ilo two- 10.62

factor interactions

Ho treatment effects Nne 7 29.58

FI is brme strength, rF2 is roast color, F3 is brand.

 U--16-

1 3 12
10.192
L p- 0 .134 tn p0 0.1038 Zn p0 *0.0096 &i p 0.0196

l ~ 15 .1243 2312
A pO 0.1924 & p -0.0238 en 123 0.0393
p0 0 000 ,

and
1 2 3 12 0.86;
p0 " 1.1658 p0 1.1094 p0 1.0097 p
p0 0 0.9806

13 a 1.2122- 23 0.9765 23 - 1.0401

p00 0P
0 pO0 0

These estimates are listed in the same order as the pertinent roWs in B7" The

remaining factorial parameter estimates are identical to or the reciprocal of

those shown for the same effect or interaction depefiding on wheihei the sum o•f
_ , 1=
the subscripts is even or odd. For example, p1 (1.1658)'1 a 0.8578,
p..I
-2
11 .......
0.9806, P001 123{.'0401)"
. . . . ..
-0.9614.";.....
. .. . . . . .
... . . . . . . . . . ......
"" l

It is seen that the analysis for the factorial treatment combinations in

paired comparisons is very similar to the analysis of variance for fact!oFials..*

Interpretatio•is are made in the usual way, particularly when the logarithms of

factorial parameter estimates are considered.

7. CONCLUDING REMARKCS

The techniques of this paper provide ioans for much new flexibility in the

analysis of paired comparisons experiments. Factorial treatnent combinations

may be used. Since specified treatment contrasts may be used generally,

fractional factorials are available also. Special treatment contrasts may have

interest in special experimental situations.

• I I • I , I , I , I mI I I IIM
 -17-

It has been noted that convergence of the iteration process given for

solution of likelihood equations, asymptotic va'riances and covariances for

treatment contrasts, and large-sample properties of test procedures will be

given in subsequent papers. The similarity ol analyses in Tables 3 and 4

suggests that conditions assumed for a test may not be constraining and that

tests in the two tables may be asymptotically equivalent. This possibility

will be investigated.

POFEREfNCES

Abelson, R. H. and Bradley, R. A. (1954). A2 2 factorial with paired compari-

sons. Biometrics 10, 487-502.

Bradley, R. A. (1953). Some statistical methods in taste testing and quality

evaluatio7L. Biometrics 9, 22-38.

Bradley, R. A. (1975). Science, statistics and paired comparisons. Biometrics

(submitted).

Bradley, R. A. and Terry, I..13.(1952). The rank analysis of incomplete block

designs. I. The method of paired comparisons. Miometriha 39, 324-45.

David, 11.A. (1963). The Mlethod of Paired Comparisons. London, Griffin.

Dykstra, 0. (1960). Ran], analysis of? incomtilote block designs: A lmethod of

paired comparisons employing unequal repetitions on nairs. Biometrics 16,

176-88.

11-11elbawy, A. T. (1974). Factorial Treatment Combinations in Paired Comparisons.

Tallahassee: Florida State University Doctoral Dissertation.

El-lielbawy, A. T. and Bradley, R. A. (1975). Treatment contrasts in paired

comparisons. II. Convergence of a basic iterative scheme for estimation.

(In preparation).

Ford, L. R., Jr. (1057). Solution of a ranking problem fror binary comparisons.

r. M.a1th. rMonth1l 64, 28-33.

Thurstone, L. L. (1927). Psychophysical analysis. Amor. Jour. Psychol, 38, 368-89.

 UNCLASSIFIED
SECURITY CLASSIFICATION OF 711IS PAGE

REPORT DOCU,.MNTATION PAGE

1. REPORT NI=ER 2. GOVT. ACCESSION NO. 3. RECIPIENT'S CATALOG NUIMER
ONR Report N~o. 99

4. TITLE (and subtitle) S. TYPE OF REPORT 6 PERIOD COVERED

Treatment Contrasts in Paired Comparisons Technical Report
I. Basic Procedures with Application 6. PERFORMING ORG. REPORT NUlER
to Factorials
FSU Statistics Report M348
7. AUTHOR(s) 8. CONTRACT OR GRANT NUIMBER(s)
Ralph A. Bradley ONR i100014-67-A-0235-0006
Abdalla T. El--elbaily
9. PERPOPIIING ORGANIZATI(1 MAEIM ADDRESS 10. PROGRA!' ELEIMET, PROJECT, TASK
The Florida Stazte University AREA UNIT NUMMS
Department of Statistics
Tallahassee, Florida 32306
11. CONTROLLING OFFICE RATE & ADDRESS 12. REPORT DATE
Statistics and Probability Program July, 1975
Office of INIaval Research
Arlington, Virginia 22217 13. NU•ER OF PAGES

14. NONITOPING AGENCY NAT F ADDRESS Cif is. SECURITY CLASSIFICATION (of this
different from Controlling Office) Report)
Unclassified
15a. DECLASSI FICATION/DOrIGRADING
SCI IDULE

16. DISTRIBUTION STATFMENT (of this Report)

Approved for public release; distribution unlimited.
T7. DISTRIBUTION STATEMENT "(o-f the abstract entered in Blockc 20, if different from
Report)

19. 11EY W'ORDS

Paired Comparisons
Treatment Contrasts
Nonparairietric Statistics
Factorials
20. AUSTRACT
A simple procedure for consideration of specified treatment ccntrasts or setsof
contrasts in a paired comparisons experiment is developed. General likelihood esti-
mation and likelihood ratio tests are [iven. Specified treatment comparisons as
appropriate in a particular experiment may be made. The procedure may be used for
consideration of tactor effects and interactions when the treatments in paired com-
parisons are factorial treatment combinations. An example is given of a taste prefer-
ence experiment on coffee with factors, brew strength, roast color and brand, each
at tto levels. Results are summarized in an analysis of chi square table very
analogous to the typical analysis of variance summary.