W L3 0 2004 OVERDUE FINES: 25¢ per day pc‘r new RETURNING LIBRARY MTEEXALS: Place in book return to mom charge from circutatmn recur A THEORY OF CITIZEN INFORMATION GATHERING IN ELECTORAL CONTEXTS By Michael M. Gant A DISSERTATION Submitted to Michigan State University in partial fulfillment of the requirements for the degree of DOCTOR OF PHILOSOPHY Department of Political Science 1980 (:> Copyright hy Michael McCoy Gant 1980 ABSTRACT A THEORY OF CITIZEN INFORMATION GATHERING IN ELECTORAL CONTEXTS By Michael M. Gant In this dissertation a formal theory of citizen information gathering strategies in electoral contexts is developed. The theory is motivated by the virtual exclusion of citizens by past efforts to describe the information provision/gathering process during elections. The theory is based on the presumption that citizens gather political information not only to help them decide whether and how to vote, but also to reduce their uncertainty about the correctness of that choice. In this context, a correct decision is one which maximizes expected utility. The theory is empirically examined in two stages, both of which use data from the 1972-l974-l976 Center for Political Studies American National Election Studies. The first stage involves testing a variety of hypotheses derived from, or suggested by, the theory. The second stage of empirical examination involves the estimation of a four-equation, non-recursive statistical model of the theory. Results of the empirical examination of the statistical model are mixed. The ability of the model to account for variation in levels of uncertainty is weak. Further, the model's ability to predict levels of political information and media utilization is moderate, at best. However, the power of the model to explain turn- out, as well as the role of the concept of uncertainty in that ex- planatory power, is quite strong. One additional focus of the theory of citizen information gathering is the role of reinforcement in electoral decision making. For the purpose of analyzing reinforcement, three measures of selec- tive perception are constructed. The three measures are based upon the assumption that the average respondent perception of a candidate's issue position is the "true" position, i.e., in the aggregate, per- ceptions of candidates' issue positions are accurate. Empirical analyses indicate that citizens engaged in selective perception (assumed to serve the purpose of reinforcement) during the 1976 presidential contest. Specifically, partisans tended to perceive the issue positions of their party's candidate as being closer to their own positions than they "actually“ were. Further, partisans tended to perceive the candidate of the opposition party as being further away on the issues than "in reality." These data describe a "classic" case of reinforcement behavior during the 1976. campaign. ACKNOWLEDGMENTS The phrase "a dissertation is a collective effort" never had much meaning for me until I embarked upon this project. Now, I hardly know which parts of this dissertation are mine, and which parts originated elsewhere. In particular, the members of my dissertation committee substantially influenced my thinking. Prof. John Aldrich, who served as chair of the committee, went so far beyond the call of duty, that words fail to describe his effort. Whatever value this dissertation may have is due largely to his refusal to accept any- thing less than my best. Prof. Charles Ostrom helped clarify my thinking on a number of points, and by his example made me understand what research is all about. Prof. David Rohde and Prof. Cleo Cherryholmes took the time from their busy schedules to point out weak spots in the dissertation, and to offer suggestions as to how to improve the analysis. A large part of this dissertation was written while I was a member of the faculty of Texas A&M University, and I would be remiss if I did not acknowledge the contributions of certain individuals there. The Head of the Department of Political Science, Prof. Samuel A. Kirkpatrick, showed tremendous patience and support. By not burdening me with many of the administrative duties that accompany a faculty position, he helped me greatly. A number of my colleagues provided not only intellectual stimulation, but emotional support as 11' well. In particular, I include here such friends as Dwight Davis, James Franke and Roby Robertson. Several of my fellow graduate students at Michigan State University, especially Dennis Simon and William Boyd, played to the hilt the dual roles of critic and com- panion. The individuals who typed the manuscript, Janie Leighman and Vickie Young, deserve mention, not only for their professionalism, but also for their patience. Both of them typed the final draft, and their work was first-rate. Ms. Leighman worked with me on the final revisions, and turned a rather harrowing experience into some- thing much less foreboding. Finally, I wish to thank my wife, Sherry, for her love and patience. Sherry is one of the most intelligent and capable people I know. Still, it is often difficult to explain to a non-academician exactly what it is we do. That she has supported me, in more ways than one, throughout this endeavor is demonstration enough not only of her love, but also of her loyalty. I wanted to dedicate this dissertation to her, but she felt that might be going a bit too far. I think that says more of her sense of perspective, and of our relationship, than anything I could write. Thanks. TABLE OF CONTENTS LIST OF TABLES ................................................. LIST OF FIGURES ................................................ CHAPTER 1: INFORMATION AND MEDIA UTILIZATION IN ELECTORAL CONTEXTS ........................................... Introduction .................................... Review of the Literature ........................ Information and Electoral Politics ........... Information Sources: Utilization and Believability ................................ The Effects of Mass Media Utilization ........ The Mass Media and Information ............... The Mass Media and Voting: Turnout and Direction ....................................... The Mass Media: Attitude Change and Reinforcement ................................... Mass Media Research: Some Organizing Structures Endnotes ........................................ CHAPTER 2: A THEORY OF CITIZEN INFORMATION GATHERING .......... Introduction .................................... A Note on Human Behavior ........................ Assumptions of the Theory ....................... An Informal Description of the Theory ........... Formal Development of the Theory ................ Reinforcement and Rational Information Gathering iv Page ll 14 19 25 29 29 3O 40 48 48 64 2.6.92 Endnotes ........................................ 73 CHAPTER 3: OPERATIONALIZATIONS AND EXAMINATION 0F ASSUMPTIONS. 78 Introduction .................................... 78 The Data and Operationalizations ................ 78 An Examination of Two Assumptions, 1976 ......... 90 Conclusion ...................................... 99 Endnotes ........................................ 100 CHAPTER 4: DERIVATION OF HYPOTHESES ........................... 105 Introduction .................................... 105 Derivation of Hypotheses ........................ 105 Summary of Hypotheses ........................... 119 Hypotheses Regarding Political Information... 119 Hypotheses Regarding Uncertainty. ............ 120 Hypotheses Regarding Media Utilization ....... 120 Hypotheses Regarding Voting .................. 121 Hypotheses Regarding the Timing of the Vote Decision ..................................... 122 Endnotes ..................................... 123 CHAPTER 5: EMPIRICAL EXAMINATION OF HYPOTHESES ................ 124 Introduction ................................. 124 Empirical Examination of Hypotheses .......... 125 Misperception, Partisanship, and Information. 155 Conclusion ................................... 174 Endnotes ..................................... 175 CHAPTER 6: A SIMULTANEOUS-EQUATION MODEL OF THE THEORY OF CITIZEN INFORMATION GATHERING ...................... 178 Introduction .................................... 178 vi Page_ The Statistical Model ........................... 179 Estimation: Some General Considerations ........ 184 Probit Analysis ................................. 187 Goodness of Fit of the Probit Model ............. 191 Problems in Estimating the Model of Simultaneous Equations .......................... 193 Additional Estimation Considerations ............ 195 Specification Error .......................... 196 Multicollinearity ............................ 198 Heteroscedasticity ........................... 199 Conclusion ...................................... 200 Endnotes ........................................ 201 CHAPTER 7: MODEL EVALUATION FOR 1972 .......................... 204 Introduction .................................... 204 The Data ........................................ 208 Discussion ...................................... 224 Endnotes ........................................ 227 CHAPTER 8: MODEL EVALUATION FOR 1974 .......................... 229 Introduction .................................... 229 The Data ........................................ 231 Discussion ...................................... 236 Endnotes ........................................ 238 CHAPTER 9: MODEL EVALUATION FOR 1976 .......................... 239 Introduction .................................... 239 The Data ........................................ 242 Discussion ...................................... 272 Endnotes ........................................ 276 CHAPTER 10: APPENDIX A: APPENDIX 8: APPENDIX C: APPENDIX D: APPENDIX E: Page CONCLUSIONS ..................................... '.. 277 Introduction ................................... 277 The Performance of the Model over Time ......... 277 Data Problems .................................. 285 Future Research ................................ 292 Conclusion ..................................... 295 Endnotes ....................................... 297 PROOF OF THEOREM I ................................ 298 A BRIEF DESCRIPTION OF BAYES' THEOREM ............. 301 TESTS OF ASSUMPTIONS FOR 1972 AND 1974....' ........ 303 1972 Tests of Assumptions ...................... 303 1974 Tests of Assumptions ...................... 306 Conclusion ..................................... 309 CONSTRUCTION OF INSTRUMENTAL VARIABLES ............ 310 Instrumental Variables for 1972 ................ 310 Instrumental Variables for 1974 ................ 311 Instrumental Variables for 1976 ................ 312 Correlations between Instruments and Endogenous Variables ...................................... 313 CORRELATIONS BETWEEN MEDIA USE INSTRUMENTS AND SIX SOCIO-ECONOMIC-POLITICAL VARIABLES, 1972, 1974, AND 1976 .......................................... 316 BIBLIOGRAPHY ................................................... 318 Table 10 11 12 13 14 15 16 LIST OF TABLES Average Information Gain on Candidate Issue Positions, 1972 Campaign .............................. Means and Standard Deviations of Uncertainty, Differential Benefit, Information and Activity Measures .............................................. Summary of Correlational Analysis between Perceptions of Candidates' Issue Positions and Media Reliance Variables (XI-X7) ..................................... Correlations between Media Use Variables and the Government Guarantee of Jobs Scale, National Health Insurance Scale ....................................... Relationships between Political Information and Media Use, and Socio-Political-Economic Variables ........... Relationships between Levels of Political Information and Degree of Media Utilization ....................... Rank-Orderings of Average Candidate-Information Levels Rank-Orderings of Average Candidate-Information Levels Rank-Orderings of Party Information Levels ............ Rank-Orderings Using PARTINFO and AFFECT .............. Relationships between Media Utilization and Timing of the Vote Decision .................................. Relationships between Uncertainty and Information ..... Relationships between Levels of Uncertainty and Degree of Media Utilization .................................. Relationships between Vote Choice and Expected Differential Benefit .................................. Test of Hypothesis 11a ................................ Relationships between Television Utilization and Voting ................................................ viii Page 10 9O 92 93 97 126 136 137 138 139 142 143 146 148 150 151 Table 17 18 19 20 21 22 23 24 25 26 27 28 29 3D 31 32 33 34 35 36 37 ix Page Relationship between Levels of Information and Voting 151 Relationship between Level of Political Information and Vote Choice ....................................... 152 Distribution of V3654 ................................. 153 Relationships between Media Choice and Social Class, Education, and Income, Controlling for Levels of Political Information ................................. 154 Means and Standard Deviations of Misperception Variables ............................................. 159 Means and Standard Deviations of Misperception Variables for Republicans, Democrats, Independents, and A11 Respondents ................................... 164 Relationships between Expected Utility and Selective Perception ............................................ 167 Relationships between Misperception Variables and Levels of Candidate-Related and Party-Related Information ........................................... 169 Relationships between Misperceptions and Information for Republicans, Democrats, and Independents .......... 171 Relationships between Media Utilization and Misperception ......................................... 173 Model Ia, 1972 ........................................ 209 Model Ib, 1972 ........................................ 212 Model 11a, 1972 ....................................... 214 Model IIb, 1972 ....................................... 216 Model 111a, 1972 ...................................... 217 Model 111b, 1972 ...................................... 220 Model IVa, 1972 ....................................... 221 Model IVb, 1972 ....................................... 223 Model Ia, 1974 ........................................ 232 Model Ib, 1974 ........................................ 234 Model Ic, 1974 ........................................ 236 T_abl_e 38 39 4o 41 42 43 44 45 46 47 48 49 so 51 52 53 54 55 56 57 58 Page Model Ia, 1976 ........................................ 243 Model Ib, 1976 ........................................ 247 Model Ic, 1976 ........................................ 249 Model 11a, 1976 ....................................... 252 Model IIb, 1976 ....................................... 254 Model IIc, 1976 ....................................... 255 Model 111a, 1976 ...................................... 259 Model 111b, 1976 ...................................... 262 Model IIIc, 1976 ...................................... 264 Model IVa, 1976 ....................................... 266 Model IVb, 1976 ....................................... 268 Model IVc, 1976 ....................................... 269 Relationships between Media Utilization Variables and Socio-Economic-Political Variables for 1972 ........... 305 Tests of the Unbiasedness Assumption, 1972 ............ 307 Relationships between Media Utilization Variables and Socio-Economic-Political Factors, 1974 ................ 309 Correlations between Instruments and Endogenous Variables, 1972 ....................................... 314 Correlations between Instruments and Endogenous Variables, 1974 ....................................... 314 Correlations between Instruments and Endogenous Variables, 1976 ....................................... 315 Correlations between Media Use Instruments and Six Socio-Economic-Political Variables, 1972 .............. 316 Correlations between Media Use Instruments and Six Socio—Economic-Political Variables, 1974 .............. 317 Correlations between Media Use Instruments and Six Socio-Economic-Political Variables, 1976 .............. 317 Figure oxooowcioiboom a—J-J A 12 13 1.4 15 16 LIST OF FIGURES Pagg_ Hypothetical Representation of Ua and Db .............. 43 The Citizen's Decision Making Process ................. 49 Hypothesis Test Loss Table ............................ 52 Hypothetical EVSI Curve ............................... 58 Cost and Benefit Curve for Si ......................... 60 Categorization of Information Sources ................. 85 Factors Influencing Information Levels ................ 117 Factors Influencing Media Use ......................... 118 Factors Influencing Media Use ......................... 118 Factors Influencing Information Levels ................ 118 Breakdown of CANDINFO by INTEREST, UTILITY, and DECIDE ................................................ 130 Breakdown of CANDINFO by INTEREST, AFFECT, and DECIDE. 131 Breakdown of PARTINFO by INTEREST, UTILITY, and DECIDE 132 Breakdown of PARTINFO by INTEREST, AFFECT, and DECIDE. 133 Equations 1-4 ......................................... 182 Equations 1-4a ........................................ 182 xi CHAPTER 1 INFORMATION AND MEDIA UTILIZATION IN ELECTORAL CONTEXTS Introduction Since the advent of research on electoral behavior, a great deal of attention has been focused on the role of information in electoral campaigns, as well as the characteristics of the well-informed voter. For instance, levels of information have been linked to such factors as voting on the basis of issue concerns (cf., Pomper, 1972; Hinckley, et a1., 1976); levels of involvement in the campaign (cf., Campbell, et a1., 1960); and levels of attitude consistency (cf., Converse, 1964, 1970). In turn, the sources of citizens' information during campaigns have been heavily researched. For instance, it is now well established that more people rely on television for electoral infor- mation than on any other medium (cf., Roper, 1977). Despite the empirical research, very few attempts have been made to build a theory which explains how and why citizens gather and use political information. The purpose of this dissertation is to build and test a formal theory of citizen information gathering strategies in electoral cam- paigns. In general, past empirical and theoretical work has focused either on a specific medium, such as television (cf., Mickelson, 1972), or on the mass media in general, treating the mass media as an undifferentiated whole (cf., Klapper, 1960). Such research has 1 generated specific statements regarding the roles of the mass media in electoral contexts. However, the use of the mass media in elec- toral contexts is primarily an information gathering activity. A major premise of this research is that media utilization cannot be fully understood outside the context of citizen information gathering, and that past research has not fully taken this context into account. By examining the process of information gathering, the theory developed in this dissertation represents a break with past attempts to delineate the role of the mass media. This dissertation primarily focuses on the roles of the mass media in the processes of information gathering. The theory devel- oped will answer two simple questions: How, and why, do citizens use the mass media in electoral contexts? This question is built upon the implicit assumption that individuals use the mass media for specific purposes; that people have intentions when utilizing tele- vision, radio, newspapers and magazines. This assumption is the primary characteristic of the g§g§_agg gratifications approach to the study of mass media. This dissertation will use and expand upon that approach. There is an alternative assumption that underlies mass media research: individuals constitute a passive audience for the mass media. That is, citizens do not use the mass media jg_g§ggr.tg achieve some goal; individuals are affected by the mass media in ways that may or may not be relevant to their own desires. This assumption underlies an approach to the study of the mass media known as the media effects school. The review of the literature, as presented below, makes it very clear that the vast majority of research conducted in the area of the mass media is a part of the media effects school. The literature review has the following format and goals. First, a brief discussion of the role of information in elections is given, followed by research bearing on the sources of information in electoral contexts. These sections, though short, demonstrate that citizens do gather information during elections, and that the mass media, particularly television and newspapers, are the principal sources of that information. The largest section in this review focuses on the effects of the mass media. The impact of the various media on information levels, turnout, the direction of the vote, attitude change and attitude re- inforcement are all systematically explored. This section makes two points. First, a great deal of research has been conducted in the media effects tradition. Second, with the exception of the rein- forcement of attitudes and the creation of new issues, mass media effects are weak, at best. In the final section I will return to the organizing concepts of uses and gratifications, and media effects. These approaches and their relative utility for understanding the roles of the mass media in information gathering will be explored. The roots of the theory of citizen information gathering, which lie in economic models of electoral choice, will be introduced. Downs' discussion of voting and elections (1957) is concerned, among other things, with the role of information in elections. In that regard, rational choice models can be shown to be a part of the uses and gratifications approach to media research. Moreover, I will demonstrate that rational choice models, insofar as they are concerned with information gathering, must be modified to account for the empirical findings stemming from both the media effects and uses and gratifications approaches to media research. Such modifications, and the resultant theory, will be presented in the second chapter. Review of the Literature Information and Electoral Politics As noted previously, information levels have long been a topic of interest among electoral scholars. Levels of information about a particular election have been linked to levels of interest and in— volvement in the campaign (Campbell, et a1., 1960; Berelson, 1966). Information levels have also been associated with intra-campaign partisan stability; citizens with little information about the cam- paign are likely to be unstable in their intra-campaign vote intentions (Converse, 1966). However, the general applicability of Converse's "J-curve“ thesis has been challenged by other analyses of electoral data (cf., Dreyer, 1974; Macaluso, 1977; Zukin, 1977). The "Michigan School," in particular Philip Converse, has attempted to show that in the 1950's voters with little information or education demonstrated little attitude consistency, or constraint, and thus were less "ideological" than their better-informed counter- parts (Campbell, et a1., 1960; Converse, 1964, 1970). As a result, many citizens were viewed as casting votes without the guidance of a cohesive framework. This conclusion has been challenged on several fronts. Converse's methodology has been attacked by Pierce and Rose (1974) and by Achen (1975). Several scholars have suggested that the lack of attitude constraint found in the 1950's did not carry over into the more issue-oriented politics of the 1960's and 1970's (cf., Luttbeg, 1968; Nie with Andersen, 1974). In turn, the "time bound" thesis has been challenged by those who suggest that changes in question wording and methodology negate any inferences of in- creased attitude constraint (cf., Bishop et a1., 1978; Sullivan, et a1., 1978). Even the form of the relationship between education and constraint, and the interpretation which equates consistency with ideological thinking, have been severely questioned (Bennet and Oldendick, 1978). Finally, the degree to which citizens vote on the basis of issue concerns has been linked to the amount of information voters possess (Pomper, 1972), and to the amount of information available to the electorate (Hinckley, et a1., 1976). This finding has been sup- ported by aggregate analyses (cf., Settle and Abrams, 1976), and by experimental evidence (Cangelosi, et a1., 1969). Clearly, there is considerable controversy regarding the specific functions of information in electoral contexts. This controversy does not, however, negate the proposition that citizens do gather in- formation during campaigns. In fact, all of the research regarding the roles of the information levels of citizens is predicated upon the assumption that citizens do gather information in electoral per- iods. The immediate question arising from this proposition involves how people go about gathering information. More specifically, what are the sources of political information for citizens? This is a question which has also attracted a great deal of scholarly atten- tion, and it is to that literature that I now turn. 6 C Information Sources: Utilization and Believability There exists a great deal of evidence which indicates that television is not only the most widely used of the major mass media, but is also the most credible source of political information during electoral periods. A study of Illinois residents in the 1972 elec- tion (Atwood and Sanders, 1975) indicates that the margin of television-to-newspapers utilization as a source of political infor- mation was about three-to—one.1 Further, respondents judged televis- ion to be a more believable source of news than either news magazines or newspapers. The most detailed data regarding the relative use of the mass media have been collected by the Roper Organization under the auspices of the Television Information Office. In a report based on data collected in l976, Roper reports that, in almOst all situations, television is preferred to newspapers as a source of political information.2 The one exception to this general statement is news regarding local elections; respondents favored newspapers over television as a source of information regarding the candidates and issues in local elections. In general, television was considered to be a more believable news source than were newspapers, by a margin of 51% to 22% (Roper, 1977).3 Even so seemingly a benign topic as media utilization has generated some controversy. Stone (1969-1970) suggests caution in drawing inferences from the Roper data, suggesting that question wording might be a problem. At least two other studies of media utilization (Clarke and Ruggels, 1970; O'Keefe, et a1., 1976) indicate that newspapers are more often used as a source of political information than is television. However, these studies were based on local samples (Seattle and Akron, Ohio, respectively), while the Roper data are based on a national sample. Further, data from the Center for Political Studies 1976 American National Election survey indicate that 64% of respondents preferred television as a source of political information, while 19% preferred newspapers. On balance, it seems safe to conclude that television is the most widely used, as well as the most believed, source of political information during electoral campaigns. While television is preferred to newspapers in general, there are differences in media utilization between several socio-economic- political categories. Past research indicates that, in general, television users are poorer (Sargent and Stempel, 1968; Greenberg and Dervin, 1970), less well-educated (Hazard, 1962; McLeod, et a1., 1965), and less politically efficacious (Robinson, 1976) than are newspaper readers. The relationship between efficacy and media utilization is weak, and at least one reanalysis of the same data used by Robinson suggests that the relationship does not exist (Miller, et a1., 1977).4 However, the relationships between media utilization and education and income are substantial, and are not a point of controversy in the literature. These data indicate the existence of differences in media utilization between various social groupings. However, regardless of whether one looks at the rich or the poor, the well-educated or the less well-educated, television is still the medium preferred for gathering political information by a majority of citizens. Given that television is reported by citizens to be their most important political information source, the bulk of the research relevant to the effects of the mass media has focused on television. As a conse- quence, the following discussion, while concerned with all mass media, will of necessity concentrate on the effects of television ut‘i '1 ization. The Effects of Mass Media Utilization One of the most pervasive and long-lasting research traditions in comunications inquiry is the attempt to identify the effects of the mass media on the viewing (listening, reading) audience. In the area of political conmunications, attention has been paid to the effects of the mass media on information learning, voting, attitude change, and attitude reinforcement. The research into information Provision and learning takes the form of comparisons between news- Dapers and television. For the other areas of interest, the results 0f research appear to be invariant with respect to medium; e.g., the effects of television on attitudes, in terms of either change or PEinforcement, are in general the same as those of newspapers. To Present a clear picture of this literature, each of these areas of research will be discussed separately. Ihe Mass Media and Information Virtually all of the research into the ability of the mass media t0 COnvey information compares newspapers and television in terms of the amount of information they provide users. Content analyses of "ewspapers and television news programs provide convincing evidence th9t newspapers not only cover more news stories than does television, but also provide more information in the stories given. This con- clusion holds in analyses of newspapers and television in non- campaign situations (Harney and Stone, 1968-1969; MacNeil, 1968), as well as during electoral campaigns (Patterson and McClure, 1976:82- 84; Patterson, 1980:21-42). Moreover, the type of information given vari es across media. During campaigns, television news stories are more concerned with the personalities of the candidates than the i ssues of the campaign, relative to newspapers (Graber, 1976; Patterson and McClure, 1976:34-42). Given the comparatively low amounts of information provided citizens by television, it is not surprising to find that those who rely on television for information during political campaigns do not gain much information about the candidates and the issues. Patterson and McClure, in their study of the 1972 presidential campaign, datermined the average amount of information gain on candidates‘ issue Positions for several categories of citizens, the categorization being made according to media use. The data are sunmarized in Table 5 l. The data demonstrate the clear superiority of newspapers in PY‘OViding readers with information useful in making a voting decision. 0f Particular interest is the fact that once newspaper readers are 113moved from consideration, regular television viewers exhibited no more issue knowledge gain than those who used neither medium. It is d1ff‘icu1t to argue with the authors when they state "[t]elevision does not help the electorate to vote on the issues . . . this con- C1“Sion is simply inescapable" (Patterson and McClure, 1976:49). 10 Table 1. Average Information Gain on Candidate Issue Positions, 1972 Campaign.* Group Average Information Gain Non-regular TV viewers 25% Regular TV viewers 28% Regular newspaper readers 35% Regular newspaper readers who were not TV viewers 34% Regular TV viewers who were not newspaper readers 19% Those who neither viewed TV regularly nor read newspapers regularly 19% ‘ *Source: Patterson and McClure, 1976:49-54. Television viewers not only gain less information than newspaper rReaders, but experimental evidence suggests that they also tend to 1ose more of the information that they do gain (cf., Wilson, 1974). The situation is compounded by the possibility that those who rely Primarily on television as a source of political information tend to overestimate their gains in knowledge. MacNeil reports a study con— duCted during the 1952 campaign6 and suggested that (MacNeil, 1968: 130. author's emphasis): [aJlthough people who watched television thought they were Inaking the largest gains of information, the researchers found television inferior to newspapers as an instrument for communicating the issues. A StUdy of the 1952 nominating conventions by the Brookings Institu- tion came to the same conclusion (Thompson, 1956:4812 11 The respondents' [television viewers'] estimates of the amount they learned seemed to run far ahead of their increase in learning as judged by improved ability to answer technical questions about the organization and procedures of the con- vention. More recent evidence confirms the proposition that television is relatively weak in conveying substantive information. Patterson and McC‘l ure found that when individuals were asked to recall information about a campaign, television viewers were less likely to recall issue positions and candidates' qualifications than were newspaper readers (Patterson and McClure, 1976:79-80). O'Keefe and Sheinkopf (1976) asked their respondents to name three issues which were important to the 1972 presidential campaign, as well as the sources of information about these issues. Of those who could recall one issue, 56% named television as the source of information, while 30% named newspapers. However, of those who could name a third issue, only 16% named tele- Vision as the news source, while 68% named newspapers.7 The literature confirms the proposition that television is an ineffective conduit for political information, relative to news- Papers. The findings reported in the next section, relevant to the effects of the media on voting, follow logically from this proposi- tion. The Mass Media and Voting: Turnout and Direction Although individuals who have trouble deciding whether and/or “PW to vote are more likely to use television than newspapers as a Source of information about politics (Merrill and PFOCtOF, 1959; O'Keefe, et a1., 1976), there is little evidence that television has ““3 great effect on either turnout or the partisan direction of the 12 vote. A study of the effects of television in Iowa in the 1952 presidential campaign found no relationship between county-wide turn- out and the percentage of homes in a county which had television (Simon and Stern, 1955). Another aggregate analysis, dealing with the 1970 gubernatorial elections, found that campaign efforts via te'levision8 were not related to turnout (Tollison, et a1., 1975). Glaser's study (1965) of the 1960 elections found that, although te‘l evision was the most effective medium in reminding people to vote, the actual relationship between television usage and voting was very weak. Newspaper readers had higher rates of turnout. A study of the 1965 College Park, Maryland election (Conway, 1968) also indi- cated that the use of newspapers, but not television, was related to turnout. Newspaper usage has also been found to exert some influence on the direction of the vote. John Robinson (1972), using 1968 SRC data, matched up newspapers with respondents, and determined the Candidates endorsed by the newspapers and voted for by the respond- ents. He found approximately a six percent swing toward the endorsed cal’tdidates. 0n the other hand, television seems to have little, if any, effect on the direction of the vote, at least in the aggregate (Simon and Stern, 1955). Michael Robinson's reanalysis of SRC data for 1960-1968 (1976) suggests that reliance on television as a pri- mary news source was weakly related to the direction of the vote for Independent respondents, but not for Republicans or Democrats. J°$1yn's study of the 1976 campaign (1977) indicates that reliance on te18Vision as a campaign information source was very weakly related to VOting in non-partisan elections. Although the relationship held 13 for all categories of partisan identification, it was strongest for independents, and weakest for strong partisan identifiers. Further- more, all of the relationships reported by Joslyn were extremely weak. Perhaps the most newsworthy events of the 1976 presidential campaign were the Carter-Ford debates. A large number of studies of the effects of the debates were conducted.9 While the findings are too numerous to catalogue here, the more important can be briefly di scussed. First, while small increases in viewers' knowledge of the candidates and issues were observed, learning was generally cor- related with viewers' pre-debates levels of information. That is, those with more knowledge about the campaign learned from the debates at a higher rate than those with less knowledge about the campaign. Small increases in viewer interest in the campaign were observed to have resulted from exposure to the debates; in turn, the debates Seemed to stimulate interpersonal communication about the campaign. Viewers were, in general, unable to say who had "won" the debates as a whole; however, viewers perceived Ford to have won the first debate, Carter the second, and neither were seen to have won the 1rl'flal contest. Most studies found that the debates had little, if any, effect upon the outcome of the election, since most respondents to these surveys had made their voting decisions prior to the cam- paign. Moreover, at least one study found that the debates fu"etioned to reinforce viewers' pre-debates vote choices. Finally, 0'11)! a small agenda-setting function was found; the debates had Httle impact upon what were perceived to be the important issues of the campaign, given that the battle lines had been fairly well drawn 14 prior to the initial meeting of the two candidates. The Mass Media: Attitude Change and Reinforcement The effects of television and the other mass media on attitudes and opinions have been the subjects of substantial amounts of research. In general, television and the other mass media have been shown to be rather ineffectual in terms of transmitting mes- sages that change users' attitudes (cf., Klapper, 1960; Katz, 1971). This is not to suggest that television programs never account for altered attitudes; there are isolated examples of the power of television to change opinions. Osgood's study of the effects of tel evising hearings by the House Unamerican Activities Comnittee in the 1950's (1966) is one example. Using semantic differential scales, Osgood measured attitudes on a variety of topics both before and after the hearings. The panel of respondents included those who followed the hearings via television and those who only read about them in the newspapers. Osgood found that the television viewers eXhibited greater attitude change than did the newspaper readers. McCroskey and Prichard (1966-67) conducted an experimental Investigation of the effects on attitudes toward various social pmgrams and policy areas of Lyndon Johnson's State of the Union addr‘ess in 1966. No comparisons between those who watched the addY'ess on television and those who did not were made. However, the authors note that those who did watch exhibited statistically sig- “lficant attitude change in six of twelve attitude areas measured. Michael Robinson's experimental investigation of the effects of the CBS special flu; Selling g_f_t_he_ Pentagon showed mixed results 15 (Robinson, 1976). While viewers' attitudes toward the Pentagon became more negative, and while viewers' sense of efficacy decreased, attitudes toward the "military-industrial complex" showed no ap- preciable change. These instances of attitude change are the exceptions, not the rule; the attitude change effects of the mass media are weak, at best. This is particularly true of media messages designed to change citizens' attitudes and votes during electoral campaigns. There are rather substantial reasons why the media cannot change attitudes. Lang and Lang (1959) suggest some reasons why the mass media in general, and television in particular, are not successful in changing attitudes or partisan choices during election campaigns. First, many citizens have already made up their minds prior to the incep- ti on of a political campaign. Second, Lang and Lang suggest that "voters are not fools," i.e. , they are well aware they are the targets 01’ attempts at conversion. Thus, defense mechanisms, such as selective perception and retention, are activated. Newcomb has suggested that ". . .a recently changed attitude is “1051: likely to persist if one of its behavioral expressions is the Se1¢-3ction of a social environment which one finds supportive of the Changed attitude" (Newcomb, 1971:361). This statement suggests that the media, especially television, can change attitudes, but only if the future environment reinforces, or gives support to, this con- version (cf., Festinger, 1972). In fact, the conmunications litera- ture suggests that this is precisely the case (cf., Katz, 1971). maPIDer emphasizes the importance of "mediating factors" in the e“V1ronment for conversion, and suggests that conversion through the 16 nuadia can occur when (1) these mediating factors are inoperative (such as in the case of new issues); or (2) the mediating factors themselves impel toward attitude change (Klapper, 1960). Patterson and McClure have suggested four conditions which enable television news programs in general, as opposed to televised |3C>l itical propaganda, to succeed in attitude conversion (Patterson and McClure, 1976:88-90): 1. The events which are broadcast are real events, telecast live; 2. The stories or events involve issues which are easily understood. A classic example is a confrontation between police and student demonstrators; 3. The story is broadcast with regular repetition over the long run; or, 4. The airwaves are saturated with the story over the short run . In the light of these conditions, the examples of attitude Change cited above are not quite so anomalous. Both the HUAC hearings and Johnson's State of the Union address were real, live eVents, and at least the hearings had an aura of confrontation about thenL. The hearings occurred over a period of several days; further, "9W8 reports and analyses of the hearings served to saturate tele- VlSion news programs with information regarding the HUAC hearings. A“Waugh the State of the Union address lasted no more than one or t"0 hours, the pre-speech publicity and post-address commentary were at such a level as to allow the judgment that television was ’1. “Id 17 saturated over the short-run with Johnson's speech to the Congress. In summary, attitude change effects of the mass media are rather weak, but the strongest single effect of messages transmitted via the mass media is to reinforce preexisting attitudes, opinions and The process of reinforcement will be more fully considered beliefs. in Chapter 2. At the present juncture, it is sufficient to note that reinforcement is the process of taking actions, such as gather- ing information, which are consistent with extant attitudes or opi nions (cf., Sears and Freedman, 1967; Freedman, et a1., 1970; Sears, 1971). The empirical evidence of the reinforcement effects of the mass media, especially television, is overwhelming; media messages have been shown to serve a reinforcement function in both campaign and "On-campaign contexts (cf., Klapper, 1960; Lazersfeld, et a1., 1968; Sherrod, 1971; Sigel, 1972). For instance, in 1972, 38% of the respondents of one survey indicated that they paid attention to the pr‘esidential campaign through the media in order to remind themselves of their candidate's strong points (Patterson and McClure, 1976:153). That is, over one-third of the sample claimed they engaged in he‘i nforcement. It is no surprise, therefore, to learn that Patterson and McClure found substantial reinforcement effects by television in the 1972 presidential campaign (Patterson and McClure, 1976, esp. 63‘69). O'Keefe's Akron, Ohio panel study also yielded evidence of he? nforcement (O'Keefe, et a1., 1976). Substantial reinforcement effects were also found in the election of 1940 (Lazersfeld, et a1., 1968:103) and, to a lesser extent, in both elections of the 1950's (Merrill and Proctor, 1959; Pool, 1966). In general (Berelson, 1966: 18 494), it appears from most studies that information and knowledge are sought and used more often as rationalization and reinforcer than as data to be used in making what might be called a free decision. Klapper's examination of the effects of the mass media offer sc>nne insights into the functions of the mass media, including tele- V"is;ion, as agents of attitude reinforcement. Any particular Ineeciium is only one of many factors which operate on an audience. Thi 5 holds true, of course, for any context, be it a political campaign or an afternoon of leisurely listening or viewing. There are several mediating factors which operate in such a way as to (2(311‘tribute to a medium's (or media's) role as an agent of reinforce- Tneeritg and against the function of attitude change. These mediating 1r“actors include (1) the existence of predispositions; (2) the psychological processes of selective exposure, perception, and retention; (3) memberships in formal and informal groups, as well as ‘zflfie existence of group norms; (4) the interpersonal dissemination of conmunicated material; (5) the possible existence of opinion I‘Eiiadership; and (6) the existence of ego-involved attitudes, which are especially resistant to change (Klapper, 1960:18-43). These findings lead to the following conclusions: 1. The mass media primarily serve a reinforcement function; 2. In the context of political campaigns, citizens utilize the mass media in order to reinforce pre-existing attitudes and electoral choices. 19 These conclusions suggest that a theory which attempts to explain the information gathering behavior of citizens during elec- tions must do two things. First, such a theory must presume purpose- ful information gathering, and hence, media use, by citizens. That is, the uses and gratifications approach, as opposed to the media effects tradition, would be most useful. Second, the theory must presume that the purpose of citizens in gathering information during campaigns is, in part, to reinforce voting choices. The theory, and the empirical examination of that theory which is presented in this dissertation, meets both of these criteria. In the last section of this chapter, the method of incorporating a rei nforcement assumption, will be systematically explored. @535 Media Research: Some Organizing Structures As noted in the introduction to this chapter, two schools of thought have served to organize and guide scholarly interest in mass media conmunications. The earliest research tradition is called the "media effects" school. The media effects approach implicitly asSumes that individuals who utilize the mass media are passive, and have no particular goals in listening to, or watching, or reading the various media. As the name suggests, the purpose of this a”Inroach was to ascertain the nature and extent of the effects of mass media comnunications on an audience, be it a viewing, listening or reading audience. The major stumbling block faced by this school 01: inquiry was that very few audience effects could be ascribed to the mass media. Much of this early research concentrated on "cam- paign effects," such as persuasion and attitude change. However, the 20 literature reviewed in this chapter documents the proposition that the media are not particularly adept at converting an individual's attitudes or opinions. During the late 1950's and early 1960's an alternate school of thought arose in reaction to the shortcomings of the "media effects" tradition of communications research. This alternative is called the "uses and gratifications" approach to mass comnunications research, and is still dominant today.10 In contrast to the media effects tradition, at the heart of this approach is the assumption of an active, rather than a passive, audience. The uses and grati- fi cations approach assumes that ". . .audience members actively form intentions, create expectations of mass media, and construct lines 01“ action in order to achieve gratifications" (Swanson, 1978:13). Thi 5 school assumes that individuals watch television news programs (Or televised political events), read newspapers, etc., with a sPecific purpose or purposes in mind. The members of the audience are viewed as having reasons for using the mass media. Since these two schools of media research make opposite as- SUmptions regarding the nature and motivation of the audience, the que3tions asked by scholars guided by the two approaches are sub- s1Zarmtially different. With the uses and gratifications approach, one asks, "What does the audience member wish to gain, or achieve, by Using a particular medium, or set of media?" If one's paradigm is the media effects school, one may ask "Has the audience member been changed by using a particular medium, or set of media, and if so ’ in what way?" The former approach focuses on the individual; the ‘atter focuses on the media, and those who control the media. 21 Uses and gratifications research reflects ". . .a desire to understand audience involvement in mass communications in terms more .perspective than the effects faithful to the . . .user's own . . tradition could attain" (Blumler, 1978:1). This desire is shared by the research presented in this dissertation. The theory of citi- zen information gathering that is developed in the next chapter represents an acceptance of the uses and gratifications approach, and an explicit rejection of the media effects school. The theory not only presumes an active audience, but also assumes that citizens use the mass media with a purpose in mind. The underlying assumption, which will be more fully explored in the next chapter, is that during election campaigns, citizens gather information through the mass media in order to help them in making voting decision, and to decrease the uncertainty that their voting choices are correct. In thi s theory, the audience is viewed as an active participant in the Triformation provision/gathering process, and not as a passive body Upon which the media work their magic. The theory presented in the next chapter is based, in part, on the rational choice model of voting behavior first proposed by Downs (1 957), and later modified by others (cf., Tullock, 1957; Riker and ordeshook, 1968; Ferejohn and Fiorina, 1974).” The theory of Ci t‘i zen information gathering combines the assumptions of the rat-‘3 onal choice model with the assumptions of statistical decision theory. Statistical decision theory is concerned with decision makl ng and information gathering under conditions of uncertainty. Thi 8 body of theory is used to incorporate the assumption that citi- z . . . . . ens use information in order to reinforce, or to reduce uncertainty 22 about, their voting decisions. Thus, the theory of citizen information gathering represents a najor modification of the usual rationality framework. This exten- sion is motivated by the following concerns. First, as was stated previously, there exists no theory which adequately explains the use of the mass media in election campaigns. It is my contention that this problem can be dealt with only through a general exami- nation of information search processes, voter uncertainty, and attempts to reduce this uncertainty. Second, presidential campaigns are not, for a majority of citizens, arenas for voting choice. In every presidential election since 1948, over one-half of those voting decided how they were going to vote before the campaign started (Flanigan and Zingale, 1979:171). Of course, it does not follow necessarily that for these It Citizens the campaign serves to reinforce their voting choices. may be the case that these voters simply drop out of political ac:‘t‘ivity from Labor Day until November. However, the most comnonly CT ted source of information in presidential campaigns is television; Farther, the major effect of television (as well as the other mass "red ‘i a) is to reinforce pre-existing choices (or attitudes and opin- 10'33). Given these considerations, it is not unreasonable to suggest the: t for many citizens the campaign does serve a reinforcement fun<:tion. Finally, I would suggest that Downs' conception of the use of 1"formation in elections is too restrictive. Downs states that .to during campaigns "[c]itizens acquire political information. he‘p them decide how to vote. . ." (Downs, 1957:238). More 23 specifically, the rational citizen ". . .is interested only in information which might change his preliminary voting decision. . indicated by his first estimate of his party differential" (Downs, 1957:241).12 Whether a bit of information will change the voting decision is determined by comparing the datum with the estimated party differential. Now, the acquisition and utilization of informa- 'tion entails some expenditures by the citizen.13 Downs suggests that before such expenditures are made the citizen calculates the exgected Laygfj of the bit (or set of bits) of information.14 If the expected pay-off exceeds the cost of the datum, the information is "purchased," otherwise, it is not. As Downs recognizes, there is a serious problem with this formulation (Downs, 1957:243-244). Neither those who are intensely committed to one party (or candidate), nor those who are apathetic have much incentive to gather additional information. Therefore, Few, if any, citizens will change their minds during the campaign; 2. Few, if any, citizens will decide whether and/or how to vote during the campaign; 3. Little, if any, learning of political information will occur during the campaign; 4. ". . .nobody has a very high incentive to acquire political information" (Downs, 1957:244, emphasis added). However, these implications of Downs' analysis of the use of “"‘=<>rmation are inconsistent with the facts. In 1976 switching. c‘ec‘iding, and learning all occurred during the campaign (Miller and 24 Miller, 1977:82-88). In short, the conclusion that "nobody has a very high incentive to gather political information" is not very helpful if one wishes to understand the processes of information gathering by citizens in election campaigns. As noted above, Downs' analysis of the use of information by citizens during elections is based on a rather restrictive assump— tion. Given the assumption that citizens use information only in order to decide how to vote, Downs' conclusions regarding information gathering are inevitable. By building a theory based on an exten- sion of the information gathering assumption, this dissertation vvi'll not only resolve this weakness in models based on rationality assumptions, but will also continue in the uses and gratifications approach which is followed implicitly by Downs. The uncertainty reduction, or reinforcement, assumption is incorporated through the use of statistical decision theory. It is expected that this modification of the standard rational choice assumptions will add substantially to the body of knowledge regarding "C>1t only voting, but also media utilization by citizens in electoral cor. texts. In the chapter that follows, this and other assumptions are made more explicit, as the theory of citizen information gather- ing is formally developed. 1The preferring The The 25 ENDNOTES actual figures were 62.7% preferring television, 22.7% newspapers as sources of information about the election. question asked of respondents was (Roper, 1977:3): First, I'd like to ask you where you get most of your news about what's going on in the world today - from the newspapers or radio or television or magazines or talking to people or where? question asked of respondents was (Roper, 1977:4): If you got conflicting or different reports of the same news story from radio, television, the maga- zines and the newspapers, which of the four versions would you be most inclined to believe - the one on radio or television or magazines or newspapers? 4Survey Research Center data for the presidential elections of 196-0, 1964 5 and 1968. In order to measure information gain, the respondents in this patlel survey were asked a series of questions about the candidates and issues before the campaign began and after it was over. The in- crease in the number of questions answered correctly at the end of .tU‘SE campaign, expressed in percentages, is the measure of information 26 gain used by Patterson and McClure (1976:43—54). 6The Influence 93: Television 95 the 1952 Election. Department of Marketing, Miami University. Oxford, Ohio: 1954. 7It would be a mistake to conclude from these findings that the mass media in general are relatively ineffectual. For instance, the Elmira study of the 1948 presidential election campaign found that overall attention to the mass media was positively related to both interest in the election and an increased accuracy of perception of the candidates' issue positions (Berelson, et a1., 1954:246-250). Becker and Preston (1969) reanalyzed 1964 SRC data, and found that increased media usage (defined in terms of the number of media em- pl oyed) was positively related to concern about the outcome of the election, interest in the campaign, efficacy, turnout, attempts at personal persuasion, and writing letters to public officials. 8The independent variables for this analysis were (1) total free non-network radio and television time, and (2) total per capita Spending for radio and television campaigning. 9The findings presented in the text are taken from a special 15308 of Political Conmunications Review. 3 (1973). edited by Lynda Lee Kaid and Keith R. Sanders. Edited volumes which present research regarding the debates include The Great Debates 1976: £931 35. Carter, Sidney Krauss, editor (Bloomington, Indiana: Indiana 27 University Press, 1979); and The Presidential Debates: Media, Electoral and Policy Perspectives, George F. Bishop, Robert G. Meadow and Marilyn Jackson-Beech, editors (New York: Praeger, 1978). 10A useful overview of this approach can be found in a special issue of Communications Research, Volume 6, Number 1 (January, 1979), edited by David L. Swanson. HThe major contribution of Ferejohn and Fiorina in the develop- ment of rational choice models of voting behavior was to shift attention from the assumption of expected utility maximization, employed by Downs, Tullock and Riker and Ordeshook, to the assumption of minimax regret. The basic advantage of the latter assumption is that it does not stipulate that citizens consider the probabilities of their votes affecting the outcome of an election, only the possi- bi l ity that a vote would affect the outcome. Reactions to this formulation were immediate (cf., Strom, 1975; Stephens, 1975; Mayer and Good, 1975; Beck, 1975; Tullock, 1975). Ferejohn and Fiorina responded to these critiques with an empirical analysis which lent substantial support to the minimax regret model. However, Aldrich (1976) reanalyzed SRC/CPS survey data for the presidential elections 01: 1952-1964 and 1972, an examination which yielded considerable evi- denCe in favor of the expected utility maximizer model. The question “enlains unresolved at this time. 12The economic theory of democracy developed by Downs has a “unlber of concerns, not the least of which is the ability and 28 willingness of citizens to acquire political information. By stating that citizens do so in order to make a voting decision, Downs is explicitly describing information gathering (and hence, media utilization) as a purposeful endeavor. Thus, insofar as economic models of political behavior are concerned with media use, they fall under the general classification of the uses and gratifications approach. 13Downs distinguishes between free and costly information. Free information entails only nontransferable costs, i.e., the costs of absorption and utilization. Costly information entails not only liontransferable costs, but also transferable costs, e.g., the costs (2f gathering, selecting, transmitting, evaluating and analyzing iriformation (Downs, 1957:221-225). 14The expected pay-off of a bit of information is E(x) = inP(xi) vvrieere x = the gain (or loss) in utility resulting from a change in the preliminary voting decision and P(x1.) = the probability the ith change will occur. CHAPTER 2 A THEORY OF CITIZEN INFORMATION GATHERING Introduction In this chapter a theory of citizen information gathering in electoral contexts will be developed. The fully developed theory will provide answers to the following questions, which were raised in the previous chapter: 1) What strategies do citizens follow in gathering information during electoral periods? 2) In general, what is the role of the mass media in these strategies? 3) What are the specific roles of the individual media in these strategies? The theory of citizen information gathering is built upon two basic premises. The first of these is that citizens actively engage 1'1 tflie information gathering/utilization process in electoral cam- pan'ens. The second premise of this research is that individuals gather information not only in order to make electoral choices, but 3150 to reduce the uncertainty about these choices experienced by the ind'ividual. That is, citizens are assumed to have two goals in gath- ering information during electoral campaigns. This presumption of a (“‘31 purpose leads to the argument, which will be more fully devel- Oped below, that the social-psychological concept of reinforcement can 29 30 be incorporated into a rational choice framework. However, while not counterproductive, reinforcement will be shown to be a suboptimal in- ‘formation gathering strategy. The two major premises of the theory--that citizens must be in- <:orporated into any theory which purports to explain the role of the nnass media during elections; and, that citizens use information to make voting decisions and to reduce the uncertainty about those deci- ssions--are reflected in the roots of the theory. The foundation of 1:he theory rests in the basic rational choice theory of voter behav- i or, which was discussed in the previous chapter. Within this general framework a model of citizen information gathering is developed using the assumptions of statistical decision theory. Given the presumption of uncertainty reduction behavior by citizens, the use of statistical decision theory is appropriate and useful, since it is designed to deal with decision making and information gathering under conditions 01" uncertainty.1 A Note on Human Behavior k Before the theory of citizen information gathering is formally developed, some extended corrments regarding the nature of human be- havior assumed by the theory are in order. Specifically, the assump- t'1 on of rationality, and the proposition that citizens gather informa- tl on purposefully, must be examined in detail. The first assumption of the theory is that people are rational. By rational, I mean that individuals have goals, can order these goals 1 n tfirms of their preference for them, and will act so as to realize theSe goals. The specific form of rationality assumed is expected 31 utility maximization, where utility is defined over all possible can- didate platforms. In concrete terms, people are assumed to be in- tendedly rational, in that if they cast a vote, they will do so with reference to the issue positions of the two candidates. The assumption that people are rational in that they maximize expected utility has been severely criticized (cf., Simon, 1961; Eflster, 1979). Perhaps the most damaging evidence against the assump— ‘tion is that, in experimental settings, people do not act so as to nnaximize expected utility, particularly as the choice situations con- 1Fronting them become increasingly more complex (Simon, 1959; Elster, l 979). Herbert Simon, perhaps the foremost critic of ”objective ration- aality," has developed the requirements of rationality as follows ( Simon, 1961:81): 1. Rationality requires a complete knowledge of the conse- quences that will follow on each choice. In fact, knowledge of consequences is always fragmentary. 2. Since these consequences lie in the future, imagination must supply the lack of experienced feeling in attaching values to them. But values can only be imperfectly anticipated. 3. Rationality requires a choice among all possible alternative behaviors. In actual behavior, only a very few of these possible alternatives ever come to mind. Be.yond these requirements, classical maximization also often requires ‘“El‘t:f1er complicated calculations by the individual who attempts to maX'imize expected utility. In short, there are limits to human capa- F’l 1 i ties regarding information gathering and processing; as a result, there are limits to objective rationality (Simon, 1957:196-206). Of the several alternatives to the principle of expected utility max‘iniization (cf., Elster, 1979:133-141), perhaps the most widely 32 accepted is Simon's notion of satisficing, or bounded rationality. The major thrust of the principle of satisficing is that people do not attempt to find the best alternative among those available, but rather choose the first alternative to come along that is "good enough'I to satisfy some set of minimal criteria (Simon, 1957; 1959; 1961). In the face of the empirical and logical forces marshalled against expected utility maximization, the principle has still been the cornerstone of much recent research in the area of voting behavior (cf., Downs, 1957; Tullock, 1967; Riker and Ordeshook, 1968; Aldrich, 1975). However, some scholars have shifted attention to the assump- ‘tion of minimax regret (cf., Ferejohn and Fiorina, 1974; 1975; .Jnldrich, 1976). Assuming that citizens making voting choices are lTfinlmBX regretters, instead of expected utility maximizers, does re- ssolve the problem associated with the calculation of probabilities of eeach alternative event occurring, a problem inherent in the expected Llirility maximization model. However, Simon's charge that an indi- \I‘i<1ua1 cannot be "objectively rational," in that he or she cannot lw all possible courses of action, nor the consequences of these aetions, is unmet regardless of whether one assumes expected utility maximization or minimax regret. Given Simon's criticisms, and given the available evidence, it i 8 clear that the rationality assumption can never be wholly justi- ‘F‘i ed. However, I would suggest that with two additional, more basic assumptions, the problems associated with the concept of rationality are eased somewhat. First, rationality requires knowledge of and a rank-ordering of a] 1 possible alternatives. This does not seem to be much of a 33 problem in presidential contests, where the viable alternatives usually number two, or at the most, three. But assuming rationality i_s problematic, even with only two or three candidates, because it must be recognized that many citizens evaluate the candidates on several criteria. These criteria might include party identification, perceptions of competence, past experience, positions on the issues of the campaign, etc. As the number of dimensions on which a citizen attempts to evaluate the candidates increases, so does the complexity of the task confronting the individual. The citizen must not only identify the criteria relevant to the voting decision, but must also rank the candidates along each dimension, determine the salience of each dimension, and weight each criterion according to its relevance to the voting decision. In short, rank-ordering two candidates is a much more complex problem than it first appears. Of course, such factors as media coverage of the primary and general elections campaigns, and the length of the entire presidential selection process, will ease the burdens of the citizen somewhat. An abundance of information will be available; further, the citizen will have a substantial amount of time to evaluate the candidates. However, just because the time and information necessary to eVa luate the candidates are available, this does not mean that (21' tizens will make such multi-dimensional evaluations. In point of fact, assuming that citizens can establish a weak preference ordering at"Orig two alternatives is non-problematic only if it is assumed that Citizens evaluate the candidates on only a few dimensions, e.g., party affiliation and/or the candidates' positions on one or two issues. Simon's criticisms regarding the difficulty of rank-ordering 34 the alternatives can be met only if it is assumed that citizens evaluate the candidates on only a few dimensions. Assuming that citizens evaluate the candidates on only a few dimensions leads to a further problem. How many dimensions must a citizen use to evaluate the candidates before the decision can be considered to be "rational"? It would seem that all the dimensions relevant to a candidate's performance in the presidency must be used, if the decision is to be considered to be what Simon considers to be "rational." However, if all relevant criteria are used, rank- ordering the candidates again becomes an extremely complex task. It would seem that, in the effort to meet Simon's criticism of this aspect of rationality, something very different from rationality must be assumed. Second, the assumption of rationality can be justified by an zappeal to another basic assumption regarding human behavior: in presidential elections, individuals vote instrumentally. If it is assumed that individuals vote in presidential elections in order to <:<>Iitribute to the (hoped-for) victory of their favorite candidate, then it follows that citizens will make some attempt to determine who t1lide candidates are, what their characteristics, issue positions, etc., are, and which of the contestants is the preferred candidate. That l 53 , if one assumes rationality, one must first assume instrumental vOting. It can also be assumed that citizens vote expressively: when citizens vote, they do not necessarily vote for (or against)7a pm‘12icular candidate. Rather, they vote because of the satisfaction F’f' "standing up and being counted," in order to contribute to the 35 maintenance of democracy, or even to make them feel good about themselves. That part of the motivation for voting is expressive is undeniable (cf., Riker and Ordeshook, 1968). However, the prop- osition that at least part of the motivation for voting in presiden- tial elections is to use voting as an instrument to some goal (i.e., (the election of one's preferred candidate) seems to be equally undeniable. The evidence of issue voting, even the continued im- portance of partisan affiliation as a guide to voting, provide support for the reasonableness of this assertion.2 In sum, it is clear that no theory of human behavior based on the assumption of rationality can completely meet Simon's objections. Appealing to the two more basic assumptions of instrumental voting, and candidate evaluations based on only a few dimensions, lessens the difficulty somewhat. However, as noted earlier, the latter assumption raises an entirely different set of questions about the appropriateness of the rationality assumption. I have used the concept of rationality as the basis for the theory of citizen information gathering for the same reason that (>“ther students of human behavior have utilized the assumption: the ¢==_0 b) where F is a non-negative level of utility akin to Downs' notion of the'party differential threshhold" (Downs, 1957:46). F represents the fixed costs of voting; thus, the difference in utility between the two candidates must be greater than the costs of merely going to the polls and casting a vote. F is assumed to be a known constant for any citi- zen. The reader will note that: 2. (Ua—Ub)-F=(Ua-F)-Ub If we let Ua—F=Ué, the hypothesis test reduces to a simple difference of means test:5 - H : 3. 1 H2: (Ua—U (ug-u ) < 0 b b) 2.0 From the calculation U5 and Ub’ and from the hypothesis test, the citizen can make an inference about the state of the world. The pos- sible inferences (and states of the world) are:’ > Ub + vote for a U b. Ua > Ub’ but U5 E-Ub + abstain c Ub > Ua’ but Ub §_Ua + abstain d. Ub > Ua +~vote for b 51 There is one other logical possibility, and that is that UaEUb. This inference would be based upon the sample observation that UaEDb. But since Ua and Ub are assumed to be continuous distributions, P(DasDb)=O. Thus, I will not consider further the case of exact equality.6 Based on the inference, the citizen makes a tentative voting deci- gigg, For instance, if the citizen finds that U; > Db, the inference would be that U5 > Ub’ and the tentative voting decision would be to vote for candidate A. However, given that the inference is based on sample data, there is some probability that any of the other infer- ences may be correct. The sum of these probabilities is the probabil- i§y_gj_grggg of the hypothesis test, and is denoted by P(E). The shaded portion of Figure 1 represents P(E). In turn, the probability that the tentative voting decision is correct, i.e., maximizes expect- ed utility, is l-P(E). Since any error will lead to an incorrect vot- ing decision, they are all of equal consequence. For instance, if voting for A is the correct decision (i.e., the one that maximizes ex- pected utility), then voting for candidate 8 or abstaining are of equal consequence, since in both cases the correct decision is not made.7 Thus, the citizen will attempt to minimize the total probabil- ity of their occurrence when conducting the hypothesis test. That is, the citizen will attempt to minimize P(E). Three features of the minimization of P(E) criterion merit imme- diate attention. First, if a citizen attempts to minimize P(E) each time an hypothesis test is conducted, the expected loss of a decision based on the hypothesis test will be minimized. Consider the loss table presented in Figure 3. 52 STATE OF THE WORLD U' > U U' < U U' < U U' > U b Ua ;-Ub Ub ;-Ua b a a b b a Ua > Ub O 1 1 1 in Bagfib 1 0 O 1 £3 a b Z a: . U ti." flog-Ua l o o l 55 b a U' > U l l 1 0 Figure 3. Hypothesis Test Loss Table The column entries represent the possible states of the world, and the row entries the possible inferences based upon sample data. For reasons previously discussed, the case of an exact tie is not con- sidered. Each cell entry represents the loss associated with an in- ference given a particular state of the world. In all cases the pro- bability associated with a loss of zero, the loss assigned to a cor- rect decision, is l-P(E). The probability that all other states of the world will occur is the probability of error, P(E). There is a loss associated with an inference only if it leads to an incorrect voting decision. The inferences in the second and third rows and the states of the world in the second and third columns will all lead to a decision to abstain. Therefore, when either one of these inferences is made, while either of these states of the world obtains, there is no loss incurred. All losses have been set equal to each other (the value of unity being purely arbitrary) since the errors are of equal consequence. The expected loss of any inference, or decision about the state 53 of the world, is given by 4. EL(Inference)=O(l-P(E))+1P(E) By A.6 in the limit the distributions of 0a and Oh will have no over- lap (assuming that Ua f Db), since they are degenerate, and thus P(E) will equal 0. Thus in the limit, 5. EL(Inference)=O(l-P(E))+lP(E)=O(1)+1(O)=O Therefore, the expected loss of basing a decision regarding the state of the world upon the hypothesis test is minimized. Further, any action which minimizes expected loss also maximizes expected utility (cf., Winkler, 1972:267-271). Hence, minimizing the expected loss will maximize the expected utility of basing an inference on the hypo- thesis test. The second feature of the minimization of P(E) criterion is that, under certain conditions, minimizing P(E) is a relatively simple task for the decision maker, once point estimates of Ua and Ub have been calculated. Consider the following Theorem. THEOREM: Assume that 02:03, and that the amounts of information regarding candidates A and B are equal. Then when conducting an hypo- thesis test, the probability of error of the test is minimized when the critical level of the test is preset at the point of intersection of the distributions of the estimators. A proof of the theorem is presented in Appendix A. The theorem dictates how the rational citizen can construct the hypothesis test so that the probability of making an incorrect 54 decision is at its lowest possible point. The theorem, as well as the form of the hypothesis test, points out a third feature of the mini- mization of P(E) criterion: unlike classical statistical inference, there is no rule for judging whether an hypothesis test has been “successfully" completed. In classical inference, there are two pos- sible errors, Type I (falsely rejecting the null hypothesis) and Type II (falsely accepting the null hypothesis). Type I error is consider- ed the more serious, so a benchmark level of the probability of Type I error, an a level, is preset. When P (Type I error) falls below this preset level, the hypothesis test is considered to be "successfully" completed, and an inference is made. However, for the statistical decision maker all errors are of consequence, so no one error can be singled out, and no benchmark for completing the hypothesis test based on only one type of error can be established. Further, presetting a ”desired" level of P(E) is incor- rect. This is so because the statistical decision maker must take some action based on the inference from the sample data. This immedi- ately implies that the rational citizen ggg_statistical decision maker not only calculates the expected utility of the action itself, but must also calculate the benefits and costs of gathering information necessary for making the decision. An action is rational only if its expected benefit exceeds the costs of making the decision to take the action. Since considering P(E) by itself does not consider the costs of decision making, a preset level of P(E) cannot serve as a criterion for completing an hypothesis test. However, considering the probability of error in conjunction with the costs and benefits of information will yield a standard for 55 completing an hypothesis test, as well as for gathering information. That is, statistical decision theory enables the rational citizen to make both his decisions, that of gathering information and that of deciding how to vote, from the results of one decision making process. Assume that the citizen has gathered information about the candi- dates in the form of Ia and Ib from any information source or sources, Si‘ On the basis of this information, he constructs the hypothesis test, and determines P(E). The citizen must now decide whether to take additional samples of information from any information source. If the decision is to collect more data, the decision based on the hypothesis test is tentative, and the entire process is repeated. If the decision is to collect no more information, the citizen makes a terminal voting decision based on the hypothesis test. That is, the citizen decides to either vote for A, vote for B, or abstain. Thus, the information gathering process also yields a voting decision. This decision is based upon the utility the citizen would gain jf_g§_were to take another sampling of information. In decision theo- retic terms, the citizen calculates the expected net gain from sampl- ing (ENGS) from any information source,i, as follows: 6. ENGSi=EVSIi-C1 The unit costs of sampling are given by the cost functions defined in A.7. The assessment of EVSIi, the expected value of sample informa- tion, is less straightforward, for it involves the Bayesian concepts of prior and posterior expected utility. Essentially, the expected value of sample information is the gain in expected utility (relative to decisions made about the state of the world, given sample 56 information) the citizen would realize if he were to take another sample. If ENGSi > O, for any $1, the citizen will gather more infor- mation, calculate new point estimates of Da and Db, conduct another hypothesis test, calculate P(E), and calculate ENGSi once again. If the initial value of ENGSi is less than or equal to zero, the citizen will gather no more information, and will make a terminal voting deci- sion. Extended discussion of Bayes' Theorem can be found in any intro— ductory decision theory text (cf., Winkler, 1972). In Appendix B, I have presented the components of Bayes' Theorem as they relate to the model. In Appendix B the following concepts are defined and ex- plained: --the states of the world, 0; --the prior probability of the states of the world, P(e); --the posterior probability of the states of the world, P(OIY); --the prior expected utility of any inference or decision, EU(Di); --the posterior expected utility of any inference or decision, EU'(D1.); --a prior optimal inference, or decision, 0*; and, --a posterior optimal inference, or decision, D**. Based on the discussion in Appendix B, the expected value of sample information from any source, Si, is equal to the difference be- tween the posterior expected utilities of the optimal decision before and after the sample is taken: 7. EVSIi=EU'(D**)-EU'(D*) 57 However, this formulation of EVSIi is not very useful in determining whether another sample should be taken. This is so because it can be calculated only after the sample has been taken. Thus the citizen cannot determine whether drawing the sample was optimal until after the sample has been drawn. Fortunately, it turns out that, ggfggg Egg sample has been drawn, the expected value of sample information can be . . .8 determined solely from the prior expected utilities of Di' 8. EVSIi=g[U(Dilej)-U(D*)]P(ej)P(y|ej) For each possible state of the world, determine the decision with the highest utility, and subtract from this utility the utility of the present optimal act. This difference is multiplied by the prior pro- bability of e and the likelihood. The sum is taken across the states of the world, and the result is the expected value of sample infor- mation (Winkler, 1972:306-325). With EVSIi calculable in terms of the prior expected utilities, the determination of ENGSi follows immediately, through application of Equation 6. Further, if the sample size is not fixed, the citizen can maximize the expected utility from sampling by determining the optimal sample size, N*, by finding the value of N which maximizes 9. ENGSixN=EVSIixN-CixN The expected value of sample information is an increasing, but marginally decreasing, function over the amount of information gathered. This reflects the fact that as the total amount of information in- creases, the value of any one sample decreases. A hypothetical EVSI curve is depicted in Figure 4. 58 A=Evs11. Amount of Information Figure 4. Hypothetical EVSI Curve Previously, the information sources and their curvilinear cost functions were defined. It is now possible to define the benefit of information from any information source, Si’ as EVSIi. Thus the bene- fit of information is monotone increasing but marginally decreasing. Define the unique benefit curves associated with each information source as Bi' Information sources differ in terms of unit costs due, in large part, to their relative levels of complexity. If one were able to define (and reach agreement upon the definition of) a "unit of information," then more complex information would contain more information relevant to the voting decision per unit of information than would less complex information. In turn, this implies that the shapes of the benefit curves for the various information categories are different. Specifically, the more complex the information derived from an information source, the faster will the benefit curve rise and then level off. The less complex the information derived from an 59 information source, the flatter will be the benefit curve, and the slower it will level off. An hypothetical representation of the cost and benefit functions for a particular information source is presented in Figure 5. The following decision rules for gathering information from in- formation sources can be derived directly from Equation 9:9 3(31.) d(c1.) d2(B,) d2(c1.) 1. When dn - dn , and when d2” < -—d§h_— , the return from that information source is maximized. Any further in- formation, if it is to be collected, should be collected from some Sj, jfi, where d C. d C. (J)< (1) dn dn 2 2 d(Bi) d(Ci) d (Bi) d (Ci) 2 When dn = dn and 2 < 2 for all d n d n Si’ the return from all information sources has been maximized. No additional information should be gathered. The rational citizen will now make a terminal voting decision. These decision rules state that when the marginal benefit of any information source equals its marginal cost, the return from that in- formation source is maximized, and a cheaper source of information should be used (if additional information is to be gathered). If no cheaper information sources exist, the citizen should gather no more information, and should make a terminal voting decision. 60 em sow m>s=o uwwmcmm use “moo .m weaned :owumEcoecm eo acaoe< Awov mumoo A_~m>mv meeceeem fie~m>m\eoc moeeeeem\memeo 61 The foregoing analysis directly implies that the citizen uses the information sources available to him/her in a sequential fashion.10 That is, expensive information sources (e.g., government publication) will be used before less expensive information sources (e.g., news- papers). In turn, the latter sources will be used before such inex- pensive information sources as radio or television. The reason that citizens will use information sources in a se- quential fashion, i.e., start with more expensive sources and move to less expensive sources, is relatively straightforward. The marginal costs of expensive information sources are, of course, higher than for less expensive sources. This is because the costs of using an infor- mation source are greater the more complex it is. By more complex, I mean, and have noted previously, having more information relevant to the voting decision per unit of infomration than less complex informa- tion sources. The cost functions for two information sources, one more complex than the other, are related as follows: 10. Ci > Cj, i E J P(E), the probability of error in making an inference from the hypothesis test about the relative utility of the election of the two candidates, is a function of information. EVSIi, the expected value of sample information, is defined as the gain in expected utility the citizen would realize if he/she were to take another sample of infor- mation. Since expected utility is a function of probabilities, one of which is P(E), then EVSIi is a function of P(E), and hence of informa- tion. Thus, the more information gained by sampling from a particular 62 information source, the higher will be EVSIi. Further, EVSI, is the benefit to be derived from a particular source of information about the candidates. Therefore, the costlier is an information source, the higher will be the benefit to be derived from its use. That is, 11. If C. >C. 1 J lfJ' Then 8. > B. 1 J Given that the marginal benefit (Bi) of more expensive informa- tion sources is greater than the marginal benefit of less expensive information sources, the citizen will be motivated to examine more ex- pensive information sources for possible use in gathering information about the two candidates before examining other, less expensive, in- formation sources. 1 However, this analysis does not imply that all (or even any) citizens will rely upon the more expensive information sources in gathering information. This is so because, in order for any informa- tion source to be utilized, the following conditions must be met: d(Bi) d(Ci) 12. d" " T > 0 d(B.) d(C ) d(B ) d(C ) ___J___ l 1, _ l_ - - 13' dn dn > dn dn 1 7‘ J These conditions state that the difference between the marginal benefit and marginal cost for any information source must not only ex- ceed 0, but must also exceed the difference between the marginal costs and benefits of all other information sources. For many citizens, 63 these conditions will never hold for expensive information sources; thus, they will never utilize these sources. In turn, if the general condition stated in Equation 12 is never met, no information will be gathered after the initial calculation of the cost/benefit curves, and the citizen will proceed directly to a terminal voting decision. Fi- nally, it is likely that if only one information source category is utilized, it will be inexpensive information sources, such as radio or television. However, this is a supposition based on the relative costs and benefits of Si’ and cannot be derived directly from the theory. Such a derivation would be possible only if the exact shapes of the cost and benefit curves were known. Further extensions of this analysis can be introduced by examin- ing the effects of various socio-political factors on the cost and benefit curves. These factors would include such citizen characteris- tics as strength of partisan identification, level of political ideol- 11 The listing of factors need not be ogy, education, income, etc. exhaustive, nor is it necessary to consider the effects of the factors individually. I shall assume that these socio-political factors, by either providing a framework for the absorption and utilization of in- formation, or by providing a relatively greater amount of assets which can be expended for information, will reduce the comparative costs of information. That is, the slopes of the cost functions for those in- dividuals with high education, income, etc., will be lower than the slopes of the cost functions for those individuals with lower levels of education, income, etc. This reduction in comparative costs will lead to a greater propensity to gather more information in general, and to utilize more expensive information sources, in particular, 64 ceteris paribus. The theory presented herein posits that the rational citizen acts as if he were a statistical decision maker. Within the framework of the theory, depicted informally in Figure 2, the citizen is able to make decisions about information gathering and voting based on the calculations derived from statistical decision theory. In essence, one set of calculations leads to a framework for making both types of decisions. The next several chapters will be concerned with an empirical examination of the theory. The assumption that media sources are un- biased will be tested. A variety of hypotheses will be derived from the model and will be empirically examined. A simultaneous-equation model of the theory, incorporating the assumption and hypotheses from the theory, will also be tested. To conclude this chapter, I will con- sider the concept of reinforcement and its compatibility with the theory of citizen information gathering. Reinforcement and Rational Information Gatherigg Any theory worthy of the name should not only increase the store of knowledge regarding a particular problem area, but must also be able to account for established findings in that area. One of the more pervasive concepts in the study of politics, especially in mat- ters regarding voting and information gathering, is the social-psycho- logical concept of reinforcement. This discussion anticipates the last chapter, wherein the applicability and utility of the theory of citizen information gathering will be evaluated. I wish to introduce the idea of reinforcement here not only because of its widespread 65 relevance and acceptance, but also because the process of information gathering described by the theory will result in observable behavior similar to that predicted by the reinforcement hypothesis. Reinforcement refers to the process of taking actions (e.g., gathering information) which are consistent with extant attitudes and/ or opinions (cf., Lajonc, 1960; Sears and Freedman, 1967; Freedman, et a1., 1971; Newcomb, 1971; Sears, 1971). The concept of reinforce- ment assumes that thoughts, beliefs, attitudes and behavior are ar- ranged in fairly sensible and orderly ways because such an arrangement is psychologically comfortable.12 Dissonance occurs when two cogni- tions of the same political object clash with one another. For in- stance, if a life-long Democrat learns that a Republican President favors wage-price controls, dissonance might result. Since such a state engenders discomfort, conscious and unconscious efforts are made to reduce the dissonance. . There are several methods of reducing dissonance (cf., Freedman, et a1., 1970:347-387); I am concerned only with those having to do with the acquisition of information. A model of cognitive dissonance would suggest that individuals would avoid information which increases dissonance, and would seek out information which decreases dissonance. More specifically, the defense mechanisms of selective exposure, per- ception, and retention will operate in situations which might gener- ate dissonance. That is, one or more of the following will occur: 1. Information, regardless of its content, will be perceived as being consistent with extant attitudes (selective perception); 66 2. An individual will voluntarily expose himself to communica- tions which are consonant with his attitude structure, and will avoid communications which are dissonant (selective ex- posure); 3. An individual will remember only that information which is consistent with his attitudes (selective retention). These phenomena are all characteristics of the process of reinforce- ment. The evidence demonstrating the existence of selective perception was discussed in Chapter 1. As noted in that discussion, this litera- ture leads to the following conclusions. First, the major effect of mass media utilization is reinforcement of attitudes and opinions. Second, in the context of political campaigns, many citizens will en- gage in activities which serve to reinforce attitudes, opinions, and/ or voting decisions. Finally, in the context of political campaigns, many citizens will utilize the various media in order to reinforce pre-existing attitudes, opinions, and/or partisan choices. That is, many citizens will engage in purposeful reinforcement. It is clear, therefore, that electoral decision making, like all decision making, consists of two stages: the actual choice, and the reinforcement of that choice. Further, it is fair to suggest that theories of electoral behavior which deal only with the choice stage of electoral decision making are incomplete. Traditional rational choice models, which have generally ignored individual behavior after an electoral choice has been made, are therefore incomplete in their treatment of electoral decision making. This is not to suggest that rational choice models of voting are useless; in point of fact, the 67 theory of citizen information gathering developed herein explicitly incorporates a simple rational model of electoral choice. The theory differs from previous economic-based models of voting in that it ex- tends the analysis to include the process of reinforcement, evidence of which, in electoral contexts, is substantial. Let us then consider the manner in which reinforcement is incor- porated into the theory of citizen information gathering. Recall that citizens in the model are presumed to gather information in order not only to arrive at electoral choices, but also to reduce the uncertain- ty, or doubt, about those choices. Electoral choices are assumed to be made on the basis of maximizing expected utility, with utility being defined over the space of all possible candidate platforms. In turn, this utility is a function of the candidates' issue posi- tions, experience, personal characteristics, etc.; the individual's own attitudes, opinions, etc., operate as the standards against which candidate attributes are judged. Correct decisions were defined as those decisions which maximize expected utility. Operationally, a correct decision is voting for that candidate "closer" to the citizen in terms of these standards. If the difference between the two can- didates is very small, i.e., they are equally ''close" or “far" from the citizen, the correct decision may be to abstain. Thus, the citizen's goal (in the model) to reduce uncertainty, or doubt, regarding electoral decisions is essentially an attempt by the citizen to increase, as much as is economically feasible, the subjec- tive probability that his preferred candidate is closer to him in terms of attitudes, policy positions, etc. Now the primary goal of reinforcement is the reduction of cognitive dissonance. In electoral 68 contexts, the purpose of reinforcement is to increase the citizen's perception that his choice is correct, i.e., closer to the citizen in terms of attitudes, policy positions, etc. In presuming that citizens gather information in order to reduce doubt, the theory explicitly in- corporates the process of reinforcement. Furthermore, the theoretical structure of statistical decision making was chosen not because it is a mathematical structure, but because it is specifically designed to deal with problems of decision making and information gathering under conditions of uncertainty, i.e., incomplete information. Further, statistical decision theory facilitates the incorporation of choice and reinforcement under one theoretical structure. Thus, the theory developed herein is not a theory of rational choice, per se, but a theory of rational decision making, incorporat- ing both choice and reinforcement. To further elaborate, let me con— sider the concept of reinforcement explicitly in terms of the theory. Recall that in building the theory, I have not assumed that in- formation gathering necessarily begins, or even occurs, during a cam- paign. It may be the case that some citizens will decide very early in the campaign, or even before the campaign, for whom they will vote, and likewise will have decided to stop gathering information. It is to these early deciders that reinforcement is particularly, al- though not exclusively, applicable. In terms of the model, after making a tentative voting decision the citizen will recalculate Ua and Db, and find that the tentative voting decision has not changed. We would then expect to find that the voting decision is "stable" through- out the campaign, just as the social-psychological hypothesis of rein- forcement would predict. Further, reinforcement implies that an 69 individual will gather information only about that candidate toward whom he is favorably disposed, or predisposed. This is also consis- tent with the theory, in that citizens are presumed to gather informa- tion about either or both candidates. While gathering information only about one candidate is consistent with the theory, it is stipu- lated that information sources one uses for the purpose of reinforce- ment do not lie about the other candidate. This information gathering strategy may be deemed productive, in terms of the model, in that gathering information about one candidate, say candidate A, will re- duce the variance of the estimator of expected utility, Da, and may or may not reduce the variance of Db. This, in turn, will lead to a re- duction of P(E), since the overlap of the distributions of Ua and Db will decrease. However, such a strategy is not optimal. Consider the following proposition: Proposition 1: Assume that a: = 05’ that U; > Db, and that the amounts of information previously gathered about each candi- date are equal. Let N denote the amount of information to be gathered about the two candidates. P(E) will be lower in the case where N/2 bits of information are collected about each can- didate than in the case where N bits of information are collected only about the most preferred candidate, A.‘3 The proposition states that, given equal variances and amounts of information regarding the candidates prior to a round of sampling, gathering equal amounts of information about both candidates will re- duce P(E) to a greater extent than will gathering information about 70 the more preferred candidate only. In reference to Figure 1, this means that the overlap of the distributions of Ga and U will be b smaller with an equal split of information than with a concentration of information on one candidate. Gathering information about both candidates, then, will yield a greater reduction of uncertainty than will gathering information only about the more preferred candidate. This result has immediate relevance to the process of reinforce- ment. Reinforcement can be defined operationally as gathering infor- mation only about the more preferred candidate. The purpose of rein- forcement, as noted above, is to increase the perception that the choice is correct; that is, the purpose is to reduce uncertainty. And yet, if only a limited amount of information is to be gathered, col- lecting equal amounts of information about both candidates will per- form the function of reinforcement, reducing uncertainty, more effec- tively than will reinforcement itself. Therefore, reinforcement is not a rational activity, in terms of information gathering, since it does not maximize the return from gathering information. That is, ENGSi, the expected net gain from sampling, is not maximized. However, reinforcement will lead to a re- duction of P(E) (uncertainty), and therefore cannot be judged a coun- terproductive information gathering strategy. The situation changes somewhat if it is no longer assumed that the variances of the estimators of utility are equal. If unequal var- iances were to be observed, it is likely that the variance of the es- timator of expected utility for the more preferred candidate would be the lower of the two. For instance, if candidate A were preferred to candidate 8, i.e., if U5 > Ub’ then it would likely be the case that 71 2 ' 2 . . . . 2 2 a < 0b. Of course, it 15 pOSSible for Ua > Ub’ and 0a > Ob' What- ever the case may be, it is possible to delineate the best strategies 0 for gathering additional information about the two candidates. Con- sider the following proposition. Proposition 2: Assume 0: f 0:. Let X denote the act of collecting a bit of information about the candidate for whom the variance of the estimator of expected utility is larger. Let Y denote the act of collecting a bit of information about the candidate for whom the variance of the estimators of expected utility is smaller. Let P(X) and P(Y) denote the probabilities of these two acts, respectively. Further, P(Y) = 1-P(X). The following statements can be made: 1. The expected net gain from sampling (ENGSi) is maximized when P(X)=1. 2. ENGSi is minimized when P(Y)=1 (or, when P(X)=O). 3. ENGS, declines monotonically from P(X)=1 to P(X)=O (or, when P(Y)=1). Of these three statements, the second is the most consistent with the process of reinforcement (assuming that the estimator of expected utility for the more preferred candidate has the smaller variance). Therefore, if the variance of the estimator of expected utility is smaller for the more preferred candidate, as is likely to be the case, reinforcement is the least desirable information gathering strategy, in terms of uncertainty reduction. However, while suboptimal (even maximally suboptimal), reinforcement will still result in a reduction of P(E), or uncertainty, and therefore cannot be judged to be 72 counterproductive to the goal of uncertainty reduction. One characteristic of reinforcement would be the existence of stable voting decisions. That is, once an individual comes to a vot- ing decision, and attempts to reinforce (or reduce the uncertainty about) that decision through information gathering, it is likely that the individual's voting decision will not change. This result is also consistent with the theory of citizen information gathering. Consider the third and final proposition: Proposition 3: Assume that o: < 0%, and U; > 66‘ Gathering information about candidate A only (i.e., reinforcing) results in a decrease in oi, and would not affect Ué (by A.5 and A.6). 2 Since, if oa decreases, P(E) decreases, and since U; and US are not affected by additional information, gathering new information about one candidate only will not lead to a change in the voting decision. In summary, reinforcement as an information gathering strategy is compatible with the theory of citizen information gathering devel- oped in this chapter. More specifically, the purpose of reinforce- ment, the reduction of uncertainty, is also a goal of citizen in the model. Further, the actual process of reinforcement--in this case, gathering information only about the more preferred candidate--has been shown to reduce the probability of error, or uncertainty, al- though not as effectively as other strategies. Finally, stable voting decisions, which would be expected of citizens engaging in reinforce- ment can, under certain conditions, also be predicted from the theory of citizen information gathering. 73 ENDNOTES 1For a good introductory treatment of statistical decision theory, see Winkler, 1972, especially Chapters 5-7. 2For some recent evidence regarding the impact of issues and/or party identification on voting, see, among others, Kessel, 1972; Kelley and Mirer, 1973; Aldrich, 1975; Miller, et a1., 1976; Frolich, et a1., 1978; Markus and Converse, 1979; Page and Jones, 1979; and Carmines and Stimson, 1980. 3For more on the strategic use of ambiguity by candidates, see Shepsle, 1972, and Page, 1976. 4Along these lines, see Tollison and Willet, 1973. Ua and Db will be used instead of E(Ua) and E(Ub), solely for the purpose of simplicity. Similarly, Ua and Ub will be used in place of E(Ua) and E(Ub), respectively. 5In conducting this test, the citizen will calculate Oa-F=Ué. Since this involves a linear transformation of a normal distribution, D; will be normally distributed, with a variance equal to that of Da. In other words, the Ua distribution in Figure 1 will be shifted 74 slightly to the left, when F > 0. If F=O, a most unlikely occurrence, the Us distribution will not move. 6The statement that P(UasUb)=O is based on the fact that, if Ga and Db are assumed to be continuous, P(Dazx)=0, and P(Dbzx)=0, where x is any exact value of expected value. Thus, P(DaEUb)=0. This, of course, is theory. In reality, one might observe Danb. However, such a result would depend upon an individual perceiving gg_ difference between the two candidates. This would occur, for example, if the citizen were totally ignorant of the characteristics, partisan affiliation, issue positions, experience, etc., of both candidates. This state of "bliss" is the point at which the theory becomes rele- vant, and the point at which the theory asserts information gathering will commence. Another situation in which one might observe Ua=Ub would be when the two candidates actually are identical, political clones, as it were. Determination of this state of affairs would re- quire perfect information, or something very close to it, an assump- tion conspicuous by its absence from the theory. Finally, one might observe Ua=Db if the two candidates were not identical to one another, but were equidistant from the citizen, one on the "left" and one on the “right" in the issue space. Again, for a citizen to determine that the two candidates were exactly equidistant from him would require nearly perfect information. 7I have assumed that all errors are of equal consequence strictly for mathematical convenience. This assumption allows the loss associ- ated with any error to be set equal to the loss of all other errors. 75 It is probably more realistic to assume that voting for the “wrong" candidate is more serious than erroneously abstaining. While both erroneously voting and erroneously abstaining will deprive the actual preferred candidate of a vote, only voting in error will give an extra vote to the less-preferred candidate. Whether errors are of equal consequence is irrelevant to the fol- lowing discussion. The results associated with Equations 4 and 5 will not change with an alternative assumption. Given that losses associ- ated with errors are always multiplied by P(E), and since P(E)+O in the limit, then the expected loss of basing a vote decision on an hy- pothesis test will always equal 0. 8This is so because the expected posterior expected utility of any decision is equal to its prior expected utility, before the sample is taken (cf., Winkler, 1972:311). 9The decision rules follow directly from a profit maximization exercise, which involves taking the first and second derivatives, with respect to n (the amount of information bits gathered), of Equation 9. Substituting Bi for EVSI(Si), the first derivative is d(EVSI(Si)*n) d(Bi) d(Ci) dn dn dn Setting the first derivative equal to zero, a candidate for the point at which profit from sampling from any source, Si’ is at a maximum is where For this point to be a maximum, the second derivative must be nega- tive. Since Bi has been defined as marginally decreasing, and since Ci has been defined as marginally increasing, 8.) d2(C1.) <0, >0 dzn d2n Thus, the second derivative can be negative only when d2(B1.) d2(C1.) dzn dzn That is, the benefit per unit of information must be decreasing at a faster rate than the unit cost of information is increasing. When this condition holds, profit from sampling from any information source is maximized when the unit costs of information are equal to the unit benefits. 10In Chapter 3, a method of categorizing information sources ac- cording to the costs associated with their use is developed. Informa- tion sources are categorized by that method as follows: Expensive Sources--candidates' campaign appearances, interest group publications and government publications; Moderately Expensive Sources--candidates' publications, party publications, newspapers and magazines; Inexpen- sive Sources--radio, television and interpersonal communications. 77 HOther factors would include overall socio-economic status, level of political activity, and level of political interest. 12Reinforcement also assumes that values have already been formed. If attitudes are in the process of formation, dissonance, pg: 5;, has little meaning (Carter, et a1., 1972). 13The validity of this proposition stems from the fact that the benefits of information are monotone increasing but marginally de- creasing. Casting this in terms of U, the variance of U will decrease with each additional bit of information, but each additional datum will reduce the variance less than the previous bit. Thus, in split- ting N bits of information between two estimators whose variances and sample sizes are identical prior to the sample of information, the net reduction in both variances, and hence the net reduction in the overlap of the two distributions, will be greater than if all N bits of information were concentrated on only one estimator. This is so because the marginal decrease in the reduction of the overlap will be lower, the smaller the number of additional bits of information. The validity of Proposition 2, below, can be similarly demonstrated along this same line of argument. CHAPTER 3 OPERATIONALIZATIONS AND EXAMINATION OF ASSUMPTIONS Introduction In this chapter two objectives will be pursued. First, a discus- sion of the data sources to be used in the empirical examination of the theory of citizen information gathering will be presented. This will be followed by a brief description of the operationalization pro- cedures employed in testing the model. Second, two basic assumptions of the theory will be empirically examined in order to determine their plausibility. The first of these assumptions states that there is no systematic distortion of informa- tion by the media; i.e., the media are unbiased. The second assump- tion states that various socio-economic-political characteristics of citizens serve to reduce the relative costs of information. The Data and Operationalizations As noted earlier, a simultaneous-equations model of the theory of citizen information gathering will be developed and empirically exam- ined in Chapters 6-9. This evaluation of the model will be carried out with data from the Center for Political Studies American National Election Studies for the elections of 1972, 1974, and 1976. Toward that end, the two basic assumptions of the theory to be examined will 78 79 be tested using data from all three years. The focal point of the discussion will be the 1976 data; the tests of the assumptions for 1974 and 1976 are presented in Appendix C. The Center for Political Studies 1972, 1974, and 1976 surveys constitute a panel study of the electorate, and while the data points are too few in number to be considered an econometric time series, the panel data offer several advantages to this analysis: 1. The model can be examined in different presidential electoral contexts. The 1976 campaign resulted in a narrow Democratic victory, while the 1972 race culminated in a Republican land— slide. The generality of the theory can be explored by em- pirically testing the statistical model under these very dif- ferent electoral scenarios. 2. The model can be examined in both presidential and congres- sional electoral contexts. There are many differences be- tween presidential and off-year elections (e.g., turnout levels, relative reliance on party identification gj§_g,yi§_ issues). Again, examination of the model to determine its applicability to these divergent electoral contexts will be useful in investigating the general applicability of the theory. 3. The possibility of over-time changes in citizens' information gathering strategies can be examined. There is at least one major disadvantage to using the CPS elec- tion surveys. The surveys for all three years are cross-sectional, single wave surveys. However, the relationships between levels of 80 political information, media use and uncertainty are dynamic, not static. For instance, level of media use at one time period is hypo- thesized to be a function of level of political information in the previous time period, i.e., Y2,t = f(Yl,t-l)' In the examination of hypotheses presented in Chapter 5, as well as the model evaluation in Chapters 7-9, these dynamic aspects of the model are lost, since mea- surement of the relevant concepts was taken at only one point in time for each of the three election year surveys. Before I consider the assumptions and hypotheses of the theory, I will describe how the major concepts of the theory have been opera- tionalized. The CPS Election Studies are, in some respects, well- suited to this analysis, in that many of the relevant concepts are either directly measured, or can be operationalized through simple manipulation of variables. In other respects, however, these data are woefully inadequate, in that some concepts are neither directly nor indirectly measured. A brief discussion of some of the more difficult operationalization procedures follows. The citizen's candidate differential benefit is not measured by the CPS surveys; however, variables can be manipulated in a variety of ways to tap this concept. One of these manipulations involves the seven-point issue scales asked of respondents in all three election year surveys.1 The other involves the respondents' placements of the presidential candidates on the one hundred-point feeling thermometers. The following formulation will constitute one measure of differ- ential benefit: 81 X. = the respondent's self-placement on the iEn-issue scale the respondent's placement of the Republican candidate on the iED-issue scale eDi = the respondent's placement of the Democratic candidate on the iED-issue scale N = the number of issue scales the respondent placed him/her- self and both candidates In words, absolute differences between the respondent's placement of the Republican and Democratic candidates on each issue scale and the respondent's self—placement on the same scale are taken. The differ— ence between these two calculations is taken, and the sum in taken across all issue scales. This score is then standardized by dividing by the total number of issue scales on which the respondent placed him/herself and both candidates. The variable measures both the strength and the direction of candidate differential benefit. A score of zero would indicate that, for the issue scales taken as a whole, the respondent sees no difference between the Republican and Democratic candidates. A score greater than zero indicates a prefer- ence for the Democratic candidate, while a score less than zero indi- cates a preference for the Republican candidate.2 By making the ex- pression additive, I have assumed, for purposes of measurement, that utility is a linear function of proximity of issue positions, and that all issues are of equal importance to the respondent.3 Another method of measuring utility involves the feeling thermom- eters. The measure used here is simply the difference between the respondent's placements of the Republican and Democratic candidates on 82 the thermometer. The placement of the Democratic candidate is sub— tracted from the placement of the Republican candidate. A score greater than zero indicates support for the Republican candidate, while a score less than zero indicates support for the Democratic candidate. As before, a score of zero indicates no preference be- tween the two candidates.4 Another concept of considerable importance is the probability of error of hypothesis testing. Obviously, it is impossible to measure the concept directly. However, the task of operationalizing the con- cept of uncertainty seems less difficult when one considers that the higher is P(E), the more "uncertain" a citizen is about the voting decision. Conversely, as P(E) decreases, the more sure, or "cer- tain," is the citizen about his or her voting choice. Thus, I shall use the term “uncertainty" to denote P(E), while recognizing that this use of the term is a departure from standard practice.5 Defining the probability of error of hypothesis testing as uncer- tainty makes operationalization possible, again through manipulation of the seven-point issue scales. To operationalize uncertainty, I have assumed that, in the aggregate, perceptions of candidates' issue positions are accurate. In turn, this implies that the average res- pondent placement of a candidate on a particular issue scale is the true candidate position. The extent to which an individual respondent places a candidate near the average placement on an issue scale, the more certain is the respondent regarding the candidate's stand on that particular issue. Given this assumption, the following formulation has been chosen to operationalize uncertainty: 83 N .— z 'x .-X. .l 2. i=1 0' 391 N where 76, = the average placement of the sample of candidate 9 on the iED-issue scale X _ th .th d . 1 f . .th jei - e j-—-respon ent s p acement 0 candidate 0 on the i-—- issue scale N = the total number of issue scales on which the respondent placed candidate 9 Taking the sum across all issues, and dividing by the number of valid issue scales, yields a measure of the respondent's uncertainty regard- ing a candidate's position on all issues.6 The higher this figure, the greater is the uncertainty of the citizen regarding the candi- date's issue positions and, therefore, the greater the uncertainty re- garding the respondent's utility for the election of that candidate. This procedure yields two uncertainty measures for each data set. For 1972, the measures are uncertainty regarding Nixon's issue posi- tions, and uncertainty regarding McGovern's issue positions (UNSURNIX and UNSURGOV, respectively). For 1976, uncertainty regarding Carter's and Ford's issue positions is measured with two operationalizations, UNSURCAR and UNSURFOR, respectively. For 1974, the two uncertainty measures deal with the issue stands of the Democratic and Republican parties; these measures are called UNCERDEM and UNCERREP. Beyond these individual measures, it will also be useful to have a measure of total uncertainty. One such measure for each of the 84 three data sets was constructed by adding the two candidate-referenced measures together. For example, the measure of total uncertainty for the 1972 survey was constructed by summing UNSURNIX and UNSURGOV. The same procedure was followed for 1976, using UNSURCAR and UNSURFOR. For 1974, UNCERDEM and UNCERREP were added together. For the three years, the variables are called UNSUR72, UNSUR74, and UNSUR76. Underlying the concept of media utilization is the notion of the costs of using various information sources. While it is not possible to directly measure the costs of media use, at least with the data used in this analysis, it is possible to categorize the various infor- mation sources available to citizens along several cost dimensions. For simplicity, I will limit the analysis to three information source categories: expensive, moderately expensive, and inexpensive. To arrive at these categories, I and three ofmy associates (who were only marginally informed about this research) rank-ordered ten sources of electoral information on three separate dimensions of cost. The ten information sources are: attendance at campaign appearances of a candidate; interest group publications; government publications; candidate publications; party publications; newspapers; magazines; television; radio; and, interpersonal communications. The three cost dimensions on which these sources were rank- ordered were taken from Downs (1957:208~212). They are procurement of information from the source; absorption of information; and, analysis and evaluation of information. A typology was constructed from the average rankings of the information sources on the three dimensions by each judge. This procedure, while somewhat gg_flgg, yielded a typology which not only reflected great similarity across judges, but which was 85 also intuitively pleasing. The categorization of the information sources is given in Figure 6. Expensive Moderately Inexpensive Expensive Candidate's Campaign Candidate Television Appearances Publications Interest Group Publications Newspapers Radio Government Publications Magazines Interpersonal Communications Party Publications Figure 6. Categorization of Information Sources Media utilization is measured in a variety of ways in the CPS surveys. Questions which are relevant to media utilization in general, media use for news, media utilization for general political informa- tion, and media use for gathering.information pertaining to a specific campaign appear on the surveys. All of these will be used to examine hypotheses which involve media utilization. Of primary importance for hypothesis testing will be the measures pertaining to the sources used to gather general political information, as well as information specif- ic to a particular campaign. A rather difficult operationalization problem arises when consi- dering the concepts of level of political information and the costs and benefits of information. Consider, for example, the problems in- volved in measuring levels of information with the 1976 data. The 1976 election survey offers only two questions designed to measure the respondent's level of political information. These questions required the respondent to name the party that controlled the House of 86 Representatives both before and after the 1974 election. There are at least two problems with using these questions as a measure of politi- cal information. First, use of these items would yield an index with a range of only 0-2. An index with such a limited range cannot finely discriminate between respondents with different levels of information. Second, and more important, the questions have nothing to do with the 1976 campaign; that is, these questions do not measure levels of poli- tical information relevant to the 1976 presidential contest. Thus, another measure of political information is needed. The measure of political information that is used in this analy- sis is based on the "likes/dislikes" questions relevant to each can- didate. In 1976, each respondent was asked to name up to five things (attributes, policy positions, etc.) he/she liked about Carter and Ford, and five things he/she disliked about Carter and Ford. I have assumed that the more a respondent knows about the candidates in par- ticular, and the campaign in general, the more positive and negative attributes of the candidates a respondent can name. Thus, I have con- structed a variable based on the total number of "likes" and "dis- likes" mentioned by the respondent; for 1976, the index has a range of zero to twenty. Similar "likes/dislikes" questions were asked Of res- pondents regarding the two major parties. A "party information" index was constructed with these questions in the same fashion as the "candi- date information" index.7 Two additional concepts, the costs and benefits of information, are neither directly nor indirectly measured by the CPS surveys. Fur- ther, there appears to be no manipulations of variables available which could reasonably approximate the costs and benefits of 87 information. The problem therefore becomes one of finding suitable surrogates for these concepts. The task of measuring the costs and benefits of information has been one which has long plagued writers in the area of electoral re- search. More often than not, benefits of information are ignored, and the ability or willingness to bear the costs of information are measured, rather than the costs pg§_§g, For instance, both Downs (1957) and Popkin and his associates (1976) suggest that information costs are reduced via strong partisan identification and/or ideologi- cal structures. Downs also suggests that individuals with high in- comes will be better able, and therefore more likely, to bear the costs of information than will individuals with low incomes. In turn, Shively (1979) suggests that levels of education be used as a measure of the ability to bear information costs in general. He further suggests that the level of political resources held by an indi- vidual is related to the ability to bear the costs of political infor- mation. In his analysis of the 1956-1960 Survey Research Center panel data, Shively operationalizes a low level of political resources through agreement with the statement, “Sometimes politics and govern- ment seem so complicated that a person like me can't really understand what's going on.“ Disagreement with the statement, for Shively, indi- cated a high level of political resources. Like his predecessors, Shively hypothesizes, and finds it confirmed by the data, that those with a low ability to bear the costs of political information (i.e., those with low levels of political resources) will rely on partisan- ship as an information shortcut. Finally Frolich, Oppenheimer, Smith, and Young (1978) tested the 88 Downsian model of voter rationality using the Survey Research Center 1964 American National Election Study. Partisan identification is again used as an information shortcut. The importance of party iden- tification in the decision calculus is discounted by the amount of information held by the citizen, with levels of information operation- alized as the number of media utilized in monitoring the campaign. This brief review of past attempts to measure the costs of infor- mation, coupled with the lack of appropriate measures, leads to the conclusion that the costs and benefits of information cannot be opera- tionalized using data from the Center for Political Studies data. However, the theory developed in the last chapter does suggest a method of circumventing the problem. The theory specifies that the amount of political information gathered by a citizen is a function of the costs and benefits of that information. Further, I have assumed that the costs and benefits of information can be defined as a function of several socio-political factors. That is, the standard practice of allowing demographic vari- ables to serve as "proxies" for information costs is assumed expli- citly. The justification for this assumption is straightforward. Socio-economic-political factors, such as income, education, party identification, etc., will serve one (or both) of two functions. First, they will provide a framework for the absorption and utiliza- tion of information. Second, they will provide a relatively greater amount of assets which can be expended on political information. The result of these functions is that the relative costs of information for those individuals with high levels of education, income, etc., will be lower than for those individuals with lower levels of 89 education, income, etc. Even with no assumptions regarding the rela- tive benefits of information,8 the expectation is that levels of po- litical information are a positive function of education, income, social class, party identification, and general political interest. On the basis of this argument, I shall substitute for the follow- ing statements: Level of political information=f(costs and benefits of informa- tion) Costs and benefits of information=f(socio-economic-political factors) an alternative specification: Level of political information=f(socio-economic-political factors) That is, socio-economic-political characteristics impact upon the ability to bear the costs of information, and can therefore be used as surrogates for the costs/benefits of information. Not only is this procedure the same as that used by other writers, but it is also jus- tifiable in terms of the theory presented herein. Finally, I have constructed an index of general political activ- ity, consisting of several questions asked of respondents each year of the panel. These questions determine whether respondents had ever en- gaged in such non-electoral political activities as writing letters to newspaper editors or national leaders, signing petitions, etc. For 1976, the index has a range from zero to five, each activity counting one point.9 This index not only measures political activity, but also is taken to be a measure of general political interest, i.e., interest in politics not specific to the campaign. For purposes of measurement, 90 I have assumed that political activity is a positive function of poli- tical interest; that is, those who have a higher general interest in politics will be more likely to engage in non-electoral political activities. Positing this relationship does not seem to be problema- tic. In turn, this operationalization is supplemented by the interest questions on all three surveys. Table 2 presents some distributional characteristics of several of the compound measures used in this analysis. Table 2. Means and Standard Deviations of Uncertainty, Differential Benefit, Information and Activity Measures Variable Mean Standard Deviation UNSURCAR 1.44 .67 UNSURFOR 1.20 .68 UNSUR76 2.35 1.16 UTIL1976 .143 1.51 DIFFEEL -2.35 39.75 CANDINFO 4.73 3.15 PARTINFO 3.15 3.12 POLACTIV ’ .41 .85 An Examination of Two Assumptions, 1976 Assumption A.5 states, in part, that there is no systematic dis- tortion of information by any information source. This assumption is crucial to the model; further, it presumes a state of affairs contrary 91 to conventional wisdom. Establishing the plausibility of this assump- tion is therefore of some importance. I have conducted several tests of this assumption. The tests for 1976 follow. The tests for 1972 and 1974 are found in Appendix C. The first test involves correlating several media utilization variables with the respondent's perception of Carter's and Ford's issue posi- tions. Issue perceptions are measured through the use of the seven- point issue scales. Since nine of these scales are used for 1976 there are, for each media utilization variable, 18 correlation coeffi- cients. Media utilization is operationalized through the use of the following seven variables: X], which measures general newspaper reading habits; X2, which measures newspaper reading for political information; X3, which measures the number of radio programs about the cam- paign to which the respondent listened; 4, which measures the number of magazine articles regarding the campaign read by the respondent; X5, which measures the number of television programs regarding the campaign watched by the respondent; X6, which measures general television viewing habits; 7, which measures the use of television by the respondent for gathering political information. The assumption being tested implies that there is no relationship be- tween these variables and perceptions of candidates' issue positions}0 The following table summarizes the results of the correlational analy- sis. The coefficients are Spearman's rhos. 92 Table 3. Summary of Correlational Analysis Between Perceptions of Candidates' Issue Positions and Media Reliance Variables (X1'X7) Independent Number of Number of Correlations Variable Correlations Where p<.05 Where p>.05 05r:.05 .O5.1 X1 9 3 4 2 X2 6 O 7 5 X3 16 O 2 0 X4 9 O 6 3 X5 6 3 5 4 X6 15 2 l 0 X7 13 2 3 0 TOTALS (126) 10 28 14 % 7.9 22.2 11.1 (Percentages do not sum to 100% due to rounding.) 93 Table 3 suggests that the degree to which a citizen relies upon any single medium does not materially affect his/her perceptions of candidates' issue positions. To give a better idea of what these cor- relation coefficients look like, I have presented the coefficients for the seven media use variables and two seven-point issue scales (gov- ernment guarantee of jobs and national health insurance) in Table 4. Table 4. Correlations Between Media Use Variables and the Government Guarantee of Jobs Scale, National Health Insurance Scale (Coefficients are Spearman's Rho's) Jobs/Ford Jobs/Carter Health/Ford Health/Carter X1 .061* .012 .059* -.10* x2 .10* -.038 .099* -.12* x3 .017 .02* .003 .06 x4 -.10* .O67* -.O86* .10* x5 -.11* .045* -.10* .15* x6 .042* .033 - 014 -.035 x7 .093* .032 .022 ' -.O65* * p §_.05 These coefficients are presented because they are gg£_representa- tive of the tests of the hypothesis that media use and perceptions of candidates' issue positions are unrelated. They are not representative in that they are larger, in general, than the coefficients between the seven media use variables and all the rest of the seven-point issue 94 scales. A brief scanning of Table 4 reveals the weakness of the rela- tionships between media utilization and perceptions of candidates' issue positions. In general, these data contribute to the plausibil- ity of the assumption that the media are unbiased. Variable X8 measures whether the respondent relied on television or newspapers as his/her primary source of political information dur- ing the campaign.H Each of the eighteen issue position perception variables were cross-tabulated with X8, and the results are similar to the results presented in Table 3. For seven of these cross-tabula- tions, x2 was not significant at the .05 level of probability. Of the eleven relationships that were significant (p < .05), the gammas for four were less than or equal to .05. Two of the statistically signif- icant relationships had gammas ranging from .05 to .10, while five had gammas equal to or greater than .10. X8 was also correlated with vote choice;12 this yielded a small but statistically significant correla- tion coefficient (r=.046; p < .05). However, controlling for either education, income, or social class obliterated the already weak rela- tionship.13 As before, the results lend support to the assumption. Finally, Xl'x8 were correlated with the utility measures based on the seven-point issue scales (Y1) and the feeling thermometers (Y2). Only two (out of eighteen) of the correlations involving Y1 were sig- nificant at the .05 level of probability; the coefficient for Y1 and X6 equals -.O96, and for Y1 and X7 the coefficient equals -.112. The correlations involving Y2 yielded four relationships significant at the .05 level of probability. The coefficients between Y2 and X1, X2, X6’ and X8 all ranged between .05 and .10. In general, media reliance variables are not good predictors of 95 respondents' perceptions of candidates' issue positions, vote choice, or expected differential benefit. Therefore, assuming that informa- tion sources do not systematically distort information is plausible. The second assumption to be examined is that demographic charac- teristics of citizens serve to reduce the relative costs of informa- tion. Given that the costs of information are not directly measur- able, it is assumed that such factors as social class, income, educa- tion, partisan affiliation, and general political interest serve to increase the ability (or willingness) to bear the costs of informing oneself. Hence, these socio—economic-political characteristics are hypothesized to be positively related to levels of political informa- tion. Testing this hypothesis will therefore provide evidence regard- ing the plausibility of this assumption. Five respondent attributes were correlated with levels of infor- mation regarding the presidential candidates: the respondent's level of education, income, and social class; the strength (not direction) of respondent's partisan affiliation; and, the level of respondent's general political interest. The first three characteristics are gen- erally thought to be relevant demographic factors; all three are hypo- thesized to be positively related to respondent's level of candidate- referenced information. As noted previously, party identification is sometimes considered to be an information shortcut; that is, party identification serves as decision making material, in order for the citizen to avoid the costs of information. This implies a negative relationship between party identification and levels of political in- formation. .On the other hand, assuming (not unreasonably) that parti- san affiliation provides a structure for the absorption and utilization 96 of information implies a positive relationship between party identifi- cation and levels of political information. Given my earlier argu- ments on this score, I am adopting the latter hypothesis as part of this assumption. By general political interest, I mean interest in the political process apart from the electoral process. As noted earlier, political interest is operationalized by an index of non-electoral political ac- tivity. Not only should such interest or activity increase one's general political knowledge, but it should also act as a vehicle, like party identification, for the absorption and utilization of political information relevant to the specific electoral campaign. Hence, it is hypothesized that a respondent's level of general political interest is positively related to his/her level of political information (spe- cifically candidate-referenced information). Simple Spearman's rhos were computed for each of these relation- ships, and are displayed in Table 5. A11 coefficients are positive, as hypothesized, and none are likely to have occurred by random chance (p §_.05 for all coefficients). The impact of strength of party affiliation on level of information is the weakest, with r=.065. This result may be interpreted as a lack of support for the positive relationship implied by the assumption, although such an interpreta- tion is arguable, at best. However, the data certainly constitute a rejection of the more traditional hypothesis of a negative relation- ship between party identification and levels of political informa- tion. Further, the relationship is positive, as implied by the assump- tion, and is statistically significant. Hence, I will argue that this implication of the assumption is plausible. Table 5. Relationships Between Political Information and Media Use, 97 and Socio-Political-Economic Variables rv Independent Variable Dependent Variable Strength of Party Identification Political Interest Education Social Class Income Political Information (.065) Political Information (.245) Political Information (.370) Media Reliance (-.200) Political Information (.223) Media Reliance (-.228) Political Information (.240) Media Reliance (-.200) (Entries in parentheses are Spearman's rhos; all coefficients are significant at the .05 level of probability.) 98 The relationships between level of political information and general political interest, education, income, and social class are of such strength as to preclude serious controversy. Certainly no causal relationships between the variables can be established. However, the fact that the measures of association range from .223 to .37 lend strong support to the implication of the assumption. The data indi- cate that assuming that socio-economic-political characteristics can stand as "proxies" for the costs of political information is plausible. One additional implication of this assumption can be tested. If socio-economic-political characteristics function to decrease the relative costs of information (or increase the ability to bear these costs), then individuals with higher levels of education, income, and social class should be observed to utilize more expensive sources of information than those with lower levels of education, income, etc. Given the argument in the preceding chapter, newspapers are more ex- pensive sources of information than is television. The hypothesis, then, is that individuals with higher levels of these attributes will use newspapers as the primary information source about the campaign to a greater degree than will individuals with lower observed levels of these attributes. The data relevant to this hypothesis are presented in Table 5. The media reliance variable is constructed in such a way as to yield a negative coefficient when a positive relationship (in the sense hypo- thesized) is observed, i.e., when newspapers are preferred to televi- sion as a political information source by those with higher levels of education, income, etc.14 The magnitude and statistical significance of the coefficients lend additional support to the assumption. 99 Conclusion In this chapter I have attempted to demonstrate, through empiri- cal examination, the plausibility of the major assumptions of the theory of citizen information gathering. In general, the data confirm the validity of the assumptions. Also, the operationalization proce- dures used to measure several key concepts have been introduced. In the next chapter a number of hypotheses will be derived from the theory. 100 ENDNOTES 1The nine seven-point issue scales from the CPS 1976 survey used throughout this analysis are: 1) government guarantee of jobs; 2) rights of accused criminals; 3) busing to achieve integration; 4) government aid to minorities; 5) national health insurance; 6) urban unrest; 7) legalization of marijuana; 8) progressive income tax; and, 9) women's equality. The government guarantee of jobs scale appeared in both the pre-election and post-election waves of the 1976 survey. Only the pre-election wave responses are used here. For 1972 and 1974 the operationalization procedure for utility is the same; however, the issue scales used are different. From the 1972 survey, the following issue scales are used to calculate differential benefit: 1) government guarantee of jobs; 2) progressive income tax; 3) governmental action against inflation; 4) legalization of mari- juana; 5) busing to achieve integration; and, 6) government aid to minorities. All but the last of these scales were asked in both the pre-election and post-election waves of the 1972 survey. The govern- ment aid to minorities scale was asked only in the post-election wave, but was asked of the entire sample. The other scales were asked of half-samples, but the samples which were asked the questions differed between the pre-election and post-election waves. 101 2In 1974, the issue scales asked of respondents were relevant to the Democratic and Republican parties, as well as to the major party candidates for Senate races, if one was being held in the respondent's state. Given the importance of partisan affiliation in off-year elec- tions, two measures of differential utility have been calculated. One of these is the expected differential benefit deriving from the elec- tion of one of the parties; the other is relevant to Senatorial candi- dates. The issue scales from the 1974 survey which are used in this analysis to measure party differential benefit are: 1) government guarantee of jobs; 2) urban unrest; 3) rights of accused criminals; 4) busing to achieve integration; and, 5) government aid to minori- ties. In measuring candidate differential, all of these scales were used with the exception of the last scale, for which respondents were given no opportunity to place the candidates. This scale was replaced with the equal rights for women scale, on which respondents could place the candidates, but not the parties. 3The salience questions asked in all three surveys dealt with issues different from those used for the seven-point issue scales. Having no measure of salience, I have assumed that issues are equally weighted by the respondent in calculating expected differential bene- fit. 4The feeling thermometers were not asked of respondents for Sena- torial or Congressional candidates in 1974. Correlating these two measures of differential benefit based on seven-point issue scales and feeling thermometers indicate that they are measuring the same thing; 102 presumably, what they are measuring is the expected utility for the election of one candidate over another. For 1976, the correlation coefficient between these two variables is -.762; for 1972, the coef- ficient is -.68. 51h formal rational choice theory, uncertainty refers to the situation in which the citizen has no knowledge regarding the proba- bility distributions over the states of the world. On the other hand, the concept of risk is used in connection with the situation wherein a probability distribution over the states of the world is known, or knowable (cf., Ferejohn and Fiorina, 1974). 6Since respondents were reacting to the issue scales in terms of different candidates in 1974, the "average placement of a candidate" for the entire sample has no real meaning. Hence, uncertainty is not measured for candidates. Rather, the procedures described in the text are used to measure uncertainty regarding the issue positions of the major parties. Thus, OD, refers to the placement of the Democrat- ic Party on issue i. It can be argued that this procedure assumes the existence of two national parties, and that such an assumption contra- dicts common knowledge regarding the organization of the party sys- tem(s) in this country. I plead guilty on both counts, while at the same time noting that this procedure is more realistic than assuming that all respondents make reference to the same candidates when res- ponding to the issue scales. This latter assumption, of course, is patently absurd. 103 7On the 1972 survey, the political information quiz yielded five questions which could be used to form an index of political knowledge. However, as in 1976, these questions measured general political infor- mation, not information specific to the campaign. The 1974 survey contained no questions comparable to these general information items. For these reasons, as well as to ensure some comparability across surveys, I have used the "likes/dislikes" questions to measure politi- cal information specific to the campaign for 1972 and 1974, as well as for 1976. 8Three such assumptions are possible: 1) The relative benefits of information increase with increasing levels of education, income, etc.; 2) The relative benefits of information are invariant with res- pect to education, income, etc.; 3) The relative benefits of informa- tion decrease with increasing levels of education, income, etc. The first two assumptions would not alter this analysis. The third assump- tion stretches the limits of plausibility. 9For 1976, the political activity, or interest, index is based on variables 3046-3050. For 1972, the index has a range of 0-4, and is based on variables 469-472. For 1974, the range of the index is 0-4, and is based on variables 2196-2199. 10The reader will notice that this procedure is not testing whether the news media are unbiased, but whether messages transmitted by the media are perceived (or received) in an unbiased fashion. In truth, it is possible for the media to transmit biased messages which are 104 selectively perceived and/or retained by the individual. In turn these individuals, in responding to issue scale questions, could give answers which make it appear that the media are unbiased. However, such a test is the best that can be achieved using survey data. A true test of the assumption would involve a content analysis of all the information sources used by respondents (including local newspapers, as well as national network news programs), a task far beyond the scope of this dissertation. 11The response categories are: l (newspapers), 2 (both are used equally) and 3 (television). In the 1976 survey, this is Variable 3654. 12Vote choice in the 1976 survey is Variable 3665. Only cate- gories 1 (Ford) and 2 (Carter) are used. . 13Controlling for education, r=.03; controlling for income, r=.03; controlling for social class, r=.02. For all three coeffi- cients, p 3_.10. Controlling for all three variables, r=.Ol, p > .30. 14See footnote #6. .a— - v“, CHAPTER 4 DERIVATION 0F HYPOTHESES Introduction In this chapter a number of hypotheses regarding information gathering, media utilization, and levels of political information will be developed. Most of these hypotheses are directly implied by the theory of citizen information gathering. However, a few hypotheses are only "suggested" by the theory; that is, arguments leading to these hypotheses can be made, but the arguments rely on assumptions outside the theory. Each hypothesis will be stated explicitly, and then the argument will be laid out in detail. Following the development of eighteen hypotheses, the statements will be summarized by classifying them ac- cording to the dependent variables of the propositions. The five de- pendent variables discussed in this chapter are levels of political information; media utilization; levels of uncertainty regarding candi- dates' issue positions; turnout; and candidate choice. Derivation of Hypotheses Hypothesis 1: There is a positive association between the expense (costliness) of the media used by a citizen to gather political information and that citizen's level of political information. 105 106 Higher levels of political information will be exhibited by citizens who use more expensive media, say newspapers, than by those who rely on less expensive media, for example television, in order to gather political information. The argument for this hypothesis is straightforward. The cost function Ci’ of any information source is a function of the complexity of the information realized from the source. Complexity is defined as the amount of information relevant to the voting decision per "bit" of information. The more complex an information source, the more rele- vant information per bit of information it has, and the more costly it is to utilize. Let Gi denote the complexity of an information source, Si' If the relation between the cost functions of two information sources, Si and Sj’ is Ci > Cj’ then 1. 61 > Gj That is, the more expensive an information source, the more informa- tion relevant to the voting decision per bit of information it will yield. Hypothesis 2: There is a negative relationship between a citi- zen's level of political information and that citizen's level of uncertainty regarding the issue positions of the candidates. Uncertainty is defined as P(E), the probability of making an error in basing the voting decision on an hypothesis test regarding the relative utility of the two candidates. Further, P(E) is defined as the intersection of the sampling distributions of Ua and Ub’ the 107 estimators of expected utility to be derived from the election of can- didates A and B, respectively. Further, such that as the variances of the estimators decrease, P(E), and hence uncertainty, decreases. Let n denote the number of bits of information gathered by a 2 citizen. Any bit of information will either decrease oa and/or OS, or will leave them unchanged; i.e., an additional bit of informa- tion will not increase 0: or 05' Formally, 3 n++oz+ 2 and/or n++ob+ Equation 2 can be written as 2. OEHPUEM 2 °b ++P(E)+ Combining Equations 2 and 3 gives 4. 0+*P(E)+ That is, increasing levels of political information will lead to de- creasing levels of uncertainty regarding the issue positions of the two candidates. Hypothesis 3: There is a positive association between the degree 108 to which a citizen uses any medium to gather political informa- tion and that citizen's level of political information. To understand the validity of this proposition, it is helpful to consider its obverse: increased media utilization will lead to either no changes in levels of political information, or will lead to de- creased levels of political information. This state of affairs will occur only when each additional bit of information gathered contains no unique information about either candidate, which in turn will hap- pen only if the citizen has gathered all possible information about the candidates, or can assimilate no more information. Given that it is highly unlikely that any citizen will have perfect (or near per- fect) information or can absorb no more information, additional bits of information will (almost) always increase an individual's store of knowledge about the candidates. Further, additional bits of informa- tion will be realized by increased media utilization. Hence, a posi- tive association between degree of media utilization and levels of political information should be observed. Note that this hypothesis applies to any information source-- television, newspapers, party publications, etc. That is, the gain in information realized by increased use of a particular medium is inde- pendent of its complexity, and of its cost. Hypothesis 4: There is a negative association between the ex- pense (costliness) of the medium a citizen uses to gather poli- tical information and that citizen's level of uncertainty regard- ing the issue positions of the candidates. 109 This hypothesis follows directly from Hypotheses 1 and 2. Hypothesis 1 states 5. (Media expense)++Information+ Hypothesis 2 states 6. Information++Uncertainty+ Therefore, 7. (Media expense)+*Uncertainty+ which is a restatement of Hypothesis 4. Hypothesis 5: There is a negative association between the degree to which any medium is used by a citizen to gather political in- formation and that citizen's level of uncertainty regarding the issue positions of the two candidates. This hypothesis follows directly from Hypotheses 2 and 3. Hypo- thesis 2 states 6. Information++Uncertainty+ Hypothesis 3 states 8. (Degree of media use)++Information+ Therefore, 9. (Degree of media use)++Uncertainty+ which is a restatement of Hypothesis 5. 110 HVpothesis 6: There is a negative relationship between an indi- vidual's level of uncertainty regarding the issue positions of the candidates and the probability that he/she will vote. According to the theory of citizen information gathering, a necessary condition for voting is that ENGSi, for all Si’ has been maximized. Further, 10. ENGS1.=-'B1.-C.i where B1=EVSIi, the expected value of sample information. EVSI, is a function of expected utility; therefore, it is a function of the like- lihoods of the states of the world occurring, given the sample results examined during hypothesis testing.1 Since these likelihoods are functions of P(E), then 11. EVSIi=f(P(E)) And 12, ENGSi=f(P(E)) Therefore, 13. P(Voting)=f(P(E)) such that as uncertainty, or P(E), decreases, the probability of voting increases. Thus, the probability of voting is a function of uncertainty. Hypothesis 7: The later in a campaign a citizen arrives at a voting decision, the lower will be his/her level of uncertainty 111 regarding the issue positions of the two candidates. This hypothesis cannot be derived directly from the theory, but is suggested by the theory. Since a voting decision will not be made until uncertainty has been reduced to as low a level as possible (sub- ject to the constraint of maximizing ENGSi for all S1), then it is presumed that citizens who have come to a voting decision late in the campaign have felt a subjective need to greatly reduce uncertainty. Hence, they would gather more information than an "early decider," and would therefore exhibit lower levels of uncertainty. Hypothesis 8: The later in a campaign a citizen arrives at a voting decision, the higher will be that citizen's level of po- litical information, ceteris paribus. This hypothesis follows directly from Hypotheses 3 and 7. By Hypothesis 7, the timing of the vote decision is a function of uncer- tainty. Further, by Hypothesis 3, 6. Informationi+Uncertainty+ Therefore, late deciders will exhibit higher levels of information than will early deciders. Hypothesis 9: The later in a campaign a citizen arrives at a voting decision, the greater will be that citizen's use of any medium for gathering political information, ceteris paribus. This hypothesis is suggested by Hypotheses 5 and 7. By Hypothe- sis 5, 112 9. (Degree of media use)++Uncertainty+ By Hypothesis 7, late deciders will have less uncertainty than will early deciders. Therefore, late deciders will have utilized any medi- um to a greater degree than will have early deciders for the purpose of gathering political information. Hypothesis 10: The later in a campaign a citizen arrives at a voting decision, the more expensive will be the information sources used to gather political information, ceteris paribus. This hypothesis follows directly from Hypotheses 7 and 4. Hypo- thesis 7 states that the timing of the vote decision is a function of the citizen's level of uncertainty. Hypothesis 4 states 7. (Media expense)++Uncertainty+ Therefore, the timing of the vote decision is a function of the cost- liness of the information sources that a citizen uses. Specifically, the more expensive the medium used to gather political information, the later the timing of the vote decision. Hypotheses 8-10 are ceteris paribus hypotheses; other things being equal, the hypothesized relationships are expected to hold. The ceteris paribus stipulation is most important for these hypotheses. The major thrust of the theory is that citizens gather political infor- mation purposively, i.e., to aid them in making a voting decision, and to reduce the uncertainty about the decision. However, people gather information for other than electoral purposes. They may gather poli- tical information in order to discuss politics with friends, or merely 113 because they enjoy observing (or participating in) electoral politics. In other words, some individuals, perhaps many, have a great interest in politics, pg; sg, We would expect these individuals to gather more information during a campaign than might be expected, regardless of such factors as the timing of the vote decision. In turn, citizens with a relatively higher stake in the outcome of the election, i.e., those with higher differential benefits, might also be expected to gather more political information than the theory would suggest. Since the theory of citizen information gathering is not relevant to these non-electorally purposive information gathering behaviors, it will be necessary to control for their effects while testing the relationship between levels of information and timing of the vote decision, and between media reliance and the timing of the vote decision. Therefore, these relationships will be controlled for levels of political activity and levels of differential benefits. Hypothesis 11: There is a positive relationship between levels of an individual's perceived differential benefit and the proba- bility of that person voting. Hypothesis 12: Individuals whose perceived differential benefit indicates that more utility will be derived from the election of the Democratic candidate will vote for the Democratic candidate. Those whose perceived differential benefit indicates that more utility will be received from the Republican candidate will vote for the Republican candidate. Both Hypotheses 11 and 12 flow directly from Assumptions 1 and 2, 114 which state that citizens are rational, and that this rationality is ’exhibited by maximizing perceived expected utility. Testing these two hypotheses may be viewed as an empirical examination of Assumptions 1 and 2. Hypothesis 13: There is a positive association between the ex- pense of the information sources a citizen uses to gather politi- cal information and the probability that the citizen will vote. This hypothesis follows directly from Hypotheses 6 and 4. Hypo- thesis 6 states 13'. Uncertainty++P(Voting)+ Hypothesis 4 states 7'. (Media expense)++Uncertainty+ Therefore, 14. (Media expense)++P(Voting)+ which is a restatement of Hypothesis 13. In more concrete terms, this hypothesis stipulates that, for example, citizens who use newspapers as their primary source of political information are more likely to vote than citizens whose primary source of information is television. Hypothesis 14: There is a positive association between the degree to which any one source is used to gather political infor- mation by a citizen and the probability that the citizen will vote. 115 This hypothesis follows directly from Hypotheses 5 and 6. Hypo- thesis 6 states 13'. Uncertainty++P(Voting)+ Hypothesis 5 states 9. (Degree of media use)++Uncertainty+ Therefore, 14. (Degree of media use)++P(Voting)+ which is a restatement of Hypothesis 14. This relationship will hold for all information sources, independent of their costs and benefits. Thus, increased use of television will lead to a higher probability of voting, an hypothesis that is contrary to the conventional literature discussed in Chapter 1. Hypothesis 15: There is a positive association between an indi- vidual's level of political information and the probability that he/she will vote. This hypothesis follows directly from Hypotheses 2 and 6. Hypo- thesis 2 states which can be restated as 4'. Information++Uncertainty+ Hypothesis 6 states 116 13'. Uncertainty++P(Voting)+ Therefore, 16. Information+*P(Voting)+ which is a restatement of Hypothesis 15. Hypothesis 16: There is no relationship between an individual's level of political information and the direction of that person's vote choice. This hypothesis follows directly from Assumption 5, which states in part that the information received from any information source is unbiased. This directly implies that 0a and 0b are not functions of either the content of information or the amount of information gathered. Therefore, for whom a citizen casts his/her vote is inde- pendent of levels of information. Hypothesis 17: Television is the most heavily utilized source of information during a campaign. This hypothesis is suggested by the theory, and cannot be derived from it. Further, it is not a unique hypothesis; it reflects conven- tional wisdom. However, most attempts at explaining this phenomenon implicitly rely on some sort of cost/benefit analysis. The theory presented in Chapter 2 uses this mode of analysis explicitly, and therefore yields a precise and explicit explanation of the predomi- nance of television as the primary source of political information for most citizens. Specifically, assuming that the benefits to be derived 117 from gathering information will be quite low for many (if not most) citizens, and given that television is a relatively inexpensive infor- mation source, television is the predominant source of political infor- mation during an electoral campaign. The final hypothesis to be presented is one which, like the pre- vious hypothesis, is only suggested by the theory. The argument for it is somewhat long; therefore, for this last hypothesis, I will pre- sent the argument before the proposition. Recall that it is assumed that certain citizen characteristics, such as education, social class, etc., serve to reduce the costs of absorbing and/or utilizing information. In turn, decreased costs will lead to higher levels of political information, since the citizen will be able to "afford" to spend more for gathering information. Because costs and benefits of information cannot be directly operationalized, this assumption has been modified to state that information levels are functions of the socio-economic-political characteristics of citizens. To clarify these relationships, a simple model is presented in Figure 7. INCOME ‘ 1 BENEFITS EXPENSE OF MEDIA INFORMATION EDUCATION 1 i UTILIZED i LEVELS ; COSTS SOCIAL CLASS y Figure 7. Factors Influencing Information Levels Next, the model implies that as more information is gathered, uncertainty is reduced, which in turn reduces the expected benefits of additional information. Since benefits are reduced, the rational 118 citizen will attempt to reduce the costs of gathering additional infor- mation; that is, the rational citizen will use less expensive informa- tion sources. Thus, the theory not only suggests that information levels are a function of media choice, but also that media choice is a function of information levels. These relationships are represented rather simplistically in Figure 8. INFORMATION ; UNCERTAINTY ; BENEFITS OF ; EXPENSE OF LEVELS INFORMATION MEDIA UTILIZED Figure 8. Factors Influencing Media Use Taken together, Figures 7 and 8 imply the following: INCOME EXPENSE OF INFORMATION EXPENSE OF EDUCATION 1 MEDIA UTILIZED i LEVELS ; MEDIA UTILIZED SOCIAL CLASS Figure 9. Factors Influencing Media Use However, as shown in the first chapter, the following relation- ships are posited in the traditional literature: INCOME EDUCATION + EXPENSE 0F , MEDIA UTILIZED 1 INFORMATION LEVELS SOCIAL CLASS Figure 10. Factors Influencing Information Levels That is, the reciprocal relationships between information levels and the expense of the information sources used to gather information has heretofore been ignored. Simple positive relationships only are hypo- thesized between these three concepts. When such an hypothesis guides 119 research, investigation of the interaction of these sets of variables is ignored. If Figure 10 is a correct representation of the manner in which these variables are related, i.e., if there is a reciprocal relation- ship between levels of information and the expense of media utilized, then the positive relationships between education, income, and social class and the expense of media used should weaken once the effects of levels of information are controlled. Therefore, the final hypothe- SES are: Hypothesis 18: The higher is a citizen's level of income, educa- tion, or social class, the more expensive will be the medium or media used by the citizen to gather political information during a campaign. Hypothesis 18.a: The relationship hypothesized above will dimin- ish when levels of political information are held constant. Summary of Hypotheses The following hypotheses are implied by or suggested by the theory of citizen information gathering. Hypotheses Regarding Political Information Hypothesis 1: There is a positive association between the ex- pense (costliness) of the media used by a citizen to gather political information and that Citizen's level of political information. Hypothesis 3: There is a positive association between the degree to which a citizen uses any medium to gather political information and 120 that citizen's level of political information. Hypothesis 8: The later in a campaign a citizen arrives at a voting decision, the higher will be that Citizen's level of political information. Hypotheses Regarding Uncertainty Hypothesis 2: There is a negative relationship between a citi- zen's level of political information and that citizen's level of un- certainty regarding the issue positions of the candidates. Hypothesis 4: There is a negative association between the ex- pense (costliness) of the medium a citizen uses to gather political information and that citizen's level of uncertainty regarding the issue positions of the candidates. Hypothesis 5: There is a negative association between the degree to which any medium is used by a citizen to gather political informa- tion and that citizen's level of uncertainty regarding the issue posi- tions of the candidates. Hypotheses Regarding Media Utilization Hypothesis 9: The later in a campaign a citizen arrives at a voting decision, the greater will be that Citizen's use of any medium for gathering political information, ceteris pgribus. Hypothesis 10: The later in a campaign a citizen arrives at a voting decision, the more expensive will be the information sources used by that citizen for gathering political information, ceteris paribus. 121 Hypothesis 17: Television is the most heavily utilized source of political information during a campaign. Hypothesis 18: The higher is a citizen's level of income, educa- tion, or social Class, the more expensive will be the medium or media used by the citizen to gather political information during a campaign. Hypothesis 18.a: The relationships hypothesized above will diminish when levels of political information are held constant. Hypotheses Regarding Voting Hypothesis 6: There is a negative relationship between an indi- vidual's level of uncertainty regarding the issue positions of the candidates and the probability that he/she will vote. Hypothesis 11: There is a positive association between levels of an individual's perceived differential benefit and the probability of that person voting. Hypothesis 12: Individuals whOSe perceived differential benefit indicates that more utility will be derived from the election of the Democratic candidate will vote for the Democratic candidate. Those whose perceived differential benefit indicates that more utility will be derived from the Republican candidate will vote for the Republican candidate. Hypothesis 13: There is a positive association between the ex- pense of the information sources a citizen uses to gather political information and the probability that the citizen will vote. Hypothesis 14: There is a positive association between the de- gree to which any one medium is used to gather political information by a citizen and the probability that the citizen will vote. 122 Hypothesis 15: There is a positive association between an indi- vidual's level of political information and the probability that he/ she will vote. Hypothesis 16: There is no relationship between an individual's level of political information and the direction of that person's vote Choice. Hypothesis Regarding the Timing of the Vote Decision Hypothesis 7: The later in a campaign a citizen arrives at a voting decision, the lower will be his/her level of uncertainty re- garding the issue positions of the candidates. 123 ENDNOTES 1From Chapter 2, Figure 3, the four states of the world of con- cern to the decision maker in the model are l) U; > Ub; 2) Ua > Ub’ but U; j-Ub' 3) Ub > Ua’ but Ug 5-Ua; and; 4) U5 > Ua' CHAPTER 5 EMPIRICAL EXAMINATION 0F HYPOTHESES Introduction In this chapter, the hypotheses derived in the last chapter will be examined with data from the 1976 Center for Political Studies American National Election Study.1 I shall pursue two major goals in this chapter, each relative to the development of the simultaneous equation model in the next chapter. First, the implications of the theory of citizen information gathering can be fleshed out in greater detail than is possible with the simultaneous-equation model. Second, testing the hypotheses derived from and suggested by the theory allows an examination of some of the assumptions underlying the specification of the statistical model. Each hypothesis will be restated in terms different from those used at the end of Chapter 3; specifically, hypotheses will be phrased in operational terms. For instance, "newspapers" will be used in place of “more expensive information sources," "television" will be used in place of "less expensive information sources," etc. Further, the hypotheses are organized and presented around major con- cepts. Hypotheses 1-4 treat levels of political information; Hypotheses 5-7 are concerned with uncertainty regarding candidate issue positions; Hypotheses 8-14 deal with voting, information, 124 125 uncertainty, and timing of the vote decision; and, Hypotheses 15 and 16 deal with media choice and media utilization. Finally, I return to the concept of reinforcement in the last section of this chapter. Within the context of this discussion, I develop three measures of citizen misperception of candidate issue positions, and examine the relations between misperception, par- tisanship, candidate preference, levels of information and media utilization. Empirical Examination of Hypotheses Hypothesis 1: Citizens who rely on newspapers as their primary source of political information will exhibit higher levels of political information than citizens who rely on television as their primary source of political information. Hypothesis 2: There is a positive association between the extent to which any one medium is utilized and levels of political information. Both hypotheses will be tested using two levels of political information variables, the party-related information variable (PARTINFO), and the candidate-related information variable (CANDINFO). The media reliance variables is V3654, whose categories are news- papers (1), television (5), and both newspapers and television equally (3). A negative relationship confirms Hypothesis 1. The correlation coefficient calculated is Spearman's rho. Correlating -.194 (p < .05); correlating -.160 (p < .05). The rela- media reliance and PARTINFO yields r media reliance and CANDINFO yields r tionships are not particularly strong, but the null hypotheses can 126 be rejected. The same information variables are used to test Hypothesis 2. The following five variables are used to operationalize degree of medium utilization: (1) number of radio programs about the campaign to which respondents listened (V3601); (2) number of magazine arti- cles about the campaign read by respondents (V3603); (3) number of television programs about the campaign respondents watched (V3605); (4) the degree to which respondents read newspaper to gather polit- ical news (POLPAPER); and (5) the degree to which respondents watched television to gather political news (NEWSVIEW).2 For V3601, V3603, and V3605, the response categories are "a good many" (1), "several" (2), and "one or two" (3); thus a negative relationship confirms the hypothesis. The latter two variables are additive indices; thus, positive relationships confirm the hypothesis. In order to simplify the presentatiOn Of the relevant data, Spearman's rho's for all of the possible relationships are given in Table 6. Table 6. Relationships between Levels of Political Information and Degree of Media Utilization PARTINFO CANDINFO V3601 (radio use) -.052* -.019 V3603 (magazine use) -.245* -.250* V3605 (television use) -.270* -.247* POLPAPER .387* .378* NENSVIEN .139* .098* *p < .05 127 The weak relationships between radio utilization and information levels are not unexpected. Relative to the other mass media, radio newscasts are rather short, and the number of campaign relevant pro- grams are few in number. However, both coefficients are in the right direction, although only one is statistically significant. For all other Operationalizations, the null hypotheses are rejected. In gen- eral, levels Of political information are a function of the citizen's reliance on the various mass media in monitoring the campaign. The strongest coefficients in Table 6 are between levels of political information and degree of newspaper utilization. The metric underlying NEWSVIEW and POLPAPER cannot be assumed to be com- mon. Therefore, any inferences comparing newspapers and television in terms of providing political information will be highly specula- tive. However, these data, in conjunction with the relationship between V3654 and levels of information, indicate not only that newspaper readers get more information than television viewers, but also that the former acquire more information at a faster rate than do the latter. That is, the more that an individual utilizes news- papers as a source of political information relative to the individ- ual who relies on television as his/her primary source of political information, the more information that person will realize. Hypothesis 3: The later in the campaign a Citizen decides whether and/or how to vote, the higher will be that Citizen's level of political information. 128 To test Hypothesis 3, the twO information level variables, CANDINFO and PARTINFO, were correlated with V3666, the variable from the 1976 CPS survey which determines when the respondent decided how to vote. The Spearman's rho calculated for V3666 and CANDINFO equals .0361, with p > .05. The coefficient is in the right direction, but is miniscule and statistically insignificant. For PARTINFO and V3666, r=-.O621, with p < .05. Not only is the coefficient in the opposite direction hypothesized, but it is statistically significant. Clearly, the null hypothesis cannot be rejected. When Hypothesis 3 was developed, it was developed with a ceteris paribus condition. Specifically, I suggested that some citizens, even though they had decided how they were going to vote very early in the campaign, might still continue to gather political information. This continued information gathering might be due to high levels of a general interest in politics, or perhaps due to high levels of expected differential benefit. To get a clearer picture of how information gathering interacts with the timing of the vote decision, the effects of these factors should be controlled. These controls were effected as follows. First, three variables - political interest (based on the political activity index), expected differential benefit based on the issue scales, and expected differential benefit based on the feeling thermometers - were broken at the median response into high and low categories. Call these variables, respectively, INTEREST, UTILITY, and AFFECT. V3666, the timing of the vote decision variable, was broken at category 5 (before the first debate) into early and late categories. Call this 129 variable DECIDE. The information variables (CANDINFO and PARTINFO) and the two differential benefit measure (UTILITY and AFFECT) yield four combinations of variables with which to conduct further in- vestigation Of Hypothesis 3. Next, the average level of political information was calculated for the different levels of political interest, expected differential benefit, and timing Of the vote decision. Finally, differences of means tests were constructed. Figures 11-14 present these data, without t-tests, in a tree diagram format. These figures are designed to answer three questions, the assumed or hypothesized answer to each being in the affirmative: Do citizens with high levels of political interest have higher levels of political information than citizens with low interest? Within categories of interest, do citizens with high levels of expected differential benefit have more political information than those Citizens with low levels of expected differential benefit? Within categories of political interest and expected differential benefit, do late deciders have more political information than early deciders? To answer these questions, each of the four figures will be discussed. In Figure 11 high interest citizens have more information than low interest citizens, but the difference is not statistically sig- nificant.3 One of the two comparisons between citizens with high and low UTILITY yields a difference in the hypothesized direction, but neither difference is significantly different from zero. Three of the four comparisons between early and late are in the hypothesized direction; the difference of means between early and late deciders in the High INTEREST/High UTILITY category is statistically 130 A.odzmozm_ came use mm_spcmv .uomuuo use .>eH4~»: .pmmmmez_ An odzmozm_ umNPmmzuoazg segue; on» new: agommumut mm.n Pm.o N.o om.o mm.m o.n mm.w mmto mofiomo manor monomo apcmm mowowo woman momomo xpgmm. mofiomo mumon monomo A—me moaomo evade womowo apemm mm.o m¢.o n—.n om.~ >e~4Hho so; >HH4HH= zmezt >h~4~h2 zoo >h~4mho gamma ~e.e _N.e emmempzc zoo emmmmpza amaze mm.e Ammmnzo o_e5am oeeeem 131 A.oazmozmp cams mew mmwsucmv .momomo new .pomau< .Hmmmmhzm x: omz_o2wp uwNPmmsuoazc smear; we» zap: xeommumut mo.m om.e ¢N.m mp.“ No.9 oo.o om.o momomo woman momomo xpcwu moHouo mung; uoHomo x—Lmu moHowo mung; moHomo zmuwm momomo manor monomo h—me Po.m homuu<.3oo mo.m mo.o mp.o mo.m Homoe< emeze pumae< zoo panama emeza om.o hmmmmhzm zoo emummezc emexe we.m Apoe_rzv oFeEem oeceew 132 A.Ouzoexmp some men mmmeucuv .mo_umo use .>»Ho_»= .pmmmmezo an oazfiemmp cmNPmmsuoax; smgaeg me» new: esommumut mm.m om.m o.e em.e mm.o N¢.m om.m mm.o moHomo mpmoa momomo Afigmu momowo manor momomo z—me moHomo mumon moHomo xamum moaomo mason moaomo Apcmm Ne.m om.e we.m . ao.e >eoooe= zoo >eoooe= emeza >eooop= zoo spooopa emoze em.e mm.m Hmmmmth zoo hmmmmezu ammzt mm.m Ammmnzo o_a5em oeeoem 133 A.odzoemmp some men mmmcucmv .uooomo new .eumau< .emmmmezn x; cczoemmp umNWmmzuoax; smzmm; mg» nae: accumumoe o~.~ mv.m eo.m nm.m pm.e mm.¢ Ne.e mm.m momomo mumot momomo macaw moHomo mumot mofiomo apcmm mouomo mamot monomo apcmm momomo mumoa momomo Apgmm a_.m mm.m om.e N_.m humod< zoo pumdoa emeze pumeae zoo eumuaa em_=e m~.m mo.m emmempzo zoo pmmmmezo emcxr em.m A_oepnzo opeEam oeeeem 134 significant (t=2.208; p < .05). In Figure 12 the difference in level of political information between High and Low INTEREST is in the hypothesized direction, and is statistically Significant (t=7.96; p < .05). One of the two com- parisons between High and Low AFFECT citizens is in the hypothesized direction; however, neither difference is statistically significant. Three of the four comparisons between early and late deciders are in the wrong direction. However, the one correct "prediction," the comparison between early and late deciders in the High INTEREST/Low AFFECT category, is also the only difference which is significantly different from zero (t=2.475; p < .05). In Figure 13 the difference in information levels between High and Low INTEREST citizens is in the hypothesized direction, and is statistically significant (t=2.36; p < .05). One of the two compar- isons between High and Low Utility citizens is in the opposite direction hypothesized (in the Low INTEREST category), and is also significantly different from zero (t=2.ll; p < .05). Only two of the four comparisons between early and late deciders are in the hypothesized direction, and none are significantly different from zero. In the last figure, Figure l4,the difference in political information levels between High and Low INTEREST respondents is in the hypothesized direction, and is statistically significant (t=lO.15; p < .05). Neither of the High/Low AFFECT comparisons are signifi- cantly different from zero, although both are in the hypothesized direction. Finally, the comparison between early and late deciders in the High INTEREST/Low AFFECT category is the only difference in 135 the "correct" direction, and is the only difference that is not statistically significant. For the other three comparisons between early and late deciders, the t values are, reading across the bottom row, and skipping the High INTEREST/Low AFFECT category, t=-2.682, t=-l.97, t=-2.92. In the four figures there are sixteen comparisons between early and late deciders in terms of levels of political information. Seven of these are in-the right direction, with two being statistically significant. Nine are in the opposite direction hypothesized, with three being statistically significant. The data indicate that Hypothesis 3 is rejected. Additional analysis of the data will not alter this conclusion. However, examining the comparisons in a dif- ferent form will yield a clearer understanding of the combined impact of political interest, differential benefit, and timing of the vote decision on levels of political information. To this end, in Tables 7-l0 I have rank-ordered each combination of INTEREST/UTILITY (or AFFECT/DECIDE)by mean level of political information. Arranged in this fashion, several interesting patterns emerge. First, the High INTEREST category appears consistently in the top four rankings of each table. In Tables 8 and l0 the top four rankings with respect to mean levels of political information are taken by the four possible High INTEREST categories. In Tables 7 and 9 three of the t0p four rankings are occupied by High INTEREST categories. In all four tables the highest ranking is taken by the High INTEREST category. Turning to the two measures of expected dif- ferential benefit (UTILITY and AFFECT), the top four rankings are evenly split between High and Low AFFECT/UTILITY in three of the four 136 Table 7. Rank—Orderings of Average Candidate-Information Levels. Rank Category@ Mean N 1* LD/HU/HI 8.38 28 2 LD/LU/HI 7.53 22 3 LD/LU/LI 7.29 29 4 ED/LU/HI 7.0 48 5 ED/HU/LI 6.59 44 6** ' ED/LU/LI 6.5l 53 7 ED/HU/HI 6.25 31 8 LD/HU/LI 6.2 30 *Highest hypothesized category **Lowest hypothesized category MB. L0: Late Deciders ED: Early Deciders HU (HA): High UTILITY (AFFECT) LU (LA): Low UTILITY (AFFECT) HI: High INTEREST LI: Low INTEREST 137 Table 8. Rank-Orderings of Average Candidate Information Levels. Rank Category@ Mean N l LD/LA/HI 7.l5 109 2 ED/HA/HI 6.26 l84 3 ED/LA/HI 6.07 ll6 4* LD/HA/HI 6.06 lZl 5 ED/HA/LI 5.24 345 6** ED/LA/LI 5.08 305 7 LD/LA/LI 4.89 175 8 LD/HA/LI 4.86 249 *Highest hypothesized category **Lowest hypothesized category @551 L0: Late Deciders ED: Early Deciders HU (HA): High UTILITY (AFFECT) LU (LA): Low UTILITY (AFFECT) HI: High INTEREST LI: Low INTEREST l38 Table 9. Rank-Orderings of Average Party Information Levels. Rank Category@ Mean N l LD/LU/HI 6.58 22 2 ED/HU/HI 6.38 3l 3* LD/HU/HI 5.76 28 4 LD/LU/LI 5.53 29 5 ED/LU/HI 5.42 48 6** ED/LU/LI 5.36 53 7 ED/HU/LI 4.34 44 8 LD/HU/LI 4.0 30 *Highest hypothesized category **Lowest hypothesized category @551 LD: ED: Late Deciders Early Deciders HU (HA): High UTILITY (AFFECT) LU (LA): Low UTILITY (AFFECT) HI: LI: High INTEREST Low INTEREST l39 Table lO. Rank-Orderings of Average Party Information Levels. Rank Categorycd Mean N l ED/HA/HI 5.58 109 2 LD/LA/HI 4.9l l84 3 ED/LA/HI 4.89 ll6 4* LD/HA/HI 4.42 l2l 5 ED/HA/LI 3.57 345 6** ED/LA/LI 3.48 305 7 LD/HA/LI 3.04 T75 8 LD/LA/LI 2.70 249 *Highest hypothesized category **Lowest hypothesized category @551 L0: ED: Late Deciders Early Deciders HU (HA): High UTILITY (AFFECT) LU (LA): Low UTILITY (AFFECT) HI: LI: High INTEREST Low INTEREST 140 tables. In Table 7 three of the top four rankings are taken by the Low UTILITY category. However, the highest ranking is occupied by the High UTILITY category. With regard to the timing of the vote decision, DECIDE, the results are mixed, but overall late deciders appear to have higher levels of political information. In Tables 8 and l0 the top four rankings are evenly split between early and late deciders. In Tables 7 and 9 three of the top four rankings are occupied by late deciders. Further, in all of the tables except Table ID the highest ranking spot is taken by late deciders. Overall, levels of political interest appear to be predominant in determining levels of political information when the relative rankings of the categories are considered. Considering the four tables as a whole, fourteen of the sixteenhighest are taken by the High INTEREST categories. Seven of the top sixteen spots are occu- pied by the high differential benefit (UTILITY or AFFECT) categories, and ten of the top sixteen rankings are taken by late deciders. While these data must be interpreted cautiously, they do suggest that an outright rejection of Hypothesis 3 may be an overly severe inference. Hypothesis 4: The earlier in the campaign a citizen decides whether and/or how to vote, the more likely will that citizen be to use television as a source of political information than will a citizen who decides later in the campaign. Hypothesis 4a: The later in the campaign a citizen decides whether and/or how to vote, the higher will be that citizen's l4l utilization of any one medium for gathering political information. Hypothesis 4 states that early deciders are more likely to use television as a primary source of political information than are late deciders. To test this hypothesis, V3666 was correlated with V3654, which compares newspaper and television utilization as sources of political information. Due to the construction of V3654, de- scribed earlier in this chapter, a negative relationship is expected. The Spearman's rho calculated equals .013, with p=.302. The coeffic- ient is not significantly different from zero, and is in the opposite direction hypothesized. The hypothesis is disconfirmed. Hypothesis 4a states that the later the timing of the vote decision, the more any one medium will be utilized to gather politi- cal information. To test this hypothesis, the timing of the vote decision variable (V3666) was correlated with variables measuring the extent to which radio, magazines, and television were used to monitor the campaign (V3601, V3603, and V3605, respectively), and with the variables POLPAPER and NEHSVIEH, defined earlier in this chapter. Due to the categorization of the first three variables, negative re- lationships are hypothesized. For the latter two variables, positive relationships are expected. The relevant data are presented in Table ll; the correlation coefficients are Spearman's rho's. The relationships reported in Table ll are weak, at best. Further, the only relationship that is statistically significant is the relationship between degree of radio utilization (V3601) and the timing of the vote decision. In turn, this coefficient is in the opposite direction hypothesized. The null hypotheses cannot be 142 rejected,.and Hypothesis 4a is disconfirmed. Table ll. Relationships between Media Utilization and Timing of the Vote Decision. Timing of the Vote Decision Y§§Ql_(radio use) r=.099* Y§§Q§_(magazine use) r=.Ol6 X3§Q§_(television use) r=.O4l POLPAPER =.0l4 NEWSVIEW =-.02 *p < .05 The tests of Hypotheses 3, 4 and 4a indicate that, in general, late deciders are not necessarily better informed than early deciders: in fact, the data indicate that virtually no relationship exists be— tween timing of the vote decision and levels of political informa- tion. Further, there appears to be no relationship between the timing of the vote decision and medium used as a source of informa- tion and degree of media utilization. In short, the data disconfirm these hypotheses drawn from the theory. The concept of citizen uncertainty is examined in Hypotheses 5- 9. Recall that three measures of uncertainty regarding the candi- dates' issue positions were constructed: UNSURCAR and UNSURFOR, which measure uncertainty regarding the issue stands of Carter and Ford, respectively; and, UNSUR76, which measures total uncertainty. UNSUR76 was constructed by adding UNSURCAR and UNSURFOR together. l43 In testing Hypotheses 5-9 substantial differences emerged in the relationships between uncertainty and information, time of the vote decision, etc., depending upon whether UNSURCAR or UNSURFOR was used as the measure of uncertainty. In order to focus the discussion on these differences, UNSUR76 is not used in the following analyses. However, UNSUR76 will be used to measure uncertainty in the model evaluation exercise presented in Chapters 7~9. Hypothesis 5: Uncertainty regarding a candidate's issue positions is negatively related to a citizen's level of political information. The reader will recall that two measures of uncertainty were constructed, one measuring the uncertainty regarding Carter's issue positions (UNSURCAR), and one measuring the uncertainty regarding Ford's issue positions (UNSURFOR). With the two measures of politi— cal information, four relationships can be examined to test the hypothesis. Table l2 presents the relevant data; as before, the coefficients are Pearson's r's. Table l2. Relationships between Uncertainty and Information. PARTINFO CANDINFO UNSURCAR =-.109* r=-.l$6* UNSURFOR r=-.0l2 =-.027 *p < .05 144 The results presented in Table l2 are intriguing. While all four coefficients are in the hypothesized direction, only the re- lationships between levels of political information and uncertainty regarding Carter's issue positions are substantively and statistic- ally significant. Levels of political information appear to have no impact upon citizens' levels of uncertainty regarding Ford's issue positions. The theory cannot account for these intriguing findings. However, we may speculate as to the reasons for these results. It is interesting to note that the coefficients linked to the non-incumbent (Carter) are significantly different from zero while those linked with the incumbent (Ford) are virtually non-existent. Prior to gathering any information, it is virtually certain that a citizen will know more about the incumbent than the non-incumbent;4 this was particularly true in 1976. This suggests that the marginal reduction of uncertainty regarding the incumbent per"unit" of information will be lower than the marginal reduction of uncertainty regarding the non-incumbent. That is, a stronger relationship between levels of information and uncertainty would hold for the challenger, relative to the incumbent. Whatever the “true" explanation may be, it is clear that the null hypothesis is rejected for uncertainty regarding Carter's issue positions, while the original hypothesis is disconfirmed regarding Ford's issue positions. Hypothesis 6: Individuals who rely on newspapers as their primary source of political information will exhibit lower levels of uncertainty regarding a candidate's issue positions than will individuals who rely on television as their primary T45 source of political information. Hypothesis 7: Levels of uncertainty regarding candidates' issue positions are negatively related to the degree to which any one medium is utilized as a source of political information. To test Hypothesis 6, UNSURCAR and UNSURFOR were correlated with V3654 (the media choice variable). Due to the categorization of the latter variable, a positive coefficient indicates that television users are more uncertain than newspaper readers. The coefficients calculated are Spearman's rho's. Correlating uncertainty regarding Ford's issue positions and V3654 yields r=.lOO (p < .05). Correlat- ing V3654 and uncertainty regarding Carter's issue positions yields r=-.03O (p=.283). The data indicate that uncertainty regarding Carter's issue positions is invariant with respect to medium utilized as the primary source of political information. However, newspaper readers exhibit significantly lower levelsof uncertainty regarding Ford's issue positions than do television viewers. Again, these anamolous findings can be accounted for, but only outside the theory. If a citizen has a substantial amount of informa- tion regarding the incumbent, "richer" information will be needed to reduce uncertainty regarding the incumbent's issue stands. Further, I have argued that newspapers do provide "richer" information than does television. On the other hand, if a citizen possessed little in- formation about the non-incumbent, as argued above, any information, "rich" or otherwise, would reduce uncertainty. In turn, uncertainty would be reduced (or remain at some level, or increase) regardless of the medium utilized to gather political information. Thus, one might expect to find that levels of uncertainty regarding the non- T46 incumbent's issue positions to be invariant with respect to medium utilized. Whatever the explanation may be, the data do indicate that the hypothesis regarding Carter's issue positions is disconfirmed, while the null hypothesis regarding Ford's issue positions may be rejected. To test Hypothesis 7, UNSURCAR and UNSURFOR were correlated with the five media utilization variables described earlier in the chapter. Spearman's rho's were calculated for each relationship. The relevant data are presented in Table l3. Table l3. Relationships between Levels of Uncertainty and Degree of Media Utilization. UNSURCAR UNSURFOR POLPAPER =-.O92* r=-.051 NEWSVIEW r=.026 r=.064 V3601 (radio use) r=.015 r=-.028 V3603 (magazine use) r=-.055 r=-.052 V3605 (television use) r=-.076 =-.l00* *p < .05 On the basis of these data, two statements can be made. First, uncertainty regarding Carter's issue positions decreases with in creased utilization of newspapers for gathering political news (POLPAPER). Second, uncertainty regarding Ford's issue positions de- creases with increased utilization of television to monitor the cam- paign. All other coefficients are either trivial or statistically insignificant, or both. In general, the data do not support the hypothesis. 147 Hypothesis 8: The greater is a citizen's uncertainty regarding the candidates' issue positions, the less likely it will be for the citizen to have voted. To test this hypothesis, the two uncertainty variables were correlated with V3655, the post-election wave turnout question. Spearman's rho's were calculated. Given the categorization of V3655, a positive coefficient indicates support for the hypothesis. Cor- relating V3655 with UNSURCAR yielded r=.096 (p < .05). Correlating V3655 with UNSURFOR yielded r=.027 (p=.269). Both coefficients are in the hypothesized direction, but only the coefficient for V3655 and UNSURCAR is significantly different from zero. The null hypothe- sis is rejected for the UNSURCAR/V3655 test, but is not rejected for the UNSURFOR/V3655 test. Hypothesis 9: The later in the campaign a citizen comes to a voting decision, the less uncertain will that citizen be re- garding the candidates' issue positions. To test this hypothesis, the two uncertainty variables were correlated with V3666, the timing of the vote decision variable. Pearson's r's were calculated; a negative coefficient indicates support for the hypothesis. Correlating UNSURCAR and V3666 yielded r=-.l03 (p < .05). Correlating uncertainty regarding Ford's issue positions and V3666 yielded r=-.067 (p=.O88). Both coefficients are in the right direction, but the UNSURFOR/V3666 coefficient is not significant at the .05 level of probability. The later in the campaign the citizen comes to a voting decision, the more certain l48 he/she will be about Carter's issue positions. However, late decid- ers will be no more or less certain about Ford's issue positions. Presumably, citizens who arrive at voting decisions later in the campaign have spent more time gathering information about the candidates, and thus should exhibit lower levels of uncertainty. Given that no relationship was found between levels of political information and uncertainty regarding Ford's issue positions, the lack of a relationship between UNSURFOR and the timing of the vote decision should come as no surprise. Hypothesis lO: The higher is a citizen's level of Expected differential benefit for a candidate, the more likely it will be for the citizen to have voted for that candidate. This hypothesis was examined using the two expected differential benefit measures, DIFFBEN (based on the seVen-point issue scales) and DIFFEEL (based on the feeling thermometers), and the post-election wave vote choice question, V3665. Due to the construction of the benefit variables, a positive coefficient between DIFFBEN and V3665 indicates support for the hypothesis; a negative coefficient between DIFFEEL and V3665 indicates support for the hypothesis. Spearman's rho's were calculated. The data are presented in Table l4. Table l4. Relationships between Vote Choice and Expected Differential Benefit. V3665 DIFFBEN r=.632* DIFFEEL r=-.699* *p < .05 l49 Clearly, the null hypotheses are rejected for both tests. Levels of expected differential benefit are strong predictors of vote choice. Hypothesis ll: Citizens who rely on newspapers as their primary source of political information are more likely to have voted than citizens who rely on television as their primary source of political information. Hypothesis lla: The relationship between media choice and voting will diminish when the effects of education, income, social class, and/or strength of party identification are held constant. Testing the first hypothesis involves correlating the media choice variable, V3654, with the post-election wave turnout question, V3655. Spearman's rho was calculated; a positive coefficient confirms the hypothesis. Correlating V3654 and V3655 yielded r=.ll4 (p < .05). The data allow rejection of the null hypothesis. The second hypothesis states that this relationship between media choice and turnout will weaken when the effect of several socio-political variables are held constant. The relevant partial coefficients are presented in Table l5. The data in Table l5 clearly indicate that the relationship between media choice and turnout weakens once the effects of educa- tion, income, social class, and/or strength of partisan identifica- tion are held constant. The data allow rejection of the null hypothesis. l50 Table l5. Test of Hypothesis lla. Controlling For: Partial Correlation between V3654 (Media Choice) and V3655 (Turnout) Education r=.073* Income r=.085* Social Class r=.O84* Strength of Party I.D. r=.l08* All four Variables r=.O4* *p < .05 Hypothesis l2: Utilization of television as a primary information source is positively related to voting. To test Hypothesis l2, V3655, the post-election wave turnout question, was correlated with V3605, the number of television pro- grams about the campaign watched by the respondent; and NEWSVIEW, which measures the degree to which respondents watch television news programs. A positive coefficient for the V3655/V3605 relationship confirms the hypothesis. Due to the construction of V3655, the turnout question, a negative coefficient for the V3655/NEWSVIEW re- lationship confirms the hypothesis. Spearman's rho's were calculated; the relevant data are presented in Table 16. Table l6 indicates that the more television programs about the campaign an individual watches, the more likely it is that the indi- vidual voted. Further, the more an individual watches television news programs, the more likely he/she is to have voted. In general, the data confirm the hypothesis. 151 Table l6. Relationship between Television Utilization and Voting. V3655 (Turnout) V3605 (Television use) r=.l9l* NEWSVIEW =-.lZl* *p < .05 Hypothesis l3: Voters are better informed than non-voters. To test this hypothesis, the two political information variables, PARTINFO and CANDINFO, were correlated with V3655, the post—election voting question. Spearman's rho's were calculated; due to the con- struction of V3655, a negative relationship confirms the hypothesis. The relevant data are presented in Table l7. Table l7. Relationship between Levels of Information and Voting. V3655 (Turnout) PARTINFO =-.248* CANDINFO r=-.263* *p < .05 Little discussion of these data is needed. The relative strength and statistical significance of the coefficients permit a straightforward rejection of the null hypothesis. Voters are better informed than non-voters. l52 Hypothesis l4: Vote choice is invariant with respect to levels of political information. The hypothesis states that levels of political information have no impact on an individual's voting choice between the two Presidential candidates. To test this hypothesis the two political information variables were correlated with V3665, the post-election wave candidate's choice question. Due to the categorization of V3665, a negative coefficient indicates a preference for Ford. Spearman's rho's were calculated. The relevant data are presented in Table l8. Table l8. Relationship between Level of Political Information and Vote Choice. V3665 (Candidate's Choice) PARTINFO =-.044* CANDINFO r=.024 *p < .05 The weak but statistically significant relationship between V3665 and PARTINFO indicates that individuals with higher levels of party- referenced political information were very slightly more likely to vote for Ford. The correlation between CANDINFO and V3665 indicates that levels of candidate-referenced political information had no impact upon the direction of the vote. In general, the data indicate that an individual's vote choice is invariant with respect to levels of political information. l53 Hypothesis l5: Television is a more frequently utilized source of information regarding politics than are newspapers. "Testing'I this hypothesis merely involves examining the distri- bution of V3654, the media utilization variable. Percentage and frequency distributions are presented in Table l9. Table l9. Distribution of V3654.* N % Newspapers 448 l9.3 Television l490 64.2 Both Equally 383 l6.5 *The question asked on the T976 CPS Survey was "Which do you rely on most for news about politics and current events - newspapers or television?" Television is clearly the preferred medium for monitoring a political campaign. Even if the "newspapers" and "both equally" categories are combined, television is still preferred to newspapers by a substantial margin. The hypothesis is confirmed. Hypothesis l6-l: Citizens of a higher social class are more likely to utilize newspapers as a primary source of political information than are citizens of lower social class. Hypothesis 16-2: Citizens with high levels of education are more likely to utilize newspapers as a primary source of polit- ical information than are citizens with lower levels of education. l54 Hypothesis l6-3: Citizens with high levels of income are more likely to utilize newspapers as a primary source of political information than are citizens with lower levels of income. Hypothesis l6-4: When levels of political information are held constant, the relationships between media choice and education, income, and social class will diminish. \ To test Hypotheses l6-l-3, V3654, the media choice variable, was correlated with V3486, a summary social class measure; V3384, the highest year of schooling completed by the respondent; and V3507, the level of respondent's family income. Due to the construction of V3654, negative coefficients indicate support for the hypotheses. To test Hypothesis l6-4, levels of political information were held constant by controlling for CANDINFO. Spearman's rho's were calcu- lated; the relevant data are presented in Table 20. Table 20. Relationships between Media Choice and Social Class, Education, and Income, Controlling for Levels of Polit- ical Information. Zero Order Controlling for CANDINFO V3654/Social Class r=-.228* r=-.187* V3654/Education r=-.20l* =-.l67* V3654/Income r=-.ZOl* r=-.l77* *p < .05 The data in Table 20 indicate that all four null hypotheses may be rejected. The data indicate that social class, education, and l55 income impact upon the citizen's media choice, and that controlling for levels of political information weakens these relationships. Further, Hypotheses l6-l-3, taken in conjunction with Hypotheses l and 2, state there is a reciprocal relationship between levels of political information and media choice (television vs. newspapers), and between levels of political information and media utilization. Misperception,yPartisanship, and Information In this section, I return to the concept of reinforcement, and inquire into the extent and nature of reinforcing behavior during the l976 Presidential campaign. There are, generally speaking, three psychological processes that are considered to be reinforcement mechanisms: Selective perception, selective exposure, and selective retention. Given the measures on the CPS l976 survey, the latter two concepts would be most difficult to operationalize. However, the concept of selective perception can be operationalized using the same seven-point issue scales used earlier to measure expected differen- tial benefit and uncertainty. Therefore, attention will be focused on the extent and nature of selective perception during the 1976 Presidential campaign. Three measures of selective perception will be used. The first operationalization is as follows: N z 'xi'eil'lxi'exil i=l N where Xi = the respondent's self-placement on the ith issue scale; 156 o. = the respondent's placement of candidate 0 on the ith issue scale; >5? the average placement of all respondents of candidate 6 on the ith issue scale; N = the number of issue scales on which the respondent placed both him/herself and both candidates. The major underlying assumption of this (and the following) measures is that, as with the measure of uncertainty, the aggregate perception of candidates' issue positions are accurate. Thus, the average respondent placement of a candidate on an issue scale is taken to be that candidate's “true" position on that issue. The absolute difference between the "true" position and the respondent's self-placement on an issue scale is taken. This indicates how far the respondent's position on an issue is from the candidate's "true" position. Then, the absolute difference between the respondent's self-placement and the respondent's placement of the candidate on an issue scale is taken. This indicates how "far away" the respondent perceives the candidate to be from the respondent's own issue posi- tion. Finally, the absolute difference between these two calcula- tions is taken, and the average misperception is calculated by dividing by the number of valid issue scales. The result is two measures, one for each candidate, of the extent to which a respondent selectively perceived candidates' positions on the various issue scales. The variables will be called MISCART and MISFORD. The lowest possible value of this measure is 0, which can occur only when ei=OYi for all N issue scales; a score of 0 indicates no misperception whatsoever. The maximum value of the measure is 6. A 157 value of 6 will occur only if, on every issue scale, the respondent places himself at one extreme on a seven-point issue scale (i.e., 1 or 7); places the candidate at the other extreme (i.e., 7 or 1); and the "true" position of the candidate is the respondent's own position i.e., if OYi=xi' The probability that this measure would equal 0 or 6, for any respondent, is virtually nil. This measure has one defect. It measures the total average misperception of respondents, but does not measure the direction of this misperception. Another variable, which measure the direction of the difference between the respondent's perception of the difference between himself and the candidate (Xi-9i) and the "actual“ difference (X.-oY ), is needed. The following operationalization fills this 1 i need. 2( IX-i’e-i I'IX-i'e'x'in N Where all the assumption, symbols and calculations (except for one) are as with the previous measure. The reader will note that the only difference between this measure and the previous operationaliza- tion is that the final difference taken is an aggual_value, not an absolute value. Thus, the measure indicates the direction of the misperception. However, in rectifying one problem, another is created. Given that actual values of differences are taken, mis- perceptions in opposite directions may cancel out. Thus, while this measure indicates the average direction of misperception, it is not an indicator of the total amount of misperception. Hence, both measures are needed. 158 As before, two measures are calculated, one for Carter and one for Ford.' These are labeled, respectively, SELECTCART and SELECTFORD. These measures have a range of -6 to 6. A score of zero, as with the previous measure, indicates no overall misperception. The same conditions which lead to a score on the first measure of zero also lead to a score of zero on this measure. Thus, for example, MISCART= 0 only when SELECTCART=O, and vice versa. Hence, zero values of this second measure are as equally unlikely as for the first measure. In turn, the probability of observing the extreme values of this variable, -6 and 6, approach zero, since these values require total misperception on the part of the respondent. In general, negative values indicate that a respondent perceives a candidate to be closer to the respondent than he "actually" is, while positive values indicate that a respondent perceives a candidate to be further away from the respondent than he actually is. Finally, a third misperception variable is constructed: All symbols have been previously defined. For each issue scale, the average respondent placement of a candidate is subtracted from the respondent's placement of the candidate. The sum is taken over all issue scales on which the respondent placed himself and the candi- date. The sum is then divided by this number of issue scales. The variables measures the direction of respondent misperception of candidate issue positions in terms of the liberal/conservative 159 continuum of the seven—point issue scales. With this measure it can be determined whether a respondent perceived a candidate to be more liberal or more conservative than the candidate "actually" was on an issue. A score of zero indicates no overall misperception. Total misperception is indicated by scores of -6 and +6. Negative values indicate the respondent perceived the candidate to be more liberal than the candidate actually was (i.e., the respondent placement is lower on a seven—point scale than the average respondent placement). Positive scores indicate the perception that the candidate was more conservative than actually was.5 As before, two variables are constructed, one for each candidate. These variables are called PERCART and PERFORD. Below, the relationships between levels and directions of misperceptions, levels of information, and partisanship will be explored. Before that, it will be usefulto look at the extent of misperception during the 1976 campaign. Some simple descriptive statistics are given in Table 21. Table 21. Means and Standard Deviations of Misperception Variables. Mean Standard Deviation MISCART 1.135 .707 MISFORD 1.232 .744 SELECTCART -.l79 .986 SELECTFORD .029 1.079 PERCART .001 .985 PERFORD .056 .995 160 Given the construction of MISCART and MISFORD, Table 21 indicates a moderate amount of total misperception by respondents during the 1976 campaign. Further, respondents do not seem to have been more likely to have misperceived Ford's than Carter's issue positions. The mean values of SELECTCART, SELECTFORD, PERCART and PERFORD are lower than those for MISCART and MISFORD. This results from the fact that they are measuring the direction of misperceptions, and misperceptions in opposite directions will tend to cancel out. Further, the size of the standard deviations for these measures, relative to the means, indicates a wide range of misperception by respondents. On average, respondents tended to misperceive Carter's issue positions as being closer to their own positions than Carter's stands actually were (i.e., than the mean respondent perception of Carter). On the other hand, according to the SELECTFORD measure, respondents do not seem to have selectively perceived Ford's issue positions. The greater misperception of Carter's issue positions may be a reflection of an advantage for the non-incumbent, which was previously discussed regarding levels of uncertainty. Respondents not only had Ford's campaign materials to rely on, but also his past performance in office. On the other hand, information regarding Carter's issue stands, competence, etc., derived almost totally from his pre-nomination and general election campaigns. Given the greater amount of information available about Ford, it is not surprising that respondents were able to selectively perceive Carter's issue posi- tions to a greater degree, relative to Ford's issue positions. 161 Carter's campaign strategy may also have had a bearing on the relative levels of selective perception. At various points in the campaign, Carter was accused of being "fuzzy" on the issues, of "waffling" on the issues, and of "trying to be all things to all people." This strategy, in conjunction with the relatively greater amount of information regarding Ford's experience, issue positions, etc., almost surely increased the opportunities of individuals to misperceive Carter's issue positions. PERCART and PERFORD measure the direction of misperceptions in terms of a liberal/conservative continuum. According to the PERCART measure, respondents did not perceive Carter's issue posi- tions as being more liberal or conservative than those positions actually were. Ford's policy stands were perceived as being slightly more conservative than"in actualityJ'but this tendency is not strong. In general, misperceptions of Carter's and Ford's issue positions were not arrayed along a liberal/conservative continuum. The data by themselves do not mean much. If a citizen engages in reinforcing behavior, in this instance by selectively perceiving candidates' issue positions, such behavior serves a purpose. In elections that purpose is, by assumption, to increase the certainty that the citizen's vote choice isva correct one, i.e., maximizes expected utility. As discussed earlier, a majority of citizens in presidential elections, including the election of 1976, make their vote choices prior to the start of the campaign. 'Given that the CPS survey pro- cess begins after the campaign has begun, it can be argued that the vote choice question on the survey does not measure a voting decision 162 that will be reinforced, so much as it measures a vote decision that has already been reinforced. However, the respondent's partisan identification can be used as an indicator of a respondent's pre- campaign vote choice. As noted in Chapter 1, the evidence of selective perception in presidential elections is quite impressive (cf., Berelson, et a1., 1954; Lazersfeld, et a1., 1968; Sherrod, 1971). Sigel's study of the 1960 presidential contest provides strong support for the proposition that partisan affiliation affects perceptions of the candidates (Sigel, 1972). Patterson's analysis of the 1976 election also demonstrates the existence of selective perception by individuals (Patterson, 1980:86-89). If it is correct to treat one's partisans affiliation as a standing vote choice which individuals will try to reinforce, then these studies suggest that people selectively per- ceive candidate characteristics in order to reinforce pre-existing vote choices. Virtually all studies of selective perception in electoral contexts, especially those noted above, deal with perceptions of candidate images. In this section the analysis is concerned with perceptions of candidate issue positions. While the objects of selective perception differ for these two approaches, it is reason- able to suggest that patterns of misperceptions regarding issue positions will be similar to those regarding candidate images. Therefore, the following hypotheses, drawn from the aforementioned literature, will be examined: - Democrats perceived Carter's issue positions as being closer to their own than did Republicans; 163 - Republicans perceived Ford's issue positions as being closer to their own positions than did Democrats; - There was no differences between Democrats, Republicans and Independents in terms of the extent to which the candidates' issue positions were misperceived. No hypotheses regarding partisanship and the ideological direction of misperceptions will be offered. It is not clear what form these hypotheses should take, nor what would constitute dis- confirmation. For instance, can it be hypothesized that conservative Democrats misperceived Carter's issue stands toward the liberal end of a seven-point continuum? Yes, but only if that individual was attempting to reinforce a vote choice for Ford. If the citizen was attempting to reinforce a vote choice for Carter, that individual would (probably) have misperceived Ford toward the liberal position, and Carter toward the conservative position. In short, presenting hypotheses regarding the relationships between partisanship and the variables PERCART and PERFORD could lead to a never-ending chain of explanation and counter-explanation. Therefore, no hypotheses will be offered; however, an analysis of partisanship and ideological misperception will be presented below. To begin this analysis, some relevant descriptive statistics are displayed in Table 22. According to the MISCART and MISFORD measures, Republicans were no more likely to misperceive Carter's than Ford's issue positions, while Democrats and Independents exhibited slight tendencies to mis- perceive Carter's issue positions. However, the differences in mean 164 mom. mwm. muo._ mmm. ecu. men. .o.m mmo. —oo. mmo. mN—.l NMN.F mmp._ cam: mhzmozommmm 44< mmm. moo.~ mmo.P Noo._ mun. mmn. .o.m cmo.u wmp. mmo.a NNN.T FPN.P ¢p_._ cum: mhzmozmmmozm com. vem. mvm. Pom. coo. mmo. .o.m po. ¢¢~.- mkm.n Ncp. mm—.P Nm—.— 2mm: mz compamucmamwz vo mcowumw>mo ucmucmum ucm meow: .NN wpamh 165 levels between partisan groups are not large. As hypothesized, the differences between Republicans, Independents and Democrats in terms of total misperceptions are virtually non-existent. When the directions of misperceptions are considered, differen- ces between partisan groups emerge. In confirmation of the first two misperception hypotheses, Democrats perceived Carter's issue stands to be closer to their own, and Ford's policies to be further away from their own, than the candidates' positions were on average. Further, these tendencies were stronger among Democrats than among Republicans or Independents. In turn, Republicans perceived Ford's issue stands as being closer to their own, and Carter's as further away, than they were on average. Finally, Independents perceived both candidates' issue posi- tions as being closer to their own positions than they were on aver- age, with the tendency being to misperceive Carter's positions to a greater extent than Ford's positions. According to the full sample means of PERCART and PERFORD, there was no ideological misperception of the candidates' policy stands in 1976. As with the other misperception variables, however, differ- ent patterns emerge when the full sample of respondents is catego- rized along partisan lines. First, Democrats perceived both Carter's and Ford's issue positions as being more conservative than they were on average. Further, there was a greater tendency to misperceive Carter's positions in the conservative direction by Democrats. Only a very tentative explanation can be offered for these results: Liberal Democrats may have been reacting to Carter's Southern origins, and as a consequence, misperceived his issue stands on the 166 conservative side. Unfortunately, this does not explain why Ford's issue stands were also misperceived in the conservative direction. Republicans and Independents did not misperceive Ford's issue positions in the conservative direction. These groups accurately perceived the incumbent's issue positions. However, Republicans did perceive Carter's stands as being more liberal than they were on average. If Republicans are more conservative than Democrats and the general population, then these data indicate that Republicans were reaffirming a standing vote choice for Ford. Finally, Independents were similar to Democrats in that they perceived Carter's issue positions as being more conservative than they were on average. Again, exactly what these data mean is unclear. However, the general pattern of misperceptions during 1976 seems to have benefited Ford at the expense of Carter. To complete this analysis of the relationship between misperception and tendency to vote for one particular candidate, the misperception variables were correlated with DIFFBEN and DIFFEEL. These correlation coefficients (Pearson's r's) are reported in Table 23. The independent variables DIFFBEN and DIFFEEL refer, respectively, to expected utility based on the seven-point issue scales, and to expected utility based on the one-hundred-point feel- ing thermometers. For DIFFBEN, positive scores indicate support for the Democratic candidate, while negative scores indicate support for Ford. For DIFFEEL, negative scores indicate support for Carter, while positive values indicate support for the Republican candidate. 167 Table 23. Relationships between Expected Utility and Selective Perception. DIFFBEN DIFFEEL MISCART .O38* -.O42* MISFORD .199* -.150* SELECTCART -.594* .368* SELECTFORD .613* -.459* PERCART .216* -.l78* PERFORD .018 -.O48* *p §_.05 The data in Table 23 indicate that there was virtually no relationship between level of expected utility and the tendency to misperceive Carter's policies. However, the relationships between expected utility and MISFORD suggest that the more favorably disposed toward Carter were respondents, the more likely they were to mis- perceive Ford's issue positions. Further, the higher was expected utility for Ford, the more likely were citizens to misperceive the incumbent's policy stands. The relationships between the expected utility measures and SELECTCART and SELECTFORD confirm the hypotheses presented earlier, and are rather impressive in their strength. Specifically, - As support for Carter increased, the tendency to perceive his issue stands as being closer to one's own increased; - As support for Carter decreased, the tendency to perceive his issue stands as being further away from one's own increased; 168 - As support for Ford increased, the tendency to perceive his issue positions as being closer to one's own increased; - As support for Ford decreased, the tendency to perceive his issue positions as being further away from one's own increased. In sum, these data describe a classic case of reinforcement. Supporters of both candidates tended to perceive their favorite candidate as being closer to them on the issues than they actually were, and to perceive the opposition candidate as being more distant on the issues than he actually was.6 Finally, the ideological misperception measures indicate that as support for Carter increased, the tendency to perceive his poli- cies as more conservative than they actually were increased. Further, as support for Carter decreased, the tendency to perceive him as more liberal than in reality increased. However, the tendency to perceive Ford as either more liberal or more conserva- tive than in actuality was not related to support for the incumbent. I turn now to a consideration of the relationships between levels and direction of misperceptions of candidates' issue stands and levels of political information and media use. If it is true that at least one motive of citizens for gathering political infor- mation, through the mass media, is to reinforce their vote choices, then it is reasonable to hypothesize positive relationships between information levels, media use, and levels of misperception. As indi- viduals gather more information or use the media to a greater extent, they are afforded a greater opportunity to selectively perceive the issue stands of the candidates. To examine this proposition, data 169 relevant to misperception and levels of political information are presented in Table 24. The coefficients reported are Pearson's r's. Table 24. Relationships between Misperception Variables and Levels of Candidate-Related and Party-Related Information. CANDINFO PARTINFO MISCART .041* .016 MISFORD .068* .045* SELECTCART .066* .106* SELECTFORD .013 -.004 PERCART -.123* -.194* PERFORD .035 .156* *p §_.05 In general, the relationships reported are weak. The coefficients associated with the MISCART and MISFORD measures indi- cate a very weak positive relationship between levels of political information and the tendency to misperceive the candidates' issue positions. There is no relationship between levels of information and the perception of Ford's issue stands as being either closer to or further away from respondent's own positions. However, there does seem to be a slight tendency for individuals with higher levels of information to perceive Carter's policies as being further away from the respondent's own issue positions than he was on average Conversely, the lower are information levels, the greater is the tendency to perceive Carter as being closer on the issues to the 170 respondent than he was on average. The relationships in the bottom quarter of Table 24 are the strongest of all (with the exception of the PERFORD/CANDINFO coefficient). These data indicate that as information levels in- creased, the tendency to perceive Carter as more liberal than he lactually wasJ'and to perceive Ford as more conservative than he 'actually was;'increased. When the means of the misperception variables (Table 21) were broken down into partisan groups (Table 22) different patterns of relationships emerged. This same result might occur if the data in Table 24 are treated in the same fashion. The relationships between levels and directions of misperceptions and levels of political information for Republicans, Democrats and Independents are reported in Table 25. As before, coefficients are Pearson's r's. The data in Table 25 indicate that Democrats made little use of information for the purpose of selective perception during the 1976 campaign. Two weak tendencies among Democrats emerge from Table 25: - As Democrats' levels of information increased, the tendency to perceive Ford as further away than he'actually was'from the respondent's positions increased; - As Democrats' levels of information increased, the tendency to perceive Carter as more liberal, and Ford as more con- servative, than they "actually" were increased. Republicans seemed to make better use of information for the purpose of selective perception during the 1976 campaign than did Democrats. The following inferences can be drawn from Table 25 171 mo. V a. som_. 44¢P.- eoo.- *eo_. smeo. m_o. apasam sza wmo. somo.- mpo. .mmo. Koo. amao.- magmacaamuca 4_q_. *Nmm.- *mmp.- som_. 0 some. 6:662_a=aam *Nm_. 4NmO.- 4_o_. ammo. *mo. o_o.- mpatuosmo mmmwmmmm mmo. «mm_.- m_o. «coo. twee. spec. mpasam Ppsa 4NN_.- .mo_.- mo. Pmo. “.0. NNo.- macmacwamuca _oo. «amp.- some“- 4wmo. ammo. *Pm.. m=662_a=amx lems. veao.- same. mmo. mmo. “No.- mgacuosao .mmmmmmmm excamma hx G-l, the equations are overidentified. If K < G-l, i.e., if an equation is underidentified, then the equation is not estimable. For Equation 1, 5. =2 G=2 G-1=l K > G For Equation 2, 6. =8 G=2 G-l=l K > G For Equation 3, 7. =7 G-l=O K > G And, for Equation 4, 8. K=6 186 G-1=1 K > G Thus, according to the necessary condition, all equations in the system are overidentified; estimation is therefore possible. However, Ordinary Least Squares is inappr0priate. Assuming that the system of equations is recursive, i.e., involves no reciprocal relationships, two-stage least squares techniques are arguably the most appropriate technique (Kmenta, 1971:559-564). However, this system of equations is non-recursive; specifically, Equations 1 and 2 involve reciprocal relationships. Therefore, equation-by-equation estimation techniques, such as ordinary least squares, are inappropriate, in that such techniques do not take into account the correlation of error terms across equations. The appli- cation of’ ordinary' least squares to this system of equations would require the assumption that )=0. 37m in 9. E(uiju an assumption which is inappropriate.3 System estimation techniques are required, since such techniques take into account the fact that the disturbance terms are correlated across equations. However, additional estimation problems arise from the presence of Y2, Y4, and Y4* in the system. Y4 and Y4, are dichotomous varia- bles. Y2 is the general media utilization variable; regardless of its operationalization, it is a categorical variable, although in many cases (e.g., when operationalized as degree of media use), Y2 has more than two possible values. Therefore, probit estimation techniques (cf., McKelvey and Zavoina, 1975), developed to estimate 187 equations involving categorical (as opposed to continuous) "left- hand side" variables, must be employed.4 A more detailed discussion of probit analysis follows below. The problem that arises from the necessity of using probit analysis is that, as far as I have been able to determine. there exists few useful system estimation procedures involving probit techniques.5 Following the discussion of probit analysis, I shall return to the problem of estimating the statistical model of the theory of citizen information gathering. Probit Analysis6 Standard regression analysis is an inappr0priate estimation technique when dependent variables are either not conceptualized at the interval level or, if they are so conceptualized, cannot be measured on an interval scale. The basic problem with such variables is that relationships involving categorical endogenous variables are nonlinear. When ordinary least squares is applied to such a re- lationship, the attempt to fit a straight line violates assumptions involving the error term. Specifically, it can neither be assumed that the theoretical error terms are uncorrelated with the independ- ent variables, nor that the variance of the error term is constant across all values of the independent variables, i.e., is homoscedas- tic. Two solutions to the problems are apparent: ". . .assume either a nonlinear model or a different error structure" (McKelvey and Zavoina, 1975:105). However, regardless of whether a nonlinear model is assumed, the error term will still be heteroscedastic. 188 Thus, probit analysis employs the second solution, assuming a dif- ferent error structure. Specifically, probit analysis is built upon the assumption that the theoretical endogenous variable of interest, call it Y, is an interval level variable; however, due to measurement difficulties the observed dependent variable, Z, is categorical. Turning again to McKelvey and Zavoina (1975:105, emphasis in orig- inal), [w]e assume that the variable of theoretical interest is interval level and would, if we could measure it, satisfy a linear model. Due to the inadequate measurement techniques, we only observe an ordinal version of Y, namely 2, for which the linear model is ngt_satisfied. Probit analysis attempts to measure the assumed underlying probability distribution of the categorically measured dependent variable. Unlike ordinary least squares, probit yields estimates of the probability of observing the different values of the dependent variable that are true probabilities, and thus fall in the range of zero to one. The transformation of the linear model to the probabil- ity model is made by assuming a cumulative normal distribution; note that this is but one probability distribution that could be as- sumed (Aldrich and Cnudde, 1975:581). With the assumption that the distribution of the theoretical stochastic term is multivariate normaL probit involves a straightforward application of maximum likelihood procedures. Specifically (Aldrich and Cnudde, 1975:580), [tJhe maximum likelihood criterion is invoked by selecting, as estimates of the true parameters, those values which have associated with them the highest probability of having ob- tained the observed sample data. 189 There are two problems inherent in applying probit estimation techniques. First, the generation of maximum likelihood estimates involves maximizing the logs of the likelihood functions of each equation with respect to the parameters of the explanatory variables. This involves solving for the partial derivatives of the logs of the likelihood functions corresponding to each parameter. To ensure a maximum, it must be shown that the matrix of second partials is nega- tive definite (McKelvey and Zavoina, 1975:107-109), i.e., all second partial derivatives must be negative. A problem arises from the fact that the equations used in this exercise are not linear in the unknowns. Therefore, algebraically determinant methods for solving systems of simultaneous equations cannot be used. Instead, itera- tive procedures must be utilized. Whenever iterative procedures are involved, there is no guarantee that the iteration will converge to a solution. However, in applying probit to Equations 2 and 4, convergence proved not to be a problem: in all cases, the procedure converged in less than ten iterations. The second problem is more serious. There is no guarantee that the iterative process used to generate the maximum likelihood esti- mates will converge to a global maximum, as opposed to a relative maximum, since the equations of first partial derivatives may contain multiple roots. This is a problem common to nonlinear maximization exercises, since the possibility of multiple roots cannot be elimi- nated (McKelvey and Zavoina, 1975:109). Two solutions are apparent: generate all local maxima, and then choose the global maximum from that set; or simply live with the problem. Since the first procedure is so obviously cumbersome, the latter strategy is generally adopted. 190 As noted previously, probit analysis generates maximum likelihood estimates of the parameters of the independent variables. As MLE's, these coefficients have the asymptotic properties of consistency, normal sampling distributions, and efficiency (minimum variance) relative to all other estimators in this class of esti- mators. The ratio of an MLE to its generated standard error approx- imates, in large samples, a standard normal distribution, or a "Z" distribution. Thus, hypothesis tests on the coefficients can be performed. In general, the interpretation of the MLE‘s generated by probit are not identical to that of the 8's generated by OLS. In particu- lar, a 8 generated by OLS represents the change in the observed value of the dependent variable for each unit change in the observed value of the independent variable(s). Given that the assumed actual values of the dependent variable are not observed in probit analysis, this interpretation breaks down. Specifically, the maximum likeli— hood estimates generated by probit are interpreted as the amount of change in the dependent variable on its assumed underlying scale for each unit change in the independent variable(s). Given the probability interpretation of probit, this translates, in terms of the observed data, to the increase in probability of being in a higher category of the dependent variable associated with a unit change in the predictor variable (McKelvey and Zavoina, 1975:114). Furthermore, this change in the probability of being in a higher category of the endogenous variable is not constant across all values of the predetermined variables. This is due to the fact that the assumed probability distribution is nonlinear (specifically, 191 cumulative normal). Thus, strictly speaking, the interpretations of the unstandardized betas generated by OLS and the MLE's generated by probit are different. However, little confusion should arise when interpreting the coefficients from probit analysis. Goodness of Fit of the Probit Model When applying OLS, the most common method of determining the goodness of fit of a model is to use R2, the proportion of the variation of the dependent variable "explained“ by the independent variables. Calculating R2 involves merely taking the ratio of the explained sum of squares to the total sum of squares. Probit analysis generates an analogous goodness of fit measure, R2; however, it is only an estimate of the "true" variation explained, since the variance of the dependent variable on its "true" underlying scale is unknown. Calculating R2 for probit models involves the following procedures. Given that the dependent variable is normalized, the variance around the regression line, 02, is equal to unity. Thus, the residual ("unexplained") sum of squares is always equal to the number of observations. The explained sum of squares can be calcu- lated as: x N x x 10. 5%: (Y.-V)2 ., 1 1-1 .aN . where Y=2 Yi/N i=1 R2 is then calculated as 11. R = where N equals the number of observations Two problems arise from the construction of this goodness of fit measure. First, given that the residual sum of squares is defined as N, R2 can never equal 1.00. That is, 100% of the variation of the dependent variable can never be "explained." In fact, the bounds of R2 are not known? nor is it known whether these bounds are fixed. Furthermore, the sampling distribution of R2 is unknown. Therefore, at the very least, ". . .R should be used with some caution until its sampling distribution is known" (McKelvey and Zavoina, 1975:112). To supplement R2, two additi0nal measures of association will be used. These are the percent of the response categories on the dependent variable predicted correctly by the probit model; and, the correlation (Spearman's rho) between the observed and predicted values of the dependent variable. These measures will be taken to be rough indicators of the goodness of fit of the probit models. Determining the statistical significance of a probit model is not so problematic. Taking the ratio of the log of the likelihood function of the equation in question to the log of the likelihood function of the null model yields the log of the likelihood ratio (LLR).8 Multiplying LLR by -2 yields a statistic, -2(LLR), that is chi-squared distributed, with the number of degrees of freedom equal to the number of independent variables in the equation (McKelvey and Zavoina, 1975:111). This measure is used to determine the prob— ability that the hypothesized model differs from the null model of 193 all Bi=0' Problems in Estimating the Model of Simultaneous Equations Earlier, I noted that for the statistical model described by Equations 1-4, it cannot be assumed that the disturbance terms are uncorrelated across equations. Accordingly, I came to the con- clusion that single-equation estimation techniques would be inapprop- riate, and that systemic estimators must be utilized. However, the appearance of categorical dependent variables in Equations 2, 4 and 4a render this strategy impossible. While some advances in estimating simultaneous-equation probit models have been made in recent years (cf., Amemiya, 1977), these techniques involve single-equation esti- mation procedures. In fact, as far as I have been able to determine, no system estimation techniques (and certainly no computer programs) exist for models involving nominally or ordinally measures dependent variables. Due to this state of affairs, it is necessary to estimate the statistical model given by Equations l-4a by estimating each equation separately. The procedure to be employed can be labeled a "pseudo two-stage least squares" technique. The procedure to be used involves the following steps; 1. Derive the reduced form equations for each of the equations in the model to be estimated, and use either ordinary least squares or probit analysis (depending upon whether the dependent variable is intervally or categorically measured) to find the reduced form estimates of the endogenous variables, Y3; 194 2. Substitute the estimated endogenous variables, Y3, into the structural equations and then, by applying either ordinary least squares or probit analysis (again depending on the level of the measurementk find the structural equation estimates. In general, using two-stage least squares to estimate models with interval level dependent variables in non-recursive systems of equations yields coefficient estimates which are consistent (Kmenta, 1971:559-564). However, in general, such an estimation technique will not yield asymptotically efficient estimates, since two-stage least squares cannot take into account the correlation of error terms across equations (Kmenta, 1971:562). Further, Maller has demonstrated that a two-stage approach applied to simultaneous probit models also yields asymptotically consistent estimates (Maller, 1977:1719); however, these estimates are also not efficient. Given the size of the CPS samples employed in this analysis, appealing to the asymptotic properties of the estimates is reasonable. Thus, the coefficient estimates to be presented in the next chapter will be considered asymptotically consistent. However, given the asymptotic inefficiency of the estimates, any inferences drawn from hypothesis testing must be treated gingerly, at best. Despite these caveats, the problems associated with the "pseudo two-stage least squares" should not be severe, if it is assumed that this approach has all the desirable properties of classical two-stage least squares. First, two-stage least squares estimation yields coefficients that are consistent, but not asymptotically efficient. This has implications for hypothesis testing, but not for the 195 overall fit of the model. Second, Kmenta (1971:581-583) reports small sample results of estimating a demand/supply system of equations using various esti- mation techniques. For the overidentified demand equation, 2SLS and 3SLS yielded identical results, for both coefficient estimates and standard errors of the estimates. For the just-identified supply equation, 2SLS yielded biased coefficient estimates (relative to 3SLS), and standard errors larger than those given by 3SLS. At this point, it would seem appropriate to derive the reduced form equations for Equations l-4a. However, for all practical purposes this is unnecessary. In order to generate estimates of Y7, Y5, and Y; for inclusion in the second round estimation of the structural equations, all that is needed is to estimate Y1, Y2, and Y3 as fonctions of all predetermined variables. Since Y4 and Y4*, the direction of the vote and the turnout variables, do not appear on the right-hand side of any equation, it is not necessary to estimate the reduced forms of Equations 4 and 4a. Additional Estimation Considerations To this point, I have concentrated on estimation problems that arise from the presence of Y2, Y4, and Y4* in the statistical model of the theory of citizen information gathering. I now turn to some potential problems in estimating Equations 1 and 3; these problems may also arise in estimating Equations 2, 4, and 4a. Specifically, I will examine in this section the problems of specification error, multicollinearity, and heteroscedasticity.9 196 Specification Error The first, and most basic, assumption of the classical linear regression is, in matrix notation, 12. Y=XB+e That is, the endogenous variables are assumed to be a linear function of the matrix of regression variables, X, and of the vector of stochastic disturbance terms, 5. Specification error refers to errors in specifying the functional relation expressed by the regression equation. The most common specification errors are; (l) omission of a relevant explanatory variable; (2) inclusion of an irrelevant explanatory variable; (3) misspecification of the form of the mathe- matical relation between Y and X (e.g., specifying a linear relation- ship, when the relationship is nonlinear); and (4) misspecification of the form of the mathematical relation between Y and the stochastic disturbance term (Kmenta, 1971:391-392). It is sufficient to note that all of these errors lead to biased and inconsistent estimates of the coefficients (Kmenta, 1971:402).10 A number of procedures designed to detect the presence of specification errors have been devised. However, their utility in this instance is questionable, for the model evaluation exercise which is presented in Chapters 7-9 must be considered, in part, as an exami- nation of the specification of the statistical model. This model is a statistical representation of the theory of citizen information gathering developed in Chapter 2. Moreover, no alternative theories will be presented for comparison against this theory, for no other theories dealing with the particular problem at hand have been 197 developed. Therefore, no "critical test", comparing two models of the same process, can be conducted. A philosophical problem emerges with the dependence on linear regression techniques to evaluate the statistical model of the theory of citizen information gathering. Classical linear regression, through the assumption of a correctly specified equation or set of equations, presumes a well-formulated theory of some process. How- ever, the theory developed in this dissertation is an initial attempt to formulate a theory which accurately describes the ways in which, and the reasons why, citizens go about gathering information during electoral campaigns. To the extent that this theory is not well-formulated, the use of regression-based techniques to evaluate its statistical representation (i.e., the model) is an incorrect use of regression analysis. In short, regression analysis is used in the next chapter to examine the initial formulation of a theory. In this analysis, a major concern is the specification of the relationships between the endogenous and predetermined variables. To the extent that the statistical model is not supported by the data, an incorrectly formulated theory, and therefore a misspecified statistical model, is indicated. ' As a consequence of these concerns, I will not present tests for particular specification errors. Rather, I will use the model evaluation exercise as, in part, a test of the specification of the model. In turn, whether or not the theory is properly formulated will be a major point of discussion following the empirical exami- nation of the model. 198 Multicollinearity Multicollinearity is the situation in which two or more explanatory variables are intercorrelated. If there is an exact linear relationship between two or more variables (i.e., perfect multicollinearity) estimation is impossible, for the data matrix X'X is singular, and its inverse does not exist. If multicollinear- ity is high, but less than perfect, the standard errors of the esti- mates of the regression coefficients are artificially inflated, leading to unduly conservative inferences regarding the statistical significance of the coefficient estimates. There is every reason to expect that multicollinearity will be a problem in estimating this system of equations. In particular, consider the reduced form of Equations 1 and 2: 13. Y$1=f(X],X2,X3,X4,X5,X6,X7,X9) l4. Y$2=f(x],X2,X3,X4,X5,X6,X7,X9) Next, consider the structural form of Equation 1, 15. Yil=f(X],X2,X3,X4,X5,X6,Y$2) and the structural form of Equation 3: 'k * 16- Yi3 fl il’ Yi2) It is most obvious that, in Equation 18, Y52 and XI-X6 will be highly intercorrelated, given that sz is estimated as a function of Yl'Y6° Further, Y?1 and Y? will be highly correlated in Equation 19. 2 Therefore, multicollinearity will be a problem in estimating the statistical model given by Equations l-4a. Given this, many 199 of the coefficient estimates will be statistically insignificant, due to the inflated standard errors of the coefficient estimates. Cau- tion will be exercised in evaluating the contributions of individual predetermined variables to "explaining" the variation of the endogenous variables. Further, where multicollinearity seems to be a problem, estimates of multicollinearity (e.g., regressions of each regressor on all other regressor variables) will be presented. Heteroscedasticity Homoscedasticity is the assumption that the variance of the error terms is constant for all observations, i.e., l7. Var (ei)=°2 That is, the variance of the error term is not dependent upon different values of the predetermined variables. If Var (e1) is not constant across all observations, the situation is called heteroscedasticity. In the presence of heteroscedasticity, least squares estimators are unbiased and consistent, but do not have the smallest variance in the class of all unbiased estimators. Further, the estimators are not asymptotically efficient (Kmenta, 1971:249-254). As a consequence, the usual tests of significance for regression coefficient estimates do not apply under conditions of heteroscedastic disturbance terms. However, there is no indication that assuming a homoscedastic error term is problematic for Equations l-4a. There is nothing about the specification of these questions which would lead one to conclude, q_priori, that the variance of the error terms is dependent 200 upon the values of the predetermined variables. Given that hetero- scedasticity is unlikely to be a problem, tests for its presence will not be presented, and I will not consider the topic further. Conclusion In this chapter a statistical model of the theory of citizen information gathering has been developed. The model presents some rather severe estimation problems. Specifically, the presence of interval level and categorical endogenous variables renders system estimation impossible. A "pseudo two-stage least squares" approach has been suggested and developed. While estimating the statistical model will be possible with this approach, the estimates of the coefficients will be consistent but asymptotically inefficient. Thus, hypothesis testing of coefficients will yield problematic results, at best. Further, it is expected that multicollinearity will be a problem, leading to further difficulties in inferring the effects of the predetermined variables on the endogenous variables of the model. 201 ENDNOTES 1To estimate the model, a ”pseudo-two-stage least squares" approach will be used. This procedure, and the reasons for its use, will be more fully elaborated below. Suffice it to say at this point that the procedure involves a mixture of probit and Ordinary Least Squares techniques. 2The relationships represented by Equations 2, 4 and 4a will be estimated through probit analysis. A probit model is of the following form: P(Y=1/a+B]X +...+Ban)=¢(a+8]X +...+ann) 1 1 P(Y=0/&+é]xl+...+énxn)=l-o(&+é1x]+...+§nxn) where Y is the dependent variable measured at a non-interval level, and 0 is the cumulative normal distribution (Aldrich and Cnudde, 1975; 582). The extension of this model to the n-chotomous dependent var- iable is straightforward. While the above equations are the correct representation of a probit model, the probit equations in the text will be represented as linear regression equations. This is done only to make all the equations in the system comparable in form. 202 3The assumption given by Equation 9 is that the theoretical stochastic terms of each equations are uncorrelated across equations. In any system of equations, this assumption is problematic; in a system of non-recursive equations, where one or more endogenous variables appear on the right hand side of another equation, the assumption is totally unfounded. Failing to take into account the correlation of error terms across equations results in asymptotic inefficiency. The usual solution to this problem is to estimate the system of equations simultaneously, i.e., by using a system method of estimation (cf., Kmenta, 1971:517-529; 573). 4In most operationalizations, Y2, the general media use variable, is measured at the ordinal level. When Y2 appears on the right hand side of Equations 1 and 3 it is assumed to be measured at the interval level, since regression techniques assume that all predetermined variables, including binary variables, are measured at the interval level. 5There are some system estimation techniques that incorporate probit analysis. However, the underlying assumptions of these pro- cedures are extremely restrictive. Further, the necessary computer software for these techniques was not available where, and when, the data analysis was conducted. As a consequence, the procedures which are outlined in this chapter were used. 6This discussion of probit analysis is taken from McKelvey and Zavoina (1975) and from Aldrich and Cnudde (1975). 203 A 7Except insofar as R2 will converge in the limit to 1.00. 8Of course, once the logs of these likelihood functions are taken, finding the ratio merely involves taking the difference be- tween the logs of the likelihood functions, the smaller being sub- tracted from the larger. 9The assumption of non-autoregressive disturbance terms will not be discussed, in that serial correlation usually occurs only when time series data are used to estimate a regression model (cf., Kmenta, 1971:201-205; 269-271). 10For a more detailed description of the problems resulting from specification errors, see Kmenta, 1971:392-402. CHAPTER 7 MODEL EVALUATION FOR 1972 Introduction In the next three chapters the results of the model evaluation procedures described in Chapter 6 will be presented and discussed. The statistical model to be evaluated is a non-recursive system of linear and additive equations which consists of the following func- tional relations: 1. Political Information=f(Strength of Party Identification, Level of Education, Level of Income, Socio-Economic Status, Interest in Politics, Interest in the Campaign, Media Utili- zation) 2. Media Utilization=f(Political Information) 3. Uncertainty=f(Media Utilization, Political Information) 4. Turnout=f(Differential Benefit, Uncertainty, Civic Duty) 4a. Candidate Choice=f(Differential Benefit) A brief summary of the hypotheses represented by Equations l-4a will be helpful. In the political information equation, all relation- ships are expected to be positive, with the exception of the campaign interest variable. Due to the coding of this measure for all three surveys, a negative relationship is hypothesized. Similarly, negative 204 205 coefficients are hypothesized for all predictor variables in the media use and uncertainty equations. Due to the coding of the turnout vari- ables for all three years, a positive relationship between uncertainty and turnout is predicted, and a negative relationship between turnout and civic duty is hypothesized. While the size of differential bene- fit is hypothesized to influence turnout, the measures of differen- tial benefit used in the evaluation of the model are designed to pre- dict the direction of the vote, not whether a vote was cast. Thus, no explicit hypothesis is offered regarding the relationship between turnout and differential benefit. Finally, for the candidate choice equation, the direction of the relationship depends upon the opera- tionalization of differential benefit used. For 1972, a negative coefficient is predicted for Utility, which is based on the seven- point issue scales; a positive coefficient is hypothesized for Diffeel, which is based on the one hundred-pointfeeling thermometers. For 1974 and 1976, different hypotheses are appropriate, and will be made explicit in Chapters 8 and 9, respectively. In this chapter the results of the model evaluation for the 1972 data set will be presented. These results, as well as the estimates for the 1974 and 1976 data sets, are based only on the set of respon- dents who participated in all three waves of the 1972-1974-1976 Ameri- can National Election Series of the Center for Political Studies. As noted previously, the purpose of testing the model with panel data is not only to examine the model in different electoral contexts, but also to examine the model at different points in time, in order to trace any changes across time among the same set of respondents. Es- timating the model using only panel participants fulfills both of 206 these purposes. The specific variables used in this analysis are as follows: Caninf: measures candidate-referenced information; Parinf: measures party-referenced information; Media Utilization: Media; measures how many of the four major news media (television, radio, newspapers and magazines) were used to monitor the campaign; V464: from the 1972 CPS survey; measures the extent of televi- sion usage to follow the campaign;1 Uncertainty: measures uncertainty regarding the candidates' issue positions; V477: from the 1972 CPS survey; measures turnout; V478: from the 1972 CPS survey; measures candidate choice; V300: from the 1972 CPS survey; measures level of education; V400: from the 1972 CPS survey; measures social class; V420: from the 1972 CPS survey; measures level of income; Party 10: measures strength of party identification; V29: from the 1972 CPS survey; measures interest in the 1972 campaign; Interest: previously described as ACTS1972; measures general interest in politics; DTerm: based on the “civic duty“ questions from the 1972 CPS survey; measures civic duty; Utility: based on the seven-point issue scales from the 1972 CPS survey; measures differential benefit; Diffeel: based on the one hundred-point feeling thermometers from the 1972 CPS survey; measures differential benefit. 207 Also used in this analysis are a number of instrumental variables for the endogenous variables of the model constructed in the first stage of the two-stage estimation procedure described in Chapter 6. The details of the construction of these instruments, as well as the correlations between the instruments and the endogenous variables for which they stand, are presented in Appendix D. It is sufficient to note here that each endogenous variable, with the exception of the voting variables (V477 and V478) has four instruments: Caninfl, Caninf2, Caninf3 and Caninf4 for Caninf; Parinfl, Parinf2, Parinf3 and Parinf4 for Parinf; Medial, Media2, Media3 and Media4 for Media; V464a, V464b, V464c and V464d for V464; and Unsurl, Unsur2, Unsur3 and Unsur4 for Uncertainty. These instruments are used when an endogenous variable appears as a predictor variable in Equations l-4a. These instrumental variables also form the basis for the manner in which the results for the 1972 modelevaluation are presented. Four models, Models I-IV, are presented, with each model consisting of estimates of Equations l-4a based on a single set of instruments. For example, for Model I Caninfl, Parinfl, Medial, V464a and Unsurl are used as the instrumental variables. In addition, two versions of each model will be presented, one version for each of the two media use variables. For example, Model Ia includes V464 and V464a, Model Ib uses Media and Media2, etc. Finally, the results from the estimation of V477, the turnout variable, are presented only for Models Ia, IIa, 111a and IVa, since the results are the same for both versions of each model. V478, the candidate choice variable, is estimated solely as a function of differ- ential benefit, a variable which does not appear elsewhere in Equations 208 l-4a. Hence, the results from estimating V478 are constant across both versions of all four models, and will thus be reported only in conjunction with the report of the estimates of Model Ia. The Data In Tables 27 and 28 are presented the results of estimating Models Ia and lb. With the exception of predicting turnout and candi- date choice, the model does not perform particularly well. In the equations predicting party- and candidate-referenced information, vir- tually none of the variation in the endogenous variables is explained by the predictor variables. In particular, the coefficients for the media use variables are of very small magnitude, and are statistically insignificant. In Table 27 much of this lack of strength of V464a is due to high multicollinearity; the correlations between V464a and V29 and Interest are .62 and -.65, respectively. For a complete report of the zero-order correlations between the six socio-economic- political variables and the media use instrumental variables, see Table 56 in Appendix E.2 The ability of the model to predict television use to monitor the campaign (V464) is also unimpressive. Both probit equations are sta- tistically insignificant, with no correlation between predicted and actual values.3 The Uncertainty equation is not much better, although the regression is significant at p=.05. Further, the coefficients for V464a, Caninfl and Parinfl are in the hypothesized direction, and are statistically significant. Using only the two differential benefit measures in separate equations, the model predicts candidate choice well, with R2 as high 209 _a_puz mo.va oe.o_nu No." m. em_.wm Ammo.v ANN..V aoaN.A.Lewcaav- amee.haama>v- ae.mn»»ewaacoo:= Pa__uz mo.va o_.c_ua cmo.nmm oe..nm Ammo.v Am~_.v soon.APceweauv- ammm.Aaaca>v- osc.mn»eewaocoo== o.oupm:uum vcm umuuwvmga :mmzumn cowumpmccou spoooccoo eoaoeaoca Na.ma Pmmuz mo.xa Nw.muwce .Vmssxm- moo.u~a mo. . eoa_.A_Lepcaav- a_a.ueee> o.oupm:uum use umHUPUmca :mmzuwo cowumpmccou spoooccoo aoooweoca Ra.ma _mmuz mo.xa wme.nwce _vmssxm- _oo.nmm mwo. - Nmo.A_Le_eauv+ Fe_.naea> peppuz mo.vQ “v.0"; mmo.u m om_.um Ame.pv Aaeo.v Ammo.v .1 wmm.fiaaaa>v- Nmo.ao~a>v+ “No.Aooa>V- “moo.v Amo_.v Apcm.v ANwN.V amoo.Aoom>v+ aaom.AoH sacaav+ “No.Aamocoo=HV- ©PN.AmN>V- ama.uceacaa Pa__nz mo.va N_.muu mmo.- m ma_.nm Aemm..v Ramo.v Ammo.v -mu upo.Aaaea>v+ .mo.flo~a>v+ mmo.aooa>v- Aaoo.v AN.P.V Amoa.v Aapm.v moo.Aoom>v+aN—N.AQH hagwmv+ omo.AummLm#:~vu mmp.Amm>vn ¢w0.uw:wcmu .Nmmp .mH Pmuoz .NN mpnmh 210 mocmuwwwcmwm mo ummu umpwmpimco .moanr pw.upm:uom can umuuwcmca :mmzumn :owumpmcgou spoooccoo eoooweoea ae._m mamuz mo.va w_~.mauwca vassxm- oom.u~m moo. . apeo.A_ooLcaov- aoo.nmua> veo.u—m:uum new cmuu_umca :mmzuma cowum_mgcou >6.8.28 empoweoca am.mm mamuz mo.va mm.mwmwfico vassxm- New." a mmo. . amme.A»pP_wo=v+ amm.-umka> wmm.upm=uum use cwuupvmca :wwzumn cowumpmggou s_oooccoo aoaoaeoca em._w .appnz mo.va Ne._e.nflca mvmssxm- mom.n~m Aomp.v Amoo.v Aomo.v . eomw.fi_c=me=v+ moo.A_ooccwov- sema.AELowov- mNK.P-nNNa> o-.u_m:uum new cmpuwcmca :wmzumn cowum_mcgou spsooceao eaaocaoca am.Pm Fa__nz mo.va Pm._a_uflca mvmssxw- Nom.nmm Aom..v AaNo.V Aomo.v . amaw.fipc=me=v+ N_o.Asaapeo=v+ ahma.A5coeov- mNN.F-uN~a> A.6.aeoov .NN o_aae 211 as .80, and a correlation between predicted and actual values no lower than .64. In the turnout equations, the coefficients for DTerm and Unsurl are statistically significant and in the hypothesized di- rection. The differential benefit measures were explicitly designed to predict candidate choice, not turnout, so the lack of statistical significance of Utility and Diffeel are not unexpected. The results presented in Table 28 are similar to those of Table 27. The model does not predict levels of information well; further media use as measured by Medial, is substantively and statistically insignificant as a predictor of information. As with V464a, this is due, in part, to collinearity with the other predictor variables. The correlations between Medial and the other left-hand-side variables range from .23 to .65. The model does a better job of predicting the variable Media, relative to the power of the model in explaining V464. Both Media equations in Table 28 are statistically significant, as are the coef- ficients for both Caninfl and Parinfl. However, while negative rela- tionships were hypothesized for the media equations, positive rela- tionships are observed. This is due largely to the fact that the dynamic relationships described by the theory of citizen information gathering are being examined with static data, a problem discussed at length in Chapters 2 and 6. Finally, the Uncertainty equations in Table 28, while substan- tively unimpressive, are statistically significant. Further, the coefficients for Medial and Caninfl are in the hypothesized direction, and are statistically significant. However, Parinfl does not perform well as a predictor of Uncertainty; its coefficient is not significant 212 mucmuwwwcmwm mo ummu umpwmuumco .mo.waa papa": mo.va No.mpnu NNo.n~m mm_.um Ammo.v Ammo.v aNeN.Apaceozv- meF.A_cePcaaV+ omm.mnxoewaacooe= _a__uz mo.va om.opuu “No.u~m Cap.nm Awao.v Akao.v aop_.A_aPaozv- aONN.A_c=Peauvu can.euxoeaaacooe= mam": mo.va ma.mawu meo.w.m mcm.um A_mp.v ame~.,.fl_ceacaav+ sap.muawaoz mam": mo.va am.amua mao.n.m mPN.nm Aoo_.v emoo._.APcePeauv+ wom._naweoz Peppuz mo.va ue.ouu Nmo.u m om_.um Acmm.v Am...v Aomm.v a~_.A.aweo=v+ anN.AoH»aeaav+ oom.fimm>v- Amm~.v Ammo.v Ap__.v Ao_o.v opo.ApmmLmu:HV1 anvo.Aomc>v+ cvo.Aooe>vn moo.Ao0m>v+ mwm.1nw:wgmm .a_Puz mo.va Np.mua mmo.u a. map.nm Ammo.v A~N_.v Aama.v oao.A_awaozv- awNN.Ao~»acaav+ eep.AmN>v- Aamm.v Ammo.v AmN..v Ampo.v mmo.fiamocoee_v- ammo.Aoma>v+ P_o.Aooe>v- aa__.fioom>v+ ama._uceceau .mump .0“ quoz .wm mpnmh 213 at the .05 level of probability, and is in the opposite direction hy- pothesized. The results of estimating Model IIa, reported in Table 29, vir- tually parallel the results presented in Table 27. For the informa- tion equations, the only difference is that the coefficient for V4640 is statistically significant in predicting Caninf; however, it is in the opposite direction hypothesized. The results of the probit analysis of V464 are the same as in Table 27--no ability to predict levels of television usage is demonstrated by levels of political in- formation, regardless of the operationalization used. The Uncertainty equations are again weak in explanatory power, although both regressions are significant at the .05 level of proba- bility. Further, all coefficients are in the hypothesized direction, and are statistically significant. Finally, both Uncertainty (Unsur2) and civic duty (DTerm) are strong predictors of turnout, while the differential benefit measures play almost no role in explaining whether a vote was cast. Little discussion of Table 30 is necessary, since the results are virtually identical to those presented in Table 28. The only point which merits discussion is the fact that information levels are posi- tively, not negatively, related to media utilization as measured by Media. Again, the necessity of examining dynamic relationships with data collected at one point in time has resulted in relationships that are in the opposite direction hypothesized by the theory. Little time also needs to be spent on Table 31, wherein the re- sults of estimating Model IIIa are reported. All conclusions drawn from Tables 27 and 29 remain intact. The only difference between 214 .a_Puz mo.va ea.m,nu Fmo.umm NmF.nm Ammo.v Amm_.v aNeN.ANceacaav- a_ca.Aaaea>v- w_m.mnseemaeeooe= peF_uz mo.va o~.c.nu mmo.umm. mc_.nm hamo.v Amwp.v amaN.A~Leeeauv- amm~.Aaaoa>v- Nuc.muzeewaocoo== o.oupm:uum vcm umuuwumca :mmzuwn cowumpmggou spoooccoo empoweoca Na.ma Pmmnz mo.xa mm._nwce .vmasxm- moo.nmm Nmo. . wo_.ANLe_eaav- aam.naoa> o.oupm:uum use cmuuwuwga :mmzpma cowumpwcgou s_aooceeo eooo_eoea Na.ma .mmuz mo.xa MNw.u%ce Pvmssxm- Noo.nmm _mo. . Nwo.fimceaeauv+ mmp.uace> _ap_uz mo.va ma.ona Nmo.u~m. cm_.nm Aamm._v Aeao.v Ammo.v mma.Aaaca>v- “mo.Aoma>v+ emo.Aooa>v- Amoo.v Aao_.v Amam.v ANNN.V ao_o.ficom>v+ wppm.Ao_sucaav+ meo.flpmocooeav- amp.fiam>v- aem.nceacaa pv—Fuz mo.va Np.muu mmo.u a mmp.um Awmo.v Ammo.v Ammo.v ammo.fieaea>v- swmo.Ao~a>v+ NNo.Aooa>v- Acoo.v Aa__.v Ammo.v “com.v *opo.Aoom>v+ *NNN.Ao—x»gmav+ mmo.flumwgmucfivl mmo.AmN>vu omm.uw:w:mu .mkmp .mmm quoz .mm mpnmh 215 mocmuwwwcmwm we pmmu umpwmuumco ,mo.Waa mm~.upm:uum vzm umuowumca cmmzumn :owumpmcgou spaoaccoo eoao_aaca am..w _a_.wz mo.va mN.PaPnALa mvaasxm- _om.umm Aom_.v Amoco.v Aamo.v . seem.A~c=me=V+ mooo.A_ooccaov- aaee.AeLoeov- amm.P-nNNa> m-.upm:uum new umaowumcn cmmzumn cowuumgcou spoooccoo eoaoeeoca aa.Pw Pa__uz mo.va mm._a_ufice mvmssxm- .om.nNm Aomp.v Aemo.v A_mo.v . . swam.fime=me=v+ N_o.A»aLPaa=v+ aaaa.AeLopov- mma._-waae> .A.a.a=oov .mN o_aae 216 8.83::me 68 33 cmifioco .8..wa _a__nz mo.va me.anu omo.nmm Nap.um Ammo.v Amao.v emFN.ANaeaozv- Nmo.amceweaav+ .wc.musoeeaaeooe= _appuz mo.va Na.e_nu “No.1Nm. o“..um Awao.v Aeao.v aamo.fima_eozv- «N_N.A~ce_eaov- .Fm.musoeaaacooe= mam": mo.va www.mawu eao.n.m m-.um AmNF.v amm~._ANceccaav+ moo.~uaaeoz memuz mo.va mm.awuu aao.u.m www.1m Aw~_.v N swam.fimceaeauv+ on._naweoz _a__nz mo.va mk.cnu amo.n m. ooN.nm Amma.v Ammo.v AeFN.V Noc.fimaeeozv+ «Pa~.Aossocaav+ m_o.AmN>v+ “a“..v A_No.v ANNO.V Amoo.v NN_.AamoeoaeHv- ammo.Aoma>v+ oo_.Aooa>v- moo.fioom>v- mac.,-nceccaa _a__uz mo.va ea.muu “No.1.m ow_.um Amaa.v Amo_.v ma~.A~aaeozv+ mm_.AaH»acaav+ mmm.amm>v- AeaF.V Ammo.v Aomo.v Aopo.v aamm.Aamocode~v- ammo.fiome>v+ o_F.Aooa>v- aoo.Aoom>v- coo.N-uceweau .Nump 2.: Eve: .8 3an 217 pa_Puz mo.va “a.o_nu e_o.n.m am_.nm Ammo.v Am_..v acNN.AmcePcaav- swam.fioaca>v- woa.mn»oewaacooe= Pap_uz mo.va mm.mpnu mmo.n m. mop.wm Ammo.v Ae_,.v emom.Amw=Feauv- mmP.Aoaoa>v- aac.mnspewapeoo== o.ou_m:uum use uwuueumca coagumn :owumpmccoo spoooccoo eoao_aoca Nu.me .mmuz mo.xa mm.auwca Pvmasxm- moo.umm moo. . mo_.Amcewcaav- mem.uaca> o.oupm:uum new umuuwcmca cowzamn scrumpmgcou speooccoo aopoaeoca N~.ma _mmnz mo.xa wow.n%cu _vmssx~- Noo.nmm coo. . ewo.Amceweauv+ FmP.uaea> peppnz mo.va No.0"; mmo.u a wm_.nm Ammm.v Aemo.v Apmo.v aam.Aoaea>v+ Nmo.AoNa>v+ amo.fiooa>v- Aaoo.v Ammo.v Apep.v Aam_.v «moo.Aoom>v+ «NNN.AQH»uLmav+ wm—.Aummgwucmv+ mmo.AmN>vi Nmo.uwcwgmm .ap_wz mo.va a..muu mNo.n~m mm..nm Amma.v ANNo.V Aemo.v aNN.Aoaca>v+ aowo.fioma>v+ mmo.Aooa>v- Amoo.v Ammo.v Amm_.v Amm_.v ao_o.fioom>v+ ao_~.Aofixecaav+ mmo.Apmocoa=Hv- amp.AmN>v- ama.ucea=au .mump .mHHH Fmvoz .—m mpaw» 218 muchwwwcmmm we ummp vmpwmuumco .mkoQw emm.npm:uum vcm tmuuwumga cmmzumn :owumpmggou s_oooceao eaeoweoca ae._m Pappwz mo.va me..epuAca mvmssxm- mom." a Aom_.v Aaooo.v Aomo.v . aaam.Amc=meav+ aooo.A_ooccaav- .Nma.AELoeov- MAN.F-HN~a> cwm.u_m=uum czm umuumumga cmmzuma cowuwpmccou s_aoaccao eoaoaeoea gm._m Pap_uz mo.va Am._a_uAca mvmssxm- No~.n~m Acm,.v Aamo.v Acmo.v . swam.Amesme=v+ c_o.A»aPFee=v+ tama.AeLoeav- -._-umae> .A.a.oeoov ._m opaae 219 Table 31 and the two previous models incorporating V464 and its in- struments is that in conjunction with Caninf3, V464c is not a statis- tically significant predictor of Uncertainty. Beyond this difference, the model performs equally well, or poorly, as Models Ia and 11a on all other counts. The only remarkable aspect of Table 32 is that the instrumental variable for Media, in this case Media3, does not appear in the Caninf and Parinf equations. In these equations, Media3 is an almost perfect linear combination of all other predictor variables; the R and R2 for the regression of Media3 on all other predictor variables both approach unity.4 In the uncertainty equations, the coefficients of both Caninf3 and Parinf3 are in the opposite direction hypothesized. Furthermore, the coefficient for Parinf3 is significant at the .05 level of probability. The results of the estimation of Model IVa, using V464d as the media use instrumental variable, are presented in Table 33. Not sur- prisingly, the results are the same as for the estimation of Models Ia, 11a and IIIa. V464d is not a significant predictor of levels of political information, nor are Caninf4 and Parinf4 significant predic- tors of media use. V464d, Caninf4 and Parinf4 are all significant pre- dictors of levels of Uncertainty, and all coefficients are in the hypothesized direction. As in the previous models, civic duty (DTerm) and uncertainty (Unsur4) are strong and statistically significant pre- dictors of voter turnout (V477). The Caninf and Parinf equations of Model IVb, in Table 34, differ from all previous information equations. In both equations, the in- strumental variable for Media, Media4, is a fairly strong predictor of 220 8531.53 66 $3 335195 8.3.. .ap_uz mo.va ON.NNHE No.6Nm NmN.uN Amoo.v ammo.ammwnmzvu rm~_.Amwchmav+ mom.mnzpcwmugmucn Fe__nz mo.va e_.Nmuu No.1Nm NNN.HN Amoo.v ANNo.V amao.Amaweozv- moo.AmceNeauv+ www.musoe_aocooe: NNmuz mo.va 6N.mauu NNo.n.m FNN.nN a: N we_N._ANLe_caav+ Nmo.Nneaaoz NNmuz mo.va Nm.mNuu wee." m. NNN.HN ANN_.V N awoo._AmL=_eauv+ _om.pnawooz _a__nz mo.va mm.Nnu mmo.nNm. cm_.wm Ammo.v Ammo.v aamN.AoH»oeaav+ aNmN.AmN>v- A_No.v AoNo.v A_mo.v Aaoo.v NNo.Aumocoo=~V+ acao.AONa>v+ NNo.Aooa>v- ao_o.Aoom>v+ Nom.uceecaa _a__nz mo.va Na.mnu cNo.nNm mN_.wN Ammo.v Ammo.v am_N.Aossacaav+ _o_.AmN>v- ANoF.V ANNo.v ANmo.v Aaoo.v .No.AemoeoaeNv- aeNO.AONa>V+ NNo.Aooa>v- ao_o.Aoom>v+ eNa.uceaeaN .mump .a~H_ quoz .Nm mpowh 221 Pe_suz mo.va Na.emns emo.n.m aeN.nN AN_P.V Ampe.v AomN.Aeaea>v- eFmF.Aaceeeaav- mNN.mnsoesaocaoe= Pap_nz mo.vs _N.m_ns aNo.nNe. Pe_.ns Am_P.V Ammo.v aeop.Aeaea>v- aPmN.Aaeseeeev- eme.mu»eseeacooe= o.ou_e=uoe use umuuwumsa smwzums so_uepmssou speooceoo eooo_eacs NN.ea .mmuz mo.xa Wm_.nwce .VNSSXN- ooo.uNN oNo. . Npo.Aaeeeeasv+ aom.uaea> o.Ou_e:pue use umuu_umee smmzumn sowuepeesou e_eooccoo eoeoeeoca NN.ma .Nmuz mo.xa wNe.n%ee .vNSSXN- Noo.uNs Nee. . Neo.fleeseeeev+ ea_.naea> .aPPuz mo.va mm.ens mmo.uNm No..ws AaNe.V A_mo.v Aaoo.v mmo.AoNa>v+ NNo.Aooe>v- aepo.flo0m>v+ ANNo.V ANa_.V AMN_.V Aemm.v aaNN.Aes»ecasv+ NNP.ASmoeoeesv+ aNam.AmN>v- _Ne.Aeaee>v+ Nee.nseseas _e__uz mo.va Np.mus aNo.nNm mN_.uN ANNo.s ANmo.s Amoe.s aNeo.AoNa>v+ NNo.Aooa>v- eo_o.AooN>v+ Aeso.v Aom_.v ANN_.V ANNe.v «mpm.AoH»useav+ puo.AummngsHvu oNP.AoN>V1 mmp.Auvou>v+ moo.nmswseu .mump .e>~ pmuoz .mm mpnep 222 museuwewsmwm we amen uwpweulmso .mo.w¢w PmN.upe:uue use umuuwumse swmzums sowue_mesou speooceoo eoeoseacs Ne._e Fa__nz mo.va _o._a_ufice mVNSSXN- NON." s AOmp.. fleece.v A_mo.v . aNae.Aac=me=V+ mooo.apoocseev- saaa.AeLoeev- NeN._-uNNa> ewm.upe=aue use umuuwumea smwzuwn sowue_mssou NPSooeeao eoeoseoea Ne._e Fas_nz mo.va eN._apnAse NVsSSXN- FON.uN Aom_.v AaNo.v Apmo.v . apae.fiac=me=v+ epo.Asue.ee=V+ saga.flscaeev- meN._-nNNa> .A.e.oeaov .NN o_eae 223 museoCmsEm so 33 uwzeuéso .mo.1v.sa Pep_uz mo.va ee.e_uu eNo.u.m Ne..us AeNo.v Amao.v aao_.haeeeeauv- an...Aeaweozv- mpe.mu»eeeaecooe= Pas—"z mo.va ee.mmns emo.nNm. NaN.ns Aomo.v Aeo.v aPON.Aacescaav- emp_.Aaaeeozv+ mem.mw»eesaaeooe= mmmuz mo.va mm.meum umo.umm. mo~.um Ammo.v «oem.A¢mswseaV+ www.muemuwz eNmnz mo.va me.mNus Nao.u.e NNN.us ANN_.V seam.fiaseeeaev+ emm.Pnaeeoz _a__uz mo.va Ne.ens mmo.n.m NeN.uN Ame..v ANmN.V AmNN.v e_P.Aessecaav+ NNe.AmN>V+ same.Aemoceeesv- A_No.v A_eo.v A__e.v “.am.v aamo.AoNa>v+ aem_.fiooe>v- o_o.Aoom>v- wNom.Aaa_eazv+ mee.N-nseeeaa _a__nz mo.va ee.mns eNo.n m. amp.us “so..v ANNN.V ANNN.V epp.Aeseecasv+ eNe.AeN>v+ aama.fiemocoeesv- AmNo.v APNe.V ANPo.V “Noe.v aNoo.AoNa>v+ wem_.Aooa>v- Npo.Aoom>v- «Nep._saaeeozv+ ea.m-nceseae .Nmmp .n>_ Pmuoz .em upseh 224 levels of political information. The coefficients for Media4 in both the Caninf and Parinf equations are statistically significant, and are in the hypothesized direction. However, the overall explanatory power of the model, relative to levels of political information, is no greater than that of the previous Caninf and Parinf equations. In the Media equations, it is once again the case that the coef- ficients for Caninf4 and Parinf4 are in the wrong direction, while still being statistically significant. Finally, in the Uncertainty equations Media4, in conjunction with Parinf4, is positively associ- ated with levels of uncertainty, a relationship which is contrary to that hypothesized. However, the relationships between Uncertainty and Media4 and Caninf4 confirm the hypotheses represented by the Uncer- tainty equations. Discussion This model evaluation exercise for the 1972 data set has not been very successful, but neither has it been a failure. I will first dis- cuss the weaknesses, and then the strengths, of the estimation of the statistical model. The model proved to have almost no ability to explain levels of political information, regardless of the operationalization of the endogenous variable. Although all regressions are significantly dif- ferent from the null model at the .05 level of probability, in no case does the adjusted proportion of variation explained (R2) reach 5%. To compound matters, the media use variables provide virtually nothing to the explanatory power of the equations. Only the fourth instrumen- tal variable for the variable Media, Media4, is statistically 225 significant and in the direction hypothesized. The strongest predictors of levels of political information are strength of party identification, education, and income. In two of the four Parinf equations, V29, the campaign interest variable, is a statistically significant predictor of information. All other socio- economic-political measures are statistically insignificant and/or in the opposite direction hypothesized. Matters deteriorate when attention shifts to the media use equa- tions. None of the probit analyses of V464, the television utilization variable, are significant at the .05 level of probability. That is, levels of political information are not related to levels of television use. On the other hand, all Media regression equations are signifi- cantly different from the null model, and all coefficients of the pre- dictor variables are statistically significant. However, all coeffi- cients for the information level measures are in the opposite direc- tion hypothesized. As noted previously, this result stems from the necessity of using static data to examine inherently dynamic relation- ships. If it were possible, for instance, to measure political infor- mation at a point in time earlier than that at which media use was measured, the hypothesized negative relationships might very well emerge. Beyond these weaknesses, some bright spots did emerge from the examination of the statistical model. For instance, while the propor- tion of variation in Uncertainty for which the model accounts is uni— formly low, the hypothesized relationships between uncertainty and media use and levels of political information are confirmed. In almost all cases the coefficients are in the hypothesized direction, 226 and are statistically significant. It is clear that the model performs best for the‘l972 data when predicting turnout and candidate choice. The coefficients reported all perform as expected. Of particular importance to the credibility of the theory of citizen information gathering is the relationship be- tween uncertainty regarding the candidates' issue positions and turn- out. A key proposition of the theory is that the higher is a citizen's level of uncertainty, the less likely it will be that the citizen voted. In all cases the proposition is strongly confirmed by the data. I will return to the performance of the model for 1972 in Chapter 10, where the results of the model evaluation for all three years will be discussed. I turn now to the model evaluation exercise for the 1974 data, presented in Chapter 8. 227 ENDNOTES 1While more than two media variables were available for analysis, only two are reported here. Media is used because it measures the use of all four news media (television, radio, newspapers and magazines) to gather political information. V464 is used because it measures the extent of television use to follow the campaign, and because similar measures of newspaper usage, which are available for the 1974 and 1976 data, were not included in the 1972 survey. The other media use vari- ables in the 1972 survey measured only whether a particular medium, e.g., newspapers, were used to follow the campaign, but did not mea- sure to what extent such media were used to gather political informa- tion. 2The usual test for multicollinearity involves regressing the suspect independent variable on the set of the remaining predictors. However, for the data sets from the 1972, 1974 and 1976 surveys, the size of the correlations between the media use instrumental variables and the other predictor variables in the political information equa- tions obviates the need for the usual procedure. It is obvious from Tables 56-58 in Appendix E that severe multicollinearity exists in the candidate- and party-referenced information equations for all three years. 228 3The probit equations for V464 in Tables 27, 29, 31 and 33 do predict a substantial proportion of the cases correctly, but these figures are misleading. The results of the probit analyses indicate that all cases were predicted to be in Category 2; further, 45.7% of the cases did fall in Category 2. Hence, 45.7% of the cases were pre- dicted correctly, and there is no correlation between predicted and actual values. 4The exclusion of Media3 from the Caninf and Parinf equations is due almost solely to the zero-order correlation between Media3 and V300 (level of education). The Pearson's correlation coefficient be- tween these two variables equals .989 (see Table 56 in Appendix E). CHAPTER 8 MODEL EVALUATION FOR 1974 Introduction In this chapter the results of estimating the statistical model of the theory of citizen information gathering with data from the 1974 CPS survey are reported and discussed. As with the 1972 test, the es- timation involves only those respondents who participated in all three years of the 1972-1974-1976 panel survey. Further, only the second stage estimates of the “pseudo two-stage-least squares" procedure are reported in this chapter. The model to be estimated is the same as that described at the beginning of Chapter 7. The following variables are used in this analysis: Info: measures political information levels; Media Utilization: V2066; measures whether respondents' primary information source about the 1974 campaign was television, newspapers or both; TVNews: measures the extent of television usage to follow the campaign; Paper: measures the extent of newspaper usage to follow the campaign; Uncertainty: measures the level of uncertainty regarding the issue positions of the two major parties; 229 230 Turnout: V2319; from the 1974 CPS survey; Vote Choice: V2322; measures the party of the candidate for whom the respondent voted in the 1974 House of Representatives elections; from the 1974 CPS survey; Education: V2418; from the 1974 CPS survey; Income: V2549; from the 1974 CPS survey; Social Class: V2525; from the 1974 CPS survey; Interest in the Campaign: V2026; from the 1974 CPS survey; General Interest in Politics: Interest; previously described as ACTSl974; PartyID: measures strength of party identification; Utility: measures differential benefit based on the seven-point issue scale placements by respondents of the two major parties.1 For each of the information, media use and uncertainty variables, only one instrumental variable has been constructed. The details of the construction of these instruments are given in Appendix D. These instrumental variables are Infol for Info; V2066a for V2066; TVNewsl for TVNews: Paperl for Paper; and Unsurl for Uncertainty. These in- struments are used whenever the endogenous variables of the model appear in equations as predictor variables. Since V2319 and V2322 do not appear as right-hand-side variables, instrumental variables for them were not estimated. Given only one set of instrumental variables only one model will be estimated in this chapter. However, three versions of the model, each incorporating a different media utilization variable, will be estimated. Model Ia uses V2066, and its instrument, V2066a; Model Ib 231 uses TVNews and its instrumental variable, TVNewsl; and Model Ic in- corporates the newspaper use variable Paper and its instrument, Paperl. The probit analyses of turnout (V2319) and vote choice (V2322) are reported only in conjunction with Model Ia, since these results are constant across all three versions of the model. The Data The results from estimating Models Ia-Ic are given in Tables 35- 37. In Model Ia, the media use variable incorporated is V2066. Due to the coding of V2066, as well as the coding of V2026, negative relation- ships between levels of political information and media use and cam- paign interest are hypothesized. All other variables in the first equation of Table 35 are hypothesized to be positively related to levels of political information. The model is able to account for about 11% of the variation in Info, a sub5tantial improvement over the political information equations examined with the 1972 data. Further, the Info regression equation as a whole is statistically significant. While all coefficients except for that of V2549 are in the hypothe- sized direction, only V2026 and V2066 are statistically significant predictors of information levels. The negative coefficient for V2066 indicates that newspaper readers have higher levels of political infor- mation than do television viewers. Thus, a key proposition derived from the theory of citizen information gathering is confirmed by these data. The probit equation for V2066 in Table 35 is significantly dif- ferent from a random model, and is able to predict correctly about 47% 0f the cases. Further, the correlation between predicted and actual 232 eeseewwwsmwm we ewe» uepweuuese .mo.w¢w mm~.u_eeuee use ueuewuese seezuee sewue—essee N_aoaccao empoweaca NNe PNNuz me.va eM.Ne.wAwe _VNSSXN- mFN.uN Nee. . aaeN.A»ew_weev- NmN.nNNNN> Nom.uwe=uee use ueuuwuese seezuee sewpepessee sweooccao eoeoweoca Ne.eN em_Pnz me.va aN.ea_que steexN- FeN.uNs Aeme.v Ame..v . a_eN.Asew_waev- «mew..w.c=meev- ee.m-e_mN> e_euz me.va N_.N_uu mNe.u.e eNP.wN AaeN.V weep.e amme._aeeeN>v+ aeNm.A_awesv+ meN.nseewaecooe= mp.uwe=uee use ueuewuese seezuee sewpe—eseee »_eoeccoo eoeoweeca Ne.ea eeenz me.va em.Nmuwwe .vseexN- me.nNs see. . wNeN.A_oweNV- Nam.ueeeN> m—wuz mo.va op.m—nm ep—.u m aem.nm ANNN.V ANNe.V Amee.v wemm.PAeeeN>v- Npe.MmamN>v- eNe.AeNmN>v+ ANNe.V APNP.V Aeee.e eme.v oeo.Awp¢N>v+ omo.Aummgmus~v+ omo.Ao~xugemv+ wo-.Am~om>v1 emm.puews_ .ewm— .ew pwuoz .mm mpeeh 233 values of V2066 is .19 which, while not overly impressive, does indi- cate a measure of explanatory power of the model. Given the coding of V2066, a positive relationship is hypothesized between levels of poli- tical information (Infol) and media use. With the results of estimat- ing the model with the 1972 data in mind, the fact that the relation- ship is in the opposite direction comes as no surprise. The Uncertainty equation, while statistically significant, can account for virtually none of the variation in the endogenous variable. The coefficients of both predictor variables are significant at the .05 level of probability; however, the coefficient for Infol is in the wrong direction. The coefficient for V2066 indicates that newspaper readers are less uncertain about the issue positions of the two major parties than are television viewers. Again, an important hypothesis regarding the effects of media utilization is confirmed by these data. The model performs best when predicting turnout (V2319) and vote choice (V2322). The probit analysis of V2319 correctly predicts 70.5% of the cases, and the correlation between predicted and actual values is equal to .30. The analysis of V2322 performs about as well, predict- ing 67% of the cases correctly, with a correlation between predicted and actual values of .29. Due to the coding of the variable Utility, a negative relationship is hypothesized between levels of utility and vote choice. This hypothesis is confirmed by the V2322 equation. Further, a negative relationship between levels of uncertainty (Unsurl) and V2319 is observed as hypothesized. The significance of the coeffi- cient for Utility in the turnout equation most likely results from the absence of the civic duty term in the 1974 equation.2 In Table 36 are presented the estimates of Model Ib for 1974. 234 Table 36. Model Ib, 1974. Info=.72 -(TVNewsl)1.158 -(V2026).424* +(PartyID).133 (.994) (.136) (.086) +(Interest).343* +(V2418).099* +(V2525).165* -(V2549).010 2 (.121) (.030) (.059) (.027) R=.346 ‘R’=.112 F=15.77 p<.05 N=819 TVNews=-.050 +(Infol).309* .2 (.063) R =.O37 -2xLLR(l df)=24.15 p<.05 N=806 50.9% predicted correctly correlation between predicted and actual=.l46 Uncertainty=l.853 +(TVNewsl)l.690* -(Infol).336* _2 (.308) (.119) R=.211 R =.O42 F=18.94 p<.05 N=819 *p5,05, one-tailed test of significance Model Ib incorporates TVNews and its instrument, TVNewsl, as the media use variables. For the Info equation, the R2 is the same as that in Model Ia. However, the media use variable, TVNewsl, has a coefficient that is statistically insignificant and in the opposite direction hy- pothesized. The coefficients for four of the predictor variables in the Info equation are statistically significant, and all of these are in the hypothesized direction. The probit analysis of TVNews presents some mixed results. While the estimated R2 is rather low, the equation is statistically signifi- cant. Further, about 51% of the cases are predicted correctly, with the correlation between predicted and actual values equal to .146. However, the political information instrumental variable, Infol, while significant at the .05 level of probability, is in the opposite direc- tion hypothesized. The instrumental variable TVNewsl is not a significant predictor 235 of political information, but is significantly related to Uncertainty. However, the coefficient for TVNewsl in the Uncertainty equation is in the opposite direction hypothesized. The Info instrument, Infol, does perform as expected; the coefficient is negative, and is statistically significant. Overall, the equation is able to account for only 4% of the variation in Uncertainty, while the regression as a whole is sig- nificant at the .05 level of probability. The final table to be presented in this chapter is Table 37. Model Ic incorporates the media use variable Paper and its instrumental variable, Paperl. While the explanatory power of the Info equation, as indicated by R2, is not much different from the information equations of the two previous models, the coefficients of none of the predictor variables are significant at the .05 level of probability. Further, only the coefficients for PartyID, V2418 and V2525 are in the hypothe- sized direction. Clearly, none of the hypotheses represented by the Info equation are supported by the data. The story does not change much when the probit analysis of Paper is considered. The equation is significantly different from a null model, and the estimated R2 of .127 is impressive when compared to the media use equations in Tables 34 and 35. However, the coefficient for the predictor variable Infol, while strong and statistically signifi- cant, is in the opposite direction hypothesized. (Finally, the Uncertainty equation of Model Ic is very similar to that of Model lb. The regression accounts for very little of the vari- ation in the dependent variable, but is statistically significant. Further, the coefficients of both predictor variables are significant at the .05 level of probability. Lastly, the coefficient for the Info 236 Table 37. Model Ic, 1974. Info=-1.068 +(V2026).138 +(PartyID).011 -(Interest).039 ( 358 083 . . (.256) +(V2418).014 +(V2525).O60 -(V2549).009 +(Paper1)1.903 2 (.063) (.064) (.026) (1.633) R=.346 R'=.112 F=15.77 p .05 N=819 Paper=.l77 +(Infol).605* 2 (.071) R =.127 -2xLLR(l df)=74.98 p .05 N=803 61.9% predicted correctly correlation between predicted and actual=.l74 Uncertainty=l.434 -(Infol).906* +(Paperl)l.830* _2 (.287) (.462) R=.166 R =.025 F=11.63 p .05 N=819 *p .05, one-tailed test of significance instrumental variable, Infol, is in the correct direction, while the coefficient for Paperl is in the opposite direction hypothesized. Discussion Examining the statistical model of the theory of citizen informa- tion gathering with the data from the 1974 CPS survey has produced re- sults remarkably similar to the evaluation of the model using the 1972 data. In this chapter, as in Chapter 7, the model performs best when predicting turnout and vote choice. However, due to the absence of civic duty measures for 1974, the predictions for V2319 are not as strong as those for the turnout variable for 1972. The 1974 data con- firmed, as did the 1972 tests, the strong negative relationship be- tween uncertainty regarding the candidates', or the parties', issue positions and the likelihood of voting. A further similarity between the 1972 and 1974 tests of the model 237 lies in the observation of positive relationships between information levels and media use in all three media use probit analyses. As noted before, these relationships are contrary to expectations. These posi- tive relationships are observed, perhaps, because the dynamic asso- ciation between information levels and media use is being examined with data collected at one point in time. If levels of information and media utilization were to be observed at different points in time, the hypothesized negative relationships might very well have been ob- served. This examination of the statistical model has been more successful than the estimations reported in Chapter 7 on at least two counts. First, all of the equations in Tables 35-37 are statistically signifi- cant, while many of the 1972 equations, particularly the probit analy- ses of V464, do not differ from a random model. Second, the media use variable Paper proved to be an impressive predictor of both levels of political information and levels of citizen uncertainty, especially when compared to the other media use variables used in this chapter, and when compared to the performance of V464 in Chapter 7. The 1974 test of the statistical model of the theory of citizen information gathering will be abandoned at this point, but will be revisited in Chapter 10. I turn now to the evaluation of the statis- tical model with data from the 1976 survey, a task which occupies Chapter 9. 238 ENDNOTES 1Given that the vote choice variable measures the party, not the candidate, for which the respondent voted, it would not be appropriate to use a differential benefit measure based on respondents' placement of candidates on the seven-point issue scales. V2322 measures the party of choice in the 1974 House of Representatives elections. How- ever, respondents had the opportunity to place only Senatorial candi- dates on the seven-point issue scales. The only sensible strategy seems to be to analyze the party voted for in the 1974 election, and to use a differential benefit measure based on the placement of the two major parties on the issue scales. The measure used, called Utility in this chapter, was described in earlier chapter as PARBEN. 2Civic duty is not included in the 1974 turnout equation because the questions necessary for the construction of this index were not in- cluded on the 1974 CPS survey. CHAPTER 9 MODEL EVALUATION FOR 1976 Introduction In this chapter the results of estimating the statistical model of the theory of citizen information gathering will be presented and discussed. The model will be examined by using the responses only of the members of the 1972—1974-1976 CPS panel survey. The results reported here are the second stage estimates from the two-stage procedure described in Chapter 6. The four equation model to be estimated is the same as that evaluated with the 1972 and 1974 data sets. The following variables are used in this analysis: Caninf: measures candidate-referenced information; Parinf: measures party-referenced information; Media Utilization: V3654; measures whether respondents relied on television, newspapers or both media as their primary source of political information about the 1976 campaign; from the 1976 CPS survey; TVNews: measures the extent of television usage to follow the 1976 campaign; Paper: measures the extent of newspaper usage to follow the 1976 campaign; 239 240 Uncertainty: measures levels of respondents' uncertainty regarding the candidates' issue positions; Turnout: V3655; from the 1976 CPS survey; Candidate Choice: V3665; from the 1976 CPS survey; Education: V3384; from the 1976 CPS survey; Social Class: V3486; from the 1976 CPS survey; Income: V3507; from the 1976 CPS survey; Campaign Interest: V3031; from the 1976 CPS survey; General Interest in Politics: Interest; previously described as POLACTIV; PartyID: measures strength of party identification; Utility: measures differential benefit based on the seven- point issue scales from the 1976 CPS survey; Diffeel: measures differential benefit based on the one hundred-point feeling thermometers from the 1976 CPS survey; Civic Duty: DTerm; based on the civic duty questions from the 1976 CPS survey. As with the 1972 and 1974 data, instrumental variables have been constructed for each endogenous variable that appears as a predictor variable in the model. Thus, instruments have not been calculated for V3655 and V3665. The details of the construction of the instru- mental variables can be found in Appendix 0. Four instruments have been constructed for each endogenous variable. These instruments are Caninfl-Caninf4 for Caninf; Parinfl-Parinf4 for Parinf; V3654a-V3654d for V3654; TVNewsl-TVNews4 for TVNews; Paperl-Paper4 for Paper; and Unsurl-Unsur4 for Uncertainty. 241 Four models will be estimated with the 1976 data, each model using a separate set of instrumental variables. For instance, Model I will incorporate Caninfl, Parinfl, V3654a, TVNewsl, Paperl and Unsurl as the instruments. Further, three versions of each model will be estimated, each version using a different media utilization variable. Models Ia, IIa, 111a and IVa will use V3654 and its instruments; Models Ib, IIb, IIIb and IVb will incorporate TVNews and its instrumental variables; and Models Ic, IIc, IIIc and IVc will use Paper and its instrumental variables. The hypotheses incorporated by the equations of the statistical model are the same as those examined in Chapters 7 and 8. In the Caninf and Parinf equations, all coefficients, with the exceptions of those for V3031 and V3654, are hypothesized to be positive. Due to the coding of V3031 and V3654, negative relationships are hypothe- sized. The relationships between Caninf and Parinf and the other media use variables, TVNews and Paper, are hypothesized to be posi- tive. For the probit analyses of V3654, positive coefficients are predicted for the political information variables. For the estima- tion of TVNews and Paper, negative coefficients for Caninf and Parinf are hypothesized. However, most of these relationships tUrn out to be in the opposite direction hypothesized, due to the necessity of estimating dynamic relationships with data collected at a single point in time. Uncertainty is hypothesized to be a negative function of the political information measures, as well as of TVNews and Paper. Due to the coding of V3654, a positive association is predicted between 242 Uncertainty and V3654. For the turnout (V3655) equations, a negative coefficient for civic duty (DTerm), and positive coefficients for Unsurl-Unsur4 are hypothesized, due to the coding of V3655. As with the previous model evaluations, no explicit hypothesis regarding the relationship between V3655 and the differential benefit measures are offered, given that the latter two variables are designed to predict candidate choice, not turnout. Finally, V3665 is hypothesized to be a positive function of Utility, and a negative function of Diffeel. Since the results of the probit analyses of the candidate choice variable, V3665, are constant across all versions of all models, these equations will be reported only in conjunction with the esti- mates of Model Ia. The analyses of V3655, the turnout measure, vary across models, but are constant across the different versions of each model. Thus, the results of estimating V3655 will be reported only with the first version of each model. The Data Table 38 gives the results of estimating Model Ia, which incorporates V3654 as the media use variable. Even the briefest look at the Caninf and Parinf equations reveals the improvement of the explanatory power of the model for the 1976 data, in comparison with the 1972 and 1974 data. The model accounts for about 20% of the variation of each of the information measures, and both equations are statistically significant. In the Caninf equation, only the coeffic- ients for Partle and V3654a, the media use variable, are in the wrong direction. In the Parinf equation, all coefficients are in the 243 .NN._wz me.vs ew.emus mee.uNe eNN.ns A_a_.v ANNe.v , wwme._waaeem>v+ weew.A_weweaav+ mme.nseewaocooe= mN._nz me.va NN.mNus eee.uNm ANNP.V Aeme.v seam._aaamem>v+ see_.w_weweaev+ NmN.n»e=_aocoo== mmm.um mwo.u~eeaee use uepewuese seezuee sewuewewsee e_eooccao eoeoweoca Ne._e F_epnz me.va we.erAwe we seexN- Nee." s mNe. . aem_.w_sewcesv- e_m._nemee> wwo.n—e:uee use ueuewueee seezuee sewuewesseu »_eoeccoo eeeoweoca Ne._e Peep": 8.x.A me.wenwwe _V NSSXN- es." s eNe. . aFNP.A_eweweaev- mNN._uamee> NN._nz me.va mm.Neus meN.n s ema.us weem._v Aswe.v Aeee.v Ame_.v NeN.NAaemem>v- eNe.ANemm>v+ eNe.AeeeN>v+ eee.waemm>v+ AP_F.V ANeN.V ANNe.V seem.Ae_wocasv+ NeN.Aemocoeesv+ weee.A_mem>v- NNN.an=wcas NN__uz ee.va eN.eenu Ne_.n.m eee.us AeNm._v AeNe.e Aeeo.e AaeP.v see.waamem>v+ e_e.ANeem>v- em_.weeam>e+ seem.weemm>e+ Aep_.v Ae_N.V ANNe.V uoo.Aa_>uweavn «nm¢.AumegmusHv+ «moo.A—mom>vn mom.puwswseu .owm— .eH pwuoz .wm mpneh 244 eeseewwwsmwm we “wee uewweu-ese .moanw .mmw.n—e:aee use ueuewuess seezuee sewue—essee epeoeceao eoeoweocs New New": me.vs NNNNmmnwwe _v «SSXN- NNN.nNs nee. . some.spoowwwev- aep.nmeem> Nwm.uweeuee use ueaewuess seezuee sewuepessee esaoaccoo empoweaca Na.eN eesnz ee.va mNNFeanwe we NSSXN- ewe." s eee. . wNae.A»ow_wu=e+ Nee.umeem> s_eooeeoo eeeoseoca N..ee Nee_nz me.va Fm.N-_wAwe me NAAXN- ae_.wNN AFNP.V APNe.V some.v . ameN.Apc=meev+ aeee.spoowwwev+ waNm.Ascawev- a_P.wmmem> wem.uwe:pee use ueuewuese seezues sewuewmssee s_oooccoo empoweoca N_.ae eae_uz ee.vs sm.NN_nAwe My «SAXN- aeP.uN AaN_.v seme.v some.e . aeNN.A_c=m==V+ e_e.seow__e=v- weNm.AesoweV- eee.ummem> .Ae.eeoov em opeaw 245 hypothesized direction. In the first political information equation, three of the seven independent variables are significant predictors of Caninf. For Parinf, only campaign interest (V3031) and PartyID have coefficients that are significant at the .05 level of probabil- ity. In the probit analyses of V3654, both Parinfl and Caninfl are significant predictors of media choice; however, both coefficients are negative, which is contrary to the hypothesis. Both equations are statistically significant and, unlike the probit analyses of the media choice variable from the 1974 data (V2066), the correlation between the predicted and actual values for the probit equation is greater than zero. In general, the model predicts media choice better for the 1976 data than for either the 1972 or 1974 data sets. The two Uncertainty equations of Model Ia are rather similar to one another in that both account for about 5% of the variation in the dependent variable, both are significantly different from the null model, and in that all coefficients in both equations are statistically significant. Further, the media choice variable, V3654a, is positively associated with Uncertainty in both equations as predicted. That is, newspaper readers exhibit lower levels of Uncertainty than do television viewers. However, the coefficients for both Parinf and Caninf are in the opposite direction hypothesized. When the statistical model was tested using the 1972 and 1974 data, the strongest performance was observed in relation to the turn- out and vote choice equations. The model performs equally well in predicting V3655 and V3665 from the 1976 data. The two turnout equations of Model Ia are virtually identical. For both equations 246 R2=.184, and 84% of the cases are predicted correctly. Further, for both probit analyses the correlation between predicted and actual values of V3655 equals .346. The coefficients for DTerm and Unsurl are in the hypothesized direction in both equations, and are sig- nificant at the .05 level of probability. Not surprisingly, the differential benefit measures do not substantially contribute to the explanatory power of the turnout equations. An examination of the candidate choice equations in Table 38 reveals that Diffeel is a substantially better predictor of candi- date choice than is Utility. However, the summary statistics of both equations, especially the correlation between predicted and actual values, indicate that both Diffeel and Utility have a great deal of power in explaining for which candidate respondents cast their votes.) Utility explains about 58% of the variation in candi- date choice, while Diffeel accounts for 79% of the variation in V3665. Both equations are statistically significant. Table 39 contains the results of estimating Model Ib. Both political information equations account for about 20% of the varia- tion in the endogenous variable, and both are statistically sig- nificant. The coefficients of very few of the predictor variables are statistically significant, but those that are significant are related to political information as predicted. The media use variable, TVNewsl, is not a significant predictor of information levels. Although the coefficients for TVNews are large in both equations, the size of the standard errors renders these coefficients statistically insignificant. This is due to high collinearity be- tween TVNewsl and the other right-hand-side variables in the Caninf 247 eeseewwwsmwm we awe» uewwewuese .mofleu MNwwuz me.va mm.mNnu eme.uNe eN.ns weaw.e weNe.e seem.wazaz>we+ «wew.wwwswsase- Nmm.wuwwewawcooe= NNwwuz ee.va wm.NNus wee.uNs eew.us wmww.e weNe.e wwa.wazoz>we+ ame.wwweweaee- one.wuwwewawcoo== w—o.n_.e:uue ace umuuwvwga :mmzumn :owuepmwgou sweoaccoo eowoweocs we.Ne wwewuz me.va we.emWwwe we seeXN- Nae.nNs mNe. . . awew.www=wcase+ wee.nmzoz>w o.on_.e:uum uce “59025.3 :wwzpmn cowuepmgwou wweooscao eowoweoss we.Ne wwewnz me.va wmm.ewuwwe we seeXN- eNe.nNs eNe. . weee.www=w=eee + Nee.umzaz>w mewnz me.va mm.Nens meN.- s ema.us wwwN.ee wwee.e waaw.e wmmw.e -N- eNe.ewmsoz>we+ mee.wwemmwe- sNem.weeam>e+ seem.waemmwe+ wsmw.e wmmN.e wmeN.we Nme.wewwwsese- mew.wwmecew=we+ eae.wwwmem>e+ wee.mw-nwewcas NNwwuz ee.va eN.eens Now." a mea.ne szN.ee wwee.e waaw.e wemw.e .1 wma.quzez>we- eme.wwemm>e+ ame.weeam>e+ NwN.waemm>e+ wwmw.e wNmN.e weeN.we mom.AoH>useee+ «wm¢.Aumewwus~e+ woo.pwpmom>ew onc.wuwswseu .wwmp .nw —ouoz .mm m—new 248 and Parinf equations. In particular, the correlation between TVNewsl and V3031, equal to -.899, leads to the coefficients of both varia- bles being statistically insignificant in both information equations} The probit analyses of TVNews are not impressive. Although both equations are statistically significant, the explanatory power of Caninfl and Parinfl in these equations is virtually nonexistent. The coefficients for both political information variables are signif- icant at the .05 level of probability; however, both coefficients are in the opposite direction hypothesized. The explanatory power of the Uncertainty equations is equally unimpressive. Both equations account for approximately 4% of the variation in Uncertainty, and both equations are statistically sig- nificant. The coefficients for Caninfl and Parinfl are in the correct direction, and are significant at the .05 level of probabil- ity. The TVNewsl coefficients are also statistically significant in the two Uncertainty equations of Model Ib; however, both are in the opposite direction hypothesized. Model Ic incorporates Paper as the media use variable. The estimates of this model are presented in Table 40. All equations in Table 40 are statistically significant. The Caninf equation accounts for about 20% of the variation in the endogenous variable. Four of the coefficients of the predictor variables are significant at the .05 level of probability, and all of these coefficients are in the direction hypothesized. The coefficient for the media use variable Paperl is not only statistically insignificant, but is also in the opposite direction hypothesized. 249 eeseewwwsmwm we pee» uewweauese .mo.we« MNwwuz me.vs Nw.wNus eee.nNe mwN.ne wmww.e wees.e seee.wwcoaese- aaem.www=wsase+ wee.wn»eewawcooe= NNwwnz me.vs Nw.wNus eee.u.m emw.ne weeN.e wwew.e see.wwwsoaaae- weea.wwweweaee+ mew.wnww=waeceo== emm.nwe:uee use ueuewuese seezuee sewuewessee wweooeeao eewoweoes ewe wwewuz me.va ewmmNNuwwe we seeXN- eem.nNs wee. . wama.www=wsase+ mw.-ucoaas mum.nweeuee use ueuewuese seezwee sewuewessee swwooceeo eowoweoca N.Ne wwewuz me.va we.aewuwwe we seexN- mwm.sz Nee. . aNea.www=weaee+ ewe.w-usoaas NNwwnz me.va am.Nanu meN.wNe ema.us wamN.we weNe.e wame.e weww.e meN.wwwcaaase+ mee.wwemm>e+ ammw.weeamwe+ Nae.waemmwe- wwew.e wweN.e weNm.e somm.wowauseee+ ~¢_.Aamwsmus~e+ mmF.Apmom>eu mow.—uwswsee NNwwuz ee.va eN.eans New.nNs ewe.ns weew.e weeN.e weNm.e eme.wewweease+ wowe.wwmacoeewe+ aeNN.wwmem>e1 weN.Nuweweee .owmw .uH wmuoz .ou w—eew 250 The Parinf equation also accounts for about 20% of the variation in political information levels. Only two of the coefficients in this equation are statistically significant, but both of these are in the hypothesized direction. The coefficient for Paperl, while in the correct direction, is statistically insignificant, due to collinearity between it and the other predictor variables in the Parinf equation.2 The probit analyses of Paper are much more powerful than those of V3654 and TVNews. In the first Paper equation, 68% of the cases are predicted correctly, and the correlation between predicted and actual values equals .378. In the second equation, 67% of the cases are predicted correctly, with a correlation between predicted and actual values equal to .334. The coefficients for both Parinfl and Caninfl are statistically significant; however, these coefficients are in the opposite direction hypothesized. The Uncertainty equations of Model 1c are rather unusual. In both equations the coefficients for the political information variables are statistically significant, but in the wrong direction. This is a reversal of the pattern found in Model Ib, but is the same as that in Model Ia. The coefficients for Paperl in both equations are significant at the .05 level of probability, and are in the hypothesized direction. This duplicates the pattern of Table 38, and is again contrary to the results of Table 39. The overall explanatory power of both equations is weak. In the first Uncertainty equation, R2=.024, and in the second equation the adjusted proportion of variation explained is 4.4%. 251 The estimates of Model II are presented in Tables 41443. Table 41 gives the results of estimating Model IIa, which incorporates V3654 as the media use variable. All equations in Table 41 are statistically significant. In the first political information equa- tion, approximately 20% of the variation in Caninf is explained by the model. With the exception of V3654b, all coefficients of pre- dictor variables are in the hypothesized direction, and several are statistically significant. The media choice variable V3654b performs much better in the Parinf equation. Not only is the coefficient for V3654b in the correct direction, it is also statistically significant. The only other variables whose coefficients are statistically significant are the campaign interest variable (V3031) and strength of party identi- fication (PartyID). The equation as a whole explains about 20% of the variation in Parinf. I The probit analyses of V3654 as functions of Caninf2 and Parinf2 are similar to the equations in Model Ia. Both equations correctly predict about 61% of the cases correctly, with a small correlation between predicted and actual values. The coefficients for both Caninf2 and Parinf2 are significant at the .05 level of probability, but are in the opposite direction hypothesized. The two Uncertainty equations of Model IIa can account for very little of the variation in the endogenous variable. The coefficients of all predictor variables are statistically significant, but those for the political information instruments are in the opposite direc- tion hypothesized. The positive relationship between media choice (V3654b) and Uncertainty indicates that newspaper readers were more 252 mewnz ee.vs ee.emns me.nNm eNN.ns wmmw.e weNe.e see.wweamem>e+ weaw.wNw=wcase+ eme.wuwwewawsaoe= MNwwuz me.va me.ers Nae.uNe waew.e weme.e swam.wweamemwe+ swew.wNweweaee+ eN.uwwewaosoo== NN.nm mnompmsaoe use Uwuuwvmwa :mmzn—wn cowampmwwou wweooccoo eowoweoca em.we wwewuz me.vs Me.erwwe we NeSXN- wee.uNs wNe. . wemw.wNwewcase- NNN.wuemem> swompezuoe tcm non—3693 :mmzama cowue—mgwou wwwoeccoo eoeoweoea Ne.we wwewuz me.va Mw.wewwwe we seeXN- mme.nNe wNe. . wwww.wNwew=aee- me.wnameN> NNwwwz me.va Nm.Nans eeN.n.m ema.ns wNew.e wemm.we wmwe.e wmee.e ewe.waemmwe- seem.Nweamem>e- mNe.wwemN>e+ awe.weeem>e+ wmew.e weeN.e wNwe.e aewm.wewwwsaae+ eNN.wwmeeowewe+ weee.wwmem>e- www.mnwewcas mewnz ee.va wN.oenu New.nNe me.ns weee.we wowe.e wmee.e wwa.weaeeN>e+ ewe.wwemm> + er.weeam>e+ wNww.e wmeN.e wNwe. weee.wewwesase+ aNNe.weaacaeewe+ seee.wwNeN>e- an.Nuw=weae .onmp .eww pmuoz .pe m—neh 253 euseewwwsmwm we ewe» uewweeiese .mones oem.uwe=uee use uewewuese seezeee sewuewessee wweooccoo eoeoweocs ww.ee eaewuz me.va mm.wquwwe me NeeXN- New.uNs weww.e wwee.e wwme.e . NNN.sz=me=e+ Neee.wwoowwwee+ «ewe.w5cowee- wwN.nmeem> mem.uwe=uue use ueuewuese seezaee sewuewessee wweoeecao eowoweoes ww.ee eeewnz ee.ve em.wquwwe me NeeXN- New.uNs weww.e wemo.e wwme.e . eaN.ch=meee+ Nwe.wsewwww=e- wwwm.wesowee- New.uemem> .we.oeooe we oweaw 254 eeseewwwsmwm we new» uewweu-ese .mo.wma NNwwnz me.vs em.eNus ee.uNm meN.us weew.e wNNe.e seww.wNm3oz>we- weww.wNwew=aee- we.wuwwewawsooee MNwwuz me.va me.NNus wee.wNe New.ne wer.e waNe.e ste.wNusez>we+ ser.wNwewcase- wNa.wnwe=waocooe= noo.uwe:uee use ueuewuese semzpee sewuewessee xwuumssee uwuuwumse aw.mo p—opnz mo.va Mm.mmwwwu we meexm- muo.u~m mmo. < «Nm~.ANwswseme+ wmm.um3oz>w o.ouwe=uee use uewewuese seezpee sewuewessee wwwooccoo eowoweoca Ne.Ne wwewnz me.va wwe.awuwwe we seeXN- Ne.uNs eNe. . wwme.wNw=weaee+ ewe.nmzoz>w mewuz me.va wN.Nauu meN.u e eea.ue wewN.we wewe.e wmee.e wNae.e ea.wwmsaz>we- seme.wsemm>e+ wee.weeam>e+ awww.waemm>e+ wmww.e weww.e wNaN.e «mwm.wowzusese+ «mom.wummsmuswe+ «Nwm.Apmom>eu «mo.enwswsee mewnz me.va eN.eeus www.6Ne mee.ns wewN.we wewe.e waee.e wNae.e emm.wmsaz>we- wwe.wwemm>e+ eme.weeem>e+ sNeN.weemm>e+ weww.e wwww.e waaN.e mwo.Aowzuweme+ *mwm.fiumwsmusme+ amen.wpmom>v1 mum.mnwswseu .onmp .eew pmuoz .me mpneh 255 eeseewwwsmwm we ewe» uewwewuese .mOaWew NNwwuz me.va me.NNus wee.nNm emw.ws wmee.e wNae.e seem.choaase- amaN.wNwewcase+ wwN.Nu»eewawcooe= mewuz me.va eN.ewus ewe.uNw sew.us wNew.e wowe.e sawm.cheaase- aeNN.wNw=w=aee+ we.wnweewawcooee mmmm—ewfiue use .5»qung :wmzumn :owuepmwwou wwwoeecao eoeoweocs wN.we wwewnz me.va ew.weNWwwe we seeXN- NNN.nNs wee. . ueue.wmwswsese+ wwm.-useees wmm.uwe=pee use ueuewuese seezuee seweewessee x—uuwssou umuuwumwn Rm.mm ppopnz mo.va “m.pkuawu we meexmu Npm.nN¢ pmo. < heme.wmwswseue+ mem.w-useees MNwwnz me.vs Nw.Nens aeN.nNm Nea.ns wwmm.e weNe.e weme.e wwee.e eem.chosase- eaee.wwemmwe+ weww.weeem>e+ wwew.waemm>e+ weme.e wmmw.e wwa.e wema.wewwwcaae+ weee.wwmoeowewe+ «New.wwmem>e- eww.wuwewcaa NNwwuz me.va wm.eans New.uNe. aaa.ue wNme.e wwNe.e weme.e wwee.e wwN.wNeoaase- ewe.wwomm>e+ smew.weeem>e+ «mom.waemmwe+ wwme.e waew.e wmmw.e omo.waea»seee+ «mme.wumewmuswe+ «owo.wpmom>en mmm.mnwswseo .owmp .ewe wmuoz .mv mwee» 256 certain of the candidates' issue positions in 1976 than were tele- vision viewers. The only significant predictor of V3655 in the turnout equation is civic duty (DTerm), whose coefficient is also in the direction hypothesized. The Uncertainty instrumental variable, Unsur2, is not a significant predictor of voting in either equation, nor are the measures of differential benefit. Both probit analyses of turnout predict over 84% of the cases correctly, with substantial corre- lations between predicted and actual values of V3655. The estimates of Model IIb, which incorporates TVNews as the media use variable, are presented in Table 42. All equations of Model IIb are significant at the .05 level of probability. With the exception of TVNewsZ, the coefficients of all predictors in both equations are in the hypothesized direction. Further, the coeffic- ients for V3031, Interest and V3384 in the Caninf equation are statistically significant. Overall, the equation accounts for about 20% of the variation in Caninf. The Parinf equation also accounts for about 20% of the variation in the endogenous variable. The coefficients of five of the seven predictor variables are significant at the .05 level of probability. Further, these five coefficients are all in the hypothesized direc- tion. However, the coefficient for the media use variable, TVNews2, is not only statistically insignificant, but is also in the opposite direction hypothesized. Both probit analyses of TVNews perform poorly. Although 62.8% of the cases are correctly predicted by both equations, there is no correlation between the predicted and actual values of TVNews. 257 Further, while Caninf2 and Parinf2 are significant predictors of TVNews, their coefficients are in the opposite direction hypothesized. Both Uncertainty equations explain about 4% of the variation in the dependent variable; the explanatory power of these equations is comparable to those of Model IIa. All coefficients in the equations are statistically significant, and those for Parinf2 and Caninf2 are in the correct direction. However, the coefficients for TVNews2 are in the opposite direction predicted. The estimates for the final version of Model II, Model IIc, are presented in Table 43. Model IIc incorporates Paper and its instru- ment, Paper2, as the media use variables. All equations of Model IIc are significant at the .05 level of probability. Paper2 does not perform well as a predictor of levels of political information in either the Caninf or the Parinf equations. In both equations the coefficients for Paper2 are statistically insignificant and in the opposite direction hypothesized. Further, in both equations the only coefficients in the wrong direction are those for Paper2. The other predictor variables in the political information equations perform well. In the Caninf equation, four of the coeffic- ients are statistically significant. In the Parinf equation all coefficients, except for the Paper2 coefficient, are significant at the .05 level of probability. Both equations explain about 20% of the variation in their respective endogenous variables. The probit analyses of Paper, utilizing Caninf2 and Parinf2, have substantial explanatory power. Both equations predict over 67% of the values of Paper correctly. Further, the correlations between the predicted and actual values for both equations exceed .33. 258 However, while the coefficients for Caninf2 and Parinf2 are statis- tically significant, they are in the opposite direction hypothesized. The two Uncertainty equations of Model IIc can account for virtually none of the variation in the endogenous variable. While all coefficients are statistically significant, those for Parinf2 and Caninf2 are in the wrong direction. The negative relationships between Paper2 and Uncertainty indicate that uncertainty regarding the candidates' issue positions decreases as newspapers are more heavily utilized as a source of political information. Tables 44-46 contain the results of estimating the statistical model with the third set of instrumental variables. Estimates of Model IIIa, which incorporates V3654 and V3654c as the media use variables, are given in Table 44. All equations of Model IIIa are significant at the .05 level of probability. The Caninf and Parinf equations in Table 44 are similar to one another in a number of ways. First, both explain about 20% of the variation in levels of political information. Second, with the ex- ception of V3654c, the coefficients for all predictor variables are in the hypothesized direction. Third, the standard errors of the coefficients of V3654c are very large; this is due, in part, to sub- stantial correlations between V3654c and V3384 and Interest.3 Further, the coefficients for Interest and V3384 are statistically significant in both equations. The only point of difference between the two equations is that income and social class (V3507 and V3486) are significant predictors of Parinf, but not of Caninf. The probit analyses of V3654 are virtually identical to the analyses of media choice in Models Ia and IIa. Both equations pre— 259 NNwwnz me.va ee.ens we.u.e eew.ue wmew.e wwNe.e sewe.woemem>e+ weee.wmw=wcase+ woN.wnww=wawsoo== NNwwuz me.va mN.eus eee.wNm wee.us wmww.e weNe.e seem.woamem>e+ eee.wmweweaee- eme.wwweewaweaoe= mwo.uwe:uee use umuewuese seezuee sewpewessee wwwooccoo eowoweaca we.we wwewuz me.vs mw.NNwwwe we NeexN- wee.uNs eNe. . «New.wmwewease- awm.wnamem> wwo.uwe:uee use ueuewuese seezuee sewue—essee wwwoeeeao eawoweasa we.we wwewuz me.vs www.quwwe we meSXN- ee.uNs NNe. . . wwww.wmweweaee- New.wuamem> NNwwuz me.va me.Naus aeN.n.e wma.us wwam.Ne wewe.e wewe.e weme.e mew.woamemwe+ seme.wwemm>e+ wwmw.weeem>e+ swew.weemm>e+ weaN.e wemw.e wmme.e wmm.wewwwcase+ seea.woaocawewe+ Nem.wwmemwe- eNe.-nwewsaa mwp—nz mo.va mm.oeuw wm—.umm mvv.nm wemm.Ne wewo.e wewe.e weee.e emw.woamem>e+ ewe.wwemN>e+ eew.weeemwe+ wwwN.waemmwe+ wwaN.e wemw.e waee.e Noo.AoHXuseme+ «mom.Aummsmu:~e+ mum.Awmom>eu umm.~uwswseu .mwmp .ewww Pwuoz .ev mwneh 260 eeseewwwsmwm we peep uewweuuese .moNWew uem.nwe:uee use ueuewuese seezuee sewuewessee wweoaccoo eoooweoca ww.ee eeewwz me.va ee.Nanwwe me seSXN- eew.uNs wwww.e wwee.e wome.e . awem.wmc=me=e+ meee.wwoowwwee+ aeNe.w5cowee- wee.nmeem> oem.uwe=aee use ueuewuese seezees sewpewessee wwwoaccao eawoweaca Nw.ee Neewnz me.va Ns.Nquwwe me seeXN- aew.uNN waew.e weme.e wome.e . «wem.wme=m==e+ eee.wwwwwww=e- weNm.wssawee- ewe.uemem> .we.oeoee as aweaw 261 predict about 61% of the cases correctly, but with little correlation between the predicted and actual values of V3654. While both Caninf3 and Parinf3 are significant predictors of V3654, the coeffic- ients for both variables are in the opposite direction hypothesized. Neither Uncertainty equation is able to explain the variation in the dependent variable. In the first Uncertainty equation, both coefficients are in the predicted direction, but only the coefficient for V3654c is statistically significant. In the second equation, both coefficients are significantly different from zero, but the coefficient for Parinf3 is in the wrong direction. The positive re- lationship between V3654c and Uncertainty confirms the proposition that television viewers are more uncertain about the issue stands of the candidates than are newspaper readers. The V3655 equations reaffirm the ability of the model to predict turnout. The two equations are virtually identical; 84.1% of the cases are predicted correctly by both probit analyses, with a cor- relation between predicted and actual values equal to .346. In both equations civic duty (DTerm) is significantly related to voting, as is level of uncertainty. The higher a citizen's level of civic duty, and the lower is his/her level of uncertainty regarding the issue positions of the candidates, the more likely it was for that citizen to have voted in the 1976 election. The results of estimating Model IIIb, which incorporates TVNews as the media use variable, are reported in Table 45. As before, all equations of the model are significant at the .05 level of prob- ability. There is only one difference between the Caninf and Parinf equations of Model IIIb and those of Model IIIa: in the information 262 eeseewwwsmwm we amen uewweuuese .mOaWew NNwwuz ee.va em.ean eae.uNm NN.us wwmw.e wwNe.e wees.wmmzaz>we+ smaw.wmwewsese- em.wuweewawcooe= MNwwuz me.va mN.Nan ewe.uNe mwN.nN wsew.e wNNe.e weew.wmmzaz>we+ szw.wNw=w=aee- eee.wnwe=waacooee mwe.uweeuee use ueuewuese seezeee sewuewessee wweooccoo eowoweoca we.Ne wwewnz me.vs mw.emwwwe we seSXN- mae.uNs eNe. . awew.wmw=wcaae+ wee.wmsoz>w o.ouwe:aee use ueuewuess seezuee sewuewessee xpuuwssou uepuwumsn xm.~o Fpopuz mo.va Wm.¢pwfiwu we meexmu omo.umm owe. a ammo.fimwswseue+ pem.um3mz>w mewnz ee.va me.Nens aeN.uNe wma.ns wweN.Ne wewe.e wewe.e wmme.e mew.wmmzaz>we+ seme.wwemm>e+ swmw.weeam>e+ awew.weemm>e+ weoN.e wemw.e weme.e wmm.wewswcase+ seea.womecewewe+ eem.wwmem>e- ewe.-nwewcaa MNwwuz me.va eN.eaus www.1Nm mee.us wmmN.Ne wewe.e wewe.e woee.e amw.wmmsoz>we+ wwe.wwemm>e+ New.weeam>e+ wwwN.weemm>e+ wme.e wemw.e weee.e Noo.Aawxuweee+ «mom.Aummwwus_e+ mum.wpmom>eu www.muwswseu .ewmp .awww wmuoz .mv apnew 263 equations of Table 44, the coefficients for TVNews are in the hypothesized direction. However, TVNewsB is not a significant pre- dictor of either Caninf or Parinf. In all other respects, including R', significant predictors, etc., these equations are virtually indistinguishable from the political information equation of Table 44. Similarly, the probit analyses of TVNews are as unimpressive as thesame estimates of V3654 in Model IIIa. While the coefficients of both Caninf3 and Parinf3 are statistically significant, both are in the opposite direction hypothesized. Further, there is no corre- lation between the values of TVNews predicted by the probit model and the actual values of the dependent variable. Finally, the Uncertainty equations of Model IIIb are nearly identical to one another. Both explain the same proportion of varia- tion in the dependent variable. In both equations, all coefficients are significant at the .05 level of probability. In both equations, Caninf3 and Parinf3 are negatively related to Uncertainty, as hypothesized. And in both equations, TVNews is positively related to Uncertainty, a relationship that is contrary to expectations. In Table 46 are presented the estimates of Model IIIc, which uses Paper as the media use variable. All equations in Table 46 are significant at the .05 level of probability. The Caninf equation accounts for about 20% of the variation of the dependent variable, as does the Parinf equation. The coefficients of all predictor variables in both equations are in the hypothesized direction. In both equations, the coefficients for V3031 (campaign interest) and V3486 (social class) are statistically significant. Additionally, 264 eeseuwwwsawm we new» uewweuuese .mONWQw . mewuz ee.vs ww.eNus eee.nNm wwN.ns wNww.e wmme.e sewe.wmsaaaae- seem.wmwewsase+ wsm.wuwe=wawsoo== MNwwnz ee.va MN.wwus ewe.n.m eaw.ns weNN.e weew.e ewmm.wesoseae- swam.wmweweaee+ mmN.wnwwewawcooe= Nem.u—e=uee use ueuewuese seezuee sewuewewsee wwwooscao empoweocs em.we wwewnz ee.vs mM.aNNwwwe we seeXN- me.uNs wee. . weme.wmwewcase+ New.-ucaaes wm.uwe:uee use uepewuese seezwee sewuewessee wwwoocsao eowoweoss wee wwewwz me.va mm.eewuwwe we seexN- ewN.nNs Nee. . smee.wmwswseee+ mNe.w-useses NNwwuz me.va ee.Naus aeN.uNm wee.ns wmww.we waNe.e wwme.e weew.e Nmm.wmcesase+ eme.wwemmwe+ weww.weeem>e+ eme.waeemwe+ weew.e wwwN.e weem.e weee.wewwwsaae+ ema.wwaocowswe+ seem.wwmem>e- Nwe.wnwewses NNwwnz me.vs eN.eeus New.nNm aaa.us weww.we wmee.e wome.e weew.e wee.wmcaaase+ «mee.wwemm>e+ weew.weeamwe+ mmN.waemm>e+ wmew.e wwwN.e wwem.e ewe.wewwwcaae+ Nmm.wwmocoeewe+ seem.wwmem>e- eem.aneweae .mwm— .uwww wmuez .oe mweeh 265 income (V3507) is a significant predictor of Caninf, while PartyID is a significant predictor of party-referenced information. Finally, in neither equation is the coefficient for the media use variable, Paper3, significant at the .05 level of probability. In the probit equations predicting Paper, both Caninf3 and Parinf3 are substantively and statistically significant predictors of newspaper use to follow the 1976 campaign. Unfortunately, both coefficients are in the wrong direction. Beyond this, the probit analyses of Paper are rather impressive. Both equations predict about 68% of the values of Paper correctly. Further, the correla- tions between predicted and actual values of Paper in both equations exceed .34. Neither Uncertainty equation performs well in predicting level of the dependent variable. In the first equation only about 2% of the variation in Uncertainty is explained. The second equation does little better, accounting for about four percent of the variation in the endogenous variable. The coefficients for all predictor variables in the two equations are significant at the .05 level of probability. The coefficients for the media use variable, Paper3, are in the hypothesized direction. However, in both equations the relationships between levels of political information and levels of Uncertainty are contrary to expectations. lThe final version of the statistical model is Model IV, which uses the fourth set of instrumental variables. Estimates of the three versions of Model IV are presented in Tables 47-49. In Table 47 are the results of estimating Model IVa; this model uses V3654 and V3654d as the media use variables. All equations in Table 47 are 266 MNwwnz me.vs ww.ean eae.uNm mwN.ns wmaw.e weNe.e sewe.wweemem>e+ weaw.waw=wsase+ wee.nwwewawcoos= NNwwnz me.va me.mNuu Nee.n.e wN.ne wmaw.e weee.e amNN.wweemeN>e+ seew.wwweweaee+ meN.nwwewawcooe= muompezuoe Ewe “590.693 5.5252“ cowue—mwwou wweooccao eowoweoss we.we wwewnz ee.va Nm.me%we we NeeXN- wee." s , mNe. . wemw.wawewcaae- wm.wwememw wmo.nwe:uee use ueaewuese seezuee sewuewessee wweooecoo eowoweeca Ne.we wwewwz me.va wNN.anwwe we seeXN- ee.nNs wNe. - swww.weweweaee- me.wuamem> mNP—nz mo.va mo.~¢uw com." a ~me.um wwNe.e weew.e wwew.e wer.e eme.wwemmwe+ Nee.weeam>e+ Nee.weemmwe+ sewa.wewwwsase+ wamN.e wewe.e weem.Ne emm.wwmosowewe+ sze.wwmom>e- «em.wweemem>e- mee.mnweweas NNwwwz me.va eN.eans New.uNm maa.ws wwNe.e weww.e wwew.e wer.e ewe.wweemwe+ Nae.weeamwe+ eweN.weemm>e+ eme.wewwwsase+ wNeN.e wewe.e wwwN.Ne mmm.numwsmuswv+ «mom.awmom>ea cmm.Auemom>en w¢.muws—seu .owmw .e>H wwuoz .we m—new 267 .eeseewwwsmwm we ewe» uewwewiese .monQw eem.uwe=eee use ueuewuess seezuee sewwewessee wwwooscoo eowoweaea ww.ee eeewuz me.vs ew.wquwwe Ne seeXN- New.nNs wmww.e wwee.e wwme.e . eaN.wec=me=e+ Neee.wwoowwwee+ wewm.w2cowee- Nww.ummem> oem.uwe:uee use uewewuese seezuee sewuewesseu wwwoeccoo eowoweoea Nw.ae eaewuz ee.va ew.wquwwe me seeXN- New.nNN wwww.e wane.e wwee.e . meN.wac=me=e+ Nee.wwwwwwo=e- smwm.wscowee- aew.uemem> .we.w=oee we oweaw 268 eeseewwwsmwm we amen uewweuuese .monQs MNwwuz me.vs me.wan eme.uNs. sew.ue wemw.e wsNe.e sewe.wemzoz>we+ wmww.wewewsaae- mm.wnww=waecooe= MNwwnz me.va Ne.eNus ee.uNm eeN.ns weew.e wNNe.e weew.wemsoz>we+.wwww.wewsweaee- wee.wnww=wawsoo== wwwoessoo eewoweoca NN.Ne wwewnz me.va we.erwwe we seexN- wee.uNs eNe. . swmw.wew=wcaae+ Nwe.umzoz>w o.ouwe:uee use umuewuese seezwee sewwewessee wweoacseo empoweaca Ne.Ne wwewnz me.vs wee.awuwwe we eeexN- Ne.uNm eNe. . awme.wew=w=aee+ wem.umzoz>w NNwwuz me.va NN.Neus eeN.u.m ema.ne wewe.e wNee.e wwee.e weew.e seme.wwemmwe+ aae.weeemwe+ eer.waemm>e+ wNmm.wewweeaae+ wmww.e wwwN.e wee.we seem.wwaocooewe+ some.wwmem>e- aew.wwaasaz>we- wma.muw=wcas MNwwuz me.va em.eanu New.nNm mee.us wewe.e wNee.e wwae.e weew.e wwe.wwemm>e+ eee.weeem>e+ wweN.weemN>e+ seem.wewwwcase+ weww.e wewN.e wmee.we seem.wwmacawewe+ wwee.wwmemwe- eem.wemzoz>we- eee.mwwsw=ae .owmw .e>w wmuez .we m—new 269 eeseewwwsmwm we umee uewweuiese .moNWsw NNwwuz me.vs ae.ewuu Nee.u.m aew.us waee.e wmee.e weeN.wew=wsase+ seem.wacesase- eww.anwewaweooee NNwwuz we.va mm.wus Nwe.u.e mww.ns wNwe.e wemw.e amew.waweweaee+ sNaa.weceaase- sNa.wuwoewaweao== Nmm.uwe:uee use ueuewuese seezuee sewuewessee wwwoaccoo eowoweoca Nw.we wwewuz ee.va eM.mwNwwwe we seSXN- eeN.uNs wee. . seNe.wewswsese+ mew.-nseses wem.uwe=wee use ueauwuese seezwee sewuewessee ew.w..wee.w.see ueuuwumLa am.wm wwowuz mo.ve Mm.mewuwwu we meexmu vwm.nmm Nmo. a «mme.wewswseue+ mom.w1uwmeee mewnz mo.va mm.~¢uw mo~.nmn mm¢.um wwNe.e weme.e wmee.e ..weee.e «mae.wwemm> + weww.weeem>e+ womN.weemm>e wesa.wewwesase+ wwmw. wemw.e wmmm.e *wmo.wumwseuswe+ woww.wwmom>e1 mum.w¢smeeeeu mwo.wuwswsee mmwwnz mo.ve wN.ovnw wow.1 u mue.um :Noe see . see see -N. ewo.Awomm>e+ «Now.wmwem>e+ aom~.wewmm>e+ wNo.AowAuseee+ wwmw.e wamw.e wwmm.e «ow¢.wummsmuswe+ «emm.wwmom>eu wo~.w¢smeeee1 om.muwswseu .mwow .u>w weuoz .me eweew 270 significant at the .05 level of probability. The two information equations are similar to those in the previous models, in that ap- proximately 20% of the variation in Caninf and Parinf are explained by the equations. The coefficients of all predictor variables are in the hypothesized direction, although only two in each equation are significant at the .05 level of probability. V3031, which measures campaign interest, is a significant predictor of information levels in both the Caninf and Parinf equations. Further, V3384 (income) is a significant predictor of Caninf, as is PartyID of Parinf. The probit analyses of V3654 are identical to those in Model IIIa. Both equations predict about 61% of the cases correctly, with very low correlations between predicted and actual values. While the coefficients for Caninf and Parinf are statistically significant, they are in the wrong direction. ’ The two Uncertainty equations each explain about four percent of the variation in the endogenous variable. The coefficients for all predictor variables are statistically significant, but those for Parinf4 and Caninf4 are in the opposite direction hypothesized. The positive relationship between V3654d and Uncertainty in both equa- tions lends additional empirical evidence to the proposition that newspaper readers were less uncertain of the candidates' issue stands in 1976 than were television viewers. The two turnout equations each predict 84% of the cases correctly, with a substantial correlation between predicted and actual values. While the coefficients for DTerm and Unsur4 are in the hypothesized direction in both equations, Only the coefficient 271 for DTerm is statistically significant. At least in Model IVa, citizen uncertainty is not a significant predictor of turnout. The estimates of Model IVb are presented in Table 48. The media use variable in this model is TVNews. All equations in the model are statistically significant. In the Caninf and Parinf equations, each of which explain about 20% of the variation of the endogenous variable, all coefficients but for TVNews4 are in the hypothesized direction. Further, in both equations the same pre- dictor variables, V3031, Interest, PartyID and V3384 have coeffic- ients that are significant at the .05 level of probability. In addition, V3507 is a significant predictor of Parinf. The probit analyses of TVNews are not as impressive as the information equations. While both equations predict about 63% of the cases correctly, there is no correlation between predicted and actual values. The coefficients for Parinf4 and Caninf4 in the TVNews equations are statistically significant, but are in the opposite direction hypothesized. Finally, in the Uncertainty equations, each of which accounts for about 4% of the variation in Uncertainty, all coefficients are statistically significant. However, only the coef- ficients for the political information measures are in the hypothe- sized direction. The final model to be estimated is Model IVc, which is presented in Table 49. All equations of the model are significant at the .05 level of probability. Again, the Caninf and Parinf equations are able to account for about 20% of the variation in the endogenous variables. In the Caninf equation, the coefficients for all vari- ables but Paper4 are in the hypothesized direction, and of these four 272 coefficients are statistically significant. Further, again with the exception of Paper4, all coefficients of predictor variables in the Parinf equation are in the correct direction; in addition, all six of these coefficients are statistically significant. As with the previous probit analyses of Paper, the equations of Model IVc are able to predict a substantial proportion of the cases correctly. Further, the correlation between the predicted and actual values of the first Paper equation equals .367, and in the second is equal to .332. The coefficients for both Caninf4 and Parinf4 are statistically significant, but in the opposite direction hypothesized. The two Uncertainty equations can account for virtually none of the variation in levels of the endogenous variable. The coefficients of all predictor variables are significant at the .05 level of probability; however, the coefficients for Caninf4 and Parinf4 are in the wrong direction. The negative relationships between Paper4 and Uncertainty indicate that as the use of newspapers to follow the 1976 presidential campaign increased, levels of uncertainty regarding the candidates' issue positions decreased. Discussion Compared to the 1972 and 1974 analyses of the statistical model of the theory of citizen information gathering, the examination of the model with the 1976 data has been somewhat successful. However, not all hypotheses represented by the model have been confirmed by the data. To detail some of the strengths and weaknesses of the model evaluation for 1976, I will briefly examine each equation. 273 The first equations to be discussed are the political informa- tion equations. In all versions of the statistical model, the Caninf and Parinf equations accounted for about 20% of the variation in the dependent variables. This represents a level of explanatory power far greater than that displayed by the political information equa- tions with the 1972 and 1974 data. Moreover, several predictor variables were consistent in their being significant predictors of political information levels. For Caninf, a citizen's level of general political interest (Interest), campaign interest (V3031) and education (V3384) were the most important predictor variables. The other three socio-economic-political variables were relatively unimportant in explaining Caninf. On the other hand, all six of the socio-economic-political variables were important predictors of Parinf. In only one case was a media use variable a significant predictor of political information; the coefficient for V3654b was a significant predictor of Parinf in Model IIa. The reason for the lack of significance of the media use variables is the high multi- collinearity between the media use instruments and the other explanatory variables in the Parinf and Caninf equations. In many equations the coefficients for the media use instruments were of substantial magnitude, and in the direction hypothesized. However, the standard errors of these estimates were so large as to preclude statistical significance. Data problems also emerged in the probit analyses of the media use variables. Unlike the 1972 and 1974 tests, all media use probit equations were significant at the .05 level of probability. Further, 274 the model was able to correctly predict a substantial proportion of the values of the media use variables. However, in many cases this predictive capability was an artifact of the skewed distributions of the media use variables: witness the zero correlations between predicted and actual values for the TVNews equations. 0n the other hand, significant correlations between predicted and actual values were observed for the Paper equations. The correlations between pre- dicted and actual values for the V3654 equations were not impressive. However, when compared with the predicted/actual correlations of zero for the probit analyses of the media choice variable from the 1974 data (V2066), the size of these correlations gains significance. The most common problem encountered in predicting levels of media use was that all coefficients for the Parinf and Caninf instru- mental variables were in the opposite direction hypothesized. It is likely that this results from not having data collected at different points in time. The model stipulates that media use at one point in time is positively related to levels of information at the next time point. This hypothesis is incorporated in the Parinf and Caninf equations. However, the media use equation hypothesizes that media use at time t is a negative function of information levels at time t-l. Having only data on media use that is coterminous with the data regarding levels of political information, and given that the former hypothesis would also hold if the relationships are tested with data that are static, estimates that are consistent with the former hypothesis, but not the latter, are observed. That is positive, not negative, relationships are observed. 275 Some rather unusual results were observed relative to the Uncertainty equations. The instrumental variables for TVNews per- formed poorly. While the coefficients for TVNewsl-TVNews4 were statistically significant in all Uncertainty equations, they were also in the wrong direction. Further, the coefficients for the Caninf and Parinf instrumental variables, in conjunction with the TVNews instrumental variables, were in the hypothesized direction, and were statistically significant. However, the coefficients for V3654d-V3654d and the Paper instruments were all in the correct direction, and all were significant at the .05 level of probability. Moreover, in conjunction with these instrumental variables, the coefficients for Parinfl-Parinf4 and Caninfl-Caninf4, while statistically significant, were all in the opposite direction hypothesized. These results suggest that television utilization, compared to newspaper usage, had no significant effect upon citizen uncertainty in 1976. Citizen uncertainty regarding the candidates' issue positions did have a strong impact upon turnout in 1976. In almost all of the V3655 equations, the coefficients for the Uncertainty instrumental variables were in the correct direction, were of substantial magni- tude, and were statistically significant. These results confirm the proposition that the more uncertain citizens are, the less likely they are to vote. Finally, using only the differential benefit measures, the model displayed a remarkable predictive capability regarding candidate choice (V3665). 276 ENDNOTES 1See Appendix E for the correlations between all predictor variables in the Caninf and Parinf equations. 2See Table 50, Appendix E. 3See Table 58, Appendix E. CHAPTER 10 CONCLUSIONS Introduction In this final chapter three goals will be pursued. First, the performance of the statistical model of the theory of citizen infor- mation gathering over time will be explored. Second, some of the data problems encountered in Chapters 7-9 will be discussed in a more systematic fashion than in these previous chapters. Last, a brief research agenda for further exploration of the theory will be offered. The Performance of the Model over Time The five key concepts of the theory of citizen information gathering are levels of political information, levels of media utilization, levels of citizen uncertainty regarding the issue posi- tions of the candidates, turnout and candidate choice. The ability of the model to explain variations in each of these concepts will be discussed, each concept taken singly for all three years. The first equation of the model specifies that level of political information is a function of media use, social class, strength of party identification, income, education, general interest in politics and interest in the campaign. Substantial differences in 277 278 the model's ability to account for variation in levels of political information were observed across the three years. In the 1972 test, the proportion of variation explained by the information equation ranged from .02 to .035. In 1974, the R2 for the information equa- tion jumped to about 11%. For the 1976 tests, the political information equations accounted for approximately 20% of the varia- tion in the endogenous variables. Differences between these tests regarding the important explanatory variables in the information equations also emerged.1 For 1972, the most important predictor variables were strength of party identification, education and income. For 1974, strength of partisan affiliation was the strongest predictor, followed again by education and income. Substantially different patterns emerged from the evaluation of the model with the 1976 data. For the Caninf equations, campaign interest and interest in the 1976 campaign were the most important predictors of levels of political information, followed by education. For the Parinf equations, all six of the socio-economic-political variables were important in explaining levels of political informa- tion. The media use instrumental variables rarely contributed to the explanatory power of the information equations. This lack of signif- icance, both substantive and statistical, held for all three segments of the panel survey. Thus, no differences across time emerged when comparing the effects of media use on levels of politi- cal information; there were no effects, thus there were no differences. As indicated in the previous chapters, a substantial 279 portion of the weakness of the media use variables in the information equations can be accounted for by the severe collinearity between the media use instrumental variables and the other predictor variables. The problem of multicollinearity will be considered in the next section of this chapter. The above comparisons reveal substantial differences in the model between the three years. First, the set of predictor variables postulated by the theory could not account for differences in levels of political information during the 1972 campaign. However, the explanatory power of these variables rises to a relatively impressive level for the 1976 data. Second, the one predictor variable that contributed to predicting political information levels for the tests of all three years was level of education. Those with higher levels of education knew more about the campaigns of all three years than did those with lower levels of education. .Third, strong partisans had higher levels of political information than did independents and weak partisans in 1972 and 1974; however, few differences in infor- mation levels between strong and weak partisans and independents emerged from the 1976 examination. Finally, citizens' interest in politics and their interest in the campaign were important predictors of political information for the 1976 campaign, but not for the 1972 and 1974 campaign. (The next equations to be considered are the media use equations. As with the information equations, estimates varied widely across the three years. The analyses of the media use variables with the 1972 data were the least impressive. For instance, none of the probit analyses of V464, the television use variable, were statistically 280 significant; further, there was no correlation between predicted and actual values for these equations. The regressions of the media use variable Media were weak, but statistically significant; the propor- tion of variation of the dependent variable explained ranged from .04 to .07. Three media use measures were used in the examination of the statistical model with the 1974 data. The probit analyses of all three variables were much better able to predict media use than were the tests using the 1972 data. For all equations, substantial pro- portions of the values of the media use variables were correctly predicted. More importantly, the correlations between predicted and actual values for the these probit equations ranged from .14 to .19. The results of the 1976 tests of the media use equations were mixed. The probit analyses of TVNews, while statistically signifi- cant, were not able to improve upon a random model in predicting values of the endogenous variable. The probit analyses of V3654, the media choice variable, were statistically significant, but weak; correlations between predicted and actual values of about .08 were observed. The probit model predicting the media use variable Paper was by far the most powerful. All of the probit equations were statistically significant; more importantly, the correlations between predicted and actual values of Paper always exceeded .33. One interesting pattern emerged in the analyses of media use for all three years: in all equations, the coefficients for the infor- mation instrumental variables were in the Opposite direction hypothe- sized. I have suggested in the previous chapters that the reason for this is that the data, which are collected at one point in time for 281 each of the three surveys, are inadequate to test the dynamic relationship between information levels and media use. I will more fully discuss this problem in the next section. The next set of equations to be considered involve the estimation of uncertainty regarding the candidates' issue positions. The equations predicting levels of Uncertainty are uniformly weak across all three surveys; the proportion of variation in the endogenous variable explained by the Uncertainty equations was in- variably equal to about .03. The differences between the three years' tests center on the significance and direction of the coefficients of the predictor variables. In very few of the equations are the coefficients for both the media use and information instrumental variables in the hypothesized direction. Rather than describe the patterns found in the Uncertainty equations for all three surveys, I will focus discussion on the 1976 equations, while making brief references to the 1972 and 1974 Uncertainty equations. The 1976 Uncertainty equations indicate that newspaper use is a relatively powerful predictor of uncertainty levels. In the equa— tions involving TVNews, which measures television use to follow the campaign, the coefficients for the information instruments are in the correct direction. However, the coefficients for the TVNews instru- mental variable are in the opposite direction hypothesized. These equations indicate that higher levels of information and lower levels of television use lead to decreased levels of uncertainty. However, this pattern was reversed for the media choice variable, V3654, and the newspaper use variable, Paper. In all of the equa- tions incorporating either V3654 or Paper, the coefficients for the 282 information instruments were in the opposite direction hypothesized, while the coefficients for the instrumental variables for V3654 and Paper were in the correct direction. These results indicate that lower levels of information and increased use of newspapers (or the choice of newspapers over television) to follow the campaign lead to decreased levels of uncertainty. The 1974 media use variables parallel those for 1976. V2066 is the same as V3654, in that it measures media choice. TVNews and Paper for 1974 are the same measures as for 1976. The results of estimating Uncertainty as a function of media use and information levels were exactly the same as for the 1976 estimates of the Uncertainty equations. The 1972 analyses used different media use variables, and only one clear pattern emerged: as reliance on television as a source of political information increased, and as the number of media used to follow the campaign increased, uncer- tainty regarding the candidates' issue positions decreased. The general pattern for these tests of the Uncertainty equation are, therefore, somewhat constant. Increased newspaper use, and the choice of newspapers over television as the major campaign informa- tion source led to decreased levels of Uncertainty. Further, the increased use of television as a source of information led to increased levels of Uncertainty. However, in 1972, increased levels of television use led to decreased levels of Uncertainty. This difference is the only major distinction between the estimates of the Uncertainty equations over the course of the panel study. The probit analyses of the turnout variables for all three surveys were even more similar than the Uncertainty regressions. All 283 equations were statistically significant; further, the coefficients for virtually all civic duty (DTerm) and Uncertainty measures were statistically significant and in the hypothesized direction. The proportions of cases correctly predicted by the turnout equations were high for all three years, and varied little. For the 1972 turnout equations, approximately 81% of the cases were predicted correctly, with a correlation between predicted and actual values equal to about .28. Seventy percent of the cases were predicted correctly by the 1974 turnout equation, with a predicted/actual cor- relation of .30. Eighty-four percent of the cases were correctly predicted by the 1976 turnout equations, with a correlation between predicted and actual values of about .35. In general, citizen uncertainty and civic duty were better predictors of turnout for the 1976 campaign than for the 1972 and 1974 campaigns, but the dif- ferences are not great. 2 Finally, substantial differences were found between the three tests of the statistical model in predicting for whom, or for what party, respondents voted. In general, the model performs better for the presidential election years of 1972 and 1976 than for the off-year election of 1974. Further, differences were found in the ability of Utility and Diffeel, the differential benefit measures, to predict candidate choice in 1972 and 1976. For 1972, the proportion of cases predicted correctly by the equation using Utility was about 85%, while 91% of the cases were correctly predicted by the equation using Diffeel. With the measure Utility, the predicted/ actual correlation equalled .64. With the measure Diffeel, the correlation between predicted and actual values of the candidate 284 choice variable was .81. With the 1976 data, the candidate choice equation correctly predicted 75% of the cases using the Utility measure, and 89% of the cases using the Diffeel measure. The cor- relations between predicted and actual votes for these two equations were .51 and .78, respectively. The model's ability to predict the party of the candidate for whom respondents voted for the House of Representatives in the 1974 election was not nearly so impressive. While 67% of the cases were predicted correctly, the correlation between predicted and actual values was only .29. This correlation, while substantial, is rather weak when compared to the predicted/actual correlations for the 1972 and 1976 turnout equations. The reason for the relative weakness of the 1974 vote choice equation undoubtedly lies in the construction of the measure of differential benefit used. This measure, called Utility, is based on respondent placement of the two major parties on the seven-point issue scales. It is not clear which Democratic and Republican parties respondents were placing on the issue scales. Respondents might have been thinking of the "national" parties, the state parties, or the local Democratic and Republican organizations. Given that the vote choice equation for the 1974 data predicted the party of the House of Representatives candidate for whom respondents voted, to the extent that respondents were placing the national, or even the state, parties on the issue scales, the Utility measure would be less relevant to the vote choice variable, and the explaina- tory power of the vote choice equation would be weakened. 285 In summary, the model performs best with the 1976 data. The explanatory power of the information equations exceeds that of the information equations for the 1972 and 1974 data. The predictive capability of the newspaper use (Paper) equations, the turnout equations and the candidate choice equations exceed the power of these equations in the 1972 and 1974 tests. The model exhibited the least explanatory power in the 1972 examination; in particular, the ability of the statistical model to explain levels of political information and media use was virtually non-existent. Data Problems The evaluation of the statistical model of the theory of citizen information gathering was plagued by several difficulties with the data used in the estimation procedures. I turn now to a discussion of each of these difficulties. » The first data problem was encountered in the estimation of the political information and media use equations. In particular, a dynamic relationship between levels of political information and media use is hypothesized, as follows: 1. Yl,t=f(Y2,t-l) 2. Y f(Y 2,t= 1,t-l) where Y1 and Y2 represent levels of political information and levels of media use, respectively. The first equation stipulates a positive function, and the second hypothesizes a negative function. The data used in this analysis, from the CPS 1972-1974-1976 panel surveys, 286 were collected at only one point in time. Thus, the following relationships were estimated in place of those above: 1*. Yl,t=f(Y2,t) * = 2 . Y ,t f(Y 2 1,t) where the relationships hypothesized by both equations are positive. Given that both Equations 1 and 1* predict positive relationships, it makes no difference whether the data used to test this association were collected at the same, or different, points in time, at least in terms of the expected relationship. However, the relationships represented by Equations 2 and 2* do differ in the direction of the relationship that is hypothesized. Thus, whether the data were collected at the same point in time or at different points in time is a matter of some concern. As a consequence of the data used in Chapters 7-9 being collected at one point in time, the relationships in Equation 2 were not estimated. Rather, the functional relation expressed by Equation 2* was examined. Further, this relationship was confirmed in almost every instance; the association between media use and information levels in the media use equations was found to be posi- tive, not negative. The fact that the data were collected at one point in time, rather than being dynamic, does not suggest that the relationship expressed by Equation 2 is inconsistent with the real world. Rather, the nature of the data prevented an examination of the relationship expressed by Equation 2, and by default, Equation 2* was tested and confirmed. 287 Therefore, the first requirement of a completely appropriate test of the theory of citizen information gathering is that the data used to estimate the statistical model be collected at more than one point in time. I will return to a fuller discussion of this require- ment in the next section of this chapter. The second problem encountered in the empirical evaluation of the statistical model was multicollinearity between the predictor variables of the political information equations. Specifically, the instrumental variables for the media use measures were highly corre- lated with one or more of the socio-economic-political variables in the information equations in the tests of the model for all three years. The major source of this problem is obvious. The media use instruments were estimated, in part, as functions of the measures of education, income, social class, strength of party identification, general interest in politics and interest in the campaign. The media use instruments were then included as explanatory variables in the political information equations along with these six socio-economic- political variables. The collinearity observed between the media use instrumental variables and the other predictor variables was virtually inevitable. The major effect of this collinearity was to render the coefficients of the media use instruments in the information equa- tions statistically insignificant. While in many instances the coefficients of the media use instruments were of substantial magni- tude, in almost all cases were the standard errors of these estimates of comparable size. Thus, the coefficients of these variables were statistically insignificant. 288 One possible solution to this problem would involve reestimating the media use instrumental variables. The instruments used in this analysis were estimated as functions of the six socio-economic- political variables described above; one of the two differential benefit measures (Utility and Diffeel); and, where appropriate and available, civic duty (DTerm). Thus, the estimation of these instru- ments was heavily dependent upon the same set of variables with which the instruments appear in the political information equations. If it were possible to estimate the media use instrumental variables as functions of a set of variables consisting of the aforementioned measures, as well as additional measures, the dependence of the instruments on the socio-economic-political variables would be lessened. Conceivably, this would produce media use instruments which are not as highly intercorrelated with the other predictor variables in the information equations. Such a procedure would require additional exogenous variables in the statistical model. However, the likelihood of specifying such measures is remote, for two reasons. On a theoretical level, the theory of citizen information gathering does not specify any additional exogenous variables, and would not without some major revisions of the theory. On a practical level, variables which are oftentimes included in a system of equations for purposes such as this, those variables being such measures as education, income and social class, are already included in the model as predictors in the political information equations. At present, therefore, a solu-- tion to the problem of multicollinearity does not seem imminent. 289 The third data problem to be discussed involves the validity of some of the measures used. In particular, the indicators used to measure the concepts of political information and citizen un- certainty are indirect and possibly invalid. The measure of political information used in this analysis is an index based on the number of "likes" and "dislikes" mentioned by a respondent about each candidate (or party). The major operating assumption is that the more knowledgeable people are about a particu- lar campaign, the more positive and negative characteristics they can ascribe to a candidate or party. It can be argued that this measure is an invalid indicator of levels of political information. This is not an unreasonable criticism to make. It can be argued that this index does not measure political information, but merely the number of positive and negative things respondents can say about particular political objects. Resorting to this indicator was necessitated by the absence of any measures on the 1972-1974-1976 surveys of political information. Whenever secondary data is employed, the problem of "making do with what you have" is always encountered. In this case, "making do" with what I had may have led to an invalid measure of political information. The same criticism can be made of the measure of uncertainty used in this analysis. This measure is based on the assumption that the average respondent perception of candidates' issue positions (on the seven-point issue scales) is the "true" candidate position. Further, deviations from the mean perception are taken to indicate uncertainty on the part of respondents. It can be argued that the average respondent perception has no meaning, much less that it represents 290 the "true" candidate positions on the issues. If this is the case, then deviations from the mean have no meaning either. In turn, one can accept that the mean respondent perception indicates the "true" position of a candidate on an issue, and still reject the argument that deviations from the mean measure citizen uncertainty. It can be reasonably asserted that deviations from the mean do not indicate uncertainty, but merely different perceptions of the issue stands of the candidates. The conclusion of both these lines of argument is that the measure of uncertainty used in this analysis is invalid. It can also be argued that there is serious slippage between the theoretical concept of uncertainty and the indicator used to measure uncertainty. Uncertainty was conceptually defined as the probability of making an incorrect voting choice, where "incorrect" means not maximizing expected utility. The variable drawn from this concept was "uncertainty regarding a candidate's issue positions;" this variable was operationalized as described in the preceding paragraph. The linkages between the concept and the variable, and the variable and the indicator, are weak. The links may be so weak as to render the indicator used not a measure of the concept of uncertainty, but of something else entirely. In sum, the possible lack of validity of the measures of political information and uncertainty poses a serious problem. It may be the case that using these measures of key concepts have made the model evaluation presented in Chapters 7-9 an unfair test of the theory of citizen information gathering. I do not wish to extend this criticism too far. I would argue that the measures used took 291 full advantage of the indicators available on the CPS surveys. However, developing better measures of the concepts of political information and uncertainty is most desirable and, in principle, possible. Some alternative measures of these concepts will be explored in the next section of this chapter. The final data problem to be discussed is that of unreliability. Of particular concern are the measures of media utilization employed in the empirical analyses of the theory of citizen information gathering. The most obvious problem involves over-reporting by respondents of their level of media use to follow a political cam- paign. It is generally recognized that self-reports of media use are a problem in survey research, in that respondents tend to give themselves credit for higher levels of media use than is warranted (Graber, 1980). The media use variables may themselves be unrealiable. In general, 3. X=X*+u where X is the "true“ value of a variable, X* is the observed value, and p is an error, or "noise" term. If it is assumed that p is random, with an expectation of zero, then it is not difficult to accept the conclusion that E(X*)=X. However, the media use measures are a fairly recent addition to the survey literature. As a conse- quence, neither the level of "noise" in these measures, nor the distribution of the error, are not known at this time. Thus, it is not possible to state with assurance that the media use measures used in this analysis are reliable indicators of media utilization. 292 Future Research The problems involved in the estimation of the statistical model of the theory of citizen information gathering are not trivial. The possibility that the measures of some of the key concepts of the theory are invalid or unreliable, coupled with the necessity of estimating a dynamic process with a set of coterminous equations and measures, suggests that the theory has not been properly operational- ized by either the data or the statistical model. In this section, I will discuss an alternative method of evaluating the theory of citizen information gathering. In order to accurately evaluate the statistical model, a multi- wave panel survey is essential. Such a survey should consist of at least two waves, although several waves would be more desirable. A survey schedule for a presidential election would ideally consist of seven waves. The first survey would be conducted well before the pre-nomination campaign; October or November would be an appropriate starting point. Another would be conducted immediately preceding the primary elections, at around the middle of January. The third and fourth waves would be conducted, respectively, in the middle of the pre-nomination campaign (around April) and to the end of the primary season (at the end of June). The fifth survey would be carried out after the major party nominating conventions, but before Labor Day, and the sixth survey would immediately precede the general election. The final wave would be conducted after the general election, in late November. 293 This constitutes an ideal, and most likely an overly ambitious, survey schedule. While the Center for Political Studies may have the resources to carry out such a research design,2 an individual may not have such resources. A more realistic design involves a three-wave panel survey. One survey would be conducted prior to the primary elections, one before the major party nominating conventions, and one prior to the general election. At the minimum, two waves are necessary in order to properly operationalize the dynamic aspects of the theory of citizen information gathering. More appropriate measures of the critical concepts are also needed. In particular, improved measures of the concepts of politi- cal information and uncertainty are necessary. Further, devising such improved measures should not prove to be an insurmountable task. The most obvious measure of political information would consist of a series of questions regarding the candidates and the campaign. These questions would be concerned with the qualifications, exper- iences and personal attributes of the candidates; the important issues of the campaign; the actual issue stands of the candidates, as opposed to the placement of the candidates on the seven-point issue scales; and the controversies of the campaign (e.g., Carter's "ethnic purity" remark during the 1976 campaign). With a substantial number of information questions, not only can indexes of political informa- tion levels be constructed, but measures of information gain and information loss can be built as well. The latter measures would involve calculating the proportion of questions answered correctly at different points of the campaign (e.g., at the beginning and end of the campaign), and then computing the differences between these 294 proportions. Alternative measures of the concept of uncertainty can also be constructed. One possible measure would involve relatively simple questions such as "How sure of (or comfortable with) your perceptions of the candidates' stands on the issues are you?" More complex Operationalizations might involve allowing respondents to indicate ranges of values of the seven-point issue scales, rather than forcing respondents to choose a specific point on the scale. Un- certainty could then be measured by the average (or total) size of these ranges on the issue scales. However, these indicators would still pose a problem in terms of the linkage between the concept of uncertainty and its measure. A more theoretically valid measure of uncertainty would consist of two questions. The first question would come toward the end of the interview schedule, and would read as follows: I would like for you to think for a moment about everything you feel is desirable about a President-issue positions, ex- perience, character, political party, or anything else you feel is important. With all of these things in mind, could you tell me which of the candidates, or , you would like to see elected as President? (Hopefully, the answer to this question will be the same as that for the candidate choice question.) The second question would involve an attempt to measure the respondent's subjective probability of the correctness of his or her choice. The interviewer would hand the respondent a card, on which would be drawn a horizontal line with one 295 hundred vertical hash marks. The left mark would be labeled “0%", and the right mark would be labeled "100%". The "25%", "50%" and "75%" marks would also be labeled. The following question would then be posed: You just indicated that would be your choice to serve as President. Now, on the card that I have just handed you, 100% means that you are absolutely sure of your desire for to be President, while 0% indicates that you have absolutely no idea who you want to be elected President. By placing a mark somewhere on this line, could you indicate for me how sure you are of your desire for to be elected to the Presidency? The percentage indicated would then be subtracted from unity to indicate the probability of error, or P(E). This measure has the advantage of confining the response to a number between one and zero, inclusive. Further, this measure yields an indicator of un- certainty that is theoretically more valid than the measures prev- iously discussed. Conclusion This dissertation began with the desire to explore how and why citizens go about gathering information during electoral campaigns. The major contribution of this research has been the development of a theory of citizen information gathering. Hypotheses were derived from the theory, and tested. A statistical model of the theory was developed, and empirically examined. 296 As with any first effort, this dissertation has not fully answered the questions of interest. Additional data need to be gathered, utilizing panel survey designs. Further, additional theoretical efforts must be carried out. The research reported here, especially the theoretical results pertaining to the im- portance of citizen uncertainty, indicates that further work is desirable. The extent to which citizen information gathering and processing in electoral contexts is not fully understood indicates that such work is necessary. 297 ENDNOTES 1By important explanatory variables, I mean explicitly those predictor variables whose coefficients were consistently signifi- cant at the .05 level of probability. 2In fact, the CPS is conducting a multi-wave panel survey for the 1980 American national election study. APPENDICES APPENDIX A PROOF OF THEOREM I THEOREM: Assume that 0: = 0%, and that the amounts of information regarding candidates A and B are equal. Then when conducting an hypothesis test, the probability of error of the test is minimized when the critical level of the test is preset at the point of intersection of the distributions of the estimators. Proof. Let c denote the critical level of the hypothesis test. Our goal is to find that value of c which minimizes the following expression: c 1 -(c-Ob)2/Zo§ c 1 -(c-Oé)2/Zo§ P(E) = 1-[00 2 6: dx + foo ——2- 5 dx /2nob /2noa . d P E l. F1nd _d c d P E 1 - '(C'fib)/Z°E 1 -2/20§ - = .. E + d C /2no§ /2noa Since it is assumed that a: = 0: . d(PEg) “(C'Ua)2 ”(C'Ub d C = E. - 8 Taking natural logs, rearranging terms, and setting the expres- sion equal to O, solve for c: 298 ‘ 2 “ 2 _ (c-Ub) - (c-Ua) — 0 2 2 “2 2 _ .' “.2 _ c - 2cUb + Ub - c - 2cU - Ua - O “ ‘. _ “.2 “2 -2cUb + 2cUa — Ua - Ub Al - A _ '2 A2 c(2Ua 2Ub) Ua - Ub _ (Us-Uuwawb) c - .rww O'+U _ a b C ’ 2 which is the intersection of the distributions of the two estimators. To determine if Oé+Ob/2 yields a minimum, it is necessary to substi- 2 d P E d tute c = O$+Dbl2 into However, this procedure is inde- 2 terminate, since gJ—Zéfll-win then equal 0. To determine if (C) d P E c = Ué+Ubl2 yields a minimum, we must examine d c when ct U$+Ob/2 (cf., Rodin, 1970:191-192). Specifically, when c < Oé+Ob/2, dchE should be less than 0, and when c > Oé+Ob/2 ddPCE should be greater than 0. It is assumed that a: 0:. Further to aid in this demonstra- tion, let us give some real values to O; and Ob, say 0; = 150 and A u = 140; thus c = 145, when c = 05+0b/2 . b 300 First, let c O$+Ub/2, say c = 146: d P E _ -(146-150)2 -(146-140)2 d c T E ' 5 (146-140)2 - (146-150)2 36-16 20>0 which satisfies our second condition. In general, of course, when c = O'+U /2 91%A§11-= 0 Thus c = O'+O /2 or the intersection of a b ’ d c ' ’ a b ’ the distributions of the two estimators, yields a minimum value of the total probability of error. Q.E.D. APPENDIX B A BRIEF DESCRIPTION OF BAYES' THEOREM In the model there are four possible states of the world (and four possible inferences about the state of the world) as defined in Figure 3. Call these states of the world 0. Based upon any infor- mation the citizen may have prior to informing himself about the cam- paign, the citizen can determine the probability of a state of the world occurring. Denote the probability distribution of o as P(o), the prior distribution of o. If the citizen has no information about the candidates, issues, etc., P(o) will be a uniform distribution. This special case of having no information prior to the first sample of information is called a state of prior ignorance. Next, assume that a sample of information has been drawn. The sample results will be an inference about the state of the world, and can be denoted, in general, as y. The likelihood of observing y, given a particular state of the world, 0, can be determined from the likelihood function, P(y/o). The likelihood function and the prior distribution are then combined to give the posterior distribution of o, P(o/y), via Bayes' Theorem: P(y/01)P(e,) P(e/Y) = ; P(Y/oj1P(oj1 J In words, the posterior probability of the particular state of the 301 302 world, given the sample information, equals the product of the prior probability of that state of the world, divided by the sum (over the states of the world) of these products. The posterior probability allows the statistical decision maker to determine the probability that his inference, or decision, about the state of the world, given the sample data, is correct. By multiplying the utility of any decision, or inference, by its prior and posterior probabilities, the prior expected utility and posterior expected utility of any decision, Di’ i=l,2,3,4 can be assessed. Denote these as EU(Di) and EU'(Di), respectively. An inference about the state of the world (based on the prior probability function) is optimal if and only if EU(D1.) 1 60(03), i f j Denote an optimal decision as 0*. Further, after a sample has been taken, the optimal act may differ from the optimal act prior to taking the sample. Denote the posterior optimal decision as D**. APPENDIX C TESTS OF ASSUMPTIONS FOR 1972 AND 1974 In this appendix two assumptions of the theory, the assumption of unbiased information sources, and the assumption that media utilization is a function of various socio-economic-political factors, will be examined for 1972 and 1974. 1972 Tests of Assumptions To examine the first assumption, that media sources do not systematically distort information, five media utilization variables will be used. The first variable is a composite index consisting of four questions which determine whether respondents monitored the campaign through the four major news media. Combining the yes and no responses yields a variable with a range of 0-8. The categories, and their meanings, are as follows: C I no media used in monitoring the campaign; 1 - radio or television the only medium used; - newspapers or magazines the only medium used; both radio and television used; .5 w'm 1 - either radio or television and either newspapers or magazines used; 5 - both radio and television and either newspapers or magazines used; 303 304 6 — newspapers and magazines only used; 7 - both newspapers and magazines and either radio or television used; 8 - all four media used. I will call this variable MEDIAUSE. The other four media variables come directly from the survey. They are variables 457, 460, 462, and 464; they measure the degree to which citizens report using news- papers, radio, magazines, and television, respectively, to follow the 1972 campaign. The test of the first assumption proceeds in three stages. First, all five media utilization variables were correlated with the differential benefit measure based on the seven-point issue scales. The coefficients for all the media utilization variables were below .05, with the exception of MEDIAUSE, and were not significant at the .05 level of probability. The MEDIAUSE/utility coefficient was sig- nificant at the .05 level of probability. Next, all five media utilization variables were correlated with the differential benefit measure based on the 1972 feeling thermometers. None of these coefficients were significant at the .05 level of probability. Finally, all five media utilization variables were correlated with the vote choice variable - for whom the respondent reported he/she voted (variable 478). Only responses 1 (Nixon) and 2 (McGovern) were used. None of the relationships were significant at the .05 level of probability. The results of these three tests indicate that, for 1972, the assumption of unbiased information sources is plausible. 305 The second assumption to be tested with the 1972 data is the assumption that media utilization is a function of several socio- economic-political variables. The same media utilization measures which were used above are used to test this assumption. The independent variables used are education (variable 300), social class (400), income (420), strength of party identification (a recode of variable 140), and general political activity. The last variable is an index with a range of 0-4, based on variables 469- 472 from the 1972 survey. The five media utilization variables were correlated with these five respondent characteristics. The results are displayed in Table 50. All coefficients are Pearson's r's. Table 50. Relationships between Media Utilization Variables and Socio-Economic-Political Variables for 1972. MEDIAUSE V457 ‘ V460 V462 V464 V300 .227* -.134* -.018 -.114* -.011 V400 .147* -.112* .024 -.10* .041 V420 .202* -.143* -.075 -.141* -.037 Strength of Party 1.0. .107* -.O63 -.095* -.10* -.081* General Political Interest .228* -.149* -.139* -.144* -.174* *Significant at .05 level of probability Of the 25 coefficients, 18 (72%) are significant at the .05 level of probability. Further, the coefficients are, in a number of cases, substantial. The data lend a good deal of support to the 306 assumption under question. 1974 Tests of Assumptions We now turn to the 1974 data. Again, we start with the assumption that the mass media transmit unbiased messages to citizens. To test this assumption, the following media utilization variables are used: TVUSE - Measures general television utilization habits. Based on Variables 2028-2033; has a range of 0-6. TVNEWS - Measures television use for gathering general political information. Based on Variables 2028, 2030, 2031, and 2033; has a range of 0-4. PAPERUSE - Measures general newspaper utilization habits. Based on Variables 2049-2056; has a range of 0-8. PAPERNEWS - Measures newspaper utilization for gathering general political information. Based on Variables 2050- 2053, 2056; has a range of 0-5. V2037 - Number of television programs regarding the campaign that the respondent watched. Values range from 1 (a good many programs) to 3 (one or two programs). V2064 - Measures the extent to which respondents used newspapers to gather information about the senatorial campaign, if the state had a senatorial election. Values range from 1 (regularly) to 4 (once in a while). V2066 - Measures whether respondents utilized television or newspapers as the primary source of information about the 307 campaign. Values are l (newspapers), 3 (both newspapers and television equally used), and 5 (television). These media utilization variables were correlated with a senatorial candidate differential benefit measure and a party dif- ferential benefit measure, both of which are based on the 1974 issue scales. Call these variables CANBEN and PARTYBEN, respectively. These media utilization variables were also correlated with respond- ent's vote choice for the House of Representatives (Variable 2322), for the Senate (Variable 2325), and for governor (Variable 2328); the latter two vote choice measures were asked only in those states that held the appropriate elections in 1974. The results are presented in Table 51. The coefficients reported are Pearson's r's. Table 51. Tests of the Unbiasedness Assumption, 1974. CANBEN PARTYBEN V2322 V2325 V2328 TVUSE -.O48* -.055* .057* .036 .026 TVNEWS -.062* -.O71* .068* .043 .023 PAPERUSE .048* .083* -.034 —.007 -.057* PAPERNEWS ~ .051* .082* -.028 -.002 -.O41 V2037 .016 .005 -.O4 —.002 -.054 V2064 .022 .098* -.01 .083* .013 V2066 .041* .045* .023 .104* .03 *p_<_. 05 308 Of the 35 coefficients, 16 (46%) are significant at the .05 level of probability. Five of the seven coefficients for candidate differential benefit, and six of the seven party differential benefit coefficients are significant at the .05 level of probability. The vote choice variables (Variables 2322, 2325, and 2328) fare somewhat better, with only 5 of the 21 coefficients (24%)significant at the .05 level of probability. While a sizable proportion of the rela- tionships are statistically significant, only one, that for V2066/ V2325, is equal to or greater than .10. The results are mixed; the data do not lend support to the assumption, although the general absence of a relationship between vote choice and media utilization is encouraging. We turn to the second assumption, that media utilization is a function of various socio-economic-political factors. The media use variables utilized here are the same as those used to test the first assumption. The independent variables are as follows: Education: Variable 2418 Social Class: Variable 2525 General Political Interest: An index based on Variables 2196- 2199, with a range of 0-4 Strength of Partisan Identification: A recode of Variable 2204 Income: Variable 2549 All media utilization variables were correlated with these five variables. The results are given in Table 52. All coefficients are Pearson's R's. 309 Table 52. Relationships between Media Utilization Variables and Socio-Economic-Political Factors, 1974. Interest PartyID V2418 V2525 V2549 TVUSE -.025 .O74* -.O43* -.024 -.ll3* TVNEWS .036* .064* .007 .008 -.042* PAPERUSE .134* .068* .32* .222* .31* PAPERNEWS .151* .076* .34* .24* .321* V2037 -.l48* -.ll4* .116* .031 .099* V2064 -.229* -.088* -.l35* -.l4l* -.12* V2066 -.155* -.016 -.l4* -.168* -.l32* *p5305 Twenty-nine of the 35 coefficients (83%) are significant at the .05 level of probability. Further, 20 coefficients (57%) are greater than .10, and 7 (20%) are greater than .20. The data lend substan- tial support to the assumption. Conclusion In this Appendix two major assumptions of the theory of citizen information gathering were empirically examined for the 1972 and 1974 surveys. The first assumption, that the message transmitted via the mass media are unbiased, received considerable support from the 1972 data, but in general were not well supported by the 1974 data. The second assumption, that media utilization is a function of education, income, social class, strength of party identification, and general political interest, received substantial support from the 1972 and 1974 data. APPENDIX 0 CONSTRUCTION OF INSTRUMENTAL VARIABLES Instrumental Variables for 1972 Instrumental variables were estimated in the first stage of the two-stage procedure described in Chapter 6 for the candidate- and party-referenced information variables; for the media utilization variables Media and V464; and for the total uncertainty measure, Uncertainty. All instruments were estimated, in part, as functions of strength of party identification (PartyID); general political interest (Interest); interest in the campaign (V29); education (V300); social class (V400); and income (V420). The instruments used in the analysis in Chapter 7 differ in terms of the differential benefit measures used (Utility or Diffeel), and whether the civic duty term (DTerm) is included. Thus, four instrumental variables were constructed for each of the endogenous variables. The following are the instrumental variables constructed, along with a listing of the differential benefit and civic duty measures used, where applicable. For the sake of brevity, the six socio- economic-political variables described above are not repeated in the following listing. 310 311 Candidate-Referenced Information Caninfl: Utility Caninf2: Utility, DTerm Caninf3: Diffeel Caninf4: Diffeel, DTerm Party-Referenced Information Parinfl: Utility Parinf2: Utility, DTerm Parinf3: Diffeel Parinf4: Diffeel, DTerm Media Utilization Variables Medial, V464a: Utility Media2, V464b: Utility, DTerm Media3, V464c: Diffeel Media4, V464d: Diffeel, DTerm Uncertainty Unsurl: Utility Unsur2: Utility, DTerm Unsur3: Diffeel Unsur4: Diffeel, DTerm Instrumental Variables for 1974 Only one set of instrumental variables were constructed for the 1974 analysis, given that only one differential benefit measure (Utility) and no civic duty term is used in this analysis. Utility 312 is based on respondents' placements of the two major parties on the seven-point issue scales. All instrumental variables were estimated as being functions of strength of party identification (PartyID); general political interest (Interest); interest in the 1974 campaign (V2026); education (V2418); social class (V2525); and differential benefit (Utility). The instrumental variables are Infol for Info; V2066a for V2066; TVNewsl for TVNews; Paperl for Paper, and Unsurl for Unsur. Instrumental Variables for 1976 The procedures used to estimate the instrumental variables for the 1972 data set were also used to construct the instruments for 1976. All instrumental variables are estimated, in part, as being functions of strength of party identification (PartyID); general political interest (Interest); interest in the campaign (V3031); education (V3384); social class (V3486); and income (V3507). The instruments differ in terms of whether the civic duty measure (DTerm) is used, and in the measure of differential benefit used. The distinction between the two measures of differential benefit are the same for these data as for the 1972 data. The following is a listing of the instrumental variables constructed, along with the differential benefit and civic duty measures used, where appropriate. For the sake of brevity, the six socio-economic-political variables described above are not repeated in the following list. 313 Candidate-Referenced Information Caninfl: Caninf2: Caninf3: Caninf4: Utility Utility, DTerm Diffeel Diffeel, DTerm Party-Referenced Information Parinfl: Parinf2: Parinf3: Parinf4: Utility Utility, DTerm Diffeel Diffeel, DTerm Media Utilization Variables V3654a, TVNewsl, Paperl: V3654b, TVNews2, Paper2: V3654c, TVNews3, Paper3: V3654d, TVNews4, Paper4: Uncertainty Unsurl: Utility Unsur2: Utility, DTerm Unsur3: Diffeel Unsur4: Diffeel, DTerm Utility ‘ Utility, DTerm Diffeel ‘ Diffeel, DTerm Correlations between Instruments and Endogenous Variables The following tables present the zero-order correlations between the instruments above and the endogenous variables for which they serve as instruments. All coefficients are Pearson's r's, and all are significant at the .05 level of probability. 314 Table 53. Correlations between Instruments and Endogenous Variables, 1972. CANINF PARINF V464 MEDIA UNCERTAINTY Caninfl .18 Caninf2 .18 Caninf3 .18 Caninf4 .18 Parinfl .20 Parinf2 .20 Parinf3 .20 Parinf4 .13 V464a .23' V464b .22 V464c .22 V464d ' .21 Medial .34 Media2 .35 Media3 .25 Media4 .34 Unsurl .26 Unsur2 .26 Unsur3 .26 Unsur4 .26 Table 54. Correlations between Instruments and Endogenous Variables, 1974. INFO V2066 TVNEWS PAPER UNCERTAINTY Infol .37 V2066a .51 TVNewsl .21 Paperl .36 Unsurl .23 315 Table 55. Correlations between Instruments and Endogenous Variables, l976. CANINF PARINF V3654 TVNEWS PAPER UNCERTAINTY Caninfl .45 Caninf2 .45 Caninf3 .45 Caninf4 .45 Parinfl .46 Parinf2 .46 Parinf3 .39 Parinf4 .46 V3654a .24 V3654b .24 V3654c .24 V3654d .34 TVNewsl .26 TVNewsZ .26 TVNews3 .26 TVNews4 .27 Paperl .48 Paper2 .50 Paper3 .48 Paper4 .50 Unsurl .27 Unsur2 .27 Unsur3 .26 Unsur4 .27 APPENDIX E CORRELATIONS BETWEEN MEDIA USE INSTRUMENTS AND SIX SOCIO-ECONOMIC-POLITICAL VARIABLES, 1972, 1974 AND 1976 There is substantial potential for collinearity between the media use instrumental variables and the six socio-economic-political variables in the information equations reported in Chapters 7-9. The following three tables present the zero-order correlations between these variables for the 1972, 1974 and 1976 data sets. All coefficients are Pearson's r's. Table 56. Correlations between Media Use Instruments and Six Socio- Economic-Political Variables, 1972. V300 V400 V420 INTEREST V29 PartyID Medial .655 .535 .343 .611 -.630 .235 Media2 .640 .521 .338 .607 -.624 .231 Media3 .989 .464 .412 .328 -.160 -.O41 Media4 .658 .530 .354 .602 -.626 .228 V464a .147 .051 .277 -.658 .620 -.159 V464b .142 .047 .271 -.660 .621 -.157 V464c .138 .030 .242 -.567 .566 -.133 V464d .134 .031 .236 -.571 .568 -.134 316 317 Table 57. Correlations between Media Use Instruments and Six Socio- Economic-Political Variables, 1974. V2066 INTEREST V2418 V2525 V2549 PartyID V2066a .183 -.558 -.668 -.728 -.595 .047 Paperl -.838 .527 .492 .401 .358 .202 TVNewsl -.785 .480 .277 .379 -.091 .379 Table 58. Correlations between Media Use Instruments and Six Socio- Economic-Political Variables, l976. V3031 V3384 V3486 V3507 INTEREST PartyID V3654a .239 -.873 -.628 -.576 -.562 .158 V3654b .237 -.872 -.624 ~.572 -.564 .161 V3654c .287 -.495 -.873 -.489 -.610 .157 TVNewsl -.899 -.O73 .027 .095 .262 .542 TVNewsZ -.867 -.151 -.O36 -.O72 .214 .543 TVNew53 -.884 -.150 -.O32 -.O65 .223 .554 TVNews4 -.862 -.147 -.O32 -.O67 .216 .539 Paperl -.635 .778 .448 .525 .543 .102 Paper2 -.624 .762 .432 .511 .525 .101 Paper3 -.631 .779 .440 .528 .543 .098 Paper4 -.629 .761 .427 .506 .526 .104 BIBLIOGRAPHY BIBLIOGRAPHY Achen, Christopher (1975). Mass political attitudes and the survey response. American Political Science Review, 69, 1218-1231. Amemiya, Takeshi (1977). The estimation of a simultaneous equation probit model. Technical Report 237, The Economic Series, Insti- tute for Mathematical Studies in the Social Sciences, Stanford University. Aldrich, John H. (1975). Candidate support functions in the 1968 election: An empirical application of the spatial model. Public .Choice,,22, 1-22. (1976). Some problems in testing two rational models of participation. American Journal of Political Science, 20, 713- 733. Aldrich, John and Charles F. Cnudde (1974). Probing the bounds of con- ventional wisdom: A comparison of regression, probit, and dis- criminant analysis. American Journal of Political Science, 19, 571-608. Atkin, Charles (1973). Instrumental utilities and information seek- ing. In Peter Clarke, ed., New models for mass communication re- search. Beverly Hills: Sage Publications. Atkin, Charles and Gary Heald (1976). Effects of political advertis- ing. Public Opinion Quarterly, 40, 216-228. Atwood, L. Erwin and Keith R. Sanders (1975). Perception of informa- tion sources and likelihood of split ticket voting. Journalism Quarterly, 52, 421-428. Beck, Nathanial (1975). The paradox of minimax regret. American Po- litical Science Review, 69, 918. Becker, Jerome D. and Ivan L. Preston (1969). Media usage and politi- cal activity. Journalism Quarterly, 46, 129-134. Bennet, Stephen Earl and Robert Oldendick (1978). The effect of edu- cation on mass belief systems: The case of issue constraint among domestic policy opinions, 1956-1976. Presented at the annual meeting of the Midwest Political Science Association, April, 1978. 318 319 Berelson, Bernard (1966). Democratic theory and public opinion. In Bernard Berelson and Morris Janowitz, eds., Reader in public qunion and communications, 2nd ed. New York: The Free Press. Berelson, Bernard R., Paul F. Lazersfeld, and William N. McPhee (1954). Voting. Chicago: The University of Chicago Press. Bishop, George F., Robert G. Meadows, and Marilyn Jackson-Beech (1978). The presidential debates: Media,,electoral,,andypolicy_perspec- tives. New York, Praeger. Bishop, George F., Alfred J. Tuchfarber, and Robert W. Oldendick (1978). Change in the structure of American political attitudes: The nagging question of question wording. American Journal of Political Science, 22, 250-269. Blumler, Jay G. (1978). The role of theory in uses and gratifications study. Political Communications Review,43, l-10. Campbell, Angus, Philip E. Converse, Warren E. Miller, and Donald E. Stokes (1960). The American voter. New York: John Wiley and Sons, Inc. Cangelosi, Vincent E., David M. Robinson, and Lawrence L. Schkade (1969). The utilization of information in rational choice: A cross-national experiment. Social Science Quarterly, 50, 78-91. Carmines, Edward G. and James A. Stimson-(1980). The two faces of issue voting. American Political Science Review, 74, 78-91. Carter, Richard F., Ronald H. Pyzska, and Jose L. Guerrero (1972). Dissonance and exposure to aversive information. In Alan Wells, ed., Mass media and society. Palo Alto, California: National Press Books. Chaffee, Steven H., Keith R. Stamm, Jose L. Guerrero, and Leonard P. Tipton (1969). Experiments on cognitive discrepancies and com- munication. Journalism Monographs, 14, 1-85. Clarke, Peter and Lee Ruggell (1970). Preferences among news media for coverage of public affairs. Journalism Quarterly, 47, 464- 471. Converse, Philip E. (1964). The nature of mass belief systems in mass publics. In David Apter, ed., Ideoquy and Discontent. New York: The Free Press of Glencoe. (1966). Information flow and the stability of partisan attitudes. In Angus Campbell, Philip E. Converse, Warren E. Miller, and Donald E. Stokes, eds., Elections and the political order. New York: John Wiley and Sons, Inc. 320 Converse, Philip E. (1970). Attitudes and non-attitudes: Continua- tion of a dialogue. In Edward R. Tufte, ed., Quantitative anal- ysis of social problems. Reading, Mass.: Addison-Wesley. Conway, M. Margaret (1968). Voter information sources in a non- partisan election. The Western Political Science Quarterly,_21, 69-77. Department of Marketing (1954). The influence of television on the 1952 election. Oxford, Ohio: Miami University. Deutsch, Morton and Robert M. Krauss (1965). Theories in social psy: chology. New York: Basic Books, Inc. Donohew, Lewis and Leonard Tipton (1973). A conceptual model of in- formation seeking, avoiding, and processing. In Peter Clarke, ed., New models for mass communication research. Beverly Hills: Sage Publications. Downs, Anthony (1957). An economic theory of democracy. New York: Harper and Row. Dreyer, Edward C. (1974). Media use and electoral choice: Some po- litical consequences of information exposure. In Norman R. Luttbeg, ed., Public opinion and public polity (rev. ed.). Homewood, Illinois: The Dorsey Press. Einsiedel, Edna F. (1975). Television network news coverage of the Eagleton affair: A case study. Journalism Quarterly, 52, 56-60. Elster, Jon (1979). Ulysses and the Sirens: Studies in rationality and irrationality. Cambridge, England: Cambridge University Press. Ferejohn, John A. and Morris P. Fiorina (1974). The paradox of not voting: A decision theoretic analysis. American Political Science Review, 68, 525-546. (1975). Closeness counts only in horseshoes and dancing. American Political Science Review, 69, 920-928. Festinger, Leon (1972). Behavioral support for opinion change. In Alan Wells, ed., Mass media and society. Palo Alto, California: National Press Books. Finney, D. J. (1971). Probit analysis (3rd ed.). Cambridge, England: Cambridge University Press. Flanigan, William H. and Nancy H. Zingale (1979). Political behavior of the American Electorate (4th ed.). Boston: Allyn and Bacon. 321 Freedman, Jonathan L., J. Merrill Carlsmith, and David 0. Sears (1970). Social psychology. Englewood Cliffs, New Jersey: Prentice-Hall, Inc. Frolich, Norman, Joe A. Oppenheimer, Jeffrey Smith, and Oran R. Young (1978). A test of Downsian voter rationality: 1964 presidential voting. American Political Science Review, 72, 178-197. Glaser, William A. (1965). Television and voting turnout. Public Opinion Quarterly, 29, 71-86. Graber, Doris A. (1976). Press and TV as opinion resources in presi- dential campaigns. Public Opinion Quarterly, 40, 285-303. (1980). Mass media and American politics. Washington: Congressional Quarterly Press. Greenberg, Bradley and Brenda Dervin (1970). Mass communication among the urban poor. Public Opinion Quarterly, 34, 224-235. Greenberg, Bradley S. and Hideya Kumanta (1968). National sample pre- dictors of mass media use. Journalism Quarterly, 45, 641-646, 705. Harney, Russel F. and Vernon A. Stone (1968-69). Television and news- paper front page coverage of a major news story. Journal of Broadcasting, 13, 181-188. Hinckley, Barbara, C. Richard Hofstetter, and John H. Kessel (1976). Information and the vote: A comparative election study. In Richard G. Neimi and Herbert F. Weisberg, eds., Controversies in American voting behavior. San Francisco: W. H. Freeman and Company. Hofstetter, C. Richard and Robert H. Trice (1977). Television news, foreign policy, and presidential campaigns: The case of 1972. Presented at the annual meeting of the Midwest Political Science Association, April, 1977. Janis, Irving L. and Leon Mann (1977). Decision making. New York: The Free Press. Joslyn, Richard A. (1977). The impact of political television on non- partisan voting behavior. Presented at the annual meeting of the American Political Science Association, September, 1977. Kaid, Lynda Lee and Keith R. Sanders, eds. (1978). Political Commu- nications Review, 3. Katz, Elihu (1971). Platforms and windows: Broadcasting's role in election campaigns. Journalism Quarterly, 48, 304-314. Kelley, Stanley, Jr., and Thad W. Mirer (1974). The simple act of voting. American Political Science Review, 68, 572-591. 322 Kessel, John H. (1972). Comment: The issues in issue voting. Ameri- can Political Science Review, 66, 459-465. Klapper, Joseph T. (1960). The effects of mass communication. New York: The Free Press. Kmenta, Jan (1971). Elements of econometrics. New York: The Mac- millan Company. Kraus, Sidney, ed. (1979). The great debates 1976: Ford vs. Carter. Bloomington, Indiana: Indiana University Press. Lang, Kurt and Gladys Engel Lang (1959). The mass media and voting. In Eugene Burdick and Arthur J. Brodbeck, eds., American voting behavior. Glencoe, Illinois: The Free Press. Lazersfeld, Paul, Bernard Berelson, and Hazel Gaudet (1968). Ip§_ people's choice (3rd ed.). New York: Columbia University Press. Luttbeg, Norman R. (1974). The structure of beliefs among leaders and the public. In Norman R. Luttbeg, ed., Public opinion and public policy (rev. ed.). Homewood, Illinois: The Dorsey Press. Macaluso, Theodore F. (1977). Political information, party identifi- cation, and voter defection. Public Opinion Quarterly, 41, 255- 260. . MacNeil, Robert (1968). The people machine. New York: Harper and Row. Mallar, Charles D. (1977). The estimation of simultaneous probability models. Econometrica, 45, 1717-1722. Markus, Gregory B. and Philip E. Converse (1979). A dynamic simulta- neous equation model of electoral choice. American Political Science Review, 73, 1055-1070. Mayer, Lawrence S. and I. J. Good (1975). Is minimax regret applica- ble to voting decisions? American Political Science Review, 69, 916-917. McCroskey, James C. and Samuel V. O. Prichard (1966-67). Selective exposure and Lyndon B. Johnson's "State of the Union" address. Journal of Broadcasting, 11, 331-337. McKelvey, Richard D. and William Zavoina (1975). A statistical model for the analysis of ordinal level dependent variables. Journal of Mathematical Sociology, 4, 103-120. McLeod, Jack, Scott Ward, and Karen Tancill (1965). Alienation and uses of mass media. Public Opinion Quarterly, 29, 583-594. 323 Merrill, I. R. and C. H. Proctor (1959). Politicalqpersuasion by television: Partisan and public affairs broadcasts in the 1956 general election. WKAR-TV Research Report 565M. East Lansing, Michigan: Michigan State University. Mickelson, Sig (1972). The electronic mirror. New York: Dodd, Mead and Company. Miller, Arthur H., Warren E. Miller, Alden S. Raine, and Thad A. Brown (1976). A majority party in disarray: Policy polarization in the 1972 election. American Political Science Review, 70, 753- 778. Miller, Arthur H. and Warren T. Miller (1977). Partisanship and per- formance: "Rational" choice in the 1976 presidential election. Presented at the 1977 annual meeting of the American Political Science Association, September, 1977. Miller, Arthur H., Edie N. Goldenberg, and Lutz Erbring (1979). Type- set politics: Impact of newspapers on public confidence. Ameri- can Political Science Review, 73, 67-84. Newcomb, Theodore M. (1971). Persistence and regression of changed attitudes: Long-range studies. In Jonathan L. Freedman, J. Merrill Carlsmith, and David 0. Sears, eds., Readings in social psychology. Englewood Cliffs, New Jersey: Prentice-Hall, Inc. Nie, Norman H. with Kristi Andersen (1974). Mass belief systems re- visited: Political change and attitude structure, Journal of Politics, 36, 540-587. Nimmo, Dan (1971). The political,persuaders. Englewood Cliffs, New Jersey: Prentice-Hall, Inc. O'Keefe, Garrett J., Harold Mendelsohn, and Jenny Liu (1976). Voter decision making in 1972 and 1974. Public Opinion Quarterly, 40, 320-330. O'Keefe, M. Timothy, and Kenneth G. Sheinkopf (1974). The voter de- cides: Candidate image or campaign issues? Journal of Broad- casting, 18, 403-412. Osgood, Charles E. (1966). The measurement of meaning. In Bernard Berelson and Morris Janowitz, eds., Reader in,public opinion and communication, 2nd ed. New York: The Free Press. Page. Benjamin 1. (1976). The theory of political ambiguity. Ameri- can Political Science Review, 70, 742-752. Page, Benjamin 1. and Calvin C. Jones (1979). Reciprocal effects of policy preferences, party loyalties, and the vote. American Po- litical Science Review, 73, 1071-1089. 324 Patterson, Thomas E. (1980). The mass media election. New York: Praeger. Patterson, Thomas E. and Robert D. McClure (1976). The unseeing eye: The myth of television power in national politics. New York: G. P. Putnam's Sons. Pierce, John C. and Douglas D. Rose (1974). Non-attitudes and Ameri- can public opinion: The examination of a thesis. American Poli- tical Science Review, 68, 626-649. Pomper, Gerald M. (1972). From confusion to clarity: Issues and American voters, 1956-1968. American Political Science Review, 66, 415-428. Pool, Ithiel de Sola (1959). TV: A new dimension in politics. In Eugene Burdick and Arthur J. Brodbeck, eds., American voting be- havior. Glencoe, Illinois: The Free Press. Popkin, Samuel, John M. Gorman, Charles Phillips, and Jeffrey A. Smith (1976). Comment: What have you done for me lately?: Toward an investment theory of voting. American Political Science Review, 29; 779-805. Riker, William H. and Peter C. Ordeshook (1968). A theory of the cal- culus of voting. American Political Science Review, 62, 25-42. Robinson, John P. (1972). Perceived media bias and the 1968 vote: Can the media affect behavior after all? Journalism Quarterly, 49, 239-245. (l976). Interpersonal influence in election campaigns: Two step-flow hypotheses. Public Opinion Quarterly,,40, 304-309. Robinson, Michael J. (1976). Public affairs television and the growth of political malaise: The case of "The Selling of the Pentagon." American Political Science Review, 70, 409-432. (1977). Television and politics: 1956-1976. The Public Interest, No. 48, 3-39. Rodin, Burton (1970). Calculus with analytic geometry. Englewood Cliffs, New Jersey: Prentice-Hall, Inc. Roper Organization, Inc., The (1977). Chaqgipg public attitudes to- ward television and other mass media: 1959-1976. New York: Television Information Office. Sargent, Leslie W. and Guido H. Stempel, III (1968). Poverty, alien- ation and mass media use. Journalism Quarterly, 45, 324-326. Schramm, Wilbur and Richard F. Carter (1959). Effectiveness of a poli- tical telethon. Public Opinion Quarterly, 23, 121-127. 325 Sears, David O. (1971). The paradox of de facto selective exposure without preference for supportive information. In Jonathan L. Freedman, J. Merrill Carlsmith, and David 0. Sears, eds., Read- ings in social psychology. Englewood Cliffs, New Jersey: Prentice-Hall, Inc. Sears, David O. and Jonathan L. Freedman (1967). Selective exposure to information: A critical review. Public Opinion Quarterly, 31, 194-213. Shepsle, Kenneth A. (1972). The strategy of ambiguity: Uncertainty and electoral competition. American Political Science Review, §§, 742-752. Sherrod, Drury R. (1971). Selective perception of political candi- dates. Public Opinion Quarterly, 35, 554-562. Shively, W. Phillips (1979). The development of party identification among adults: Exploration of a functional model. American Poli- tical Science Review, 73, 1039-1054. Sigel, Roberta S. (1972). Effects of partisanship on the perception of political candidates. In Samuel A. Kirkpatrick and Lawrence K. Pettit, eds., The social psychology of political life. Belmont, California: The Duxbury Press. Simon, Herbert A. (1957). Models of man. New York: John Wiley and Sons, Inc. (1959). Theories of decision-making in economic and be- havioral science. American Economic Review, 49, 253-283. (1961). Administrative behavior. New York: The Macmillan Company. Simon, Herbert A. and Frederick Stern (1955). The effect of televi- sion upon voting behavior in Iowa in the 1952 presidential elec- tion. American Political Science Review, 49, 470-477. Stephens, Stephen V. (1975). The paradox of not voting: Comment. American Political Science Review, 69, 914-915. Stone, Vernon A. (1969-70). Sources of most news: Evidence and infer- ence. Journal of Broadcasting,,l4, 1-4. Strom, Gerald S. (1975). On the apparent paradox of participation: A new proposal. American Political Science Review, 69, 908-913. Swanson, David L., ed. (1976). Communication Review, 6. (1978). Political communications research and the uses and gratifications model: A critique. Political Communications Re- view, 3, 11-17. 326 Sullivan, John L., James E. Pierson, and George E. Marcus (1978). Ideological constraint in the mass public: A methodological cri- tique and some new findings. American Journal of Political Science, 22, 233-249. Thompson, Charles A. (1956). Television and presidential,politics. Washington, D.C.: The Brookings Institute. Tipton, Leonard P. (1970). Effects of writing tasks on utility of information and order of seeking. Journalism Quarterly, 47, 309-317. Tollison, Robert, Mark Crain, and Paul Pautler (1975). Information and voting: An empirical note. Public Choice, 24, 43-49. Tollison, R. D. and T. D. Willet (1973). Some simple economics of voting and not voting. Public Choice, 16, 59-71. Tullock, Gordon (1967). Toward a mathematics of politics. Ann Arbor, Michigan: The University of Michigan Press. (1975). The paradox of not voting for oneself. American Political Science Review, 69, 919. Underwood, Geoffrey, ed. (1978). Strategies of information proces- 51 9. London: Academic Press. Wilson, C. Edward (1974). The effect of medium on loss of information. Journalism Quarterly,_51, 111-115. Winkler, Robert L. (1972). Introduction to Bayesian inference and decision. New York: Holt, Rinehart and Winston, Inc. Zajonc, Robert B. (1961). The concepts of balance, congruity and dis- sonance. Public Opinion Quarterly, 41, 280-296. Zukin, Cliff (1977). A reconsideration of the effects of information on partisan stability. Public Opinion Quarterly,,4l, 244-254.