An Impression-Driven Model of Candidate Evaluation Author(S): Milton Lodge, Kathleen M

An Impression-Driven Model of Candidate Evaluation Author(s): Milton Lodge, Kathleen M. McGraw, Patrick Stroh Source: The American Political Science Review, Vol. 83, No. 2 (Jun., 1989), pp. 399-419 Published by: American Political Science Association Stable URL: http://www.jstor.org/stable/1962397 Accessed: 23/02/2010 11:56

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=apsa.

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 [email protected].

American Political Science Association is collaborating with JSTOR to digitize, preserve and extend access to The American Political Science Review.

http://www.jstor.org AN IMPRESSION-DRIVENMODEL OF CANDIDATE EVALUATION MILTON LODGE KATHLEEN M. McGRAW PATRICK STROH State Universityof New York Stony Brook

T e describeand test two process models of candidateevaluation. The memory-basedmodel holds that evaluations are dependent on the mix of pro and con informationretrieved from memory. The impression-drivenmodel holds that evaluations are formed and updated "on-line"as informationis encountered. The results provide evidencefor the existenceof stereotypingand projectionbiases that renderthe mix of evidenceavailable in memorya nonveridicalrepresentation of the informationto which subjectswere exposed. People do not rely on the specific candidateinformation available in memory. Rather,consistent with the logic of the impression-drivenprocessing model, an "on-line"judgment formed when the informationwas encounteredbest predicts candidate evaluation. The results raise both methodologicaland substantive challenges to how political scientists measure and model the candidate evaluation process,

O ur primaryaim because citizens are exposed to informa- is to develop and test two competitive tion (be it party affiliation, character,or models of the candidateevaluation proc- policy stances)over an extendedperiod of ess. We focus on information,our general time, and some of that informationmust perspective being the information-proc- be stored in, and retrieved from, long- essing approach identified with Herbert term memory when an evaluationor vote Simon (1981, 1985; Newell and Simon is necessary. 1972). All contemporarypolitical science Despite the lack of explicit attentionto models of vote choice are information- memory in contemporaryvoting models, processing models in that candidate eval- they appearto share the common implicit uation is treated as a function of the mix assumption that the direction and of information (e.g., Brody and Page strength of candidate evaluation is in 1972; Campbell et al. 1960; Enelow and large part determinedby the "evidence" Hinich 1985; Kelley 1983; Kelley and availablein memory. Forexample, Kelley Mirer 1974; Lazarsfeld, Berelson, and and Mirer (1974) state the case for a Gaudet 1944). However, none of them memory-basedmodel of vote choice: 'The provide an accounting of the memory voter canvasses his likes and dislikes of processes involved in candidate evalua- the leading candidatesand major parties tion. Yet memory processesare criticalto involved in an election. Weighing each understanding candidate evaluation, like and dislike equally, he votes for the

AMERICAN POLITICAL SCIENCE REVIEW VOLUME 83 NO. 2 JUNE 1989 American Political Science Review Vol. 83 candidatetoward whom he has the great- problems found to undermineseriously est numberof net favorableattitudes" (p. the veridicall"and "representative"mem- 574). ory modelsrelates to cognitiveconstraints This memory-basedmodel of evalua- in the processingof information;that is, tion makestwo criticalassumptions about memoryoften reflectsone's prior expecta- human informationprocessing. First, the tions, with the bulk of the empiricalevi- information recalled by respondents is dence strongly implicatingthe impact of thought either to be a more or less veridi- prior beliefs on attention and memory cal representationof the informationthey processes (Fiskeand Taylor 1984). A sec- were actually exposed to (Kelley 1983) ond issue we considerfocuses on evalua- or-more reasonably-to reflect the sa- tive constraintsin the processingof new lient information they were exposed to information. Is the evidence culled from plus what they "reasonably"inferred memory a reflection of rationalization about the candidates(Conover and Feld- processesthat serve to "fit"the evidenceto man 1986; Granberg1985). Second-and "match"one's current evaluation of the from our perspectivemost critically-the candidate?Again, the empiricalevidence model assumes that citizens base their documentsan evaluativebias in informa- evaluationon the evidenceretrieved from tion processing, the basic finding being memory. Moreover, the methodologies that people are prone to recall evidence used to assess the validity of candidate that is consistent with current attitudes evaluation models also reflect the impor- (Chaikenand Stangor1987). tance of memory. The reliance on re- In additionto problemsrelated to mem- sponses to the Survey Research Center ory per se, we are concerned with the (SRC) open-ended like-dislike questions judgmentprocess itself: How do citizens and candidate issue proximity scales re- use whatever informationis available in flects assumptions that (1) the citizen is memory to inform an evaluation?To ex- willing and able to conduct a search for aminememory processes as they apply to relevant informationin memory, (2) this candidate evaluation, we consider two search will generate a veridical or (more competing information-processing plausibly) a representativesample of the models: the memory-basedmodel of the informationto which the citizen was ex- judgment process and the "on-line,"or posed, and (3) this available information impression-driven, model (Hastie and is then used to inform an evaluation and Park 1986; Lichtensteinand Srull 1987). vote choice. We test the validity of these assump- Memory-basedprocessing. The memory- tions by examiningtwo importantaspects based model of candidate evaluation of the memory-judgmentrelationship as aligns with common sense in positing a appliedto the process of candidateevalu- positive relationship between memory ation. The first aspect deals with memory and judgment.When a personis calledon retrieval processes: How much of what to make an evaluation, whether at the kind of candidateinformation do citizens ballot box or in reply to an interviewer's retrievefrom memory?The key question request,it is reasonableto assumethat the being, Is the "configuration"of evi- overall evaluation is based upon the ac- dence representedin long-term memory tual mix of pro and con evidenceavailable a veridical mapping or (more plausi- in memory. Commonsense, unfortunate- bly) a representativesample of informa- ly, is a fallible guide. A great deal of tion to which the citizen was exposed, or research indicates that there is typically is memory for a political candidate sub- little or no relationshipbetween the blend ject to systematic biases? One class of of pro and con informationavailable in

400 CandidateEvaluation

memory and the direction or strengthof haps, for the paucity of specifics in the evaluation (see Hastieand Park1986 for a SRC open-ended like-dislike questions review). Apparently, memory-based (Gant and Davis 1984). evaluation is too cognitively taxing, rely- Forthe memory-basedmodel of evalua- ing on the effortfulretrieval of specificin- tion, a summaryjudgment is thought to formation from long-term memory, the be "computed"from the specific memory weighing of evidence, and then the com- traces recalledat the time the assessment putation of a summary evaluation, all is called for. To the extent that this repre- common complaints leveled against ra- sents an accurateportrayal of the evalua- tional choice models in generaland issue tion process, it is arduous, time-consum- voting models in particular (Herstein ing, and unreliable,requiring more effort 1981). and attentional resources than all but Downsian man could or would be willing Impression-drivenprocessing. Given the to expend. On the other hand, the on-line failure to find unequivocal support for model of the judgmentprocess is psycho- memory-basedprocessing, a second class logically realisticin proposing that when of models has been proposed. Impression- asked to voice an opinion, people typical- driven, or "on-line,"processing occurs ly retrievetheir summary evaluation from when a judgmentis made as relevant in- memory, only dredgingup specificmem- formationis encountered.A key concept ory traces when pressed to give reasons in understandingthe on-line model of why they favor or oppose a candidateor evaluationis the notion of an "evaluation issue. On-line processing appears to be counter"or "judgmenttally"-what Hastie more cognitively efficacious than mem- and Park (1986) call a "judgmentopera- ory-basedevaluation, in that the individ- tor," and Wyer and Srull (1986) term an ual is formingthe judgment"on the fly"as information "integrator."It is conceived evidence is encounteredand updatingthe as a counterin workingmemory that inte- on-line evaluation immediately, without grates new information into a "running having to first store each piece of evidence tally" of one's current impression. What in long-termmemory and then later labo- makes on-line processing so psychologi- riously compute a summary evaluation cally compelling is that unlike the from whatever memory traces are still memory-basedmodels of the judgmental available. process, on-line processing does not tax Clearly,people sometimesrely on their the cognitivelimits of the humaninforma- memoryof likes and dislikesto informan tion processingsystem unduly. When ex- opinion, while at other times they can posed to new information, people can simply retrieve their on-line judgments. operatenaturally as "cognitivemisers" by What conditions produce impression- simply retrievingthe evaluation counter driven or memory-based information from memory, updating this summary processing? The critical mediating vari- tally, storing the new value in long-term able appears to be the individual'sproc- memory, and then in the name of cogni- essing objective or "goal"when informa- tive economy "forgetting"the actual tion is initially encountered.If, as is com- pieces of evidencethat contributedto the monplace, the person acquires informa- evaluation. Thus it is that people can tion with the explicit or implicitobjective often tell you how much they like or dis- of forming an impressionor making an like a book, movie, candidate, or policy evaluative judgment, impression-driven but not be able to recount the specific processingoccurs. If, on the other hand, whys and wherefores for their overall one's goal were to rememberas much in- evaluation-a partial explanation, per- formation as possible (as is typical of

401 American Political Science Review Vol. 83 laboratory tasks but thought to be rela- tween what is available in memory and tively rare in real world situations)-or, judgmentis dependentupon the citizen's perhapsmore likely, if one were exposed processingobjective when informationis to informationwith no specific objective encountered. in mind and only later asked to renderan evaluation of the target-then memory- based processingoccurs: one's evaluation ExperimentOverview would be based on whatever bits and pieces of informationare still availablein We set as our task an examinationof memory. the relationshipbetween a person'smem- At stake is how we model and interpret ory for a candidateand his or her evalua- the correlatesand pathways of candidate tion of that candidate. First, we examine evaluation. When, as here, the citizen is the content of memory and its impact on conceived as a bounded rationalist, it is the candidate evaluation, asking, What incumbent on us to construct models of kind of informationdo people remember candidate evaluation that do not require about a candidateunder different process- more capacity and processingskills than ing conditions? What relationshipexists human beings are known to possess. To between the mix of informationavailable the extent that the evaluation of political from memory and candidateevaluation? figuresis much the same as the evaluation Next we turn to tests of the memoryof ordinarypeople by psychologists, our based versus the impression-drivenmodel discipline's interpretationsof the hows of candidate evaluation. In line with a and whys of candidate evaluation and dynamic information-processingperspec- vote choice appear to demand too much tive, we incorporatesome of the most im- of the voter, and our techniqueof asking portant factors known to underlie the respondentsfor their recollectionof likes evaluationof political candidates.An im- and dislikes (as well as their recollection plicit assumptionof on-line processingis of the candidates'position on the issue that "firstimpressions" (that is, the initial scales used to develop proximity scores) input into the judgmentcounter) should may well be misdirected. be a particularlypowerful determinant of In line with contemporarythinking in judgments. Within the electoral context, the cognitive sciences,we focus on the ac- the political candidate'sparty affiliationis tual processes involved in decision mak- certainlya critical"first impression" factor ing. From the cognitive perspective, a and is accordinglyincorporated into the psychologicallyinformed model of candi- informationstream. Next, given the im- date evaluation must satisfy the con- portanceof "character"on the evaluation straintsof "boundedrationality," the most of political candidates,we systematically importantbeing that (1) the information manipulatedthe second piece of informa- in memory available for recall is rarely a tion our subjects'were exposedto with an veridical representationof what the citi- endorsementappraising them of the can- zen was exposed to but, rather, is cogni- didate'scompetence and integrity. Final- tively constrained by inferences drawn ly, we examinethe impactof a candidate's from the stereotypic default values of policy statementson the evaluationproc- one'srepresentation; (2) the interpretation ess. Presumably,the voter's on-line inte- and recollection of new information is gration of the candidate'sissue positions also constrained evaluatively, as people will serve as a better explanation of the are prone to rationalizenew information candidateevaluation process than will the to make it better conform to their current information that the voter can later evaluation; and (3) the relationship be- rememberabout the Candidate,

402 CandidateEvaluation

Experimental Design and Method tion of a campaign brochure of a congressman seeking reelection. They first A nonprobability sample of 422 non- read the five-page campaign brochure, student adults from Long Island, New which describeda CongressmanWilliams York were interviewedby trainedcollege of Troy, New York as a long-timeRepub- students in partial fulfillment of a class lican who worked actively in the Repub- project in the spring of 1987. Fifty-two lican party since the 1950s, had defeated percent of the subjects were male and the incumbentDemocratic representative 88% were white. The median family in- in 1980, and who had held the congres- come was $40 thousand. The sample was sional seat since then. (Williamsis a ficti- well educated, with 68% reportingsome tious person; the subjectswere not aware college or a college degree. While the of this). Following this one-paragraph respondents in this sample were better biography (which mentionedhis Republi- educated and enjoyed more income than can affiliation four times), the brochure the average U.S. voter, they also ex- cited 40 policy positions attributedto the pressed-more important for our pur- congressman,each of them a declarative poses-similar levels of political interest statementof the type, CongressmanWil- and correspondinglylow levels of politi- liams "favorsmajor cuts in federalspend- cal activity beyond voting. In terms of ing on social programs."All 40 policy party affiliation,29% of subjectsreported statementswere in fact selected from the being Republican, 35% Democrat, and Lodgeand Hamill(1986) inventory of pol- the remainderIndependent or of no party icy statements.Thirty had been evaluated affiliation. The experimentproceeded in earlier as Republicanitems by a similar six stages, as summarizedin Table 1. sampleof New Yorkers,while the remaining 10 were policies characteristicof the Stage1. Campaignbrochure. The subjects Democratic party. Thus Congressman were recruited to participatein a study Williams'soverall policy posture in the that had as its ostensibleaim the evalua- campaign booklet was 75% consistent

Table 1. Experimental Design

ExperimentalConditions Impression-driven Stagesand Tasks Memory-based Positive Neutral Negative Campaignbrochure Republicancongressman yes yes yes yes Endorsementby civic organization no yes (+) no yes(-) Dimensionfor statementsrating readable/not like/dislike like/dislike like/dislike Distractorword test yes yes yes yes Congressmanevaluation yes yes yes yes Policy position recall yes yes yes yes Policy recognition yes yes yes yes Finalquestionnaire Republican-Democraticrating of 60 policies yes yes yes yes Like-dislikeratings of remaining statements(no. cases) 60 20 20 20 Partisanknowledge, political interest, party ID, ideology, demographics yes yes yes yes

403 American Political Science Review Vol. 83

and 25% inconsistentwith his Republican or dislike the Congressman').Our pur- affiliation. pose here was to mimic the real world sit- For the experimentaltreatment we sys- uation in which people are thought to in- tematically varied the subjects'informa- tegrate new information spontaneously tion-processinggoal in reading the cam- into an evaluation (Hastie and Park paign brochure. Approximately one- 1986). quarter of the subjects (N = 100) were In short, in the three impression-driven randomly assigned to the memory-based conditions we attempt to trigger the im- condition in which they first read the pression formation process explicitly short biography describingCongressman while also providing some subjects with Williams as a Republicanand then read information (civic association endorse- the 40 policy statementsattributed to the ment) designedto bias theirinitial impres- congressman. To thwart the subjects in sion in a positive or negative direction. this condition from forming an impres- We expect the impressionformation proc- sion of the candidate, they were told to ess to guide subsequentmemory and eval- rate each policy statement in terms of uation processes-hence the designation readability-how "easy or hard" "the impression-driven.Conversely, the sub- statement[was] to understand"on a five- jects who evaluate the policy statements point scale. The stated rationale for this at the syntactic level (in terms of read- rating task was that the congressman's ability) are not expected to form an im- campaignstaff wanted to know "whether pressionand will, we expect, be forced to the averageAmerican voter would be able rely on informationavailable in memory to read and understandthe campaignbro- when called on to evaluate the congress- chure." man-hence the designation memory- The remaining322 subjects were ran- based. Whetherthe set of 40 policy state- domly assigned to one of three impres- ments in the campaign brochure repre- sion-driven conditions: a positive en- sentsmore than or less than a normaldose dorsement condition, in which the con- of exposure about a candidate is a moot gressman was described by the "Troy question here. What is clear is that while Civic Association"as a hardworking,"in- the amount of informationpresented here touch," honest representative;a negative about the congressmanis well within the endorsementcondition, where it was said bounds of long-term memory, it is well CongressmanWilliams had a poor attend- beyond the limits of short-termworking ance and voting record and was "out of memory. Given this all-too-familiarpre- touch"with his constituency;or a control dicament, people are expected to act conditionin which no endorsementby the "reasonably"-albeitnot by the dictatesof Troy Civic Association was provided.' unboundedrationality. They take short- The subjectsin these threeconditions then cuts, use heuristics, make guesses, and read the same 40 policy statementsin the otherwise compensate for the fact that campaign booklet as did those in the they must make do with less than all the memory-based condition but evaluated informationthey were exposed to. the statementsin termsof how much (on a five-point scale) they personally liked or Stage 2. Distractor task. After having disliked the congressman'spolicy posi- read and rated the statementsin the cam- tion. The subjects in these three impres- paign booklet, we had all subjects com- sion-drivenconditions were also instruct- plete a 44-item Readers'Digest vocabu- ed to try to form an overall impressionof lary test. This distractortest was designed the candidatewhile ratingthe issues ("Try to purgefrom short-termmemory the pol- to reacha judgmentof how muchyou like icy statementsmade in the campaignbro-

404 CandidateEvaluation chure and simulate the more natural modelsrelies on the recognitiondata. For- course of events where nonrelevantinfor- ty policy statements were presented in mation intercedesbetween exposure and serial fashion, each followed by a five- evaluation. Respondents were allowed point scale ranging from (1) 'Yes, Con- five minutes to complete the multiple gressmanWilliams definitely said this"to choice test. Verbaltest scoresdid not con- (5) 'No, he definitely did not say this." tributeeither substantively or statistically Twenty of the policy statements were to the models of evaluationand are there- "old"items, that is, policies that actually fore droppedfrom furtherdiscussion. appearedword for word in the campaign booklet and 20 were "new"policies, policy Stage 3. Evaluationof the congressman.2 statementswhich the congressmannever Followingthe distractortest, two types of made. Half of the old and half the new candidate evaluation were obtained. A statementswere consistentwith Congress- global evaluation question asked subjects man Williams' Republicanparty affilia- for an "overall impression"of the con- tion (policies judged characteristicof the gressmanon a five-pointscale that ranged Republicanparty) and half were incon- from "extremelyfavorable" to "extremely sistentpolicies (identifiedwith the Demo- unfavorable."Subjects were also asked to cratic party). rate CongressmanWilliams on the 24 trait adjectives used in recent SRC National Stage 6. Final questionnaire. Following ElectionSurveys. Two trait dimensions- the recognitiontask, subjectswere asked competence(e.g., hardworking,not qual- to characterizeall 60 policy statements ified) and integrity (e.g., compassionate, (the 40 from the brochureas well as the 20 dishonest)-were evident from factor new recognition test policies) on a five- analyses and consistent with data ob- point scale as characteristicof either the tained in prior research (e.g., Kinder Republicanor Democratic party. These 1986; Kinderand Abelson 1981; Markus ratings allow us to confirm the partisan 1982). Because the dimensionality of directionof the policy statements.Follow- political characteris not directlyrelevant ing these partisanratings, subjectsin the to our concerns here and because the memory-based condition were asked to 24-itemcharacter index and global evalu- rate all 60 policy statementson the like- ation measurewere highly correlated(r = dislike scale, while subjectsin the impres- .73), we combined the two into a single, sion-drivenconditions, having evaluated equally weighted measure of candidate the 40 statementsearlier in stage 1, now evaluation. Also included in this section rated just the 20 new policies. were three checks on the endorsement As a measureof political knowledgeall manipulation. Examination of the re- subjects were also asked to identify the sponse to these manipulationchecks indi- party affiliationof contemporarypolitical cated that this informationwas attended leaders and groups. The responsesto the to by the two groups receiving the en- leaders-and-groupsquestions were com- dorsements. bined with the partisan ratings of the 60 policy statementsto form an additive in- Stage 4. The recall of policy statements. dex of partisan knowledge. Each correct In a surpriserecall test, the subjectswere response (defined by sample consensus called on to recall "as accuratelyas pos- opinion, which also correspondedto the sible"the actualpolicy statementsthat ap- authors'"expert" opinion) added a point peared in the campaignbrochure, to the knowledge score, whereaseach in- correct response resulted in the loss of a Stage 5. The recognitionof policy state- point. Don't know or a missing response ments. Our major test of the processing was scored as zero,

405 American Political Science Review Vol. 83

The final series of questions tapped positions that they personally prefer ac- political interest, party identification, curately?Do subjectserroneously impute ideological persuasion, conventional po- ("project")preferred issue stancesonto the litical activities,and demographics,all us- candidate? ing the standardSRC formats. Next, having determinedthe content of available memory, we move to incorpo- Analytic Plan rate memory variables into a process model of candidateappraisal. In keeping We address two distinct questions in with this dynamic perspectiveon human our empiricalanalyses. The first is con- informationprocessing, we manipulateor cernedsolely with memory:How muchof measure "impression"variables at differ- what kind of informationis available in ent points in time. The manipulated"first memory? The second is concerned with impression"variables consist of the candi- the relationship between memory and date'spartisanship and the positive, nega- judgment:How is the informationavail- tive, or no endorsementof the candidate able in memory used in candidateevalua- madeby the fictitiousTroy Civic Associa- tion? tion in the brochure.Moreover, the sub- The analysis of the recognition data jects' ratings of their personal liking for will allow us to determinethe content of the candidate'sissue stances (under the available memory.3 The most stringent impression formation instruction)forces criterionwould be a veridicalrepresenta- them to revise their assessmentcontinu- tion-a "snapshot"-inwhich all 40 policy ously, presumablythrough the activation positions endorsedby the candidatewere of the evaluation counter. The summary correctly recognized. After eliminating measureof the liking for the 40 campaign this unrealisticpossibility, we turn to ex- policies is used as an indicatorof this up- amine in depth whether the information dated impression(hereafter referred to as available in memory is a representative the on-line evaluation). Finally, the subset of the candidate's message or dependent measure is candidate evalua- whether it reflects a systematically dis- tion (derivedfrom the traitratings and the tortedversion of his positions. Two mani- "overall"impression question). festations of biased memory are of interest to us here. The first is the extent to which voters create a stereotyped repre- What Is Available in Memory? sentation of the candidate.Looking at all 40 policies (those consistent with his Re- We begin with the question, 'How publican label and those inconsistent),is much of what kind of informationfrom there evidence of a stereotypingbias in the campaignbrochure leaves a recogniz- memory such that subjects correctly able trace in memory?Keep in mind that remember-and perhapserroneously im- to measure available memory 40 policy pute-more consistent (Republican)than statementswere presentedin the recogni- inconsistent (Democrat) policies to the tion task, 20 of them old items (a sample candidate?The second line of questioning of statementsthat had actually been pre- examines the impact of the individuals' sented in the campaignbrochure) and 20 own policy preferenceson memory for of them new policies (statementsthat had the candidate'spolicies in a direct test of not appearedin the brochure).Of the old projection bias (Brody and Page 1972; and new items 10 of each type of state- Conover and Feldman1986; Markusand ment were Republican,and 10 were Dem- Converse 1979). Under what conditions ocraticpolicies. Subjectswere requiredto are subjects likely to remember policy indicate whether CongressmanWilliams

406 CandidateEvaluation

Table 2. Recognition Task Responses As a Function of Experimental Condition and Partisan Knowledge

ExperimentalCondition PartisanKnowledge Memory- Impression- Issues based driven F(1,418) Low Medium High F(2,418) Republican Old (in brochure) Hitsa .55 .69 40.22*** .62 .64 .71 7.97*** Misses .23 .21 ns .23 .22 .20 ns New (not in brochure) False alarmsa .26 .30 4.91* .27 .28 .30 ns Correct rejections .45 .52 6.30** .47 .49 .56 6.88*** Democratic Old (in brochure) Hits .64 .55 ns .67 .63 .63 ns Missesa .18 .26 17.03*** .20 .24 .28 8.53*** New (not in brochure) False alarms .15 .20 8.25** .20 .19 .17 ns Correct rejectionsa .52 .60 5.47* .51 .55 .68 17.33*** Note: The values are the proportionof responsesof each type, calculatedseparately for type of issue (Repub- lican or Democrat).Within each group of nonindependentresponses (e.g., hits and misses)the datado not sum to 1.0 because don't know responsesare not included. The values in the F columns are the main effect F-statisticsand significancelevels (ns = not significant). aTheseresponses are the componentsof the measureof stereotyping. *p < .05. **p < .01. ***p < .001. had or had not endorsedthe policy in his rately. The results are summarized in campaignbrochure. Table 2. Four types of responses are possible Table 2 reportsthe mean percentageof when analyzing recognition memory each type of responseas a function of ex- (Murdock1982). There are two types of perimentalcondition -because therewere correct responses. A hit occurs when a no recognition differences among the subjectresponds, 'Yes, he said that"to an three impression conditions, those data old policy statement. A correct rejection are combined-and level of partisan occurs when a subject says no to a state- knowledge. The results clearly eliminate ment that had not appearedin the bro- the implausiblestandard that recognition chure. Similarly, there are two types of memory is a veridicalmapping of the in- errors. A miss occurs when a subject re- formationin the campaignbrochure: ap- sponds "No, he didn'tsay that"to a state- proximately40% of all responsesare er- ment that was actually presented in the rors or don't know answers. brochure,and a false alarmoccurs when a Our primaryconcern is with how one's subject errs by responding'Yes, he said information-processing goal (memory- that"to a new item. For the presentdata, based versus impression-driven)affects we tally the responsesfor the Republican memory for campaign information. As and Democratic policy statements sepa- expected,recognition varied as a function

407 American Political Science Review Vol. 83 of experimentalcondition. On threeof the Stereotyping. One heuristic device that four correct responseindicators (hits and individualscan use to deal with complex correctrejections), subjects in the impres- informationin the social world is through sion-drivenconditions outperformed sub- the use of stereotypes (Hamilton 1981; jects in the memory-basedcondition. Par- Lippmann 1922; Taylor and Crocker ticularlystriking are the differencesin hits 1981). Stereotype refers to an abstract for the party-consistentitems (55%as op- knowledge structure representingbeliefs posed to 69%). This patternis consistent about the members of social groups with a large body of research showing (Hamilton 1981). Once categorized, the that the simple act of forming an impres- individual "inherits"the representative sion facilitates memory for information characteristicsof that group. The stereo- (Fiskeand Taylor 1984). Recognitionalso typical cue evoked in the presentstudy is varied as a function of level of sophistica- Republican,as all subjectswere told that tion. Consistent with prior research the candidatewas a Republicancongress- (Lodge and Hamill 1986), those with man. higher levels of partisanknowledge tend- The stereotypemeasure is based on the ed to be more accurate (in terms of hits recognitionresponses that reflect the exi- and correct rejections) than those with tent to which the congressmanwas per- lower knowledge levels.4 ceived to be a prototypical Republican, In sum, it is clear that the amount of Specifically,the -measure consists of four policy information representedin mem- independentcomponents (as indicatedin ory differssignificantly from that present- Table 2). ed in the campaign booklet. Memory varies as a function of the processinggoal 1. hits to old consistent items (correct adopted when originally exposed to the recognition of Republican policies information,as well as the level of parti- from the brochure) san knowledge. Forthose who treatedthe 2I missesof old inconsistentitems (fail- policy statements in the campaign bro- ure to recognize Democratic state- chure as mortarfor constructingan over- ments from the brochure) all impressionof the candidate, the pat- 3. false alarmsfor new consistentitems tern of remembrancesis significantlydif- (incorrectattribution to new Repub- ferent from subjectsin the memory-based lican issues) conditions, who were restrained from 4. correctrejection of new inconsistent forming an initial impression.The simple items (correct rejection of Demo- (and, we believe, natural)act of integrat- cratic policies not in the brochure), ing the policy positions into an on-line candidate evaluation influences memory The stereotype measure was created by processes. summingup the numberof each of these At this juncture-still focusing on the four responses. The higher the score on information available in working mem- this measure,the strongerthe representa- ory -let us explore the more provocative tion in memory of the congressmanas a issue of whetherwhat is availablein mem- prototypicalRepublican. ory is a representativesample of the en- The results of the 4 (processingcondi- counteredinformation or whetherthere is tions) X 3 (partisanknowledge) analysis evidenceof systematicdistortions of real- of variance of the stereotypemeasure in- ity. Two manifestationsof biased mem- dicate that stereotypingwas significantly ory can be examinedsystematically with- more likely in the impression-drivencon- in this experimentaldesign; stereotyping ditions (averageM = 17.67) than in the and projection. memory-based condition (M - 14.49,

408 CandidateEvaluation

F[3,410] = 14.90, p < .001). In addition, Projection.In addition to stereotyping(a highly knowledgableindividuals engaged bias based on expectanciesderived from in more stereotyping (M = 18.88) than groupmembership cues), we also examine did those with moderateor low amounts the extent to which our subjects'memo- of partisan knowledge (Ms = 16.42 and ries are biasedby projectionof personally 15.34, respectively, F[2,410] = 25.77, p preferredissues onto the candidate (see < .001). The processing condition and Krosnick1988a for a review and critique knowledge variables did not interact. of the projectionliterature). Evidence for Analyses of the two separatedimensions projection typically rests on proximity of the stereotype measure-accurate scores (the differencebetween self-place- responses and errors-yielded identical ment and candidate placement on issue results. Thus, the highly significant preference scales). In contrast to this stereotypingdifferences observed here are proximity approach, we are interestedin not due to eitherenhanced accuracy or to the impact of personalpreferences on the misperceptionsalone but to both. attributionof issue positions to the candi- The most direct explanation for these date (as evident from "Yes, he said that" differencesin stereotypingis that individ- responsesin the recognitiontask). Given uals who see their information-processing experimentalcontrol over the content of goal as forming an impressionof the can- the campaignmessage and direct tests of didate must integrate a large number of memory accuracy, our projection meas- separate pieces of information-in this ures are considerablydifferent from pre- study 40 issues plus a few pieces of per- vious operationalizations. sonal information-into a coherent In order to assess projection,two ratio whole. For these subjectsRepublican pro- measureswere computedfor each subject: vides an initial cue on which to build this (1) the number of 'liked"-or positively impression.Apparently, individuals with evaluated-statements attributed to the higher levels of knowledge have devel- congressman in the recognition test, oped a partisanknowledge structurethat dividedby the total numberof policies the is readily available for organizing this subject liked (labeled positive memory); new information.In the process of form- and (2) the numberof "disliked"-ornega- ing an impression, some of the incoming tively evaluated-policy statements at- informationwill nicely fit this schema-a tributed to the congressman,divided by Republicancandidate espousing Republi- the total number of disliked policies can policies-and be retainedin memory (negative memory). These two variables (resultingin hits), while otherpieces of in- range from 0 to 1.00, with higher values formation-Democratic policies-will reflectinga greaterlikelihood of retrieving not fit. These incongruent policies are each item type. The use of the ratios con- more likely to be discarded because of trols for initial differencesin evaluations theirlack of matchto expectations(result- of the policies, resultingin greaterpreci- ing in the misses). Moreover, because sion in tests of memory biases.5Differen- stereotypes contain expected values tiating between positive and negative (Taylor and Crocker 1981), when faced memory allows us to examine whether a with the uncertaintask of teasing out old projectionasymmetry is evident in mem- from new issue statements, individuals ory, specifically,whether positive projec- rely on these "default"values to guide tion (or assimilation) is more common theirjudgmental process, thereby yielding and more powerful than negative projec- a high false alarmrate to the new Republi- tion (or contrast). In other words, the can issues and the high probabilityof re- subjectsshould be more likely to retrieve jectingnew Democraticissues. from memory policies they personally

409 American Political Science Review Vol. 83

Table 3. Projection As a Function of Experimental Condition and Partisan Knowledge

ExperimentalCondition Positive Memory NegativeMemory PartisanKnowledge Memory-based Impression-driven Memory-based Impression-driven Low .42 .57 .36 .46 Medium .40 .54 .37 .43 High .43 .57 .40 .43

Note: The valuesare the proportionof positivelyand negativelyevaluated issues attributed to the candidatein the recognitiontask. Highervalues reflecta greaterprobability of attributingeach issue type to the candidate. prefer than policies they dislike, and, personally preferredissues provides pre- more importantly, the positive memory liminaryevidence for a projectionbias in projection bias should be more strongly memory. However, projection entails related to measuresof candidateevalua- more than the individual'sown issue pref- tion than negative memory projection erences:affective orientationstoward the bias (Krosnick1988a). candidateare also important (Brody and Positive and negative memory were Page 1972; Markus and Converse 1979). first analyzed by a 2 (processingcondi- In the present context, subjectswho like tion) X 3 (partisan knowledge) X 2 the candidate should retrieve more posi- (memory measure: positive or negative) tive policies and fewer negative policies analysis of variance, with repeatedmea- than do subjects who dislike the candi- sureson the last factor. The resultsof that date. In orderto examinethis relationship analysis are summarized in Table 3. between biased memory and candidate Across all subjects, positively evaluated evaluation, correlationsbetween the ap- policies were more likely to be attributed propriate measures were computed and to the congressman(M = .48) than were summarizedin Table 4. negatively evaluated policies (M = .41, Note that three different measures main effect for memorymeasure F [1,416] related to "candidate evaluation" are = 30.20, p < .001). In addition,a main ef- analyzed: the manipulated endorsement fect for processing condition indicates (coded 1, 0, and -1), the on-line evalua- that the imapression-drivensubjects at- tion of all policy statements,and the final tributeda higherpercentage of both posi- candidate evaluation. In general, the tive and negative policies than did the results conform to expectations. In the memory-based subjects (F [1,416] = case of the positive memory ratio, the 39.19, p < .001). However, this main ef- more positive the on-line and the candi- fect is qualifiedby a conditionby memory date evaluations, the greater the likeli- measureinteraction (F [1,416] = 6.16, p < hood of retrieving personally preferred .01), which indicatesthat the two process- policies and attributingthem to the candi- ing groups differ more on the positive date. Conversely, those with more posi- memorymeasure than the negative mem- tive evaluations (stemmingfrom the en- ory measure.Finally, partisan knowledge dorsement,on-line evaluation,and candi- was not significantlyrelated to the mem- date evaluation) were less likely to re- ory measureseither as a main effect or in trievenegatively evaluated policies. Final- interactionwith processingcondition. ly, as would be predictedfrom the asym- The increasedlikelihood of retrieving metry hypothesis, the magnitude of the

410 Candidate1 4 Evaluation

Table 4. Relationships between Evaluation and Projection Measures

MemoryAvailability Evaluation Positive Memory Negative Memory Endorsement .04 -.10* On-line .33*** -.10* Candidate .22*** -.10* Note: Entriesare correlationcoefficients. The evaluationmeasures are coded so thatpositive values reflect the positive endorsementcondition and morepositive evaluations.Higher values on the memorymeasures reflect a greaterprobability of attributingpositively and negativelyevaluated issues to the candidate.Only subjectsin the impression-drivenconditions are included. *P < .05. **p < .01. ***p < .001. relationship between candidate evalua- pared to examine the conditions under tion and memory is strongerfor positive which the mix of liked and dislikedissues memory projection than for negative in memory predicts the direction and memory projection for two of the three strengthof candidateevaluation. We then measures(the exceptionbeing the weaker proceed to develop and estimatea model endorsementvariable). of the candidateappraisal process. To sum up the resultsof the analysesof availablememory, we find clear evidence The memory-judgment relationship. between that the informationretrieved from mem- Recallthat the crucialdistinction ory is neithera veridicalnor even a repre- the two models in termsof understanding sentative sample of the candidate'sissues the memory-judgmentrelationship con- positions. Rather, the memory represen- cerns the activation of the evaluation tation of the candidate'sissue positions is counter. When, as is typically the case systematically biased. The tendency to (Hastieand Park 1986), an individualac- createa stereotypicalmemory representa- quiresinformation with the goal of form- tion is particularlypronounced for those ing an impression(our impression-driven in the impression-drivenconditions, espe- conditions), the judgment counter is cially so for political sophisticates.Final- thought to be activatedimmediately, and stored ly, to close the circle, individuals'own the on-line summaryevaluation is information policy preferencesalso bias memory, as separately from the specific subjectsare more likely to attributeissues used as its inputs. Underthese conditions, they personallyliked to a positively eval- there is no reason to expect a correlation uated candidate.6 between the specific facts retrievedfrom memory and either the strengthor direction of judgment. On the other hand, if the information is acquired without the objective of forming an impression (our Candidate Evaluation: memory-basedcondition), the evaluation The Memory-Judgment Link counteris not activated,and no summary evaluation is stored in memory. When At this point we shift focus from mem- later asked to make an evaluative judg- ory per se to the candidate evaluation ment, the individual is forced to go into process itself. First, the memory-based long-termmemory to retrievepreviously and impression-drivenmodels are com- acquiredinformation and then use what-

411 American Political Science Review Vol. 83

Table 5. PredictingCandidate Evaluation: A Comparisonof Available Memory and On-LineEvaluation

ExperimentalCondition Impression-driven Variable Memory-based Positive Negative Neutral Available memory .20* .13 -.01 -.01 On-line evaluation .26** .54*** .46*** .51***

Note: Entriesare standardizedregression coefficients. The variablesare coded so that positive values reflect greaterproportions of positivelyevaluated issues available in memory,more positive on-lineevaluations, and more positive candidateevaluations, *P < .05. **p < .01. ***p < .001. ever memory traces are presently avail- The resultsare straightforward.As pre- able for computingthe evaluation. Under dicted, a significantrelationship between this difficult condition, there should be a available memory and candidateevalua- positive relationbetween candidate evalu- tion was obtained only in the memory- ation and the evaluative mix of informa- based condition. This resultcomplements tion retrievedfrom memory. the findings in nonpolitical domains To comparethe two processingmodels (Hastie and Park 1986; Lichtensteinand optimally, a summary measure of the Srull 1987) demonstratingthat the rela- evaluationcounter as well as a measureof tionship between memory and judgment the information available in memory is depends on processing objectives. More- needed. We use the on-line evaluation over, the on-line evaluation counter (the sum of the likes [1] and dislikes [-1] proves to be a much strongerpredictor of for all 40 of the candidate'sissue stances) candidate evaluation in the impression- as an indicatorof the tally in the counter, driven conditions. Finally, the significant as this score best representsthe subjects' (albeit weaker) relationshipbetween on- overall impressionof the candidateafter line evaluation and candidateevaluation their exposure to all of the policies. The in the memory-basedcondition suggests available memory measureis a ratio: the that we may have not been entirely suc- number of positively evaluated policies cessfulin inhibitingthe impressionforma- attributedto the congressmanin the rec- tion process among those subjects. (See ognition test, dividedby the total number Hastie and Park 1986 for a discussionof of positive and negative policies attri- the difficultyin creatingtasks that inhibit buted in the recognitiontest (Hastie and spontaneousimpression formation). Park 1986). A ratio of .50 correspondsto an equal numberof liked and dislikedpol- Candidateevaluation. Our model of the icies availablein memory, a score greater candidateevaluation process includesthe than .50 indicatesa greaterproportion of mainstayvariables of party identification, liked policies, and a score less than .50 re- ideology, and partisanknowledge, as well flects a greaterproportion of dislikedpol- as the endorsementby the civic associa- icies.7 These two measureswere used to tion (which was experimentallymanipu- predict candidateevaluation. The regres- lated to have three values-positive, sion analyses are summarizedin Table 5, negative, and neutral-coded 1, -1, and

412 CandidateEvaluation

Table 6. Estimatesof the CandidateAppraisal Model

Maximum Variable LikelihoodBeta StandardError t-Value StandardBeta On-line evaluation Endorsement 1.28 .92 1.39 .07 Party identification 3.07 .84 3.66 .20 Ideology 5.91 .87 6.79 .37 "Old"issue memory Endorsement .01 .01 1.53 .07 Party identification -.00 .01 -.24 -.01 Partisanknowledge -.01 .00 -2.87 -.13 Ideology -.02 .01 -1.62 -.09 On-line evaluation .01 .00 11.35 .60 "New"issue memory Endorsement .04 .02 1.45 .08 Party identification -.05 .02 -2.11 -.13 Partisanknowledge -.01 .00 -3.40 -.19 Ideology -.01 .03 -.54 -.04 On-line evaluation .01 .00 4.18 .26 CandidateEvaluation Endorsement .29 .04 6.63 .30 Party identification -.03 .04 -.64 -.03 Ideology .13 .04 2.89 .15 On-line evaluation .02 .00 6.59 .41 "Old"issue memory .21 .27 .76 .04 "New"issue memory .16 .10 1.64 .08*

Note: Chi-squared(3 df) = 4.98, p = .173; goodness-of-fitindex = .996; adjustedgoodness-of-fit index = .953; root mean squaredresidual = 4.079. *p < .05. **p < .01. ***p < .001.

0). These exogenous variables are ex- all of these variables are used to predict pected to be related to three components the dependentmeasure, candidate evalua- of the candidateevaluation process. The tion.8 first is the on-line evaluation, which we The rationalefor the model is straight- assume is the key variable influencing forward. Party identificationand ideol- subsequentmemory and judgments.The ogy should be related to the endogenous second is the ratio measurereflecting the variables because the candidate and his mix of positive and negative policies issueswere portrayedas Republican.Par- availablein memory (as in Table5). How- tisan knowledge is includedin the model ever, here we distinguish between old because, as reportedearlier, it reliablyin- issues (the proportionof liked policies at- fluences the representationof the candi- tributedto the candidatefrom those that date in memory. Most important,accord- actually appeared in the campaign bro- ing to the dynamicprocessing logic under- chure) and new issues (the proportionof lying the impression-drivenmodel, im- liked policies attributedto the candidate pressions formed early in the process that he never advocated, that is, false at- guide both the subsequentencoding and tributionsof his policy stances). Finally, retrievalof information,as well as evalu-

413 American Political Science Review Vol. 83

Figure1. The CandidateEvaluation servatives feel more positive about the Model candidate'sissue positions than do Demo- crats and liberals. Also, ideological per- Endorsement suasion has a significantdirect influence -.132 + . 295 on candidateevaluation. -A8S Memory related Partyll) .260.20 Partisanknowledge is inversely

On-Line * 0+S51 Candidate to both memory variables. Individuals Evaluation Evaluation with high levels of partisan knowledge 'Partsan,,J Knowledge . , were less likely to retrievea high propor- I-^.133 b "Olde"154 Memory tion of old issues that they personally liked, and also less likely to falsely project their preferrednew issues onto the candi- Note: Entries are standardizedregression coeffi- date than were individuals with lower cients. Only significantpaths are shown. levels of partisanknowledge. The intrigu- ing relationshipsthat exist between partisan knowledge and the various memory ations. Therefore, we expect that the measures are worth noting. Whereas manipulated endorsement will be posi- stereotyping (an expectancy-based,non- tively relatedto subsequentjudgment and affective bias) occurs more frequently memoryprocesses. The same logic applies among those with high levels of partisan to the effects of the on-line evaluation knowledge (Table 3), personal policy (wherea positive relationshipbetween the preferences have a stronger impact on on-line counter and the subsequentmem- memory among the less-knowledgeable. ory and evaluationmeasures is expected). This pattern of results is consistent with Although the two variables representing recent findings indicatingthat those with the informationavailable in memory are high levels of political expertiseare more not expectedto be significantlyrelated to likely to draw on partisanschemata when candidate evaluation in the impression- making political inferences(Conover and drivenprocessing conditions, particularly Feldman1986; Hamill, Lodge, and Blake in comparisonto the impactof the on-line 1985). The overall effect is an increased evaluation (as seen in Table 5), we esti- likelihood of creatinga stereotypicalrep- mate those links in order to determine resentationof the candidate.On the other whetherthe old and new (i.e., false) mem- hand, those with low levels of politicalex- ory measuresdifferentially predict candi- pertiseare more likely to rely on personal date evaluation. preferences, resulting here in a greater The parametersobtained from the full likelihood of projectingtheir policy pref- information maximum likelihood esti- erencesonto the candidate. mation of the recursive model are The pathways depictedin Figure1 con- reported in Table 6. The model itself is firm our expectations about impression- displayed in Figure1, which includes all driven processing.First, the endorsement path coefficientsthat are significantat the had a directand highly significantimpact p < .05 level (accordingto either one- or on candidateevaluations and is also asso- two-tailed significance tests, as dictated ciated with more positive on-line evalua- by the logic of the processes). Consider tions (althoughthis link does not reachan first the effects involving party identifica- acceptablelevel of significance). tion, ideology, and partisan knowledge. Considernow the relationshipbetween As expected,both party identificationand the memory and judgment variables. ideology exert significant effects on the Note the criticalresult: the highly signifi- on-line evaluation. Republicansand con- cant direct link between on-line evalua-

414 CandidateEvaluation

tion and candidate evaluation (a replica- General Discussion tion of the simple regressioneffects provided in Table 5 within the context of the Becausecitizens are exposedto political completeappraisal model). Moreover-as informationover extendedperiods of time predicted-no link exists between prefer- but are only periodically called on to ences for the candidate's actual issue voice an opinion or cast a ballot, a psy- stances that are availablein memory (old chologically realistic model of candidate issue memory) and candidateevaluation. evaluation must be able to account for There is, however, a modest but reliable what informationthe citizenhas available link between evaluation and "memory" in memory as well as how the mix of evi- for the issues falsely attributedto the can- dence in memory contributesto evalua- didate. In other words, whereasmemory tion. In this study we examinethe validity for the candidate'sactual policy positions of the assumption of a direct memory- has no discernible effect on candidate judgmentrelationship underlying contem- evaluation, the false belief that the candi- porary models of candidate evaluation date voiced personallypreferred stands is and (findingit wanting)propose an alter- directlylinked to evaluation.We interpret native methodof the candidateevaluation this linkage as evidenceof projectionbias process. The results raise both method- in the candidate evaluation process. ological and substantive challenges to Nevertheless,it should be noted that the how we measureand model the candidate magnitudeof this relationshipis relatively evaluation process. weak, consistent with such effects as Focusingfirst on memory-on what in- reportedby Markusand Converse(1979). formationabout the candidateis storedin According to the observed chi-squared memory and ostensibly available for statistic (X2[3 df, N = 319] = 4.98, p = making an evaluation-two findings .173), the model provides a reasonably emerge. First, people whose goal is to good fit to the data. The adjustedgood- form an impressionof the candidatetend ness-of-fit index coefficient is .953, indi- to constraintheir memory representation cating that the majority of the variance cognitivelyby attributingto the candidate and covariance is accounted for by the policy positions consistentwith their own model. A comparison of the estimated stereotypical preconceptions. Second, model to an alternativeprovides a strong- people tend to constrain the evidence in er method for assessing model validity memory evaluatively by attributing to (Anderson and Gerbing 1988). To test favored candidatespolicy positions they directlyour assumptionthat on-line eval- personally like. The effect of both biases uation is the critical component of the is to renderthe mix of evidence in mem- candidate appraisal process, we reesti- ory a nonveridical representationof the mated the model after eliminatingthe on- information subjects were actually ex- line evaluation-candidateevaluation link. posed to. The differencebetween the chi-squareds The primarythrust of this researchwas obtained from the full and "nested" directedtoward the evaluationprocess it- models was significant (chi-squareddif- self: how informationis integratedinto a ference[1 df, N = 319] = 40.77, p < .001; summaryevaluation. The major findings adjustedgoodness-of-fit index of the alter- of this study challengethe assumptionof nate model = .70), indicating that the a directrelationship between memory for model is significantlyimproved by con- a candidate'spolicy statementsand evalu- sideringthe path betweenthe on-line tally ation of that candidate. We find-in and candidateevaluation. direct contrast to the common sense as-

415 American Political Science Review Vol. 83

sumption underlying political science to the open-endedprobes are not specific models of candidate evaluation-that personal or policy attributes but rather when called on to form a generalimpres- "diffuse evaluative judgments" (Stokes sion of the candidate,people do not rely and Miller 1962; see also Gant and Davis on the specific issues available in mem- 1984), correspondingto our notion of re- ory. Rather,the evidencesuggests that the trieval of the on-line tally from memory. "judgmentcounter," activated when infor- Moreover,strong partisans (who, we sup- mation was initially encountered, is re- pose, are most likely to be chronicimpres- trieved and used to renderan evaluation. sion-drivenprocessors because of theirin- The results may appearparadoxical on terest in politics) are more likely to re- two levels. First, although two theoreti- spond to the open-endedprobes with non- cally meaningfulmemory biases (stereo- specific and generally affective expres- typing and projection) were identified, sions, whereasthose with but weak parti- neither is a strong predictorof candidate san feelingsare more likely to provide ex- evaluation. There was no evidence that plicit, detailedresponses (Gant and Davis stereotypingis relatedto evaluation,9and 1984). In addition, Kelley and Mirer the impactof projectionis relativelyweak (1974)report that the predictiveaccuracy (in comparisonto the magnitudeof the re- of the like-dislikeindex actually declines lationshipsbetween candidateevaluation as the numberof specificlikes and dislikes and the on-line evaluation, endorsement, reportedby the voter increases,a finding and ideology). However, this pattern is that is consistentwith our contentionthat indeed consistentwith the logic of the im- specific memory traces are not related to pression-drivenprocessing model: the on- evaluation. line tally should be -and is -a stronger In sum, what we see as our contribution determinantof evaluation than memory to this literatureis the idea that voters do variables. We are not concluding that not typically rely on their memory for issues involving the encoding,representa- specific issues to inform their evaluation tion, and retrievalof politicalinformation but instead call up their summary tally are unimportant.Rather, we would like when asked for an evaluation. This on- to make the case for greaterprecision in line processing is, at least in its broad specifying the conditions under which a strokes, psychologicallyrealistic in plac- strong relationship between memory ing minimal information-processingde- biases and candidate evaluation is likely mands on voters. Indeed, the model sug- to occur. gests an even simpler decision rule than More important, these results are in that thought to guide "thesimple of act of strong contrastto the majorityof studies, voting"(Kelley and Mirer1974), namely, which find a strong positive relation be- retrieve the on-line tally. To the extent tweena respondent'sspecific likes and dis- that this is an accurateportrayal of the ac- likes about a candidate and evaluation tual evaluation process, it challengesthe (e.g., Kelley 1983; Kelleyand Mirer1974; logic of askingrespondents their likes and Lau 1986; Miller, Wattenberg, and dislikesas indicatorsof the reasonsunder- Malanchuk1986). The numerousmethod- lying their vote choice. While political ological differences between this study scientists have long been aware that and investigationsthat rely on the NES voters often forgetand fabricatetheir past like-dislikemeasures compel caution. A partisanship, attitudes, and voting be- tentativeresolution of these disparatere- havior, this concern should be extended sults requires a considerationof the na- to the informationvoters reportas under- ture of the like-dislikesurvey responses. lying evaluationsof political candidates. A substantialproportion of the responses It is increasinglyevident that citizens

416 CandidateEvaluation

are not well informedand cannot engage measureof recognitionaccuracy, d', which assesses in the computations required by most memory"strength" as a functionof the ratio of hits theories of voter to false alarms (Murdock1982). Analyses of d' for rationality. We add to these data (conductedafter dichotomizingthe five- this the caveat that candidatememory is point recognition responses and considering the distorted by systematic biases that stem Republicanand Democraticissues separately)con- directly from impression-drivenprocess- firm our conclusions. The impression-drivensub- ing. However, from a cognitive perspec- jects were significantly more accurate than the memory-basedsubjects (Es [3,418] = 10.74 with p < tive, these consequences of impression- .001 and 2.54 with p < .05, for the Republicanand driven evaluation are a naturaland (dare Democraticissues, respectively).Similarly, subjects we say) "reasonable"way to compensate high in partisanknowledge were significantlymore for the severe limitationson humaninfor- accurate than subjects with low or moderate amountsof partisanknowledge (Fs [2,419] = 12.39 mation processingthat render models of with p < .001 and 4.73 with p < .01). unbounded rationality psychologically 5. The stereotypingand projectionbias measures unrealistic. are independent(r = .04 for stereotypingand positive projectionmemory, r = .15 for stereotyping and negativeprojection memory). 6. We do not mean to imply that stereotypingand Notes projectionare the only biasesthat have an impacton political candidate memory. At least two other An earlier version of this paper was presented at the biases could be contrastedwith the veridicallmem- 1987 Annual Meeting of the American Political ory"model. First,we suspectthat more salient or im- Science Association, Washington. We thank Stanley portant issues have an advantagein memory over Feldman, Shanto Iyengar, Jon Krosnick, Rick Lau, less importantissues (Krosnick1988b). In addition, and Bernadette Park for helpful comments. the orderin which informationis receivedtypically 1. Note that the absence of an endorsement means has an impacton memorysuch that more recentin- that the control-condition subjects received the same formationhas an advantageover older information information about the candidate as did the subjects (Anderson and Hubert 1963; Dreben, Fiske, and in the memory-based condition. Those two condi- Hastie1979). We are not able to examinethe impact tions differ only in processing objective. of salienceand order in the context of this experi- 2. At this juncture in the experiment we included a mentaldesign. See, however, McGraw,Lodge, and task order manipulation. Half the subjects (regard- Stroh 1988 for an extensionof the presentwork ex- less of processing condition) were asked for their aminingsalience and orderbiases. evaluation of Congressman Williams immediately 7. This availablememory measure is similar,but after the five-minute distractor task, followed by the not equivalent,to the projectionmeasures used in recall and recognition tasks (as outlined in the text the analyses reportedon p. 408 and Table 3. The and Table 1). This group probably best represents projection measuresreflect the probability of re- the normal course of events, where the citizen is first trievingeither liked and dislikedissues, controlling exposed to information and later called on to make for initial evaluationsof the issues when first en- an evaluation. The other half of the subjects com- counteredin the brochure.The memoryavailability pleted the two memory tasks before giving their measurereflects the mix of liked and dislikedissues evaluation. However, because preliminary analyses availablein memory. The correlationsbetween the did not reveal any reliable differences due to the task availablememory ratio, and positive-negativepro- order manipulation, all reported analyses collapse jectionmemory ratios are .50 and -.39, respective- the data across this factor. ly. 3. Use of the recognition data is necessary to con- 8. At this point, we focus exclusively on the sider the first issue of memory accuracy. For the sake impression-drivenprocessing condition and do not of consistency and brevity we rely on the recogni- estimatethe modelfor the memory-basedcondition. tion data, rather than the recall data or some com- Inclusionof that condition was critical for under- bination of recognition and recall, for all subsequent standing the impact of processing objectives on analyses. The analyses reported in Tables 3-6 were memory, as well as the relationshipbetween avail- also done substituting comparable measures derived able memory and judgment.However, the impres- from the recall data for the recognition-based mea- sion-drivenconditions represent the most common sures. The conclusions drawn from all the recall- and realisticinformation-processing mode, whereas based analyses correspond to those reported in this the memory-basedcondition was purposefullyarti- paper. A complete summary of the recall analyses is ficial. Therefore, the impression-drivencondition available from us on request. data provide the optimalcircumstances for estimat- 4. Signal detection theory provides a summary ing the parametersof the candidateappraisal model.

417 American Political Science Review Vol. 83

Moreover,the stereotypemeasure was not includ- FHamill,Ruth, Milton Lodge, and FrederickBlake. ed in the candidateappraisal model. As operation- 1985. 'The Breadth,Depth, and Utility of Parti- alizedhere, stereotypeis purelya cognitivemeasure, san, Class, and IdeologicalSchemas." American and there are no a priorireasons to expect that this Journalof PoliticalScience 29:850-70. affect-freevariable will (by itself)predict an evalua- Hastie, Reid, and BernadettePark. 1986. 'The Rela- tive judgment.Consistent with this reasoning,pre- tionship between Memory and JudgmentDe- liminaryanalyses revealed that stereotypingwas not pends on Whetherthe Task Is Memory-basedor relatedto candidateevaluations. For the sake of par- On-Line."Psychological Review 93:258-68. simony, stereotypingwas omittedfrom the complete Herstein,John A. 1981. "Keepingthe Voter'sLimits model estimation. in Mind: A CognitiveProcess Analysis of Deci- 9. Seen. 8. sion-Makingin Voting."Journal of Personality and Social Psychology40:843-61. References Kelley, Stanley.1983. InterpretingElections. Prince- ton, NJ: PrincetonUniversity Press. Kelley, Stanley,and Thad Mirer.1974. 'The Simple Anderson,James C., and DonaldW. Gerbing.1988. Act of Voting." American Political Science "StructuralEquation Modeling in Practice: A Review 61:572-91. Review and Recommended Two-Step Kinder,Donald R. 1986. "PresidentialCharacter Re- Approach."Psychological Bulletin 103:411-23. visited."In Political Cognition: The Nineteenth Anderson, Norman H., and S. Hubert. 1963. "Ef- Annual CarnegieSymposium on Cognition,ed, fects of ConcomitantVerbal Recall on Order RichardR. Lau and David 0. Sears. Hillsdale, Effects in Personality Impression Formation." NJ: Erlbaum. Journalof VerbalLearning and VerbalBehavior Kinder, Donald R., and Robert P. Abelson. 1981. 17:1-12. "AppraisingPresidential Candidates: Personality Brody, RichardA., and BenjaminI. Page. 1972. and Affect in the 1980 Campaign."Presented at 'The Assessment of Policy Voting."American the annual meeting of the American Political PoliticalScience Review 66:450-58. ScienceAssociation, New York. Campbell,Angus, Phillip E. Converse, WarrenE, Krosnick,Jon A. 1988a. "PsychologicalPerspectives Miller, and Donald E. Stokes. 1960. TheAmeri- on PoliticalCandidate Perception: A Review of can Voter. New York:Wiley. the Literatureon the Projection Hypothesis." Chaiken,Shelly, and CharlesStangor. 1987. "Atti- Presentedat the MidwestPolitical Science Asso- tudes and Attitude Change."Annual Review of ciation Meeting,Chicago. Psychology38:575-630. Krosnick,Jon A. 1988b. 'The Role of AttitudeIm- Conover, Pamela J., and Stanley Feldman.1986. portancein Social Evaluations:A Study of Pol- 'The Role of Inferencein the Perceptionof Politi- icy Preferences,Presidential Candidate Evalua- cal Candidates."In Political Cognition: The tions, and Voting Behavior."Journal of Person- Nineteenth Annual Carnegie Symposium on ality and Social Psychology55:196-210. Cognition,ed. RichardLau and David 0. Sears, Lau, RichardR. 1986. "PoliticalSchemata, Candi- Hillsdale,NJ: Erlbaum. date Evaluations, and Voting Behavior." In Dreben, Elizabeth K., Susan T. Fiske, and Reid PoliticalCognition: The Nineteenth Annual Car- Hastie. 1979. 'The Independenceof Item and negie Symposiumon Cognition,ed. RichardR, EvaluativeInformation: Impression and Recall Lauand David 0. Sears.Hillsdale, NJ: Erlbaum. OrderEffects in Behavior-basedImpression For- Lazarsfeld,Paul F., BernardBenelson, and Hazel mation."Journal of Personalityand Social Psy- Gaudet. 1944. The People's Choice: How the chology 37:1758-68. Voter Makes Up His Mind in A Presidential Enelow,James M., and MelvinJ. Hinich. 1984. The Campaign.New York:Duell, Sloan, and Pearce, Spatial Theory of Voting: An Introduction, Lichtenstein,Meryl, and Thomas K. Srull. 1987. Cambridge:Cambridge University Press. "ProcessingObjectives As a Determinantof the Fiske,Susan T., and ShelleyE. Taylor. 1984. Social Relationship between Recall and Judgment.' Cognition.New York: RandomHouse. Journal of Experimental Social Psychology Gant, Michael,and Dwight F. Davis. 1984. "Mental 23:93-118. Economyand Voter Rationality:The Informed Lippmann, Walter. 1922. Public Opinion, New CitizenProblem in Voting Research."Journal of York:Harcourt, Brace. Politics 46:132-53. Lodge,Milton, and RuthHamill. 1986. 'A Partisan Granberg,Donald. 1985. "AnAnomaly in Political Schema for Political InformationProcessing." Perception,"Public Opinion Quarterly49:504- AmericanPolitical Science Review 80:505-19. 16. McGraw, Kathleen, Milton Lodge, and Patrick Hamilton, David L. 1981. Cognitive Processes in Stroh. 1988. "On-LineProcessing in Candidate Stereotypingand IntergroupBehavior, Hillsdale, Evaluation:The Effects of Issue Order, Issue NJ: Erlbaum. Shambaugh Conference on Communication,

418 CandidateEvaluation

Cognition, and PoliticalJudgments and Action, Simon, Herbert.1981. TheSciences of the Artificial, Universityof Iowa. 2d ed. Cambridge:MIT Press. versity of Iowa. Simon, Herbert.1985. "HumanNature and Politics: Markus,Gregory B. 1982. "PoliticalAttitudes dur- The Dialogue of Psychology with Political ing an ElectionYear: A Reporton the 1980 NES Science." American Political Science Review PanelStudy." American Political Science Review 79:293-304. 76:538-60. Stokes, Donald E., and Warren E. Miller. 1962. Markus,Gregory B., and PhillipConverse. 1979. "A "PartyGovernment and Saliency of Congress." Dynamic SimultaneousEquation Model of Elec- Public Opinion Quarterly26:531-46. toralChoice." American Political Science Review and Jennifer Crocker. 1981. 73:1055-70. Taylor, Shelley E., "SchematicBases of Social InformataionProc- Miller,Arthur H., MartinWattenberg, and Oksana In Social The OntarioSym- Malanchuk. 1986. "SchematicAssessment of essing." Cognition: posium, ed. EdwardTory Higgins,C. PeterHer- Presidential Candidates." American Political man, and Mark P. Zanna. Hillsdale, NJ: Law- ScienceReview 80:521-40. rence Erlbaum. Murdock,Bennet B., Jr. 1982. "RecognitionMem- Wyer, Robert S., and Thomas K. Srull. 1986. ory." In Handbook of Research Methods in "HumanCognition in Its Social Context."Psy- HumanMemory and Cognition,ed. C. Richard. chologicalReview 93:322-59. Puff. New York:Academic. Newell, Allen, and HerbertSimon. 1972. Human Problem Solving. Englewood Cliffs, NJ; Prentice-Hall,

Milton Lodge is Professor, KathleenM. McGraw is Assistant Professorof Political Science, and PatrickStroh is a Ph.D. candidatein Political Science, State Universityof New York, Stony Brook, NY 11794,

419