INFORMATION TO USERS

This manuscript has been reproduced from the microfilm master. UMI films the text directly from the original or copy submitted. Thus, some thesis and dissertation copies are in typewriter face, while others may be from any type of computer printer.

The quality of this reproduction is dependent upon the quality of the copy submitted. Broken or indistinct print, colored or poor quality illustrations and photographs, print bleedthrough, substandard margins, and improper alignment can adversely affect reproduction.

In the unlikely event that the author did not send UMI a complete manuscript and there are missing pages, these will be noted. Also, if unauthorized copyright material had to be removed, a note will indicate the deletion.

Oversize materials (e.g., maps, drawings, charts) are reproduced by sectioning the original, beginning at the upper left-hand corner and continuing from left to right in equal sections with small overlaps. Each original is also photographed in one exposure and is included in reduced form at the back of the book.

Photographs included in the original manuscript have been reproduced xerographically in this copy. Higher quality 6" x 9" black and white photographic prints are available for any photographs or illustrations appearing in this copy for an additional charge. Contact UMI directly to order.

University Microfilms International A Bell & Howell Information Company 300 North Zeeb Road, Ann Arbor, Ml 48106-1346 USA 313/761-4700 800/521-0600

Order Number 0211208

Associative memory structure and the evaluation of political leaders

Reed, David Russell, Ph.D. The Ohio State University, 1991

Copyright ©1991 by Reed, David Russell. All rights reserved.

UMI 300 N. Zeeb Rd. Ann Arbor, MI 48106

Associative Memory Structure and the Evaluation of Political Leaders

Dissertation

Presented in partial fulfillment of the requirements for the degree of Doctor of Philosophy in the Graduate School of The Ohio State University

By

David Russell Reed, B.A.

»»»>►>► «««««

The Ohio State University 1991

Dissertation Committee: Approved by

Herbert F. Weisberg

Jon A. Krosnick Co-chairs

William G. Jacoby Department of Political Science Copyright by

David Russell Reed 1991 Acknowledgments

1 owe a debt of gratitude to many people, who supported me in this endeavor and in my earlier studies. My first debt lies with Lucinda Callender, David Leuthold, and Dean Yarwood. They gave me my start in the study of politics and encouraged me to continue with graduate work. My gratitude also goes to Bill Jacoby, who taught me about the study of political attitudes and the roots of public opinion. He oversaw my early exploration of trait judgments and did much to help me realize their importance. Jon Krosnick has been so instrumental in the development of my thoughts on the psychological underpinnings of trait structure, its measurement and meaning, that I may only begin to acknowledge all that I owe him. He also spent many hours helping me design and interpret the studies included here, and it was through his auspices that I was able to obtain lab space to conduct them. Herb Weisberg has been my mentor in the true sense of the word—a guide, friend, and counselor on this long search. I need say no more; no matter what I could say, it would be inadequate. I owe Marijke Breuning a great deal. She was always there, with a sympathetic ear, to talk with me about the problems I encountered and to help me solve them. My fellow travelers, John Bruce and Charles Smith, kept my spirits up throughout the months when progress was slow and the writing difficult. My thanks also go to the Ohio State University Graduate School, which provided a grant to help offset a part of the costs of this research. V ita

January 27, 1962 Bom - Alton, Illinois

1985 B. A., University of Missouri at Columbia; Columbia, Missouri

1986-1987 Graduate Teaching Assistant, Department of Political Science, The Ohio State University, Columbus, Ohio

1987-1991 Graduate Research Assistant, The Polimetrics Laboratory for Social and Political Research, The Ohio State University, Columbus, Ohio

Publications

Niemi, Richard G., David R. Reed, and Herbert F. Weisberg. 1991. "Partisan Commitment: A Research Note." Political Behavior. Forthcoming.

Fields of Study Major Field: Political Science

iv Table of Contents

Acknowledgements ...... ii V ita...... iv List of Figures ...... vii List of Tables ...... viii

Chapter Page I. Introduction ...... 1 II. Candidate Trait Evaluation...... 9 Understanding the determinants of evaluation 9 A question of structure...... 1 2 The study of structure and trait judgments 1 6 Summary...... 24 Notes...... 26

III. Accessibility and Trait Evaluation...... 27 Accessibility in social ...... 2 8 Reaction time as a measure of accessibility...... 3 2 Accessibility in ...... 3 5 Conclusions—The need for accessibility...... 45 Notes...... 4 9

IV. A Dimensional Analysis of Memory Structure...... 5 1 Introduction ...... 5 2 M ethod...... 5 5 Plan of analysis...... 5 8 Model development...... 5 9 Exploratory factor analysis...... 6 4 Parameter estimates...... 7 4 Discussion ...... 8 1 Notes...... 8 4

v V. An Experimental Approach to Assessing Trait Structure ...... 109 Introduction ...... 110 Priming as a test of relatedness...... 112 Procedure ...... 116 A typology of trait relationships...... 121 A check on non-word and same-word differences...... 124 Testing differences in response times...... 126 Negativity, positivity, and trait structure...... 137 Conclusions ...... 140 Notes...... 144

VI. Political Involvement and Trait Structure...... 155 Introduction ...... 155 Televison news and accessibility...... 157 Involvement and trait structure...... 161 Measuring involvement...... 164 Involvement and trait structure: accessibility... 165 Involvement and trait structure: evaluation.....169 Discussion...... 173 Notes...... 177

VII. Conclusion ...... 184 Traits and the political system...... 184 Traits and the individual...... 191 Implications—structure and accessibility...... 194 Implications—positivity and negativity...... 197 Implications—the sources of trait structure ...... 201 This study and future research...... 201 Conclusion ...... 205 Notes...... 210

Appendix 1: Questionnaire...... 211

Appendix 2: Regression Estimates...... 222

List of References...... 225

vi List of Figures

Figure Page 4.1 An example of a network of activation...... 8 7

4.2 The factor analytic view of a network of activation...... 8 8

4.3 The 24 traits used in the study...... 89

4.4 An example of how the stimuli appeared on the computer screen...... 90

4.5 Nesting the various models...... 9 1

5.1 Trait words used in Study 2...... 147

5.2 Artificial words used as primes in Study 2...... 148 List of Tables

Table Page 4.1 Exploratory factor analysis of the trait attributions ...... 9 2

4.2 Exploratory factor analysis of the standardized latencies...... 9 6

4.3 Assessing the fit of the five models...... 100

4.4 Maximum likelihood estimates of the loadings for the traits...... 102

4.5 Non-standardized loadings of the negative traits on the negativity factor...... 104

4.6 Standardized loadings of the negative traits on the negativity factor...... 105

4.7 Maximum likelihood estimates of the error variance for the traits...... 106

4.8 Estimates of the Inter-factor Correlations...... 108

5.1 Categorizing the relationship between various primes and the trait "Hardworking"...... 149

5.2 Means for the reaction time measures, Comparing artificial-word, same-word, and other trials...... 150

5.3 Expectations for the hypothesis tests under the different models of trait structure ...... 151

viii 5.4 Differences in reaction times due to primes...... 152

5.5 Differences in reaction times due to primes, Evaluatively incongruent and congruent trials 153

6.1 Indicators of political involvement...... 178

6.2 Interaction of priming effects and involvement...... 179

6.3 Comparing low and high involvement: priming...... 180

6.4 Comparing low and high involvement: the fit of the models to the evaluative judgments...... 182

6.5 Comparing changes in fit...... 183

ix I. Introduction

The list is endless: liberal, experienced, reactionary, a man of the people, crook, front-runner, in labor’s hip pocket, trigger happy, Ronnie Ray-gun. These and other labels are applied to political candidates by their foes or by their own campaign organization, by the media or by the average citizen. Some convey subtle shades of political meaning, others paint with the broadest of strokes. All are used to define candidates, to identify them as worthy of support or of scorn. All are a part of candidate evaluation, the process by which voters and potential voters decide who they like and how much, and if it is even worth voting at all. Voters consider not only the partisanship or policy positions of a candidate, they also want to know what kind of person he is. In a culture where people are often judged by their handshake or by the amount of eye contact they make in a conversation, it is no surprise that voters consider the character of the candidate to be important in and of itself. As it has often been pointed out, this tactic is a reasonable one for voters to use (c.f., Glass 1985; Shabad and Anderson 1979). There are no guarantees that, once elected, the promises made will be acted upon, the planks in the platform built into a more permanent structure. But, or so the popular perception goes, if a person is honest, or a hard worker, or possesses whatever trait is deemed to be important to the performance of presidential duties, that is not likely to change overnight. While most people believe, then, that the 'leopard doesn’t change his spots,' few actually get close enough to see the spots themselves. At best, they see different images of the candidate on the evening news. Often, they only see a part of the picture and, through choice or necessity, must guess at the rest. Some aspects of a candidate's character remain ambiguous, perhaps as the result of deliberate cover-up. Voters were left to wonder, during the 1988 nomination campaign, if Gary Hart the victim of a hounding press or blatant womanizer. Frequently, there are conflicting explanations of a candidate's behavior. How were voters, watching the 1988 presidential debates, to interpret Michael Dukakis' answer to a lurid question about violence against his wife: as cold and uncaring or as rational and level-headed? The process of inference is not limited to perceptions concerning the character of the candidate; issue positions are also "filled in" by voters and both are linked to partisanship (Conover and Feldman 1989, 1986; Rahn 1989; Lodge, McGraw and Stroh 1989). 1 In some cases a voter may consciously or unconsciously sample from the information available, looking at only enough to be reasonably confident of having made a good choice, inferring the pattern of the unexamined information from that readily available or most important. Whether we characterize this process as procedural (Simon 1985) or as cognitive miserliness (Fiske and Taylor 1984), the upshot is that at least some of the information a citizen has about the candidate's personality, some of the voter's , are the product of inference from limited information. Voters also take shortcuts in applying standards of evaluation to political leaders. Just as they do not have the time, energy, or willingness to seek out all relevant information about leaders, they also do not seek to apply every relevant standard in making electoral choices. Partisanship, ideology, and past record have been seen as shortcuts, or perhaps an evaluative shorthand, of this sort (Downs 1957; Campbell, Converse, Miller, and Stokes 1960; Converse 1964; Key 1966; Fiorina 1981). These two phenomena are central to our understanding of political decision making. What do people want to know about their leaders and which pieces of that information actually influence how citizens vote? Ultimately, the answer we arrive at has as much to say about normative questions of democracy in America as it does about individual voters.

Accessibility and evaluation. The expectations we may draw from past findings about evaluation and choice processes indicate that the information most likely to be used is that which is accessible in memory. (Fischoff, Slovic and Lichtenstein 1980; Higgins and King 1981; Tversky and Kahneman 1981; Iyengar et al. 1984; Iyengar and Kinder 1987; Rahn 1989; Krosnick and Kinder 1990)2 The importance of the accessibility concept to this discussion is that it introduces a new and important notion of a dimensional structure, very different from the simpler types of dimensions or factors commonly found in political analyses. As it has been used by cognitive and social psychologists, cognitive structure means something more than dimensions or categories. As outlined by Anderson (1983), one way of discussing the structure of memory is as a system of nodes in a network connected by associational links. These links, over which activation (i.e., increased ease of accessibility) spreads, are established over time as different nodes are made active in conjunction in working memory. Over repeated conjunctions, a "transient working memory structure will be turned into a permanent long-term memory trace" (Anderson 1983, 262; emphasis in original).

The structure and content of this document This dissertation takes as its focus structure in this latter sense. In it, I delineate a theory of cognitive organization that will allow the assessment of connections among candidate character judgments in the aggregate. The obvious place to start such an assessment is in the pattern of accessibility among cognitions. Working from a theory of associative memory structures as developed in cognitive and social , I apply the notion of networks of accessibility. These networks, which shape the relative accessibility of stored cognitions, are measured using the response latencies (e.g., reaction times) of subjects making judgments on a series of personality traits. 5

Chapter Two reviews more thoroughly the literature on candidate effects in voting and evaluation. In it, I discuss the history of research on candidate perception and evaluation, noting the special place of personality characteristics in this research. Much effort has gone into identifying the specific character traits most frequently applied to political leaders by citizens. The last decade has seen attempts to identify the structure of those traits, in terms of patterns of covariation in evaluation. I refer to the type of structure identified in this manner as the evaluative trait structure. This search for structure is an attempt to specify the evaluative standards applied to the personalities of leaders. I find this attempt wanting, due to its failure to answer the most compelling questions about traits and trait structure. Chapter Three is my attempt to formulate another approach to the identification of trait structure and trait standards. In doing this, I rely on a model of associative memory borrowed from work in cognitive . This memory model specifies that the relationship between traits may be measured using their relative accessibility. I refer to structure identified using this type of relationship as the accessibility or activational structure. Related traits act as primes for one another; thinking about or even seeing one makes others more accessible; they become activated. Those traits for which this effect is observed are related. The special nature of this relationship has major implications for the development and use of evaluative standards in processing political information. The fourth chapter uses this theory for two purposes. First, the notion of accessibility networks is used to identify the accessibility structure of personality traits for several political figures in a group of subjects. Then, this structure is compared to the evaluative structure found in those same subjects. It is found that the same model of the relationship between the traits is required for both types of structure. This model is the same as that found in analyses of the evaluative structure in national samples, consisting of Competence, Leadership, Integrity, Empathy and a Negativity dimension. Chapter Five uses an experimental technique to assess activational structure. Subjects are primed with trait words and their response time while making judgments on other traits is measured. Judgments primed with related traits should be made faster than those primed with less related or unrelated words. The analysis reveals additional confirmation of the model. Also found are previously unexpected inhibiting effects when a positive trait is primed with a negative trait. In the sixth chapter, the relationship between trait structure and political involvement is explored. A related line of research in cognitive organization employing the notion of political schemata has shown differences based on expertise. Those with higher levels of involvement should tend to unitize the traits, combining them into broader categories. Also, recent work by Iyengar and Kinder (1987), employing similar notions of accessibility and priming, has shown that political involvement mediates the effect of some types of priming. Expectations are confirmed, in both the activational and evaluative structure, for one of the leaders evaluated by the subjects, but disconfirmed for another leader in the evaluative structure. An explanation for this is sought in the nature of how the two structures are believed to be formed. In the final chapter, I discuss the implications of the theory of associative memory structures, and my specific findings, for our understanding of the American polity. I argue that traits, as standards of evaluation, are important not only in assessing how citizens behave, they are also important to our understanding of how leaders behave and the type of policies endorsed by elites. I close by discussing the implications of my findings for several aspects of political cognition. 8

Notes

1. In an attempt to reduce the problem to a manageable size, I would like to restrict the discussion, for the most part, to the area of candidate-centered judgments, leaving aside considerations of issue stance and partisanship. Of special interest are evaluations based on personality traits and personal characteristics, but not including judgments based on the physical appearance of a candidate. There are several reasons behind this decision. First, the areas of partisanship and issue-based evaluations seem to have received more attention than has candidate-based evaluation, and so I believe more resources should be devoted to its development. Furthermore, attempting to include all of these topics in the current research project would stretch available resources too thin for adequate coverage of any. Although issue and party considerations have an effect on and are affected by trait evaluations, the study of one class of determinant by itself is an essential first step in understanding the entire picture.

2. For a general discussion of accessibility aimed at a political science audience, see Ottati and Wyer (1990). II. Candidate Trait Evaluation

Understanding the determinants of evaluation If we are to understand how voters evaluate candidates and, ultimately, how they make their vote choice, it is essential that we ask and answer several questions about different aspects of evaluation.

Identifying the determinants. First, which specific judgments about the candidate have an effect on overall evaluation? It helps us very little to know that people judge a candidate to be patriotic, or in favor of nuclear power, if that judgment has no effect on final vote choice. Since the publication of The Voter Decides (Campbell, Gurin, and Miller 1954), students of American politics have identified three major classes of determinants of vote choice, three major factors in the process of candidate evaluation. The first of these, partisan identification, has to do with feelings of attachment to a group object (Campbell et al. 1960) or perhaps a retrospective evaluation or running tally of the performance of the two parties (Fiorina 1981). The second type of consideration on which people base their evaluations of candidates is issue position. Where do the candidates 10 stand on abortion, gun control, aid to the homeless, or any matter of public policy? Finally, the personal and physical qualities of the candidate also have an effect on political choice (Kinder 1986; Kinder and Abelson 1981; Markus and Converse 1979; Page and Jones 1979; Page 1978; Miller and Miller 1976; Kagay and Caldiera 1975). It matters to voters not only that the potential leader is of the correct party, or that he has come out in favor of the "right" policies, but also that the candidate is the right kind of person to lead the nation. In a country where people are often judged by their handshake or by the amount of eye contact they make in a conversation, it is no surprise that voters consider the character of the candidate, the personality presented in ads, press releases, and speeches, as important in and of itself.

The relationship among the determinants. An understanding of how the types of determinants are interrelated is also imperative. What effect do they have upon each other, or do they merely share a common source? Early attempts to explain the vote (e.g., Lazarsfeld, Berelson, and Gaudet 1944) found the origins of attitudes and the vote in a limited set of demographic variables. In the funnel of causality, described by Campbell, et al. (1960), the demographic characteristics located at the wide end of the funnel also played a major role in establishing later attitudes toward parties and candidates. 11 The three major categories of determinants do not act independently to shape overall evaluation, they are interrelated in a complex web not yet fully understood (c.f., Page and Jones 1979; Markus and Converse 1979). For example, they are linked in ways that cause persons to 'remember' different things about a candidate based on the presence or absence of information about party affiliation, even to the point of 'remembering' information that they were not actually given (Rahn 1989; Lodge 1989). Issue positions are also "filled in" by voters (Conover and Feldman 1989; 1986; Rahn 1989; Lodge, et al. 1989).

The relationship within a class o f determinants. To appreciate fully the role of a class of determinants, it is also essential to penetrate the inner structure of the class. To use a more technical term, we need to grasp the dimensional structure of party, issue, or personality evaluation. These dimensions constitute a picture of the organization of political evaluation. This approach has progressed furthest, or at least has had the most resources devoted to it, in connection with judgments on policy issues. Here, it has taken the form of looking at the ideological constraint present in mass belief systems (Converse 1964; Marcus, Tabb, and Sullivan 1974; Nie and Anderson 1974; Nie, Verba, and Petrocik 1976; Erickson 1979). This search for structure has involved more than simply outlining the consistency among issue attitudes in the mass public. As Converse himself states, it is a search for "some superordinate value or posture toward man and 12 society" that provides an organizing theme to issue attitudes (1964, 211). The shape of the inner dimensions of specific judgments or attitudes has implications for the process of inference as well, as work on cognitive economy in political evaluation suggests (Conover and Feldman 1989, 1986, 1984; Hamill and Lodge 1986; Lodge and Hamill 1986; Hamill, Lodge, and Blake 1985). How does a candidate's stand on one issue influence the perception of his stand on other issues? If a leader is seen as honest, does that have implications for judgments about competence? The process of 'filling in' that occurs as people make evaluations using limited information occurs within classes as well as across them.

A question of structure This dissertation is an effort to gain insight into the last of these three points, the relationship of elements within a class of determinants. Two conceptions of structure are at the heart of this work. First, I wish to know about structure as it signifies an organizing force in political life, along the lines of ideology or partisan conflict. Secondly, I deal with structure at a more concrete level, in terms of the interdependence of specific elements of evaluation. In this sense, I wish to know which specific judgments affect which others, and how they do so. In an attempt to reduce the problem to a manageable size, I would like to restrict the discussion, for the most part, to the area of candidate-centered judgments, leaving aside considerations of issue 13 stance and partisanship. Of special interest are evaluations based on personality traits and personal characteristics, but not including judgments based on the physical appearance of a candidate. There are several reasons behind this decision. First, the areas of partisanship and issue-based evaluations seem to have received more attention than has candidate-based evaluation, and so I believe more resources should be devoted to development of this area of research. Furthermore, attempting to include all of these topics in the current research project would stretch available resources too thin for adequate coverage of any. Although issue and party considerations have an effect on and are affected by trait evaluations, the study of one class of determinant by itself is an essential first step in understanding the entire picture.

Candidate trait evaluation As Asher discusses in his 1983 “state of the discipline” review, until the 1980’s research on many aspects of candidate-centered evaluation and perception had not kept pace with research on partisanship and issue attitudes (Asher 1983, 360). Of concern to Asher is the paucity of research on the development of candidate images; this is a concern that 1 share. There has been no shortage in the supply of descriptive studies documenting the effects of candidate trait evaluations on voting, of demographic variability in trait judgments and cognitions, or of the effects of other variables on candidate image (cf. Stokes 1966; Sigel 1964; 1966; Shabad and Anderson 1979; Kinder, et al. 1979; 1980; Kinder 1986; Glass 1985). These studies have established that candidate trait evaluation has a fairly constant structure. The dimensions of candidate trait evaluation (i.e., which trait judgments have the strongest relationships with each other) are usually found to be the same at different points in time (see Kinder and Sears 198S; Kinder 1986; Kinder and Fiske 1986). More difficult to locate are explanations of how candidate images evolve (compare this to the fascination for monitoring the evolution of partisanship across multiple generations) or why certain candidate traits should be expected to 'go together1 in peoples' minds (compare this to the decades long discussion of issue attitude constraint; but see Kinder 1986).

The need to go beyond description. These "how" and "why" questions are of major importance if we are to be able to claim we understand candidate evaluation or vote choice. It is not enough to say that citizens like a candidate whom they judge to be competent or honest. If we are unable to speculate upon the reasons for the coalescence of specific trait judgments into these dimensions, we will have to start from scratch every time the situation changes. A theory, containing an explanation of the phenomena that includes answers to the why's and how's, is necessary. Of course, theories abound. One early attempt to develop a theory of candidate perception, to explain how citizens formulate impressions and images of presidential candidates, can be found in the work of Nimmo and Savage(1976). Drawing on the work of Tagiuri (1958), they outlined a theory of candidate perception 15 consisting of three basic components: the object of perception, the perceiver, and the situation in which the perception is taking place. Unlike earlier theories which place the source of the candidate image in either the perceiver or the perceived (Nimmo and Savage refer to these as "perceptual balance theory" and "image theory" respectively; 1976, 31), the image of the candidate held by voters is described as being a result of a transaction between perceiver and perceived within a distinct situation. The third component, the situation in which the perception is taking place, is an important one in any theory of candidate perception and evaluation. Perception and evaluation do not take place in a vacuum (Huckfeldt and Sprague 1990; Huckfeldt 1983; Sprague 1982). Candidates try to shape the environment in which the perception takes place, evoking themes they expect will benefit their standing or harm their opponent's (MacDonald et al. 1989). The situation is important because it provides “relatively well-defined expectations about seeking and holding public office that structure any candidate’s relationship with potential voters” (p. 46). The work of Iyengar and his colleagues (Iyengar 19909; Iyengar and Kinder 1987; Iyengar, Kinder, Peters, and Krosnick 1984; Krosnick and Kinder 1990) has focussed on the effects of media priming in the interaction between perceiver and perceived. They have found that certain issues and themes are made more prominent as evaluative standards as a result of exposure to television news. These situational influences may do more than simply provide constraints on the transaction taking place between the voter 16 (perceiver) and the candidate (perceived). It will be my contention that there is an intimate and reciprocal relationship between the perceiver’s definition of the situation (Fazio 1986) and his or her perception of a political candidate. This relationship operates through the relative accessibility of candidate and campaign related concepts as the potential voter processes incoming information from the political environment. Concepts and traits which are brought to the conscious attention of the voter by the situation, or concepts which are chronically accessible (Bargh, et al. 1986), will have a greater impact on the overall evaluation of the candidate. Moreover, traits which are connected in a voter's mind will tend to pull each other into consciousness, increasing the likelihood that a particular category or dimension will have a greater effect on evaluation and on vote choice. Before going into the notion of accessibility too deeply, I would first like to explore the current state of knowledge concerning which traits are important and how trait judgments are used by voters in forming their overall evaluations of candidates.

The study of structure and trait judgments The study of candidate images is essentially the study of how candidates are represented in the minds of the populace, how they are thought of. The study of candidate perception is the study of the process by which these representations come to be. Both of these are important because they are the determinants of candidate evaluation and, ultimately, of vote choice. The categories into which 17 information is sorted and stored during perception shape the standards by which candidates are evaluated.

A dominant framework. The focus in the discipline on information variables and memory processes has led one group of scholars to declare recently that “all contemporary political science models of vote choice are information-processing models” (Lodge, et al 1989, 399; emphasis in original). One might also say the same thing about research on candidate perception. While this statement might be hyperbole, it is true that almost all of the contemporary literature on candidate perception has components which are focused on the processing of information. Yet this does not make every research effort an information-processing model of candidate perception. It is nonetheless an important observation, as much for what it says as for what it glosses over. Few could dispute the importance of such a ubiquitous framework, or of the implications of this ubiquity. Like the air we breathe, anything that exists everywhere we look but is never looked at for its own sake may be taken for granted. Unfortunately, this is often what happens when political scientists take an information processing approach to the study of political phenomena. As pointed out by Rahn, et al. (1989), there are two reasons for the fact that political scientists frequently talk about cognitive processes but are less likely to actually investigate them. First, the processing part of information processing is often treated as an 1 8 assumption and, as such, is left outside of the model, untested. Second, overwhelming reliance on cross-sectional mass surveys makes the study of cognitive processing difficult at best. Thus, much of the emphasis in political perception has been on cognitive structures such as schemata (Miller, et al. 1986; Lodge and Hamill 1986; Hamill, et al. 1985), stereotypes (Conover 1981; Conover and Feldman 1986; Feldman and Conover 1983; Rahn 1989), and prototypes (Kinder et al. 1980). A look at the development of the field makes the reasons for this abundantly clear.

Dimensional analyses. To begin with, dimensional analysis of candidate evaluation and perception has a long and respected history in political science, which predates the current fascination with the cognitive framework. Dimensional analysis in general and of issue stances in particular provided an obvious avenue of attack for scholars of candidate evaluation and candidate perception, building as they did upon earlier work on the dimensions of party competition (e.g., Downs 1957; Stokes 1963; Converse 1966). One of the seminal works in this field had the eponymous title “Dimensions of Candidate Evaluation” (Weisberg and Rusk 1970). In this article, Weisberg and Rusk used non-metric multidimensional scaling on feeling thermometers to locate a number of candidates in a two- dimensional space best described in terms of older New-Deal issues (strongly related to partisanship) and a cluster of issues of more recent vintage such as urban unrest and the Vietnam War. The quest then, as now, was to delineate not simply the connections 19 between elements (issue stands in this case), but the basic organizing forces of American politics. Thus, there was a heavy reliance on party and issue factors. Of course, image-focused studies are by no means new; several studies in the early 1960’s (McGrath and McGrath 1962; Sigel 1964; 1966) focussed on candidate image. The dimensional approach has dominated this aspect of the study of candidate perception as well.* This has had an obvious impact on the types of questions asked by researchers and the types of assumptions they must make in their research, compounded by the fact that the major source of data has been mass surveys and standard image rating questions. Specifically, researchers from a tradition such as this, when confronted with a desire to investigate how voters process information, often do so in terms of structure-as-categories-of- evaluation. They then assume the underlying cognitive connections between concepts mirror those they observe in the covariation of affect. This approach is adequate if all one wishes to do is outline the important determinants of evaluation. It comes up short, however, if one wishes to go beyond this to topics such as the process of evaluation and the nature of the relationship between traits.

The content o f trait dimensions. This can best be illustrated by taking a close look at some of the most prominent work done on candidate image perception in the last decade. I am referring here to the work by Kinder and his colleagues centered on the candidate trait battery included in one form or another in all recent American 20 National Election Studies. The inclusion of this battery has made these measures the most visible and most usable measures of candidate perception at the disposal of students of American politics. The trait battery had its beginnings in the 1979 NES Pilot Study. In their report to the NES Board of Overseers, Kinder Abelson and Fiske reported their desire to develop an “ensemble of survey instrumentation designed to assess the American public’s perceptions of its national political leadership” (1979, 1). One aspect of this instrumentation was framed in terms of the “personality characteristics citizens ascribe to leaders” or, as it came to be operationalized, to candidates for the presidency (1979, 2). The traditional method used by scholars in tapping candidate imagery, at least in terms of measures taken from the National Election Studies, had been an analysis of the open ended likes and dislikes about the candidates. Kinder et al. noticed that analysis by many scholars (Campbell, et al. 1960; Miller and Miller 1976; 1977; Nimmo and Savage 1976) of open-ended questions in various formats produced many mentions of the personal characteristics of candidate. These analyses tended to yield interpretations roughly paralleling expectations generated by social psychological research. Kinder et al. noted that "four autonomous lines of social- psychological research in fact converged...[indicating that] evaluations should fall roughly along the largely independent dimensions of competence and sociability (1979, 6)." Based on this, the authors developed and tested a trait inventorydesigned to cover these recurring themes of competence and sociability and to work well in a 21 dimensional analysis (1979, 7-8). The investigative thrust of the research was to find out which personality traits were important to citizens when evaluating their leaders. The question format used in the 1979 study was designed to determine how well the respondent believed any trait described a political figure. Analysis of the data obtained from this series of questions indicated a significant relationship between trait ascription and overall candidate evaluation and preference (as measured by candidate feeling thermometers and rank ordering of candidates, respectively). While attempting to avoid entanglement in causal inferences, rightly recognizing the role of traits as both reason for and rationalization of evaluations and preferences, Kinder et al. report that the traits accounted for 30% to 4S% of the variation in evaluation and 15% to 50% in preference (1979, 10). This level of prediction was hardly affected by the addition of party identification to the equation. They also found what they described as a “mix of the generic and the particular” in terms of the factor structure of the traits. Unsurprisingly, considering the construction of the trait battery, the generic part consisted of two mostly independent dimensions best labeled as competence and integrity. There was also a third factor idiosyncratic to each candidate. Later efforts have elaborated on this picture of structure. Current wisdom indicates that these two dimensions may be further subdivided into four dimensions: competence, leadership, integrity and empathy (Kinder 1986; Kinder and Fiske; 1986) As can readily be seen, this picture of the important categories of evaluation is truly 22 an elaboration of our earlier understanding; two of the categories have the same labels. Essentially, what has happened is that sub­ categories have been distinguished within the initial dimensions. So it is that evaluations of competence and leadership are highly associated with each other, and integrity and empathy have been found to be very highly associated (Kinder 1986). Typically, models incorporating trait effects do not make this further distinction, preserving only the categories of competence and integrity or competence and "personal qualities" (Sniderman, et al. 1990; Rahn et al. 1990).

Traits as an organizing force. Kinder and others have attempted to place presidential character traits within a broader framework of impression formation and evaluation. In general, traits have come to be understood as "cognitive categories that observers apply to behavior" (Kinder 1986, 234). In the case of citizens and presidents (and presidential candidates), citizens do this in order to better manage and understand the political world. Inundated with complex and changing information, which most pay only limited attention to, citizens can make sense of the world by seeking the causes of events in stable personality characteristics of prominent public figures (Heider 19S8; Kinder and Fiske 1986; Kinder 1986). One common explanation of how they do this is found in the notion of implicit personality theories, which allow people to infer stable character traits from overt behavior (Kinder 1986; Kinder and 23 Fiske 1986; Abelson et al. 1982). In essence, these provide information that tells an observer which traits are associated with a behavior. Furthermore, citizens can and do elaborate on these initial inferences because the theories also contain information on which traits are most likely to occur together. The importance of the dimensions to overall evaluation changes from candidate to candidate, election to election. Kinder and Sears report that competence seems to weigh more heavily in general. This has been found to change in response to several factors. These include candidate histories (Kinder and Sears 198S; Kinder and Abelson 1981), voters whose picture of an ideal president centers more on integrity (Kinder et al. 1980), or current societal concerns and events (Barber 1980; Page 1978). In a slightly different context, ideological identification, education, gender and other sociodemographic characteristics of citizens have been found to have an impact on which characteristics are emphasized as definitive of an ideal president (Kinder et al. 1980). Evidence concerning the extent to which and the process by which ideally definitive traits influence evaluation of actual candidates is mixed. It is not clear if presidential prototypes impact only upon incumbent candidates running for re-election, or if the failure to find much connection between prototype and non- incumbent is due to the impact of the current incumbent on definitions of the ideal (Kinder et al. 1980). 24 Summary The analyses of the structure of trait judgments have been, to this point, an attempt to identify stable groups of traits, that cohere across elections, voters, and candidates. As such, they have identified the categories or dimensions of importance in terms of the covariation o f evaluative measures. For example, competence is deemed an important broad category because evaluations of candidates on traits such as intelligence, experience and knowledgeableness follow a pattern; individuals who rate a candidate as likely to possess one of these traits rate them as more likely to possess others. These results have been repeated over and over for approximately a decade, and my intention is not to call them into question. Instead, I wish to point out that the covariation of evaluation provides an incomplete picture of the structure of evaluation, and that an alternate measure will supplement and extend our understanding of both structure and process. When we ask ourselves about the dimensions of trait structure, what is it we want to know? I believe we want to know which traits go together in people's minds, consciously or not. When a voter thinks about honesty, which other traits is she likely to think about at the same time? The analyses have, to this point, shown that some traits 'go together' and the dimensions they form are important predictors of attitudes and behavior. They have not shown that these traits are used together, or indeed have any impact on each 25 other. Standard factor analyses attribute the covariation of evaluation to underlying, unmeasured constructs. The supplement that I wish to propose is centered on the concepts of construct accessibility and networks of activation, the topics of the next chapter. Using these concepts, it is possible to outline and measure some of the effects a trait has on other traits. These effects have important implications for our understanding of structure, in both senses of the word, as well as evaluation, the use and availability of political information, and many other aspects of political cognition. Notes

Of course, the predominance has not been complete. Experimentation, for example, is becoming more common among students of public opinion in general and political person perception in particular. Several recent examples include, but are not limited to, Rahn, et al. 1988; Rahn 1989; Lodge, et al. 1989; Stroh 1989. III. Accessibility and Trait Evaluation

In the previous chapter, I made the case that if we want to know 'what goes with what' in people’s perception of political candidates, we must supplement current efforts by developing an explanation of why certain traits become associated to form dimensions of evaluation. Furthermore, it is not enough to know the categories of evaluation used by voters, we must also be curious about the categories of perception which act as guides to evaluation. The approach I suggest be used as a supplement is one based on the concept of the relative accessibility of trait judgments, which can be measured using response latencies. While this concept is now familiar to most students of political behavior, its use and even mention by political scientists has been rare until relatively recently. After discussing accessibility research in general, the chapter includes an examination of the theoretical underpinnings of the accessibility concept—response time measures in general. The discussion of response latencies is a necessary precursor to any research which directly measures accessibility, since reaction time is one of the most common ways to measure accessibility.

27 28 The concept of construct accessibility will bring with it the notion of a new type of structure, different from the simpler type of structure that was the main focus of discussion in the previous chapter. As it has been used by cognitive and social psychologists, cognitive structure means something more than dimensions or categories. As discussed by Anderson (1983), one way of discussing the structure of memory is as a system of nodes in a network connected by associational links.1 These links, over which activation (i.e., increased ease of accessibility) spreads, are established over time as different nodes are made active in conjunction in working memory. Using these ideas as a foundation, the discussion will turn to recent applications of accessibility in the study of political perception and evaluation.

Accessibility in There are two strands of accessibility research that need to be considered here. Although these two strands may be considered conceptually distinct, they are often woven together or not even distinguished as separate. I am referring to research on the accessibility of constructs, best exemplified by the work of Higgins and his colleagues (Higgins, Rholes and Jones 1977; Higgins and King 1981; Higgins, King and Mavin 1982; Higgins, Bargh and Lombardi 1985), and work on attitude accessibility, especially that of Fazio and his colleagues (Fazio 1986; Fazio, et al. 1982; Fazio et al. 1983; Fazio, et al. 1986; Powell and Fazio 1984). Attitude accessibility may be considered a special case of construct accessibility, where the 29 attributes stored as part of the construct are evaluative. The two are often intermingled and have similar findings concerning correlates of accessibility, in terms of both antecedents and consequences. To show that this is so, and to introduce some important findings about the causes and consequences of accessibility, let us now turn to a discussion of this research.

Accessibility as a guide. Findings that indicate current attitudes and cognitions may act to guide perception are not new (e.g., cognitive consistency theory). The problem has been to find an adequate theory explaining when, why and how they do so. Researchers involved in the study of accessibility have attempted to use findings based on reaction time studies to formulate answers to these questions. One of the most comprehensive attempts to develop a theoretical framework that ties together accessibility research is found in Fazio's article on attitudes as guides to behavior (1986). The proximate guide to behavior, in Fazio's model, is the definition of the event, which is in turn a function of the immediate perception of the attitude object and the definition of the situation in which the attitude object is encountered (1986, 212-213). The perception of the attitude object may or may not be guided by currently held attitudes about that object, depending upon the state of activation of the attitude(s) toward that object. An attitude which is not activated will not be used as a guide, even though it is available in memory. The likelihood of an attitude being activated (i.e., the accessibility of 30 the stored affect regarding the attitude object) is influenced by cues which make the attitude relevant and by the chronic accessibility of the attitude. In his description of the process by which attitudes come to guide behavior, Fazio (1986) describes an attitude as essentially an association between a given object and a given evaluation. The strength of the association is analogous to the strength of the attitude and is measured by the accessibility of the attitude. At the top of the scale in terms of strength are attitudes which are automatically accessed or activated upon merely encountering the attitude object (Fazio, et al. 1986). Automatic activation, according to Fazio and his colleagues, is unlike standard activation in that it is not a controlled process requiring reflection or attention but instead occurs spontaneously (Fazio, et al. 1986, 229). The idea of automatic activation of attitudes is an important one because it leads us to a major part of accessibility theory—the network of spreading activation.

Activational networks. The notion that activation spreads through a network of associations in memory is one that can be found throughout the cognitive psychology literature (Higgins and King 1981; Anderson 1983; Fazio 1986; Fazio, et al. 1982; Fazio et al. 1983; Fazio, et al. 1986) and, as we shall see below, it is often assumed by researchers in public opinion, especially those making use of priming procedures. In the spreading activation model, memory may be represented in terms of a structured network.2 The 3 1 nodes of the network are objects, events, concepts, types and categories3 which are connected by associational links over which activation spreads. When a particular level or threshold of activation is reached, the object centered at the node enters consciousness (or, as Anderson refers to it, working memory) and is, to use the familiar terminology, accessed. Furthermore, the presence of some level of activation that does not reach this threshold facilitates activation of that node, just as it is easier and quicker for an automobile to reach the speed limit from a rolling start than from a dead stop (see Bargh et al. 1986; Higgins et al. 1985 for more on this point).4 Higgins and King, in their 1981 article on social constructs, identify a number of factors affecting accessibility. Among these are expectations that the construct will occur, the belief that certain goals and needs will be facilitated by some construct, recency of use of the construct, frequency of use of the construct, distinctiveness of the construct, and the number and strength of the construct's associational linkages with other accessible constructs. To this list, we may add cues which make the construct relevant (Fazio 1986), the prominence of the construct in self-schema (Markus 1977; Markus and Smith 1981) direct experience with the object (as opposed to indirect non-behavioral experience; see Fazio et al. 1982) and in the case of attitudinal constructs where the attributes in question are evaluative, repeated expression of the evaluation (Powell and Fazio 1984), and perhaps attitude extremity (Powell and Fazio 1984).s 32 Before turning to the application of accessibility to political cognition, a short discussion of reaction time measures is in order. The key to the appropriate use of such measures is a carefully grounding of their interpretation in theory.

Reaction time as a measure of accessibility Reaction time measures, or response latencies as they are also called, have a long history in psychological research, dating back at least to the mid 19th century (see Kantowitz, et al. 1988, Appendix A, for a history of the development of experimental psychology). By the early 1970's, reaction time measures were characterized by one experimental psychologist as having "become about as common a dependent variable as there is in human experimental psychology" (Pachella 1974, 41). As Pachella points out, the popularity of reaction time measures (RTM's) has followed in the wake of an increased interest in cognitive psychology. In contrast to behaviorist approaches confined to studying the observable, cognitive psychology is devoted to the study of information (i.e., cognitions) and what is done with it. This means that the focus of cognitive research is often on an unobservable event taking place inside of people's heads. One of the few things measurable about such events are their duration. As any sampling of research reports dealing with RTM's will show, they are used in a variety of ways to test a variety of hypotheses. This is true even though the latency measure is quite 33 uniform, usually being defined as the time interval between the presentation of some stimulus and the response to that stimulus. The "trick" to using such a measure correctly is to ground its interpretation in well defined and testable hypotheses about relative lengths or changes in reaction times based on the experimental condition. On this point, Pachella informs us that

experimenters often desire to interpret the reaction times obtained in an experiment as "the time to recognize," "the time to deduce," "decision time," or "the time to search memory." In order to allow any of these specific interpretations, an experiment must employ a design that is suitable to the conclusion that the obtained variation in reaction time is related to the variation in the duration of the particular mental process under study (Pachella 1974, 45).

The solutions most often chosen, Pachella goes on to point out, are converging designs used to break the reaction time into constituent parts of interest (1974, 46-53). Such methods are not required in what Pachella refers to as the "molar" approach, in which the time interval as a whole is all that is of interest (1974, 46). Fortunately, since Pachella devotes the rest of his article to pointing out the problems with the use of converging methods with latency measures, researchers in political perception are not often interested in breaking down the interval into constituent parts. This is the case because the interpretation of the time interval in political perception is usually in terms of therelative accessibility of some attitude or construct. If the interest were to 34 shift slightly, for example to an interest in the actual time it took to retrieve an attitude from memory, the methods discussed by Pachella would become necessary since the retrieval process would have to be separated out from other processes accounting for some portion of total response latency. It is important to keep in mind that accessibility is not the same thing as response latency or reaction time, it is but an interpretation of these measures. Since the rest of this paper deals strictly with accessibility, however, 1 shall at times use accessibility as a synonym for latency or reaction time. Accessibility should also not be confused with availability (Higgins and King 1981). While accessibility refers to the readiness or ease of retrieval of an attitude or construct from memory, availability refers to its presence as such (i.e., whether or not it is there to be retrieved at all). Availability is thus a dichotomous variable, while accessibility is, at worst, measured at the ordinal level.®

The previous pages have pulled together a large amount of information about research on accessibility and RTM's in general. Their purpose has been to place the idea of accessibility as a dependent variable in perspective, to show how it may be used to explore the relationships between many different concepts used in cognitive research and to test hypotheses about the processes involved in human cognition. With this background information in mind, we may now take a look at several applications of the accessibility concept in the study of political cognition. Several 35 studies involving candidate perception in which the ideas of accessibility and activation are inherent but never fully exploited will also be reviewed.

Accessibility in political cognition Candidate perception is a process characterized by a give and take between perceiver and perceived within an environment that also has an impact on how the process unfolds. Often, it is difficult to study such processes as they unfold and we are left instead with the option of studying the results of the process and the traces it has left.7 The dimensional analyses of evaluation reported above do just that: by looking at how evaluative ratings are structured, we hope to make inferences about not only the causes (presumably, the perception of candidates' personalities, issue stances and physical appearance) but the processes (inference, comparison to ideals, implicit personality theories) that produced them. Measurements of reaction time may be thought about in this way, also. Though any measurement strategy using reaction times assumes some internal process is occurring at the time the measurement is being made, it is also the case that RTM's may be used to get a picture of the structure of cognition, structure produced by some process which occurred in the past. Thus, when we measure accessibility we may think of it as a measure of a trace left by some past process. The importance of adding this conceptualization to our studies is that we may use accessibility as a measure of a structure that guides evaluation, and not simply gloss it as a process that affects evaluation in some 36 unspecified manner. In short, networks of activation allow us to explain the memory processes underlying accessibility and to formulate testable hypotheses about the relationship between accessibility and activation. This latter point is, I believe, one that is often missed by students of political perception, who have tended to treat accessibility as a process and overlook the possibilities of treating it as evidence left by other processes. However, the concept itself has found a place in the political perception literature, indeed an ever more prominent place as 'information-processing' has come to be recognized as a coherent 'approach' by the discipline as a whole. What has this place been? Often as not, the accessibility concept has been implicit in the writings of scholars of political cognition. Yet, several authors have given it a central role in their research. It is to the latter I wish to turn first, examining how the concept has been applied. After this, the discussion will focus on research in which the concept of accessibility plays a less prominent role (or, at most, is mentioned in passing) in the theoretical underpinnings of the work. This will set the stage for a detailed look at how response latencies may be used as traces of the process of candidate perception. This will supplement the study of evaluative traces which often assume accessibility effects but seldom test for them.

Accessibility, Consistency, and Attitude Importance. Perhaps the best place to begin looking at accessibility and political cognition is in the work of Fazio, mentioned above as one of the prominent 37 researchers involved in the more general study of attitude accessibility. In an attempt to assess the effect of accessibility on attitude-behavior consistency, Fazio and Williams (1986) conducted an investigation into the relationship between attitudes toward presidential candidates, subsequent perceptions, and voting behavior in the 1984 election. Accessibility was operationalized as response latency to two identical statements about candidates Mondale and Reagan. These statements were of the form "A good president for the next four years would be Ronald Reagan." The statements were presented in the form of sound recordings; the candidate name came at the end of the statement so that the response latency could be measured from the end of the acoustic signal to the commencement of the subject's response (Fazio and Williams 1986, 506-507). Responses were made by striking one of five keys labeled strongly agree through strongly disagree. The model underlying the research is that described above (Fazio 1986). Within the process model, as Fazio and Williams title it, attitudes are simply object-evaluation associations and the accessibility of the attitude is a function of the strength of this association (1986, 505). Also, as you will recall from the discussion above, attitudes are believed to guide behavior only if they are activated. The stronger the association, the more likely they are to be activated; the more likely to be activated, the more likely to guide perception and to be predictive of behavior. Thus, subjects who were quicker at accessing their attitudes toward the candidates as potential presidents were expected to exhibit a stronger relationship 38 between these attitudes and later perception of these candidates (during televised debates) and voting behavior. The results of their analysis support the model. Those individuals with relatively speedier responses to the attitude question had perceptions of the debate outcomes that accorded more highly with these attitudes, measured two or more months earlier. Those in the high accessibility group also voted in a manner more consistent with their attitudes about the candidates. As important as these concrete findings are, they are overshadowed by the fact that they provide support for the model, which can be applied to many such attitude-behavior problems. The importance of these findings for the present research effort is also to be found in their support for the model, and for the evidence they provide that our understanding of political cognition may benefit from the further application of it. Scholars studying political information in its various guises have long sought explanations for selective perception, for models explaining effects such as persuasion and projection (Brody and Page 1972; Krosnick 1988c; Stroh 1989), for explanations of the conditions under which these occur. The accessibility concept, and the model of perceptual Fazio has built around it, provides such an explanation. Embedded as it is in other models of cognitive processes, such as the spreading activation model, it also brings many opportunities in the form of unanswered research questions. One such question is the connection between attitude accessibility and attitude importance. Krosnick (1988a, 2) notes that accessible attitudes and attitudes that are personally important bear 39 a striking similarity to each other on three important characteristics. Both types are distinct relative to other attitudes, both are frequently topics of thought and discussion, and both possess extensive linkages with other constructs in memory. This suggested to Krosnick the likelihood that important attitudes are stronger determinants of perception and behavior because they are more accessible. In a pair of studies focusing on issues such as abortion, defense spending and integration, he found a significant relationship between self-reported attitude importance and accessibility. Krosnick discusses this in terms of two alternative processes which may be occurring (1988a, 12). First, it may be that the relative clarity of internal cues causes people to take longer to consider an attitude object in the case of unimportant attitudes because of the difficulty of integrating the ambiguous information. On the other hand, speculates Krosnick, the cues themselves may take longer to retrieve in the case of unimportant attitudes even though the attitudes are no less clear than important attitudes. In either case, the author admits that the causal nature of the connection cannot be established with his correlational data and he calls for experimental manipulations of attitude importance to see if this affects response latency. There is an interesting possibility that is overlooked here. In both of the processes described by Krosnick, it is importance that is the suspected cause of accessibility, though the author does leave room for spurious correlation to be the culprit (1988a, 13).8 The possibility that is overlooked is that accessibility determines ratings 40 of importance. Experimental evidence demonstrates that behavioral experience and frequency of activation lead to lessened response times and, it is theorized, something approaching automatic activation (Bargh et al. 1986; Fazio 1986; Fazio et al. 1986). Fazio has offered (and shown support for) his theory that accessible attitudes guide behavior. Other evidence has shown that people make inferences about the importance of their own attitudes based on cues much as they would about others (Nisbet and Wilson 1977). Thus, it is possible that certain attitudes are made more accessible, even automatically accessible, through repeated exposure to the attitude object (say reports on integration in the evening news, or conversations with friends and co-workers interested in the topic). As accessible attitudes, they begin to guide perception and, in the long run, behavior. When the subject is asked to make an assessment of the importance of the attitude, he or she recalls thinking often about the attitude object or remembers behaving in accordance with the attitude and infers that it must be important. Of course, what is most likely is a reciprocal causal connection between accessibility and assessment of attitude-importance, with one of the processes Krosnick describes (importance leads to increased thought, salience, etc. which leads to increased accessibility) occupying one side of a causal loop with the process I have just described. This is an important empirical question that deserves exploration.

Accessibility, the Media, and Evaluative Standards. At least one aspect of it has been explored to some extent by researchers 41 investigating the effects of the media on political perception. Commonly called the agenda-setting function, and more recently known as priming, the effect of the mass-media on 'what people think about1 has long been of interest to students of public opinion. In a recent article, Entman (1989) discusses the differences between media influence as "telling people what to think" and "telling people what to think about" in terms of two models which will sound familiar from our earlier discussion of candidate evaluation (see the discussion of Nimmo and Savage in Chapter 1). Entman labels these models the "autonomy model”, which hypothesizes that at most the media influence what people think about, and the "interdependence model", which assumes that the media's influence is greater than minimal and is produced by an interaction between the public and the information put out by the media (349). This argument directly parallels the earlier discussion of perceiver-determined versus stimulus-determined images; Entman's interdependence model is very similar to the interaction model discussed by Nimmo and Savage. In the autonomy model, cognitive screening of information (e.g., selectivity and inattention) are emphasized. In Entman's model of interdependence, which adopts an information processing point of view, information is checked for salience rather than its match with currently held beliefs. Salience is determined through comparison with cognitive structures "which organize their thinking...[and contain, among other things]...rules for linking different ideas" (Entman 1989, 349). 42 Entman adopts the term schema, but only out of convenience. (See his Note 4.) The concept of message salience plays a role similar to that of attitude importance in Krosnick's work, in that it is a product of interest, strength of belief, centrality of attitude object to macroconstructs such as liberalism/conservatism, and familiarity of the attitude object discussed in the message (Entman 1989, 352- 353). This sounds very much like a list of traits describing important attitudes or of accessible attitudes discussed above (e.g., familiarity as frequency of exposure, interest as distinctiveness, centrality as extensive linkages). Thus, it is quite possible that the salience of the message is bound up with the accessibility of the object of the message as a construct in the receiver's memory. Although Entman never raises this possibility himself, the idea that construct accessibility plays a role in media influence on public opinion is examined in the work of others. For example, Iyengar and his colleagues explore the effect of TV news coverage on presidential evaluations, making explicit and extensive use of the accessibility concept to do so (Iyengar, Kinder, Peters, and Krosnick 1984; Iyengar and Kinder 1987; Krosnick and Kinder 1990; Iyengar 1990). Their thesis is, to use the familiar terminology, that news coverage influences 'what people think about' and, by doing so, "might also determine the standards by which presidents are judged" (Iyengar et al. 1984, 779). In later works, this process is given the name priming (Iyengar and Kinder 1987; Krosnick and Kinder 1990). The mechanism through which this is hypothesized to occur is accessibility. 43 In their experimental studies, Iyengar and his colleagues use clips gathered from actual evening news broadcasts. Subjects were exposed to differing levels of information about issues such as energy, inflation and defense. It was assumed that the information was made "accessible by its recency" (Iyengar et al. 1984, 779); actual response latencies were not measured. Given the assumption that exposure increases accessibility, the evidence provided by their experiments did support the hypotheses. Ratings of the president's performance in the issue area 'primed' did, in general, have a greater relationship to relevant evaluations (e.g., overall evaluation, competence evaluations) than to nonrelevant evaluations (integrity evaluations), and had significantly different impacts in the experimental group than in the control group.9 Moreover, this effect was not mediated by a change in evaluation of the president's issue performance, since the exposure to the issue did not seem to cause any such change. The hypothesis here is that accessibility increases the likelihood that the accessible construct will be used as an evaluative standard. The importance of the construct, at least as measured by increased ability to predict overall evaluation, increases. The process by which these evaluative standards change may or may not be a conscious one; Iyengar and his colleagues are correct in reporting that their "results are noncommittal on this point (1984, 785)." Thus, while they do not claim to answer the question of causal precedence (accessibility leading to increased importance or increased 44 importance leading to accessibility) they do acknowledge both possibilities.10

Implications. What is most important to take away from this research is the evidence that exposure to media information does change the relevant evaluative standards used by viewers without changing the extremity of the evaluations. This establishes three points very important for subsequent research. The first of these points is that exposure to mediated information and, by inference, the accessibility of the constructs contained therein, plays a role in shaping evaluation.11 This indicates a continued need to explore the relationship between construct accessibility, construct/attitude- object importance, and evaluation. Secondly, it provides additional evidence that accessibility is not the same as evaluative extremity (recall the earlier discussion concerning the slight correlation between the two). Thus, conclusions based on the dimensional analysis of attitude extremity may only provide a part of the picture of the interrelationships between candidate image traits. The third point is a more subtle one: the relevance of the message to particular trait constructs plays an important role in mediating the effects of the exposure. Implicit in this statement is the proposition that there are links between a message and some trait constructs but not others. In this case, performance in some policy area is linked with the category of evaluation labeled competence but not with the category labeled integrity. Putting this into a spreading activation model results in the hypothesis that 45 exposure to information about issue performance produces not only an increase in the accessibility of that information when making evaluations but an increase in the accessibility of certain evaluative standards as well.

Conclusion—The need for accessibility From these three points, I conclude that a research program designed to describe the process by which evaluation takes place, how evaluative standards come into being or change over time, and the role of candidate perception in evaluation must take as its first step the description of the linkages in networks of activation. The linkages that should be explored include—but are not limited to— those between the evaluative standards, image traits, and other determinants of overall candidate evaluation such as issue stance and performance, partisan affiliation, and liberal/conservative placement. The focus of this dissertation is on only one of these topics—candidate character traits as a determinant of overall evaluation In a recent review of candidate perception research, Kinder and Fiske found many similarities of results in terms of candidate personality characteristics. Across research designs employing disparate methodologies, the same traits and the same dimensions show up time and again. Their interpretation of this is that people have implicit personality theories about presidents, just as they do about their friends (Kinder and Fiske 1986, 205-206). Judgments 46 about presidents are "rough-and-ready comparisons" based on accessible information that "happens to come to mind (201)." This is often the tack taken in research of this type. Whether the process is described in terms of prototypes or stereotypes, cues or inferences, the availability and/or the accessibility of information is usually acknowledged as a key factor. Several recent examples should show this to be the case. Rahn (1989), for example, manipulates the accessibility of partisan stereotypes in order to examine their effects on issue stance perception and trait ratings, finding accessibility does make a difference in some circumstances.12 Lodge, McGraw and Stroh (1989) are concerned with determining if the available information stored in memory is in the form of particular judgments that must be tallied when an evaluation is called for or is instead simply a stored evaluation called up already formed. They deal with the availability side of the question and find evidence to indicate voters rely on summary tallies stored as such. Although researchers often state that it is the accessible information that has the greatest influence, few take the trouble to find out what information is accessible or how the accessibility of any one piece of information affects the accessibility of any other. All too often, political scientists anxious to describe the process in information processing do not go beyond conceptualizing accessibility as a (Fischhoff, Slovic and Lichtenstein 1980) to explain how voters, acting as cognitive misers, can reduce their costs of making decisions. Furthermore, they seldom look for empirical evidence showing that changes in accessibility are truly present. To 47 use the concept in this manner is productive but, it is important to go beyond this, to what is implied about the connections between cognitions in voters' pictures of the world. If accessibility of information plays the role attributed to it, there should be some trace of its influence to be found in the way information is stored in memory, there should be an identifiable structure of accessibility.13 This structure is important because it will not only help us understand the categories used by citizens in evaluating public figures, it will shed a great deal of light on the process by which these evaluations are formed. Does the mere mention of some personality trait influence the ratings on other traits, or even whether or not those other traits are used in the judgment at all? The evidence obtained should speak chiefly to concerns about which judgments are used at all in making the evaluation, and which judgments may be used as cues in making other judgments. There are certain basic inconsistencies in the way the field has approached the idea of structure. On the one hand, so many of our models insist that accessibility shapes the information environment; what is available is often what determines how people see things and how those things are evaluated. Yet, when it comes to measuring the structure of that information environment, to see "what goes with what", the idea of applying accessibility as a measurement technique is grossly underutilized. In the next section, a research program designed to perform just such a task will be outlined. The hypotheses to be tested will come directly from current models of 48 trait structure, models derived from analyses of evaluation. By setting the research up in this way, we can obtain direct comparisons of evaluatively derived and accessibility derived structures, and so compare networks of activation to dimensions of evaluation. 49

Notes

1. Related to this idea of structure are, as Kinder puts it, a whole family of theories based on the idea that new events are "interpreted in terms of old knowledge" (Kinder 1983,414). In Anderson's terms, and in terms that I choose to follow, the "old knowledge" is embodied in long-term memory traces produced by the conjunction of concepts in working memory. (See below.) The idea may be more familiar to students of public opinion from work employing the schema concept.

2. The discussion of the general spreading activation model owes much to Anderson (1983) and to Posner and Snyder (1975).

3. Category here refers to a construct consisting of information about a class of objects, events or properties (Higgins and King 1981).

4. Posner and Snyder point out that there are two possible mechanisms that may account for differences in reaction time after exposure to related and unrelated items. The alternative to the spreading activation model is one in which the slower response after unrelated items is due to "a limited-capacity system that can read out of only one memory location at a time. Time is required to shift from one location to another, and the shifting time increases with the distance between locations" (1974, 71). The evidence they present provides much more support for the spreading activation model. This has been the model of choice for students of social cognition and is the one adopted here.

5. Powell and Fazio (1984) report that attitude accessibility and attitude extremity are correlated slightly, yet insist that "the strength of the object- evaluation association is not confounded with attitude extremity....repeated [attitudinal] expression affected attitude accessibility without having any discernible effect upon attitude scores (146)." So, although the exact relationship is unclear, accessibility and extremity are quite different animals. The astute reader will wonder if this covariance affects in any way my earlier claims for the accessibility measure as a supplement to evaluative measures in candidate perception research. Indeed, it does not. It would be more troubling if there were no correlation between the two, since they are to be used to explore the relationships among the same trait ratings. So, the low correlation is a reassuring sign that I am on the right track.

6. Accessibility, measured in terms of reaction time, would seem to be an interval if not a ratio level variable. Pachella, however, presents evidence that it may best be treated, in certain instances, as only ordinal. 50

7. The concept of a trace is taken from Anderson's spreading activation theory, where he describes the process by which a "transient working memory structure will be turned into a permanent long-term memory trace" (Anderson 1983, 262; emphasis in original).

8. In another, related article (Krosnick, 1988b) the author also emphasizes importance as causally prior to accessibility, this time directly in terms of candidate evaluation, while allowing for other possibilities. He states, for example, that H[v]oters seem to simplify this task [the generation of candidate evaluations] by concentrating only on those policy attitudes that they consider important. This may occur either as the result of deliberate, conscious decisions to focus on these attitudes, or simply because these attitudes are especially accessible in memory and come to mind automatically as candidate evaluation standards(207)." The causal precedence is unclear in the case of automatically activated standards. Interestingly, the author has been involved in other work that looks at the causal relationship from both angles (see Iyengar et al. 1984), which is discussed below.

9. It is important to note here the acceptance of the findings reported earlier concerning the dimensions of candidate image evaluation without question (see above discussion of Kinder et al. 1979). The measurement of integrity and competence were based on a battery of six trait adjectives, three assumed to measure each of the two concepts. No independent test of the dimensionality was reported.

10. Actually, if mere exposure leads to a conscious choice on the part of a person to increase the importance of the attitude, accessibility need not follow - except that later Krosnick (1988a) showed the correlation between reported attitude importance and attitude accessibility. The data presented in the Iyengar paper cannot speak to this because they assume exposure led to increased accessibility.

11. This accessibility or priming effect also has an effect on character assessm ents, a point made in the 1984 Iyengar et al. piece but emphasized much more in the later book length treatment (Iyengar and Kinder 1987; see esp. Chapter 7).

12. Note that her design does not actually measure changes in accessibility of the stereotypes. The presence or absence of partisan labels in the presentation of the stimuli is assumed to affect the availability of the stereotypes.

13. I shall leave the question of where to look for this structure, whether at the individual or the aggregate level, for the next section. IV. A Dimensional Analysis of Memory Structure

Abstract. In order to understand how citizens perceive and evaluate political leaders, it is important to develop an accurate picture of how they perceive the personalities of those leaders. The research presented takes an approach both familiar and very novel to outlining the connections between judgments about personality traits. Traits are treated as nodes in associative memory networks, which are akin to the standard constructs of personality dimensions. It is hypothesized that the levels of activation of any pair of trait nodes in a single sub-network are likely to covary: when one is highly activated, others in that sub-network will be relatively highly activated and thus relatively accessible. The level of activation between nodes in unrelated or less related sub-networks will be less likely to covary, there being fewer connections over which activation might spread. In general, if a trait thought to be associated with any dimension A is relatively accessible for whatever reason, other traits in that dimension should be relatively more accessible. If this is so, a dimensional analysis technique (such as factor analysis or covariance structure modeling) should reveal an interpretable factor structure using response latencies of traits as measures or indicators of the 52 accessibility of the latent trait dimensions. The results indicate four distinct but highly related sub-networks are required to account for the observed patterns in accessibility.

Introduction In the previous chapters, I outlined past efforts at understanding how the public sees and understands the political world and the people in it. One important finding has been that, frequently, people seek this understanding by attributing stable personality characteristics to prominent leaders. In political science, one of the main lines this research has taken has been dimensional studies of personality traits, most notably those by Kinder and his colleagues. In that discussion, I made the argument that such analyses can only take us so far in our understanding, for they offer few clues concerning how these dimensions are formed or how they shape the processing of information and, ultimately, evaluation of leaders. The alternative offered was to use the psychological theory of associative memory networks as a tool to test current notions about the organization of personality traits. The key to this approach is the use of the twin notions of accessibility and activation to appraise the connections between traits. The rationale for the use of accessibility data is very straightforward. Although there are many sources of variation in the accessibility of trait judgments of political candidates, one of the most theoretically elegant explanations of the patterns of covariation 53 is found in the memory models discussed in the previous chapter. In short, the greater the degree of relationship between any pair of traits, the more ties they have to the same traits in an accessibility network, and the more likely their accessibilities are to covary. The study described in this chapter begins with the collection of data on the accessibility of a number of trait judgments for several political figures. The interpretation of these data will depend on a slight variation in the two types of structure, activation and evaluation. In Figure 4.1.1, showing an activation network, the ‘thing’ that ties together the various traits is not portrayed. This missing entity is the concept or over-arching personality dimension latent in the network itself. Its relationship to the network is portrayed in Figure 4.1.2, which has very obvious similarities to the earlier figure. In the theory as I have reconstructed it, these over-arching labels are not only methodologically latent, they are cognitively latent as well. That is to say, I wish to treat the factors that will eventually emerge from the model of the covariances not simply as latent variables but as being truly latent in the minds of the subjects themselves. There is simply no need to posit an intermediate level of cognitive processing that goes on when a person encounters the phrase ‘Candidate Jones worked late into the night stuffing envelopes alongside the campaign office workers’ that requires him or her first to categorize that behavior as indicative of some over-arching trait called ‘capability’ or ‘sociability.’ There are many cues in the information and many possibilities about how it will eventually be 54 encoded and interpreted. These depend on the relative accessibility not of over-arching personality dimensions (e.g., ‘Capability’ or ‘Sociability’) but of particular nodes within the networks-cum- dimensions (e.g., ‘Hardworking’ or ‘Friendly’). Thus, the standard interpretation of a factor-analysis model such as the one depicted in Figure 4.1.2 is that the covariance in accessibility of traits loading on the same factor is due to their relationship to the underlying construct. In this case, the construct is the network itself, periodically being charged with activational energy then subsiding back to an unexcited baseline state. A model of the structure of the covariance between measures of accessibility of various traits should enable us to obtain a picture of these networks. With this in mind, a study was designed to collect measures not only of standard trait ratings of several political leaders but of the accessibility of those trait judgments as well. The covariance of these measures is then modeled and several hypotheses about the structure are tested. It is hypothesized that the levels of activation of two related nodes in a single sub-network are likely to covary: when one is highly activated, others in that sub-network will probably be relatively highly activated. The level of activation between nodes in unrelated or less related sub­ networks will be less likely to covary, there being fewer connections over which activation might spread. The level of activation can be measured using response latencies to stimuli containing the traits. It is not essential to our purposes here that the absolute levels of activation for all traits within any sub-network be equal, only that, 55 in general, if a trait thought to be associated with any dimension A is relatively more accessible, other traits in that dimension should be relatively more accessible. It should be possible to get a picture of the entire network, including information about the sub-networks (dimensions of trait evaluation) using a dimensional analysis technique such as factor analysis or covariance structure modeling on response latencies.

Method The study was conducted over a period of two weeks in May, 1990. Subjects were 165 volunteers from introductory political science courses who received extra credit for their participation. Groups of four to eight persons were introduced to the study using a prepared statement explaining that the study concerned their “opinions on political candidates and some political issues.” Subjects were assigned to individual cubicles so as to reduce distractions that might cause increases in the measurement error associated with the accessibility of their judgments. Accessibility was operationalized in terms of response latencies to 24 traits (listed in Figure 4.2) for each of four political leaders. The trait phrases were designed to be similar to existing batteries of traits commonly used in the American National Election Studies. The exact batteries as they currently exist could not be used in the context of measuring response latencies because they vary from short single words to longer phrases. This broad range in the length of the stimuli would have magnified disparities in word-recognition 56 time for the various traits. Instead, the trait word list was designed to contain single words rather than phrases as is more commonly the case in standard survey contexts. The four political figures used in the study were drawn from a larger pool of names pretested on a separate group of subjects. The decision to use a particular political leader was based on high rates of recognition, accurate identification, and the willingness of pretest subjects to evaluate the particular person. The decision was also made under the restriction that an equal number of Republican and Democrat political leaders were to be chosen. Based on these criteria, George Bush, Michael Dukakis, Dan Quayle, and Jesse Jackson were used as the four political figures. The study was conducted through the use of microcomputers. When subjects entered their cubicles, their initial oral instructions were to press the key marked ‘Yes’ to begin the study. Further instructions were then presented on the computer screen. The initial task presented to the subjects was to make a yes or no decision (by pressing one of two keys appropriately labeled) on whether or not each trait described each of the four persons. The person-trait pairs were presented in a random order generated anew for each subject. Subjects were instructed to ‘not take too much time’ making their decisions, and to go with their first response.1 Latencies were recorded directly by the computer. The person-trait pairs were presented in the following manner. A name appeared near the top of the blank screen, followed one second later by a trait near the bottom of the screen (see Figure 4.3). 57 Subjects were given a chance to practice the task on four pairs of unrelated persons and traits. After making half of the judgments (48 person-trait pairs), they were given a one minute break to minimize fatigue and inattentiveness. The computer signaled the end of the break and then gave the subjects the remaining 48 person-trait pairs to be evaluated. Following the last judgment, subjects completed another task on the computer. This task was designed to measure their response speed so that this factor could be taken into account in the analysis of the latency data. It was necessary to have several measures of the relative speed of each subject in simply reading a word and pressing a button without making any evaluations or drawing upon prior evaluations. To measure this, subjects were told to identify a word flashed on the screen as either “Yes” or “No” by pressing the corresponding button as quickly as possible. There were six repetitions of this task, three “Yes” trials and three “No” trials. Finally, subjects completed a paper and pencil questionnaire containing, among other things, the 96 person-trait pairs in a five- point response format, with end-points labeled “does not describe” and “describes very well”. This was designed to replicate the American National Election Study format. (See Appendix for the full text of this questionnaire.) These measures of a subject’s willingness to attribute each trait to each political figure provide a method of assessing the evaluative structure of the traits independently of the measures of trait accessibility. Subjects were debriefed after completing the questionnaire. 58

Plan of Analysis This research design results in two sets of data on the trait judgments, one based on five-point evaluations and the other on the accessibility of the judgments. The plan of analysis for the data is, therefore, two-pronged. First, in order to assess the factor structure in a way that establishes a type of baseline for this particular group of subjects, an exploratory factor analysis of the evaluative data is presented.2 To parallel this, an exploratory factor analysis of the accessibility data is presented as well. An alternate method of assessing the structure among a group of variables is to use confirmatory factor analysis. Confirmatory factor analysis allows the researcher to specify, prior to the actual analysis, different models of the structural relationships between the measured variables and test those models just as one would test any other hypothesis. Moreover, each and every decision about the structure can be based on theoretical or empirical statements about the expected relationship between pairs of variables, groups of variables, or between the latent factors. For example, if there are theoretical reasons to believe that some of the latent factors are correlated but others are not, this can be specified in the model and the likelihood that any particular inter-factor correlation differs from zero can be tested. Furthermore, careful forethought in model construction allows the researcher to compare the adequacy of different but similar models and even to conduct statistical tests concerning the relative increase or decrease in fit of the various 59 models to the data. Following the exploratory analyses, several models of trait structure are tested using both the evaluative and accessibility data.

Model Development The models used are developed from findings in the literature based on five-point rating data from national samples. The first step in developing models of the underlying structure among measured variables in terms of latent or unmeasured variables is to develop hypotheses about which of the measured variables lie together in the same network. The associative memory model implies that traits (nodes) in a network will covary in accessibility due to the transference of activational energy through the memory traces between the traits. In this case, hypotheses are needed that detail which traits occur together in the same activation networks. The only real source of such hypotheses is the literature on presidential character trait ratings, using the five-point response mode in which respondents report how well they believe a particular trait describes a candidate, i.e. how strongly they attribute that trait to that candidate. That the structure of trait accessibility would be similar to the structure of trait evaluation is not an outrageous notion. While this research can do no more than raise the question of the relationship between the two types of structure, it is entirely possible that a causal relationship exists between the two structures. The simplest causal mechanism is that of inference. When a person is asked to 60 evaluate a candidate on some trait, he or she may or may not have ever thought about that particular trait or stored an evaluation on that particular trait. In making a judgment on the trait, the respondent would need to make inferences from currently stored cognitions; the most likely cognitions to be used are accessible ones (Fischoff, Slovic and Lichtenstein 1980; Higgins and King 1981; Tversky and Kahneman 1981). The cognitions most likely to be accessible at this point are ones that are related to the current target trait, which has been brought into working memory and thus sent out its burst of activational energy. Thus, there should be a great deal of similarity between the two types of structure.

The Two and Four Factor Models. The most sophisticated analyses of trait structure are to be found in the various works by Kinder and his colleagues (Kinder, Abelson and Fiske 1979; Kinder 1986). Borrowing a term from the original analysis of the 1979 ANES Pilot study data, the models presented are designed to test the presence of ‘generic factors’ that are present across candidates (Kinder, et al. 1979,13). The first model I wish to test is based on a simplification of recent analyses of the five-point data that hearkens back to the original 1979 results. This centered upon two generic underlying factors.3 The first factor is best characterized as a Capability or Capacity factor. Traits such as intelligent, inspiring, and hardworking were associated with this factor. The second generic factor is best referred to as a Personability or Sociability 61 factor. Characteristics such as honest, caring, and corrupt loaded on this factor. The two generic factors were correlated relatively highly. The second model is a more articulated version of the first in which the 24 traits break out in to four factors: Competence, Leadership, Integrity and Empathy. (This model is fully developed in Kinder’s 1986 analysis.) Competence and Leadership are, more or less, subcategories of the original Capability factor; Integrity and Empathy are refinements of the Personability factor. The two pairs of sub-factors should be expected to be quite highly correlated, and there should be sizable correlations across the pairs as well.4

Accounting for negative phrasing. To these models must be added the complicated set of issues related to how the traits were phrased. For several reasons, it may be expected that some of the common variance among the trait accessibility measures is due to the fact that some of the traits are phrased in terms of negative characteristics (e.g., weak, foolish) and others are not. In his 1986 analysis of evaluative data, Kinder added a factor for negatively phrased traits, and guessed that the reason had to do with respondents’ level of difficulty in making negative judgments. There are at least three possible explanations for this phenomenon. First, the variation due to the phrasing of the traits and the need for a negativity factor in the evaluative data may be due to a general bias against making negative judgments of public figures. Evidence related to this is found in the work of Sears and his colleagues on the so-called positivity bias (Sears 1976; Lau, Sears 62 and Centers 1979). If people are motivated to evaluate people positively, responses to the negatively phrased traits may be a product not only of evaluations on that trait and its underlying factor, but of this bias as well. Another possible reason, and one discussed briefly by Kinder (1986, note 6), is related to the mental gymnastics expected to be involved in comparing negative traits with their positively phrased counterpart. In other words, the over-arching trait category is probably triggered more readily by traits reflecting competence rather than ineptitude, integrity rather than roguery. A third explanation, and one that raises possibilities for further research, makes use of an interesting finding in other work on spreading activation. Fazio and his colleagues have shown that activation can spread on basis of affective tags carried by attitude objects, i.e. from one positively evaluated object to another positively evaluated object or from positive word to positive word (Fazio, et al. 1986). Being primed with a word that elicits a negative affective reaction primes other such words. Aside from its many consequences for information processing in general, this would also have consequences for the structure that emerges in dimensional analyses of evaluation if that structure is shaped by activation networks. So in addition to the simple two and four factor models found in the literature, two models incorporating a negative factor as well as the substantive factors were also developed. No correlation between the negativity factor and the substantive factors was allowed.3 These models were specified in such a way that the 63 loadings of the twelve negatively phrased were constrained to be equal in the unstandardized solution. This means that the amount of variation in each of the twelve negatively phrased traits associated with the negativity factor is assumed to be equal across all traits. While the various negatively phrased traits are undoubtedly of different levels of negativity, it is not at present understood exactly how that would affect the accessibility of the judgments (especially given that accessibility is related to other factors such as extremity of the evaluation, which cannot be incorporated into the aggregate model of accessibility structure). The equality constraint may be overly conservative but, it is the least likely method of those available to result in meaningless covariation being captured by the negativity factor.

A unidimensional model . In addition to these models, a model based on the notion that there is only one underlying personality dimension for all 24 of the traits was also specified. This model implies that there are no distinctions between the traits on basis of Capability or Sociability, much less a more articulated version of these two dimensions. Substantively, this would mean that the activation of any of the personality traits was just as likely to cause an increase in the accessibility of any other trait in a citizen's mind. In terms of public opinion formation, this model implies that personality judgments are not organized around questions such as ‘Is that candidate capable of performing the functions of the office?’ but instead around broader questions of what the candidate is like as a 64 person. In a sense, this model provides a baseline to which the fit of the other models can be compared. If, for example, a two-factor model fits well, yet not significantly better than the one factor model, the need for complicated multi-factor models of trait organization must be questioned.

Nested models. The different versions of the models were developed in such a way as to make each model nested within the four factor plus negativity version of the latent structure. Nesting simply means that one model can be obtained from the other by freeing or fixing parameters in the model. Only free parameters are estimated in the analysis. In this case, by simply setting the value for the correlation between Competence and Leadership to 1.0 on one hand and between Integrity and Empathy to 1.0 on the other, one may cause the four factor model to become a two factor model, et cetera. This is illustrated in Figure 4.4.6

Exploratory Factor Analysis Evaluative Structure. The factor pattern matrices produced by the exploratory factor analyses of the evaluative data are presented in Table 4.1. Factors were extracted using maximum likelihood factoring. Initial solutions were obtained for solutions from zero to six factors. The patterns of eigenvalues as well as the Tucker-Lewis coefficients were taken into consideration in making a decision on factor retention. A four-factor solution was chosen for each leader. The factors were then rotated to a general oblique solution. 65 In many respects, the results of these analyses parallel the findings of similar analyses of national samples. For Bush, Jackson, and Quayle two factors veiy similar to the original ‘generic’ factors in the work of Kinder, Abelson and Fiske (1979) are definitely present. The Capability dimension (which here encompasses the six traits typically assigned to Competence as well as four of the six associated with Leadership) fails to emerge for Dukakis and seems to be replaced by a ‘Wimp’ factor, with loadings for weak, aimless and hesitant. For all of the four leaders, a factor having large loadings for most of the twelve traits associated with Integrity and Empathy, i.e. a Sociability factor, also emerges. This factor is somewhat different for each of the four, but in general the three positively-phrased empathy traits have relatively small loadings for this factor. The size of these loadings is best explained by the presence of a factor that has solid loadings, for each of the four political figures, for the positively-phrased traits. The presence of this factor is perhaps attributable to the unwillingness of the subjects to make extreme negative judgments about these political figures. It is interesting to note here that the Positivity factor does not seem to account for the positively-phrased leadership traits for the two Republicans but does do so for the two former Democratic presidential contenders. Finally, in each of the factor rotations, there emerged an Experience factor (or perhaps ‘Lack of Training’ factor would be a better characterization; unfortunately no positively-phrased traits antonymous to these two were included), with high loadings for just two traits: Inexperienced and Unqualified. 66 The structure exhibited in the evaluative responses from the 165 subjects did not differ overly much from that commonly found in national samples. Capability and Sociability are the principle structuring concepts underlying the individual trait judgments, with the two idiosyncratic latent concepts of Positivity and Experience also showing up for this sample. Overall, I believe this similarity provides reassurance that these results are generalizable, at least in terms of the similarity of the underlying processes at work.

Accessibility Structure. The central purpose of this study was to assess the structure of accessibility among the traits, specifically the covariance of the response latencies. The first step in doing this was to standardize the measures so as to account for individual differences in average response speed.7 There are several different and equally valid ways of doing this. The goal of the standardization is to re-express each individual’s response latencies on each of the 96 trait judgments as either ratios or differences from some measure of response speed for that individual. The study incorporated a measure, actually multiple trials of a measure, of reading speed and manual dexterity. After making the 96 trait judgments, the subjects were asked to complete six trials of a task that measured the amount of time it took them to read a word from the computer screen and identify it as either ‘yes’ or ‘no’. For each subject, an average response speed on these six trials was computed. This measure was then used to standardize the reaction times for each individual using the equation SXj = Xj / RS, where Xj 67 is the original latency on trait i, SXj is the standardized Xj, and RS is the response speed. Thus, the latency is re-expressed as a ratio based on the subject’s response speed.8 The next step in assessing the structure of the accessibility of the trait judgments was an exploratory factor analysis of the reaction time data. The first and foremost question to be asked was if there was any recognizable factor structure in this data at all. Factors were extracted using a maximum likelihood estimation technique.9 The resulting factor pattern matrices are presented in Table 4.2. The latent structure revealed in these exploratory factor analyses conforms in part to expectations about the underlying causes of the variability in accessibility of the various trait judgments. There are some parallels between the exploratory factor analysis of the accessibility data and of the five-point judgments. These parallels are most obvious in the results for Dan Quayle. Here, it is very clear that the two factors recovered from the data closely resemble the Capability and Sociability factors found in the literature and in the factor analyses of the five-point judgments by these subjects. Ten of the twelve traits associated with Capability load together on one factor and eleven of the twelve Sociability traits load on a second factor. Only four of the 24 traits cross-over with large loadings on the opposite factor. The resemblance is less clear for the other candidates. For Bush, Dukakis, and Jackson there is a very dominant first factor with sizable loadings for a majority of the 24 traits. This is, perhaps, due to the imperfection present in the standardization formula, which simply fails to eliminate all of the common variation attributable to measurement artifacts. However, the second factor in each case gets the bulk of its loadings from either the 12 traits associated with Capability or from the 12 traits associated with Sociability. In the case of George Bush, five of the seven traits having relatively large loadings on the second factor come from the Sociability traits. For Michael Dukakis, six of the eight traits loading on the second factor do so. Turning to the results for Jesse Jackson, the loadings for the second factor tend to come predominantly from the Competence traits (five out of six do so). Thus, while it seems that the structure derived through exploratory factor analyses of the accessibility data is not as clear cut as it is for the five-point judgments, there is some evidence that the two are similar. For Quayle and Jackson, there is evidence of a network of traits consisting mainly of those previously identified as Capability traits, while for Bush and Dukakis a network centering in traits associated with Sociability is present. The structure seems less defined than that found in the evaluative data; the presence of the single dominating factor is evidence that there are certainly factors other than Capability and Sociability networks at work here that have not been eliminated by the standardization for response time within individuals. The second factor, however, confirms prior expectations and puts to rest fears that the accessibility data are too complex to be treated with factor analytical techniques. 69 Comparing Models with Confirmatory Factor Analysis The exploratory analyses reported so far give reason to believe that both the evaluative and accessibility data are structured in much the same way. Minimally, one may expect to see a Capability/Sociability split in the accessibility data and perhaps a more complex four-factor version of this in the trait evaluations. Confirmatory factor analysis allows the fit of the various models of the trait structure, described earlier, to be tested using both types of data. The fit of the models, within each type of data, may then be compared. Each of the models was fit to both the five-point evaluation data and the measures of trait judgment accessibility. Parameters were estimated using the maximum likelihood estimation technique available in LISREL 7.10 The matrix analyzed was the covariance matrix for the 24 trait judgments, based on 157 cases. In order to provide scales for the latent variables, their variances were fixed at 1.0. No covariance between the error terms for the measured variables was allowed in the models. For each of the models fit to the data, the LISREL program calculates a value that may be used to assess the fit of the model. Typically, x2 /df ratios of 2.0 or less are indicative of a good fit of the model to the data. Moreover, when covariance structure models are constructed so as to be nested (i.e., one model may be created from another merely by freeing or constraining one or more parameters) the difference in for the two models is itself distributed as with degrees of freedom equal to the difference in 70 degrees of freedom between the two models. This allows a significance test by which one may assess the need for any of the parameters by examining the worsening in fit going from a less restricted model to a more restricted one. The various models estimated for the two sets of data examined here were constructed in this manner, and the values and degrees of freedom are reported in Table 4.3.

Evaluative Structure. There are several interesting results revealed in this table. Unsurprisingly, the best fit is obtained for the most complex model, which contains the four substantive factors as well as the negativity factor. For both Bush and Dukakis, the fit of the model reaches the 2.0 threshold for the x^ /df ratio. While leaving room for improvement, the fit is acceptable for Jackson and Quayle, with the ratio taking on values of 2.4 and 2.5, respectively.11 At the other end of the spectrum, the fit for the One Factor model, which specifies a structure that does not distinguish between overarching personality dimensions, fits very poorly. This offers additional confirmation that there is some structure to the trait evaluations. In terms of relative fit of the models for the evaluation measures, in every case the change in x^ is statistically significant. There is a definite difference in the size of the loss in fit that depends on the type of change in the model. Consider the changes in X^ when changing the number of substantive factors (i.e., moving from the 4 Factor to the 2 Factor model, or from the 4 + Negative to 71 the 2 + Negative model) versus dropping the negativity factor while retaining the same number of substantive factors. The biggest loss of fit comes with the loss of the negativity factor, which typically results in a change of 150 or more points in the value of Halving the number of substantive factors produces a change of only 50 points or so. This is not to say that the addition of the negativity factor is more important than breaking the two substantive factors of Capability and Sociability into their constituent parts. It is, however, ample evidence of the utility of the negativity factor in obtaining an adequate model of evaluative trait structure. Moreover, every additional restriction placed on the model results in a significant decrease in fit. This replicates and confirms Kinder’s finding that the most complex model, that which has four substantive factors as well as a negativity factor, is the best model of the underlying evaluative structure.

The accessibility structure. The figures presented in Table 4.3 for the accessibility measures look very different. Here, the smallest X^/df values are found for the most complex of the models only for two of the four candidates. In the case of both Bush and Jackson, there is no change in the value of the ratio as one moves from complex to simple models. However, the actual values of this ratio are, for all models and for all candidates, close to the 2.0 level. For Dukakis, Quayle and Jackson the ratio takes on values of 2.2 or less except in the case of the one factor model, where the worst fit is 72 about 2.3. For Bush, the value of the ratio stays steady at 2.4. While one could wish for slightly lower values in this case, this is still not an inadequate fit. Even though all five of the models fit adequately, it is possible to determine if any one model fits better than the others. Using the fact that changes in x^ across nested models are themselves distributed as x^> it becomes apparent that there are real differences in fit across the models of structure. Generally, the difference in x2 between the two and four factor models is significant; the change between two models whose only difference is in a Negativity factor is not significant. More specifically, for Dukakis and Quayle, both of the four factor versions of the model are significantly better fitting than all other models at the .OS level or better. In the case of George Bush, all models are significantly better than the One Factor model but, there is no significant improvement obtained after the Two Factor model. For Jackson, neither of the four factor versions of the model differ significantly from their two factor counterpart. There is, then, evidence that a four factor structure is needed to account for the connections between the traits. This is true for only two of the four candidates but, because the picture of structure sought is a generic one, this is not overly troubling. Removing just the negativity factor from either of the basic models does not result in a significant change in fit for either of the two 1988 major party nominees. The negativity factor does, however, play a role in how the activation networks are structured 73 for both Quayle and Jackson. For Quayle, the changes are significant at better than the 0.01 level. Jackson is a borderline case; for both forms of the model the difference does just make the .10 level of significance. (The critical value of x2 with 1 degree of freedom is 2.706; the value here is 2.8 for both the two and four factor comparisons.) Therefore, I conclude that the Negativity factor, as well as the four substantive factors of Competence, Leadership, Integrity, and Empathy are needed to account for the structure in the accessibility data. To sum up the discussion of the fit of the various models, the pattern revealed in the x^ values should be interpreted as support for a four factor model of the structure of trait judgments, using either the or the accessibility measures. In six out of eight cases, the exceptions being the Bush and Jackson accessibility data, this form of the model does fit significantly better than a two factor version. Moreover, there is a special role played by the negativity of a trait. Although the differences are more dramatic for the evaluation data, the utility of a negativity factor in modeling variation in accessibility is evident. In short, the structure for the two types of data are very similar to each other and both fit with results reported in the literature using national samples and survey methodology.

Parameter Estimates Having established which models fits the data best, it remains to discuss the actual estimates obtained. The estimates may be 74 broken into four main categories: loadings of the traits on the substantive factors, loadings of the negatively phrased traits on the negativity factor, the error variance for the trait measures, and the correlation between the latent variables or factors. The discussion will focus on each of these categories in order, considering the estimates themselves, the significance of the estimates, as well as potential problems with the respective matrices in which each set is found.

Substantive Factor Loadings. The estimates for the factor loadings of the traits on their respective substantive factors, for both data types, are presented in Table 4.4. The results are presented only for the Four + Negative model, the model exhibiting the best fit to the data. The values presented are for completely standardized solutions, so that, after the estimation itself is accomplished both the measured variables and the latent factors are standardized to have variances of 1.0. This makes the output conform more closely to that typically presented for exploratory factor analyses, in which the measured variables are standardized. This could be accomplished by analyzing matrices of correlation coefficients in the confirmatory context, as it is in the exploratory context, but this is not recommended as it may result in incorrect goodness of At estimates and/or standard errors.12 There are few surprises to be found in the estimates themselves. All were highly significant, much more than twice the size of their standard errors. Most ranged above 0.5, with the 75 average loading slightly higher for the evaluative data than the accessibility data. Perhaps the most surprising thing is the steadiness of the estimates across the various leaders; most traits have a spread of only 10 to 15 points. Large differences in loadings across candidates would indicate that different traits stand out as definitive of the trait dimension for the different leaders. The small differences demonstrate how well the generic model not only fits the structure but truly sums up the dimensions. Further information about the relationship between the measured variables and the latent factors can be found by examining the modification indices for the matrix of factor loadings. For every measured variable not allowed to load on a factor, LISREL computes an estimate for the change in fit for the whole model that would result if the variable were to be allowed to load on that factor. These values may be examined for patterns indicating specification problems. For example, if there are consistently high modification indices for the parameter corresponding to the loading for ‘moral’ on the leadership or capability factor, this would indicate a systematic relationship between morality judgments and the overarching capability factor, a relationship not expected to occur given previous findings. A close examination of these indices reveal no such systematic relationship across candidates. While there are certainly some parameters which have been fixed at zero that have relatively large modification indices, no single trait has such large loadings across several candidates. For example, removing the restriction that 76 judgments regarding cruelty are not at least in part due to underlying judgments regarding capability would, for the Bush accessibility data, typically result in an improvement in fit of about 12 points, a statistically significant improvement. This is not the case for either of the other three candidates or for any of the candidates for the trait attribution data. The differences across candidates fit in with the notion that each of the candidates has particularized patterns of loadings across the different factors, or even that there are candidate specific factors. There is no evidence in the modification indices that one could make across the board adjustments in which traits load on which factors that would result in large improvements in fit.

The negativity factor. Besides the loadings on these substantive factors, estimates were obtained for the loadings of the negatively phrased traits on the negativity factor. As explained earlier in this chapter, this factor was included to account for two possible processes which may affect both trait attributions and trait accessibility as measured using reaction times. First, it is possible that judgments about negative or undesirable traits are more difficult to make simply in terms of the processes through which people must go in order to make the judgment. For example, the overarching trait may be more likely to be stored in terms o f a desirable rather than an undesirable category (e.g., competence rather than incompetence). Thus, to make a judgment on a negative trait requires a greater expenditure of resources. Or, findings 77 concerning the need for a negative trait dimension may be explained by a socially programmed unwillingness on the part of subjects to express a negative judgment of a public figure, especially one that will be recorded. The origins of such a hesitancy may be rooted in such maxims as “If you can’t say anything good about a person, don’t say anything at all.” The loadings of the negatively phrased traits were restricted to be equal in the respective models. This guaranteed that the amount of variance in each trait judgment associated with the underlying dimension was equal. This does not, however, force the proportion of variance to be equal. Thus, in a standardized solution in which the values have been rescaled to account for the differing variances of each variable, the loadings will no longer show up as equal. The values in Table 4.5 are the values from the unstandardized solution and thus represent, when squared, the amount of variance in the negative traits associated with the negativity factor. The values presented in Table 4.6 are from the standardized solution and, when squared, are the proportion of variation associated with the negativity factor. The most obvious differences are between the two sets of loadings for the accessibility and attribution data, respectively. Across the board, the values are much higher for the five-point evaluations than for the accessibility measures. Indeed, the loadings barely attain significance for Bush in the accessibility data and do not do so for the accessibility of judgments about Dukakis. In all other cases, the t-test for the null hypothesis that the parameter is zero easily reaches the 0.05 level, indicating that the 78 measured variables do have a non-zero loading on the factor. In most cases, the proportion of variance in a trait associated with the negativity factor approaches 15 percent for the attribution measures but is more usually in the range of 2 to 4 percent for the accessibility measures.13 Overall, there is very little added to the model of the structure of accessibility by the negativity factor, in terms of these loadings. However, while not much of the variation in accessibility is dependent upon the negative nature of the trait words, the loadings are significantly different from zero and the fit statistics indicate it is a necessary factor. The decision to retain a negativity factor is the correct one.

Error Variances. There are noticeable differences in the amount of unique variance associated with the traits in the two sets of data and the size of the estimates of the unique variance across the different leaders within a single set of data. Table 4.7 presents the standardized estimates of the variance in the unique factors, also known as the error variance for the traits. Each value is the proportion of variance in the trait measure that is not due to any of the common factors but may instead be attributed to measurement error, the unique nature of the trait, or other common factors omitted from the model. Typically, the estimates obtained indicate that about one-third to one-half of the variance in the evaluation measures is unique variance. For the accessibility measures, the range is about one-half 79 to two-thirds. There are many explanations for the larger proportion unique variance in the accessibility measures, ranging from explanations based on the chronic accessibility of some traits for some of the subjects to the rather rough nature of the standardization to account for differences in reading speed and dexterity. In general, the difference in the average proportion of unique variance is not troublesome to the arguments presented in this research because they have no direct bearing on the fit of the model. If the four factor model, for example, is an adequate model of the covariance between the measured variables, this will be true whether the unique variances are 10 percent or 80 percent.

Inter-factor correlations. One of the central concerns of modeling the dimensional structure of traits must be with the necessity of each additional dimension in the final structure. Do people really divide judgments about the capability of a political figure into judgments about competence and judgments about leadership? Are the two sub-dimensions separable and distinct? One answer to this question was provided by the tests of fit comparing the various models. There, comparisons of indicated the Four + Negativity model was necessary to account for the generic structure across the candidates. Another way of assessing this is to look at the correlations between the factors and make substantive judgments about the tradeoffs between added complexity and the simplicity of a model with fewer dimensions. 80 In the model of interest here, in which the Capability and Sociability factors have been broken into two sub-factors, it is the correlations between Competence and Leadership on one hand and Integrity and Empathy on the other that are of special interest. The estimates of these correlations are presented in Table 4.8. The estimates for the evaluation data are quite high, hovering around 0.90. These compare favorably with estimates obtained by Kinder using a national sample (1986).14 The correlations between the pairs of sub-factors in the accessibility data are higher. The estimates for the correlations for Jesse Jackson are, after rounding, a perfect 1.0, indicating that the sub-factors are perfectly correlated and thus the four factors are really two factors. For Bush and Dukakis, the estimate just misses perfect correlation at 0.99. Only in the case of Dan Quayle does the estimate drop noticeably below 1.0. It is difficult to know what to make of these high correlations in and of themselves. In terms of substantive rather than statistical significance, it seems pointless to preserve the distinction between competence and leadership or between integrity and empathy. With respect to the rating data, the correlations are in the gray range and it is difficult to say, as Kinder said, that these narrower dimensions are “related but distinct elements of character (1986, 248).” Yet, the statistical tests indicate that, for Dukakis and Quayle, there is an important difference between the two and four factor versions of the structure. 81 D iscussion The search for a model of trait structure that tits across a range of political leaders has produced a picture that divides personality characteristics into four groups: competence, leadership, integrity and empathy. Additionally, it is known that there is something about negatively phrased traits that sets them apart from positive traits. The research presented here provides two-fold confirmation of this. The analysis of trait ratings closely replicates findings reported in the literature; the covariance of trait judgments may be attributed to the five latent factors. Analysis of the accessibility data tells much the same story. But more importantly, knowing the dimensional structure of the accessibility of trait judgments provides a more direct answer to the question of which traits are connected in the minds of citizens. It is the desire to answer that question, not simply a desire to know which evaluations covary, that should be the driving impetus behind dimensional approaches to the study of traits. After all, what does it mean if we know which trait judgments covary? The only real practical use that has been made of the information is in data reduction, allowing regression analyses to proceed using trait factor scores built around competence, leadership, integrity and empathy. On the other hand, knowing which traits are linked in accessibility networks will lead down several new paths by which we can understand political evaluation. Various aspects of the memory model underlying the results may be used to formulate hypotheses about how traits become linked. At the other end of the evalutation 82 process, questions concerning inference and evaluation may be addressed using the activation network concept, especially as it ties to notions of accessibility already in use in the discipline (Rahn, et al 1990; Iyengar 1990; Iyengar and Kinder 1987; Krosnick 1988a; ). Using the accessibility measures offers a straight-forward theoretical framework in which one may ground interpretations of factor structure. Traits which are associated in networks of activation, which tend to occur together in working memory and are stored together in long-term memory, will covary in accessibility. The evidence from these data certainly indicates there is good reason to preserve the distinctions, found in studies analyzing evaluative ratings, between competence and leadership and between integrity and empathy. The fit of the model with only two factors borders on adequate but certainly leaves room for improvement. Improvement is found by making the distinction between the four sub-networks. That this is statistically significant for only two of the four leaders examined is not overly damaging to the claim, because the search is for a generic model of structure. The generic model requires four trait sub-networks, as well as the negativity dimension. Given the present results, it may now be said that thinking about a trait from one of the four sub-networks makes a person more likely to think about other traits from that sub-network. This could not be said with any assurance when all that was known about trait structure was based on analyses of trait ratings. This is extremely important, given all that we already know about the impact of accessibility upon information use and, ultimately, 83 evaluation. Traits that are more accessible are more likely to be used for evaluation, information searches, or categorization of observed or reported behaviors. This chapter has relied upon one facet of associate memory networks, spreading activation, to test the notion that trait judgments are organized into four main groups. At what level does the activation spread—is it through the primary networks of Capability and Sociability or the sub-networks of Competence, Leadership, Integrity and Empathy? Exactly what role does the positivity or negativity of a trait play in the spread of activation among traits? The next chapter takes on these and other questions, using an experimental approach. There, further confirmation is found to support the picture of organization advanced in this chapter. 84 N otes

According to Fazio (1990), such instructions help minimize measurement problems due to inattentiveness and other processes not related to accessibility per se.

It is important to remember that the structure of the traits in these subjects need not be identical to that found in national samples. The real question here is if the structure found in the two types of trait data parallel each other. Because the expectation is that the evaluative structure and the associative memory networks are tightly related, it is also expected that these two will look more or less alike. (Interestingly enough, many of the analyses of the evaluative data do show a high degree of similarity to those done on national samples.)

In that early study, based on a more limited set of trait stimuli, Kinder et al. found two generic traits related to Capability and Sociability. These were Competence and Integrity.

The four-factor versions of the model are actually specified in such a way that the correlations between the two pairs of factors were constrained to be equal. This was done in order to simplify the models being tested. The analyses were run without this constraint and these results were compared to the constrained model. There was remarkably little difference between the two estimates. Typically, the correlation between competence and leadership was approximately 0.03 or so below the constrained correlation, while the correlation between integrity and empathy would be about 0.03 above it. Tests comparing the fit of the constrained and unconstrained model showed that, for the latencies data, the model did not fit significantly better for any of the four political figures (at the 0.01 level). For the evaluative data, only the case of Quayle showed significant differences at the 0.01 level. Given this, the simpler constrained model was retained.

In his 1986 work on trait structure, Kinder allows the negativity factor to be correlated with the substantive factors. While this will produce a better fitting model, simply because it has more free parameters, there seems to be little theoretical 85

justification for allowing this. The correlation estimates he presents are quite small.

6 . The ability to specify the models in such a manner is very important in assessing the relative fit of the models. As discussed by Hayduk (1987), the difference between values for nested models is itself distributed as x^ with the degrees of freedom for the difference equal to the difference in degrees of freedom between the two models. This is inherent also in the modification indices computed as a part of the LISREL output, which for fixed or constrained parameters are estimates of the expected change in x^ when one frees a parameter that is currently fixed in the model. This is the very definition of nested models. As Joreskog and Sorbom (1989 p. 45) discuss, the modification indices are themselves distributed as x^ with one degree of freedom. There may be small discrepancies between actual reduction in x^ and the modification index given that the modification index is only an approximation of the fit function (Joreskog and Sorbom 1989, 103). So, rather than simply using the values of the modification indices, each nested model was actually estimated and the x^ values themselves were used in calculations in differences of fit.

7. Actually, a prior consideration involved outliers and missing data. To begin, the distribution of response latencies was examined to ascertain if some subjects had had their attention wander during the measurement task. The distributional pattern indicated that the vast proportion of subjects had made their decisions in well under 5 seconds. To help eliminate problems with measurement error, response latencies above 8.0 seconds were coded to missing. (The total number of trials recoded was only 44 out of a possible 15360.) The three respondents with more than two missing values out of 96 trials were eliminated from the following analysis entirely.

8. Several methods of transformation are possible. Another form that was examined used the equation SXj = Xj - RS. Both forms of the standardization were computed and the resulting standardized latencies were entered into exploratory and 86

confirmatory models of the type described below. Because there was no a priori reason for choosing one form of the standardization over the other, the form which expresses the original latency as a ratio of the average Yes/No word recognition speed was chosen on the basis of the relative interpretability of the two sets of solutions.

9. The same procedure was followed as with the evaluation data. In this case, a two factor solution was deemed adequate by these criteria.

10. The maximum likelihood technique in LISREL 7 is an iterative procedure which minimizes the fit function F = log I III I + triSZ'l) - log I Isl I - (p + q). For a full discussion, see Joreskog and Sorbom (1989). This procedure has been shown to be fairly robust given small violations of the normality assumption. 11. These values compare favorably with the benchmark results achieved by Kinder (1986) in a study of 1983 evaluations of then-president Ronald Reagan. There, he reports obtaining a X^/df value of 2.02 (see his Table 10.2). It should be noted here that Kinder’s analysis incorporated a model with a slightly higher number of free parameters, since the negativity factor was allowed to be correlated with the other factors. Freeing these parameters in the current analysis would, by definition, result in a lower x^/df value than is presented. Thus, the fit here compares more than favorably with Kinder’s.

12. The analysis of correlation matrices is especially problematic given the constrained parameters in the models including the negativity factor. See Joreskog and Sorbom, Section 1.21 (1989) for a discussion.

13. The proportion of variance in any measured variable associated with a factor is obtained by squaring the loading of that variable on the factor in the standardized solution.

14. In that analysis, Kinder found correlations around 0.90. See his Table 10.3. Commands R espect f Strong

in s p irin g

Easily Influenced Weak

No D irection

FIGURE 4.1

An example of a network of activation. Solid arrows represent connections within a sub-network or dimension. Dashed arrows represent connections with other sub-networks. Commands R espect Strong

Inspiring

LEADERSHIP Easily Influenced Weak

No D irection

FIGURE 4.2 The factor analytic view of a network of activation.

Solid arrows represent connections within a sub-network or dimension. Dashed arrows represent connections with other sub-networks. 89

Hardworking Intelligent Knowledgeable Inexperienced Unqualified Foolish

Influential A Leader Inspiring Weak Aimless Hesitant

Decent Moral Trustworthy Power-Hungry Dishonest Corrupt

Compassionate Friendly Understanding Prejudiced Unkind Cruel

FIGURE 4.3

The 24 traits used in the study. 90

DAN QUAYLE

TRUSTWORTHY?

Figure 4.4

An example of how the stimuli appeared on the screen. 91

Figure 4.5 Nesting The Various Models

Phi for One Factor Model Competence Leadership Integrity Empathy Negativity Competence 1.0 Leadership 1.0 1.0 Integrity 1.0 1.0 1.0 Empathy 1.0 1.0 1.0 1.0 Negativity 0.0 0.0 0.0 0.0 1.0

Phi for Two Factor and Two + Negative Models Competence Leadership Integrity Empathy Negativity Competence 1.0 Leadership 1.0 1.0 Integrity ¥ ¥ 1.0 Empathy ¥ ¥ 1.0 1.0 Negativity 0.0 0.0 0.0 0.0 1.0

Phi for Four Factor and Four + Negative Models Competence Leadership Integrity Empathy Negativity Competence 1.0 Leadership ¥ 1.0 Integrity Free Free 1.0 Empathy Free Free ¥ 1.0 Negativity 0.0 0.0 0.0 0.0 1.0

Notes: Cells containing "1.0" or "0.0" indicate that the correlations were fixed to those values. "Free" indicates that the correlations were estimated by LISREL without further constraints. "¥" indicates that the estimates of the correlations were constrained to be equal for the model(s) in question. 92 Table 4.1 *

Exploratory Factor Analysis of Trait Attributions

George Bush

Trait Factor 1 Factor 2 Factor 3 Factor

Hardworking .49 .31 . _ Intelligent .38 .39 - - Knowledgeable .31 .49 - - Inexperienced -- - .69 Unqualified -- - .73 Foolish - -.31 - - Influential - .50 .33 - A Leader - .72 - - Inspiring .47 .61 - - Weak - -.71 - - Aimless - -.48 .50 - Hesitant - -.50 - - Decent .54 -- _ Moral .50 - -- i Trustworthy .66 - O - Power-Hungry - - .50 - Dishonest -- .74 - Corrupt -- .60 - Compassionate .83 -- - Friendly .66 - - - Understanding .68 -- - Prejudiced -- .50 - Unkind -- .61 - Cruel -- .56 -

Inter-factor Correlation

Factor 1 Factor 2 Factor 3 Factor 2 .30 Factor 3 -.42 -.11 Factor 4 -.19 -.27 .28

* Only loadings greater than 0.30 are shown. 93 Table 4.1 (Continued)

Exploratory Factor Analysis of Trait Attributions

Michael Dukakis

Trait Factor 1 Factor 2 Factor 3 Factor

Hardworking - .55 -- Intelligent -.33 .51 -- Knowledgeable - .63 - - Inexperienced - - - .73 Unqualified - -- .92 Foolish .41 --- Influential - .64 -- A Leader - .72 - - Inspiring - .74 -.53 - Weak -- .35 - Aimless .30 - .55 - Hesitant -- - - Decent -.43 .37 -- Moral -.63 .31 -- Trustworthy -.44 .57 - - Power-Hungry .55 - -- Dishonest .60 - -- Corrupt .57 - -- Compassionate -.61 .37 -- Friendly -.50 .36 -- Understanding -.63 .31 -- Prejudiced .52 -- - Unkind .67 - - - Cruel .73 - - -

Inter-factor Correlation

Factor 1 Factor 2 Factor 3 Factor 2 -.37 Factor 3 .11 -.09 Factor 4 .30 -.28 .12 94 Table 4.1 (Continued)

Exploratory Factor Analysis of Trait Attributions

Dan Quayle

Trait Factor 1 Factor 2 Factor 3 Factor

Hardworking .50 - .39 - Intelligent .34 - .43 - Knowledgeable .43 - .41 - Inexperienced - - - .98 Unqualified - -- .66 Foolish -.36 .46 -- Influential .75 - - - A Leader .75 --- Inspiring .77 --- Weak -.57 --- Aimless -.60 -- - Hesitant -.55 --- Decent - -.35 .53 - Moral -- .60 - Trustworthy -- .61 - Power-Hungry - .51 - - Dishonest - .70 -- Corrupt - .69 - - Compassionate -- .79 - Friendly - -.41 .47 - Understanding -- .64 - Prejudiced - .61 - - Unkind - .68 -- Cruel - .85 --

Inter-factor Correlation

Factor 1 Factor 2 Factor 3 Factor 2 -.19 Factor 3 .37 -.40 Factor 4 .03 -.08 .11 95 Table 4.1 (continued)

Exploratory Factor Analysis of Trait Attributions

Jesse Jackson

Trait Factor 1 Factor 2 Factor 3 Factor

Hardworking _ .44 -.33 - Intelligent - .38 -.35 - Knowledgeable - .53 -- Inexperienced -- - .82 Unqualified - - - .87 Foolish .45 --- Influential - .50 -.47 - A Leader - .36 -.52 - Inspiring - .57 -.34 - Weak -- .64 - Aimless - - .65 - Hesitant -- .49 - Decent -.34 .60 - - Moral -.48 .42 -- Trustworthy - .67 -- Power-Hungry .54 - - - Dishonest .69 - - - Corrupt .71 - -- Compassionate - .68 -- Friendly - .65 -- Understanding - .62 - - Prejudiced .58 - -- Unkind .66 -- - Cruel .72 - --

Inter-factor Correlation

Factor 1 Factor 2 Factor 3 Factor 2 -.41 Factor 3 .24 -.39 Factor 4 .19 -.22 .20 Table 4.2 *

Exploratory Factor Analysis of Standardized Latencies

George Bush

Trait Factor 1 Fac

Hardworking .33 .49

Intelligent .57 -

Knowledgeable .49 -

Inexperienced .46 -

Unqualified .71 -

Foolish .53 - Influential .37 - A Leader .40 -

Inspiring - .50

Weak .46 - Aimless .36 -

Hesitant .41 - Decent -- Moral - .69

Trustworthy .58 -

Power-Hungry .63 -

Dishonest .49 -

Corrupt -- Compassionate .32 .36

Friendly .43 • Understanding .31 .47

Prejudiced .45 - Unkind - .42 Cruel - .87

Inter-factor Correlation = .48

* Only loadings greater than 0.30 are shown. Table 4.2 (Continued)

Exploratory Factor Analysis of Standardized Latencies

Michael Dukakis

Trait Factor 1 Factor

Hardworking - .34

Intelligent - -

Knowledgeable .50 -

Inexperienced .55 -

Unqualified .55 -

Foolish .35 - Influential .64 -

A Leader .43 -

Inspiring .53 -

W eak .54 - Aimless .41 - Hesitant - .33 Decent .41 - Moral - .53

Trustworthy .63 -

Power-Hungry - -

Dishonest - .55 Corrupt .41 .38

Compassionate - .64 Friendly .34 .35

Understanding - .84

Prejudiced .51 -

Unkind .44 - Cruel .54 -

Inter-factor Correlation = .47 Table 4.2 (Continued)

Exploratory Factor Analysis of Standardized Latencies

Dan Quayle

Trait Factor 1 Factor 2

Hardworking - .32 Intelligent - .48 Knowledgeable - .69 Inexperienced .36 .35 Unqualified - .65 Foolish .38 Influential - .39 A Leader - .59 Inspiring - .64 W eak - .48 Aimless - .49 Hesitant .46 Decent .44 Moral .48 Trustw orthy - .31 Power-Hungry .59 Dishonest .51 Corrupt .78 Compassionate .41 Friendly .44 Understanding .33 Prejudiced .43 Unkind .49 Cruel .47

Inter-factor Correlation = .44 Table 4.2 (Continued)

Exploratory Factor Analysis of Standardized Latencies

Jesse Jackson

T rait Factor 1 Fac

Hardworking .60 _

Intelligent .57 -

Knowledgeable .40 -

Inexperienced - .43

Unqualified .32 -

Foolish - .68 Influential - .83 A Leader - .58

Inspiring - .39

W eak .41 - Aimless .55 - H esitant .48 - Decent .47 - Moral .73 -

Trustworthy .39 -

Power-Hungry .50 -

Dishonest .46 -

Corrupt .41 - Compassionate .37 .31

Friendly .48 -

Understanding .60 -

Prejudiced .41 -

Unkind .58 - Cruel .50 ~

Inter-factor Correlation = .44 Table 4.3

Assessing the Fit of the Five Models

The Evaluative Structure

Model Bush Dukakis Quayle Jackson

df X2 X2/df X2 X2/df X2 X2/df x2 X2/df 1 Factor 252 956.1 3.8 807.6 3.2 1129.7 4.5 913.6 3.6

2 Factor 251 753.2 3.0 707.1 2.8 873.8 3.5 756.4 3.0

2 + Negative 250 542.8 2.2 523.0 2.1 690.8 2.8 645.3 2.6

4 Factor 247 715.7 2.9 669.9 2.7 814.7 3.3 719.6 2.9

4 + Negative 246 490.5 2.0 465.4 1.9 625.0 2.5 592.8 2.4 Table 4.3 (Continued)

Assessing the Fit of the Five Models The Accessibility Structure

Model Bush Dukakis Quayle Jackson

d f X2 X2/d f X2 x 2/d f X2 X2/d f X2 X2/d f

1 Factor 252 608.2 2.4 556.1 2.2 591.6 2.3 504.5 2.0

2 Factor 251 590.4 2.4 543.6 2.2 547.6 2.2 495.8 2.0

2 + Negative 250 589.4 2.4 543.5 2.2 538.2 2.2 493.0 2.0

4 Factor 247 585.1 2.4 518.9 2.1 529.5 2.1 493.4 2.0

4 + Negative 246 584.1 2.4 518.8 2.1 519.2 2.1 490.6 2.0 102 Table 4.4

ML Estimates of Factor Loadings for the Traits Confirmatory Factor Analysis

Evaluative Measures

Trait Bush Dukakis Quayle Jackson

Hardworking .77 .76 .82 .67 Intelligent .74 .73 .74 .80 Knowledgeable .70 .83 .79 .77 Inexperienced .42 .39 .48 .45 Unqualified .60 .43 .64 .49 Foolish .57 .55 .66 .70 Influential .65 .71 .86 .74 A Leader .77 .81 .80 .84 Inspiring .83 .71 .89 .82 Weak .67 .53 .66 .63 Aimless .55 .46 .71 .59 Hesitant .48 .26 .55 .49 Decent .74 .74 .75 .82 Moral .71 .79 .80 .77 Trustworthy .89 .84 .88 .85 Power-Hungry .48 .46 .50 .58 Dishonest .79 .66 .74 .73 Corrupt .62 .52 .63 .68 Compassionate .82 .83 .77 .79 Friendly .78 .75 .75 .78 Understanding .82 .81 .77 .72 Prejudiced .50 .53 .62 .56 Unkind .66 .54 .65 .68 Cruel .60 .59 .63 .68 103 Table 4.4 (Continued)

ML Estimates of Factor Loadings for the Traits Confirmatory Factor Analysis

Accessibility Measures

Trait Bush Dukakis Quayle Jackson

Hardworking .67 .50 .53 .45 Intelligent .67 .51 .60 .55 Knowledgeable .66 .63 .68 .51 Inexperienced .54 .60 .52 .49 Unqualified .75 .61 .56 .54 Foolish .64 .51 .52 .59 Influential .57 .59 .49 .54 A Leader .53 .54 .65 .63 Inspiring .46 .64 .69 .49 Weak .34 .57 .49 .53 Aimless .39 .55 .56 .52 Hesitant .44 .50 .60 .54 Decent .57 .56 .56 .57 Moral .65 .52 .63 .63 Trustworthy .59 .66 .51 .53 Power-Hungry .57 .59 .64 .57 Dishonest .52 .53 .49 .64 Corrupt .66 .61 .63 .53 Compassionate .64 .62 .56 .58 Friendly .54 .62 .63 .68 Understanding .64 .60 .56 .65 Prejudiced .56 .55 .54 .50 Unkind .51 .67 .62 .48 Cruel .61 .64 .61 .50 104 Table 4.5

Non-standardized Loadings of Negative Traits on Negativity Factor

Evaluative Measures

Two + Negativity Four + Negativity

Bush .36 .36

Dukakis .42 .42

Quayle .45 .45

Jackson .40 .42

Accessibility Measures

Two + Negativity Four + Negativity

Bush .17 .18

Dukakis -.11 .14

Quayle .41 .42

Jackson .29 .29 Table 4.6 105

Standardized Loadings of Negative Traits on Negativity Factor

Evaluative Measures

Trait Bush Dukakis Quayle Jackson

Inexperienced .41 .40 .32 .28 Unqualified .47 .39 .31 .29 Foolish .42 .41 .33 .35 Weak .34 .37 .31 .34 Aimless .40 .42 .31 .37 Hesitant .31 .39 .36 .35 Power-Hungry .32 .33 .36 .27 Dishonest .41 .45 .38 .35 Corrupt .47 .43 .42 .37 Prejudiced .35 .43 .42 .27 Unkind .48 .42 .48 .42 Cruel .52 .51 .53 .42

T-Values 13.62 13.34 13.15 12.03

Accessibility Measures

Trait Bush Dukakis Quayle Jackson

Inexperienced .10 .06 .19 .13 Unqualified .11 .07 .18 .15 Foolish .12 .08 .24 .14 Weak .08 .07 .21 .18 Aimless .09 .06 .19 .14 Hesitant .09 .06 .17 .16 Power-Hungry .10 .08 .26 .18 Dishonest .10 .07 .21 .17 Corrupt .15 .07 .23 .13 Prejudiced .13 .08 .17 .12 Unkind .13 .08 .23 .14 Cruel .12 .07 .25 .17

T-Values 2.01 0.78 5.62 3.17 106 Table 4.7

ML Estimates of Error Variance for the Traits

Evaluative Measures

T rait Bush Dukakis Quayle Jackson

Hardworking .41 .42 .34 .56 Intelligent .45 .47 .45 .36 Knowledgeable .51 .32 .38 .41 Inexperienced .66 .69 .67 .72 Unqualified .42 .66 .50 .68 Foolish .51 .53 .46 .39 Influential .58 .49 .26 .45 A Leader .40 .35 .36 .30 Inspiring .31 .49 .21 .33 W eak .44 .59 .48 .49 Aimless .54 .62 .41 .52 Hesitant .67 .78 .56 .64 Decent .45 .45 .43 .32 Moral .49 .38 .36 .41 Trustworthy .21 .29 .22 .28 Power-Hungry .67 .68 .62 .60 Dishonest .22 .36 .31 .35 Corrupt .40 .54 .42 .41 Compassionate .33 .31 .41 .38 Friendly .39 .44 .43 .40 Understanding .33 .34 .40 .49 Prejudiced .63 .53 .45 .61 Unkind .34 .54 .35 .37 Cruel .36 .39 .32 .35

Note: The values presented are from the completely standardized solution; thus they represent the proportion of variance in the measured variables that is unique variance. 107 Table 4.7 (Continued)

ML Estimates of Error Variance for the Traits

Accessibility Measures

T rait Bush Dukakis Quayle Jackson

Hardworking .56 .76 .72 .80 Intelligent .55 .74 .64 .70 Knowledgeable .57 .61 .54 .74 Inexperienced .69 .63 .70 .74 Unqualified .43 .62 .65 .69 Foolish .57 .73 .68 .64 Influential .68 .66 .76 .71 A Leader .72 .71 .57 .61 Inspiring .79 .59 .52 .76 W eak .88 .67 .71 .69 Aimless .84 .69 .65 .71 Hesitant .80 .75 .61 .69 Decent .68 .68 .69 .67 Moral .58 .73 .60 .61 Trustworthy .65 .57 .75 .72 Power-Hungry .66 .65 .52 .64 Dishonest .72 .72 .71 .56 Corrupt .54 .62 .55 .70 Compassionate .59 .62 .69 .67 Friendly .71 .62 .61 .54 Understanding .59 .64 .69 .58 Prejudiced .67 .70 .69 .74 Unkind .73 .55 .57 .75 Cruel .61 .59 .57 .72

Note: The values presented are from the completely standardized solution; thus they represent the proportion of variance in the measured variables that is unique variance. Table 4.8

Estimates of Inter-factor Correlations Confirmatory Factor Analysis

Competence with Leadership and Integrity with Empathy

Data Bush Dukakis Quayle Jackson

5 Pt. Attributions .91 .87 .92 .90

Accessibility .99 .99 .92 1.0

Note: The estimates of the correlation of the Competence and Leadership factors and of the Integrity and Empathy factors were constrained to be equal in the various models. V. An Experimental Approach to Assessing Trait Structure

Abstract. The associative model of memory implies that related traits will act as primes to each other during information processing tasks. As a consequence, one way of assessing the relatedness of traits is to assess their utility as primes for each other. An experiment was conducted in which the accessibility of trait dimensions was manipulated. Subjects were primed with a trait word then asked to make a judgment about a political candidate on another personality trait. Primes that are closely related to the trait being evaluated should have an effect on the time it takes to make the evaluation. Regression analysis was used to test hypotheses about the trait structure. Are traits organized along the dimensions of Capability and Sociability; is there a more detailed distinction between Competence, Leadership, Integrity and Empathy; or do both of these models fail to account for priming effects among the traits? The analysis indicates that a two-stage four-factor model of trait structure is necessary to account for the results.

109 110 Introduction Ultimately, the topic to which this research is addressed is how judgments of personality are made and how they affect overall evaluation and choice of political leaders by average citizens. More narrowly, this research is focused on how trait evaluations are linked together to form a cohesive picture of a leader’s entire personality, or at least of those aspects of personality which prior research has found to have a bearing on vote choice. These linkages are important in that they may provide, consciously or not, cues to an individual as he or she processes new information or seeks to make an overall judgment about a familiar or unfamiliar political person. In a naive reflection of the terminology of associative networks, we often ask, both of ourselves and of others, what was going through our minds as some decision was made or some action taken. In the context of survey research, questions which ask respondents to reveal reasons for behaviors or attitudes do this same thing. The reason for this concern is the belief that knowing what people were thinking about can help us understand how and why they came to the conclusions that they did. Knowing what was going through their heads as they took in new information, even if they didn’t realize it, can help us understand how voters came to hold certain beliefs and attitudes. In studies outside of the political context, accessibility has been repeatedly shown to have a significant impact on interpretation and recall of information (Srull and Wyer 1979, 1980, 1986; Lingle and Ostrom 1979; Higgins, Rholes and Jones 1977; Higgins, Bargh, and 111 Lombardi 198S; and many others). The determinants of construct accessibility are myriad. They include, but are not limited to, recency of activation, frequency of activation, affective or emotional states, and the level of activation of related constructs. (For a full review, see Higgins and King 1981.) In terms of the topic at hand, judgments about specific personality traits and overall personalities of political figures, the approach I have taken has focused on only one of these determinants, the activation of related traits. I have done so because of the importance of understanding which traits are related, which groups of traits form dimensions of judgment and networks of activation. When a voter considers a candidate for office, what aspects of the candidate’s personality weigh in the decision and how do they do so? When any citizen considers a political leader, along what lines do his or her thoughts form? This is an age old question, one of interest to political practitioners as well as students of politics. Machiavelli, in his instruction manual for political leaders, debates such questions as wether it is better for a prince to be loved or to be feared. One possibility is that voters simply think in terms of the likableness of a candidate, considering various personality traits in terms of their positivity or negativity, whether the trait inspires love or hate. Machiavelli, had his interests been in explaining the behavior of individual citizens, would have recognized this model of trait structure. Or, it is possible that groups of traits cluster together to form overarching personality dimensions that are used as evaluative standards. In this case, citizens have learned, one way or 112 another, that certain traits occur together in their political and social environment. The research reported in the previous chapter lends some support to already existing pictures of the trait structure underlying how citizens apprehend the personality of political figures. Dimensional analyses of the accessibility of trait judgments roughly parallel the results obtained from analyses of trait evaluation data. There are two highly related dimensions best described as Capability and Sociability. These are made up of four trait sub-networks, two within each of the main dimensions. These have been labeled Competence, Leadership, Integrity and Empathy, following the lead of Kinder (1986). That research relied upon dimensional analysis techniques that attempt to reproduce the correlation between the individual measures of accessibility based on models of trait structure. While it was conducted in a laboratory setting and used measures not ordinarily used by political scientists, the study was not experimental in nature; no independent variable was manipulated. It is, however, possible to use an experimental approach to test hypotheses about the structure of trait accessibility. This chapter reports on an experiment involving priming subjects with trait words and measuring changes in the accessibility of other traits.

Priming as a test of relatedness One of the fundamental implications of the associative network model of memory is that activational energy is transmitted 113 throughout the network when any node of the network is activated. The more related the nodes, the greater the increase in accessibility of the as yet unactivated node or, put another way, the greater the chance that the second node will be activated. Activation itself means that the node, the information, is brought from the more-or- less unlimited storage area of long-term memory into the more limited aspect of consciousness referred to as working memory. This model allows us to test the relationship between nodes experimentally by manipulating the accessibility of some node Nj and then measuring the accessibility of a different node N2. If the nodes are highly related, there should be some change in the accessibility of N2. If they are not, no change should occur.

There are a number of priming strategies that might be used to manipulate accessibility. Typical strategies used by psychologists involve the use of a cover story to hide from the subjects the fact that they are deliberately being exposed to primes. In unrelated- task priming, subjects are told that they are participating in several different short studies. In the “first” study, subjects perform tasks that expose them to the priming stimuli in such a way that the stimuli are an unremarkable part of the study. For example, Higgins, Rholes, and Jones (1977) told subjects they would be engaging in two unrelated studies, one on color perception and another on impression formation. As part of the color perception study, subjects were given a “memory” word that they had to repeat after reporting the color of each stimulus. The word was in actuality a trait prime. Srull and Wyer (1979) devised an 114 interesting task that had subjects construct three word sentences from a list of four words. Only one sentence was possible from each set. This sentence exemplified the trait to be primed. In both of these cases, the subjects were more or less aware of the trait information presented. In subliminal priming, this is not the case. Bargh et al. (1986) has used subliminal priming to present trait words outside of conscious awareness. This was done by asking subjects to identify the location of flashes on a computer screen. The flashes, unbeknownst to the subjects, were actually trait words presented for less than 100 milliseconds. After identifying the location of 100 flashes, the subjects participated in a different study, this one involving impression formation. These types of primes, while very useful in studying the effects of accessibility on evaluation directly, are of limited usefulness here. First, they can only be used to manipulate the accessibility of one trait category at a time. The category must be defined in advance by the researcher. Thus, this strategy is not suited to the present problem, in which the categories themselves are to be defined. Second, the manipulations are separated in time from the measurement of the dependent variable. It is unclear how noticeable the changes in response latency would be after such long delays, due to the quick decay of rather high levels of activation needed to affect response time (Fazio 1990; Lingle 1982). In studies in which the priming stimulus is not rehearsed, activation levels (as measured using response latencies) have been found to 1 15 decay within four to fifteen seconds (Warren 1972; Meyer, Schveneveldt, and Ruddy 1975). An alternate strategy, one that allows the prime and the trait to be of indeterminate relationship and brings the two closer together, is required. Fortunately, one exists. This alternate priming paradigm is typically used by theorists developing or testing memory models per se, rather than how those models affect evaluation. In those types of studies, as in the current one, the focus is on structure and the measured dependent variable, an indicator of accessibility, is response latency. In a task commonly known as probe recognition, subjects are primed directly with a category word, either consciously or subliminally, then asked to identify another word presented on a computer screen (Lingle 1982; Lorch 1982; De Groot 1983). Members of a category should be, and are, recognized faster if preceded by their category label. In a related fashion, Fazio (1990) reports on a study in which probe recognition time was used to assess the strength of the association between an exemplar of a category (brands of a product) and the category label (the type of product, e.g. credit cards), recasting the interpretation of response latency. As Lingle (1982) points out, using probe recognition time as a measure of activation has several advantages. Chief among these is that it may be used to determine “whether or not a single concept has been activated in memory during a designated time span (483).” Moreover, it can detect “a thought briefly activated in close 116 proximity to other thoughts (483).” This is exactly what is needed in the present research, where the accessibility check is to be made on a fairly large number of trait words in quick succession. The probe recognition paradigm may be used to formulate a priming strategy to assess the relatedness of various personality traits and to test hypotheses about if and how those traits are organized in memory. In short, the accessibility of some trait Tj may be manipulated by exposing a subject to the trait word and testing to see if such exposure creates changes in the amount of time it takes to process other trait words. Priming with words that are highly related to the trait being assessed should produce faster judgment times, since the time spent retrieving the second word from long-term memory is shortened due to increased accessibility. With enough subjects and careful design, it would be possible to sketch out whole networks, indicating the strength of the ties between trait nodes within a network and across networks. While this latter notion presents the researcher with interesting possibilities, it is not within the scope of this research to attempt such a task. Instead, the goal here is simply to test to see if a simple two factor or two network model of the structure of trait judgments is sufficient or if a more complex four factor model is necessary to obtain a truer picture of trait organization. Along the way, comparisons will be possible across various candidates and types of traits as well. 117 P ro c e d u re The study was conducted over a period of 3 weeks in October and November of 1990. Subjects were volunteers from introductory political science courses who received extra credit for their participation. A total of 164 subjects participated in the study. Three of the subjects were excluded from the analysis after it was determined that these subjects had failed to follow the directions and/or were not native speakers of English. Groups of subjects, ranging in size from four to eight persons, were introduced to the study by the reading of a prepared statement explaining that the study concerned their “opinions on political candidates and some political issues.” They were again instructed that they should not take too long in making up their minds since the researchers were interested in their first reactions. Subjects were also told that part of the study involved their ability to evaluate political candidates in the presence of “distractions.” They were told that there would be other things flashed onto the computer screen as they made their judgments, but they were to not pay attention to this. They were also told that the trait on which they were to evaluate the political person would only be on the screen for a very short while, so they were to watch the screen at all times. These instructions were designed to keep the subjects’ attention on the computer screen but not necessarily on consciously processing the priming word.1 Exposure to the word in and of itself should be enough to elicit the priming effect. 118 In fact, this is a modification of a method used frequently to assess the automatic spread of activation (Fowler, Wolford, Slade, and Tussinary 1981; De Groot 1983; Bargh, et al. 1986). In studies of that type, primes are typically presented in such a way as to avoid conscious perception by the subjects. This insures the activation is due to mere exposure to the prime. In the present research, a cover story and short prime duration was used so that the subjects did not dwell on the fact that they were being primed; this might produce a change in the way they react to the judgment task. The presentation is not, however, subliminal. As in the study reported on in the previous chapter, subjects were assigned to individual cubicles in order to reduce distractions that might cause increases in the measurement error associated with the accessibility of their judgments. The research was again conducted through the use of microcomputers. Subjects were seated in front of the computer and activated the program themselves by striking a key. After being reminded of the instructions by the machine, they were given a chance to go through four practice trials. For this study, a trial consisted of the presentation of a candidate name, the flash of a prime, and the presentation of a trait for evaluation. The name of the person was presented near the top of the screen and was visible for the duration of the trial. At the same time, an "X" appeared at the center of the screen. After two seconds, to allow the subject time to focus on the person named, the priming word appeared at the center of the screen, in the place marked by the X. This word stayed visible for 100 milliseconds then 119 was written over by a blank field for 100 milliseconds. The blank field allows the subject to continue processing the prime without interference from new information, thus completing the activation process, without allowing him or her to continue to consciously identify the word. This is frequently used in subliminal priming studies (Bargh and Pietromonaco 1982; Bargh, Bond, Lombardi, and Tota 1986; De Groot 1983). After this, the trait word appeared at this same location. The trait word stayed visible for one second, at which time both it and the candidate name were removed from the screen. The subject responded by pressing one of two keys marked “YES” and “NO”, indicating that the word described or did not describe the person named. Following a delay of five seconds after the subjects response, the next trial was initiated. Each subject made a total of 144 judgments in this manner, in three sets of 48 with a one minute rest period between the sets.

Stimuli development. The trait word list developed for the first study was used as a source of trait words for this study as well. Since it was anticipated that only a rather limited number of subjects would be available for the study, and that asking each subject for a very large number of judgments would certainly introduce problems such as respondent fatigue into the data collection, the number of traits used in the study was reduced, as was the number of political figures. It was decided that the maximum number of trials for each subject should be approximately ISO. Since other aspects of the 120 original design included having a trial for every subject in which each person-trait pair was primed with a non-word, the number of traits that could be used was fairly limited. Sixteen traits from the original 24 were chosen, two positively phrased and two negatively phrased from each of the four factors of competence, leadership, integrity and empathy. (See the list in Table 5.1.) The number of political figures used was cut from four to three. I wished to drop neither Bush nor Dukakis, these two having been the most recent nominees form the two major parties. Jackson, it was felt, might have interesting differences due to his prominence in American presidential politics without ever having been nominated either for president or vice-president by the Democrats. This left Vice-President Dan Quayle, who was dropped as the least interesting of the four persons included in the first study. A total of 48 person-trait pairings were possible with these stimuli. Each of these pairs appeared once in each of the three sets of 48 trials that made up the experiment. The 48 were randomized prior to the study, so that no pattern existed in how they were presented. The trait pairs always appeared in the same order in the set of 48 and in the same order for all subjects. This procedure held constant any contaminating priming effects due to previous judgments. Thus if the person-trait pair Jackson-Honest appeared before the pair Bush-Intelligent on a trial in which the prime was highly related, it did so for every trial. A total of ten primes were included in the experiment, at least one positively phrased and one negatively phrased from each of the 121 four networks or factors found in the literature. (See Table 5.1.) Also, a list of non-words (Table 5.2) was created as part of the original design to assess a baseline for each subject. These non­ words were designed to resemble words (rather than using a string of XXXX’s, for example); this gives a better baseline because the word still demands the subject’s attentional resources and triggers a memory search for the word (de Groot, Thomassen and Hudson 1982). Each subject rated every candidate on every trait three times over the course of the 144 trials, twice being primed with actual trait words and once with a non-word. There were a total of 384 combinations of person-prime-trait (3 X 8 X 16). Approximately 40 subjects saw each person-prime-trait combination.

Other measures. Following the 144 priming trials, subjects were asked to complete another short task on the computer. As in the first study, this task was designed to measure their manual dexterity and overall response speed so that these factors could be controlled in the analysis of the response latencies. There were six trials in which the subject was asked to strike the “YES” or the “NO” button depending on which of those words appeared on the screen. After this the subjects completed a pencil and paper questionnaire on which was included the trait evaluations for the political figures in a five-point response format (ranging from “does not describe” to “describes very well”). Other questions were also included on the questionnaire. (See appendix for the full text of the questionnaire.) 122 A typology of trait relationships The analysis of the reaction time data is initially aimed at answering one very straightforward question - does the nature of the prime, in terms of its relationship to the trait being judged, affect the time it takes to make the judgment? The relationship between the prime and the trait can, in this study, fall in one of eight categories, as shown in Table S.3. The first degree of relationship is merely that the word used as a prime be an actual trait word. The second dichotomous characteristic that can be used to assess the degree of relationship is the evaluative direction of the prime and trait. If the words are both positive or valued traits, they may be expected to be more highly related than if one word is indicative of a desirable personality trait and the other is a negative or undesirable characteristic. Evidence of this type of relationship has been found by Fazio in his work on accessibility in connection with attitude strength (Fazio, Sanbonmatsu, Powell, and Kardes 1986). More substantively interesting for the project at hand are the next two categories. The central concern here is the idea that the traits are organized into associative memory networks. Given the current literature on trait structure, as well as the results of the first study reported in this work, there are two possibilities for what these networks might look like. The simpler picture is one in which trait judgments are organized around two overarching trait constructs: capability and sociability. This model will be called the Primary Network Model. The second, more complex model, is one in which the trait structure is more highly developed, being organized 123 around the four macro-traits of competence, leadership, integrity and empathy. This model is most aptly called the Sub-network Model since, as it was outlined by Kinder (1986) and further developed in the previous chapter, competence and leadership may be said to be sub-factors or sub-networks of capability while integrity and empathy are sub-networks of sociability. So, for any prime-trait pairing it is possible that the prime and trait are from the same primary network or not, or from the same sub-network or not. Of course, any prime and trait which are from the same sub-network are necessarily from the same primary network. This nesting presents troublesome possibilities for the analysis of priming effects that must be carefully considered. Even if the Sub-network Model is the more accurate model, it is still possible that there will be no observed differences between priming effects for traits related at the sub-network level and traits related at the network level due to the high degree of relationship between the competence and leadership sub-networks on the one hand and the integrity and empathy sub-networks on the other. It will be important then, when setting up the analysis, to do so in a way that allows for detection of network level effects separate from sub­ network effects while still insuring that sub-network effects are not really due to the second order relationship. Results which show effects at both levels are indicative of a two-stage model of trait structure but substantively there is little reason to differentiate between this and the Sub-Network model. 124 The final column in Table 5.3 indicates that there may be even greater priming effects if the word used as a prime is the same word that is being judged. While this is not as substantively interesting as the other types of relationships, this type of trial was included in the design as well. One reason for this was as a check on the priming technique used in the study. It was a possibility that no effects would be present for the other types of relationship (same evaluative direction, same network, et cetera). If this has happened but there were effects for the same word trials, the failure for the other types of relationship could have been attributed to the failure of the models rather than a failure of the technique. If no priming effects were found for any of the types of relationship between prime and trait, including the same word trials, the technique itself would be suspect.

A check on nonword and same-word differences Before actually testing the various relationships between traits, it is perhaps appropriate to take a look at the unstandardized mean response times across several of these categories of relationship to see if there are any observable differences. Table 5.4 shows the breakdown in mean response times for all trials, as well as a comparison of means for real-word trials and same-word trials, by political person.2 The first thing one notices about the means is that there are certain disparities in decision time across the three leaders. Judgments concerning Bush are typically made the fastest, taking on 125 average 1.38 seconds, no doubt due to the general prominence of the current incumbent president in the minds of the subjects. Jackson ratings are made the second fastest and Dukakis judgments take the longest to make. Aside for the fact that Bush is better known, not too much can be made of this. It should be noted that these differences must be taken into account when considering all of the cases as a whole.3 More interesting are the differences in response time between trial in which the prime was a fabricated non-word and those in which it was a trait word. In the overall comparison, not separating the trials by leader, there is a small difference in the expected direction. The mean response time, and thus the level of activation of the trait being evaluated, is shorter when the subject has been primed with another trait word (1.48 seconds to 1.46 seconds, significant at the 0.02 level). Separating the trials by person, this holds true for two of the three comparisons. The differences are in the expected direction for both Bush and Dukakis, but are in the opposite direction in the case of Jesse Jackson. At least part of the explanation for this can be found in the last two rows of Table 5.4. These rows present the mean response latencies for trials in which the prime was the same word as the trait being evaluated. In the overall comparison, as well as in the comparison for two of the three candidates, the response time for the same-word trials is significantly larger than for the other trials. This difference is especially large for Jackson and contributes to the longer mean for real-word trials.4 This is exactly the opposite of 126 what was expected, and what is predicted by the associative memory model. The culprit here is faulty design of the presentation of the prim es.5 The problem is that the trials in which the prime and the trait are the same word are so visibly different from the other types of trials that it throws the subjects off their stride, so to speak. In cases where the prime is another word, the presence of the prime is detectable, though only a few subjects reported that they could consciously recognize the prime. In same-word trials, there was no obvious flash—the "X" marking the center of the screen was replaced by a trait word and that was it. Subjects had been told to expect distracting information, they were focused on the "X" and were, after a few trials, no doubt cueing on the flashed prime as an indication that the trait word was coming up. When this cue failed to appear, as was the case in same-word trials, they were slowed down rather than speeded up. Given this probable design flaw, the same-word trials have been excluded from all subsequent analyses.

Testing differences in response times It is possible to do more than simply describe differences between the means across groups of trait judgments. The next section develops and tests a model to explain the response latencies of each judgment. Of course, the central interest here is not to predict or explain how long it takes someone to make judgments about the personal qualities of a political candidate; it is to determine if priming with a trait that is hypothesized to be related to the trait 127 being evaluated has an effect on the time it takes to make such an evaluation. The presence of significant priming effects for words from the same network or sub-network constitutes evidence in support of the hypothesized trait structure. Rather than looking for priming effects over and above a baseline established using non­ words, response latencies for trials in which the primes came from outside of the network should be used to establish significance differences. The technique for doing this is to incorporate a pair of dichotomous independent variables in a regression equation, along with other regressors thought to be causally related to response latency. The first dichotomy ("Network") is coded to 1 if the trial used a prime that was in the same primary network as the trait being evaluated in the trial. It is coded 0 for all other trials. The second dummy variable ("Sub-network") is coded to 1 if the trial was for a prime from the same sub-network as the trait. Only three types of trials are included: trials using traits from outside of the primary network as primes, same primary network primes, and same sub-network primes. The test that is set up with this situation can best be seen from the following equation:

(Eq. 1) RT = a + bj (Network) + b 2(Sub-network) + other variables.

If the prime is neither from the same primary network nor the same sub-network, the intercept is simply a, as seen in Equation 2. (Eq. 2) Intercept = a + bj(0) + b2(0) = a. 128 Equation 3 shows that the intercept takes on the value of a + bj when the prime is from the same primary network, but not the same sub-network. (Eq. 3) Intercept = a + bj(l) + b2(0) = a + bj.

And finally, if the prime is from the same sub-network, and therefore from the same primary network, the intercept is computed as

(Eq. 4) Intercept = a + b j(l) + b 2(l) = a + bj + b 2- Thus, the assumption is that the form of the relationship is the same for all three types of trials. For example, if one of the other variables included in the equation is a measure of the interest of the subject in politics, the effect of interest on reaction time should not be expected to be different in trials where the subject has been primed with a same sub-network trait versus traits from outside of the primary network. The only difference is in the level of the dependent variable, all else equal.

The hypothesis tests. There are four possible outcomes to the analyses. First it is possible that there is no significant difference in accessibility between judgments primed with out-of-network traits and those primed with within-network traits or within-sub-network traits. Secondly, it may be that the two factor model is the most accurate picture of trait organization. In this case, one would rind significant effects for same-primary-network priming but not for any additional change due to same-sub-network priming. A third possibility is that the four factor model is the more accurate. Under 129 this scenario, there would not be significant changes due to priming with traits from within the primary network but there would be effects for same-sub-network priming. Finally, it is possible that there are two stages of effects, a change due to associations within a primary network and within the sub-network. Either of the last two findings would be consistent with Kinder’s more complex model (1986), since it is impossible to predict from the evaluative data and from the theory if the accessibility will spread in two stages.6 Several tests should be performed on the coefficients for the dummy variables to determine which of these four models prevails. The simplest test is to see if bi differs significantly from zero. If it does, then the absolute level of accessibility of a trait judgment is affected by the fact that the prime comes from the same primary network (including the same sub-network), confirming the validity of the current picture of trait organization, or at least the two factor version of it. The next test is to see if being from the same sub-network causes the absolute level of accessibility to be changed from that found in trials using primes from outside of the network. Since the intercept for same sub-network trials is actually computed using both bi and b2, the test here is to determine if the sum of bi and b2 is significantly different from zero. In this case, the signs of bi and b 2 should be carefully considered. The third test that must be conducted is to determine if there is an additional effect for primes from the same sub-network over and above the effect from simply 130 being in the same primary network. The test involves checking to see if b 2 itself is significantly different from zero.7 The pattern of expectations for these hypothesis tests under the four structural models is shown in Table 5.5. If, in general, the null hypothesis involving the effect for being in the same primary network (bi) can be safely rejected but the additional effect (b2) for being in the same sub-network cannot, the evidence points to a simple two-factor structure based on capability and sociability. Also in this case, the test for the overall effect for primes in both primary network and the sub-network (bi + b2) should be rejected, although this is not a necessary condition.8 The case is stronger for a four factor model based on competence, leadership, integrity and empathy if the sub-network effect (bi + b 2) is found to differ significantly from zero and the primary network effect (bi) does not. In this case, it is immaterial if b2 itself differs significantly firom zero if bi and b2 have the same sign.9 If activation is found to increase due to both same primary network and same sub-network connections, a two-stage structure is the best model of trait judgment organization. If none of the null hypotheses can be rejected, there can be no positive conclusions regarding the trait structure as measured by changes in accessibility. There may still be network structure to the traits but the hypothesized structures are absent. Another possibility to be considered in such a case is that the priming effects are so small that the method itself is ill-suited to detecting the structure among the traits. 131

Model specification. The fully specified model, of which the dichotomous variables for the priming effects are only a part, includes a number of variables that can be expected to have a bearing on response time. These variables must be included for the sake of producing accurate estimates for the dichotomous variables of interest, as well as accurate standard errors for those estimates. The intent here is not to predict or explain reaction times, but to assess the effect of certain classes of primes on those reaction times. One of the most important of these other variables is a measure of individual differences in dexterity. The measure used was the average response time to the six word-recognition trials that each subject completed at the end of the computer part of the study. This is the same measure used in the study reported in the previous chapter to standardize the response times. Standardization was not necessary in this case because including the term in the regression equation controls for these individual differences in response time. Also included to help account for variation across individuals were several measures of political interest and exposure to the media. Subjects had been asked how interested they were in following political campaigns, how frequently they watched both local and national television news broadcasts, and whether or not they were registered to vote. A check on gender differences indicated that female subjects tended to report higher interest in politics and have a lower average on the measure of dexterity, so gender was also included in the equation. 132 In the overall model, there were also controls for the person and the trait being evaluated. Two dummy variables for the three leaders, as well as an indicator of how well the subject liked the leader being evaluated in a particular trial, were included to account for across-candidate variation. A set of fifteen dummy variables was included for the sixteen traits. It was also expected that there would be differences based on whether or not a subject’s response to the query was affirmative or negative, so this was included in the model as well. Finally, to account for learning over the course of the trials, an indicator of where the trial fell in the sequence was included in the model. The estimates for these variables across the various analyses to follow will not be presented, since they have no direct bearing on the topic at hand. In general, the estimates were usually significant at better than the .05 level and were almost always in the expected direction. An example of the estimates for one leader is presented in Appendix 2.

Results. In principle, the detection of priming effects seems straight-forward; in fact, it is not. There are same-word primes, non-word primes, out-of network primes, and several permutations on the sub-network relationship. The prime may have the same or the opposite evaluative connotation as the trait being evaluated. The trait itself may be positive or negative. All of this makes the analysis to follow more complicated than it might otherwise be. 133 The model was estimated using ordinary least squares regression. Estimates were obtained separately for evaluatively congruent and incongruent trials. The estimates and the associated p values for all leaders as well as for each of the three leaders separately are presented in Table 5.6. The results reveal a confirmation as well as an elaboration of the pattern observed in the means. For the estimates using all trials regardless of candidate, the effect associated with being in the same primary network is significant for both congruent and incongruent trials. The difference is, of course, that for trials in which the prime and trait were evaluatively incongruent, there was an inhibiting effect (indicated by the positive sign) and for the congruent trials there was an activating effect (a negative sign for the estimate). This is repeated for both Bush and Jackson, though one or the other fails to make it to an adequate level of significance for each of these two leaders. In the case of Dukakis, neither the inhibiting nor the activating effect is present at a significant level. The second hypothesis, that there will be significant effects due to being primed with a trait from the same sub-network, finds mixed support. Recall that the direction and size of this effect is obtained by adding the estimates for same primary network effects and same sub-network effects, since the comparison here is not between primary network trials and sub-network trials but between sub­ network and out-of-network trials. In the overall model, and for two out of three of the leaders, the estimate obtained is in the expected 134 direction—the total effect indicates inhibition in the incongruent trials and activation in the congruent trials. A closer look at the significance of these estimates and where the differences lie is very revealing. For the congruent trials, the activation effects are only significant for the overall model and for Jackson. In the case of Bush and Dukakis, where the signs of bi and b 2 do not agree, the effect is not significant. For the evaluatively incongruent trials, the picture is even less straightforward. In three of the four cases, Jackson being the exception, the signs of the estimates indicate an inhibiting effect due to the primary network which is counteracted by the sub-network effects. For the overall equation and the Bush equation, this produces an intercept for the sub-network trials that is not significantly different from that of out- of-network trials. For Dukakis, the inhibiting effect is very weak and the sub-network activating effect is quite strong, producing significantly quicker response times. In the case of Jackson, both of the separate effects are inhibiting, neither is significant in and of itself, but together they produce a significantly slower response time. The last effect to be considered involves comparing the effect for sub-network connections over and above primary network effects. In six out of the eight cases in Table S.6, the sign for this effect is negative, indicating that primes from the same sub-network reduce response time. The effect is significant or marginally significant for all three candidates. For Bush and Dukakis, the significant effects occur for the evaluatively incongruent trials. In the case of Jackson, the effect is marginally significant for the 135 evaluatively congruent trials. This should be interpreted as evidence that traits from the same sub-network act to increase the accessibility of each other.

Discussion. What does all of this mean in terms of the structure of trait organization in long-term memory? Looking first at the evaluative congruent trials, where positive traits were used as primes for positive traits or negative traits were used for negative traits, the pattern of estimates points to a two-factor or two network organization of traits. There are significant priming effects for the same primary network trials in the full set of trials and for Jackson. None of the estimates are significant for Bush or Dukakis. For Bush, the primary network effect does have the expected sign. Turning to the evaluatively incongruent trials, where a positive trait was used as a prime for negative traits and negative for positive, the picture is in some ways clearer and in some more clouded. For the complete set of trials, as well as for Bush and Dukakis, it appears as though being primed with a related trait both inhibits and activates judgments. If the trait is from the same primary network, inhibition occurs and it actually takes longer to make the judgment. If the trait used as the prime is also from the same sub-network, this inhibition is in large part mitigated by an activation effect. In the case of Dukakis, the initial inhibiting effect is not present. In the case of Jackson, the inhibiting effect does not occur (at least not to a significant degree) at the primary network level but is present at the level of the sub-network. 136 The general picture that emerges is one that fits quite well with the current literature on trait organization. There seems to be a definite role played by both the smaller, more constrained sub­ networks of competence leadership, integrity and empathy and by the broader networks organized around capability and sociability. When the prime is of a different evaluative connotation, the similarity of the prime and the trait being evaluated causes problems in making the final evaluation, producing longer judgment times than if the prime was an unrelated trait word. If the prime also happens to be from the same sub-network, the increased closeness of the relationship overcomes somewhat this inhibiting effect. When the prime and the trait have the same evaluative connotation, the judgment is facilitated if the prime is from the same primary network. This facilitation is not usually further increased if the trait is from the same sub-network. There are obviously differences across the leaders included in the study. Unfortunately, it is difficult to assess at this point just why those differences occur. The analysis so far has focused on the relationship between the prime and the trait used as the evaluative standard, and has given special consideration to one aspect of the positivity or negativity of the evaluative connotation of these words, namely, their congruence. Important differences were found between congruent and incongruent priming effects. It is also possible that there are differences simply because the trait being evaluated is a desirable trait or an undesirable one. In the previous chapter, and in the work of Kinder (1986), there is certainly an indication that the trait 137 structure is affected by the positivity or negativity of the traits. Strictly from a methodological point of view, perhaps undesirable traits are unprintable. People may just have a hard time deciding if a person possesses a negative trait no matter how they are primed, thus muddying the results in the previous section. The following section reexamines the priming effects due to shared primary networks and sub-networks, breaking apart the congruent and incongruent trials based on the evaluative connotation of the trait about which the judgment is being made.

Negativity, positivity and trait structure Again, the model was estimated using ordinary least squares regression, this time for subsets of the trials based on both the evaluative connotation of the trait being evaluated as well as the match between the evaluative connotation of the prime and the trait being evaluated. The results for the three estimates of interest are presented in Table 5.7. The model has been estimated for all three political figures overall as well as for each separately.

Results. Looking first at the evaluatively incongruent trials, it becomes evident that there are differences in priming effects between positive and negative traits. Glancing down the estimates for the negative traits, the first thing that one notices is that there are very few significant effects. The only estimates which even approach statistical significance are for the same sub-network trials when compared to out-of-network trials for Bush and Dukakis. For 138 these two political figures, there is a small priming effect at the primary network level and an additional degree of activation at the sub-network level that together produces marginally significant faster responses for the same sub-network trials for negative traits. So, some support for the four factor model. The picture presented in the estimates for the positive traits primed with negative traits confirms the findings in the previous section. Judgments about positive traits, when preceded by the presentation of a negative trait, are inhibited. This is true for the overall model as well as for the three leaders separately, though the effect is stronger for Bush than the other two figures included in the study. Additionally, this effect is mitigated at the level of the sub­ network—the negativity of the prime is overcome by the close ties between the prime and the trait being evaluated. This mitigating effect is not present for Jackson. Yet, the general picture here is of a two stage structure, as it was without the breakdown into positive and negative traits. This is a structure in which both the primary network and the sub-network are important. The results for the evaluatively congruent trials also differ for positive and negative traits. When a negative trait is used as a prime for another negative trait, there is a significant degree of activation that takes place at the level of the primary network, as seen in the estimates for the entire set of trials as well as those for Bush and Jackson. There is no evidence of additional activation at the sub­ network level. Note, however, that the activation effect takes place 139 at the sub-network level for Dukakis. Thus, there is more mixed evidence for a two stage structure. The estimates for the subset of trials in which positive traits were used as primes for other positive traits are quite odd and, in at least one case, somewhat inexplicable. One would expect that this would be the subset of trial where an activation effect would be most pronounced. Yet, for the complete set of trials, none of the estimates are significant. This result is repeated for the Michael Dukakis judgments. There is a large priming effect at the sub-network level for Jackson but for Bush there is an inhibiting effect at the primary network level. If the results observed in the complete set of congruently primed judgments on positive traits were repeated for all three of the leaders, there would be a straightforward explanation: these words are already so activated that the manipulation strategy used here can’t make a difference. Priming positive traits with other positive traits just cannot tell us much about structure in this case. Priming positive judgments with a negative word from the same primary network can inhibit the judgments, as shown above. Furthermore, this slowdown can be partially overcome if the word is very closely related (i.e., lies in the same sub-network). Thus, the positive words are connected to the network structure but are simply so accessible that the manipulation to increase accessibility has no effect. However, this explanation only receives support in the subset of trials in which Dukakis was the target of the evaluation. The Bush 140 trials show very definite evidence of inhibition of access and the Jackson trials of facilitation of access. The Bush estimates are quite inexplicable; the Jackson results do not need much explanation by themselves but are odd since they are not part of any pattern for this type of trial. So, while there are primary network and sub­ network effects present, it is difficult to accept them as confirming or disconfirming evidence of the hypothesized two or four factor structure.

Conclusions Fortunately, the evidence from the other trials is clear and interpretable. Network effects are to be found at both the primary network and sub-network level. The primary network effects tend to be found in the trials for Bush and Jackson (incongruently primed positive traits and congruently primed negative traits) while the sub-network effects are more frequently found for the Dukakis trials. Judgments on negative traits more readily show evidence of the priming effects, though these are not statistically significant when the prime used is a positive word. Judgments on positive traits are inhibited by priming with negative words from the same primary network but if the prime is highly related (i.e., from the same sub-network) this is partially overcome. Overall, the results support a picture of trait organization much like that found in the current literature. Traits do appear to be organized around the overarching personality dimensions of capability and sociability and the sub-dimensions of competence and 141 leadership on one hand and integrity and empathy on the other. When a person thinks about some aspect of competence, the likelihood that other aspects of competence will be thought about at the same time is affected. How it is affected depends on the evaluative connotations of the traits as well as the evaluative congruence of the two specific traits.

Other possibilities. Obviously, there are many other questions about trait accessibility that this research does not address. For example, it is possible that certain categories of traits are more closely tied to one leader or candidate than another, either in terms of the overarching trait dimensions or in terms of the positivity and negativity of the traits. Perhaps just thinking about George Bush is enough to activate the positive trait nodes in the associative memory networks of most people, or the positive nodes for people who like him and the negative nodes for those who do not. This research was not designed to find answers to questions such as these, important as they are. The question addressed here is the structure among the traits themselves, not connections between traits and leaders. Inter-leader differences in accessibility may, in one particular situation, produce incorrect inferences from the results. If the ties between some category of traits and a particular leader are so strong that priming effects cannot occur at significant levels with the current priming strategy, then it is possible that the null hypothesis might be incorrectly accepted. This seems not to have occurred—the one set of questionable results (for congruently 142 primed positive trait trials) showing unexplainable but significant effects. The results presented do answer the question the research was designed to answer. And, as most research does, it also raises new questions. Exactly how does the difference in priming across positive and negative traits, congruent and incongruent pairs, affect evaluation itself? Obviously, there are ties between the positive and negative traits that produce changes in accessibility. Do these changes produce actual changes in how well a candidate is liked? This seems very likely, especially in situations where a citizen is attempting to produce inferences about a relatively new figure on the political scene. Typically, the public must rely on mediated information about candidate behavior, 30 second sound bites, and other forms of incomplete or ambiguous information. The way in which that information is Erst interpreted and categorized is going to have an effect on how subsequent information is categorized. Consider the case of Michael Dukakis in the 1988 election. The Bush organization had taken great pains to label him as cold, lacking in empathy, a technocrat. This may have had the effect of constantly priming the empathy sub-network for judgments about Dukakis. Now, what happens when ambiguous information or behavior is observed? The perfect example of what occurs is to be found in Dukakis’ answer to the opening question in the second presidential debate. When asked what his reaction to the rape and murder of his wife would be, Dukakis answered in a rather cold, rational way, discussing 143 his opposition to the death penalty even in those circumstances. Now, this could have been interpreted as evidence of his capability to hold high office, remain calm under fire, to remain unaffected by personal tragedy and lead the nation. Instead, it was more evidence of his lack of empathy, of human feelings. The Republican spin- doctor’s job was no doubt made easier since this network of traits was already tied to the candidate. Less dramatic scenarios are no doubt played out all the time during election campaigns and throughout the course of any administration, for any public figure. Once the connections among the various traits and between the traits and prominent political figures are understood, the detailed work of exploring the effect of the networks on evaluation and information processing can take place. The current research has done much to map out the first of these, the connections between the traits. The other questions await future efforts. 144 Notes

After the subjects had completed all portions of the study, they were informed of the true purpose of the study. None of the subjects indicated that they had suspected the cover story. Many of the subjects believed they had been asked about person-trait pairs more than once, but many were not certain of this. A very few indicated that they believed this repetition was a “check” on the stability of their answers. Such speculation may be expected to contribute to a general lengthening of reaction times in later trials, as the subjects try to remember if they have seen this pair before and if so, what their response was the first time. This may be true even though later trials were found to have quicker response latencies. In any case, such problems are not correlated with the independent variables of interest since the ordering of the primes was randomized.

For this and all subsequent analyses, several data cleaning procedures were used to eliminate outlying or odd cases. Ten people with over four trials that took eight seconds or longer were completely eliminated from the analysis, leaving 154 subjects. Other trials taking longer than eight seconds were set to missing; this resulted in a loss of 77 trials out of over 22,000. For a detailed and practical discussion of these and similar procedures, see Fazio (1990).

It is important to remember that these differences in average speed of judgment do not make a difference to the question at hand and, neither do many other sorts of cross-candidate differences. For example, it may be the case that hardworking is particularly tied to the personality image of George Bush across most people. Thus, the baseline (or unprimed) activation level for Hardworking is higher than it is for other traits. The question is, which primes cause this level of activation to increase to even higher levels? The only time this higher baseline could become problematic is when the trait becomes so chronically accessible that it is automatically activated whenever one even thinks about George Bush. This would seem to be a very unlikely state of affairs for most subjects. 145

4. Even with these trials removed, the means for the real-word trials is 1.49 seconds, still substantially larger than for non­ word trials. As is discussed below, the final explanation for this is to be found in the inhibiting effect of primes that have the opposite evaluative direction of the trait in some situations.

5. This methodological explanation is the likely one for reasons other than its failure to provide support for the associative memory model as the researcher would like. The evidence does support the model in the other cases (i.e., other types of priming trials) as shown below.

6. Research on other forms of two-stage priming reports failure of priming to spread over two stages or levels of memory structure (de Groot 1983).

7. More formally, the null hypothesis for this test is bj + 2 b = bj. Since bj is on both sides of the equation, it can be subtracted from both sides, leaving the standard null hypothesis b2 = 0.

8. There is a consistent problem with this test across each each of the models when the signs of bj and b 2 differ. Imagine the situation in which bj is significant and has a negative sign, indicating a priming effect for same primary network traits. If b 2 has a positive sign, indicating an inhibiting effect for traits within the same sub-network, the sum of bj and b 2 may not differ by a significant amount from zero. This may be the case whether or not b 2 itself differs significantly from zero. Similar situations may occur when examining the other models. Throughout the analysis, careful attention will be given to possibilities such as this.

9. A slightly altered set of circumstances holds if the estimates have opposite signs. In this case, the failure to reject the null hypothesis that the sum of the two estimates differs significantly from zero may be misleading. For example, if bj has a positive sign indicating an inhibiting effect when the prime is from the same primary network as the trait being evaluated, and b 2 has a negative sign indicating an activation 146 effect, the overall sum of the two may not differ significantly from zero. This is an indication that the group of trials in which the prime and the trait came from the same sub­ network were not responded to any faster than the trials using primes from outside of the network. In this case, this does not necessarily mean that there has been no activation effect since the activation effect may have been counteracted by an inhibiting effect at the level of the primary network. This may occur even though bj itself does not differ significantly from zero. Figure 5.1

Trait Words Used in Study 2

Hardworking! Intelligent! Inexperienced! Unqualified A Leader! Inspiring Aimless Hesitant! Decent Moral! Dishonest! Corrupt! Friendly! Understanding Prejudiced Unkind!

Indicates that the trait was used as a prime, in addition to being used as a characteristic to be evaluated. Figure 5.2

Artificial Words Used as Primes in Study 2

Obreken

Densis

Opargen

Terecht Table 5.1

Categorizing the Relationship between Various Primes and the Trait Hardworking

Example Trait Evaluatively Same Same Same Primes Word Congruent Network Sub- Word Network 1 Obreken No No No No No 2 Unkind Yes No No No No 3 Honest Yes Yes No No No 4 Hesitant Yes No Yes No No 5 A Leader Yes Yes Yes No No 6 Inexperienced Yes No Yes Yes No 7 Intelligent Yes Yes Yes Yes No 8 Hardworking Yes Yes Yes Yes Yes Table 5.2 Means for Reaction Time Measures Comparing Artificial Word, Same Word, and Other Trials

All Candidates Bi Dukakis Jackson Mean N Mean N Mean N Mean N

All Trials 1.47 22176 1.38 7392 1.54 7392 1.48 7392

Artificial Words 1.48 7392 1.41 2464 1.65 2464 1.40 2464 vs. Real Word Trials 1.46 14784 1.37 4928 1.49 4928 1.51 4928

Significance of the Difference in Means 0.02 0. 0.001 0.001

Same Word 1.60 960 1.35 347 1.62 308 1.87 305 vs. All Others 1.46 21216 1.38 7045 1.54 7084 1.46 7087

Significance of the Difference in Means 0.001 0. 0.13 0.001 Table 5.3

Expectations for Hypothesis Tests Under the Different Models of Trait Structure

Model H0: bi = 0* H0: bi + b2 = Of Ho: b2 = 0

No structure; Fail to reject. Fail to reject. Fail to reject. Structure other than hypothesized

Two Factor Structure Reject. Reject. Fail to reject.

Four Factor Structure Fail to reject. Reject. Unknown.

Two-stage Structure Reject. Reject. Reject.

* Here, bi is the coefficient for the dichotomous variable SAME PRIMARY NETWORK and b2 is the coefficient for the dichotomous variable SAME SUB-NETWORK. Since all primes in the same sub-network are necessarily in the same primary network, the change in the intercept is computed by summing bi and b2.

f These expectations may not hold if the coefficients are of opposite signs. See text. TABLE 5.4 1 Differences in Reaction Times Due to Primes

Evaluatively Evaluatively Incongruent Congruent Estimate Estimate Prime and Trait are from the: (Prpb. > Q (Prj?b, >_1) ALL N=7218 N=6299 SAME PRIMARY NETWORK 0.045 -0.041 (b l) (0.04) (0.06) SAME SUB-NETWORK 0.018 -0.048 (bl + b2) (0.42) (0. 11) SAME SUB-NETWORK ONLY -0.027 -0.007 (b2) (0.30) (0.83)

BUSH N=2404 N=2076 SAME PRIMARY NETWORK 0.091 -0.042 (bl) (0.009) (0.24) SAME SUB NETWORK 0.018 0.018 (bl + b2) (0.60) (0.73) SAME SUB NETWORK ONLY -0.073 0.060 (b2) (0.07) (0.29)

DUKAKIS N=2408 N=2109 SAME PRIMARY NETWORK 0.007 0.009 (bl) (0.87) (0.79) SAME SUB-NETWORK -0.054 -0.006 (bl + b2) (0.18) (0.91) SAME SUB-NETWORK ONLY -0.061 -0.015 (b2) (0.19) (0.78)

JACKSON N=2406 N=2114 SAME PRIMARY NETWORK 0.048 -0.075 (bl) (0.22) (0.08) SAME SUB-NETWORK 0.077 -0.163 (bl + b2) (0.05) (0.007) SAME SUB NETWORK ONLY 0.029 -0.086 (b2) (0.53) (0.18) The excluded category of primes are those trait words outside of the primary network of the trait being evaluated. TABLE 5.5 1 Differences in Reaction Times Due to Primes Evaluatively Incongruent Trials

Negative Positive Traits Traits

Estimate Estimate Prime and Trait are from the: (Prob. > t) (Prob- > t) ALL N=3607 N=3611 SAME PRIMARY NETWORK -0.033 0.116 (bl) (0.31) (0.0001) SAME SUB-NETWORK -0.034 0.067 (bl + b2) (0.28) (0.03) SAME SUB-NETWORK ONLY - 0.001 -0.049 (b 2) (0.97) (0.16)

BUSH N=1202 N*1202 SAME PRIMARY NETWORK -0.026 0.194 (bl) (0.60) (0.0001) SAME SUB-NETWORK -0.063 0.99 (bl + b2) (0.20) (0.04) SAME SUB NETWORK ONLY -0.037 •0.095 (b2) (0.52) (0.09)

DUKAKIS N=1204 N=1204 SAME PRIMARY NETWORK -0.061 0.077 (bl) (0.31) (0.16) SAME SUB-NETWORK -0.084 -0.031 (bl + b2) (0.16) (0.57) SAME SUB NETWORK ONLY -0.023 -0.108 (b2) (0.74) (0.09)

JACKSON N=1201 Nsl205 SAME PRIMARY NETWORK 0.010 0.074 (bl) (0.86) (0.18) SAME SUB-NETWORK 0.036 0.126 (bl + b2) (0.51) (0.02) SAME SUB-NETWORK ONLY 0.026 0.052 (b2) (0.69) (0.43 TABLE 5.5 i (Continued) Differences in Reaction Times Due to Primes Evaluatively Congruent Trials

Negative Positive Traits Traits Estimate Estimate Prime and Trait are from the: (Proh. > f> (Prob. > t)

ALL N=3173 N=3126 SAME PRIMARY NETWORK -0.078 -0.008 (bl) (0.02) (0.79) SAME SUB-NETWORK -0.101 0.001 (bl + b2) (0.02) (0.98) SAME SUB-NETWORK ONLY -0.023 0.009 (b2) (0.63) (0.85)

BUSH N=1056 N=1020 SAME PRIMARY NETWORK -0.111 0.083 (bl) (0.05) (0.08) SAME SUB-NETWORK -0.088 0.159 (bl + b2) (0.26) (0.02) SAME SUB-NETWORK ONLY 0.023 0.076 (b 2) (0.79) (0.31)

DUKAKIS N=1055 N=1054 SAME PRIMARY NETWORK 0.003 0.030 (bl) (0.96) (0.55) SAME SUB-NETWORK -0.135 0.086 (bl + b2) (0.06) (0.22) SAME SUB-NETWORK ONLY -0.138 0.056 (b2) (0.08) (0.23)

JACKSON N=1062 N=1052 SAME PRIMARY NETWORK -0.108 -0.047 (bl) (0.07) (0.44) SAME SUB-NETWORK -0.92 -0.276 (bl + b2) (0.26) (0.002) SAME SUB-NETWORK ONLY (0.016 -0.229 (b2) (0.86) (0.02) VI. Political Involvement and Trait Structure

Abstract. Associative memory structure, built up from re­ peated juxtaposition of traits in working memory, is expected to dif­ fer based on the experiences of citizens. Taking exposure to television news and self-declared interest in politics as indicators of differing political experience, this chapter examines the fit of the sub-network model within groups of high and low involvement subjects. Evidence from the priming study (Study 2) indicates that the two groups do not differ with respect to Bush or Jackson and differ a great deal in the organization of trait information concerning Dukakis. These findings are compared to findings on the evaluative structure using data from the same groups of subjects. The two types of structure are found to be similar for Dukakis and Bush, but different for Jackson. It is hypothesized that these differences occur as a result of the nature of the processes that produce the two types of structure.

Introduction In the development and discussion of the associative memory model underlying this research, the connections between traits were

155 156 accounted for with the notion of repeated conjunction. As Anderson (1983) lays out this model of memory, the connections between nodes in long-term memory are built up over time as the nodes, in this case traits, are repeatedly brought into working memory at the same time. Eventually, the connection between the traits is stored as a memory trace. It is over these traces that activational energy spreads, causing changes in the accessibility of related nodes in a network. What, though, produces the repeated conjunction of traits? It may be any number of things—something as simple as seeing the words "decent" and "moral" together in this sentence would require that they both be brought into working memory. Or, the situation may be more complex. For example, an observer in Iowa or New Hampshire who sees a presidential candidate attending church one Sunday morning might be motivated to try to understand the rea­ sons for this behavior. The observer might wonder if this indicates that the candidate and her husband are decent, God-fearing folk. On the other hand, the observer's mind might take a cynical turn and guess that the candidate is doing this to gamer political support, leading to the conclusion that the candidate is actually just another power-hungry pol. In any case, the mere act of observing and cate­ gorizing the stimuli in one's environment produces conjunctions of many things in working memory and, over time, this will leave traces in long-term memory.1 157 The process is not, as these examples might lead one to believe, completely idiosyncratic. If it were, there would be little reason to look for common patterns in memory networks of the general populace . One important source of similarities in trait organization is undoubtedly the media. We live in the age of mediated politics, with the vast proportion of political information obtained by citizens coming from radio, television, newspapers, and magazines. Televi­ sion has a particularly prominent place in this scheme. With about two-thirds o f respondents in National Election Study surveys indicating that television is their most important source of political news, television certainly provides the opportunity for mass exposure to the same sets of trait conjunctions. Mass media coverage of presidential campaigns contains much direct information on candidate traits. Joslyn finds that, as part of the dramatic nature o f news coverage, candidates are given thematic roles in a campaign (1984). These roles often are characterized in terms of the personality of the candidate: "humorless, somber, and dull...a bumbler and fumbler...eccentric...foolish and guileless... vindictive...manipulative (13). Analysis of the content of campaign coverage indicates that much of the information presented is directed to the personal qualities of the candidates (Patterson and McClure 1976; Patterson 1980; Joslyn 1984; Graber 1980). 158 Television news and accessibility And, as Iyengar and Kinder have shown, the news, at least television news, certainly 'matters' (Iyengar and Kinder 1987; see also Iyengar and Kinder 1986; Iyengar, Kinder, Peters, and Krosnick 1984). In a series of studies that considers priming and accessibility and their effects on evaluation, the authors found that television news coverage not only plays an agenda setting role, but also has an effect on which standards are used for evaluation. This effect, which they call the priming effect, works through the mechanism of differential accessibility. Issues that are covered more frequently on television news become more accessible in the minds of viewers and thus come more readily to use as standards of evaluation. It is important to note that in Iyengar and Kinder's work, the terms priming and accessibility are used in a very different way than they are being used in this research. In the research being pre­ sented here, priming is a tool used to introduce different experimen­ tal conditions. Changes in accessibility, as measured using reaction time, are indicators of connections between traits. In Iyengar and Kinder's work, priming is their name for a process that occurs in the real world, and that has important effects on evaluation. Accessibility is the mechanism by which exposure to television cov­ erage is translated into priming effects; it is not measured but rather assumed to have occurred. Citizens, motivated by the desire to un­ derstand their world while conserving limited cognitive resources, 159 use accessibility as a heuristic (Iyengar and Kinder 1987; Fischoff, Slovic, and Lichtenstein 1980). Yet, the parallels between my work and theirs are more than in terminology. Although not specifically testing for priming effects & la Iyengar and Kinder, the studies I present here demonstrate a differ­ ent type of priming effect. Mere exposure to a trait word exemplify­ ing leadership makes other leadership traits more accessible and, according to Iyengar and Kinder's findings, accessibility has an effect on political evaluation. Accessibility is the key in both sets of studies.

Exposure, involvement and priming effects. One of the impor­ tant intervening variables in the priming and agenda-setting effects examined by Iyengar and Kinder was political involvement. They found the relationship between involvement, agenda-setting, and priming to be very complex. In terms of agenda-setting, the lesser involved may be less likely to be exposed to the coverage but are more likely to be swayed by it. Those with a higher level of political involvement are more likely to see the news but less likely to accept its version of the world without question (p. 60-61). The relationship is complicated even further in the area of priming effects. The authors find no difference between the low involvement subjects and the high involvement ones and develop an explanation of this based on a two-step process of priming effects. The first step involves acceptance of the coverage, being "captured 160 by the focus of the news (p. 95)." Here, as with agenda setting, the highly involved are less susceptible. The second step involves recognizing the fact that the problem or issue being covered does or should have an effect on political evaluation. Because they have more resources and are more likely to think about politics, the highly involved are more likely to make this second step. So, while the more involved are not as likely to have the importance they assign to issues affected by television news (agenda-setting), they are just as likely as the lesser involved to think about and use different standards on the basis of news coverage. There is an alternate explanation available for the failure to find differences in susceptibility to priming effects. While the politically involved do have the resources to avoid agenda-setting effects, they are able to bring these resources into play because agenda-setting refers to assigning importance, a more-or-less conscious task. Priming effects involve use of a standard, and the accessibility of the standard, which is perhaps less conscious. Repeated exposure and attention to television news will result in unconscious and probably unrecognized changes in the accessibility of various evaluative standards. It is the highly involved who are more likely to have high levels of exposure and attention and, since the process is an unconscious one, they should not be able to use their more sophisticated resources to form a bulwark against this effect, as they do with agenda-setting. 161 Involvement and trait structure Given the evidence of the relationship between political involvement and other aspects of priming and accessibility, a closer look at how involvement and media usage affect trait structure is in order. To paraphrase Iyengar and Kinder, it is possible that, by repeatedly presenting certain aspects of candidate personalities in conjunction with each other, "television news sets the terms by which political judgments are rendered and political choices made (1987, 4)."2

Predictions based on exposure. There are several predictions about the differences that might exist between highly involved and less involved groups of citizens. In the simplest scenario, television news acts as the repository of socially defined connections between the traits; higher involvement, through exposure to television news, leads to unconscious absorption of those connections. If these conditions hold, the generic model developed to account for trait structure across all candidates and all levels of involvement should fit better for the highly involved than for the lesser involved. If there is no difference between the two groups, then it is probably the case that television news coverage has no effect, merely acting to support already widely held trait connections. Without further testing, however, this can remain only a guess. Other possible explanations for failure to find differences include lack of real variation in involvement due to the nature of the sample. 162 Another possibility is that television news is so pervasive that there is really little variation in the population itself. If the fit is better for low involvement subjects, an alternate explanation will have to be sought. Perhaps television news coverage includes connections that are not characteristic of the Sub­ network model of trait organization. It would be reasonable to assume that high involvement exposes one to a very different set of trait conjunctions, through the use of other media and through engaging in more active political participation, producing more and stronger inter-network connections between traits in long-term memory. Without an empirical study of trait conjunction in television news coverage, it would be impossible to choose definitively among these explanations.

Predictions based on cognitive differences. What sort of differences might be expected between high-involvement and low- involvement subjects on the basis of cognitive differences? A significant amount of research has been done on novice/expert differences in the use and organization of another type of cognitive structure, political schema. While the novice/expert distinction is, at best, only moderately similar to the distinction of high and low political involvement, it does offer valuable comparisons. The political schemata of experts are more detailed, contain more hierarchical organization, and come more readily to use than do the schemata of novices (Lau and Sears 1986; Fiske and Taylor 1984; 163 Fiske, Kinder, and Larter 1983). Schematics are also more likely to make use of more of the detailed and inconsistent information they encounter (Fiske et al. 1983; Lau and Sears 1986). These findings may be used to formulate several hypotheses about involvement-related differences in the associational structure of traits. As discussed by Lau and Sears (1986), increases in expertise eventually produce such tight linkages between the elements of a schema that they become 'unitized1, and multiple elements are processed as 'chunks' or unitary wholes. In terms of the model of trait structure based on activational networks, this leads to the expectation that experts should have fewer dimensions in their structure, tending to process and classify candidates in terms of broader personality clusters. Ultimately, the very politically involved might process traits in terms of a single positive-negative dimension. Another possibility is that chronicity (Bargh, et al. 1986; Lau1990) plays a greater role than does involvement per se. In this model of individual differences, some people are more prone to use personality traits as cues in information processing, others to use party or issue cues. The politically involved are more likely to have a category that is chronically accessible, but only a portion of these will be trait chronics. If this is the case, there should be few differences between high and low involvement citizens as a whole, since the effect will be washed out among the highly involved. 164 A reconciliation. So, one the one hand, the politically involved have a greater degree of exposure to the mass media and its subtle priming effects, so might be expected to be better fit by the generic model of structure that fits the general population the best. On the other hand, the highly involved are more likely to unitize this information, producing a simpler organization than the rather complicated Sub-network Model found to fit so well in the previous two chapters. There should, then, be fewer indications of a need for four separate dimensions to account for the trait structure of the politically involved.

Measuring involvement Iyengar and Kinder, having at their disposal a rather rich data set designed to examine several aspects of involvement, preserved distinctions between interest, media exposure, informal communication, political activism, and political expertise. Unfortunately, the information available on subjects in the present set of studies is limited to the first two of these, as well as a measure of participation. Subjects were asked about their interest in politics, frequency o f viewing local and national television news, and if they were registered to vote. Their responses are shown in Table 6.1. These four measures were combined, using factor analysis, into an overall scale of political involvement.3 This strategy ensures that the measure used is not simply one of exposure to television news, since it incorporates both exposure and interest. The resulting scale, 165 with a mean of 0.0 and standard deviation of 0.84, was used to partition the subjects into high and low involvement groups. Of the 151 subjects with non-missing values for the scale, the 77 below the mean were classified as low involvement, and the 74 above the mean were put into the high involvement group.

Plan of analysis. The analysis will proceed by first showing the differences in priming effects between high and low involvement subjects across the three political figures asked about in Study 2. From this, predictions will be made about the fit of the models to the evaluative data from these same subjects. Again using confirmatory factor analysis, the models will be applied to these data and the fit compared to the predictions. This not only serves as a check on involvement differences, it also acts as another way to test the similarity of the activation structure and the evaluative structure.

Involvement and trait structure: accessibility Data for the subjects were divided into high and low involvement subsets, based on each subject's score on the political involvement scale, as discussed above. The regression model used in Chapter 4 was again used here.4 Preliminary to estimating the priming effects separately for the two groups, the regression was run with additional terms for involvement (dichotomized) and the interaction between involvement and the primary network and sub­ network effects. If the interaction effects are significant, there is 166 reason to run the analyses separately for the two groups. There were two significant interactions present for each of the political figures across the four types of trials, as shown in Table 6.2. This pattern warrants further analysis with the breakdown by involvement.

Results. The estimates for the different types of trials, based on the evaluative direction and congruence of the prime-trait pair, are presented in Table 6.3. The estimates obtained for the two groups are, for the most part, very similar to those found for the subjects as a whole. There are sharper differences among the three leaders, and several interesting differences between high and low involvement groups within leaders. In two cases, priming a negative trait with a positive word produced a significant change in accessibility. Both occurred for the low involvement subjects. As in the main analysis, the effects were due to activation within the same sub-network when the effect for the primary network was added in as well (i.e., bi + b2). There are a larger number of significant effects for trials in which a negative prime was used for a positive trait. Here again, there was evidence of both inhibition as well as activation of related traits. For Bush judgments, all three types of priming effects were significant for both levels of involvement. When the traits were more broadly related, coming from the same primary network, an inhibiting effect was present. This was mitigated, but not completely 167 overcome, if the trait was also from the same sub-network. This exactly replicates the pattern observed in the subjects as a whole. There are several differences between the high and low involvement groups for the Dukakis and Jackson trials. For Dukakis, the inhibiting effect observed in the aggregate analysis disappears when involvement is considered. The activation effect found for same sub-network only primes is now observed to be present only for the low involvement subjects. While the sign for this effect is negative for the high involvement group, it is far from significant. For Jackson, only the inhibiting effect is present, just as in the aggregate analysis. However, this occurs at the level of the primary network for the low involvement subjects and is associated with the sub-network for the high-involvement subjects. The effects for the set of trials in which both the prime and trait were negatively phrased are surprising for one of the three political figures. When making judgments of Bush and Jackson, there are accessibility effects at the level of the primary network, as in the analysis of the sample as a whole. For Dukakis, the situation is a bit different. Here, when not considering differences in involvement, a priming effect was observed at the sub-network level. This falls apart when the subjects are divided on the basis of political involvement. There are still accessibility effects—for the low involvement people they still occur at the sub-network level but, for the high involvement group, there is evidence of activation at the level of the primary network. Most odd is the previously 168 unsuspected presence of an inhibiting effect among the low involvement subjects at the primary network level. There is no obvious explanation for this effect. Finally, the equally inexplicable results of the evaluatively congruent positive trials are still present in the sub-group analysis. Though the effect tends to be significant only for the lesser involved subjects, there is an inhibiting effect to being primed with a positive trait and then being asked to judge a leader on another positive trait. This effect is quite significant for the Bush trials for the low involvement group; it approaches significance for Bush among the highly involved and for Dukakis among the less involved. For Jackson, the expected priming effects occur, at the level of the sub­ network, but only for the low involvement subjects.

Similarity of the two groups. In general, there is a greater degree of similarity between the two groups for Bush than either of the other leaders. Out of thirteen significant or near significant effects for the Bush trials, seven are for the low involvement group and six for the high involvement; the split is nearly equal in the number of effects that occur at the primary and sub-network levels. The sub-network model, in short, would seem to fit equally well for the high and low involvement subjects in thinking about George Bush. While there are fewer significant effects in total, the same balance more-or-less holds for Jesse Jackson. Significant effects were 169 present for low involvement subjects in three of the four types of trials, for high involvement subjects in two types. If there is some difference in the fit of the model between the two groups, it is small and it slightly favors the low involvement subjects. The largest differences in the effects for the two groups are for Michael Dukakis. For all four types of trials, there are significant or nearly significant effects at the subnetwork level for the low involvement subjects. For the evaluatively congruent negative trials, there are also effects at the primary network level, although these are somewhat counter to expectations. Among the estimates for the high involvement subjects, there are effects only for one type of trial, and these seem mostly due to primary network connections. If there are any differences in the fit of the model between the highly involved and the less involved, these differences are in the trait structure involving Michael Dukakis.

Involvement and trait structure: evaluation One of the important findings of the previous chapters was the similarity of the trait structure using accessibility measures and evaluative measures. The division of the subjects into high and low involvement groups offers a further chance to test this similarity, this time between priming results and five-point judgments for the same group of subjects, while extending our understanding of the relationship between involvement and structure. If the trait structure for the five-point evaluations is in accord with the results 170 just reported, there will be greater confidence in both the assertion that evaluative and accessibility structure are intertwined and in the findings concerning political involvement. Failure to confirm the predictions should cause us to question both.

Predictions o f Jit. There is little reason to believe that, in general, there will be vast differences in fit for the two groups of subjects. For both George Bush and Jesse Jackson, the results presented indicate that the Sub-network model fits equally well at both levels of involvement. Michael Dukakis, however, presents a different story indeed. The Sub-network model held up very well for the subjects with low levels of involvement but, for the more involved, failed to receive much support. Based on these predictions, in the next section of this chapter, confirmatory factor analysis should reveal differences in the fit of the four factor Sub-network model to evaluative data for Dukakis as well. Differences in fit for the two groups should be negligible for the other two political figures. The analysis of the evaluative data for the subjects in Study Two takes the form of confirmatory factor analysis. The models applied are in all respects identical to those outlined in Chapter 3. The data for the subjects were again divided into high and low involvement groups, and the models were fit to each separately. The resulting fit statistics are presented in Table 6.4. 5 171 Results. The overall fit of the model across the three candidates resembles that found in Study 1. Generally, if one considers an average fit for a leader, the fit of any particular model is best for Bush and worst for Jackson. By groups, however, the best fit is typically found for the low-involvement subjects in connection with judgments about Dukakis. A comparison of absolute fit across groups shows that there are very few differences for two of the three candidates. The x2/df ratio differs by only about one-tenth of a point for Bush, if at all. For the Jackson judgments, there are slightly larger differences, averaging about two-tenths of a point. In four out of five cases, these differences favor the low-involvement group. In the case of Michael Dukakis, all models fit much better for the low involvement group than for the high involvement subjects. In absolute terms, the difference averages close to six-tenths of a point on the x2/df ratio. While this finding is expected for the models with a larger number of factors, it is surprising to see the difference exists for the One Factor model as well. The important test, however, is the significance of the change in fit moving from one model to the next within the columns of Table 6.4. If the expectations about the similarity of the two types of trait structure are to be borne out, there should be few differences in these changes for either Bush or Jackson. In other words, if the Sub­ network Model fits better than the Primary for the highly involved, it should fit better for the group with a lower level of involvement as 172 well. For Dukakis judgments, some differences should be present. Based on the results of the analysis of accessibility structure, the degree of improvement in moving from the Primary Model to the Sub-network Model should be larger and more significant for the low-involvement subjects than for the highly-involved.

Comparing changes in InformationJit. on the changes in fit is presented in Table 6.5. Statistics are presented for the comparison of the Primary Model and the Sub-network Model (the two and four factor models), both with and without the negativity factor. As can be seen, the evidence here gives mixed support to the predictions. By far, the largest changes in the chi-square values, and the changes in the x2/df ratio, occur where expected — for the low-involvement subjects for Dukakis judgments. While the significance of the changes indicates the need for the four-factor model for both groups, the significance of these changes is much larger for those subjects who are less politically involved. This finding vindicates the assertion that the accessibility and the evaluative structure are intertwined, and supports the finding that there are differences between high and low involvement structure for Dukakis traits. The results for Bush are somewhat different than expected. Based upon the results from the priming study, no differences were expected between the two groups. Instead, small differences were found indicating a slightly larger improvement for the highly involved in going to the Sub-network Model. The differences in chi- 173 square are smaller than those observed for Dukakis, and the x2/df ratio is essentially unchanging across the models. The analysis of the Jackson judgments goes completely against expectations. The significance test of the nested models indicates that the Primary Network model fits just as well for the low- involvement subjects as does the Sub-Network Model. The change for the more-involved is, on the other hand, highly significant. Nothing in the previous results would have lead to this expectation. It indicates that, for Jackson at least, not only are there differences between the subjects on the basis of involvement, these differences are in the opposite direction as those found for Dukakis. Moreover, there is less similarity between the two types of structure for Jackson than for the other political leaders in the study.

D iscussion For two of the three political figures included in the study, there seem to be real differences in trait structure due to differing levels of political involvement. The bothersome thing about these differences are that they send contradictory messages about involvement-based differences for the two leaders. In the case of Dukakis, the Sub-network model seems more applicable for the less involved in both the accessibility structure and the evaluative structure. These findings support the hypothesis that those with higher levels of political involvement organize trait information in a more unitized fashion, along fewer dimensions. For 174 Jackson, there were few differences between the high-involvement and low-involvement subjects in terms of accessibility. For the evaluative judgments, however, the analysis showed that a larger number of dimensions was needed for the highly involved. How can this be explained? One possibility is that accessibility structure, dependent only upon repeated conjunction of constructs, may be less likely to vary across candidates and more susceptible to the organizational forces set in place by increased involvement and expertise. This does not deny the possibility of a causal link between activational networks and the dimensional structure observed in evaluations themselves; it merely loosens that link. Evaluative structure, while not consciously built, is more likely to be influenced by more-or-less conscious processes such as rationalization and explanation, discussions of why one likes or dislikes a candidate. The very process of measuring evaluative structure, asking subjects to rate candidates on batteries of traits, causes them to be conscious of how the traits fit together. A subject who has just rated the candidate on several traits may take those ratings into account when making subsequent ratings. This is likely to have an effect on the structure of their answers. The accessibility structure, on the other hand, is not measured in a way that is likely to produce this type of effect, even if the subjects were aware that the speed of their responses w as being measured. 175 There is another possible reason for the Dukakis results. Michael Dukakis was already ancient history to the twenty year-old subjects in this study, having lost his bid for the presidency a full two years before and fallen quickly from the limelight. How many of the subjects had direct personal memories of the 1988 presidential campaign? Probably not many—some had not even been eligible to vote. Of those who had been eligible, it had been their first opportunity and they, like many eighteen year-olds, had not bothered. Most likely, what set the highly involved apart was that they had voted in that election, or at least had paid a relatively larger amount of attention to it than their less involved peers. Why, then, does the Sub-network Model fit better for the less involved, better than for any other candidate? The answer is that the model is a generic model, one specifically designed to paper over the differences between candidates, and thus the differences that come about as a result of knowing more about a candidate. Seen in this light, it is only logical that, for a less well known political figure, among a group of less- knowing subjects, the generic model fits admirably well. They simply have less distinguishing information to muddle the clear baseline generic structure. The involved, on the other hand, know and remember some things that set Dukakis apart. The problem with this explanation is that it cannot account for the Jackson results. The four-factor model fits remarkably better for the politically involved, not at all better for the uninvolved. 176 Unfortunately, there are no indicators in this dataset to seek out further explanations. Perhaps the measure of political involvement is too truncated to capture differences for the better known political figures? Direct measures about levels of information on specific leaders are needed to test the involvement hypotheses further. Additional studies, employing similar methodologies using a broader range of subjects, would prove invaluable to our understanding of how the generic sub-network structure is differentiated for better known leaders. Other possibilities, such as exploring the relationship between development of structure and evaluation, perhaps using Iyengar and Kinder's notions of priming effects, should also be pursued. Data on the pattern of trait conjunctions on network newscasts and in the print media should also be obtained, so that the fit of the Sub-network model and the roots of the division of political personality judgments into competence, leadership, integrity and empathy may be better understood. 177 Notes

1. Of course, any traces already present in long-term memory will affect, but not necessarily determine, which categories the observer considers. Other factors, such as the recency of activation of the various trait nodes and the presence of chronically accessible trait categories in the mind of the observer, will also have an effect. Thus, the relationship between accessibility, activation, and structure is a non­ recursive one.

2. Much of the following discussion focuses on television news coverage. This is so for several reasons. First, television news is the dominant source of political news for most Americans, especially campaign news. Second, the measures I have available do not ask about other media usage. While much of what I say about the media as a source of activational networks may well applicable to other mass media, either print or broadcast, the measure of involvement used is based in large part on television news viewing habits.

3. Three cases with missing values on one or more of the variables were eliminated from the subsequent analysis. The loadings for the four measures on the factor were: National Television, 0.74; Local Television, 0.61, Registration Status, 0.12; Interest in Politics, 0.55.

4. The terms for interest, exposure to television news, and registration status were left in the regression model. Since the involvement scale is not used as a term in the model but simply to divide the subjects into two groups, this strategy allows the individual variables, which do have an effect on reaction time, to be included in the model.

5. The full set of results for each of the models is not presented here. Overall, the estimates for factor loadings, error variances, and correlations between factors were very similar to those presented in the appropriate sections in Chapter 4. 178

Table 6.1 Indicators of Political Involvement

Frequency Percent Local Television News Never 3 1.9 1-2 Days / Week 48 31.2 3-5 Days / Week 81 52.6 6-7 Days / Week 22 14.3 Total 154

National Television News Never 8 5.2 1-2 Days / Week 72 46.8 3-5 Days / Week 62 40.3 6-7 Days / Week 12 7.8 Total 154

Interest in Politics Most of the Time 40 26.1 Some of the Time 78 51.0 Now and Then 27 17.6 Hardly at All _8 5.2 Total 153

Registration Status Registered 110 72.4 Not Registered 41 27.0 Total 151 Table 6.2

Interaction of Priming Effects and Involvement

Evaluatively Inconguent Evaluatively Congruent

Interaction With Negative Traits Positive Traits Negative Traits Positive Traits

Bush Significance Significance Significance Significance (Prob. > t) (Prob. > t) (Prob. > t) (Prob. > t) Primary Network Effect n.s. n.s. n.s. .08 Sub-Network Effect n.s. n.s. .20 n.s.

Dukakis

Primary Network Effect n.s. n.s. .01 n.s. Sub-Network Effect n.s. .15 n.s. n.s. Jackson

Primary Network Effect n.s. n.s. n.s. n.s. Sub-Network Effect n.s. .08 n.s. .03 Table 6.3

Comparing Low and High Involvement

Evaluatively Incongruent Trials

Negative Traits Positive Traits

Low Involvement High Involvement Low Involvement High Involvement

B ush Estimate (Prob. > t) Estimate (Prob. > t) Estimate (Prob. > t) Estimate (Prob. > t)

Same Primary Network -0.029 (0.71) -0.024 (0.70) 0.215 (0.01) 0.211 (0.01) ( b l ) Same Sub-Network -0.109 (0.15) -0.044 (0.51) 0.090 (0.19) 0.107 (0.10) (bl + bl) Same Sub-Network Only -0.080 (0.37) -0.020 (0.80) -0.125 (0.11) -0.104 (0.18) (b2)

D ukakis

Same Primary Network -0.061 (0.48) -0.042 (0.60) 0.095 (0.25) 0.065 (0.34) ( b l) Same Sub-Network -0.172 (0.05) 0.003 (0.97) -0.083 (0.31) 0.032 (0.63) (bl + b2) Same Sub-Network Only - 0.111 (0.27) 0.045 (0.63) -0.178 (0.06) -0.033 (0.67) (b 2 )

Jac k so n

Same Primary Network -0.010 (0.90) 0.034 (0.63) 0.124 (0.09) 0.032 (0.69) (bl) Same Sub-Network 0.013 (0.86) 0.067 (0.35) 0.064 (0.38) 0.222 (0.01) (bl + b2) Same Sub-Network Only 0.023 (0.80) 0.033 (0.69) -0.060 (0.50) 0.190 (0.05) (b2) ISO Table 6.3 (Continued)

Comparing Low and High Involvement

Evaluatively Congruent Trials

Negative Traits Positive Traits

Low Involvement High Involvement Low Involvement High Involvement

Bush Estimate (Prob. > t) Estimate (Prob. > t) Estimate (Prob. > t) Estimate (Prob. >

Same Primary Network -0.138 (0.09) -0.115 (0.10) 0.209 (0.01) -0.063 (0.31) (bl) Same Sub-Network -0.046 (0.70) -0.149 (0.12) 0.220 (0.03) 0.080 (0.35) (bl + b2) Same Sub-Network Only 0.092 (0.48) -0.034 (0.74) 0.011 (0.92) 0.143 (0.14) (b2)

Dukakis

Same Primary Network 0.137 (0.07) -0.144 (0.06) -0.009 (0.90) 0.050 (0.45) (bl) Same Sub-Network -0.028 (0.77) -0.273 (0.01) 0.153 (0.12) 0.105 (0.23) (bl + b2) Same Sub-Network Only -0.165 (0.14) -0.130 (0.24) 0.162 (0.13) 0.055 (0.57) (b2)

Jackson

Same Primary Network -0.120 (0.18) -0.138 (0.07) -0.006 (0.95) -0.095 (0.22) (bl) Same Sub-Network -0.083 (0.49) -0.104 (0.33) -0.484 (0.01) 0.077 (0.49) (bl + b2) Same Sub-Network Only 0.037 (0.78) 0.034 (0.76) -0.478 (0.01) 0.018 (0.88) (b2) Table 6.4

Comparing Low and High Political Involvement

The Fit of the Models to the Evaluative Judgments

Bush Dukakis Jackson Model Low High Low High Low High One Factor X2 584.8 543.9 609.7 737.1 747.3 793.2 X2/df (252) 2.3 2.2 2.4 2.9 3.0 3.1

Primary X2 525.9 526.1 546.8 644.5 631.7 684.2 X2/df (251) 2.1 2.1 2.2 2.6 2.5 2.7

Primary + X2 493.5 497.2 455.6 589.6 559.4 634.3 X2/df (250) 2.0 2.0 1.8 2.4 2.2 2.5

Sub-Network X2 518.1 511.1 485.0 632.5 629.3 658.5 X2/df (247) 2.1 2.1 2.0 2.6 2.5 2.7

Sub-Network+ X2 482.0 473.2 367.2 571.3 554.0 601.1 X2/df (246) 2.0 1.9 1.5 2.3 2.3 2.4 Table 6.5

Comparing Changes in Fit

Bush Dukakis Jackson Comparison Low High Low High Low High

Primary Change in x2 7.8 15.0 61.8 12.0 2.4 25.7 vs. significance .10 .01 .001 .02 n.s. .001

Sub-network Change in x2/df 0.0 0.0 0.2 0.0 0.0 0.0

Primary + Change in x2 11.5 24.0 88.4 18.3 4.4 33.2 vs. significance .05 .001 .001 .01 n.s. .001 Sub-network+ Change in x2/df 0.0 0.1 0.3 0.1 +0.1 0.1

oo u> VII. Conclusion

In 1960, Angus Campbell and his co-authors produced the seminal empirical work on American voting behavior, The American Voter. In the concluding chapter of that volume, the authors turned for a moment from explaining the behavior and attitudes of individ­ uals to discussing the effects of individual voting decisions on the American polity itself. They recognized the short-sightedness of ad­ dressing only the causes of political behavior and disregarding the consequences of it. Before discussing the implications of my findings for our understanding of individual political attitudes, it is important to take a page from that book and consider the significance of traits and trait structure to our understanding of the American political system in general.

Traits and the political system "The popular election," wrote Campbell et al., "is a device of control (1960, 543)." Yet the type of control they found exercised was not, one is led to believe, the type commonly thought to be an integral part of representative democracy. Control over public policy was found to be limited. In part this was because the American 184 185 public was so ill-informed on policy questions and had little real feel for the political divisions between the parties on many policy issues. Control instead was found at the level of broad goals and means of achieving them (545).1

Traits as demands, traits as policy. Failing to find the bulk of the population in possession of a healthy democratic concern for and information on issues, Campbell et al. focused on image and affect, especially affective orientations to the political parties. They refer to "a comparison of the total image of one of the candidate-party alternatives with the image of the other (546)." These images, they go on to point out, are made up in large part of judgments about honesty, dependability, capability, or general positive reactions. The vagueness of policy demands, coupled with the prominence of image, produces 'policy' demands that have much more to do with successfully displaying the right type of feeling or personality in an administration than might otherwise be the case. Consider several of the cases discussed by Campbell et al. Their "classic example" of public opinion as a guide for governmental goals is federal responsibility for the economy. "What the public en­ dorsed was the New Deal's belief that the federal government should assume responsibility for the nation's economic welfare (545)." To apply the language of personality characteristics, what the electorate demanded was that the administration not only care for them but care about them. They demanded a policy of empathy. The result 186 was an ever increasing level of intervention in the economy and the growth of the social welfare system. Another case in point, also taken from the discussion in The American Voter, is that of the Eisenhower administration and the Korean War. "Their mandate to the president-elect was to produce a solution and, far from prescribing specific policies he should follow, they gave him virtually unlimited freedom of action (546)." This demand may be expressed in terms of capability, the will and skill to succeed at something, while the Democrats seemed to be falling into a stalemate. In a second war a little more than a decade later, in another Asian country, another Republican presidential candidate would also play upon the demands for capability. The Democratic president, Lyndon Johnson, would bow out of office, as much a prisoner of the conflict in Vietnam as any POW wasting away in a camp in the jungle. To bring things forward in history, consider the events in the third year of the Bush administration. Bush, still haunted by ques­ tions dating to the early days of his vice-presidency under Ronald Reagan, faced possible investigations by both the House and Senate into the Iran Hostage crisis and allegations that the Reagan-Bush campaign negotiated with the Iranians to postpone the timing of the release of the hostages for political gain. On the domestic front, his chief of staff, John Sununu was embroiled in an ethics controversy over the use of publicly funded transportation to private or political events, as well as the solicitation of transport from wealthy 187 Republican supporters. All of this reflected poorly on the integrity of the Bush White House. Perhaps more important was the question of the leadership qualities of Bush the person. During the election campaign, he had to battle against an image of the pampered, preppie wimp. His campaign handlers took great pains to see that he was coached to avoid ineffectual-sounding phrases in his extemporaneous remarks. By the summer of 1991, his speaking style was a running gag on Saturday Night Live. The administration was one in which a large emphasis had been placed on character, integrity, and traditional ethics. One must wonder, then, how the president and his advisers perceived the pub­ lic's reaction to the scandals and to continued reference to the lack of true leadership in the White House. Did these character issues, to mix the terms up a bit, have an effect on policy decisions? If Bush had not had to worry about the perceived absence of a domestic pol­ icy, if he had not been tarnished with the wimp label earlier on, would he have taken such a strong line with Manuel Noriega in Panama and Saddam Hussein in Iraq? This possibility alone is enough to warrant further research into trait perception and trait structure, and especially into the role of the media and political involvement in strengthening the connec­ tions between traits in associative memory. Regardless of the direc­ tion of causal flow, from the media to the mass public or vice-versa, it is important to understand how individual traits become connected into bundles, especially if these bundles then become demands 188 placed upon those in power. This is true even if those demands are only perceived by the politicians, and have little to do with the vote.

Ideology, trait structure, and democratic ideals. However, personality judgments do have an impact on the vote. Recent models of the vote have given them a prominent place in the vote decision (c.f., Rahn et al. 1990; Sniderman, Glaser and Griffin 1990), or used them as unique measures of overall candidate evaluation (Lodge, McGraw, and Stroh 1989). Typically, this takes the form of computing an overall measure of candidate affect based on the direction and strength of trait attributions, combining one or more of the trait dimensions. If trait evaluations are as important to the vote as are issue attitudes and partisanship, it behooves us to know how and where citizens obtain the information upon which they are based, by what processes they are formed, and, like issue attitudes, how they cluster together to form dimensions. If, as with the term ideology in the case of issue attitude structure, trait structure were given a concise title, perhaps research on it would boom as it did for issue attitude structure during the 1960's and 1970's. Perhaps not. Personality judgments have always been seen as too easy, not demanding enough to be used as a basis for sound democratic government. Unlike issue attitudes, it seems unlikely that a researcher would be tempted to set up a series of hurdles (cognizance, intensity of feeling, differentiation of parties) for candidate-based behavior as the authors of The American Voter did to gauge issue-based behavior. Ideology has, since before 189 Campbell and his colleagues penned that work, been seen as the "medium of political translation par excellence (202)." However, the more we have learned about the vote decision, the more we have realized the importance of candidate affect and evaluation, not only as products of issue judgments and partisan politics, but as determinants of them as well. One obvious extension of the trait structure research should be to investigate the connections between traits and the political parties. Just as earlier research found ties between the parties and social groups, and between the parties and policy positions, there are probably important ties between parties and the traits typically used to describe political leaders. The interconnections among all of the determinants of the vote indicate that, should we look, we would find many traits have greater associations with one party over the other. To pursue this line of inquiry moves beyond treating trait judgments as a determinant of the vote to treating them as an organizing force on a par with issue attitudes or partisanship.

Political communication, trait structure, and standards of evaluation. Finally, the topics of media influence, levels of information, priming, and trait structure come together to form an interesting and troublesome nexus of questions about political communication and decision making. Iyengar and Kinder, working with a different type of priming and accessibility than I have used here, have shown the "insidious" power of television to prime certain evaluative standards, making them more likely to come to mind and 190 thus to be used (1987 p. 4). Their demonstration involved issues and events, which are more complex standards to employ. What effect does priming, in the Iyengar and Kinder sense of the word, have on trait judgments and the use of different evaluative standards? Who is more likely to be affected? How does this affect the nature of the competitive struggle between the parties? This is especially pertinent if certain traits are already tied to parties and are more easily activated to begin with.

Traits and policies. So, the study of traits and trait structure is important not just for understanding how individual voters make up their minds about who they will vote for, it is also important to our understanding of the American polity as a whole. Traits may act as a demand upon policy-makers, influencing not just the goal choices (policy objectives) but the manner in which those goals are reached. The latter choice has serious implications for the policy's chance of success, effectiveness, and the presence of unforeseen consequences, not to mention its effects on political culture and the society as a whole.2 Traits, used as broad standards by voters evaluating issue groups, parties, and candidates, may also act to tie together the polit­ ical arena in a less demanding fashion than do issue stances. Trait structure, in this case, might serve in the same capacity as issue structure does for some people. Finally, still on the topic of traits as evaluative standards, we must consider the role of the mass media and its impact on traits. If 191 television primes certain issues to be used as standards, it is likely that it does so for traits as well. We must know which traits are tied together, how closely they are bundled* and what types of differences in those bundles are present in different subgroups of the electorate if we are to understand the impact o f the mass media on politics. To focus only on policy issues, whether as a source of policy demands, a general organizing force, or an avenue of media effects, will do ourselves and our theories an injustice.

Traits and the individual Trait evaluations and trait structure are also an integral component of our understanding of voting behavior at the level of the individual voter. It is important to distinguish trait structure from trait judgments, and the present work on accessibility networks from earlier studies on the dimensions of evaluative judgments.

Trait judgments and trait structure. The simplest reason for being concerned with trait structure, instead of simply looking at trait judgments separately, is that we wish to know the evaluative standards applied to candidates. How are the separate judgments organized into a coherent whole, or arc they not organized at all? Evidence of connections between the tfaits tells us that citizens are applying or finding, consciously or not* a unifying theme to trait judgments. Over and above simply knowing what these unifying themes are, there is also a wealth of information to be gained by considering 192 what the structure helps the citizen do, how it affects the processing of information. One example of this is in trait inference. Consider a voter who organizes trait judgments along the lines of the Sub­ network Model. If the voter believes that a politician is well- described by the word intelligent, he or she might infer that the leader is also possessed of other traits indicative of competence. The information on intelligence may be of less help in deciding if the political figure is a strong leader, and of even less help on decisions regarding integrity and empathy. If, on the other hand, the voter does not employ such a complicated organizational structure, but instead simply organizes traits into those that an elected official should have and those she shouldn't have, the voter may have a very different time of it. In this case, information on intelligence makes it more likely to infer any other positive trait, not just those identified earlier as exemplifying competence or capability. This becomes especially important if we also consider that the different dimensions may not be weighted equally, a topic not well covered in the current literature on political trait evaluation. Concern with the complexity and dimensionality of trait structure brings to mind an earlier debate over complexity and constraint in belief systems. Of special significance is the question of how these differ across individuals. The earlier debate, cast in terms of ideology, focused on differences in degree or type of organization based on measures of ability (Converse 1964; Marcus, Tabb and Sullivan 1974; Stimson 1975). Here, I have considered differences <

193 in complexity of trait structure based on involvement in politics. Unfortunately, analysis of the present data does more to raise new questions than answer old ones, presenting conflicting results for two of the three political leaders used in the study. Data measuring different aspects of the subjects' general political involvement and level of candidate-specific information, which might provide a clearer explanation, are simply not available. Obtaining an explanation of these anomalies is important because the answer will have a direct bearing on our understanding of democratic government in the United States. Do those with a higher level of political involvement, information, or cognitive ability unitize trait information? Do those with lower scores on those measures have a simpler view of character traits, organizing them simply in terms of their desirability? The first of these indicates the presence of a group of people who, while not necessarily using the policy information that democratic theorists might prefer, are applying their cognitive might to political decision-making in a valid and useful way. If large numbers of people fall into the other category, we are presented with a picture of a populace that, already failing to pass the higher hurdle of ideological thinking, also fails to apply anything more than simple evaluative cues in their use of trait information. Yet another blow would be struck against the model of the rational (however limited) voter necessary to democratic theory.

Networks of activation and dimensions of evaluation. The study of trait structure and inference takes on added importance 194 when one considers the findings on priming and accessibility. Patterns of inference are likely to follow patterns of accessibility, because people use information that is more accessible. This leads to the need for the distinction between networks of activation and dimensions of evaluation. Previously, our information about the evaluative standards used by citizens to assess the personal qualities of leaders came from one of two sources: open-ended likes and dislikes questions about the candidates and trait batteries in which the respondent was asked how well he or she thought the word described the political figure. The first of these was fine for assessing which qualities people thought of when they thought about specific political figures. It did little to help us with our questions about structure. The trait batteries, on the other hand, did allow researchers to look for patterns of covariation in the trait judgments that could be attributed to underlying evaluative standards. The problem with that type of analysis was that it only told us which evaluations covaried. This information could then be used to reduce a large number of judgments about specific traits to one or two pieces of information on candidate evaluation, useful in a model to predict the vote, but not as informative as one might wish. As students of individual political behavior we are obliged to formulate theories about why people behave the way they do, or think the things they do, and so we must move beyond mere description. Trait judgments must become more than an entry in a model used to predict the vote. We must deal with topics such as the 195 sources of standards of evaluation and the effect of organization on how the individual handles new information, makes evaluations, and eventually behaves. Treating trait structure in terms of networks of activation allows us to do this. Information regarding activational relationships, as the evidence presented in the preceding chapters shows, not only allows a description of trait connections but also permits us to understand how they are related. When people think about a trait such as 'Hard­ working, other traits become more accessible. This phenomenon, known as spreading activation, is the basis for the relationships between groups of traits that form activational networks. Coupled with what is already known about the propensity of decision-makers to use accessible information, especially in situations that do not offer special rewards for added investment of attentional resources, we can now begin to investigate the effects of trait structure on information processes such as inference, categorization, recall, and evaluation.

Implications - structure and accessibility When people evaluate the personal qualities of political leaders, their evaluations tend to be organized along two major dimensions, each of which has two sub-dimensions. Some traits belong to a group best described as capability traits, which can be further broken down into competence and leadership. Others fall into the category of sociability, the components of which are integrity and empathy. This finding, produced by other researchers using 196 various political figures and national samples, was reproduced in my sample of college students with a surprising degree of similarity. More importantly, the results presented indicate that when people use these traits to evaluate political leaders, the traits that have been found to be evaluatively related become more accessible. There is a structure of accessibility and it closely resembles the structure of evaluation. The results presented in the first study indicate that the level of accessibility of traits within a dimension covary, in the same way that the likelihood of ascribing traits to a leader does. The two types of structure, resembling each other so closely, no doubt bear some causal relationship to each other.

Trait structure, accessibility, and inference. The most likely causal mechanism is the process of inference from other trait evaluations. As Conover and Feldman (1986) have pointed out in the context of political schemas, voters sometimes need to make guesses about objects in their political environment, whether because of incomplete information, time constraints, or simply the wish to conserve cognitive resources. They used the schema concept to explain how currently held information was used as a basis for the inferences. The associative memory model offers a more concise statement of this process, grounded in the related notions of activation networks and accessibility. Inference in this framework works in the following manner.3 When an individual wishes to make a judgment on a personality characteristic, the mere process of 197 thinking about the characteristic increases the level of activation of related traits. Some of these are brought into working memory, along with their affective tags. Once present in working memory, these related traits are used as guides in making the current judgment. O f course, other information is used as well, so the traits are not perfectly related, and the exact mix of information is dependent upon circumstantial factors related to accessibility.

Inference and accessibility outside the trait context. Trait- related information and evaluation does not exist in a vacuum. Although not dealt with directly in this research, the notion of activational networks and associative memory structure is more broadly applicable to political attitudes on a wide range of objects. Like the schema concept, it is applicable to parties, class, race, issues, and ideology. Since the basis upon which the links in an activational network are formed is repeated conjunction, any context which repeatedly brings the same constructs into working memory at the same time may eventually lead to activational linkages. The network need not be limited to traits, but may include nodes from any category of socio-political constructs: issues, political leaders, movements, organizations, emotions, traits, and everyday people such as friends or co-workers. Any of these things may act to make the others more accessible, once the activational link is established in long-term memory. 198 Current efforts at understanding accessibility effects in political science have focused on other determinants such as recency of use or frequency of use, as in the study of priming effects by Iyengar and Kinder (1987; see also Iyengar 1990; Krosnick and Kinder 1990), and the chronicity explanation used by Lau (1990). While there is profit to be found in looking at those types of variables as sources of accessibility effects, activational networks should not be ignored. Recency and frequency are useful only in terms of a single concept at a time; activational networks allow us to explore connections between groups of concepts. They allow us to explore the inner structure of political thought, not just the outer contours.

Implications - positivity and negativity Perhaps the most surprising discovery in my research came out of Study Two, and deals with the priming and inhibiting effects of positive and negative traits. In general, using a positive prime for a negative trait judgment had less effect than any of the other three types of pairings. The simplest explanation, that the relationship is weaker in the case of evaluatively incongruent pairs, is not adequate to explain this. It is inadequate because some of the strongest effects were found for the exact opposite pairing, in which a negative prime was used prior to making a judgment on a positive trait. In this case, both an activational effect as well as a previously unsuspected inhibiting effect were found. The general pattern of these effects was that seeing a negative but somewhat related trait caused a subject to make the decision 199 more slowly. This inhibiting effect was somewhat overcome if the trait was negative but highly related. One explanation for this is that the evaluative information in the negative trait prime activates other negative traits much more strongly than the positive traits, tying up limited resources. When the positive word comes into view, the resources must be freed up before it can be categorized and evaluated. If the trait was very closely related, this latter step is somewhat easier, thus the mitigating effects found for primes from the same sub-network. That negative trait-words are strong primes for other negative words is shown in this same study, where strong effects were found for that type of pairing. So, it seems that activation spreads through two types of connections, associational relationships and evaluative ones. The strong activation of other affectively negative traits produces a log-jam in working memory, preventing positive traits from being as readily recognized and evaluated as they might otherwise be. Positive primes, on the other hand, do not produce an inhibiting effect for negative words, nor do they typically produce much of an activational effect for other positive words, with the exception of judgments about Jackson.

Priming affect. There are a number of intriguing possibilities brought to mind by these findings. First, they offer additional support to the possibility, discussed by Ottati and Wyer, that a person's present affective state or the affect elicited by accessible words may be confused with affect felt for an object of evaluation 200 (1990). Consider two objects, George Bush, about whom I have little or no affective information, and the American flag, for which I have strong positive feelings. What happens if, after being repeatedly exposed to campaign commercials in which I see both Bush and the flag, I am asked to evaluate George Bush. If conditions are right and the associative link has been established, simply by thinking about 'George Bush' makes 'the flag' more accessible, perhaps so accessible that it automatically pops into my mind. At this point, I may confuse my feelings for the flag with my feelings for George Bush.4 Obviously, this is an over-simplification, and there would be a number of other factors that would come into play, but the possibility of such a process occurring is not to be ignored. My findings indicate that this connection would be easier to make for particular types of pairings. The effects should be larger if the objects already elicit similar evaluations, especially if both are negatively evaluated. The inhibiting effects present for some trials using positive-positive pairings are inexplicable at this time, so further research needs to be done before predictions about those effects may be made. Negative-positive pairings provide an especially fascinating set of possibilities. My results show that priming with a negative word can make positive words less accessible, unless they are very closely related. So, to return to the example, if I already have a negative evaluation of Bush, I am less likely to experience the accessibility effects programmed by the repeated conjunction with the positively evaluated flag. The negative evaluation blocks positively evaluated 201 associates. The activation of positive words did not inhibit judgments on negative words, so we would expect that, were the evaluations of Bush and the flag to be reversed, accessibility effects resulting from the campaign commercials would be present.

Priming and inhibiting standards. There might also be more direct effects that do not depend upon the confusion of affect across attitude objects. Recall that people are simply more likely to use accessible information, either as evaluative standards or as guides to inference (Tversky and Kahneman 1974; Kahneman, Slovic, and Tversky 1982; Iyengar and Kinder 1987). I have shown that priming people with a negative trait makes it more difficult, more time consuming, for them to make a judgment on a positive trait; in short, it makes the positive traits less accessible. Repeated exposure to negative standards not only makes other negative standards more accessible, it blocks positive standards as well. Now, I have not been concerned here with the effects of priming on direction and extremity of evaluation, but with mapping out the connections between the traits themselves, knowing from other studies that traits do have an effect on evaluation. Yet, if accessibility has an effect on the standards used, and the positivity or negativity of a prime has a differential effect on the accessibility of positive and negative traits, there is certainly the possibility that mere exposure to negative primes causes differences in evaluations on positive traits. I commend the test of this hypothesis to students of political behavior. The present research has established methods 202 of priming that produce the necessary differences in accessibility; their appropriate application should produce a valid test.

Implications - the sources of trait structure Future research effort should be directed to the discovery of the sources of trait structure. Here, I have speculated that the generic model fits less well for the highly-involved in spite of, rather than because of, more frequent exposure to and interest in mass media news coverage. This is not to say that I have given up on the hypothesis that the conjunction of trait information in news broadcasts parallels that found in the generic model. The fact is that the model of trait structure confirmed here is a generic model, and is likely to have a general source. The most logical source is the mass media. Currently, there is no research on the development of trait structure as a candidate becomes better known. Can we expect judgments about almost unknown candidates to be completely unstructured, or will their organization parallel existing generic ties? It would also be interesting to observe the differentiation of candidate structure, in the popular mind, over the course of a campaign and see if or how this is paralleled in campaign news coverage.

This study and future research The findings presented in the current research certainly have implications for the research programs of many scholars in all 203 corners of the discipline. There are also several things that should be said about my own particular research agenda as it relates to the topic of trait structure. Doing research lets a good researcher finds answers to the substantive questions he has about political phenomena and also points the way to better methods for finding those answers the next time he sets out on the quest.

Subjects. A general complaint about social psychological research is the narrow subject base upon which many of its conclusions rest (Sears 1986). This research, too, is based on the responses of college students and conducted in a laboratory setting. I believe that this was, in general, not problematic. The key aspect of the present study was a comparison of the fit of the evaluative structure and the activational structure in a group of subjects. This does not in any way depend upon the level of affect for any of the figures being evaluated, which is likely to be different for a group of college students than for a sample of the populace as a whole. It does not even matter if the evaluative structure of trait judgments in my subjects matches that found in the general populace. In fact, however, it was found to match quite well. The fit of the Sub-network Model, with the negativity factor, was very similar to that reported by Kinder for a national sample (1986). The loadings on the factors were in the same range, as were the reliabilities. Again, this similarity is not necessary for the success of 204 the comparison, but it should act to lend greater weight to the findings. What is necessary to the success of the study is that the subjects process political information in a manner no different than the average citizen. Is there any reason to expect the relationship between evaluation and activation to be any different for college students? I have found no reason to expect this. There are differences in styles and capability of information use based on expertise, and there may be a narrower range of expertise or involvement in these subjects than in a national sample. However, what these students lack in one (involvement, having less to gain or lose in the political arena), they may make up for in the other (expertise, since all were enrolled in an introductory American Politics course at the time of the studies). So, while future research should use a wider range of subjects for purposes of exploring differences in political involvement, the choice of subjects in this case should have had no impact on the ability of the research to answer the questions it was designed to address.

Political leaders and election cycles. Other possible limitations on the generalizability of the results are difficult to overcome within the scope of a single study. Given a limited number of subjects, who can answer only a limited number of questions, the pool of leaders that could be included as stimuli was restricted in the first study and became even smaller in the second. Yet, even with greater resources, one would have been hard-pressed to find a larger pool of leaders of 205 'presidential' caliber. The only other obvious choice during this time period was Lloyd Bentsen, Dukakis' running-mate during the ill-fated 1988 presidential campaign. The other option available would have necessitated inclusion of state-level leaders from Ohio, such as Senators John Glenn and Howard Metzenbaum, or Governor Richard Celeste. This strategy has several drawbacks, the most telling of which is that little is known about trait judgments at other than the presidential level.5 Furthermore, the subjects could be expected to have much less information about some of these other politicians, as was determined in a pre-test that did include a number of holders of state-wide office. Another strategy that could have been used to create a broader range of candidates with known qualities would have been to create a set of artificial political figures. This strategy would not have been appropriate, given the nature of this research. The associative model of memory portrays the networks in long-term memory as products of repeated conjunction built up over time. Judgments of artificial candidates may well be structured in some manner, but this organization probably bears a closer relationship to some ideal or generic candidate network. Again, that in itself is an interesting line of research, but it was not within the scope of this research. Future efforts, however, should include aspects of this in the design. Again, I believe that the use of artificial candidates may not be appropriate. Instead, I believe a repeated cross-section, using measures such as those used here, at four or five time points over the course of the nomination and general election campaign, would prove immensely valuable. Such a design could not only ascertain differences across candidates in trait structure but trace the evolution of trait structure with respect to a single candidate over time. Coupled with data on campaign themes, media coverage, and media usage by the subjects, such a design could answer questions about the sources, development and impact of trait structure and trait evaluation.

Experiments, surveys and measuring accessibility. This pair of studies benefitted by incorporating survey methods and the experimental method, albeit both in a laboratory setting. Study One, though using a rather different data collection method than is normally encountered in a survey context, was survey research pure and simple. The computer itself was set up to do the interviewing and record the responses, a combination of CAT1 and self­ administered survey techniques. Indeed, the research could have been conducted in the subjects' own homes, had I had access to a more portable computer and the time to visit each subject separately. This possibility should certainly be kept in mind for future efforts, where the researcher wishes to use a broader range of subjects. The technique requires nothing more demanding of the subject than the ability to push one of two buttons, clearly marked. With the proper introduction and computer interface, some subjects might even enjoy this more than a face-to-face or telephone interview. 207 The second study, using a very similar data collection method, was a true experiment. The accessibility of trait information was manipulated and changes in the dependent variable measured. This information was coupled with data collected from the same subjects in a pencil and paper survey, and predictions made on the basis of the experimental findings were confirmed in the familiar survey data. The coupling of two methods of data collection, as well as two types of research design, is useful for several reasons. The first study, which simply compared the structure found in two very different types of data by applying the same analysis technique to both, demonstrated not only the similarity of the two types of structure but also the utility of the new measure. The second study, incorporating data collected through two types of research designs, allowed me to compare the results from a relatively new priming technique to a tried and proven survey technique. The use of experiments in political science research seems to have come into its own in the last decade, no doubt due to the influence of social and cognitive psychology. The research presented here has certainly benefitted both from the theory and the experimental methods and measures borrowed from that discipline. It has also benefitted because the results of the two different types of studies, and the two different types of measures, could be compared. The results of each have aided in the understanding of the other, just as the work in the two disciplines has done and will continue to do. 208

C onclusion This dissertation has melded psychological theory and political science research concerns to extend our knowledge of one aspect of political evaluation—the structure of trait judgments. Prior to this, our understanding of how traits were organized in the minds of citizens was limited because we had not yet asked the right questions. There was interest in how traits were organized, but researchers had gone only so far as describing this organization and had not asked why this was so, how it came about, or what this organization really meant. By reframing the question and seeking out a way to answer it in its new form, this research has both confirmed existing results and extended our knowledge in several ways. The question asked about trait structure prior to this was simply, "Which traits go together to form dimensions of evaluation?" This is a simple enough question to answer using standard dimensional analysis techniques. Unfortunately, those techniques do not tell us what is meant by 'to go together1 or 'form dimensions.' The analysis merely tells the researcher which judgments covary. The answer to a question is only as good as the question itself. In a sense, I was not unhappy with the answers I was reading in the literature on trait judgments, but with the questions being asked. So, I sought to refine the question. It became, "Which traits are connected in people's minds, are more likely to be used together, or form cohesive units for a reason that I can explain?" 209 I came to realize, as I looked for an answer, that what I was after was a picture of the associational structure of traits. I wanted to know which traits caused which other traits to come into people's minds. Once I had found this tool, the theory of associative memory structure, I knew the answer was near. In finding my answer, I have also raised many other questions. Most of these are brought about by speculation on the application of this model of memory structure to other aspects of evaluation. So it is that by worrying about questions already answered are advances made and old answers confirmed or rejected. 210 Notes

1. Of course, this finding has sparked decades of debate over the relative importance of policy voting and other determinants of the vote, as well as policy representation and the influence of the public on policy. On the latter, see Monroe 1979; Page 1983; Erickson 1976; Weissberg 1976.

2. For several excellent discussions of this topic, see Simon 1960; Simon 1975.

3. Of course, much of the following discussion bears the same caveat that a schema-based explanation of inference would. At different times, in different situations, people process information with more or less attention to detail. While there are frequently rewards for being a cognitive miser, there are sometimes punishments associated with shortcuts. For a full discussion of these factors, see Fiske and Pavelchak, 1985.

4. The relationship between an attitude object and its evaluation is explored in several places by Fazio and his colleagues (Fazio, et al. 1982; Fazio et al. 1986; Fazio and Williams 1986).

5. It would be interesting to examine these judgments in other electoral contexts, but this was not within the scope of this research. Appendix 1

The following pages contain the questionnaire used for Studies One and Two. The only difference between the questionnaires across the two studies is the absence of the trait rating battery for Dan Quayle in Study Two. Political Candidate Study Protocol 90B0022-2

Please complete the computer part of the study before opening this booklet.

To insure credit for your participation, record your Student ID number here: 213

1. Some people seem to follow what's going on in government and public affairs most of the time, whether there's an election going on or not. Others aren't that interested. Would you say that you follow what's going on in government and public affairs

most of the time. some of the time. only now and then. hardly at all.

2. Please rate the following people on a scale from 0 to 100. A "0" means that you do not like the person at all and a "100" means that you like the person very much. A "50" means that you neither like nor dislike the person.

George Bush ______

Jesse Jackson ______

Michael Dukakis ______

Dan Quayle______

3. In an average week, how often do you watch the

local television news? national television news? Never Never One or two days One or two days Three to five days Three to five days Six or seven days Six or seven days 214

4. Generally speaking, do you consider yourself a Republican, a Democrat, an independent, or what?

Republican Democrat Independent Other Please specify______Don't know

At the top of each of the following pages is a name. Below each o f the names is a list of words that might be used to describe that person. Indicate whether or not you think the word describes the person by circling a number.

Circling the numeral 1 means that you believe the word does not describe the person at a ll.

Circling the numeral 5 means that you believe the word does describe that person extremely well.

The numbers 2, 3, and 4 should be used to indicate intermediate levels between "not at all" and "extremely well".

It is important that you take your time and think about each ward for each person. Consider each word carefully before making your decision. 215

George Bush

Does not D escribes describ e Extrem ely W ell Decent 1 2 3 5 P rejudiced 1 2 3 5 Influential 1 2 3 5 Compassionate 1 2 3 5 Trustworthy 1 2 3 5 In sp irin g 1 2 3 5 C orrupt 1 2 3 5 Hard Working 1 2 3 5 Intelligent 1 2 3 5 Foolish 1 2 3 5 Inexperienced 1 2 3 5 Unqualified 1 2 3 5 Understanding 1 2 3 5 Knowledgeable 1 2 3 5 Moral 1 2 3 5 Power Hungry 1 2 3 5 Weak 1 2 3 5 A Leader 1 2 3 5 A im less 1 2 3 5 Unkind 1 2 3 5 F rien d ly 1 2 3 5 Cruel 1 2 3 5 H esitant 1 2 3 5 D ishonest 1 2 3 5 216

MICHAEL DUKAKIS

Does not Describes Describe Extremely Well Decent 1 2 3 4 5 Prejudiced 1 2 3 4 5 Influential 1 2 3 4 5 Compassionate 1 2 3 4 5 Trustworthy 1 2 3 4 5 Inspiring 1 2 3 4 5 Corrupt 1 2 3 4 5 Hard Working 1 2 3 4 5 Intelligent 1 2 3 4 5 Foolish 1 2 3 4 5 Inexperienced 1 2 3 4 5 Unqualified 1 2 3 4 5 Understanding 1 2 3 4 5 Knowledgeable 1 2 3 4 5 Moral 1 2 3 4 5 Power Hungry 1 2 3 4 5 Weak 1 2 3 4 5 A Leader 1 2 3 4 5 Aimless 1 2 3 4 5 Unkind 1 2 3 4 5 F riendly 1 2 3 4 5 Cruel 1 2 3 4 5 Hesitant 1 2 3 4 5 Dishonest 1 2 3 4 5 217

Jesse Jackson

Does not Describes Describe Extremely Well D ecent 1 2 3 4 5 P rejudiced 1 2 3 4 5 Influential 1 2 3 4 5 Compassionate 1 2 3 4 5 Trustworthy 1 2 3 4 5 In sp irin g 1 2 3 4 5 C orrupt 1 2 3 4 5 Hard Working 1 2 3 4 5 In tellig en t 1 2 3 4 5 F oolish 1 2 3 4 5 Inexperienced 1 2 3 4 5 U nqualified 1 2 3 4 5 Understanding 1 2 3 4 5 Knowledgeable 1 2 3 4 5 Moral 1 2 3 4 5 Power Hungry 1 2 3 4 5 Weak 1 2 3 4 5 A Leader 1 2 3 4 5 A im less 1 2 3 4 5 Unkind 1 2 3 4 5 F rien d ly 1 2 3 4 5 C ruel 1 2 3 4 5 Hesitant 1 2 3 4 5 D ishonest 1 2 3 4 5 218

8. On the next page is a list of words people sometimes use to describe themselves. Indicate how well you think each word describes yourself by circling the appropriate num ber.

Please take your time and consider each word carefully. 219

YOURSELF

Does not Describes Describe Extremely Well Decent 1 2 3 4 5 P rejudiced 1 2 3 4 5 Influential 1 2 3 4 5 Compassionate 1 2 3 4 5 Trustworthy 1 2 3 4 5 In sp irin g 1 2 3 4 5 C orrupt 1 2 3 4 5 Hard Working 1 2 3 4 5 Intelligent 1 2 3 4 5 Foolish 1 2 3 4 5 Inexperienced 1 2 3 4 5 Unqualified 1 2 3 4 5 Understanding 1 2 3 4 5 Knowledgeable 1 2 3 4 5 Moral 1 2 3 4 5 Power Hungry 1 2 3 4 5 Weak 1 2 3 4 5 A Leader 1 2 3 4 5 A im less 1 2 3 4 5 U nkind 1 2 3 4 5 Friendly 1 2 3 4 5 C ruel 1 2 3 4 5 Hesitant 1 2 3 4 5 D ishonest 1 2 3 4 5

I 220

Now we would like to know something about the feelings you have toward several different political figures, feelings caused either by something about the person or by something they have done. Please indicate, by circling YES or NO, whether the person named in each section has ever made you feel the emotion presented in the left-hand column.

9. Michael Dukakis

Has he ever made you feel A ngry Yes No Not sure Hopeful Yes No Not sure A fraid Yes No Not sure P roud Yes No Not sure

10. George Bush Has he ever made you feel A ngry Yes No Not sure Hopeful Yes No Not sure A fraid Yes No Not sure P roud Yes No Not sure

11. Jesse Jackson Has he ever made you feel A ngry Yes No Not sure H opeful Yes No Not sure A fraid Yes No Not sure P roud Yes No Not sure 221

In this final section, please answer a few questions about yourself.

12. Your year in school is (circle one): 1 2 3 4 Other

13. Gender Male ______Female

14. Your year of birth: .

15. Your race: .

16. Are you currently registered to vote? Yes No

Thank you for your participation in this study. Please close the booklet and leave it on the desk. Don't forget to check out with the experimenter

before you leave. Appendix 2

The following two tables are an example of the full results from one of the regression analyses run in Chapter 5. The results presented are for George Bush and show the estimates for the evaluatively incongruent and congruent trials separately. These results are meant to be illustrative of the estimates for the variables included in the model that are not of substantive interest for the problem at hand.

222 223

Evaluatively Incongruent Trials

Dependent Variable: Time R-square 0.1308 N=2404

Param eter Estimate Prob > T Intercept 1.61 0.0001 Trial number -0.004 0.0001 Avg. Response time 0.53 0.0001 Response (yes/no) -0.16 0.0002 Interest 0.02 0.2868 Local Television 0.07 0.0055 National Television -0.14 0.0001 Gender 0.08 0.0059 Registration 0.13 0.0001 Hardworking 0.06 0.5081 Intelligent -0.03 0.7157 Inexperienced 0.23 0.0044 Unqualified 0.15 0.0676 A Leader -0.13 0.1229 Inspiring 0.15 0.0895 Aimless 0.04 0.6223 Hesitant 0.21 0.0151 Decent 0.17 0.0460 Moral 0.04 0.6401 Dishonest -0.08 0.3336 Corrupt -0.13 0.0970 Friendly -0.00 0.9619 Understanding -0.01 0.8880 Prejudiced -0.20 0.0122 Thermometer Rating -0.003 0.0001 Primary Network 0.09 0.0091 Sub-Network Only -0.07 0.0707 Primary plus Sub-network 0.02 0.6022 224 Evaluatively Congruent Trials

Dependent Variable: Time R-Square 0.1489 N=2076

Param eter Estimate Prob > T Intercept 2.04 0.0001 Trial number -0.004 0.0001 Average Response Time 0.64 0.0001 Response (yes/no) -0.03 0.4904 Interest -0.07 0.0035 Local Television -0.06 0.0178 National Television -0.04 0.1292 Gender 0.02 0.4683 Registration 0.20 0.0001 Hardworking -0.25 0.0121 Intelligent -0.47 0.0001 Inexperienced -0.02 0.8521 Unqualified -0.11 0.2125 A Leader -0.36 0.0004 Inspiring -0.27 0.0041 Aimless -0.26 0.0037 Hesitant 0.22 0.0185 Decent -0.31 0.0011 Moral -0.38 0.0002 Dishonest -0.19 0.0475 Corrupt -0.14 0.1145 Friendly -0.25 0.0155 Understanding -0.29 0.0028 Prejudiced -0.23 0.0083 Thermometer Rating -0.003 0.0001 Primary Network -0.04 0.2360 Sub-Network Only 0.06 0.2938 Primary plus Sub-network 0.02 0.7317 References

Abelson, R. P., D. R. Kinder, and S. T. Fiske. 1982. "Affective and Semantic Components in Political Person Perception." Journal o f Personality and Social Psychology 42:619-30.

Anderson, J. R. 1983. "A Spreading Activation Theory of Memory." Journal o f Verbal Learning and Verbal Behavior 22:261-95.

Asher, H. B. 1983. "Voting Behavior research in the 1980's: An Examination of Some Old and New Problem Areas." in Political Science, The State o f the Discipline, ed. A. Finifter. Washington, D. C.: American Political Science Association.

Barber, J. D. 1980. The Pulse o f Politics. New York: Norton.

Bargh, J. A., and P. Pietromonaco. 1982. "Automatic Information Processing and Social Perception: The Influence of Trait Information Presented Outside of Conscious Awareness on Impression formation." Journal of Personality and Social Psychology 43:439-449.

Bargh, J. A., N. A. Bond, W. J. Lombardi, and M. E. Tota. 1986. "The Additive Nature of Chronic and Temporary Sources of Construct Accessibility." Journal o f Personality and Social Psychology 50:869-78.

Bentler, P. M., and C. Chou. 1988. "Practical Issues in Structural Modeling." in Common Problems/Proper Solutions, ed., J. Scott Long. Beverly Hills, CA: Sage.

Brody, R. A., and B. I. Page. 1972. "Comment: The Assessment of Issue Voting." American Political Science Review 66:450-58.

Campbell, A., G. Gurin, and W. E. Miller. 1954. The Voter Decides. Westport, CT: Greenwood. 225 226

Campbell, A., P. E. Converse, W. E. Miller, and D. E. Stokes. 1960. The American Voter. New York: Wiley.

Conover, P. J. 1981. "Political Cues and the Perception of Political Candidates." American Politics Quarterly 9:427-48.

Conover, P. J., and S. Feldman. 1984. "How People Organize the Political World: A Schematic Model." American Journal of Political Science 28:95-126.

Conover, P. J., and S. Feldman. 1986. "The Role of Inference in the Perception of Political Candidates." in Political Cognition: The Nineteenth Annual Carnegie Symposium on Cognition, ed. R. R. Lau and D. O. Sears. Hillsdale, NJ: Erlbaum.

Conover, P. J., and S. Feldman. 1989. "Candidate Perception in an Ambiguous World: Campaigns, Cues and Inference Processes." American Journal of Political Science 33:912-40.

Converse, P. E. 1964. "The Nature of Belief Systems in Mass Publics." in Ideology and Dissent, ed. D. E. Apter. London: Collier- McMillan.

Converse, P. E. 1966. "The Problem of Party Distances in Models of Voting Change." In The Electoral Process, ed. M. K. Jennings and L. H. Zeigler. Englewood Cliffs, NJ: Prentice-Hall. deGroot, A. 1983. "The Range of Automatic Spreading Activation in Word Priming." Journal o f Verbal Learning and Verbal Behavior 22:417-436. deGroot, A., A. Thomassen, and P. Hudson. 1982. Associative Facilitation of Word Recognition as Measured for a Neutral Prime." Memory and Cognition 10: 358-370.

Downs, A. 1957. An Economic Theory o f Democracy. New York: Harper.

Entman, R. M. 1989. "How the Media Affect What People Think: An Information Processing Approach." Journal o f Politics 51:347- 70. 227

Erickson, R. S. 1976. "The Influence of Newspaper Endorsements in Presidnetial Elections: The Case of 1964." American Journal o f Political Science 20:207-233.

Erickson, R. S. 1979. "The SRC Panel Data and Mass Political Attitudes." British Journal of Political Science 9:89-114.

Fazio, R. H. 1986. "How do Attitudes Guide Behavior?" In Handbook o f and Cognition: Foundations o f Social, ed. Behavior R. M. Sorrentino and E. T. Higgins. New York: Guilford.

Fazio, R. H. 1990. "A Practical Guide to the Use of Response Latency in Social Psychology Research." InResearch Methods in Personality and Social Psychology, Volume 11 o f Review o f Personality and Social Psychology, ed. C. Hendrick and M. S. Clark. Beverly Hills, CA: Sage.

Fazio, R. H., and C. J. Williams. 1986. "Attitude Accessibility as a Mediator of the Attitude-Perception and Attitude-Behavior Relations: An Investigation of the 1984 Presidential Election." Journal of Personality and Social Psychology 51:505-14.

Fazio, R. H., D. M. Sanbonmatsu, M. C. Powell, and F. R. Kardes. 1986. "On the Automatic Activation of Attitudes." Journal o f Personality and Social Psychology 50:229-38.

Fazio, R. H., J. Chen, E. C. McDonel, and S. J. Sherman. 1982. "Attitude Accessibility, Attitude-Behavior Consistency, and the Strength of the Object-Evaluation Association." Journal of Experimental Social Psychology. 18:339-57.

Fazio, R. H., M. C. Powell, and P. M. Herr. 1983. "Toward a Process Model of the Attitude-Behavior Relation: Accessing One's Attitude Upon the Mere Observation of the Attitude Object." Journal of Personality and Social Psychology 44:723-35.

Feldman, S., and P. J. Conover. 1983. "Candidates, Issues and Voters: The Role of Inference in Political Perception.” Journal o f Politics 45:810-39. 228 Fiorina, M. P. 1981. Retrospective Voting in American National Elections. New Haven: Yale University Press.

Fischhoff, B., P. Slovic, and S. Lichtenstein. 1980. "Knowing What You Want: Measuring Labile Values." In Cognitive Processes in Choice and Decision Behavior, ed. T. Wallsten. Hillsdale, NJ: Erlbaum.

Fiske, S. T., and M. A. Pavelchak. 1985. "Category-based versus Piecemeal-based affective responses: Developments in Schema-triggered affect." In Handbook o f Motivation and Cognition: Foundations o f Social Behavior, ed. R. M. Sorrentino and E. T. Higgins. NY: Guilford.

Fiske, S. T., and S. E. Taylor. 1984. Social Cognition. New York: Random.

Fiske, S. T., D. R. Kinder, and W. M. Larter. 1983. "The Novice and the Expert: Knowledge-based Strategies in Political Cognition." Journal of Experimental Social Psychology 19:381-400.

Fowler, C. A., G. Wolford, R. Slade, , and L. Tassinary. 1981. "Lexical Access With and Without Awareness." Journal of Experimental Psychology: General 3:341-362.

Glass, D. P. 1985. "Evaluating Presidential Candidates: Who Focuses on their Personal Attributes?" Public Opinion Quarterly 49:517-534.

Graber, D. A. 1980. Mass Media and American Politics. Washington, DC: Congressional Quarterly Press.

Hamill, R. and M. Lodge. 1986. "Cognitive Consequences of Political Sophistication." in Political Cognition, ed. R.R. Lau and D. 0. Sears. Hillsdale: Lawrence Erlbaum.

Hamill, R., M. Lodge, and F. Blake. 1985. "The Breadth, Depth and Utility of Partisan, Class, and Ideological Schemas.” American Journal of Political Science 29:850-70. 229 Hayduk, L. A. 1987. Structural Equation Modeling with LISREL: Essentials and Advances. Baltimore: Johns Hopkins University Press.

Heider, F. 1958. The Psychology of Interpersonal Relations. New York: Wiley.

Heider, F. 1958. The Psychology of Interpersonal Relations. NY: Wiley.

Higgins, E. T. , W. S. Rholes and C. R. Jones. 1977. "Category Accessibility and Impression Formation." Journal o f Experimental and Social Psychology 13:141-54.

Higgins, E. T., and G. King. 1981. "Accessibility of Social Constructs: Information Processing Consequences of Individual and Context Variability." In Personality, Cognition and Social Interaction, ed., N. Cantor and J. F. Kihlstrom. Hillsdale, NJ: Erlbaum.

Higgins, E. T., G. King, and G. H. Mavin. 1982. "Individual Construct Accessibility and Subjective Impression and Recall." Journal o f Personality and Social Psychology 42:35-47.

Higgins, E. T., J. A. Bargh, and W. Lombardi. 1985. "Nature of Priming Effects on Categorization." Journal o f Experimental Psychology: Learning, Memory, and Cognition 11:59-69.

Huckfeldt, R. R. 1983. "The Social Context of Political Change: Durability, Volatility, and Social Influence." American Political Science Review 77:929-944.

Huckfeldt, R. R., and J. Sprague. 1990. "Social Order and Political Chaos: The Structural Setting of Political Information." in Information and Democratic Processes, ed. J. A. Ferejohn and J. H. Kuklinski. Chicago: University of Illinois Press.

Iyengar, S. 1990. "Shortcuts to Political Knowledge: The Role of Selective Attention and Accessibility." in Information and Democratic Processes, ed. J. A. Ferejohn and J. H. Kuklinski. Chicago: University of Illinois Press. 230 Iyengar, S., and D. R. Kinder. 1987. News That Matters. Chicago: University of Chicago Press.

Iyengar, S., D. R. Kinder, M. D. Peters, and J. A. Krosnick. 1984. "The Evening News and Presidential Evaluations." Journal o f Personality and Social Psychology 46:778-787.

Joreskog, K. G., and D. Sorbom. 1989. LISREL 7. Chicago: SPSS.

Joslyn, R. A. 1984. Mass Media and Elections. Reading, MA: Addison-Wesley.

Kagay, M. R., and G. A. Caldiera. 1975. "I Like Looks of His Face: Elements of Electoral Choice, 1952-1972." Presented at the Annual Meeting of the American Political Science Association, San Francisco.

Kahneman, D., P. Slovic, and A. Tversky. 1982. Judgments Under Uncertainty: and Biases. NY: Cambridge University Press.

Kantowitz, B. H., H. L. Roediger, and D. G. Elmes. 1988. Experimental Psychology, 3rd ed. New York: West.

Key, V. O., Jr. 1966. The Responsible Electorate: Rationality in Presidential Voting, 1936-1960. Cambridge: Harvard University Press.

Kinder, D. R. 1983. "Diversity and Complexity in American Public Opinion." in Political Science, The State o f the Discipline ed., A. Finifter. Washington, D. C.: American Political Science Association.

Kinder, D. R. 1986. "Presidential Character Revisited." in Political Cognition: The Nineteenth Annual Carnegie Symposium on Cognition, ed. R. Lau and D. O. Sears. Hillsdale, NJ: Erlbaum.

Kinder, D. R., and D. O. Sears. 1985. "Public Opinion and Political Action." in The Handbook o f Social Psychology, 3rd ed., ed.,G. Lindzey and E. Aronson. New York: Random House. 231 Kinder, D. R., and R. P. Abelson. 1981. "Appraising Presidential Candidates: Personality and Affect in the 1980 Campaign." Paper delivered at the Annual Meeting of the American Political Science Association, New York.

Kinder, D. R., and S. T. Fiske. 1986. "Presidents in the Public Mind." in Political Psychology ed., M. G. Hermann. San Francisco: Jossey-Bass.

Kinder, D. R., M. D. Peters, R. P. Abelson, and S. T. Fiske. 1980 "Presidential Prototypes." Political Behavior 315-37.2:

Kinder, D. R., R. P. Abelson, and S. T. Fiske. 1979. "Developmental Research on Candidate Instrumentation: Results and Recommendations." Report to NES Board, Center for Political Studies, ISR, University of Michigan.

Krosnick, J. A. 1988a. "Attitude Importance and Attitude Accessibility." Paper delivered at the Fifty-ninth Annual Meeting of the Midwest Psychological Association, Chicago.

Krosnick, J. A. 1988b. "The Role of Attitude Importance in Social Evaluation: A Study of Policy Preferences, Presidential Candidate Evaluations and Voting Behavior." Journal o f Personality and Social Psychology 55:196-210.

Krosnick, J. A. 1988c. "Psychological Perspectives on Political Candidate Perception: A Review of the Literature on the Projection Hypothesis." Paper delivered at the Annual Meeting of the Midwest Political Science Association, Chicago.

Krosnick, J. A., and D. R. Kinder. 1990. "Altering the Foundations of Support for the Presidnecy Through Priming." American Political Science Review 84:495-512.

Lau, R. R., and D. O. Sears (eds.). 1986. Political Cognition. Hillsdale: Lawrence Erlbaum.

Lau, R. R., D. O. Sears, and R. Centers. 1979. "The Positivity Bias in Evaluations of Public Figures: Evidence Against Instrument Artifacts." Public Opinion Quarterly 43:347-358. 232 Lazarsfeld, P. F., B. Berelson, and H. Gaudet. 1944. The People's Choice. NY: Columbia University Press.

Lingle, J. H. 1983. "Tracing Memory-Structure Activation During Person Judgments." Journal of Experimental Social Psychology 19:480-496.

Lingle, J. H. and T. H. Ostrom. 1979. "Retreival Selectivity in Memory-based Impression Judgments." Journal of Personality and Social Psychology 37:180-194.

Lodge, M., and R. Hamill. 1986. "A Partisan Schema for Political Information Processing." American Political Science Review 80:505-19.

Lodge, M., K. M. McGraw, and P. Stroh. 1989. "An Impression-driven Model of Candidate Evaluation." American Political Science Review 83:399-420.

Lorch, R. F. 1982. "Priming and Search Processes in Semantic Memory: A Test of Three Models of Spreading Activation." Journal o f Verbal Learning and Verbal Behavior 21:468-492.

MacDonald, S. E., J. W. Prothro, G. Rabinowitz, and K. J. Brown. 1988. "Political Evocation and Styles of Candidate Evaluation." Political Behavior 10:117-35.

Marcus, G. E., D. Tabb, and J. L. Sullivan. 1974. "The Application of Individual Differences Scaling to the Measurement of Political Ideologies." American Journal o f Politcal Science 18:405-420.

Markus, G. B., and P. E. Converse. 1979. "A Dynamic Simultaneous Equation Model of Electoral Choice." American Political Science Review 73: 1055-1070.

Markus, H. 1977. "Self-schemata and Processing of Information about the Self." Journal o f Personality and Social Psychology 35:63-78.

Markus, H., and J. M. Smith. 1981. "The Influence of Self-Schemas on the Perception of Others." In Personality, Cognition and 233 Social Interaction, ed., N. Cantor and J. F. Kihlstrom. Hillsdale, NJ: Erlbaum.

McGrath, J. E., and M. F. McGrath. 1962. "Effects of Partisanship on Perceptions of Political Figures." Public Opinion Quarterly 26:236-48.

Meyer, D. E., R. W. Schvaneveldt, and M. C. Ruddy. 1975. "Loci of Contextual Effects on Visual Word-recognition." in Attention and Performance, ed. P. Rabbitt and S. Domic. NY: Academic Press.

Miller, A. H., and W. E. Miller. 1976. "Ideology in the 1972 election: Myth or reality?" American Political Science Review 70:832- 49.

Miller, A. H., and W. E. Miller. 1977. "Partisanship and Performance: 'Rational' Choice in the 1976 Presidential Election." Paper delivered at the Annual Meeting of the American Political Science Association.

Miller, A. H., M. P. Wattenberg, and O. Malanchuk. 1986. "Schematic Assessments of Presidential Candidates." American Political Science Review 80:521-40.

Monroe, A.D. 1979. "Consistency between Public Preferences and National Policy Decisions." American Politics Quarterly 7:3-19.

Nie, N. H., and K. Anderson. 1974. "Mass Belief Systems Revisited: Political Cahnge and Attitude Structure." Journal of Politics 36:540-591.

Nie, N. H., S. Verba, and J. R. Petrocik. 1976. The Changing American Voter. Cambridge, MA: Harvard University Press.

Nimmo, D. D., and R. L. Savage. 1976. Candidates and Their Images. Pacific Palisades, CA: Goodyear.

Nisbett, R. E., and T. D. Wilson. 1977. "Telling More Than We Can Know: Verbal Reports on Mental Processes." 84:231-59. 234 Ottati, V. C., and R. S. Wyer, Jr. 1990. "The Cognitive Mediators of Political Choice: Toward a Comprehensive Model of Political Information Processing." in Information and Democratic Processes, ed. J. A. Ferejohn and J. H. Kuklinski. Chicago: University of Illinois Press.

Pachella, R. G. 1974. "The Interpretation of Reaction Time in Information Processing Research." in Human Information Processing: Tutorials in Performance and Cognition, ed. B. H. Kantowitz. Hillsdale, NJ: Erlbaum.

Page, B. I., and C. C. Jones. 1979. "Reciprocal Effects of Policy Preferences, Party Loyalties, and the Vote." American Political Science Review 73:1071-1089.

Page, B. I. 1978. Choices and Echoes in Presidential Elections. Chicago: University of Chicago Press.

Patterson, T. E. 1980. The Mass Media Election. NY: Praeger.

Patterson, T. E., and R. D. McClure. 1976. The Unseeing Eye: The Myth o f Television Power in National Politics. NY: G. P. Putnam's Sons.

Posner, M. I., and C. R. R. Snyder. 1975. "Attention and Cognitive Control." in Information Processing and Cognition: the Loyola Symposium, ed. R. L. Solso. New York: Wiley.

Powell, M. C., and R. H. Fazio. 1984. "Attitude Accessibility as a Function of Repeated Attitude Expression." Personality and Social Psychology Bulletin 10:139-48.

Rahn, W. M. 1989. "The Role of Partisan Stereotypes in Information Processing About Political Candidates." Paper delivered at the Annual Meeting of the American Political Science Association, Atlanta, Georgia.

Rahn, W. M., E. Borgida, J. Aldrich, and S. Klein. 1988. "The Process of Candidate Appraisal: An Experimental Investigation.” Paper delivered at the Annual Meeting of the American Political Science Association, Washington, DC. 235 Rahn, W. M., J. H. Aldrich, E. Borgida, and J. L. Sullivan. 1990. "A Social-cognitive model of Candidate appraisal." inInformation and Democratic Processes, ed. J. A. Ferejohn and J. H. Kuklinski. Chicago: University of Illinois Press.

Sears, D. O. 1983. "The Person-Positivity Bias." Journal of Personality and Social Psychology 44:233-250.

Sears, D. O. 1986. "College Sophomores in the Laboratory: Influences of a Narrow Data-Base on Social Psychology's view of Human Nature." Journal of Personality and Social Psychology 51: 515-530.

Shabad, G., and K., andersen. 1979. "Candidate Evaluations by Men and Women." Public Opinion Quarterly 43:18-35.

Sigel, R. S. 1964. "Effect of Partisanship on the Perception of Political Candidates." Public Opinion Quarterly 28:485-493.

Sigel, R. S. 1966. "Image of the American Presidency - Part II of an Exploration into Popular Views of Presidential Power." Midwest Journal of Political Science 10:123-37.

Simon, H. A. 1960. 7 he New Science o f Management Decision. NY: Harper and Row.

Simon, H. A. 1975. "Style in Design." in Proceedings o f Second Annual Environmental Design Research Asociation Conference, 1970, ed. J. Archea and C. Eastman. Pittsburg: Camegie- Mellon University Press.

Simon, H. A. 1985. "Human Nature in Politics: The Dialogue of Psychology with Political Science." The American Political Science Review 79:293-304.

Sniderman, P. M., J. M. Glaser, and R. Griffin. 1990. "Information and Electoral Choice." in Information and Democratic Processes, ed. J. A. Ferejohn and J. H. Kuklinski. Chicago: University of Illinois Press. 236 Sprague, J. 1982. "Is There a Micro-theory Consistent with Contextual Analysis?" In Strategies o f Political Inquiry, ed. E. Ostrom. Beverly Hills: Sage.

Srull, T. K., and R. S. Wyer. 1979. "The Role of Category Accessibility in the Interpretation of Information about Persons: Some Determinants and Implicaitons." Journal of Personality and Social Psychology 37:1160-1172.

Srull, T. K., and R. S. Wyer. 1980. "Category Accessibility and Social Perception: Some Implications for the Study of Person Memory and Interpersonal Judgment." Journal of Personality and Social Psychology 38:841-856.

Srull, T. K., and R. S. Wyer. 1986. "The Role of Chronic and Temporary Goals in Social Information Processing." In Handbook o f Motivation and Cognition: Foundations o f Social Behavior, ed. R. M. Sorrentino and E. T. Higgins. New York: Guilford.

Stokes, D. E. 1963. "Spatial Models of Party Competition." American Political Science Review 57:368-377.

Stokes, D. E. 1966. "Some Dynamic Elements of Contests for the Presidency." American Political Science Review 60:19-28.

Stroh, P. K. 1989. "Candidate Ambiguity and Voter Projections." Paper Delivered at the Annual Meeting of the American Political Science Association, Atlanta, GA.

Tagiuri, R. 1958. "Introduction." in Person Perception and Interpersonal Behavior, ed., R. Tagiuri and L. Petrullo. Stanford, CA: Stanford University Press.

Tversky, A., and D. Kahneman. 1981. "The Framing of Decisions and the Psychology of Choice."Science 211:453-58.

Warren, R. E. 1977. "Time and Spread of Activation in Memory." Journal of Experimental Psychology: Human Learning and Memory 3: 458-466. 237 Weisberg, H. F. 1989. "Some Perspectives on the 1988 Presidential Election: The Roles of Turnout and Ronald Reagan." Paper delivered, at the Annual Meeting of the American Political Science Association, Atlanta, GA.

Weisberg, H. F., and J. G. Rusk. 1970. "Dimensions of Candidate Evaluation." American Political Science Review 64:1167-1185.

Weissberg, R. 1976. Public Opinion and Popular Government. Englewood Cliffs, NJ: Prentice-Hall.

Wheaton, B. 1988. "Assessment of Fit in Overidentified Models with Latent Variables." in Common Problems/Proper Solutions, ed. J. Scott Long. Beverly Hills, CA: Sage.

Zajonc, R. B. 1980. "Feeling and Thinking: Preferences Need No Inferences." American Psychologist 39:151-175.