StatisticsResearchLetters(SRL),Volume5,2016www.srljournal.

org
doi:10.14355/srl.2016.05.003

DecompositionofSymmetryUsing
CumulativeSubAsymmetryModelfor
SquareContingencyTables
KoujiTahata1,TomokiOyamaandSadaoTomizawa2
DepartmentofInformationSciences,FacultyofScienceandTechnology,TokyoUniversityofScience,NodaCity,
Chiba,2788510,Japan
kouji_tahata@is.noda.tus.ac.jp;2tomizawa@is.noda.tus.ac.jp

Abstract
Forsquarecontingencytableswithorderedcategories,thepresentpapergivesthetheoremthatthesymmetrymodelholdsif
and only if four models of (1) cumulative subasymmetry, (2) palindromic symmetry, (3) global symmetry, and (4) marginal
meansequality,holdsimultaneously.Itanalyzessomedatausingthisdecomposition.
Keywords
OrdinalData;PalindromicSymmetry;Separation

Introduction
Consider an r r square contingency table with ordered categories. Let pij denote the probability that an
observationwillfallinthe i throwand j thcolumnofthetable( i 1 r j 1 r ).Considerthesymmetry(S)
model,definedby

pij p ji

(i j )

(Bowker,1948;Bishop,FienbergandHolland,1975,p.282).
Let
i

Gij pst

(i j )

s 1 t j

and
i

G ji pts

(i j )

s 1 t j

TheSmodelcanbeexpressedas
Gij G ji

(i j )

McCullagh(1978)consideredthepalindromicsymmetry(PS)model,definedby
Gij

i
G
j 1 ji

(1 i j r )

where 1 1 withoutlossofgenerality.AspecialcaseofthePSmodelobtainedbysetting 1 and 1 r 1

istheSmodel.
Let X and Y denotetherowandthecolumnvariables,respectively.Considerthemarginalmeansequality(ME)
model,definedby E ( X ) E (Y ) ,where
r

E ( X ) ipij
i 1 j 1

17

www.srljournal.orgStatisticsResearchLetters(SRL),Volume5,2016

and
r

E (Y ) jpij
i 1 j 1

Also,theMEmodelcanbeexpressedas
r 1

r 1

i 1

i 1

Gii 1 Gi 1i
(see,Tahata,YamamotoandTomizawa,2012).
As a model which indicates the structure of Gij with j i 2 , namely, pij with j i 2 , Tomizawa,
MiyamotoandOuchi(2006)proposedthecumulativesubsymmetry(CSS)model,definedby
Gi i 2 Gi 2i

(i 1 2 r 2)

Tahataetal.(2012)gavethetheoremthattheSmodelholdsifandonlyifallthePS,MEandCSSmodelshold.
Tomizawa,Miyamoto,YamamotoandSugiyama(2007)consideredthecumulativetworatiosparametersymmetry
(C2RPS)model,definedby
Gij
G ji

j i

(i j )

Theglobalsymmetry(GS)modelisdefinedby

U L
where

U pij L pij
i j

j i

(Read,1977).Tahata,YamamotoandTomizawa(2013)gavethetheoremthattheSmodelholdsifandonlyifall
theC2RPS,GSandMEmodelshold.
Themarginalhomogeneity(MH)modelisdefinedby

Gi i 1 Gi 1i

(i 1 2 r 1)

(Stuart, 1955). The model indicates that the marginal distribution for the row variable is equal to that for the
column variable. Note that the model indicates the structure of Gij with j i 1 . The extended MH model is
definedby
Gi i 1 Gi 1i

(i 1 2 r 1)

(Tomizawa, 1993). This model indicates the asymmetric structure of Gij with j i 1 . Therefore, we are
interestedinconsideringtheextendedCSSmodelastheanalogoustoTomizawa(1993).Inthispaper,wepropose
amodelwhichindicatesthestructureofasymmetryfor Gij with j i 2 .Also,weshowthattheC2RPSmodel
canbeseparatedintotheproposedmodelandthePSmodel.Usingthisresult,wegivethedecompositionoftheS
model.
Decompositions
Thissectionproposesamodelbasedon Gij ,andgivesadecompositionoftheC2RPSmodel.
Consideramodeldefinedby
Gi i 2
Gi 2i

18

(i 1 2 r 2)

StatisticsResearchLetters(SRL),Volume5,2016www.srljournal.org

Weshallreferthemodelasthecumulativesubasymmetry(CSAS)model.Wenowobtainthefollowingtheorem.
Theorem1.TheC2RPSmodelholdsifandonlyifboththePSandCSASmodelshold.
Proof.IftheC2RPSmodelholds,thenthePSmodelholds.AssumethattheC2RPSmodelholds,andthenweshall
showthattheCSASmodelholds.WeseefromtheC2RPSmodelthat
Gi i 2
Gi 2i

(i 2) i 2

(i 1 2 r 2)

ThustheCSASmodelholds.
AssumethatboththePSandCSASmodelshold,andthenweshallshowthattheC2RPSmodelholds.Weseefrom
thePSmodelthat
Gi i 2
Gi 2i

i
i 1

(i 1 2 r 2)

Then,wecanset 1 1 withoutlossofgenerality.SincetheCSASmodelholds,
G13

1
1

2
2
G31

So,weobtain

Similarly,weobtain 3 () 2 because
G24

2
2

G42
3 3

Therefore,wecansee

i 1

(i 1 2 r 2)

Thus
Gij
G ji

i
2

j 1

j i

(i j )

TheC2RPSmodelholds.Theproofiscompleted.
WenowobtainthefollowingtheoremfromtheresultgivenbyTahataetal.(2013)andTheorem1.
Theorem2.TheSmodelholdsifandonlyifallthePS,CSAS,GSandMEmodelshold.
Goodness-of-Fit Test
Let nij denotetheobservedfrequencyinthe i throwand j thcolumnofthe r r tablewith n nij ,andlet

mij denotethecorrespondingexpectedfrequency.Assumethatamultinomialdistributionappliestothe r r table.

Wecantesteachmodelforgoodnessoffitbyusingthelikelihoodratiochisquaredstatistic G 2 definedby
r
r
nij
G 2 2 nij log

i 1 j 1
m ij

where m ij isthemaximumlikelihoodestimateof mij underthemodel.Table1presentsthenumbersofdegreesof

freedomfortestinggoodnessoffitofthemodels.

19

www.srljournal.orgStatisticsResearchLetters(SRL),Volume5,2016

TABLE1THEDEGREESOFFREEDOMOFMODELSFOR r r TABLE

Models

Degreesoffreedom

r (r 1) 2

C2RPS

r (r 1) 2 2

PS

(r 1)(r 2) 2

CSAS

r 3

ME

GS

An Example
Table 2, taken directly from Tominaga (1979, p.130), is the crossclassification of Japanese fathers and his sons
academic background which were examined in1955.Note that category(1)is elementaryschool;(2) junior high
school;(3)highschool;and(4)university.Table3givesthevaluesofthelikelihoodratiochisquarestatistic G 2 for
modelsappliedtothesedata.TheCSASmodelfitsthedatainTable2well.
TABLE2CROSSCLASSIFICATIONOFJAPANESEFATHERSAND
HISSONSACADEMICBACKGROUND;TAKENFROMTOMINAGA(1979,P.130)

Sonseducationallevel

Fathers
educationallevel
(1)

Total
(1)

(2)

(3)

(4)

374
(374.0000)
18
(18.0000)
4
(6.9571)
2
(1.5482)

602
(602.0000)
255
(255.0000)
23
(23.0000)
6
(3.4947)

170
(167.3926)
139
(139.0000)
42
(42.0000)
17
(17.0000)

64
(64.6918)
71
(72.9156)
55
(55.0000)
53
(53.0000)

398

886

368

243

(2)
(3)
(4)
Total

1210
483
124
78
1895

Note: EstimatedexpectedfrequenciesfortheCSASmodel.
a

272872 . Hence, under this model, the

Under the CSAS model, the maximum likelihood estimate of is
probabilitythatafathersacademicbackgroundstatusis i orbelowandhissonsacademicbackgroundstatusis
272872 times higher than the probability that the fathers academic
i 2 or above is estimated to be

backgroundstatusis i 2 oraboveandhissonsacademicbackgroundstatusis i orbelow (i 1 2) .

Moreover, from Table 3, we see that the S model fits these data poorly, yielding G 2 115123 with 6 degrees of
freedom.Also,eachofthePS,MEandGSmodelsfitthesedatapoorly,buttheCSASmodelfitsthesedatawell.By
usingthedecompositionoftheSmodel(i.e.,Theorem2),weshallconsiderthereasonwhytheSmodelfitsthese
data poorly. We can see that the poor fit of the S model is caused by the poor fit of the PS, ME and GS models
ratherthantheCSASmodel.
2

TABLE3LIKELIHOODRATIOCHISQUAREVALUES G FORMODELSAPPLIEDTOTHEDATAINTABLE2

Models

Degreesoffreedom

G2

1151.23*

C2RPS

31.20*

PS

31.13*

CSAS

3.18

ME

700.17*

GS

1093.22*

Note:*Significantatthe0.05level.

20

StatisticsResearchLetters(SRL),Volume5,2016www.srljournal.org

Concluding Remarks
We have proposed the CSAS model which indicates the structure of Gij with j i 2 , namely, pij with
j i 2 .Usingthismodel,wehavegivenTheorem1.Thetheoremmaybeusefultoseethereasonforthepoorfit

whentheC2RPSmodelfitsthedatapoorly.Inpractice,wecanseefromTheorem1thatthepoorfitoftheC2RPS
modeliscausedbythepoorfitofthePSmodelratherthantheCSASmodelinexample(seeTable3).Forsquare
contingencytableswithorderedcategories,Theorem2wouldbeusefulforseeingthereasonforthepoorfitwhen
theSmodelfitsthedatapoorly.
ACKNOWLEDGMENT

Theauthorswouldliketothankananonymousrefereeforhelpfulcommentsandsuggestions.
REFERENCES

[1]

Bishop, Y. M. M., Fienberg, S. E. and Holland, P. W. (1975). Discrete multivariateanalysis: theory andpractice. MIT Press,
Cambridge.

[2]

Bowker,A.H.(1948).Atestforsymmetryincontingencytables.JournaloftheAmericanStatisticalAssociation43,572574.

[3]

McCullagh,P.(1978).Aclassofparametricmodelsfortheanalysisofsquarecontingencytableswithorderedcategories.
Biometrika65,413418.

[4]

Read,C.B.(1977).Partitioningchisquareincontingencytables:Ateachingapproach.CommunicationsinStatisticsTheory
andMethods6,553562.

[5]

Stuart,A.(1955).Atestforhomogeneityofthemarginaldistributionsinatwowayclassification.Biometrika42,412416.

[6]

Tahata,K.,Yamamoto,K.andTomizawa,S.(2012).Decompositionofsymmetryusingpalindromicsymmetrymodelina
twowayclassification.JournalofStatisticsApplicationsandProbability1,201204.

[7]

Tahata,K.,Yamamoto,K.andTomizawa,S.(2013).Decompositionofsymmetrymodelintothreemodelsforcumulative
probabilitiesinsquarecontingencytables.EuropeanJournalofPureandAppliedMathematics6,299306.

[8]

Tominaga, K. (1979). Nippon no Kaisou Kouzou (Japanese Hierarchical Structure). University of Tokyo Press, Tokyo (in
Japanese).

[9]

Tomizawa, S. (1993). A decomposition of the marginal homogeneity model for square contingency tables with ordered
categories.CalcuttaStatisticalAssociationBulletin43,123125.

[10] Tomizawa,S.,Miyamoto,N.andOuchi,M.(2006).Decompositionsofsymmetrymodelintomarginalhomogeneityand
distancesubsymmetryinsquarecontingencytableswithorderedcategories.Revstat:StatisticalJournal4,153161.
[11] Tomizawa,S.,Miyamoto,N.,Yamamoto,K.andSugiyama,A.(2007).Extensionsoflineardiagonalparametersymmetry
andquasisymmetrymodelsforcumulativeprobabilitiesinsquarecontingencytables.StatisticaNeerlandica61,273283.

21