Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp
.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.
Taylor & Francis, Ltd. and American Statistical Association are collaborating with JSTOR to digitize, preserve
and extend access to The American Statistician.
http://www.jstor.org
fH5(X) 16At({1})
Suppose thata Bayesian observesdata X = x and tests
a hypothesisH using a loss functionthatsays the cost of
1. INTRODUCTION typeII erroris some constantb overthealternative and the
Considertosses of a coin knownto be eitherfair,two- cost of typeI erroris constantover the hypothesisand is
headed,or two-tailed.There are six nontrivialhypotheses c x b. The posteriorexpectedcost of rejectingH is then
about 0, theprobabilityof heads: cbPr(H is truelX = x), while the posteriorexpectedcost
of acceptingH is b(1 - Pr(H is truelX= x)). The formal
H1:0=1 H2: 0 = 1/2 H3: 0 = O Bayes rule is to rejectH if the cost of rejectingis smaller
H4:0#71 H5:0 7#1/2 H6:0O. thanthe cost of accepting.This simplifiesto rejectingH
if its posteriorprobabilityis less than 1/[1+ c], whichis
Jeffreys (1960) introduceda class of statisticsfor testing
equivalentto rejectingH if the posteriorodds in its favor
hypothesesthatare now commonlycalled Bayes factors.
are less thanl/c. This, in turn,is equivalentto rejectingH
The Bayesfactorforcomparinga hypothesisH to its com-
if theBayes factorin favorof H is less thansome constant
plement,thealternative A, is theratioof theposteriorodds
k implicitlydetermined by c and thepriorodds.
in favorof H to thepriorodds in favorof H.
It wouldseemthenthata Bayesiancould declineto spec-
To make thismoreprecise,let Q be theparameterspace
ifypriorodds, interpret theBayes factoras "theweightof
and let QH c Q be a propersubset.Let ,ube a probability
evidencefromthedatain favourof the... model"(O'Hagan
measureover Q and, foreach 0 c Q, let fx IE ( I0) be the
1994, p. 191); "a summaryof theevidenceprovidedby the
densityfunction (or probabilitymass function)forsomeob-
data in favorof one scientifictheory... as opposed to an-
servableX givene = 0. The predictivedensityof X given
other"(Kass and Raftery1995,p. 777); or the"'odds forHo
H: e) C QH is fH(x) equal to the average of fxie(xj0)
to H1 thatare givenbythedata' " (Berger1985,p. 146) and
withrespectto ,t restricted to QH. Similarly,thepredictive
testa hypothesis"objectively"by rejectingH if theBayes
densityof X givenA: e , QH is fA(x) equal to the av-
factoris less thansome constantk. In fact,Schervish(1995,
erage of fx Ie(x 0) withrespectto Atrestricted to QA (the
p. 221) said "The advantageof calculatinga Bayes factor
complementof QH). That is,
overtheposteriorodds ... is thatone need notstatea prior
I odds..." and then (p. 283) thatBayes factorsare "ways
fH (X) fQHfxE (x 0)d-t(0) to quantifythedegreeof supportfora hypothesisin a data
1-t(QH)
set."Of course,as theseauthorsclarified, suchan interpreta-
and tionis notstrictlyjustified.While theBayes factordoes not
fx IE(xI0)d-t(0) dependon thepriorodds,it does dependon "how theprior
fA (X) fQlA
mass is spreadout overthetwo hypotheses"(Berger1985,
Pi(QA)
p. 146). Nonetheless,it sometimeshappensthatthe Bayes
If p is the priorprobabilitythatH is true-that is, p factor"will be relativelyinsensitiveto reasonablechoices"
,-(QH)-then the posteriorodds in favorof H is the ra- (Berger 1985, p. 146), and thena commonopinionwould
be that"such an interpretation is reasonable"(Berger1985,
p. 147).
Michael Lavine is AssociateProfessor,Instituteof Statisticsand Deci-
sion Sciences,Duke University,Durham,NC 27708-0251 (Email: michael
We show,by example,thatsuch informaluse of Bayes
@stat.duke.edu).MarkJ.Schervishis Professor,Departmentof Statistics, factorssuffersa certainlogical flawthatis not suffered by
CarnegieMellon University, PA 15213.
Pittsburgh, using the posteriorodds to measuresupport.The removal
120 General
122 General