Feldman Zaller Article

A Simple Theory of the Survey Response: Answering Questions versus Revealing Preferences
Author(s): John Zaller and Stanley Feldman

Source: American Journal of Political Science, Vol. 36, No. 3 (Aug., 1992), pp. 579-616
Published by: Midwest Political Science Association
Stable URL: http://www.jstor.org/stable/2111583
Accessed: 27/10/2010 03:23
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=mpsa.
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.
Midwest Political Science Association is collaborating with JSTOR to digitize, preserve and extend access to
American Journal of Political Science.
http://www.jstor.org
A SimpleTheoryoftheSurveyResponse:Answering
QuestionsversusRevealingPreferences*
ofCalifornia,Los Angeles
JohnZaller,University
StanleyFeldman,StateUniversity
ofNewYorkat StonyBrook
Opinionresearchis besetbytwomajortypesof "artifactual"variance:hugeamountsof over-

timeresponseinstabilityand thecommontendency trivialchangesin questionnaire
forseemingly
formto affect We proposea simplemodelthatconverts
theexpressionof attitudes. thisanomalous
; errorvariance"intosourcesof substantive insightintothenatureof publicopinion.The model
abandonstheconventional butimplausiblenotionthatmostpeoplepossessopinionsat thelevelof
of typicalsurveyitems-and insteadassumesthatmostpeopleare internally
specificity conflicted
overmostpoliticalissues-and thatmostrespondto surveyquestionson thebasisof whatever ideas
areat thetopof theirheadsat themoment of answering.Numerous areshown
empiricalregularities
tobe consistentwiththeseassumptions.
Virtuallyall publicopinionresearchproceedson theassumption thatciti-

zens possessreasonably well formedattitudeson majorpoliticalissuesandthat
surveysarepassivemeasuresof theseattitudes. The standard viewis thatwhen
surveyrespondents saytheyfavorX theyaresimplydescribing a preexistingstate
offeelingfavorably towardX.
Accumulating evidenceon thevagariesofmasspoliticalattitudes, however,
has madethisviewincreasingly dubious.If,as is wellknown,peopleareasked
the same questionin a seriesof interviews, theirattitudereportsare highly
changeable.Many,as muchevidencealso shows,reactstrongly to thecontextin
whichquestionsare asked,to theorderin whichoptionsare presented, and to
whollynonsubstantive changesinquestionwording.Thesephenomena aremore
thanmethodological curiosities;theyraise seriousdoubtsabout whatpublic
opinionsurveysmeasure.
In view of this,we proposea new understanding of themass surveyre-
sponse.Most citizens,we argue,simplydo notpossesspreformed at
attitudes
thelevelof specificitydemandedin surveys.Rather,theycarryaroundin their
headsa mixof onlypartiallyconsistent ideas and considerations. Whenques-
preparedfordeliveryat thefifth
*Originally annualmeeting of thePoliticalMethodologySo-
ciety.The dataused in thepaperwerecollectedbytheBoardof Overseersof theNationalElection
Studiesundera grantfromtheNationalScienceFoundation.Collectingthesedatamadeunusually
heavydemandson thestaff forSocial Researchat theUniversity
of theInstitute ofMichigan,where
thestudywas conducted.In thisconnectionwe are grateful to ZoAnneBlackburn,StevePinney,
and SantaTraugott.We also thankBarbaraGeddes,ShantoIyengar,Don Kinder,MattLyons,and
Douglas Riversfortheirhelpfulcomments on variousdrafts
of thepaper.Finally,we wouldliketo
thankHiroakiMinatoforinvaluableassistancein preparing thedataforanalysis.Neither theseindi-
vidualsnorinstitutionsbearanyresponsibilityforanyerrorsofjudgment or factthatmayappearin
thispaper.
AmericanJournalofPoliticalScience,Vol. 36, No. 3, August1992,Pp. 579-616

? 1992bytheUniversityofTexasPress,P.O. Box 7819, Austin,TX 78713
580 JohnZallerand StanleyFeldman
tioned,theycall to minda sampleof theseideas, includingan oversampleof

ideas madesalientbythequestionnaire andotherrecentevents,anduse themto
chooseamongtheoptionsoffered. Buttheirchoicesdo not,inmostcases,reflect
anything thatcan be describedas trueattitudes; thethoughts
theyreflect
rather,
thataremostaccessiblein memory atthemoment ofresponse.
A modelbasedon theseclaimscan, as we show,providea farbetteraccount
of theexistingevidenceon politicalattitudes,
includingsuchrelatedmattersas
attitudeconsistency and theeffectsof politicalawareness,thancan currently
dominant modelsofthesurveyresponse.
We beginwitha reviewof existingmodelsofpublicopinionandthenout-
modelanddrawupona rangeofnewandexistingevidenceto
linean alternative
demonstrate itsvalue.
LimitsofExistingTheories
Responseinstability
One ofthemostunsettling ofopinionresearch
findings hasbeenthediscov-
eryofa largecomponent ofrandomness inmostpeople'sanswerstosurveyques-
tions.Ifthesamepeopleareaskedthesamequestioninrepeatedinterviews, only
abouthalfgivethesameanswers.The datainTable1, basedon interviews ofthe
samepersonssix monthsapart,illustrate theproblem.As can be seenfromthe
entrieson the maindiagonals,only45% to 55% gave the same answerboth
times,eventhoughabout30% couldhavedoneso bychancealone.' The amount
differs
ofresponseinstability fromone issueto another(see Feldman1989),but
thecases showninTable 1 arefairlytypical.
In his famouspaper"The Natureof BeliefSystemsinMass Publics,"Con-
verse(1964) arguedthatresponseinstability is due mainlyto individualswho
lack meaningfulattitudesbut nevertheless indulgeinterviewers by politely
choosingbetweentheresponseoptionsputin front ofthem-butchoosingin an
almostrandomfashion."Largeportions of an electorate,"
he suggested,"simply
do not have meaningful beliefs,even on issues thathave formedthe basis
forintensepoliticalcontroversy amongelitesforsubstantial periodsof time"
(1964, 245).
Thisconclusionhas beenstrongly challengedbyscholarswhocontendthat,
althoughpeople's"surveyresponses"fluctuate citizenshaveunderlying
greatly,
"trueattitudes"thatare overwhelmingly stable(Achen 1975, 1983; Dean and
Moran 1977; Erikson1979; Feldman1989). The fluctuations thatappearin
people's overtsurveyresponsesare attributed to "measurement error,"where
'Given data fromonlytwo pointsin time,it is impossibleto distinguish

systematicattitude
changefromrandomfluctuation.However,analysisofdatafromthree- andfive-wavepanelsstrongly
suggeststhatalmost all responseinstability
represents randomratherthan systematicchange
(Feldman1989).
A THEORY OF SURVEY RESPONSE 581
Table 1. ResponseStabilityoverRepeatedInterviews:
TWoExamples
American Relations withRussia

(Corner Percentaging)
Attitudesin January1980
Cooperate Middle Tougher Unsure
June 1980:
Cooperate 25% 8 8 2
Middle 7 4 5 1
Tougher 5 5 17 2
Unsure 4 2 3 4
N = (338) (153) (266) (74)
Level of GovernmentServices
(Corner Percentaging)
Attitudesin January1980
Cut Middle Keep Same Unsure
June 1980:
Cut 24% 6 5 6
Middle 8 4 2 2
Keep Same 5 4 15 3
Unsure 8 2 3 6
N = (362) (122) (208) (138)
Note: The exactquestionswere:"Some people feelit is important forus to try

veryhardto getalongwithRussia. Othersfeelit is a big mistaketo trytoo hard
togetalongwithRussia.Wherewouldyouplaceyourself on thisscale,orhaven't
youthought aboutthis?"Respondents werethenaskedto place themselves on a
seven-point scale. In thistable,points1, 2, and 3 havebeencountedas "cooper-
ate";4 is countedas middle;5, 6, and7 havebeencountedas "tougher."
The secondquestionwas: "Somepeoplethinkthegovernment shouldprovide
fewerservices,even in areas such as healthand education,in orderto reduce
spending.Otherpeople feel it is important forthegovernment to continuethe
servicesitnowprovidesevenifit meansno reduction in spending.Wherewould
you place yourselfon thisscale, or haven'tyou thought aboutthis?"Itemwas
recodedas above.
Source:NationalElectionStudies,1980PanelSurvey.
sucherroris said to stemfromtheinherent of mappingone's attitudes

difficulty
ontotheunavoidably vaguelanguageofsurveyquestions.
Both approachesto responseinstability have criticaldeficiencies.Con-
verse'sthesis,whichtakesanyinstabilityas evidenceofa "nonattitude,"
was an
extremeclaimintendedto characterize attitudesonlyon highlyabstractissues.
On moretypicalissues,as ConverseandMarkus(1979) argue,people'sattitudes

maybe moreor less"crystallized" andare,as a resultofthis,moreorlessstable.
But thisonlyraisesthequestion how crystallization
of can be measuredapart
fromitssupposed on
effect response Since
stability. no one haseversaid,attitude
remains,
crystallization as Krosnick and Schuman (1988) have pointedout,more
a metaphor thana testabletheoryof attitude
stability.
The newer"measurement error"theoryof responseinstability appears
equallyunderspecified core. When,as all estimates
at itstheoretical agree,mea-
surement "error"typically one-half
constitutes or moreofthevarianceoftypical
attitudeitems,one naturally wonderswhatexactlythis"error"consistsof and
howithas beengenerated. Yetwe presently knowso littleaboutthesequestions
thatthetermremainsessentially an alternativenamefor"unexplained variance."
ResponseEffects
In additiontotherandomresponsevariancethathasbeenattributed to mea-
surement error,thereexistssystematic variancefromartifactual "responseef-
fects."Considera well-known Cold Warexperiment on attitudestowardSoviet
journalists.In a split-halfsample,37% of respondents werewillingto allow
communist reportersin theUnitedStates.Yetwhen,in theotherhalf-sample,
respondents werefirst askedwhether U.S. reporters shouldbe allowedin Russia
(whichmostfavored),thepercentage agreeingto allow Russianreporters here
doubledto 73%.
Thereare numerousotherfindings of thistype:people are less likelyto
describethemselves inpoliticsjustaftertheyhavebeenaskedabout
as interested
obscureissues(Bishop,Oldendick,andTuchfarber 1984); people'sattitudesto-
wardabortionare affected by thekindsof items(concerning, e.g., religionor
women'srights)thatprecedeit (Tourangeauand Rasinski1988; Tourangeau et
al. 1989); people give quite different answersto open-endedquestionsthan
to questionsthatask themto choose amonga .eriesof prespecified options
(Schumanand Scott 1987). Seeminglyirrelevant features of surveydesigndo
not,as mightbe suspected,affectonlyunsophisticated people who mightbe
suspectedofhavingnonattitudes; theyensnareall typesofrespondents (Krosnick
andSchuman1988;see also Bishop1990).
The literatureon responseeffects thusmakesit clearthatsurveyquestions
do notsimplymeasurepublicopinion.Theyalso shapeand channelit by the
mannerin whichtheyframeissues,orderthealternatives, andotherwise setthe
context ofthequestion.Thishas led researchersto a conclusion thatseems indis-
putable but thatis fundamentally at odds with the assumptions of most political
scientistsaboutthenatureof politicalattitudes: namely,peopledo notmerely
revealpreexistingattitudesonsurveys;to someconsiderableextent, peopleareusing
thequestionnaire to decide whattheir"attitudes" are (Bishop,Oldendick,and
Tuchfarber 1984;Zaller1984;Feldman1990).
Psychologists makethisargument mostdirectly. Tourangeauand Rasinski
(1988), forexample,arguethatresponsestoattitude questionscan be understood

as theoutcomeofa question-answering processin whichpeople(1) decidewhat
theissueis; (2) canvastheirmindsforrelevant thoughts;(3) combineideas into
a coherentattitude;and (4) map theresulting attitudeontoavailableresponse
options.Because, as theymaintain, features of theinterview processcan affect
each of thesesteps,thequestionnaire can also readilyaffectwhatgetsreported
as publicopinion.
Morerecently, Wilsonand Hodges(1991) haveproposeda modelin which
attitudesare"temporary constructs" thataremadeup at themoment ofresponse
on thebasis of ideas in a largebutinternally conflicted"database." In perhaps
the mostambitiousattempt to deal withthe question-answering process(and
muchelse), WyerandSrull(1989) offer an information-processingmodelhaving
morethanthreedozenelements.
WhatNeedstoBe Done
Despitetheevidencefrompsychologists andsurveymethodologists, public
opinionresearchers largelyignoreboththe longstanding problemof massive
over-time responseinstability and thenewerfindings on questionnaire effects.
Moreover, manyofthosewhorecognizetheproblems makewhatamountessen-
tiallyto patch-upsof thetraditional view.In thecase of responseeffects, the
patch-upconsistsof trying to prevent theproblemfrombecomingconspicuous;
thisis doneby,forexample,makingsureto keepquestionorderconstant across
timeseriessurveysor,in somecomputer-assisted surveys, randomizing question
orderacrossrespondents. In thecase ofresponseinstability, thepatch-up consists
of statistical
corrections
formeasurement error,corrections thatcreatethecom-
fortableillusionthatfixed"trueattitudes" existbeneaththeenormoussurface
noise.
The challenge,then,is to devisea theory thataccommodates bothresponse
instabilityand responseeffects and thatis craftedto thekindsof problemsand
data facinganalystsof publicopinion.This is whatwe attempt to do in this
paper.The theorywe proposeis, we admit,simplerthanwouldbe necessaryto
explainall ofthefindings thatpsychologists havenowdocumented. Buta theory
sufficientlycomplextodo thiswouldhavelittlevaluetomostpoliticalscientists,
and our aim is, above all, to reachthisgroup-convincingit thattheconven-
tionalunderstanding of publicopinionis unworkable and thata practicalalter-
nativeis available.
An AlternativeModel oftheSurveyResponse
Accordingto conventionalattitudetheory,individualschoose whichever
optioncomesclosestto theirownposition.But if,as we contend,
prespecified
do nothavefixedpositionson issues,howdo theymaketheir
peopletypically
choices?
Since mostsurveyresearchtakesthe"trueattitude"as itsprimitive
unitof
analysis,littleattention has beendevotedtothisquestion.If,however, oneturns

to studiesthatemploydepthinterviews, one findsmuchusefulevidence.Among
thebestoftheseis Hochschild's(1981) studyofattitudes towardequality, What's
Fair?Fromherinterviews with28 persons,Hochschildfoundthatpeoplewould,
ifaskedto do so, readilyanswerfixed-choice questions,butthatgiventheop-
portunity to talk,"peopledo notmakesimplestatements; theyshade,modulate,
deny,retract, orjustgrindto a haltin frustration. Thesemanifestations ofuncer-
taintyarejustas meaningful and interesting as thedefinitive statementsof a be-
liefsystem"(238).
Hochschildparticularly emphasizestheambivalence ofmanyofherrespon-
dents.Thisambivalence frequently leadsthemtocontradict themselves-which
is to say,to givetemporally unstableresponsesin thecourseofa singleconver-
sation.Considerthisaccountof theattitudes of one ofhersubjectstowardgov-
ernment incomeguarantees: "VincentSartoricannotdecidewhether or notthe
government shouldguaranteeincomes,because he cannotdecide how much
weightto giveto thevalueof productivity. He believesthattherichare mostly
undeserving and. . . yethe is angryat 'welfarecheats'whorefusetowork....
Caughtbetweenhis desireforequalityand his knowledgeof existinginjustice,
on theone hand,andhisfearthata guaranteed incomewillbenefit evenshirkers,
on theother,he remainsambivalent aboutpoliciestowardthepoor"(252).
Current attitudemodelsseemquiteirrelevant totheseobservations. Therea-
sonforSartori'svacillation is not,as students ofConversemightsay,thathe has
no opinionon thisquestion,noris itthat,as usersofmeasurement errormodels
mightsay,Sartorihas a "trueattitude" thatHochschildis unableto measurereli-
ably It is rather thatSartorihas conflicting opinions,or at leastconflictingcon-
siderations, that lead him to givedifferentresponsesatdifferent times,depending
on howhe thinksabouttheissue.
It is easy to objectto thelimitations on rigorinherent in depthinterviews
suchas Hochschild's.Nonetheless, we are persuadedthatthebasic pointabout
ambivalence-thatindividualspossess multipleand oftenconflicting opinions
towardimportant issues-representsan important insight.Muchpsychological
research reinforces thisview.Memoryresearchers, forexample,haveshownthat
peoplestorehugeamountsof information in theirlong-term memories, butcan
retrieve and use onlya fraction of itat one time.The particular materialtheydo
recalldependson a combination of chanceand recencyof activation.Hence,
peoplemakequitedifferent judgmentsand beliefstatements, depending on the
information theyhappento recall fromlong-term memory(Raaijmakersand
Shiffren 1981;WyerandHartwick1984).
Another researchtradition, mainlyconcerned withsocialcognition, focuses
on theorganization of ideas in themind.A centralconceptin muchof thisre-
searchis the"schema,"a termthathas beenadaptedfromcognitive psychology.
A schemais a cognitive structurethatorganizespriorinformation andexperience
arounda centralvalueor idea andthatguidestheinterpretation ofnewinforma-

tionandexperience.
A criticalpointaboutschemasis thatpeopletypically haveseveralofthem
availableforunderstanding any givenphenomena.For example,an individual
beingintroduced to a "professor" wouldreactquitedifferently ifthenewperson
were insteaddescribedas "a motherof four."That is, different associations
wouldcometo mind,different qualitiesofthepersonwouldbe noticed,different
conclusionswould be drawnfromtheperson'smannerisms, and so forth.In
short,theperceiver's attitudetowardthepersonwouldbe different. Thus,Tesser
(1978), in statements thatrepresentthedominant thrustof muchcognitivepsy-
chologyandthatnicelycapturea centralfeature ofthemodelwe propose,writes:
"An attitudeat a particular pointin timeis the resultof a constructive pro-
cess. . . . And, thereis nota singleattitudetowardan objectbut,rather, any
numberof attitudes dependingon thenumberof schemasavailableforthinking
abouttheobjects"(297-98). And"personsdo nothavea singlefeelingorevalu-
ationofan object.Feelingsvarydepending upontheparticular cognitive schema
we 'tunein'" (307).
These studiestendbothto corroborate Hochschild'sinsightsconcerning
ambivalenceand to undermine the conventional politicalscienceassumption
(whichis at theheartof bothConverse'sblack-and-white modeland Achen's
"trueattitude" measurement errormodel)thatitis normalforindividuals tohave
a single, coherent attitude on issues. In view of this, our model will follow
in that
Hochschild assuming peoplecarry around in their
heads a mix ofmore or
less consistent "considerations," where a considerationis defined as a reasonfor
favoringone side of an issue ratherthananother.(E.g., a personwho thinks
about"Pentagonwaste"whiledecidinga questionaboutdefensespendinghas
raiseda consideration thatmaywellcontrolherdecisionon thatissue.)
The first axiomofourmodelmaynowbe statedas:
AXIoM1: The ambivalenceaxiom.Mostpeoplepossessopposingconsid-

thatmightlead themto de-
erationson mostissues,thatis, considerations
cidetheissueeitherway.
We emphasizethattheconceptof consideration, as used in thisaxiom,is

notjust anotherwordforschema.First,it is cast in thelanguageof everyday
politicaldiscourse(see Kelley1983),as befitsa termintended forpoliticalrather
thanpsychological analysis.Second,it makesno reference to mentalstructures
or operations,suchas theinterpretationof rawsensoryinput,thatarecentralto
theconceptofschema.
Ournextproblemis todecidehowindividuals transformthediverseconsid-
erationsin theirheads intoclosed-endedresponses.One possibility, as Taylor
and Fiske (1978) suggest,is thatindividualsmakechoices"offthetop of the
head"on thebasis of thefirst

idea thatcomesto mind.Thus,peoplemaymake
socialjudgmentsby seizingon "a single,sufficient
and salientexplanation
...
oftenthe firstsatisfactoryone thatcomes along. . . . [I]nsteadof employingbase
rateor consensusinformation
logically,peopleare moreofteninfluenced
by a
single, colorfulpiece of case historyevidence. . . . Instead of reviewingall the
evidencethatbearsupona particular problem,peoplefrequently use theinfor-
mationwhichis mostsalientor availableto them,thatis, thatwhichis most
easily broughtto mind"(251). Tverskyand Kahneman's(1982) well-known
workon framing effectsreinforcestheviewthatindividuals oftenareoverlyin-
fluencedbya single,dominant consideration.
Atthesametime,muchdatainbothpoliticalscience(Campbelletal. 1960)
andcognitivepsychology (Anderson1974) indicatethaton otheroccasions,in-
dividualsreachdecisionsbyaveraging acrossa rangeofcompeting ideas. Thus,
Kelley(1983) showsthatvotersseemto decidewhichpresidential candidateto
supportby summing up all of their"likes"and "dislikes"abouteach partyand
candidateandchoosingtheone withthehighest
presidential nettotal.
The axiomswe proposeallowindividuals to respondto surveyquestionson
thebasis ofeitherone or manyconsiderations, depending on howmanyhappen
tobe readilyaccessiblein memory atthemoment thequestionis posed:
AxIoM2: The responseaxiom.Individualsanswersurveyquestionsbyav-
eragingacrosstheconsiderationsthathappento be salientatthemoment
of
response,wheresaliencyis determinedbytheaccessibility axiom.
AxIoM3: The accessibilityaxiom.The accessibility
of anygivenconsider-
ationdependson a stochasticsamplingprocess,whereconsiderationsthat
havebeenrecentlythought aboutaresomewhat morelikelytobe sampled.2
For thecase in whicha persondevotesgreatthought and attentionto an issue,
Axiom3 impliesthattheremaybe multiple considerations
salientin memory at
themomentof answering questionsabouttheissue and hencemanyconsidera-
tionsto be averagedacross.But a personwho rarelythinksaboutan issue and
who is confrontedby an interview situationthatrequiresa successionof quick
answers(Feldman1990)mayhaveonlyoneconsideration immediately available
inmemory, in whichcase theaveraging rulereducesto answering on thebasisof
a single"top-of-the-head"
consideration,as suggestedbyTaylorandFiske.
Thesethreeaxioms,although spareandinformal, can be usedbothtoorga-
nize muchexistingresearchand to generatetestablenewhypotheses aboutthe
natureofthemasssurveyresponse,as we shallnowbeginto show.
2Theempiricalwarrant
forthisaxiomis extremely
strong(see Barghet al. 1986;Higginsand
King 1981;BodenhausenandWyer1987).
Data
Sincewe base muchofouranalysison datafromthe1987PilotStudyofthe
NationalElectionStudies(NES), it is worthpausingbriefly to describethis
study.The surveyattempted to measure,interalia, the"considerations" thatun-
derliepeople'sresponsesto standardclosed-ended surveyitems.The studywas
conducted in twowavesa monthapart;457 personswereinterviewed in theMay
waveand 360 in theJunewave. All hadpreviously participatedin the1986Na-
tionalElectionStudy.Othertechnicaldetailsof thestudyare availablethrough
theNES attheUniversity ofMichigan.
The basicmethodwas to ask peoplea closed-ended policyitemandthento
ask themto talkin theirownwordsabouttheissuesitraised.The closed-ended
itemsweretelephoneversionsofthestandard NES itemsonjob guarantees, aid
to blacks,and government servicesand spending.In formA, respondents were
asked the open-endedprobesimmediately afteransweringthe givenclosed-
endedpolicyitem.Theexactformofthe"retrospective" open-ended probeswas:
aboutthequestionyoujust answered,I'd likeyouto tellme

Stillthinking
whatideascametomindas youwereanswering thatquestion.Exactlywhat
thingswentthroughyourmind.(Anyothers?)
This probewas designedto elicita "memorydump"of theconsiderations im-

mediatelysalientin people'sminds.Priorworkby Ericssonand Simon(1984)
showsthatsuchprobescan workeffectively if askedimmediatelyaftera given
taskhas beencarriedout.
In formB, interviewersreadtheitemsintheusualway,but,without waiting
foran answer,theyaskedrespondents togivetheirreactions
totheprincipalidea
elementsin thequestion.For thejob guarantees question,theprobeswereas
follows:
Beforetellingme howyoufeelaboutthis,couldyoutellme whatkindsof

thingscome to mindwhenyou thinkaboutgovernment makingsurethat
everypersonhas a goodstandardofliving?(Anyothers?)
Now,whatcomes to mindwhenyou thinkaboutlettingeach personget
ahead on theirown?(Anyothers?)
Immediately followingtheseprobes,the interviewer rereadthe original

questionandtooktheperson'sreplyto it. (Fullquestionwording
closed-ended is
intheappendix.)
The two typesof probesare clearlynot equivalent.The "retrospective"
probes,whichwereposedafterpeoplehad answeredthequestionin thenormal
way,weredesignedtofindoutwhatexactlywas on people'smindsatthemoment
of response.The "prospective" or "stop-and-think"3 probes,on theotherhand,

weredesignedto inducepeople to searchtheirmemoriesmorecarefully than
theyordinarily wouldforpertinent considerations. Notethatthestop-and-think
probesdo notraisenew ideas or pushtherespondent in a particulardirection;
theysimplyrequiretherespondent to say explicitlywhatmeaninghe or she at-
tachestothedefining phrasesofthequestion.
Respondents wererandomlyassignedto questionformand answeredthe
same typeof thequestionin each wave of thestudy.The threetestitemsand
associatedopen-ended probesappearedneartheendofeach waveofthesurvey.
Interviewers wrotedownas faithfully as possibleall responsestotheopen-ended
probes,includingincidental side comments (e.g., "This is a toughone"). The
transcribed comments weresubjectedto an elaborateclassification scheme,with
as manyas fourcomments codedforeach probe.4Respondents on thestop-and-
thinkside averagedabout3.7 codablecomments perpolicyitem,withalmostall
respondents offeringatleastonecodablecomment. Theaverageon theretrospec-
tivesidewas 2.9.
All comments,includingside comments, wereratedon severalvariables
by staffcodersat theInstitute forSocial Researchat theUniversity of Michi-
gan. Becausethecodingprojectwas considered a difficultone,onlyexperienced
coderswereused. The mostimportant variablewas "directional of com-
thrust
ment,"whichindicatedwhichside of theissue,ifany,theremarkfavored.Al-
thoughthisvariablenotedambivalence, confusion, andnonissueconcerns,75%
of comments had a cleardirectional thrust. The otherkeycodingclassification
was "frameof reference," a variablethatincludedmorethan140 categoriesand
triedto capturethesubstantive content ofeachremark. The framecodesreferred
to generalprinciples (e.g., equality, theroleof government), individualismand
the workethic,the fairnessof theeconomicsystem,particular groups(e.g.,
blacks,theelderly),personalexperience,and particular government programs.
(Further information aboutthesecodes maybe foundin Table5 and associated
text;all codesarefullydescribedinthestudycodebook.)
Thesedataarenotwithout limitations. The mostobviousis coderreliability.
Abouta tenthofall interviews weredouble-coded, and,although exactreliability
dataarenotavailable,thecodingsupervisor reported a "difference" betweencod-
erson 10% to 15% of all cases. Thisdifference ratewas regardedas normalfor
material of thistype,5butitis higherthanone wouldhopefor.In addition,10%
of remarks wereso unclearthattheycouldnotbe assigneda directional thrust.
A finallimitation is thedifficulty in confidently distinguishing one "considera-
3Thisaptdesignationis theinvention
ofKathleenKnight.
4Fortheaid to blacksitem,therewereup to six probes:threequestions,each followedby a
queryfor"anyothers?"As manyas fourremarks werecoded in connection witheach of thesesix
probes.On theotheritems,thereweretwoinitialprobes,eachwithfollow-up probes.
5Personalcommunication fromStevePinney,whosupervisedourprojectat ISR.
tion"fromanother.Whentworemarks haveclearlyopposingthrusts,thisis no
problem.Butpeoplesometimes offera seriesof remarks
on thesamesideofan
issue.Do suchremarks represent orjustelaborations
separateconsiderations, on
a singleidea? Even a personlistening to theinterviews
as theyoccurred,as we
did,wouldsometimes be uncertain;codersworking froman imperfecttranscript
wouldexperience greateruncertainty.
Even in retrospect,we are notsurehow theseproblemscould have been
ameliorated
significantly in thecontextof a masssurvey. We thusfeelthePilot
Studydataarethebestthatarereadilyattainable fordirectly
examiningtheidea-
tionalunderpinnings of massattitudes, butwe admitthattheyare,nonetheless,
farfromperfect.
TestsoftheModel
Axiom
CheckoftheAmbivalence
Preliminary
We beginassessmentof themodelby makinga plausibility checkof the
holdmultiple,
axiomthatclaimsthatmostindividuals ideason most
conflicting
issues.Ourdatagiveus threewaysto measuretheextentof ambivalencein the
public,as follows:
1. A countofthenumberofopposingremarks byeachpersonthatcan be
pairedagainstone anotherIf,forexample,a respondent makestwocomments
witha liberalthrustand twowitha conservative hisscoreon theconflict
thrust,
scale is two.Ifhe makesthree(ormore)on one sideoftheissueandonlytwoon
theother,theconflictscoreis stilltwo,sincethenumber ofopposingcomments
thatcan be pairedremainstwo.
2. A countofthetimespeople spontaneously expressambivalenceor diffi-
cultyin makingup theirminds.A specialcode was createdto capturesuchre-
marks;itreadsas follows:"Mentionindicatesambivalence,conflict (e.g., 'I see
meritin both sides'; 'That's a toughquestion'; 'Depends'; 'Both are valid
points')."
3. A countofthenumberoftimespeople make"two-sided comments." In-
cludedin theframeof reference codes are special"starcodes" thatindicatea
thrust
directional tothecomment butalso ambivalencewithrespecttothatdirec-
tion.Starcodes applyto cases in whichrespondents had a preferencebutwere
clearlypayingsomeattention totheothersideoftheissue.Instructions tocoders
foruse ofstarcodesreadas follows:
A star code is used onlyforcases in whichthereis a singlethoughtor
comment thatencompassestwoopposingelements(e.g., "Although I think
favorY" Starcodesareusedforcomments
X, I nevertheless inwhichR sees
twosidesto an issue.
Examplesof starcodes are "people shouldtryto get ahead on theirown, but

government shouldhelp whennecessary"and respondent "admitsproblem(s)
withanyprogramor typeof program, butinsistsit is worthwhile anyway."A
countof thestar-coded remarksmaythusbe considereda measureof ambiva-
lence.
Fromthesethreemeasureswe createda fourth: a summary oftheindiceson
whicha personscored + 1 or higher.Because conflictand ambivalenceare
equallyconsequential whether theyoccurwithinthecourseof one interview or
acrossseparateinterviews, all indicesare calculatedacrossbothwavesof the
survey.
Frequencydistributions on thesefourmeasuresare shownin Table 2. As
can be seen,each measurecapturessubstantial amountsofambivalence, a result
thatis consistentwiththefirst axiom.Evenon themoreconservative evidenceof
theretrospective probes,whichinvolveonlyone queryin each wave,6thesum-
marymeasureindicatesthat36% to 48% ofrespondents areto somedegreeam-
bivalenton thesethreeissues. And thisis surelyan understatement. Whatthe
retrospectiveprobescapture,as explained,is thereasonthepersonhas answered
theitemas he justhas; theycannotcaptureanything likethefullrangeof ideas
in theperson'shead. However,theprospective probesweredesignedto tap a
widerrangeoftheideasinpeople'sminds;on evidencefromthem,roughly 75%
ofrespondents areat leastsomewhat conflictedon the three issues.
These resultsprovideclear initialsupportforthemodel'sfirstaxiom,the
ambivalence axiom.Sincetheothertwoaxiomscannotbe testeddirectly, we turn
nowto an examination of thedeductiveimplications of thethreeaxiomstaken
all together.
fromtheModel
FirstDeductions
We beginwithdeductions fromthemodelthatareentirely straightforward
andperhapsuninteresting andproceedto moreusefulandimportant ones. If,as
axiomclaims,theaccessibility
theaccessibility ofa givenconsideration depends
on theamountof thought devotedto an issue,we shouldfindthatpeoplewho
are,in general,morepoliticallyinvolvedhavemoreconsiderations at thetopof
theirheadsand availableforuse in answeringsurveyquestions.This is thefirst
of 18 deductionsfromthemodelthatwe makeandtest(Deduction1). (To keep
each willbe numbered
trackof thedeductions, in parentheses,
as here.)Despite
someindication ofnonmonotonicity inthedata,Table3 essentially
confirms this
expectation.7
we wouldexpectpersonswhohavegreater
Similarly, in an issueto
interest
6Moreprobesweremade,butsincetheyweredirective, we makeno use ofthem.

7Thenonmonotonicity versionofourmodelleadsone
is probablyreal,as a moresophisticated
toexpect(see Zaller,inpress,chap. 8).
Table 2. Expressionsof Conflictand Ambivalenceon PoliticalIssues
Probes
Retrospective Stop-and-ThinkProbes
Considerations
Conflicting Considerations
Conflicting
Jobs Services Aid to Blacks Jobs Services Aid to Blacks
Count
0 73.9%a 57.8 73.4 36.9% 30.7 29.0
1 22.6 33.6 22.6 27.3 29.0 21.6
2 3.5 5.2 4.0 22.2 21.6 25.0
3+ 0.0 3.4 0.0 13.1 18.7 24.4
of Ambivalence
Expressions ofAmbivalence
Expressions
0 76.9 83.5 78.8 63.2 72.2 71.1

1+ 23.1 15.5 27.8 36.8 27.8 28.9
Two-SidedRemarks(StarCodes) Two-SidedRemarks(StarCodes)
0 75.0 91.7 81.4 64.9 85.2 72.3

1+ 25.0 8.3 18.6 35.1 14.8 27.7
TotalIndications TotalIndications
0 60.2 51.4 63.6 25.9 24.3 24.7

1 15.7 37.6 16.9 39.1 50.9 44.0
2 18.5 7.3 12.7 9.8 14.8 12.0
3 5.6 3.7 6.8 25.3 10.1 19.3
Note:aMeasuresaredescribedin text.
Source: 1987NES PilotStudy.
Table 3. Effectof PoliticalAwarenesson Volumeof Open-EndedComments
Level ofPoliticalAwareness
Low Medium High
Standardof living 2.3 2.9 3.5 3.5 3.2

Government services 2.3 3.2 4.3 4.6 4.7
Aid toblacks 3.3 3.9 4.4 5.2 4.4
N= 44 53 54 38 46
remarksin thegivencell. These

Note:Cell entriesare theaveragenumberof discretesubstantive
data are derivedfromstop-and-think probesshowthesame
probes,butresultsfromretrospective
pattern.The awarenessmeasureis describedin theappendix.
Source:1987NES PilotStudy.
have,all else equal, morethoughts aboutthatissuereadilyaccessiblein memory

thanotherpersons(Deduction2). Since thePilotStudydid notdirectly ask re-
spondentshow important or interesting each of thepolicyissues was to them,
ourabilityto testthisexpectation is limited.We did, however, findthatblacks
raisedmoreconsiderations thanwhitesin connection withtheaid to minorities
item(p < .01). Government employeesalso hadslightly moreto saythanother
personsaboutthegovernment servicesitem(p = .07). Unemployed persons,
however,were not more likelyto raise considerations pertinent to the job
guarantees item.
Axioms2 and 3 claimpeopleanswersurveyquestionsby averaging across
whatever considerations aresalientinmemoryIfthisis so, we shouldfindstrong
correlationsbetweenmeasuresofpeople'sthoughts as theyanswera surveyitem
and thedirection of decisionon theitemitself(Deduction3). Thus if,forex-
ample,a personmakestworemarks thatfavortheliberalside of theissue and
one thatfavorstheconservative side,we wouldexpectthatthepersonwould,on
average,takethe liberalside of the issue. Althoughthisinference may seem
hardlyworthtesting,it is by no meansobviousthatit can be confirmed. Social
psychologists, workingin thedomainof social cognition, haveturnedup cases
in whichthedirection of people'sopen-ended thoughts is uncorrelated, or even
negatively correlated,withevaluations ofthegivenissue.
As Hastieand Park(1986) havecontendedin an influential essay,thesur-
prisingnoncorrelations occurbecausepeopletypically do notconstruct attitude
statements fromideas theycan retrieve frommemoryas theyare questioned.
Instead,peoplerecallattitudes formedat an earliertime.Thus,theymaintain,
thereis no necessarycorrelation betweentop-of-the-head ideasandattitude state-
ments(see Lodge,McGraw,andStroh1989).
Notwithstanding this,ourdataindicatesubstantial correlations betweenthe
ideas mostaccessibleto individuals at themoment ofresponseandtheresponse
given.To showthis,we createdadditiveindicesofpeople'sopen-ended remarks,
coded fordirectional thrust.We thencorrelated theseindiceswithresponsesto
theclosed-ended items,as shownin Table4. On thestop-and-think side,corre-
lationsbetweentheindicesand theirassociateddichotomous itemin each wave
ofthesurveyaveragedabout.40. Whenan indexofall remarks overbothwaves
of thesurveyis correlated witha scale thatconsistsof responsesto theclosed-
endeditemsfromthetwowaves,thecorrelations averageabout.50. In theother
halfofthestudy, thecorrelations betweenindividuals'retrospective remarks and
theirclosed-ended responsesin thesamewaveofthesurveyaveraged.70. When
retrospectiveremarks anditemsweresummedandcorrelated acrosswavesofthe
survey,the correlations averaged.80. (Much of thedifference in correlations
betweenthestop-and-think side and theretrospective side appearsto be due to
errorin thecodingof theclosed-endedresponses,which,as discussedbelow,
seemsto have been higheron thestop-and-think side.) Giventhattheclosed-
Table 4. RelationshipbetweenAvailableThoughtsand Closed-EndedItems
Wave 1 Wave2 Combined

Correlations
withRemarksMade Justbefore
Answering Closed-EndedQuestion
Jobguarantees .39 .39 .50
(212) (161) (173)
Government
services .31 .36 .41
(187) (153) (165)
Aid to blacks .57 .48 .63
(220) (165) (166)
Correlations
withRemarksMade Justafter
Answering Closed-EndedQuestion
Jobguarantees .79 .70 .79
(126) (123) (105)
services
Government .79 .70 .78
(137) (105) (106)
Aid to blacks .67 .83 .83
(144) (114) (112)
Note:Cell entryis Pearsoncorrelation

betweenopen-endedremarksmadejustbefore(orjustafter)
answeringcloseditemandscoreson thecloseditem.Firstcolumnshowscorrelations
fromfirstwave
of survey;secondcolumnshowscorrelations fromsecondwave; thirdcolumnshowscorrelations
betweenremarks frombothwavesandcombineditemscoreson bothwaves.
endeditemsin thesetestsareessentially dichotomies,8andthefragilenatureof

ouropen-ended data,thesearesizablecorrelations.9
We shouldadd thatthe findings of correlationsbetweentop-of-the-head
thoughts and attitudereportsare not,takenalone, clinchingevidencethatthe
former havecausedthelatter.Rather,thecorrelations-which, as we haveindi-
cated,couldnotbe takenforgranted in lightof pastresearch-simplyrepresent
one of morethana dozencases in whichwe havebeenable to developevidence
thatis consistent
withthethree-axiom model.
8A fewrespondents volunteered "itdepends"middleresponses,whichwereaccepted.

9The substantialmagnitudes of thesecorrelations do not,however,show thatclosed-ended
surveyresponsesare,afterall, perfectly validindicatorsof theconsiderationsin people'sheads.As
we noteimmediately below,differentconsiderations are salientat different
times,leadingpeopleto
makedifferentattitudereportsat differenttimes.WhatTable4 showsis onlythatclosed-ended atti-
tudereportsare reliableindicatorsof theconsiderations thatare salientat themoment ofmakinga
response.
ExplainingResponseInstability
Responseinstability overrepeatedinterviews is, as we haveindicated, one
of themostimportant and disturbingempiricalregularities associatedwiththe
mass surveyresponse.In thissection,we attempt to use ourmodelto explain
We beginwithsimpleillustrations
thisinstability. of ourapproachandthenpro-
ceedto moresystematic analysis.
Whenasked in the May interview abouttheproperlevel of government
services,one respondent, identifiedas a teacher,emphatically favoredhigher
levelsof servicesand spending.The country was facingan educationalcrisis,
theteachersaid, and moreexpenditures foreducationweredrastically needed.
Anycuts in federalservicesor spendingwouldinevitably reducethealready
inadequatefundsavailableforeducation.Justa monthlater,however, thesame
individualfavoredcutsin government spending.Government was too big and
had to be cut back. Therewas no reference to theeducationalcrisisthathad
preoccupied thisindividualjusta fewweeksearlier. 10
Researchershave long knownthatdifferent people can answeridentical
questionsas iftheyconcerneddifferent topics.Whatthevignette of thevacillat-
ingteachershowsis thatthesamepersoncan answerthesamequestionat differ-
enttimesas if it involveddifferent topics.This can happen,accordingto the
model,becausetheconsiderations thatdetermine people'ssurveyanswersvary
acrossinterviews. Thus,peoplecan give strongly felt,contradictorysurveyre-
sponseswithout eitherchangingtheirmixoffeelingson theissueor consciously
experiencing anyambivalenceor conflict-iftheparticular considerationsthat
determine theirsurveyresponseshaveshifted.
Our data werecollectedwitha specificview to detecting and measuring
suchshifts.Table5 presentsclosed-endedsummaries of thesedataforfourre-
spondents." Notethatthesedataare fromthestop-and-think side,in whichre-
spondents wereencouragedto thinkaboutissuessomewhat morefullythanthey
ordinarilywould.
Considerrespondent reactionto a guaranteed
A. His first standardofliving
was thatit was inconsistentwithAmericanideals;he was also bothered by the
unfairnessof supporting thosewhorefuseto work.Yethe worriedaboutletting
individualsget ahead on theirown, sayingthatsome peopleneed specialhelp
and thatsocietyhas an obligationto help theneedy.In thesecondinterview,
however, therewas no signof thisambivalence.Respondent A gavesix reasons
whyindividuals oughtto getaheadon theirown,including a restatementof his
'?We would like to presentverbatim of whatthisand otherrespondents

transcriptions said.
However,theHumanSubjectscommittee ofMichiganhasdetermined
at theUniversity thatsuchuse
oftherawprotocolswouldbe an invasionoftherespondents' righttoprivacy.
" The selectionsweresubjecttotheconstraint
thatwe wantedtworespondentswhowerestable
on theclosed-ended itemsandtwowhowereunstable.
Table 5. Contentof Open-EndedResponseson Job GuaranteesQuestion
Couldyoutellme whatkindsofthingscometomindwhenyouthink
about...
FirstWave SecondWave
... government mak- eachper-
... letting ... government mak- ... letting
eachper-
ingsurethatevery son getahead on ingsurethateach son getahead on
personhas a good theirown? personhas a good theirown?
standardofliving? standardofliving?
A:
Respondent on FixedResponse)
(StableConservative
110. Idea is un- 180. Some people 158. Tax burdentoo 140. Individualism/
American needhelp(Lib.) great(Con.) workethic
(Con.) (Con.)
161. Dutyto helpthe 110. Idea is un- 133. Equal opportu-
needy(Lib.) American nityexistsfor
(Con.) all (Con.)
136. Unfairifsome 346. Program/food 344. Program/educa-
don'twork stamps(Con.) tion(Con.)
(Con.)
145. All shouldmake
italonebut
someneedhelp
(ETU)
B:
Respondent on FixedResponse)
(StableConservative
158. Tax burdentoo 344. Program/educa- 344. Program/educa- 140. Individualism/
great(Con.) tion(Lib.) tion(Lib.) workethic
(Con.)
156. Gov't. redtape 344. Program/educa- 219. Groupref.to people
142. Shiftless
(Con.) tion(Lib.) middleclass deservefate
(Con.) (Con.)
144. Valueofcom- 110. Idea is un- 152. Limitedgov't. 147. Motivationto
petition American (Con.) work(Con.)
(Con.)
140. Individualism/
workethic
(Con.)
C:
Respondent (Unstable,Conservative
to Liberal)
140. Individualism/ 140. Individualism/ 151. Gov't. mustin- 140. Individualism
workethic workethic sureequal opp. workethic
(Con.) (Con.) (Lib.) (Con.)
161. Dutyto helpthe
needy(Lib.)
D:
Respondent (Unstable,Liberalto Conservative)
344. Program/educa- 140. Individualism/ 150. Idea of welfare 111. Fairnessof
tion(Lib.) workethic state(Lib.) Amer.system
(ETU) (Con.)
Table5 (continued)
Couldyoutellmewhatkindsofthingscometo mindwhenyouthink
about.
FirstWave SecondWave
... government mak- . . . letting
eachper- . . . government
mak- . . . letting
eachper-
ingsurethatevery son getahead on ingsurethateach songetahead on
personhas a good theirown? personhas a good theirown?
standardofliving? standardofliving?
357. Program/hous- 348. Program/health

ing(Lib.) (Lib.)
356. Program/wel- 143. Work& welfare
fare(Lib.) (Lib.)
ofthesecodescan be foundin theICPSR codebookforthe1987PilotStudy.

Note:A fulldescription
Each remarkis identified
in parentheses
as havinga liberal,conservative,
or uncertaindirectional
thrust
(ETU indicatesthattheevaluational
thrustoftheremark was unclear).Evaluational
thrust
was
codedindependentlyof substantive
contentof remark.
feelingthatjob guarantees are un-American, without raisinganyopposingcon-

siderations.
RespondentA, who opposed government job guaranteesin bothof his
closed-ended attitude reports,is thusless stablein hisreactionto theguaranteed
standardof livingquestionthanhis stableclosed-endedresponseswouldsug-
gest. He wentfrombeingan ambivalent conservativeon thisissue to beinga
confident conservative. Giventheambivalence oftheinitialinterview,one would
notbe surprised if,overmanyencounters withthisquestionorslightly rephrased
versionsofthequestion,he occasionallysaw thecentralissueas aid totheneedy
rather thanun-American idealsand,on thisbasis,expressedsupportfortheso-
cial welfareoption.Certainly he hasrealimpulsesinthatdirection.
Nowconsiderrespondent C. Although she is ambivalentat bothinterviews
on thebasis of similarconsiderations, she changesherclosed-endedresponse
fromconservative to liberal.One can imaginethatshemayhave,ineffect, tossed
a mentalcoin in decidinghow to answerthefixedquestion-not because,as
Converse'snonattitudes thesissuggests,she had no meaningful reactionto the
issue (clearly,she did), and notbecause,as themeasurement errormodelssug-
gest,she was notquitesurewhatthequestionasked(she saw itthesamewayat
bothinterviews), butbecauseshewas undecidedbetweenlargelystablebutcon-
flictingimpulses.
These cases showthateven if, as we contend,people base theirattitude
reports on theideasthataremostimmediately salientto them,itis byno means
easy to explainover-time In themoststraightforward
responseinstability. case,
instabilityarisesfromchangesin theconsiderations thatare immediately salient
at thetimeof makingan attitude report,as inthecase ofthevacillating teacher.

Butother,quitedifferent scenariosarealso possible,as we haveseen.Withthese
complexities in mind,let us proceedto a moresystematic examination of re-
sponseinstability.
If, as themodelclaims,individualspossesscompeting considerations on
mostissues,andiftheyansweron thebasisofwhatever ideashappentobe atthe
topoftheirmindsat themoment ofresponse,one wouldexpecta fairamountof
over-time inpeople'sattitude
instability reports (Deduction4). Thereasonis that
theconsideration(s) thatare stochastically accessibleat one interview mightnot
be so prominent at thenext.This inference is strongly supported by a massof
existing evidence(e.g., Table 1).
The modelnotonlyanticipates responseinstability, butalso expectsit to
havea definite structure. Supposethat80% oftheconsiderations inoneperson's
headinducehertowarda liberalresponseon a givenissue,while20% induceher
towarda conservative response;andsupposethatfora secondperson,thesepro-
portionsare reversed.If each based her surveyresponseson a one-element
samplefromthedistribution of considerations in herhead,each wouldexhibit
responseinstability overtime,butoverthelongrun,thefirst personwouldbe
liberal80% of thetimeand thesecondwouldbe conservative 80% ofthetime.
Thus,citizenswouldhavecentraltendencies thatare stableovertime,buttheir
attitude statements wouldfluctuate greatlyaroundthesecentraltendencies (De-
in the
duction5). Thisis, fact,exactly pattern thathas been obtained repeatedly
byresearchers in the"measurement error"tradition(Achen1975;Erikson1979;
Juddand Milburn1980; Judd,Milburn,and Krosnick1981; Feldman1989;
Zaller1990;see, however, Krosnick1988).
If,as shown in Table 3, morepolitically awarepersonshavea largernumber
of considerations at thetop of theirhead and accessibleforuse in answering
questions,theyshould,all else beingequal, exhibitgreaterstabilityin their
closed-ended responses.The reasonis thatattitude reports formed froman aver-
age of manyconsiderations will be a morereliableindicator of theunderlying
population ofconsiderations thanan averagebasedonjustone ortwoconsidera-
tions(Deduction6).
Although initialresearchfailedtoconfirm thisexpectation, Feldman(1989)
and Zaller (1990) havemorerecently shownin separatedata setsthatpolitical
awarenessis, in fact,associatedwitha reduction in thechancevariation asso-
ciatedwithpeople'sattitude reports.(The difference betweentheinitialandlater
testsofthisexpectation is thattherecentworkusestestsofpoliticalinformation
as themeasureofpoliticalawareness.)
By parallellogic,peopleshouldbe morestablein theirresponsesto close-
endedpolicyitemsconcerning doorstepissues-thatis, issuesso closetoevery-
dayconcernsthatmostpeopleroutinely givesomethought tothem(Deduction7).
Evidencetending to supportDeduction7 maybe foundinthediscussionofracial

issues in Converse(1964), the discussionof moralissues in Converseand
Markus(1979), andgenerally inFeldman(1989).
Anotherimplication of our modelis thatgreaterambivalence oughtto be
associatedwithhigherlevelsofresponseinstability (Deduction8). Since,as we
justsaw,someindividualswhoexhibitno apparent ambivalence withina single
interview maynonetheless be quiteconflicted, this
it is essentialthat,in testing
implicationof themodel,we employa measureof ambivalence thatspansboth
interviews.Accordingly,we havebuilta measureoftheextentto whichan indi-
vidual'sconsiderations favorone or theotherside of a givenissue
consistently
acrossbothwavesof thesurvey.We constructed thismeasureby meansof the
following formula:
I (liberalremarks)- X (conservative remarks)
X (liberal) + E (conservative)+ X (ambivalent)
A scoreof one on thismeasurewouldindicatethattheperson'sremarks were
either all liberalin their thrust or all conservative, while a score of zero would
indicatethatthepersonhad madean equal numberof liberaland conservative
remarks. We expecthigherstability on theclosed-ended itemsforcases in which
all oftheperson'sremarks runinthesamedirection. The datainTable6 support
thisexpectation. In fiveof six trials,thismeasurewas associatedwitha statisti-
callysignificant increasein responsestability; in thesixthcase, therelationship
achievesmarginal statistical
significance (p = .07).
The resultsare not,however, as strongas theymightbe. If peopleformed
responsesby averagingacrossaccessibleconsiderations, stabilityratesshould
varybetween50% (forpeopleevenlydividedin theirconsiderations) to 100%
(forpeople withperfectly consistent considerations). Particularlyon thestop-
and-think side,thedatasignificantly departfromthisexpectation. How can this
be explained?
The mostlikelyreasonfortheshortfall fromexpectations is codingerror.
As reported earlier,randomcheck-coding indicatedthatcodersdisagreedon the
codingof 10% to 15% ofall open-ended remarks, andsuchmiscoding obviously
impairsourabilitytodetermine whichrespondents shouldbe perfectly stableand
whichperfectly random.
To testthisexplanation fortheimperfect resultsinTable6, we developedan
item-level measureof codingerrorand correlated it withthemagnitude of the
stability-consistency relationships shownin Table6. The correlation was, as ex-
pected,quitehigh.12Moreover, codingerrorwas consistently higheron thestop-
errorrates,we used item-level

'2Toestimateitem-level percentages thatthecoders
of remarks
foundtobe uncodable.The rateofuncodableremarks variedfroma lowofabout5% forone itemto
a highofnearly20% foranother.Weexpectedthatthemagnitude relation-
ofthestability-consistency
Table 6. ResponseStabilityand Consistencyof Considerations
Job Government Aid to

Guarantees Services Blacks
RetrospectiveConsiderations
Consistency
ofconsiderations: N N N
.00 .50 (7) .59 (11) .57 (7)
.01 to .50 .80 (20) .70 (25) .71 (19)
.51 to .99 .77 (15) .78 (16) .80 (15)
1.00 .91 (63) .87 (54) .96 (71)
(p < .01) (p < .02) (p < .01)
Stop-and-ThinkConsiderations
Consistency
ofconsiderations: N N N
.00 .63a (16) .54 (14) .57 (14)
.01 to .50 .68 (74) .77 (63) .83 (66)
.51 to .99 .73 (37) .80 (50) .84 (44)
1.00 .88 (45) .73 (37) .88 (41)
(p < .02) (p < .07) (p < .01)
Note:Cell entriesareproportion itemsfromwave 1 towave

stablein theirresponsestoclosed-ended
2. Measureof consistencyis describedin text.P-valuesarebasedon uncollapsedmeasure.
and-think side, whichexplainswhytheshortfall fromexpectations is higherin

thatformofthesurvey. 13
Withinthelimitsofthedata,then,theresultsinTable6 areprobably about
as strongas could be expected.We obtainedsizablestabilityeffectsat conven-
tionallevelsofstatistical infiveofsixtrialsandwerecloseto signif-
significance
icanceon thesixth,despitesignificant measurementerror,a smallsample,and
abnormally owingto therelatively
low base ratesof instability shorttimebe-
tweenreinterviews.
One additionalpointneeds to be made. The measureof consistency in
shipsin Table 6, as summarized by unstandardized regressionslopes,wouldbe greatest forthose

itemswiththelowestratesofuncodableremarks. Thiswas strongly thecase (r2 = 0.93 on a logafith-
micfit;df= 5; p < .05). Further detailsofthistestarereportedinZaller(in press,chap.4).
'3Itis, we mightadd, notsurprising thaterrorratesare higheron thestop-and-think side. It
wouldseemeasier,a priori,to determine thedirectionalthrustof a remarkiftheremarkhas been
madeinjustification ofan opinionjustrendered,as on theretrospectiveside,thanifithasbeengiven
as partofa discussionofwhata givenphrasemeansto therespondent, as on thestop-and-think side.
6oo JohnZallerand StanleyFeldman
Table6 was calculatedoverbothwavesofthesurveyin orderto captureconsist-

encyof considerations bothwithininterviews andacrossthem.The consistency
measurecan, however,be calculatedwithina singleinterview. Whenthisis
done,we comeup witha strong, clearfinding:theone-wavemeasureofconsist-
encyof considerations has almostno capacityto predictover-time responsein-
stability(in eitherform).
This finding indicatesthattheconflictmostresponsible forresponseinsta-
bilityis conflictthatoccursacrossrather thanwithininterviews andthatrespon-
dentsare oftenunawareof theirconflictas theyanswerquestions.Thus, the
vacillating teacherexhibitedno conflict overgovernment serviceswithineither
interview, butsubstantialconflictacrossinterviews. Mostlikely,oncetheteacher
beganto viewthegovernment servicesitemthrough theprismofeither"bloated
government" or "educationcrisis,"he or she fell intoa mindsetthatblocked
thinking aboutthe otherpointof view.'4This suggestionfitsnicelywithour
model,whichholdsthatpeopleanswersurveyquestionshastilyandon thebasis
ofincomplete memory searches.
Explaining"ResponseEffects"
The modelcan also explainseveralimportant responseeffects in masssur-
veys,thatis, cases in whichseemingly irrelevantfeatures of questionorderor
designaffects theresponsesgiven.In thissectionwe surveya variety of these
responseeffects.
Considerfirst"ordereffects," suchas theeffect of a priorquestionon sup-
portfortherightsof communist reporters,as discussedearlier.If,as themodel
claims,people are normallyambivalent on issues butansweron thebasis of
whateverideas are mostaccessibleat the momentof answering, raisingnew
considerationsin immediate proximityto a questionshouldbe able to affect the
answersgivenbymakingdifferent considerationssalient(Deduction9).
The intrusionof unexpectedor novel considerations into the question-
answeringprocesswould not, however,be expectedto affectall respondents
equally.Some peoplemaypossessconsiderations thatare so consistent in sup-
portof one side of an issue thattheadmission of one competing consideration
shouldhaveno effect. Others,however, maybe deeplyambivalent on theissue-
thatis, maypossessa roughly evenbalanceofproandconconsiderations. These
are thepersonswho shouldbe moststrongly affected changesin
by artificial
questionorder(Deduction10).
Tourangeau et al. (1989) havereported supportforthisexpectation. People
whoreported thattheyhad mixedfeelingsaboutan issue(and whoalso said the
'4Tourangeau and Rasinski(1988) distinguish between"interpretation"
of a questionand "re-
trieval"of information
relevantto answering it. In termsof thisdistinction,
ourresultssuggestthat
is themoreimportant
interpretation sourceofresponseinstability.
A THEORY OF SURVEY RESPONSE 6oi
issue was veryimportant to them)werequitesusceptibleto questionorderef-

fects-or whattheauthorscall "carryover effects" whileotherpersonswere
notsusceptible at all. The carryover effectsin thevulnerablegroups(wherethe
groupsvariedfromissueto issue)rangedfromlowsof4 and8 percentage points
on abortionand welfare,to highsof 34 and 36 percentage pointson aid to the
Contrarebelsanddefensespending,withan averagecarryover of 19 percentage
points.
Anotherimportant typeof responseeffectis whatmaybe called the"en-
dorsement effect."In thiscase, reference to a politicalfigureor groupsystem-
aticallyaltersthepublic'sresponsesto a givenpolicydilemma.For example,
Mueller(1973) foundthatquestionsthatindicated thatPresident Johnson favored
a particularpolicyoptionled to greatersupportforit. Similarly, references to
communism systematically increasedsupport forU.S. involvement inKoreaand
Vietnam.This anticommunism endorsement accordingto Mueller,sug-
effect,
gested"somewhat conflicting observations.On theonehand,support forthewar
was clearlytiedto theanti-Communist spiritin Americaatthetime.To generate
a kindof warfever,one merelyhad to toss thewords,'Communist invasion,'
intothediscussion.On theotherhand,theCommunist elementwas notentirely
builtintotheresponseto thewarbecauseAmericanshad to be reminded of it
beforetheiranti-Communism was fullyactivated"(1973, 48). The tendency of
peopleto base attitude reportson theideas thatare mostimmediately salientto
them,as specifiedin Axioms2 and3, wellexplainssucheffects (Deduction11).
Althoughmostresponseeffects are considered"methodological artifacts,"
theyaresometimes givensubstantive namesorinterpretations. Considerthese:
1. Race of interviewer effects.Shortlyaftera 1986 New YorkTimespoll
foundthatPresident Reagan'sapprovalrateamongblackswas 37%, a Washing-
tonPostpollestimated thatblackapprovalofReaganwas only23%. The differ-
ence was tracedto thefactthat,whiletheTimesfollowednormalinterview pro-
cedures,thePostused blackinterviewers whoinformed theirblackrespondents
thattheywouldbe participating in a studyof theattitudes
of blackAmericans.
As Sussman(1986) maintains, thelikelyeffect of thiswas to induceblackre-
spondents to "thinkblack"intheirevaluationsofReagan'sperformance (thefirst
itemon thesurvey).15
'5Thisandotherrace-of-interviewer effectsmightbe interpreted as "socialdesirability"

effects.
If by social desirability one meanscases in whichpeopleconsciouslymisreport
effects in
attitudes
orderto avoid embarrassment, a social desirability is at variancewithourargument.
interpretation
Butif,as seemsequallyconsistent withavailableevidence,socialdesirability is takento meancases
inwhichpeopleareunsurewhattheirattitudes arebutareinfluenced bytheimmediate context togive
greater weightto a particular
consideration, thensocial desirability
effectsrepresent simplyanother
typeof situationin whichpeoplemakeattitude reportson thebasis of theideas mostimmediately
salienttothem.
2. Referencegroup effects.In a classic social psychologyexperiment,

Charters and Newcomb(1958) foundthatRomanCatholicsweremorelikelyto
stateattitudesthatwereconsistentwithchurchdoctrine(e.g., on birthcontrol)
their
if,justpriorto questioning, religionwas madesalientto them.Thiseffect
was takenas evidence oftheimportance "reference
of groups"on attitudes.
3. "Primingeffects"of televisionnews.Accordingto Iyengarand Kinder
(1987) and Iyengar(1991), televisionnewsoftenfunctions to "prime"certain
ideas, therebymakingthemmoreaccessibleforuse in evaluating presidential
performance, decidingbetweencandidatesin elections,and assessingthena-
tion'smostimportant problems.In thisway,TV newsis said to affect attitude
reports permanently
without changing attitudes.
individuals'underlying
4. "Framingeffects" ofquestionwording and questionorderIftheconsid-
erationspeopleuse in answering questionscan be primedbyTV news,theycan
also be primedby howquestionsare wordedor framed.Forexample,Lau and
Sears(1983) and Lau, Sears,andJessor(1990) haveexperimentally shownthat
questionsaboutpersonalfinancial statusare morestrongly correlatedwitheval-
uationsofincumbent politicianswhenthequestionsareaskedincloseproximity.
The reason,presumably, is thata considerationused in answering one question
remainsavailableforanswering subsequent questions,thereby inducing a corre-
lation.16
The mechanism responsibleforeach of theseeffects appearsto be a ten-
dencyforpeopletoanswerquestionsatleastpartly on thebasisofideasthathave
been made momentarily salientto them.As such,theymaybe countedas re-
sponseeffects forwhichourmodelof thesurveyresponsegivesan explanation
(Deductions12, 13, 14, 15).
Milbum(1987) has documented yetanotherway in whichquestionorder
effects can occur.He foundthataskingrespondents to "tellme everything that
comesto mindwhenyou thinkof a Liberal(Conservative)" causedsubsequent
attitude reportsto be moreideologicallyconsistent withone anotherthanwere
theattitude reportsof a controlgroup.Milbumobtainedthiseffect, however,
onlyforpersonswho had eitherliberalor conservative leaningsto beginwith.
Price(1991) has recently replicatedthesefindings in an experiment conducted
on collegestudents and in nonexperimental datafromnationalsamples.In both
cases, Pricefoundthatsimplyaskingpersonsto place themselves on a liberal-
conservative to
ratingscale was sufficient increasethe consistency
ideological of
responsesto subsequentpolicyitems-but, again, only forpersons who pos-
sessed a clear ideologicalleaning(i.e., were notcentrists or unableto place
themselves on the self-ratingscale).
on whether
'6Thereis disagreement theseeffects
haveartificially of"pocket-
estimates
inflated
bookvoting"in NES surveys(see Lewis-Beck1985;Lau, Sears,andJessor1990),butno disagree-
mentthatframing effects do so.
could,inprinciple,
This typeof questionordereffect is also explainablefromthemodel.Hav-

inghadtheirideologicalorientations madesalienttothemjustpriortoanswering
policyitems,thoserespondents whopossesssuchorientations aremorelikelyto
relyon themas a consideration in formulating responsesto subsequentpolicy
questions,thereby makingthoseresponsesmorestrongly correlatedwiththeir
ideologicalpositionsand hencealso moreideologically consistent withone an-
other(Deduction16).
Thissectionhas examineda variety ofempiricalregularities-question or-
dereffects,endorsement race-of-interviewer
effects, effects,
reference groupef-
fects,priming effects,questionframeeffects, andtheeffect of makingideology
salient.Withinconventional attitudetheory, someoftheseempiricalregularities
aretakenas substantive findingsand someas methodological artifacts.
We have
shown,however, thatall maybe explainedas manifestations ofa commontheo-
reticalmechanism, namely, thenormaltendency of peopleto respondto survey
questionson thebasis of theideas thathappento be, forwhatever reason,im-
mediately salienttothem.
ExplainingtheEffects
ofExtraThought
Surveyresponses,as conceivedhere,are not"attitudes" per se; theyare
unreliableindicators of themixof considerations in theperson'smind-unreli-
able because,amongotherthings,people normallyanswerwithoutretrieving
frommemory all relevantconsiderations.If,however, peoplecouldbe artificially
inducedto retrieve a largerthannormalnumber ofconsiderations,itshouldim-
provethereliability oftheirresponsestoclosed-ended items.
Our intentin designingthestop-and-think probeswas to createsuchan in-
ducement.By requiring to discusstheelementsof a questionbefore
individuals
answering it,we wereinducingthemtocall to mindandtakeaccountofa wider
rangeof ideas thantheynormally would.We therefore expectedthatresponses
followingthestop-and-think treatmentwouldbe, all else equal, morereliable
indicatorsof the set of underlyingconsiderations thanresponsesmade in the
standard way,thatis, intheretrospectivecondition (Deduction17).
Unfortunately,ourabilityto testthisexpectation is compromised by an ar-
tifact.Because of theuse of an explicit"no interest"optionin theretrospective
conditionbutnotin thestop-and-think condition, morerespondents failedin the
retrospectiveconditionto respondto the issue item.Low-awarenesspersons
weremostaffected by thisquestiondifference; theirno opinionrateaveraged
38% intheretrospective conditionbutonly4% inthestop-and-think condition.'7
Thismeansthatretrospective respondents, especiallyless-awareones,area more
selectedgroupandwould,forthisreasonalone,be expectedtobe moreideolog-
icallyconsistentthantheirstop-and-think counterparts. Thisartifact
runsagainst
in no opinionratesacrossformsaveraged13% (16% vs. 3%) in
thedifference
'7Bycontrast,
thehighestinformation
quintile.
thegrainoftheanticipatedstop-and-thinkeffect,makingitmoredifficulttodem-
onstratetheeffect,
particularly
forless-awarerespondents.
We developedtwotestsof theexpectation of increasedresponsereliability
in thestop-and-think test,we expecteda measureofsocial
condition.In thefirst
welfareideology(see appendix)to be morestrongly correlatedwiththetarget
items(jobs, governmentservices,andaid toblacks)inthestop-and-think condi-
condition.
tionthanintheretrospective Weusedthefollowing interactive
regres-
sionmodeltotestthisexpectation, whereFormrefers toquestionform:
Item= bo + b, x Form+ b2 x Ideo. + b3 x Form x Ideo.
Whenwe estimated thismodelforrespondents whoscoredintheupper40%
of ourmeasureof politicalawareness,we foundthatthecriticalcoefficient, b3,
ranin theexpecteddirection forall threeitems,butachievedstatistical
signifi-
cancein onlyone case. To increasethestatistical poweroftheinteraction testin
our smallsample-the numberof cases in each testaveragedabout 140-we
reestimated themodelundertheconstraint thatall coefficients
be equal across
thethreeitems.The results,shownin thetoppanelofTable7, confirmed expec-
tations:theeffect of ideologyis twiceas largein thestop-and-thinkcondition, a
differencethatis statistically
significant.
We also estimated themodel,underthesame constraints, forrespondents
scoringin thebottom40% oftheawarenessmeasure.Herewe foundthat,as also
on
Table 7. The EffectofStop-and-Think
IdeologicalConsistency
Low Awareness HighAwareness

Intercept 0.01 - .20
Ideology 0.89 0.62
(.21) (.14)
Form -0.36 0.00
(.21) (0.20)
Ideologyx Form -0.24 0.62
(.27) (.26)
N= 434 437
Note:Model is shownin text;modelwas estimated acrossthe

simultaneously
jobs, services,and minority aid items.Data are takenfromwave 1 of the
survey, answered"don'tknow"on wave
exceptin cases in whichrespondents
1, inwhichcase theirwave2 responses,ifany,wereused.Testwas conducted
forrespondents scoringin thebottom40% and top40% of awarenessscale.
Measuresaredescribedin theappendix.
shownin Table7, thestop-and-think testnotonlyfailedto increaseconsistency,

butmightactuallyhavereducedit. We shallreturn to thisapparent reversalin a
moment.But first,we havea secondtestof Deduction17. If extrathought in-
ducesmorereliableattitude reports,itshouldenhancenotonlycorrelations with
ideologybutalso theover-time stabilityof theseresponses.As can be seen in
Table 8, however,thedata completely failto supportthisexpectation. In fact,
less-awarepeopleexhibitless consistency inthestop-and-think condition, while
more-aware onesshowno effect.
It is essentialto evaluatetheseresultsin lightof theselectionartifact we
havedescribed.The gaininreliability amonghighlyinformed personsinTable7
runsagainstthegrainoftheartifact andso is especiallylikelytobe real.The null
Table 8. The EffectofStop-and-Think

on Test-RetestCorrelations
Retrospective Stop-and-Think
Low Awareness
Jobguarantees .68 .45
(40) (62)
Government
services .56 .43
(41) (58)
aid
Minority .79 .53
(53) (57)
MiddleAwareness
(29) (60)
services
Government .48 .38
(31) (51)
aid
Minority .81 .51
(33) (58)
HighAwareness
(39) (51)
services
Government .61 .58
(37) (48)
aid
Minority .70 .86
(32) (45)
Note: Cell entriesare test-retest

Pearsoncorrelations.
Numbersare in paren-
theses.
6o6 JohnZallerand StanleyFeldman
findings and apparentreversals,because theymighthave been caused by the

artifact,are moresuspect.Butmightnotthereversals also representrealeffects
of thestop-and-think treatment? Indeed,theymight.In an impressive seriesof
experiments, Tim Wilson and colleagues (Wilson et al. 1989; Wilson and
Hodges 1991) haveshownthat,contrary to ourmodel,askingpeopleto articu-
latethereasonsfortheirattitudes consistently reducesthepredictive reliability
of attitudereports,especiallyforpersonsless knowledgeable aboutthegiven
attitudeobject.
The explanation forthedisruptive effectsofthought, as Wilsonet al. main-
tain,is thataskingpeople to thinkaboutthereasonsfortheirattitudes causes
themto sampleideas thatare too heavilyweightedin thedirection of cognitive
reactions to theattitude
objectrather thanaffective ones.Attitude reports thatare
based on thisunrepresentative sampleare, as theyconclude,less reliablethan
reportsbasedon theideasthatareotherwise mostaccessibleinmemory
Notethatthisargument acceptsthecentralassertion ofourmodel,whichis
thatpeople'sattitude reports(and also, as theWilsonet al. studiesshow,behav-
iors)reflect theideas thatare at thetopof thehead at themomentof decision
ratherthanany deepertypeof "trueattitude." In fact,it is preciselybecause
attitudereports dependon immediately salientideasthatextrathought, in bring-
inga biasedsampleofideastothetopofthehead,provesdisruptive.
The argument ofWilsonetal. aboutoversampling cognitions attheexpense
of feelingscouldexplaintheunexpected resultswe obtainedin Tables7 and 8.
Sniderman, Brody,andTetlock(1991) haveamassedconsiderable evidencethat
less-awarepersonsare morelikelyto base attitude reports on feelings,whereas
more-aware ones tendto respondon thebasisof ideologicalprinciple.It would
followfromthisthatthestop-and-think treatment, withitsrelativelycognitive
flavor,should prove most disruptive to the attitudereportsof theless aware
whichis whatwe havefound.
Sincethisposthocexplanation runsinthedirection oftheselectionartifact,
we cannotbe surethatit is correct.It nonetheless makestheusefulpointthat
failureto confirm (Deduction17) maybe due to a thought manipulation that
failedtomimicpeople'snaturalthought processes.Ifso, a morecarefully crafted
manipulation mightyetproducethereliability gainthatourmodelanticipates.'8
Summary
of what
The modelwe haveproposedis, like all models,a simplification
actuallyoccurs. In addition,it has less formalprecisionthanconventional
studiesshowingthatthought
'8Thereare,in fact,psychological manipulations thatstresseither
theaffective aspectsofsituations
orcognitive producepredictably
different
effects,
disrupting attitude
reports in somesituationsbutnotothers(see MillarandTesser1986). Forfurther
discussionofthese
issuesas theyrelatetothemodel,see Zaller(forthcoming,chap.5).
measurement errormodels(see Achen 1975) and paysless attention to mental

processesthando the attitudemodelsof psychologists. But, despiteits limi-
tations,the simple three-axiommodel is, we believe, uniquelysensitive
to thewide rangeof empiricalregularitiesassociatedwiththemass surveyre-
sponse.Our hope, therefore,is thatas otherscriticizeour modeland propose
theykeep theirtheorizing
alternatives, broadplanethatit can
on a sufficiently
accommodate therangeofempiricalregularitieswe haveexamined.
The empiricalphenomena forwhichourmodeloffers an explanation,which
we summarize inTable9, maybe groupedunderthreegeneralheadings:
1. Dependenceofattitude reports onprobabilisticmemory search.Because
attitude reportsare based on memorysearchesthatare bothprobabilistic and
incomplete, reports
attitude on the
tendto be (1) unstableovertime;(2) centered
meanof theunderlying considerations; withtheoutcomesof
and (3) correlated
memory searches(Deductions3-5). Thisis also whypeoplewhoaremorecon-
flictedin theirunderlyingconsiderations aremoreunstableintheirclosed-ended
surveyresponses(Deduction8).
2. Effects madesalient.The notionthatindividuals'survey
ofideas recently
responsescan be deflected ofideasmaderecently
inthedirection salienthasbeen
used to explainquestionordereffects,endorsement race-of-interviewer
effects,
reference
effects, groupeffects,questionframing andTV newspriming
effects,
(Deductions9-16).
effects
3. Effectsofthought on attitudereports.The notionthatthinking aboutan
issue,as gaugedbygenerallevelsofpoliticalawareness,enablespeopletorecall
a largernumber ofconsiderations andhenceto makemorereliableresponseshas
beenusedtoexplainwhymorepolitically awarepersonsexhibitgreater response
stabilityandwhythepublicas a wholeis morestableon "doorstep"issues(De-
ductions6, 7). It also explainswhymorepolitically awarepersons,andpersons
especiallyconcernedaboutan issue,are able to recallmorethoughts relevantto
it (Deductions1, 2). Finally,thenotionthatgreaterthought makes attitude
re-
portsmorereliablehas beeninvoked,withonlylimitedsuccess,to explainthe
effects ofextrathought atthemoment ofresponding to an issue(Deduction17).
Althoughitis easyto imagine alternativeexplanations formanyofthepar-
ticularphenomena we haveexplained with our model, itwould not,we think,be
easyto develop an alternativethathas therange and simplicityof ourmodel.Yet
our model does, as we have admitted,representa simplificationof a processthat
mustbe muchmorecomplicated. Whatsortsofsimplifications havewe made?
One of themostimportant involvestheissue of "on-line"processing.As
indicatedearlier,HastieandPark(1986), amongothers,arguethatpeopleoften
use a "judgment operator" toupdatecontinuously "on-line"as they
theirattitudes
acquirenew information. People are said to store their updated attitudesin
6o8 JohnZallerand StanleyFeldman
Table 9. List of EmpiricalPhenomenaThat theModel Claims to Explain
1 Peoplewhoare,in general,morepolitically awarehavemoreconsiderations at the

topoftheirheadsandavailableforuse in answering surveyquestions.
2. Peoplewhohavegreater interestin an issueshouldhave,all else equal, more
thoughts aboutthatissuereadilyaccessiblein memory thanotherpersons.
3. Thereshouldbe strongcorrelations betweentheideasat thetopof people'sminds
as theyanswersurveyitemsandtheirdecisionson theitemsthemselves.
4. Thereshouldexista fairamountofover-time instabilityin people'sattitudereports.
5. Opinionsthataresubjectto repeatedmeasurement shouldhavecentraltendencies
thatarestableovertime,butshouldfluctuate aroundthesecentraltendencies.
6. The attitudereportsof politicallyawarepersonsshouldexhibitgreater over-time
stabilitythanthoseof less-awarepersons.
7. Peopleshouldbe morestablein theirresponsesto closed-ended policyitemscon-
cerningdoorstepissues-that is, issuesso close to everyday concernsthatmost
peopleroutinely paysomeattention to them.
8. Greaterambivalence oughtto be associatedwithhigherlevelsofresponseinsta-
bility.
9. Raisingnewconsiderations in immediate proximity to a questionshouldaffect the
answersgivenby makingdifferent considerations salient.
10. Peoplewhoareambivalent on an issueshouldbe mostaffected bymanipulations
thatraisenewconsiderations in immediate proximity to a questionabouttheissue.
11. Inserting thenameof a prominent politicianor groupintoa questionshouldaffect
thepublic'sresponsesto thequestion(the"endorsement effect").
12. The raceof an interviewer shouldat leastsometimes affect theresponsesto ques-
tionswhichhe or she asks.
13. Manipulations thatraisethesalienceof a reference groupcan affect responsesto
questionson whichthereference grouphas a well-known position.
14. Newsreports can "prime"certainideas,thereby makingthemmoreaccessiblefor
use in formulatingattitudestatements on relatedsubjects(the"priming effect").
15. Questionordercan "prime"certainideas,thereby inducing correlations withproxi-
materelateditems.
16. Inducingindividuals to thinkabouttheirideologicalorientation in close proximity
to questionshavingideologicalcontent can "prime"ideologyforuse in answering
thosequestions.
17. Inducingpeopleto thinkmorecarefully aboutan issuebeforestating an opinion
shouldenhancethereliability oftheopinionreport.(Notconfirmed.)
long-termmemoryand retrievethemas required,ratherthan, as in our model,

create attitudestatementson the spot as theyconfronteach new surveyquestion.
We are sympatheticto on-lineinformation processingand believe thatsome
formof it eventuallymustbe included in models of political attitudes.Yet even
the strongestadvocates of on-line processing so farhave foundno evidence of
such processingin the formationof policy attitudes(as against the formationof
evaluationsof persons). And this is scarcely surprising.The issues about which
citizensmustanswersurveyquestionsaretoo numerous, too multidimensional,

and,in mostcases, too obscureforitto be feasibletoengageinon-lineprocess-
ingofall relevant
information.
Discussion
In closingwe wouldliketoconsiderthreebroadquestionsaboutourmodel.
Is it plausible?Whatare its implications
forfuture
opinionresearch?Whatare
itsnormative fortheroleofmassopinionina democracy?
implications
Substantive oftheModel
Plausibility
Of thevariousclaimswe havemade,theambivalence axiom(alongwithits
implication thatindividualsnormally do nothavea single,fixed,and firmatti-
tudeon issuesbutinsteadhavemany,potentially opposing"considerations") is
perhapstheleastfamiliarand hencetheleastintuitive. It also flatlycontradicts
thedominant academictheories ofpoliticalattitudes.Achen(1975) assumesthat
all respondents have"trueattitudes" and maintains thatthesetrueattitudes, al-
thoughappearing to vacillatebecause of measurement are
error, overwhelmingly
stable.Converse(1964) likewiseassumesthat,although manyrespondents have
nonattitudes, thosewhodo haveattitudes haveperfectlyfixedandstableones.
Despitethebreadth of itsappeal,theconventional idealoffixedand stable
"trueattitudes" is well pastdue forstrongquestioning. One needonlyconsider
that,in bothpublicopinionsurveysand in elitedebate,policyquestionsare
typicallyframedin termsof summary judgments(whetherabortionshouldbe
permitted or not;whether schoolchildrenshouldbe busedto promote racialin-
tegration or not)so thatthemakingof thesejudgments requiresan aggregation
ofone'sfeelingsacrossfrequently diverseconcerns.Thereis absolutely no rea-
son to supposethata personmustfeelconsistently abouteach of theelements
thathe or she aggregates acrosswhenmakingsuchsummary judgments. Thus,
someonewhosupports a woman'srighttocontrol herreproductive scheduleneed
notalso feelcomfortable withaborting fetuses(Scott1989); advocatesof gov-
ernment aid to theneedyneednotfeelcomfortable withbig government (Feld-
manandZaller1992).
The heartofourargument is thatformostpeople,mostofthetime,thereis
no needto reconcileor evento recognizetheircontradictory reactions to events
andissues.Each represents a genuinefeeling,capableofcoexisting withoppos-
ingfeelingsand,dependingon itssaliencein theperson'smind,controlling re-
sponsesto surveyquestions.Analystsof publicopinionlonghavebeen aware
thatfewcitizensare "ideologicallyconsistent" in theirresponsesto different is-
sues (Converse1964). Our argument is thatmanycitizensareequallyinconsis-
tentin theirreactionsto different aspectsof thesameissue,wherean "issue"is
simplyany bundleof concernsthata pollsterasks respondents to aggregate
acrossinthecourseofanswering a question.
Although thislineofargument mayseemunobjectionable, itis easytoover-
6to JohnZallerand StanleyFeldman
look its implications.The mostimportant is thatindividualstypically do not

develop"trueattitudes" of thetypethatopinionanalystsroutinely assume,but
possessa seriesof autonomous andofteninconsistent reactionsto thequestions
askedbypollsters.Or,to putitanotherway,mostopinionson mostissueshave
botha centraltendency anda variance.
Weemphasize,however, thatnothing inourmodeldeniesthatsomepersons
maydevelopwhollyconsistent setsof considerationswithrespectto some is-
sues. Indeed,thereis no doubtthatsomepeople,including manypoliticalactiv-
istsandotherswhomightbe driventowardcognitive consistency,
do exactlythis.
Suchpeoplearewellcapturedbyconventional "trueattitude"models.Butthese
peoplecan be accommodated equallywellbyourmodel,which,although hold-
ingthatmostindividuals areto someextentambivalent, allowssomeindividuals
tobe unambivalent. Thus,ourmodelaccommodates boththemajority ofpersons
whoare,as Table2 indicates,inconsistent in theirreactionstodiverseaspectsof
issuesandtheminority whoarestableandconsistent.
EmpiricalPotentialoftheModel
The value of our modelforimproving researchon publicopinionlies in
threedirections.One involvesopeningup themicrofoundations of attitudere-
portsto empiricalscrutiny andanalysis.Forexample,researchers oftenreferto
attitudesthataremoreor less "crystallized," moreorless ideological,or inother
waysheterogeneous acrosspeopleand issues(Rivers1988; Sniderman, Brody,
andTetlock1991). But sincethe"attitude" is alwaystheprimitive termof anal-
ysis,theyhavehad no wayof directly verifying theseimputations. Withinour
model,however, a crystallizedattitude
mightbe one thatis foundtobe basedon
a largeror morehomogeneousset of underlying considerations; similarly,an
ideologicalattitude mightbe one basedon abstract orprincipled considerations.
The empiricalconsequencesof thesemeasureddifferences in attitudinalmicro-
foundations couldthenbe investigated.
Anotherdirectionforfuture researchis in thearea of communication and
persuasion. Studies of politicalpersuasiontypically assumethatattitude change
involvesa conversion experiencein whichone crystallized attitude structure
re-
places another.Our model,however, raisesdifferent most
possibilities, notably
"persuasion byframing." The idea hereis thatifan eliteor someotherpersuader
discussesan issueso as tohighlight one setofconsiderations rather thananother,
thepublicwouldbe expectedto respondbybasingitsopinionstatements on the
considerations thusmadesalient.Edelman's(1977) and Bennett's(1980) work
on theeffectson massopinionof symbolicissuemanipulation byelites,Kinder
and Sanders's(1990) workon the effectsof different questionframes,and
Popkin's(1991) observations on howcandidates use issuestomobilizevotersup-
portall affirmtheimportance ofmanipulating massopinionbymanipulating the
considerationsthataresalienttothepublic.
A THEORY OF SURVEY RESPONSE 6i i
Traditional persuasionremainspossiblewithinourconception of attitudes,

butitassumesa somewhatdifferent form.As Zaller(1984, forthcoming) main-
tains,attitudechangewithinourframework mustnotbe understood as an all-or-
nothing shiftin a "trueattitude,"butas an adjustment in themixof considera-
tionsrelatingto an issue.
Finally,thenotionof a fundamentally ambivalent publiccan helpdevelop
moreeffective linkagesbetweenpublicopinionand thepolicymaking process.
Forexample,V. 0. Key (1961) foundinPublicOpinionandAmerican Democ-
racythata sizablefraction of theelectoratewantedstronger social welfarepro-
tectionswhilealso favoring taxcuts"evenifitmeansputting offsomeimportant
thingsthatneedto be done."In a comment muchinthespiritofourmodel,Key
observedthatbothattitudes, thoughappearingcontradictory, werereal. As he
wrote,"a simplecalculusof self-interest makessimultaneous supportof taxre-
ductionand expansionof social welfareactivitiesentirely consistent.. . . For
thesystemas a whole,however,thistypeof opinioncombination is irrational
and createsproblemsin programmaking.. . . The balance of forcesdrives
policymakers backtowardconcealedandindirect taxation, whichmaybe regres-
siveinitsincidence"(168).
A similarpointhas been made by Free and Cantril(1967), who in their
studyThePoliticalBeliefsofAmericansfoundthatthepublicexhibited a com-
of
bination philosophical oppositionto welfarestate and
policies operational sup-
portforvirtually thefullrange of welfareprograms. This "schizoid"belief pat-
tern,Free and Cantrilargued,made it extremely forgovernment
difficult to
engageinrationalpolicy-making.
Thus,we feelthata conception thatemphasizesthat"opinion"on an issue
a rangeofreactionsrather
is generally thana single"trueattitude"willbe more
forinvestigating
fruitful linksbetweenmassopinionandthepoliticalprocess.
Normative oftheModel
Implications
The normativeimplications of Converse'snonattitudesthesisareextremely
bleak.In thelimitingcase of a publicwithout attitudes-aclaimthatConverse
did not actuallymake-self-government makeslittlesense. As Achen(1975,
1,227)putit,"Democratictheoryloses itsstarting point."In contrast,theimpli-
cationof Achen'sempiricalinvestigation was relatively
optimistic: mostmem-
bersofthepublichavetrueattitudes thatarealmostperfectly stable.Thedisturb-
inglyhighlevelsofresponsechangediscovered byConverse,Achenargued,are
due to thevaguequestionsof surveyresearchers ratherthanto thevagueminds
ofcitizenrespondents.
Our positionfallssomewhere betweentheConverseand Achenpositions.
We agreewithConversethatthereis a greatdeal of uncertainty, tentativeness,
andincomprehension in thetypicalmasssurveyresponse.The problem,we fur-
theragree,is muchdeeperthanvaguequestions.Andyet,withAchen,we reject
thepremiseof Converse'sblack-and-white model,whichis thatmostresponse

fluctuationis due to essentially randomguessingbypeoplewho haveno mean-
ingfulopinions.
Ourclaimis thatevenwhenpeopleexhibithighlevelsofresponseinstabil-
ity,theopinionstheyexpressmaystillbe based on real considerations. Even
whentheseconsiderations turnoutto be transitory, theopinionstatements they
generatearenot,forthatreason,necessarily lackinginauthenticity.
This argument extendsto the interpretation of aggregatesurveyresults.
Thus,if55% of Americansreportthattheyapproveof thewayGeorgeBushis
doinghisjob as president, it shouldnotbe takenas evidencethata majority of
thepublicis unequivocally supportive ofthepresident. Rather,itshouldbe taken
to meanthat55% of Americansare on balancepositivetowardBush'sjob per-
formance-eventhoughtheparticular 55% whoexpressapprovalwillnaturally
changefromone surveyto thenext,depending on thecross-cuttingaccidentsof
manydifferent people'smemory searches(Page andShapiro1992).
Thereis, then,no inconsistency betweenourassertions of individual-level
ambivalence andinstability, on theone hand,andbeliefinthemeaningfulness of
aggregate-level poll results,on theother-providedone readspollsas revealing
a balanceofconsiderations rather thanas countsofpeople's"trueattitudes."
Thisconception of an ambivalent publicmayfallshortofouridealof what
publicopinionoughtto be like,as thisidealis expressedinpoliticaloratory and
democratic mythology. Butifdemocracy is possiblein a countrythatbothglori-
fieseconomicindividualism and demandsthe welfarestate;thatprofessesto
cherishequalityand practicesracialdiscrimination; thatinsistson bothhigher
levelsof government servicesand lowertaxes;and thathatesCongressbutre-
electscongressional incumbents at extremely highrates-thenitis also possible
underourunderstanding ofmasspublicopinion.
Manuscriptsubmitted2 August1991
received4 December1991
Final manuscript
APPENDIX
Attitude
Items
Jobguarantees.Some people thinkthatthegovernment in Washingtonshouldsee to it that
of living.Othersthinkthatgovernment
everypersonhas a job and a good standard shouldjust let
eachpersongetaheadon theirown.
Beforetellingmehowyoufeelaboutthis,couldyoutellmewhatkindsofthings cometomind
whenyouthinkaboutgovernment makingsurethateverypersonhas a goodstandardofliving?(Any
others)
Now,whatkindsof thingscometo mindwhenyouthinkaboutletting eachpersongetahead
on theirown?(Anyothers)
Government services.Some peoplethinkthegovernment shouldprovidefewerservices,even

in areassuchas healthand education,in orderto reducespending.Otherpeoplefeelit is important
forthegovernment toprovidemanymoreservicesevenifitmeansan increaseinspending.
Beforetellingmehowyoufeelaboutthis,couldyoutellmewhatkindsofthingscometomind
whenyouthinkaboutfewer government services?(Anyothers)
Now,whatkindsof thingscome to mindwhenyouthinkaboutincreasesin government ser-
vices?(Anyothers)
Aid to blacks. Some peoplefeelthegovernment in Washington shouldmakeeveryeffort to
improve thesocialandeconomicpositionofblacks.Othersfeelthatgovernment shouldnotmakeany
to helpblacksbecausetheyshouldhelpthemselves.
specialeffort
Beforetellingmehowyoufeelaboutthis,couldyoutellmewhatkindsofthings cometo mind
whenyouthinkabout"thesocialandeconomicpositionofblacks?"(Anyothers)
Whatcomesto mindwhenyouthinkaboutefforts toimprove thesocial and economicposition
ofblacks?(Anyothers)
And (whatcomes to mind)whenyou hearthephraseblacksshouldhelp themselves? (Any
others)
PoliticalAwareness
Scale
A 19-pointscale, havingan alphareliability
of .85, as follows:v635to v642,v202(up to four
points).Two recognition tests,v242, v244 (one itemeach). Fourcomparative candidatelocation
items:v730 and v731; v749 and v750; v811 andv812; v831 andv832;plusv723 (a noncomparative
locationtest).
IdeologyScale
Social Welfare
The social welfareideologyscale consistsof 14 itemsconcerningindividualismandequality,
plus two measuresof ideologicalself-designation,as follows:v620 to v622, v624, v626, v701 to
v706,v2176,v2178,v2179,plusv722andvI 010.
REFERENCES
Achen,Christopher. 1975. "Mass PoliticalAttitudes and theSurveyResponse."AmericanPolitical
ScienceReview69:1218-31.
. 1983. "TowardTheoriesofData." Presented at theannualconvention oftheAmerican Polit-
ical ScienceAssociation,Chicago.
Anderson, Norman.1974. "Information Theory:A BriefSurvey."InAdvancesinMathe-
Integration
maticalPsychology, vol. 2, ed. David Krantzetal.
Bargh,John,R. N. Bond,W. J.Lombardy, andM. E. Tofa. 1986. "TheAdditiveNatureofChronic
and Temporary ConstructAccessibility." Journalof Personality and Social Psychology50:
869-78.
Bennett,Lance W. 1980. Public Opinion in AmericanPolitics. New York: HarcourtBrace
Jovanovich.
Bishop, George. 1990. "PoliticalInvolvement and ResponseEffects."Public OpinionQuarterly
53:209-18.
Bishop,George,RobertW. Oldendick,andA. J.Tuchfarber. 1984."WhatMustMy Interest inPoli-
ticsBe IfI JustToldYou 'I Don't Know'?"PublicOpinionQuarterly 48:510-19.
Bodenhauser, GallenV., andRobertS. Wyer.1987. "Social Cognition andSocial Reality."In Social
Information Processingand SurveyMethodology, ed. Hans J.Hippler,NorbertSchwarz,and
SeymourSudman.New York:Springer-Verlag.
Campbell,Angus,PhilipE. Converse,DonaldStokes,andWarren Miller.1960.TheAmericanVoter.

New York:Wiley.
Charters,W. W., Jr.,and TheodoreNewcomb.1958. "Some Attitudinal Effects of Experimentally
IncreasedSalienceof GroupMembership." In Readingsin Social Psychology, 3d ed., ed. E.
Macoby,TheodoreNewcomb,andE. Hartley. New York:Holt,Rinehart.
Converse,PhilipE. 1964. "The Natureof BeliefSystemsinMass Publics."In Ideologyand Discon-
tent,ed. David Apter.New York:FreePress.
Converse,PhilipE., andGregory Markus.1979."Plus ,a Change. . : The NewCPS PanelStudy."
AmericanPoliticalScienceReview73:32-49.
Dean, Gillian,andThomasMoran. 1977. "MeasuringMass PoliticalAttitudes: Changeand Unreli-
PoliticalMethodology
ability." 4:383-401.
Edelman,Murray.1977.PoliticalLanguage.New York:AcademicPress.
Ericsson,K. Anders,andHerbert Simon. 1984.ProtocolAnalysis.Cambridge: MIT Press.
Erikson,Robert.1979. "The SRC PanelData andMass PoliticalAttitudes." BritishJournalofPoliti-
cal Science9:16-49.
Feldman,Stanley.1989. "MeasuringIssuePreferences: The ProblemofResponseInstability." Politi-
cal Analysis1:25-60.
. 1990. "Theory,Measurement, andtheSurveyResponse."Presented attheannualmeeting of
theAmericanPoliticalScienceAssociation,Atlanta.
Feldman,Stanley,and JohnZaller. 1992. "PoliticalCultureof Ambivalence: IdeologicalResponses
totheWelfareState."American JournalofPoliticalScience,36: 268-307.
Free,Lloyd,and HadleyCantril.1967. The PoliticalBeliefsofAmericans.New York:FreePress.
Hastie,Reid,andBernadette Park.1986."TheRelationship betweenMemoryandJudgment Depends
on Whetherthe Judgment Task Is On-lineor MemoryBased." PsychologicalReview93:
258-68.
Hastie,Reid, and NancyPennington. 1989. "Noteson theDistinction betweenMemory-based and
On-lineJudgments." In On-lineCognitive PersonPerception, ed. JohnBassili. Hillsdale,NJ:
Erlbaum.
Higgins,E. Tory,and Gary A. King. 1981. "Accessibilityof Social Constructs: Information-
ProcessingConsequencesof IndividualandContextual In Personality,
Variability." Cognition,
and Social Interaction,ed. NancyCantorandJohnF. Kihlstrom. Hillsdale,NJ:Erlbaum.
Hippler,Hans J.,and Norbert Schwarz.1987. "ResponseEffects in Surveys."In Social Information
Processingand SurveyMethodology, ed. Hans J. Hippler,NorbertSchwarz,and Seymour
Sudman.New York:Springer-Verlag.
Hochschild,Jennifer.1981. What's Fair? Americans'AttitudestowardDistributiveJustice.
Cambridge:HarvardUniversity Press.
Iyengar, Shanto.1991.Is AnyoneResponsible?Chicago:University ofChicagoPress.
Iyengar,Shanto,and Donald Kinder.1987. News ThatMatters.Chicago: University of Chicago
Press.
Judd,Charles,andCynthiaLusk. 1984. "KnowledgeStructures andEvaluativeJudgments: Effects of
Structural Variableson AttitudeExtremity." Journalof Personality and Social Psychology
8:331-38.
Judd,Charles,andMichaelMilburn.1980."TheStructure ofAttitude SystemsintheGeneralPublic:
A Comparison ofa StructuralEquationsSystem."American SociologicalReview45:627-43.
Judd,Charles,MichaelMilburn,andJonKrosnick.1981."PoliticalInvolvement andAttitude Struc-
tureintheGeneralPublic."AmericanSociologicalReview46:660-69.
Kelley,Stanley, Jr.1983.InterpretingElections.Princeton:Princeton UniversityPress.
Key,V. O., Jr.1961.PublicOpinionandAmerican Democracy.New York:Knopf.
Kinder,Donald,and LynnSanders.1990. "Mimicking PoliticalDebatewithSurveyQuestions:The
Case ofWhiteOpinionon Affirmative ActionforBlacks."Social Cognition 8:73-103.
A THEORY OF SURVEY RESPONSE 6i5
Krosnick,Jon. 1988. "Attitude Importance and Attitude Change."JournalofExperimental Social

Psychology 24:240-55.
Krosnick,Jon,andHowardSchuman.1988. "Attitude Intensity,Importance,andCertainty andSus-
ceptibility
toResponseEffects." JournalofPersonality and Social Psychology 54:940-52.
Lane, Robert.1962.PoliticalIdeology.New York:FreePress.
Lau, Richard,andDavid0. Sears. 1983."InducingApparently Self-interested
PoliticalPreferences."
American JournalofPoliticalScience27:233-52.
Lau, Richard,David 0. Sears,andTomJessor.1990. "FactorArtifact Revisited:SurveyInstrument
EffectsandPocketbook Politics."PoliticalBehavior12:217-42.
Lewis-Beck,Michael. 1985. "PocketbookVotingin U.S. NationalElections:Fact or Artifact?"
Lodge, Milton,KathleenMcGraw,and PatrickStroh.1989. "An Impression-Driven ModelofCan-
didateEvaluation."American PoliticalScienceReview83:399-420
Loftus,ElizabethF., StephenE. Feinberg,andJudith M. Tanur.1985."CognitivePsychology Meets
theNationalSurvey." AmericanPsychologist 40:175-80.
Milbum,Michael. 1987. "IdeologicalSelf-schemata and Schematically-Induced Attitude Consis-
tency."JournalofExperimental Social Psychology 23:383-98.
Millar,Murray, and AbrahamTesser.1986. "Effects of Affectiveand CognitiveFocus on Attitude-
BehaviorRelationship." JournalofPersonality and Social Psychology 51:270-76.
. 1989. "The Effectsof Affective-Cognitive Consistencyand Thoughton the Attitude-
BehaviorRelation."JournalofExperimental Social Psychology 25:189-202.
Mueller,John.1973. War,Presidents, andPublicOpinion.New York:Wiley.
Page, Benjamin,and RobertShapiro.1992. TheRationalPublic. Chicago:University of Chicago
Press.
Popkin,Sam. 1991. TheReasoningVoter.Cambridge: Cambridge UniversityPress.
Price,Vincent.1991. "PrimingIdeologicalLeanings:The Impactof Salience and Commitment."
ReporttotheNationalElectionStudiesBoardofOverseers, AnnArbor,Michigan.
Raaijmakers, J.,andRichardShiffren. 1981. "SearchofAssociatedMemory." Psychological Review
88:93-134.
Rivers,H. Douglas. 1988."Heterogeneity inModelsofVoterChoice."American Journal ofPolitical
Science.
Schuman,Howard,andStanleyPresser.1981.QuestionsandAnswers inAttitudeSurveys.NewYork:
AcademicPress.
Schuman,Howard,andJacqueline Scott.1987."ProblemsintheUse ofSurveyQuestionstoMeasure
PublicOpinion."Science236: 957-59.
Scott,Jacqueline.1989. "Conflicting BeliefsaboutAbortion:Legal Approvaland MoralDoubts."
Social Psychology Quarterly 52: 319-26.
Sears,David O., andRichardLau. 1983. "InducingApparently Self-interestedPoliticalResponses."
Sniderman,Paul, RichardBrody,and PhilipTetlock.1991. Reasoningand Choice. Cambridge:
CambridgeUniversity Press.
Sussman,Barry.1986. "Do BlacksApproveof Reagan?It Dependson Who'sAsking."Washington
Postweeklyed. 10 February.
Taylor,ShelleyE., and Susan Fiske. 1978. "Salience,Attention, and Attribution:Top of theHead
Phenomena." In AdvancesinSocial Psychology, ed. LeonardBerkowitz. New York:Academic
Press.
Tesser,Abraham.1978. "Self-generated Attitude Change."In Advancesin Social Psychology, ed.
LeonardBerkowitz.New York:AcademicPress.
Tourangeau,Roger,and KennethRasinski.1988. "CognitiveProcessesUnderlying ContextEffects
inAttitude Measurement." Psychological Bulletin103:299-314.
6i6 JohnZallerand StanleyFeldman
Tourangeau,Roger,KennethRasinski,NormanBradbum,and Roy D'Andrade. 1989. "Carryover

EffectsinAttitude Surveys."PublicOpinionQuarterly 53:495-524.
Tversky, Amos, and Daniel Kahneman.1982. "The Framingof Decisionsand thePsychologyof
Choice."In QuestionFramingand ResponseConsistency, ed. RobinHogarth.San Francisco:
Jossey-Bass.
Wilson,Timothy, Dana S. Dunn, Dolores Kraft,and D. Lisle. 1989. "The DisruptiveEffectsof
ExplainingWhyWe Feel theWayWe Do." In Advancesin Social Psychology, vol. 19, ed.
LeonardBerkowitz.Orlando,FL: AcademicPress.
Wilson,Timothy, and SarahHodges. 1991. "Attitudesas Temporary In The Construc-
Constructs."
tionofSocial Judgment,ed. AbrahamTesserandL. Martin.Hillsdale,NJ:Erlbaum.
Wyer,Robert,andJonHartwick.1984. "The RecallandUse ofBeliefStatements as Bases forJudg-
ments."JournalofExperimental Social Psychology20:65-85.
Wyer,Robert,andThomasSrull. 1989.Memoryand CognitioninIts Social Context.Hillsdale,NJ:
Erlbaum.
Zaller,John.1984. "Towarda Theoryof theSurveyResponse."Presented at theannualmeetingof
theAmericanPoliticalScienceAssociation,Washington, DC.
. 1990. "PoliticalAwareness,Elite OpinionLeadership,and theMass SurveyResponse."
Social Cognition8:125-53.
. 1992. TheNatureand OriginsofMass Opinion.Cambridge:CambridgeUniversity Press.
Forthcoming.

Feldman Zaller Article

Enviado por

Dados do documento

Descrição original:

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

Feldman Zaller Article

Enviado por

Direitos autorais:

Formatos disponíveis

A Simple Theory of the Survey Response: Answering Questions versus Revealing Preferences

Author(s): John Zaller and Stanley Feldman

Opinionresearchis besetbytwomajortypesof "artifactual"variance:hugeamountsof over-

Virtuallyall publicopinionresearchproceedson theassumption thatciti-

AmericanJournalofPoliticalScience,Vol. 36, No. 3, August1992,Pp. 579-616

tioned,theycall to minda sampleof theseideas, includingan oversampleof

'Given data fromonlytwo pointsin time,it is impossibleto distinguish

American Relations withRussia

Note: The exactquestionswere:"Some people feelit is important forus to try

sucherroris said to stemfromtheinherent of mappingone's attitudes

On moretypicalissues,as ConverseandMarkus(1979) argue,people'sattitudes

(1988), forexample,arguethatresponsestoattitude questionscan be understood

analysis,littleattention has beendevotedtothisquestion.If,however, oneturns

arounda centralvalueor idea andthatguidestheinterpretation ofnewinforma-

AXIoM1: The ambivalenceaxiom.Mostpeoplepossessopposingconsid-

We emphasizethattheconceptof consideration, as used in thisaxiom,is

head"on thebasis of thefirst

aboutthequestionyoujust answered,I'd likeyouto tellme

This probewas designedto elicita "memorydump"of theconsiderations im-

Beforetellingme howyoufeelaboutthis,couldyoutellme whatkindsof

Immediately followingtheseprobes,the interviewer rereadthe original

of response.The "prospective" or "stop-and-think"3 probes,on theotherhand,

Examplesof starcodes are "people shouldtryto get ahead on theirown, but

6Moreprobesweremade,butsincetheyweredirective, we makeno use ofthem.

0 76.9 83.5 78.8 63.2 72.2 71.1

0 75.0 91.7 81.4 64.9 85.2 72.3

0 60.2 51.4 63.6 25.9 24.3 24.7

Table 3. Effectof PoliticalAwarenesson Volumeof Open-EndedComments

Standardof living 2.3 2.9 3.5 3.5 3.2

remarksin thegivencell. These

have,all else equal, morethoughts aboutthatissuereadilyaccessiblein memory

Table 4. RelationshipbetweenAvailableThoughtsand Closed-EndedItems

Wave 1 Wave2 Combined

Note:Cell entryis Pearsoncorrelation

endeditemsin thesetestsareessentially dichotomies,8andthefragilenatureof

8A fewrespondents volunteered "itdepends"middleresponses,whichwereaccepted.

'?We would like to presentverbatim of whatthisand otherrespondents

357. Program/hous- 348. Program/health

ofthesecodescan be foundin theICPSR codebookforthe1987PilotStudy.

feelingthatjob guarantees are un-American, without raisinganyopposingcon-

at thetimeof makingan attitude report,as inthecase ofthevacillating teacher.

Evidencetending to supportDeduction7 maybe foundinthediscussionofracial

errorrates,we used item-level

Table 6. ResponseStabilityand Consistencyof Considerations

Job Government Aid to

Note:Cell entriesareproportion itemsfromwave 1 towave

and-think side, whichexplainswhytheshortfall fromexpectations is higherin

shipsin Table 6, as summarized by unstandardized regressionslopes,wouldbe greatest forthose

Table6 was calculatedoverbothwavesofthesurveyin orderto captureconsist-

issue was veryimportant to them)werequitesusceptibleto questionorderef-

'5Thisandotherrace-of-interviewer effectsmightbe interpreted as "socialdesirability"

2. Referencegroup effects.In a classic social psychologyexperiment,

This typeof questionordereffect is also explainablefromthemodel.Hav-

Low Awareness HighAwareness

Note:Model is shownin text;modelwas estimated acrossthe

shownin Table7, thestop-and-think testnotonlyfailedto increaseconsistency,

Table 8. The EffectofStop-and-Think

Note: Cell entriesare test-retest

findings and apparentreversals,because theymighthave been caused by the

measurement errormodels(see Achen 1975) and paysless attention to mental

Table 9. List of EmpiricalPhenomenaThat theModel Claims to Explain

1 Peoplewhoare,in general,morepolitically awarehavemoreconsiderations at the

long-termmemoryand retrievethemas required,ratherthan, as in our model,