Employment Discrimination and Statistical Science

Author(s): Arthur P. Dempster

Source: Statistical Science, Vol. 3, No. 2 (May, 1988), pp. 149-161
Published by: Institute of Mathematical Statistics
Stable URL: http://www.jstor.org/stable/2245557 .
Accessed: 17/04/2011 07:34

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.

Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at .
http://www.jstor.org/action/showPublisher?publisherCode=ims. .

Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.

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.

Institute of Mathematical Statistics is collaborating with JSTOR to digitize, preserve and extend access to
Statistical Science.

http://www.jstor.org
StatisticalScience
1988, Vol. 3, No. 2, 149-195

and
EmploymentDiscrimination
StatisticalScience
Arthur
P. Dempster

Abstract.Goodpracticeoflegalstatisticsdependson a foundation ofgood

statisticalscience.Causal inference
fromstatisticaldata dependsbothon
understanding of substantivecausal processesand adequateprinciplesof
statisticalinference.The paper makesa case that Bayesianreasoningis
neededforstatisticalstudiesofemployment A modelbased
discrimination.
on Bayesianprinciplesis developedin detailand is usedto showthatany
statisticalestimateof the effectsof employment discriminationmustbe
adjustedfromsourcesof knowledge outsidethe statistician's
data. Econo-
metricanalyses,whichsuggestotherwise, are analyzedand criticized.
Key wordsand phrases:Bayesianinference, causal analysis,employment
directand reverseregression.
discrimination,

1. INTRODUCTION judgment requiredofa modelbuilderwhoseeksobjec-

tivitywhenselectinga particularprobability model
Statisticalscienceis aimed at understanding real fora specificrealworldanalysis.Becausean essential
worldphenomena, whethernaturalor social,through featureofscienceis thatfindings and knowledge can
a combination ofempiricaldata analysisand statisti- be communicated, we need to makeevidentialbases
cal modeling. To approachscientific analysisthisway andjudgments and be preparedto arguetheir
explicit,
is in thetraditionofJerzy Neyman's work,and hence use in reasoneddebate.
the following studyof rolesforstatisticalarguments In recentyears,law courtshave struggled to define
in evaluatingregression-based claimsof employment standardsforthe use of statisticalevidencein legal
discrimination reflects manyscientific principles that proceedings. Many statisticians and econometricians
in
Neymansupported his long and admirable career haveparticipated in theprocessby servingas expert
as an innovatorand leader.In particular, I sharehis witnessesin adversary proceedings.Recentpapersby
beliefsthatprobabilities ought ideallyto be obtained Meier (1986) and Fisher (1986) surveythe current
fromempirical frequencies, andthatan important test stateoftheartand discussmanyproblemsofprofes-
ofreported inferences is long run conformity to spec- sional standardsfacedby statisticalexpertsin the
ifiedproperties suchas 0.05 levelsor expectations. field.My concernhereis less withlegalprocessesper
fromNeymanin usingthesubjectiveinter-
I differ se, butratherwithstatisticalreasoningprocessesthat
pretationofprobability, and in givingBayesianrea- in my view shouldbe pursuedwhen analyzingthe
soningratherthansamplingarguments a morecentral circumstances surrounding legal cases. My remarks
place in the practice of statisticalinference and reflectexperienceas an expertmainlyconfinedto a
decisionmaking.I see no contradiction betweenthe typeofemployment discrimination case wherea pro-
subjective interpretation of -probability and the tectedclass of employeesalleges systematicunfair
desirabilityof objectiveevidentialbases forassumed withholding of rewardssuch as compensationand
modelsthat incorporatesubjectiveprobabilitiesas promotion.Similar model buildingideas could be
technicalconstructs. In myopinion,themajorconcern applied more broadlyin the social sciences,most
aboutobjectivity arisesnotovertheissueofsubjective directly whenthe phenomenainvolverepeateddeci-
versusnonsubjectiveinterpretations of probability, sion making.DeGroot,Fienbergand Kadane (1986)
but ratheroverthe empiricalevidenceand scientific containsreviewsofbroadertypesofstatisticalissues
in legalsettings.
ArthurP. Dempsteris Professor ofTheoreticalStatis- A previouspaper (Dempster,1984) set fortha
tics,Harvard University, Cambridge, Massachusetts framework with three levels of information: facts
02138. This articleis a revisedversionof the IMS abouteach employeeknown to the facts
statistician,
NeymanLecturepresentedat the JointStatistical aboutthe employeeknownto the employer/decision
Meetings, Chicago,August20, 1986. makerand an ideal completeset of factsabout the

149
150 A. P. DEMPSTER

employee. The framework was usedto describea col- Are actions or decisionsby an employercausing
lectionof obstaclesfacinga statisticianseekingto employees to suffer ill effectsofdiscrimination? Stat-
inferthepresenceor absenceofdiscrimination, orthe isticiansand theirclientsclearlyneedsharedworking
extentof discrimination, practicedby an employer principles leadingtoagreedunderstanding ofpotential
againsta legallyprotectedclass ofemployeessuchas causal inferences.One set of such principles,ably
women,blacks or olderpersons.I also describeda reviewed byHolland(1986a),represents a majorintel-
controversy over.whichof two formsof regression lectualcontribution of statisticsto science.The idea
analysis,called directregression and reverseregres- is to identify randomization-based studiesas situa-
sion,is thecorrectformto applyto employment data tionswherecausal inferences maybe soundlydrawn,
whenseekingto estimatea discrimination effect,and and to evaluateotherattemptsat causal inference by
I suggested thatmyframework couldbe usedto resolve judgingwhethertheirnonrandomized circumstances
thecontroversy. Pratt(1986)concludeda briefdiscus- adequatelyconform to criticalassumptions whichare
sion of Dempster(1984) withthe remark,"But we transparently satisfiedin randomized studies.Statis-
mustkeep strivingtowardsensiblemodes of using ticianssuchas Prattand Schlaifer(1984) or Holland
statisticsin legal and public arenas-think of the (1986a)are skepticalthatmanyclaimedcausaleffects
alternatives!" The presentanalysisattempts construc- in econometric or othersocial sciencemodelscan in
tivedevelopment oftheideas in thepreviouspaper. factmeettheassumptions.
The goal ofSection2 is to definea modeofstatis- Because the data sets on whichclaimsof employ-
tical thinkingthat I believeto be appropriateand mentdiscrimination are usuallybased are adminis-
sufficientto deal withtheproblem.Then,in Section trativerecords,and thereforeabout as far from
3, I takethemodeling processbackto firstprinciples, randomizedstudiesas one can get, it mightseem
and set out the reasoningbehindthe typeof model that statisticiansconcernedabout theirreputations
whichI advocate.Simplemathematical analysisofmy forscientificcredibility woulddeclarecausal infer-
modelindicatesthateitherdirectorreverseregression enceimpossibleor worse,and leavethematterthere.
givesunbiasedestimatesof discrimination effectsif Nevertheless, whileI acceptthe cogencyand impor-
corresponding assumptionshold,but that in general tance of the negativearguments, I believethat we
neitheris valid,becausethecorrectanalysisdepends bear professional responsibility forcarrying the dis-
on thevalueofa parameter thatcapturesan essential cussion further, but along lines complementary to
aspectof the real world,but whosevalue cannotbe themainissuesraisedso far.
estimatedfromthe statistician's data. Two scientific Holland(1986a) classifieddiscussionsofcausation
hurdlesare raised by this situation.One involves intothoseaddressing themeaningofthetermcause,
teachingstatisticiansand othersto see clearlyand those attempting to understandor establishcausal
explicitlythat statisticalanalysison its own rarely mechanisms in relationto a specificclass ofphenom-
offers completesolutionsto externally specified prob- ena and thosetryingto identify and measurecausal
lems.Instead,statistical analysisbringsus tothepoint effects in specificsituations.Althoughonlythethird
ofseeingclearlywhatthegap in ourknowledge is, so typeis directly concerned withinferring causaleffects,
thatwe maythenaddressthe secondhurdle,namely, impliedattitudesand understandings involvingthe
the difficult task of lookingoutsidethe data forevi- othertwo typesmust underlieany specificcausal
dencebearingon the missinginformation. Section3 analysisofstatisticaldata. For myviewson theoper-
also relates the new models to more traditional ational meetingof causation,I referto Dempster
descriptionsof econometricmodels as typifiedby (1987). An essentialpoint is that causal language
Goldberger(1984), and argues that Goldberger's impliesthepresenceofsomeactionmechanism oper-
framework locksin arbitrary assumptions whichprej- ating in the real worldand yieldingconsequences
udice his conclusionsand lead him to proposean called causal effects.The secondtypeof discussion
irrelevant test forthe validityof reverseregression. picks up fromhere and asks, for example,in an
Finally,in Section4, I discussa conceptualdilemma, employment discrimination case,whatare the causal
whichcan scarcelybe avoided,eitherbyemployers or mechanismsoperatingto determinerewards of
regulators. The questionis: whatto do if economic employment? My claimthattheanalysisofSection3
efficiency and legal avoidanceof discrimination col- belowmakesa contribution tothestatisticsofemploy-
lide? I call thisthe problemofjudgmentaldiscrimi- ment discrimination rests on the adoptionof an
nation.I suggestthatthemodelofSection3 provides explicitviewofthebasic mechanism ofrewarddeter-
a basis foranalyzingthedilemma,and leadsto policy minationwhichis at best left implicitin traditional
attitudesconsistent withfairand reasonablewaysto econometric models.
handletheproblem. The distinction doesnotinvolvethediscrimination
Causalthinking to theissues.Is discrim-
is intrinsic mechanism itself,because itseemssafeandreasonable
inationcaussingreducedpay forcertainemployees? to regardthediscrimination causaleffectas a quantity
 EMPLOYMENT DISCRIMINATION 151

deliberatelysubtracted fromwhatwouldotherwise be subjec-

quentlyare adequatein practiceto represent
a fairand properreward,on thebasis ofsomeirrele- Considera modeloftheform
tiveuncertainty.
vantemployee characteristic.The problemarisesover (1) Y= Ga + X: + e,
specifyingthecausal mechanismor mechanisms that
determine the fairand properstandard.As nearas I whereeach termin the equationdenotesan n x 1
can understand,the traditionaleconometricmodel vectorand the ith rowin thevectorequationmaybe
conceivesthe fairrewardas determined by a mathe- written
matical formulawith added random disturbance, (2) Yi = Gia + Xi: + ei,
whichin mechanistictermsinvolvesthe action of
substitutingin a formula, throwing diceand perform- wherewe supposethatYidenotesthesalaryoftheith
ing a computation.Even though such activities, employee, Gi denotesthe genderofthe ith employee
exceptingthe dice throw,are essential,theyofferin (1 formale and 0 forfemale)and Xi denotesa 1 x k
themselves no explanationofhowfairness is achieved. vectorofk characteristicsoftheithemployee. Gender
I proposeto add anotheringredient to thedescription is used instead of, say, race or age, because it is
ofthecausalmechanism; namely,theconception that convenientto work with a concretehypothetical
theemployer/decision makeractivelyengagesin deci- examplewhichis a clearcutdichotomy. The elements
sionmakingunderuncertainty, so foreach employee e1,e2, ..., en of e are conventionallysaid to be inde-
computesa posteriorexpectedrewardbased on a pendentN(0, a2) randomvariables,whereasa, fdand
defensiblemethodology forassessinglikelycontribu- a-are said to be parameters ofthemodel.
tions of the employee.I do not claim that such a There is both consensusand disagreement over
mechanismis entirelyrealisticor factual,althoughI different aspectsoftheinterpretation oftheequations
do believethatit adds essentialexplanatory powerto (1) or (2). There are two major noncontroversial
the traditionalmodel. I do claim, as set forthin points.One is thepurelymathematical contentofthe
Section3, thatthe proposedmechanismhas conse- model,includingthe mathematical meaningofprob-
quencesforthewaya statistician shouldbuildmodels ability,whichhas shiftedlittlesince the measure-
for assessingpotentialdiscrimination effects,and theoretic revolutionalteredmathematical statisticsin
thenceforclarifying to the statisticianwhatare the the 1930s and 1940s.The otheris that,quite apart
gapsin information whichmustbe filledfromsources fromany probabilistic interpretation,the equations
externalto the data if the desiredinferences about have a well understooddeterministicmeaning,
discrimination are to be completed. wherebyeach of the n employeesis thoughtto be
associatedwithactualnumerical values Yi,Gi,Xi and
ei,whichobeythelinearrelationspecified in equation
2. THE MEANINGOF STATISTICALMODELS (2). Whenwe attemptto definethe realworldinter-
The use ofcausallanguageis notordinarily a source pretationof the probabilistic contentof the model,
ofmisunderstanding communication.
iinscientific The however,the consensusamongtheoreticalstatisti-
same is not at all trueofprobabilistic language,and cians disappears.The two main contendersmaybe
henceI disagreewithCox (1986)ontherelativeimpor- labeled(i) thechancemechanism meaningand (ii) the
tanceofestablishing philosophical underpinnings for personalmeasureof uncertainty meaning.As elabo-
causationand probability. Even in the causal arena, ratedbelow,I believethat (ii) is the onlysatisfying
confusionovertherolesofvariationand uncertainty meaning.Interpretation (i) is valid and useful in
generallydominatesdifficulties over understanding restricted circumstances, but thenis bothsubsumed
themeaningofthecausal mechanisms. underand clarifiedby (ii). As remarkedby Savage
I believethat statisticianswho make theirpeace (1967),"The foundational inthedefinition
difficulties
withconceptions ofprobability intermswhichexclude of personalprobability are less than those of other
orignoresubjective interpretationaredepriving them- attemptsto defineprobability, and the truthbehind
selves of structureneededforan explicitand fully other attemptsto define probabilityis correctly
comprehensible accountofuncertaininference. More- expressiblethroughthe theoryof personalprobabil-
over,whenmathematical modelsareusedto represent ity."See also Savage (1962,pages 102-103).
deterministicphenomena, thereis littledisagreement To apply the normal linear model (1) in the
aboutwhattheymean,but,whenprobabilistic models chance mechanismmode we need to believe that
are used to reflectchanceand uncertainty, disagree- e1, e2, ..., en are generatedby a random mechanism
mentusuallylies close to the surfaceand is oftena whichoperatesindependently ofX and G. Thereis a
serious impedimentto scientificdiscourse.I wish powerful attractionto thinking thisway,because (2)
thereforeto establishwithcaremytermsofdiscourse. thenbecomesa completeexplanationofhoweach Y1
Simple normal linear models are adequate to is determined, to wit,one firstoperatesthe chance
capturetheessentialtypesofvariation, and notinfre- mechanismto findei,thendetermines Xi and Gi for
152 A. P. DEMPSTER

theithemployee, henceYiis computed from(2). Ifan and mightseriously challengemanyuses ofstochastic

employer actuallyoperatesthisway,thena finding of modelsin naturaland engineering sciencescontexts.
a $ 0 is interpretable as a finding ofsex discrimina- I believethereis a betterway,namely,to recognize
tion, because the chance mechanismis sex-blind thatchancemechanisms areonlyonewayto establish
by definition, whereasXi: representsobjectivejob- scientificallyacceptablepersonalmeasuresofuncer-
relatedcharacteristics, hencetheremaining termGia tainty,thusturning attentionto thedifficult butnec-
is purelydetermined bygender,and so is discrimina- essarytaskofestablishing broaderstandards.
tory. I believethatreasonableprobability modelsmaybe
Courtshave been hearingarguments close to this constructed bothforsituationswherechancemecha-
fromstatisticaland econometric expertsfora decade nismsare identifiable in therealworldand forsitua-
or so. These arguments oftenlead to a curioussortof tionswhereno suchmechanisms are implied.In fact,
legalmischief, becausea judgemayrulethata plain- thereis a criticalcommonelementofhumanjudgment
tiff'sfitted modelofform(1) establishesa primafacie in bothvarieties.For example,in a cointossingsitu-
case fordiscrimination, thus puttingthe burdenof ation,wherethe conceptof physicalchancemecha-
proofon theothersideto showtheabsenceofdiscrim- nism may be plausible,applicationto specificreal
ination.The rebuttalis mostoftenattempted froma worldcircumstances requiresa judgmentthat the
similarcausal model,butaddinga term,sayX':l', on uncertainty representation is symmetric underall per-
the rightside of (1) to represent unobservedcharac- mutationsof the orderof the tosses. On the other
teristics ofeach employee. It is thenargued,correctly, hand,ifwe wishto learnfrom100pastexperiences of
thata valueofa estimatedfroma regression analysis a typeof situation,in orderto quantifyuncertainty
omittingX' is quite likelyto be badlybiased. But aboutthe nextexperienceofthe same type,then,in
such arguments are oftenrejectedby courtsbecause circumstances whereno actual chancemechanismis
the rebutterscannot actuallyproducethe proper remotelyin view,we may oftenbe willingto agree
regressionanalysisincludingX' (because X' is by thatprioruncertainty aboutthe 101 experiences(i.e.,
definition unobservable!), so are reducedto indirect uncertainty beforeexperiencing any ofthem)should
argumentwhichis judgedtoo weak to overturnthe treatthemsymmetrically. The objectivity ofprobabil-
moreconcrete primafaciecase.Another rebuttalstrat- itymodelsconstructed foreithervarietyof situation
egyis to producea largenumberof regressions with requiresthatourcolleaguesacceptthecorresponding
alternative choicesofX, whichmayexhibitvariations symmetry judgments as reasonable,and eithertypeof
in bothmagnitudeand signof estimatedsex coeffl- symmetry shouldbe scrutinized inmuchthesameway
cients,and so mayconfusetheissue. by tryingto identify asymmetries in the real world
All of these arguments, both forand againstthe repetitions involved.Jacob Bernoulli'sstudyof the
primafaciecase, dependin an essentialway on the binomialdistribution leadingto thelaw oflargenum-
chancemechanismas a causal factor.I believethata bers(Bernoulli,1713)explicitly usedtheurnmodelas
damagingconfusion operatesherebetweentheuse of an abstractrepresentation oftheproblemoflearning
chancemechanisms as analogiesandtheuse ofchance frompast experience.We shoulddo the same,while
mechanismsas realities,the former arisingwhenwe explicitly recognizing whatwe are doing.
saythatuncertainties in somerealworldsituationare The attractiveforcethat binds the two typesof
likethe uncertainties in a dice game.The purposeof situationtogether is the interpretation ofprobability
the analogyis mental illumination.Actual chance as a personalmeasureof uncertainty. For example,
mechanismsin real worldemployment contexts,if returning to model(1), the subjectiveinterpretation
theyexist at all, have nothinglike the major role is relevantto an individualanalyst who knows
prescribed in widelyappliedstatisticalmodels.Mak- thenumerical valuesofselectedXi and expressesper-
ing this negativeargumentto a judge is not easy, sonal uncertaintyabout the unknownnumerical
however, becausethe otherside can pointto a huge valuesofthe corresponding Yi by describing themas
literatureon causal modelingin econometrics and independent withN(a + Xi,B,a2) distributions. Spe-
sociology whichevidently restson similarassumptions cificnumerical probabilities calculatedfromthesedis-
ofrandomness. I believethatmuchstochasticmodel- tributions according totheusualrulesofmathematical
inginthesedisciplines is undermined bya dependence probability are specificpersonalmeasuresof uncer-
on fictitious chancemechanisms. tainty.In the coin tossingexample,the probabilities
How shouldwe proceed?One positionis to reject are relevantto the analyst'suncertainty about pro-
all scientific use ofprobabilistic reasoningunlessthe spectiveoutcomes.And so on, acrossall applications
real worldbasis of the chance mechanismis firmly offormalprobability models.
established.Certainlythis wouldrule out all use of The historical, recordas chronicledby Hacking
statisticalinference procedureson data fromobser- (1975) showsthat computedprobabilities werepro-
vationalstudiesin the social and medicalsciences, posed for decisionmakingin games of chance by
 EMPLOYMENT DISCRIMINATION
distinguished scholarsbeginning in themid17thcen-
tury,and thatwithindecadesthe further proposalto
extendsuchuses to widerscientific sphereswas well
established.In otherwords,probabilityis a basic
quantitativetechnicalconstruct likeforceor temper-
aturethatwe learn.tomeasure,to enterintocompu-
tationsand to applyin variousclasses of situations.
Individualprobabilitiesare meaningful because we
havelearnedhowtousethemin appropriately circum-
scribedsituationsto represent ouruncertainty about
unknownfactualcircumstances.
We do need to put moreeffort intobetterunder-
standingof what constitutesacceptablejudgmental
and evidential bases.Although thetheoryofprobabil-
ity is in a class by itselfwhenit comesto guiding
analyststo inferencesand decisionsunder uncer-
tainty,I do notagreewiththeviewsofsomeBayesian
advocatesthat subjectiveprobability shouldbe used
universally, makinglittleor no distinctionbetween
sourcesthat are trulysubjectivein the ordinaryas
opposedto technicalsense of the term,and sources
that are shared among communicatinganalysts.
Althoughcaution is advisable,we should explore
frameworks whichweaken Bayesian technologyin
returnformoreacceptableaccountsof modelcon-
struction,whileretaining subjective probability inter-
pretationas centralto the semantics.The theoryof
belieffunctionsis an alternativeI take seriously
(Dempster,1968;Shafer,1976,1987),butno detailed
modelsareas yetinplaceforstatistical practice,while
Bayesis at lastbecomingpractically feasible.
The subjectiveinterpretation requiresthat every
probability be understood as a conditional probability
giventhecurrent stateofknowledge oftheinterpreter.
Oftenthe intendedcurrentstateis onlyimplicit, but
mustbe clearfromthe contextofthe discussion,as,
forexample,inthecase ofa chancemechanism, where
theassociatedprobabilities are understood to be con-
ditionedon prospectiveunbiased operationof the
mechanism.On the otherhand,if the worldunder
studyis plainlynotgoverned by chancemechanisms,
thenstochastic modelersneedtothinkexplicitly about
whoseuncertainty theyaredescribing andunderwhat
conditioning.I often find probabilitymodeling
assumptionsin econometricstudies to be opaque
because no attentionis givento describingwho is
supposedto be subjectivelyinterpreting the prob-
abilitiesgivenwhatcurrentstateof knowledge, and
henceno clearbasis existsforhowthe modelswere
constructed, bywhom,or why.
3. AN ECONOMETRICMODEL FOR
EMPLOYMENTDISCRIMINATION
Stochasticmodelsare neededin employment dis-
criminationstudiesbecause the economicprocesses
153
are not fullyknownor predictableto any actoror
observerin the system.I assumethe presenceof a
primary actorcalledthe analystor statisticianwhose
tasksare firstto formulate modelsrepresenting hisor
her uncertain knowledgeabout relevant factual
aspects of the system,and secondto computeand
reportposteriorprobabilities and expectations given
theavailabledata,including theposterior distribution
of the effectsof discrimination in the systemunder
study.It is understood thatthemodelsareconstructed
to be as objectiveas possiblein thesenseoflayingout
the evidenceand arguments in favoroftheproposed
representations,so thatcoexperts andconsumers alike
maybe encouragedto place credibility in the infer-
ences.In the sequel,if an assumptionis referred to
withoutspecificattribution, as thoughmade by the
author, thereadermayassumethattheauthoris being
identified withtheanalyst/statistician,as a matterof
convenience.
Thereis also a secondary actorintheformalsystem
called the employer,who representsthe decision
makeror set of decisionmakerswithresponsibility
forsettingthe rewardof each employee.The basic
unit of statisticalanalysisis a decisionto supplya
rewardYito a particular employee, leadingto a sample
ofn suchdecisionsand associatedrewardsY1,Y2,. . .,
Yn. The reason behind identifying the units with
decisionsratherthanwithemployees, besidesthefact
thatthe data base mayreferto severaldecisionson
the same employee, is to emphasizethateach unitis
associatedwiththe activationof a decisionmaking
mechanism. Specifically,it is assumed,as partofthe
analyst'smodel formulation, that fairnesson the
employer's partmeansthatthe employer determines
Y1, Y2, ..., Yn by calculating n separate Bayesian
posteriorexpectationsof the true worthmeasures
Y1*, 2 *,*..., Yn whichare assumedto exist,butto
have values unknowneitherto the employeror the
analyst.Note the implication thatthe employer sets
up n different probabilistic representationsofuncer-
tainty,one foreach unknownY.**
The realismoftheassumedYi**maybe questioned,
forexample,because it is impossibleto disentangle
thecontributions ofindividuals fromthoseofa group,
or becausethe utilityof an employee'scontributions
is just too hard to defineand measure.It may be
difficultto measurerewardYitoo,especiallyifintan-
giblesare included.For purposesofthepresenttheo-
reticaldiscussion,I assumethe existenceof Yi and
Yi**as a minimalsetofconstructs necessaryto getat
a difference Y* - Yi whichis to be interpretedas a
measure of employee deficit. Note the implicit
assumption herethatYiand Y", and Yi*tobe defined
later,aremeasurements on a scalesuchthatsumsand
differences are meaningful.
As proposedin Dempster(1984),the statistician's
154 A. P. DEMPSTER

modelassumesthata vectorXi of employeecharac- samplesize is largeenoughto renderparameters such

teristicsassociatedwiththe ith decisionis available as a, f and a2 effectively knownprecisely. In practice,
bothforpurposesofmodelconstruction and ofdraw- ofcourse,samplingerroris usuallylargeenoughto be
ingfinalinferences.The statistician
also assumesthat interesting, and influentialregardingmodel choice,
theemployer possessesa moreextensivevectorXi*of but herethe intentionis to focusthe mainattention
characteristicsassociatedwiththe ith decision,and on bias effects whichare unrelatedto samplesize. A
thatbothbelievethereis a stillmorecomprehensive similarassumptionwas madeby Prattand Schlaifer
vectorXi* of characteristicsincludingthe informa- (1984).
tion needed to determinethe true worthmeasure In econometrics, a linearmodelsuchas (1) is often
Yv**
Ii describedas a causal model.One motivation forthe
Althoughthe basic units of analysisare called a termis simplythat economicscienceis directedat
sample of decisions,theyare not assumedto be a elucidatingthe causes of economicphenomena,and
randomsamplefromany definedpopulation.They econometric modelsare seen as contributing in some
are,however, assumedto be exchangeable inthesense wayto this goal. But whatis the causal mechanism
ofde Finetti,forpurposesofconstructing a probabi- reflectedin (1)? A possiblecausal interpretation of
listicmodel.This device,whichis surelythesinequa partofthe modelis to describee as arisingfromthe
non of Bayesianmodelconstruction, simplymeans operationofa chancemechanism. I arguedin Section
that the formulation musthave mathematical sym- 2 thatbeliefin suchmechanisms in theworldofreal
metry underall permutations oftheindicesi,implying employment decisionsis questionable, andI suggested
thatdifferences amongthecircumstances surrounding thatthe probability aspectsof the modelshouldbe
individualdecisionsmustbe explicitly builtintothe interpreted noncausally.Perhapssensingthe weak-
subscripted scalarsand vectorsappearingin the for- nessofthechancemechanism view,Goldberger (1984,
mulation.Becausethelatterconditionmaybe contra- page 108) firstdescribedthe terme in the model(w
dictedbyexperience, therecan be nothingimmutable in his notation)as a "disturbance," and thenwenton
abouta particular assumption ofexchangeability, but to identify the "disturbance" as representing "addi-
whateverdegreeof objectivity moststatisticalinfer- tionalinformation availableto the employer but not
encesbasedon sampleshas is dependent on a consen- to the statistician."Pratt and Schlaifer(1984,page
sus aboutsomeexchangeability hypothesis. 11) quote similarphrasesfromseveralwell-known
The famousde Finettitheoremon exchangeable econometrics textbooks.
randomvariablespermitsus to thinkofthesampleof What causal factorsshouldentera causallyinter-
decisionsas theuncertainty equivalentofa multivar- pretablestochasticeconometric modelofemployment
iaterandomsamplewherethevariablesarethelistof discrimination? Although causalanalysescan be com-
propertiesof each decision.Decidingto represent plicatedendlesslyby diggingdeeperand deeper,I
uncertainty as equivalentto a multivariate sampling believewe need onlyconsidertwo primaryfactors,
hypothesisleads to the furtherquestionsof what corresponding to controland treatedin the language
familyof multivariate distributions to specify,and, of experimental statistics,wherethe controlcausal
forBayesianstatisticalanalysis,whatpriordistribu- mechanism definesa standardagainstwhichpossible
tionoverthefamilyto specify. The maintraditionin discrimination is to be determined. the
Specifically,
appliedeconometrics is to specifysamplingdistribu- controlmechanismwhichI proposeforadoptionby
tions crudelyfrom first and second conditional the statisticianhas the employer activelyprocessing
momentswhosevalues are establishedby a mixture the information availableon each employee(i.e., G
ofassumption andestimation fromdata.Forpurposes and X*), and computing
of Bayesian analysis,momentassumptionscan be
(3) Y* = E(Y**IG, X*).
enhancedto normalmodels.The robustnessof the
resulting simpleanalysesin actualpracticeremainsa The secondcausal mechanismis overtsex discrimi-
trickyquestionwhich,froma Bayesianperspective, nationwhereby theemployer adds a' to Y* to deter-
can only be addressedon a case by case basis by minethepay ofmaleswhilepayingfemalesonlyY*,
computingwith more realisticmodels, something i.e.,theemployer sets
whichis impossiblegiventhe currentprimitive state
(4) Y=Ga/'+Y*.
oftheart.Because thepresentdiscussionrefersonly
to a hypothetical situatiorn,however, thereis no ques- Both of the causal mechanismsjust described
tionof faithfulness to an actualpopulationdistribu- are implicitin standardeconometric analyses.For
tion,and I assumesimplenormalmodels. example,Goldberger(1984) uses a as an "assess-
Anothersimplifying assumptionmade hereabout mentof discrimination," whereI use a withvari-
the hypothetical situationis that the statistician's ous superscriptsfor several different measuresof
 EMPLOYMENT DISCRIMINATION 155
He usesp forthe "employer's
discrimination. assess- where
mentofproductivity," to
whichI believecorresponds a* = a' + a".
(7)
my Y*. He does not introduce Y**, however,and so
does notexplicitly represent the mechanism (3). Ab- One consequenceof (6) is thatthereis no waygiven
sentthe mechanism, in myopinion,the varioussci- Y,X* and G,letalonegiventhemorerestricted actual
entific,ethicalor legal interpretations of the model data Y, X and G, to separatea* intothetwocompo-
cannotbe understood. In particular,thebasicdistinc- nentsa' and a". I call a' the effectof prejudicial
tionbetweenjudgmentaland prejudicialdiscrimina- discriminationbecause it carriesthe overtdiscrimi-
tion,whichI introducebelowand discussfurther in nationrepresented in (4). I call the othercomponent
Section4, requiresexplicitrepresentation of Y**. a" judgmental discrimination becauseit is a technical
Next I explainhow the modeldefinedby (3) and formof gendereffectcreatedas a byproduct of the
(4) can providea statisticianwitha framework to employer'suse of a basic tool of decisionanalysis.
estimatethe effectsof employment discrimination. The problemsposed by the presenceof judgmental
The precisedetailsof the modelare as follows.The discrimination are deferredto Section 4. Here, we
characteristicsof individualemployees,otherthan proceedto the problemof howto estimatethe total
gender, aretreatedas jointlynormalwithinemployees discrimination effecta*.
and independentacross employees.The covariance The problemwouldofcoursebe trivialiftheavail-
matrixis takento be thesameforall employees, while able data were Y, X* and G, insteadof the actually
thereare twopossiblydifferent meanvectors,one for availableY, X and G, i.e.,we wouldsimplysolvethe
each gendergroup.Because all conditionaldistribu- equations(6) and obtainvaluesfora* and ,B*.What
tionsofone subsetofcharacteristics giventhevalues statisticalexpertsactuallydo, literallythousandsof
of anothersubsetof characteristics are normalwith times,is carryout regression analysesbased on the
meanlinearin theconditioning setand residualinde- model(1) whichsuperficially resembles thepostulated
pendentofthe conditioning set,it is possibleto con- truemechanism(6). Can one relatethe a from(1)
structlinearmodelsgalore.Although all suchmodels withthea* from(6)? Ifso,can oneadjusttheincorrect
havepredictive interpretations, onlythoseassociated a to obtaina*? The answerto the firstquestionis
withidentifiedcausal mechanisms meritcausal inter- yes,easily.Formulasforbias in estimatedregression
pretation. coefficientswhenthe wrongX is used have a long
The translationof(3) intoa linearmodel,namely, history,and I derivesimpleexpressions belowformy
model.The answerto the secondquestionis no, not
(3*) Y** = Ga" + X*B* + e**, easily.
Surprisingly,many statisticiansand econometri-
does havean identified namely,
causal component, ciansappearuntroubled bythefactthattheavailable
a and the truea* are different and seemcontentto
(5) Y*-Gao"+X*#*
reporta as a valid measureof discrimination. One
in (3*) reflecting
the employer's conditionalexpecta- possiblereasonis thatthebias a - a* is judgedsmall
tionjudgment, whereasGa' in (4) reflectsovertsex enoughto ignore.A secondreasonmaybe confusion
discrimination. over the relativescientificmeritof the alternative
The interpretation of e** is that it consistsof models.Afterall, model(1) can be derivedrigorously
employee characteristicsnotavailabletotheemployer fromthe normalassumptions, as a decomposition of
at thetimeofdecisionmaking,i.e.,it couldbe expres- Y whosetermsdescribeprobabilistic prediction ofan
sibleas e**= X**:**.A criticalmathematical property uncertainY givenan observedX. Even thoughthe
of the modelis that the conditioning operation(3) validinterpretation of (1) is based onlyon statistical
guaranteesthate** is independent ofX* and has the associationand not on establishedcause, the term
samemeanforbothgendergroups(whichcan be taken causal modelis so oftenappliedindiscriminately in
to be 0). Because in generalthereis no reason to econometrics thatmanypractitioners maynotdistin-
expect employeecharacteristics to have the same guishbetweenlabeledcausalandactualcausal.A third
meansin bothgendergroups,and thereis plentyof reason,whichcan operatesynergistically with the
empiricalevidenceto the contrary, it is importantto second,is a typeofdefaultlogic.If a modelspecifies
stressthespecialreasonfortheequal meansproperty parameters whichcannotbe estimatedfromthedata,
here,namely,thatthecausal componentY* in (3*) is or in econometric termsare not identifiable, thereis
obtainedfromconditionalexpectation. a tendencyto blamethe modelratherthan the real
Combining (3*) and (4) intoa singleequationyields world,and to choosethemodelwhoseparameters are
The firstlineofreasoning
identifiable. maysometimes
(6) Y = Ga* + X*,8*, be successfully appliedon a case by case basis (cf.
 156 A. P. DEMPSTER

Kraskerand Pratt, 1984). I regardthe othertwo ofinformation on X2,and thenceon gM2 and AF2. It is
arguments as specious.Causal interpretation is essen- easy to constructartificialX* consistentwithgiven
tial to inferring discrimination as a cause and cannot data Y and X such that gM2 and AF2 have arbitrary
be taken lightly.There is no logical connection values.Such arbitrary valuesmayoftenbe unreason-
betweenwhetheror not a parameterexists and is able on a priorigrounds,butthereis no magiccureto
important in the real world,and whetheror not we be foundin thedata forthebias in (12). In thissense,
are luckyenoughto have data permitting estimation we knowthatit is hopelessto tryto solvetheproblem
of the parameter.If the data are not there,the only by replacingthe traditional"direct"regression based
courseopento theobjectively orientedBayesianis to on (1) withan alternativeformcalled"reverse" regres-
seekbetterdata,or,muchthesame,to seekobjective sion, whateverthe definitionof reverseregression.
sourcesofpriordistributions. Still,the reverseregression storyis fascinating,and
Despitethesenegativeremarks, thereare excellent leads me to concludethat reverseregressionhas a
reasonsforbeginning a studybyfitting model(1) and possiblerolein helpingthestatistician whois serious
estimating a. One reasonis to findoutwhatthedata aboutdeveloping a prior(= posterior)distributionfor
can legitimately tell us, namely,how well Y can be the bias a - a*.
predictedfromX. A secondreasonis moretechnical. The originalmotivation forthe statisticalmethod
It turnsout to be easierto giveinterpretable mathe- calledreverseregression, as wellas forthecontrasting
maticalexpressionsforthe bias a - a* than fora* termsdirectand reverse,comes fromcontrasting
directly. This is dueto thesimplicity oftheprocessof definitions of "fairness"whichare virtuallyfreeof
addingvariablesto a regression analysis. stochasticor causal modelingassumptions.Both
Ifwethinkhypothetically ofcarrying outtheregres- approachesagreethateach employeeshouldbe paid
sion analysisindicatedby (3*) in two stages,first exactlywhathe or she deserves,and thenask fora
regressing on X, and secondbringing in the informa- substitute principleto be usedin therealworldwhere
tionin X* notcontainedin X, we maywrite such perfectionis not achievable. In the first
X*f* = X1 + X2
approach,theprincipleis to requirethat,givenequal
(8)
qualifications, malesand femalesshouldbe paid the
whereX1 has the formX#%',whileX2 has the form sameon average.Ordinary, ordirect,regression based
X*,8 andthelinearcompounds represented byX2are on (1) is seen as a meansto obtaina suitablequalifi-
withX. Substituting
uncorrelated (8) into(6) leadsto cationsmeasureto be usedas a practicalstandardfor
judgingsuch equalityof pay averagesover gender
(9) Y= Ga* +X1 + X2,
groups.It followsthat a = 0 is the criterionforno
hencecomparison
with(1) showsthatW.'= fiso that discrimination, and a is the amountto be added to
femalepay to achieveparitywithmales.This simple
(10) X1 = XJ,,
lineofreasoningsoundsappealing,untilit is realized
and thatthechoiceofstandardis farfrominnocuous.For
example,I also supportthe principle,but withthe
(11) Ga* + X2= Ga + e.
more appropriateX*,8*used in place of X,8. The
If we denotethe male and femalepopulationmeans secondapproachreversesthe rolesofpay and quali-
ofX2by gM2 and AF2, and takepopulationaveragesof ficationsand suggeststhat the criterionforno dis-
(11) forG = 1 and G = 0, then,because e has zero crimination shouldbe that males and femaleswith
meansforbothgendergroups,we findthatAF2 = 0 givenpay shouldon averagehaveequal qualifications
and a = a* + ,UM2,hence measure.Statisticianscontemplating reverseregres-
sion are also naturallydrawnto X,8 from(1) as a
(12) a = a* + -
(AMe AF2)
suitablecompositequalifications measure,hencetest-
In words,thebias fromusinga in place ofa* is given ingfordiscrimination requireslookingat the"reverse"
bythedifference ofmaleandfemalepopulationmeans regression of X,8on Y and G. If the gendergroups
ofthe additionalpredictive variableX2 knownto the haveparallelbutdifferent regressionlines,as implied
employerbut not to the statistician.The reasonfor by the assumptionof commoncovariancestructures
retainingAF2 in (12) is thatthe conditionLF2 = 0 iS in thetwogendergroups,theremedy is to adjusteach
an artefact
oftheparticular choiceofgenderindicator female'spay by the amountrequiredto bringthe
G. Formula(12) holdsforany genderindicatorsuch regression linesintoconformity. The problemis that
that the male-femaledifference has absolutevalue the requiredshiftis different in the case of reverse
unity. regression fromthecase ofdirectregression.
data Y and X providea and X1,
The statistician's All statisticiansknowthattheregression linesof Y
assumingeffectively infinitesamplesize,butis devoid on X and of X on Y have different slopes,but the
 EMPLOYMENT DISCRIMINATION 157

apparentlyparadoxicalconsequencesformultigroup tance betweenthese lines measuredon the y scale.

datahaveprovoked extendeddiscussiononlyrecently. by aR, it followseasilyfrom(13)
Denotingthiseffect
The proposalto use reverseregression as a standard and (14) that
foremployment discrimination was putforth byBirn-
(15) aR = a* + - - -
baum (1979). The idea was independently conceived (AM2 AF2) [r2/_r1](M1 AlFi)

by HarryRobertsin the summerof 1979,at a time
when he was cailryingout massive analyses of
employeerecordsat the HarrisBank in Chicago,in
preparationfora hearingon chargesthat the bank
was practicingdiscrimination. Therehas been a long
subsequentliterature, includingConwayand Roberts
(1983),withsubsequentcommentaries and rejoinder
in theApril1984issueoftheJournalofBusinessand
EconomicStatistics, and Goldberger (1984). The key
pointis thatdirectand reverseregressions oftengive
conflictingmessages.In situations wheremalesareon
averagemorequalifiedthanfemales,it oftenhappens
that the average male salary exceeds the average
femalesalaryamongemployeeswitha givenqualifi-
cationsmeasure,suggesting discrimination againstfe-
males,whereassimultaneously amongemployees with
a givensalarythe averagequalificationmeasureof
males exceedsthe averagequalificationmeasureof
A comparisonof (12) and (15) showsclearlyhow a
changefromthe"direct"assessmenta tothe"reverse"
assessmentaR can easilyswitchthesignoftheeffect.
Note thatall the quantitiesappearingin the addi-
tionalterm[T2/T1](AM1 -AF1) are determined by the
data Y, X and G, so thattheproblemof
statistician's
fromtheprob-
assessingthebias in aR is no different
lemofassessingthebias in a, i.e.,fora personalistit
is stillthe problemof assessinga probabilitydistri-
butionfor(M2 - F2), givenwhateverevidencecan
be broughtto bear on the question.I do thinkthat
some statisticiansfacingthis admittedly formidable
taskwouldbe helpedby havingthe additionalquan-
titiesassociatedwithreverseregressionin view.In
particular,the conditionthat aR is unbiasedcan be
written
(16) /M2 -
LF2 = [T2t1](AM1
-
AF1)

females, suggesting discrimination againstmales.The This says that an unknowngenderdifference is a

papers cited above containmany examplesof the specifiedmultipleof a knowngenderdifference. I
phenomenon derivedfromactualdata. Simpletables mightor mightnotbe able to findevidencethatthe
ofaveragesoverqualifications variableswithinsalary specifiedmultipleis correct,but I believeI would
categories,whencomparedwithaveragesoversalaries oftenfindit easier to thinkabout whatthe correct
withinqualification categories, are sufficientto showmultipleshould be than to thinkabout the more
striking results. abstractLM2 - AF2 directly.
I have alreadyarguedabove that neithermethod Finally,I wish to reconcilemy analysisof direct
can be provedto give an unbiased assessmentof versus reverseregressionwith that of Goldberger
discrimination whenallowanceis madeforthediffer- (1984). Goldberger clearlysets out to refute"various
ence betweenthe employer'squalifications measure claims"and "certainimpressions" made and leftby
X*,8*and the statistician'sconstructed X,8. I now "criticsofdirectregression and proponents ofreverse
derivean expression forthebias in thereverseregres- regression." His approachis to describethreepossible
sion measureto complement the expression(12) for models,each meantfor"causal" interpretation and
thedirectregression bias. In additionto thenotation furnished accordingly witha "pathdiagram."Under
AM2 and AF2 forthe male and femalemeansofX2,we Model A, "MultipleCauses" (Goldberger's quotation
needthenotationgM1 and AF1 forthemaleand female marks),the directregression estimateof discrimina-
meansofX1,and notationri and r2 forthevariances tioneffectis shownto be unbiased.UnderModel B,
of X1 and X2 whichthe simplemodelrepresents as"MultipleIndicators," thereverseregression estimate
commonto bothgendergroups.To obtainequations is shownto be unbiased.UnderModel C, "Errorsin
fortheregression linesofX1 on Y, we mayuse (9) to Variables,"neitherestimateis generallyunbiased.
expressthe regressioncoefficient of X1 on Y as Goldberger arguesthatModel B "is the onlyknown
COV(X1, Y)/VAR( Y) = T1 (T1 + T2)-1. Thus, the equa- specificationunderwhichreverseregression provides
tionsofthe male and femalereverseregression lines a validestimator ofa," wherea denotestheeffectof
are discrimination againstfemalesin all his models.He
shows thatModel B has a testableproperty, and gives
(13) x - AM1 = r1(r1 + T2) [Y -a - /M2]
-m,
an examplewherea significance teston the statisti-
and cian's data rejectsthetestableproperty, and another
examplewhereit does not,perhaps in the lattercase
(14) X - ALF1 = r1(r1 + T2)[[Y -
AF2I.
AF1
because the sample is too small. He concludes that
The reverseregression estimateofthediscrimination "reverseregressionshould not be taken seriously
effectin favorof males is the male-less-female dis- unlessaccompanied bytheinformation neededto test
 158 A. P. DEMPSTER
the restrictionsofthe multipleindicatormodel."He
refutescertainclaimsforreverseregression madeon
the basis of Model C, but I did not find a clear
discussionofhowhe proposesto deal withthebias of
the directregression estimatorunderModel C. The
readeris leftwithlittledoubtthatGoldberger consid-
ersdirectregression to be themaincontender.
I findGoldberger's analysisto be flawedin both
generaland specificways.The generaldefectis the
(forme) mysterious way in which"causal" models
appearand departwithoutanydetailedconnection to
causal processesin the realworld.Goldberger (1984)
misinterprets a remarkofDempster(1984):by being
"somewhatskepticalaboutthe existenceof a chance
mechanism," I was notarguingthatan errortermin
a certainlinearmodelhas zero variance,but rather
thatthewholeconceptoferrortermis beinginappro-
priatelyborrowedfromstochasticmodelsadoptedin
quite different fieldsof science. Goldberger(1984)
strengthens myskepticism.
More specifically, Goldberger'sModels A and B
makestrongassumptionswhichare essentialto his
claimthatModelA justifiesdirectregression, andthat
Model B facilitatesrejectingreverse regression
througha significance test.Whentheseassumptions
are removed,both models are consistentwith my
causal model.If condition(16) is assumed,mymodel
providesa counterexample to Goldberger's
hypothesis
that Model B is the onlymodelsupporting reverse
regression.I demonstrate belowthatGoldberger's test
can onlyinvalidateunnecessaryassumptionswhose
removalleadsbackto mymodel.
ConsiderModelA, "MultipleCauses,"whichI take
to be Goldberger's modelofchoice.ModelA is speci-
fiedbythreeequations,
"(30a,b,c) y=p+az, p=,B'x+w,
x = liz + U."
Equation (30a) is trulycausal because it represents
overtdiscrimination, and is the same as my(3) with
thedifferent symbolsy forY,p forY*, a fora" and z
forG.The secondandthirdequationsareformal linear
models of a sort widelyencounteredin statistics,
wherethe termsw and u are random"disturbance"
termsassignedzero means,while f3'xand,z are
systematic components.
Equation(30c) is easilyunderstood in a descriptive
senseas sayingthatthe qualifications x (or X in my
notation) differin gender means. Because x is
observed, thisgenderdifference is an empirical factin
mostapplications. But thepossiblecausal interpreta-
tionsof (30c) are unspecified,and hence,in myview,
the label causal shouldbe avoided.For example,can
it be meaningful to say that a genderdifference in
yearsofschooling is causedbygender?I do thinkthat
suchterminology can conveyscientific meaningas a
kindofunderstood shorthand forcausal mechanisms
relatedto gender.It is no doubttruethat complex
social forcesoperateto createeducationaldisadvan-
tagesto femalesin certainlabormarkets, and advan-
tages to femalesin othermarkets,witha net bias
towardhelpingmalesto obtainmoreprestigious and
higherpayingjobs.But all statistical among
variation,
individualsas well as amonggroups,is ultimately
susceptibleto such open-endedcausal explanation.I
suggestthat it is preferableto maintainthe term
descriptiveforvariationonly vaguelyexplainedin
causalterms, whilesavingthetermcausalforspecified
explanations.
Equation (30b) is the mostinteresting of the set,
becauseit mustcarrywhatI calledabovethecontrol
causal mechanism. Myversionof(30b) is
(17) Y* = Ga' +X1 + X2,
whichfollowsformally from(5) and (8), andoriginally
from(3) whichspecifiesthe decisionmakingmecha-
nismofthe employer. By definition,Y* and p coin-
=
cide,and myX1 X,8and his fl'xare thesame.And,
becausew is describedas "theadditionalinformation
availableto theemployer butnotto thestatistician,"
itmustbe thatwandX2arethesame.The differences
between(17) and (30b) are therefore thatthe former
includesthe termGa", whichintroduces judgmental
discrimination,and that I allow fornonzerogender
differencein X2, whereas Goldbergerarbitrarily
assumeszerogenderdifference in w. The resultis his
conclusionthat directregressionleads to unbiased
estimatesof discrimination effects,whereasmyfor-
mula(12) leavesthestatistician notknowing thebias.
Thereis typically a substantialobservedgendercom-
ponentin X1t;hence it is plausiblethat a similar
componentwould be found in X2 if it could be
observed.I suggestthatdefaultlogicis inappropriate
whenthereis goodreasonto doubtthedefaultoption.
Considernext Goldberger'sModel B, "Multiple
Indicators,"specifiedby
"(40a, b, c) y =p + az, x = yp+ e,
p = liz + U."
Again,the firstequationrepresents overtdiscrimina-
tion,whilethethirdequationmakesdescriptive sense,
but lacks explicitcausal explanation.Model B is
unlikeModelA in thatthesecondequationalso lacks
causal sense,because the employer'sassessmentof
productivity cannotbe a causal determinant ofjob-
oftheemployee.
relatedcharacteristics
Goldberger a quoteof
motivates(40b) as reflecting
HarryRobertsto theeffect thatjob-relatedcharacter-
istics are "surrogatesor proxiesfor productivity,"
whichsuggests to methat(40b)is meantforpredictive
 EMPLOYMENT DISCRIMINATION 159

Earlierin hispaper,on page104,Gold-

interpretation. would offerno help at all in decomposinga* into
bergernotes"confusion" overan "elementarydistinc- a' + a", whichwouldrequirefurtherempiricalmeas-
tion"between"a proxyforp(intheimperfect correlate ures of Y -Y*.
sense)" and "a falliblemeasureof p (in the strict Is it important to separatea* intoitscomponents?
errors-in-variable
sense)." By omittingsystematic I believeit is scientifically important to do so,evenif,
effects
otherthanp from(40b),he evidently optsfor as a matterof law or policy,a* is definedto be the
the fallible measures interpretation of Roberts's effectof discrimination. I believealso that,if legal
proxies. expertsorpolicymakerswereto understand thetech-
Supposewe back up fora momentand ask what nicalaspectsofjudgmentaldiscrimination thatI will
would have becomeof Model B if nonzerogender now attemptto explain,theymightwish to work
effectshadbeenallowedineachoftheequations(40b). toward replacinga* by a differentstandard.In
Justification of reverseregressionwould not have Dempster(1984),I tracedtheidea ofjudgmental dis-
followed.But the modelwouldhave been stochasti- crimination to Phelps(1972),whocalledit "statistical
callyequivalentto mygeneralmodel,whichjustifies discrimination." The term"statistical discrimination"
neither directnorreverseregression. Goldberger could is a goodone,but couldbe confusing in mycontext,
thenhavesoughttheconditionforreverseregression becausethe"statistical" refers to statisticalreasoning
to be valid,whichwouldhaveledto a singlecondition by the employer, whereasmuchof mypaperfocuses
equivalentto (16), notto k conditionsas in (40b). By on a statisticianengagedin a verydifferent sortof
arbitrarilyassumingzero gendereffectsin all of the statisticalreasoningfromthat of the employer.
equations(40b),he introduces a strongmathematical WhereasPhelps (1972) statedthatjudgmentaldis-
assumption, namely,thatthemeangenderdifferences crimination is "damaging"and "important forsocial
ofeach component ofx shouldbe proportional to the policyto counter,"Dempster(1984) suggestedthat
regressioncoefficients of these componentson p, thequestionneeded"further elucidationand debate,"
wherethe constantof proportionality is the mean whichI nowattempt.
genderdifference on the employer's assessedproduc- The expectation operationin equation(3) expresses
tivity.The strongassumptionof equalityamongk an essentialprinciplegoverning applicationsof the
quantitiesis indeedtestable,butfailureofthepropor- theoryof personalprobability. There are two basic
tionalitydoesnotinvalidate(16),whichis nottestable ways of gettinginto troublein the long run if the
becauseit containstermswhichcannotbe estimated. principleis consistently violated.Both are conse-
In short,I believethat the modelsA and B of quencesof the property that,if the principleis fol-
Goldberger (1984) can and shouldbe reconciledwith lowed,then in a long sequenceof trialsthereis no
mymodelinsucha wayas toundermine hisarguments wayto selecton thebasis ofXi*and Giat each triala
fordirectand againstreverseregression. Both forms subsequencewhoseaveragediffers fromthatofaver-
ofregression are subjectto bias thatpurelystatistical agingthepredictions ofthe expectation operation.It
methodsare unableto correct. follows that,ifin a longsequenceoftrialsan employer
consistently pays womenan amounta" morethan
4. JUDGMENTAL DISCRIMINATION AND OTHER indicated by (3), whilepayingmenexactlytheamount
PROBLEMS indicated by (3), thenin the longrunwomenwillon
average receive a" more than their true Y**, while
If the statistician's
problemis to learnas muchas men on averagewill receiveexactlytheirtrue Y**.
possibleabouttheemployer's rewardsettingprocess, This willhappenwithinanyqualifications class,and
thenwe need to addressboththe problemdiscussed willsurelymakea prima facie case fordiscrimination
in Section3 of adjustinga* forstatisticalbias, and againstmen (assumingpositivea"). The seconddif-
the problemof separatinga* into its components ficultywithusing a" to adjust femalesalaries for
definedin (7); namely,a' whichI call theprejudicial judgmentaldiscrimination is that a competitor can
discrimination effectbecause it is a simple,across- selectagainstthe rewardsettingmechanismso that
the-board,gender differencehaving no basis in in the long run the employerwill be leftwithonly
productivity assessmentand a" whichI call thejudg- femaleemployees paid on averagean amounta" more
mentaldiscrimination effectbecauseit arisesin the thantheiractual averageproductivity, whilethe all-
courseof a presumedhonestattemptto assess pro- malecompetitor has the advantageofpayingexactly
ductivity.The firstproblemcould be solved by the rightamount.These hypothetical arguments are
eliminatingthe information differential betweena based on highlyidealizedcircumstances, of course,
statistician's
X and an employer's X*, by somemix- but the phenomenaare surelyreal. It followsthat
tureof gettingmoreinformation to the statistician adjustmentsfor judgmentaldiscrimination require
or restrictingtheemployer to moreobjectively meas- regulationswhichoperatesuccessfully againstmar-
urableX*. But, knowledgeof a* obtainedthis way ketforces.
160 A. P. DEMPSTER
A different sort of problemwithabandoningthe
principleof equation(3) is the lack of any principle
to put in its place. For example,it mightbe thought
thatreplacinga" byzeroin equation(5) woulddefine
an excellentnondiscriminatory rewardsettingmech-
anism. Such mechanismscan only lead to endless
controversy, however, overwhatvariablescan legiti-
matelybe includedin X*, because some variables
wouldbe criticizedas intrinsically taintedcarriersof
discrimination. I argued in Dempster(1984) that
manyvariablescould legitimately be regardedas so
affected,invarying degrees, thusposingdifficult prob-
lems of judgmentfora regulator.Notice,however,
that,ifthe definition offairnessimplicitin equation
(3) is maintained, thereis no restriction at all on the
variablesadmittedto X*.
Therearewaysto addressrealproblemsofdiscrim-
inationwithoutviolatingthe theoryof probability.
For example,iftheproblemis to redressthenegative
effectsof past overtdiscrimination on the qualifica-
tionsof a protectedclass, thenthe cost of doingso
could be assessed as a tax on employers. Assuming
that reasonableequityamongcompetingemployers
can be maintained,then overpayingmembersof a
disadvantaged class whiletheyimprovequalifications
is certainlya reasonableformoftax. Variousequiva-
lent forms,both voluntaryand enforced,are easily
conceived.Affirmative actionprograms can be ration-
alizedin thisway.
Anotherway to proceedis to exploitsubjective
uncertainty in theuse ofthetheoryofpersonalprob-
ability.An employer'sassessmentof the expected
productivity ofemployeesin a protected class maybe
wrongin a way that social experimentation could
effectivelydemonstrate, ifsuchexperimentation were
feasible,whichit rarelyis due to timelags and pro-
hibitiveexpense.The regulatormay arguethat an
employer'sproductivity assessmentbased on past
experience is invalidifit does notincorporate certain
mechanisms, suchas, forexample,theeffect ofhigher
rewardsthemselves on theproductivity of employees
in a disadvantaged class. If thereis an honestdiffer-
ence of personalprobabilities,each backed by an
explicitandcredibleanalysis,thenpoliticalforcesmay
permitthe regulatorto prevail,no doubtsometimes
correctly,and sometimeswrongly, as can be demon-
stratedonlyin the longrun.Effectiveregulationof
thissortobviouslyrequiresskillin causal modeling of
thesocialphenomenaconcerned. Again,I believethat
existingaffirmative actionagreements couldbe inter-
pretedas legitimateguessesat fairprocedures under
assumptionsaboutcausal processesnot yetformally
expressedas computations.
Finally,a briefcommenton technicalstatistics,
specificallyon implications of mymodelingstrategy
forstatisticalanalysis.An important pointis thatthe
principle(3) is in no waytiedto normalmodels,but
it doesimplythatmodelingneedsto be carriedouton
an additively meaningful rewardscale.Anotherpoint
is thatequation(4) is too simplisticformostapplica-
tionsin that the effectsof overtdiscrimination are
unlikelytobe equalacrossa protected class.Statistical
modelingthus needs to becomeincreasingsophisti-
catedaboutcopingwithnonnormal, including nonlog-
normal,distributions, and in allowingforinteraction
effects.These complicationscan only increasethe
burdens of expressingprior uncertaintiesabout
quantities not econometrically identifiedby the
data. Formaleducationof applied statisticiansand
econometriciansneeds reconstruction along lines
whichincludeuncertainty assessmentof gaps rou-
tinelyleftbystandardstatisticalanalyses.
ACKNOWLEDGMENTS
I thankthe Instituteof MathematicalStatisticsfor
theopportunityto presentthe 1986NeymanLecture,
onwhichthepaperis based.Comments byJohnCarlin
and StephenStigleron earlierdraftshave helpedto
improvethepresentation.The workwas facilitatedin
partbyNational.ScienceFoundationGrantDMS-85-
04332.
REFERENCES
BERNOULLI, J. (1713).ArsConjectandi.
BIRNBAUM, M. H. (1979).Procedures and correc-
forthedetection
tionof salaryinequities.In SalaryInequities(T. R. Pezzullo
andB. E. Brittingham, eds.).Lexington Books.
CONWAY, D. A. and ROBERTS, H. V. (1983). Reverseregression,
fairnessand employment discrimination.J. Bus. Econ.Statist.
1 75-85.
Cox, D. R. (1986).Commenton "Statisticsand causal inference,"
byP. W. Holland.J.Amer.Statist.Assoc.81 963-964.
DEGROOT, M. H., FIENBERG, S. E. and KADANE, J.B., eds. (1986).
and theLaw. Wiley,NewYork.
Statistics
DEMPSTER, A. P. (1968). A generalization of Bayesianinference
(withdiscussion).J. Roy.Statist.Soc. Ser. B 30 205-247.
DEMPSTER, A. P. (1984).Alternative modelsforinferring employ-
mentdiscrimination fromstatisticaldata. In W G. Cochran's
Impacton Statistics(P. S. R. S. Rao and J. Sedransk,eds.)
309-330.Wiley,NewYork.
DEMPSTER, A. P. (1987). Causality,uncertainty, and statistics.
Paper presentedat the Symposiumin Honorof I. J. Good,
Dept.Statistics,VirginiaPolytechnic InstituteandStateUniv.,
Blacksburg, Va.
FISHER, F. M. (1986).Statisticians,
econometricians, andadversary
proceedings.J.Amer.Statist.Assoc.81 277-286.
GOLDBERGER, A. S. (1984).Reverseregression and salarydiscrim-
ination.J. HumanResources19 293-318.
HACKING, I. (1975).TheEmergence ofProbability.Cambridge Univ.
Press,Cambridge.
HOLLAND, P. W. (1986a).Statistics and causalinference. J.Amer.
Statist.Assoc.81 945-960.
KRASKER, W. S. and PRATT,J.W. (1984).Bounding of
theeffects
 EMPLOYMENT DISCRIMINATION 161
proxyvariables on regressioncoefficients.Graduate School of SAVAGE, L. J. (1962). The Foundations of Statistical Inference.
Business Administration,Harvard Univ. Methuen,
London.
MEIER, P. (1986). Damned liars and expert witnesses. J. Amer. SAVAGE, L. J. (1967). Implications
ofpersonalprobability
forin-
Statist.Assoc. 81 269-276. duction. J. Phil. 64 593-607. Reprinted in The Writingsof
PHELPS, E. S. (1972). The statisticaltheoryof racism and sexism. Leonard JimmieSavage-A Memorial Selection,published by
Amer.Econ. Rev. 62 59-61. ASA and IMS.
PRArr, J. W. (1986). Review of "W. G. Cochran's Impact on SHAFER, G. (1976). A MathematicalTheoryofEvidence. Princeton
Statistics."J. Amer. Statist.Assoc. 81 565-566. Univ.Press,Princeton, N. J.
PRATT, J. W. and SCHLAIFER, R. (1984). On the natureand discov- SHAFER, G. (1987). Probability
judgmentin artificial
intelligence
eryof structure.J. Amer. Statist.Assoc. 79 9-21. and expertsystems. Statist.Sci. 2 3-16.

Comment
FranklinM. Fisher

ArthurDempster'spaper has a good deal to say estimatesthemodel

constant,a statistician
about the interpretation of probability modelsand
(1) Y= Ga + X: + e,
causal thinking, muchof it uncontroversial. Rather
than discusssuch mattersin the abstract,however, where (letting i denote values for a particular
let's considerthe exampleof employment discrimi- employee),Yi denotessalary,Giis 0 forfemaleand 1
nationthatDempsteruses and see whatit is thathe for male employees,Xi is a vector of observed
is reallysaying. employeecharacteristics age,
(education,experience,
This is not hard for me to do, because I have etc.),and the ei are assumed to be randomvariables
encountered Dempster'sviewson previousoccasions. (usually taken to be independentN(O, c2), although
I was a witnessforthe plaintiff in two employment thiswillplayno rolein thepresentpaper).a, f and o-
discriminationcases,OFCCP v.HarrisTrustandSav- areparameters to be estimated,and it willaid discus-
ingsBank (Department ofLaborCase No. 78-OFCCP- sionto assumethat the sample size large
is sufficiently
2) and CynthiaBaran v. The RegisterPublishing to enableus to take such parameters as known with
Company(Civil N. 75-272,U. S. Districtof Conn.). A
certainty. positivevalue ofa is taken to be evidence
In both cases, I testifiedon mattersof econometric ofdiscrimination againstfemales.
principleratherthanputtingforward
and statistical a What is wrongwithsuch a procedure?Dempster
studyof my own, and Dempstertestifiedfor the pointsout severalpossibilities.In the firstplace,he
defendant. This paperis largelybased on myexperi- suggestsinterpretingthe stochasticelementinvolved
enceand testimony in thosecases. (I believe-butdo by assumingthatthe nondiscriminatory employer is
not knowforsure-that,just as myown experience computing
in employment discrimination cases has been as an
expertassistingplaintiffs'counsel,Dempster'sex- (2) Y* = E(Y**I G,X*),
perience,to whichhe refers,has been as an expert whereX* is a vectorof employeecharacteristics
assistingcounselfordefendants.) knownto the employer(butpossiblynotto the ana-
A particularemployer is accusedofsex discrimina- lyst),Y** denotes"true"employeeproductivity and
tion. (As does Dempster,I take this as a leading YI denotes employee productivity as estimated by
example.)In general,thismeansthatsalariespaid to the employerin (2). Both Y** and Yi*are assumed
femaleemployees averagelessthanthosepaidto male measuredin monetaryunits to be comparableto
employees. One possiblereasonforthisdiscrepancy is wages,Yi. Discrimination as pay-
is to be interpreted
discrimination; anotheris that male employeesare ingmales more than Yr',i.e.,
moreproductive thanfemaleones.
To examinethe questionof whetherthereis a (3) Y = Ga' + Y*,
gender-basedwage difference holdingproductivity
witha' > 0.
This is not the onlyformthatdiscrimination can
ofEconomics,
FranklinM. Fisheris Professor Depart- take. Dependingon the state of the outsidelabor
mentofEconomics,Massachusetts Instituteof Tech- market,discrimination is morelikelyto consistof
Cambridge,
nology, Massachusetts 02139. payingfemalesless than the employertrulythinks