Campaign+Timing+and+Vote+Determinants

Campaign Timing and Vote Determinants David A.M. Peterson Department of Political Science University of Minnesota 1414 Social Sciences Building 267-19th Avenue South Minneapolis, MN 55455-0410 dpeters@polisci.umn.edu Paper prepared for the 1999 annual meeting of the American Political Science Association, Atlanta, GA. This paper was supported by a National Science Foundation Graduate Research Fellowship. I would like to thank Phil Paulino, Kevin Quinn, Simon Jackman, John Freeman and Brad Carlin for their helpful comments and suggestions. Any errors that remain are solely my responsibility. Abstract Questions about the role of campaigns in making different considerations more important for voters have been central to the study of political behavior for fifty years (Lazardsfeld et al 1948). The basic concern is does the information presented during the campaign alter how voters evaluate and choose between candidates. This paper develops a random coefficient or hierarchical logit model to analyze the 1984 NES Continuous Monitoring Survey. The specification treats the effect of partisanship, policy distance and candidate character traits as a function of the campaign timing. Of the theories tested in this paper, the attitude strength model best predicts the changes in vote determinants across the campaign. 1 1 Introduction Elections are the only formal link between the will of the people and the policy government enacts. Political campaigns educate and inform voters about what the candidates would do if elected. The effectiveness of campaigns is a prominent concern not only of political scientists, but also of journalists, politicians, and political reformers. Generally, the question is does the content and style of the campaign allows voters to make an informed choice? V.O. Key (1966) used the oft-repeated metaphor of the election as an echo chamber, where the voice of the people in November is merely a reverberation of the campaign. If the campaign focuses on policies, the voice returning will be based on issues. If the input is about the character or personality traits of the candidates, the sound will be too. Thus, the voters use the candidate’s policy positions when forming an evaluation of him or her, but only insofar as the presidential campaign highlights the policy information. The political science literature regularly examines this hypothesis, or another similar to it. Most of the tests of presidential campaigns, however, rely on cross sectional differences in information or uncertainty to test for campaign effects (e.g. Bartels 1996, Alvarez 1998). While these results illustrate the importance of information differences in the use of issues, they do not address campaigns’ influence because campaigns can alter the difference in three ways. First, campaigns could exacerbate preexisting differences in information, where the most informed at the start of the campaign learn the most from it. Second, the campaign environment could minimize differences. If the content of information is redundant and there is a limit to what can be learned, those with less information at the beginning catch up. Finally, the campaign can have parallel effects on all voters, where the campaign maintains but does not exacerbates the differences in information. An analysis based on cross sectional data cannot separate these effects. What is needed is to study voters at different points during the campaign. In this paper, I consider how political campaigns alter what information voters use to evaluate candidates. To examine this question, I rely on the 1984 National Election Study (NES) Continuous Monitoring Survey (CMS). The CMS sampled voters by weeks, and this paper exploits this clustering to test for changes in the weights applied to the determinants of vote choice. To examine the effects of the campaign on the relationships I use a hierarchical model, treating the coefficients between the predictors partisanship, policy proximity, and candidate trait assessments, and the dependent variable vote choice as a function of time and the aggregate levels of learning in the campaign. This approach allows me to test several competing theories of campaign effects. I find that the time of the campaign does alter how weight the determinants of vote choice. The flow of campaign information affects the use of partisanship, and personality traits. Policy distance, however, is unaffected by the campaign, either through a general time trend or an aggregate measure of campaign learning. These results support some theories of campaign learning, while challenging other prevailing theories. 2 2 Current Theories The effectiveness of campaigns is an important component in determining whether or not our representative system work. Elections serve as the sole direct connection between the preferences of the electorate and the preferences of the people who enact legislation. Our micro level theories hold that voters do not perform well. Media campaigns produce muddled images of candidate policy stances. News coverage emphasizes candidate strategy and focuses on personal characteristics more than the policy debate. Voters, it is commonly believed, are ill informed and inattentive, unable employ their vote for a reasoned response to the campaign debate. Nationally, the electorate listens to what the parties offer, and consistently selects the platform that is closest to the collective preferences (Erikson, MacKuen and Stimson forthcoming). The choice of the voters at the start of the campaign is in line with the micro-level results; it seems to have little to do with the policy positions of the candidates. The tracking polls vary widely during the campaign, which seems to be in line with the pessimistic view outlined above. Nevertheless, these data show that the voters learn. By the end of the campaign, the tracking poll numbers converge to what it should, based on the policy preference of the voters and the positions of the candidates (Gelman and King 1994). The resolution of the micro and macro level performance of voters, suggested by Gelman and King, lies in the effectiveness of the campaign. Voters are able to learn enough about the candidates through the campaign that by November, they cast a reasoned ballot. However, Gelman and King do not test a micro level theory of learning. Without such a model, the disjuncture between individual and aggregate voting patterns remains. 2.1 Minimal Effects The empirical study of the effects of presidential campaigns on the attitudes and behaviors of voters began with the work of the Columbia scholars (Berelson, Lazardsfeld, and McPhee 1954; Lazardsfeld, Berelson, and Gaudet 1944). This pair of works, The People’s Choice and Voting, establishes the questions asked and methods used by students of political behavior to this day. Because they are interested in learning and persuasion over the course of the campaign, Berelson et al. develop a four-wave panel survey of respondents from Erie County Ohio to track the attitudes and opinions during the election. These works contain two conclusions that address the question of this paper. They find no changes in the preferences of the voters from the campaign. In November, voters base their decisions on predispositions expressed six months earlier. Less than 15 percent of their respondents change their minds about which candidate to support during the campaign. That does not imply that the campaign is meaningless or worthless. Instead, they argue that the role of the campaign in the political process was to inform voters and “bring voter’s predispositions to the level of visibility and expression” (1944 3 p. 75). This conclusion resulted in the notion that there are “minimal effects” of campaigns on attitudes and behaviors of voters. If the campaign does not alter the content of the attitudes or their weights, the coefficients should be constant across campaigns. The only change that should occur is the precision of the estimates. The model should simply be heteroskedastic. As the campaign progresses, the error variance in the model should decrease. 2.2 Enlightenment The micro level model of campaign effects in Gelman and King (1994) is a simple enlightenment model. They ask why do campaign polls vary so considerably when the election outcome is so predictable? Their answer is the campaign serves to teach voters about the appropriate weights to apply to ideology and the candidates' policy positions. The variability in tracking polls occurs because voters are not properly enlightened; they apply an incorrect weight to ideology. As the campaign progresses, voters alter the weight applied to the fundamental variables until they incorporate them accurately based on their enlightened preferences. The changes in the candidate’s vote share stems from changes in the weights applied to different considerations, and not changes from the content of the considerations. They are more specific about the changes they expect. Based on this model they suggest that “if voters are becoming enlightened, then the fundamental variables should be increasing in importance over the campaign (p. 441).” They are most interested in growth in the use of ideology and issue positions. The implication for this paper is that the coefficients on the policy distances should increase over time. The signs of the campaign effects should be the same as the individual level variables. But what of the other components? They find that party exerts a relatively constant effect over time. They do not discuss the role of the personality characteristics of the candidates. Their model is about the regularity and importance of issue positions in elections. Presumably, the use of candidate character information is not what they mean by fundamental variables. Therefore, the campaign should not heighten its use. 2.3 Uncertainty Alvarez (1998) posits a similar model. Drawing on a simple spatial model of voter decision making under imperfection information, Alvarez (1998) proposes a model of candidate evaluation that explicitly incorporates the uncertainty voters have about candidates. He asks: how does the amount of information alter a voter’s decision making and views of candidates? By using a decision theoretic definition of uncertainty, he finds that uncertainty and evaluations are negatively associated. More important for this research, he finds that uncertainty moderates the relationship between issues and evaluations. The higher a voter’s certainty the more weight he or she applies to issues. Alvarez is explicit about what he means by uncertainty. Voters are unsure of exactly where the candidate stands on a particular issue. This uncertainty stems from “the voters’ own disincentives to gather and process costly information about candidates 4 as well as the candidates’ incentives to disseminate ambiguous information” (p. 30). This is distinct from other concepts used by political scientists such as cognitive complexity, political awareness, political knowledge, political sophistication, political interest, political involvement, attentiveness, and engagement (see Zaller 1992; Delli Carpini and Keeter 1996; Luskin 1987; Stimson 1975; and Sniderman, Tetlock, and Brody 1991). Uncertainty is not an inherent quality of the person, but is an interaction between the information about the candidate that is available and the person’s ability to use it. A voter’s level of uncertainty differs across candidates, and changes over the campaign. Based on this model, Alvarez generates three hypotheses that relate to this question. First, campaigns decrease voters’ uncertainty about candidate policy positions. Second the greater a voter’s uncertainty about the candidate, the less he or she uses policy information. Third, the greater the policy uncertainty, the more he or she relies on other, non-policy attitudes such as party or the candidate traits. 2.4 Attitude Strength Social psychological models of attitude strength posit that how an attitude is used is moderated by how strong (i.e. accessible, salient, or stable) it is (Petty and Krosnick 1995). Unlike, Alvarez’s uncertainty theory, attitude strength is often a relative concept. That is, when determining what attitudes to rely on when evaluating something, a person is more likely to use the ones that are strong than those that are weak. While the campaign is likely to increase the policy attitude’s strength, it may not do so to the same degree as other, competing considerations such as the personality traits. For instance, Johnston, Blais, Brady, and Crête’s (1992) examination of the 1988 Canadian election find that party, policy, and personality are altered by the content of the campaign. As one is primed, it increases in importance. The attitude strength model is very similar to the uncertainty model. In fact, certainty is often used as a measure of attitude strength in the social psychological literature. There is an essential difference between the theories tested in this paper. Alvarez’s argues that as policy uncertainty decreases, voters will rely on other information less. The attitude strength model, because it is not derived from a utility function, can be generalized to other, valance based attitudes. If attitude strength is the driving force, the uncertainty about the candidate’s character traits should moderate the use of trait perceptions. As the campaign progresses, the respondent should weigh the traits more heavily. 2.5 Campaign learning All of these theories are silent about changes in the content of the voter attitudes. None of these models imply any type of change in the actual perceptions of the candidates, only the certainty that voters have about these attitudes. It is plausible that the actual content of the attitudes changes. For instance, Mebane and Wand (1997) indicate that who identifies as a Democrat during the primary season changes as Mondale and Hart’s electoral fortunes change. If party identification means slightly different 5 things at different times in the campaign, this could manifest itself as different change in the weights applied to partisanship. The specific implication for this work is that the effect of partisanship should be strengthened as Mondale’s chances improve. Mebane and Wand find that when Hart is a viable threat for the Democratic nomination, independents are more likely to identify as Democrats are because the party standard bearer is more likely to be a moderate. Once Mondale won the nomination, these independents move back to the middle, or even slightly to the Republicans. This movement essentially watered down of Democratic identification during the primaries—especially of partisanship as a predictor of Mondale voting. This change in the content of party identification implies that the effect of partisanship will change as Mondale’s change. When he is likely to be the candidate, party should be a stronger predictor. When the race is up in the air, the relationship between party and the vote should be attenuated. Expected changes in the other coefficients depend on the change in the content of the attitudes. If the content of the information environment is one sided, or if it changes dramatically at some point in the campaign, we should see a change in the content of the attitudes. The direction of this change, however, is difficult to predict. 3 Modeling campaign effects. The hypothesis tested in this paper requires a specific type of data. As outlined earlier, the standard cross sectional design cannot separate effects of the campaign from preexisting differences in information. To test the effects of the campaign on individual voters, the data must have samples of voters taken at various times in the campaign. It is also necessary to identify when in the campaign the survey takes place, and collect enough different time points to have significant variation in the timing. While the standard NES datasets do identify the date of the interview, the preelection wave covers only the final few months of the campaign. If most of the learning about the candidates occurs early in the campaign as is suggested by both Bartels (1993) and Alvarez (1998), these data will understate the effects of learning. The only large survey of an American presidential election that has a large number of independent cross sections at numerous times during the campaign is the 1984 CMS. The CMS contains 46 differently independent cross sections of roughly 80 respondents. Each sample is an independent random sample of the American electorate. 3.1 Measurement While the structure of the CMS is unique, the questions asked are the standard ones developed by the NES. Therefore, deciding how to measure the variables is relatively easy. The dependent variable is the vote choice. Unfortunately, this measure is not included in the survey until the twenty-third week of the CMS—before it is certain that Mondale is the Democratic nominee. Rather than throw out the first half the sample, I construct a choice variable based on the relative feeling thermometers. A respondent is a "1" if he or she rates Mondale higher than Reagan, a zero if reversed. Respondent who 6 rate the candidates equally are treated as missing (as are those who cannot decide later in the sample). This is not perfect. Of those respondents who are asked to make a choice, only two of the over 700 respondents choose the candidate they rate lower on the thermometers. The micro-level independent variables of interest are partisanship, the amount of policy disagreement between the candidate and the respondent, and the respondent’s assessment of the candidate’s personality traits. Partisanship is the standard seven-point scale. The policy measures are the average squared differences between the respondent’s self placement and the candidate’s position estimated by the mean of all respondent’s placement of the candidate on the policy questions. The personality measures are an additive scale based on the battery of questions asking the respondent to rate the candidate on traits such as intelligent and knowledgeable. Finally I include the following demographic characteristics: race (white=1), gender (male=1), age in years, and education (a four point scale, the same as Alvarez 1998). The macro level predictors are measures of time and the certainty of the corresponding micro level predictor. The campaign time measure is the log number of weeks since the start of the survey.1 For this paper, I assume that as the campaign progresses, voters are exposed to more information. The simple time measure captures this increase. The macro level uncertainty measures for partisanship is the aggregate subjective likelihood of Mondale being the Democratic nominee. For the pre-convention period, this is calculated the same as Bartels (1988). The policy uncertainty measures are combination of the measures of Campbell (1983) and Alvarez (1998). Campbell’s’ measure for each particular policy is the sample variance. Alvarez’s measure for an individual is the sum of the squared differences between the respondent’s placement of the candidate and the true position, measured by the sample mean. This measure used here is the sample variance standardized and averaged cross all issue scales.2 The trait uncertainty measure follows the same logic. For each candidate, the aggregate amount of uncertainty about his character is the sample variance, averaged across the traits. Clearly, these macro level measures are not the uncertainty that the individual respondents have about their attitudes. These measures serve as proxies of the information flow in the campaign environment. Changes in the macro-level uncertainty about a candidate stem from the content of information that is available. If the information context of a week is particularly clear about where Reagan stands on the Time is included as a log measure because it is hypothesized that the impact of a week’s worth of learning will decrease as the campaign progresses. Other specifications, such as including an untransformed or squared measure of time, were also tested. The model presented here best fits the underlying theory, but a different specification does not change the substantive interpretation. The distance is standardized to a zero one scale because the scales are of differing lengths. Using the unstandardized distance would essentially weight the scales differently. 2 1 7 issues, it should translate to a change in the uncertainty. While a direct measure of the campaign environment could be developed, this measure better fits theoretically. It directly measures the amount of learning rather than observes the information and assumes that voters use it. 3.2 Statistical model The hypothesis developed above is not a standard “X causes Y” relationship that political scientists normally test. The usual approach is to hypothesize about a direct (usually linear) relationship between two variables and use a technique such as OLS, logit, or probit (depending on the dependent variable) and make inferences based on the coefficient(s) of interest. The hypotheses I develop, however, are not directly about the relationship between the micro level predictors and vote choice. They are about how the campaign alters this relationship. Using standard regression based techniques, this is impossible to test; the coefficients are constant across observations by assumption and necessity. What is necessary is an approach that treats the coefficient as “random” in that it is drawn from a random distribution whose mean is a function of the sample week. The micro level specification of vote choice is straightforward: Logit (Voteij ) = β 0 j + β 1 j * PIDij + β 2 j PolicyWM ij + β 3 j * PolicyRRij +, (1) β 4 j * TraitsWM ij + β 5 j * TraitsRRij + β 6 Raceij + β 7 Genderij + β 8 Ageij + ε ij where i = 1,…,nj denotes the individuals within sample weeks and j = 1,…,46 denotes weeks. Treating time and the aggregate uncertainty terms as determinants of β 0 j , β 1 j , β 2 j , β 3 j , β 4 j , and β 5 j results in the following macro-level equations: β oj = γ 00 + γ 01 * ln(Week j ) + υ 0 j β 1 j = γ 10 + γ 11 * ln(Week j ) + γ 12 * UCPID j + υ1 j β 2 j = γ 20 + γ 21 * ln(Week j ) + γ 22 * UCPolWM j + υ 2 j β 3 j = γ 30 + γ 31 * ln(Week j ) + γ 32 * UCPolRR j + υ 3 j β 4 j = γ 40 + γ 41 * ln(Week j ) + γ 42 * UCTraitWM j + υ 4 j β 5 j = γ 50 + γ 51 * ln(Week j ) + γ 52 * UCTraitRR j + υ 5 j where the υ ' s are the macro level error terms and the “UC…” terms represent the aggregate uncertainty about each of the variables. Equations (1) and (2) define a multilevel model that reformulates to a single equation by substitution: (2) 8 Logit (Voteij ) = γ 00 + γ 01 * ln(Week j ) + (γ 10 + γ 11 * ln(Week j ) + γ 12 * UCPID j ) * PIDij + (γ 20 + γ 21 * ln(Week j ) + γ 22UCPolWM j ) * PolicyWM ij + (γ 30 + γ 31 * ln(Week j ) + γ 32 * UCPolRR j ) * PolicyRRij + (γ 40 + γ 41 * ln(Week j ) + γ 42 * UCTraitWM j ) * TraitWM ij + (γ 50 + γ 51 * ln(Week j ) + γ 52 * UCTraitRR j ) * TraitRRij + β 6 Raceij + (3) β 7 Genderij + β 8 Ageij + ε ij + υ 0 j + υ1 j * PIDij + υ 2 j * PolicyWM ij + υ 3 j * PolicyRRij + υ 4 j * TraitWM ij + υ 5 j * TraitRRij The inclusion of error terms in the macro equation (2) makes (3) difficult to estimate. If the υ terms were excluded, this becomes a fixed effects model and specification through interaction terms is relatively easy. Unfortunately, that also implies a deterministic relationship between the campaign and the use of the determinants by voters. The statistical estimator must estimate a model with mixed-level errors, a random specification of the coefficients, and a binary dependent variable. To the best of my knowledge, this estimator is absent from the political science literature. Wong and Mason (1985) develop a two-stage, Empirical Bayes (EB) model to account for exactly this data structure and theoretical specification. The first stage a straightforward logit model is used for the micro-level observations (respondents) within each of the J macro units (weeks). For the ith respondent in the jth week we observe: Yij = 1 if i voted for Mondale 0 if i voted for Reagan and assume: Yij | pij ~ Bernoulli ( pij ) (4) independently for i = 1, , n j and j = 1,  , J . pij is the probability that respondent i in week j votes for Mondale. This probability, pij , is then model in the standard logit set up, with the regressors outlined above. The second stage uses some macro-level regressor to explain the variation in the J micro coefficients. Because the macro level variables in this model are continuous, this second stage is analogous to a linear regression. The variability in the kth macro-level equation is: β jk = Z ′jk  k + υ jk , υ jk ~ N (0,η kk ) (5) where Z jk represents the observations of the macro variables and  k is matirx of macro level coefficients. The covariance between the error terms is assumed, for k ≠ k ′ , 9 cov(υ jk ,υ j′k ′ ) = η kk ′ if j = j ′ 0 otherwise (6) The full macro model can be re-written as:  j = Z j + . j . j ~ N (0, +) The likelihood function for this model is: f (Y |  .) = ∏ exp(− X′ Z j − X′ . j )   ij ij   ∑ Y ′ X j Z j   + ∑ Y ′ X j . j  j j j 1 + exp(− X′ Z j − X′ . j )  j   ij ij (7) which has no closed form solution. As the model is current specified, there is no unbiased maximum likelihood estimation technique available. The best approach is to treat this as a classic discrete choice model where the  are fixed effects and the υ kj terms in (3) as random effects. The model can be estimated by a restricted maximum likelihood (REML) (Patterson and Thompson 1971; Harville 1974; Dempster and Monajemi 1976). While the REML results are less biased, than ordinary maximum likelihood, they are still not unbiased. A better solution is a Bayesian (either empirical or full) approach. In order to specifiy either an empirical or a fully Bayesian model, priors need to be placed on the macro parameters. Without strong intution about the macro parameters, the priors are assumed diffuse. The particular structure is:  j ~ MVN (0, ∑ γ j ), ∑ −1 γj 0.001 0.0 0.0 = 0.0 0.001 0.0 0.0 0.0 0.001 The inclusion of these priors leads to the full posterior of: (8) f ( Y |  ,. ) = ∏ (9) exp(− X′ Z j − X′ . j )  1  ij ij  −1  ∑ Y ′ X j Z j  + ∑ Y ′ X j. j − ∑. ′j′+ . j  j j j 1 + exp(− X′ Z j − X′ . j )  j 2 j   ij ij Wong and Mason (1985) use an EB approach to estimate this. The basic explanation of the EB approach is to obtain + , the maximum likelihood estimate of + , and then obtain posterior interval estimates of the coefficients with + fixed at + . The problem associated with this technique is that it does not account for the confidence in 10 ~ ~ the variance parameters. That is, + is treated as if it were a known quantity and not an estimate (Dempster 1987, Seltzer et al 1996). In a simple linear regression model, there are adjustments, but in a discrete choice set up, these corrections are prohibitively difficult. The effect of this omission is to have too great a confidence in our inferences about  j and  . An additional problem arises from irregularities in the distribution of the variance of  j . If J is small, the ML estimate of the + matrix can be highly asymmetric but the point estimate will not be able to reflect this. Thus, the values of + used to calculate the parameters of interest may be a poor summary of the information available about + , despite being the single most likely value. The solution to these problems lies in a fully Bayesian approach where the joint posterior of the full model includes the conditional posterior for + . This allows one to integrate over the variability in the + matrix and results in a more accurate estimate of the confidence in the macro-level parameters. The problem associated with this approach is that it is computationally intensive. The high order of integration involved is difficult in all but the simplest models. Fortunately, advances in Markov chain Monte Carlo (MCMC) techniques, most notably the Gibbs sampler, has made this approach tractable (see Gelman et al 1995; Western 1998; and Jackman, forthcoming for introductions to MCMC techniques). While this paper is not about MCMC techniques, their application in political science is still rare enough to warrant a brief discussion. Define θ as the vector of all of the unknowns (here the  j ,  , . j , and + ) where each element of θ is one of the elements of the unknown vectors. The joint posterior for the vector θ given the data and priors, is the product of the conditional density of each of the elements of θ , given the data, priors, and the true value of every other element of θ . Of course, the true values of the elements of θ are unknown. However, it is possible to construct a Markov chain whose stationary distribution is equivalent to the posterior distribution of θ . Markov chain theory holds that this chain will converge to the joint posterior distribution as long as the chain meets a few mild regularly conditions and iterates long enough. The iterations occur by positing starting values of each of the unknowns, then sampling from the conditional density of each element given the data, priors, and the current values of i every other element of θ . Let θ n represent the ith iteration of unknown n, where the vector θ 0 is the starting values of the chain. The chain then iterates in the following manner by drawing successively from the distributions: ~ θ1i | y, x, z ,θ 2i −1 ,θ 3i −1 ,,θ ni −1 i θ 2 | y, x, z ,θ1i ,θ 3i −1 ,,θ ni −1 i i θ 3i | y, x, z ,θ1i ,θ 2 ,,θ n−1  i θ n | y, x, z ,θ1i ,θ 2i ,,θ ni −1 (10) 11 That is the value of θ 1i is drawn from the conditional distribution of θ 1 given the data, priors, and the current value of every other element of θ for the i-1th iteration. θ 2 is then drawn from its conditional distribution given the data, priors, the value of θ 1 just drawn, and the i-1th value of the remaining elements of θ . This continues until θ N is drawn given the ith value of θ 1 through θ N −1 . At this point, the iterations start over for θ 1i +1 . The values of each of the draws are collected and used to make inferences about the quantities of interest. 4 Results The analysis of the CMS proceeds in two parts. The first section takes a simple descriptive look at the how the key variables change over time. The second section presents the results from the hierarchical logit model. 4.1 Policies, personalities, and uncertainty over time. As outlined above, the different theories of campaign effects imply different temporal patterns in the independent variables. A simple descriptive look at these patterns supports some theories over others. Unfortunately, the weekly averages of the variables are not that informative. Each week’s sample is based on an average 80 respondents. The aggregate measures, therefore, have a tremendous amount of sampling variability that masks the underlying temporal pattern. In an effort to remove this randomness and better capture the underlying changes, a non-linear Hanning smoother has smoothed the aggregate data. For this section, this applies to every figure shown. In the hierarchical logit, only the macro -level uncertainty measures that are smoothed. Figure One presents the smoothed values of the vote share for each Reagan. While there is some movement in the graph, at no point does Reagan do worse than 55% of the sample vote. This data mimics the patterns in other tracking polls—Reagan led from the beginning and his reelection was never in jeopardy. 12 0.8 0.75 0.7 Percent Vote for Reagan 0.65 0.6 0.55 0.5 1 3 5 7 9 11 13 15 17 19 21 23 Week 25 27 29 31 33 35 37 39 41 43 45 Figure One Reagan's Vote Share by Week The average policy distance and trait assessments are presented in Figure Two. The solid lines denote the Mondale, while the dashed are the patterns for Reagan. There is clearly some week to week movement in these, but there does not seem to be any systematic pattern across time. If respondents were changing their policy positions consistently as a response to campaign message, we would expect to see the policy distance for one candidate grow while the other shrinks. This does not appear to be the pattern. Finally, the start and end points of each of these series are roughly equal. Based on these graphs, if the impact of the components change over the course of the campaign, the change is probably not the result of systematic movement in the content of the policy or trait attitudes. What is interesting to note is that the distance between the respondent and Reagan is large than the distance between the respondent and Mondale. Reagan’s lead is despite this difference. In the final few weeks, this gap is closed. This movement precedes the spike at the end of the vote series. 13 3.5 3.3 Trait Assessment, Reagan Trait Assessment, Mondale 3.1 2.9 2.7 2.5 Policy Distance, Reagan 2.3 2.1 1.9 1.7 Policy distance, Mondale 1.5 1 3 5 7 9 11 13 15 17 19 21 23 week 25 27 29 31 33 35 37 39 41 43 45 Figure Two Independent Variables Over Time The macro level measures of campaign learning are derived from the aggregate level uncertainty measures discussed above. Figures Three and Four present the patterns in the policy and personality uncertainty for Mondale and Reagan. The policy uncertainty variables are bounded by zero and one to control for the different lengths of the issue scales in the CMS. The scale of the trait uncertainty ranges from zero to, roughly, ten. The trait scales do not vary in length, so they are not standardized to the same zero-one scale. The campaign does not appear to decrease the uncertainty about either Reagan’s policy positions or his character traits. While the trait uncertainty is lower by the end of the campaign, it appears to cycle some throughout and the uncertainty from the first week is one of the lowest of the campaign. The policy uncertainty has a slight upward trend. From week nine until 34, the level of uncertainty is either constant or increasing. There is some decrease in the final ten weeks of the campaign, but even at this point, the electorate appears to be more uncertain than at the start. The trends in uncertainty about Mondale fit the expectations more closely. While the changes are not monotonic, there is a clear downward trend in both the policy and trait uncertainty. The change in the trait uncertainty is quite dramatic. At the start of the campaign, the electorate is not very consistent in its depiction of Mondale’s character. By the end of the campaign, this has changed. Note that the uncertainty terms are not comparable across candidates. The terms are based on the variance in the candidate placements. This, in turn, is based on the mean placement. Because the scales are bounded, a candidates whose true position is near the mean of the scale has a lower maximum squared difference than those who are at the extreme. On a five-point scale, the largest deviation from a true position of “3” is two points. On the same scale the largest error for a candidate who is truly is a “5” is four point. Once squared, these differences make comparisons problematic. Thus, while the 14 electorate seems to be more certain about Mondale’s traits than Reagan’s, this may simply be a function of Mondale’s placement on the scale. 0.45 0.4 Mondale Uncertainty 0.35 0.3 0.25 Reagan Uncertainty 0.2 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 Figure Three Policy Uncertainty Over Time 2.2 2 Trait Uncertainty, Mondale 1.8 1.6 1.4 Trait Uncertainty, Reagan 1.2 1 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 Figure Four Personality Uncertainty Over Time These results are problematic for the uncertainty and enlightenment models. Each of these models assumes or implies that the campaign serves to educate voters about where the candidates stand. While this appears to be the case for Mondale, the Reagan evidence is less clear. If the campaign did alter voter’s uncertainty about Reagan, it increased it. Of course, these simple time trends do not get at the heart of the comparison 15 between the theories. The real test lies in how the campaign alters the importance placed on the determinants of vote choice. 4.2 Campaigns learning and vote determinants. MCMC based results differ from traditional frequentist or likelihood based models. While it is possible to develop point and variability estimates, these are not based on a distributional assumption in the same way. Instead, they are based on a random sample drawn from the marginal posterior distribution. The point estimate for a coefficient is the sample mean (or median) of this posterior. The variability is the standard deviation or confidence region of the posterior. Additionally, the posterior is not necessarily a smooth curve—it may be skewed or even multi-modal. A more accurate, but less interpretable way to represent the marginal posterior distributions of the coefficients is through its kernel density plot based on the draws from the chain. Again, the key tests of the effectiveness of campaigns are lies in the estimates of the macro parameters. The attitude strength and other models of campaign effects make different predictions about how we should expect the important determinants to change. The results of these tests are found in Table Two and the density plots are in the Appendix. Table Two: Vote Determinants (Mondale Vote = 1, Reagan Vote = 0) Variable Mean Standard Confidence Interval Deviation Constant Intercept 0.50 0.71 -0.78 2.24 ln(week) 0.47 0.34 -0.13 1.21 Partisanship Intercept -0.49 0.16 -0.81 -0.18 ln(week) 0.08 0.08 -0.07 0.24 Certainty -0.44 0.23 -0.91 -0.03 Reagan Traits Intercept -0.94 0.49 -1.98 -.07 ln(week) -0.43 0.15 -0.71 -0.16 Uncertainty -0.61 0.29 -1.16 -0.02 Mondale Traits Intercept 0.65 0.53 -0.29 1.84 ln(week) 0.45 0.15 0.15 0.72 Uncertainty 0.58 0.25 0.11 1.09 Reagan Policy Intercept 0.39 0.24 0.03 1.00 ln(week) -0.01 0.06 -0.13 0.10 Uncertainty -0.16 0.62 -1.83 0.59 Mondale Policy Intercept -0.05 0.21 -0.46 0.34 ln(week) -0.07 0.06 -0.19 0.06 16 Uncertainty Age White Male -0.03 0.002 -1.05 -0.10 0.34 0.006 0.35 0.18 -0.66 -0.01 -1.73 -0.46 0.57 0.01 -0.36 0.25 The results in Table Two are somewhat difficult to interpret. The bottom three rows are straightforward. Whites are more likely to vote for Reagan while age and gender are unrelated to vote choice. The estimates for the other variables are not as simple. For each of the variables, including the constant, the micro and macro parameters are collected under the italicized headings. For the constant, the intercept in the macro equation is 0.50, while the coefficient for the log of the number of weeks is 0.47. Neither of these estimates is distinguishable from zero, as indicated in the standard deviation and confidence interval of the posterior distribution. The impact of partisanship changes over the campaign. Again, the macro certainty measure is the sample’s assessment of Mondale’s chance for the nomination. The expectation of the campaign learning model is as Mondale’s chances increase, partisanship should be a stronger predictor of vote. This is what we see in these results. The effect of partisanship separate from the time of the campaign and the status of the Mondale campaign is negative and strong. The macro parameter indicates that as Mondale’s chances improve (or he is officially declared the nomination) the effect of partisanship increases dramatically. Figure Five plots the coefficient across time.3 Note that this is not a predicted probability graph that is common with a logit model, but changes in the actual coefficient based on changes in the macro level variables of interest. The curve’s step is the point in time where Mondale is officially endorsed by the convention. From that point on, the probability that he will be the nominee is one. 3 The effect of the logged week is set to zero. 17 0 -0.1 -0.2 -0.3 -0.4 -0.5 -0.6 -0.7 -0.8 -0.9 -1 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 Figure Five Effect of Partisanship Over Time The trait coefficients also show a tremendous amount of movement, but are difficult to interpret directly. All six of the macro coefficients are significant. Regardless of the level of certainty or the time of the campaign, the traits exert an impact on the vote choice. Those who like Reagan’s character are more likely to vote for him. Likewise, the more a voter likes Mondale’s traits, the more likely they are to vote for him. The campaign bolsters these effects. Both of the time coefficients are of the same sign as the intercept. The combined effect of time and changes in the uncertainty is to more than double the magnitude of the coefficients from January to November. The uncertainty terms present something of a conundrum. The hypotheses generated by the theories of campaign influence hold that the as the campaign clarifies the candidates’ characters the coefficients should increase in importance (the coefficients should be of the opposite sign of the intercepts). The results from the random coefficient model, however, suggest that the voters use the traits more when the electorate is uncertain about the traits. As the variability decreases, the traits are worse predictors of vote choice. This is explained, in large part, because the vast majority of the movement in the coefficients is determined by the campaign timing variable. Additionally, the collinearity between the two terms actually serves to alter the sign of the macro level coefficients. If uncertainty is included as the sole macro-level predictor for both trait measures, it is correctly signed. What is important is the combined effect of the macro level changes that serves to bolster the trait assessments in the decision making process. 18 4 3 Mondale Trait Coefficient 2 1 0 -1 -2 Reagan Trait Coefficient -3 -4 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 Figure Six Personality Coefficients Over Time The policy coefficients do not change over the campaign. The only macro coefficient for the policy terms that is significant is the intercept for distance from Reagan. The greater the issue distance from Reagan the higher the probability of supporting Mondale. The average policy distance from Mondale is not an independent predictor of vote choice. Some of this is due to the relationship between this and the other micro level variables (distance to Mondale is highly related to distance from Reagan), but the added complexity of the model makes it difficult to separate out the direct impact of this variable. 4 Conclusion Which of the models of campaign effects is the most accurate? The results are not completely conclusive, but some models clearly fair better than others. The minimal effects model is clearly not supported. The weights applied to partisanship and the perceptions of the candidate personality traits do change throughout the campaign. While this is not the resounding differences that Berelson et al were looking for, it does appear that the campaign is an important element in the final vote outcome. Of course, the minimal effects have undergone serious criticism for almost two decades now, so this is not surprising. The enlightenment and uncertainty models also fair poorly. Unlike Gelman and King’s results, party does increase in importance throughout the campaign. The essential fundamental variables, the policy distance between the voters and candidates are not weighed more heavily—in fact, one of the two issue distances is insignificant. The changes in the support for Reagan is driven mostly by changing the weight applied to candidate character. The uncertainty model fares even worse because it posits the exact opposite relationship for the non-policy information. The campaign information increases the use of the character attitudes. 19 Support for the attitude strength model is mixed. The content of the campaign information works for some attitudes (the trait assessments) but not for others. This points to the strength of a model that is not based on voter utilities. It is important to uncover the heterogeneity in the voting rules used by the electorate. While Alvarez’s model takes an important step in the moderating role of uncertainty, the specification is too limiting. An attitude strength approach to the same general relationships generates hypotheses to situations where a utility function does not necessarily fit. The increased importance of the personality information suggests that the social psychological approach may provide a better fit with how voters weigh different attitudes when voting. Finally, the learning model is also supported. As would be expected with the changes in partisanship found in Mebane and Wand (1997), the use of partisanship in Mondale evaluations is strengthened by the campaign. At the same time, the learning model cannot account for the other results. There is no consistent pattern for changes in content of the policy or personality attitudes over the campaign. The changes reported above stem from changing weights, and not changing content. This works is not definitive. First, the nature of the 1984 campaign limits the generalizability of the results. Reagan’s tremendous, and nearly constant lead may have mitigated the amount of change. If the election outcome was never in question, the campaign may be a poor choice to study. Unfortunately, the only data of American elections has the necessary structure. It may be necessary to branch out to other countries (the Canadian data used by Johnston et al) or data that do not contain all of the relevant measures, perhaps a set of tracking polls like those used in Gelman and King (1994) Additionally, there is very little information about the actual campaign included in the macro level model. The aggregate level uncertainty and timing measures serve as proxies for the campaign information and voter learning, but it does not directly measure the campaign flow. Incorporating a direct measure of campaign information, or even indicators of important campaign dates (debates, conventions) could provide a broader set of tests of campaign effects. The random coefficient logit model makes it relatively easy to incorporate tests of a broader range of campaign effects. Finally, the statistical model developed in this paper is worthy of mention because it can be applied to a host of other questions in political behavior. Alvarez’s work on uncertainty and candidate evaluation test the same model on six different election studies. Rather than estimate six different models, and presume each contains no information about the other, it is possible to specify a single model that would encapsulate every election study and specify how the specifics of the election alter the relationships. The works of Branton and Jones (Branton and Jones 1999) have made use of the geographic clustering in NES sampling to test for contextual effects. Social and political psychologist are becoming more interested in the moderating role of strength on attitudes. It is relatively simple to re-specify the clustering in the hierarchical model from something based on the sampling technique, to some function of the underlying psychological processes. 20 References Alvarez, R. Michael. 1998. Information & Elections. Ann Arbor, MI: University of Michigan Press. Second edition. Bartels, Larry M. 1988. Presidential Primaries and the Dynamics of Public Choice. Princeton, NJ: Princeton University Press. Bartels, Larry M. 1993. “Messages Received: The Political Impact of Media Exposure.” American Political Science Review. 87:267-284. Bartels, Larry M. 1996. “Uninformed Votes: Information Effects in Presidential Elections.” American Journal of Political Science. 40:194-230. Berelson, Bernard R., Paul F. Lazardsfeld, and William N. McPhee. 1954. Voting. Chicago: University of Chicago Press. Branton, Regina P., Bradford S. Jones. "Multiculturalism, Diversity, and Prejudice." Paper presented at the annual meeting of the Western Political Science Association. Campbell, James E. 1983. “Ambiguity in the Issue Positions of Presidential Elections: A Causal Analysis.” American Journal of Political Science. 27:284-293. Cowles, M.K. and Brad P. Carlin. 1996. “Markov Chain Monte Carlo Convergence Diagnostics: A Comparative Review.” Journal of the American Statistical Association. 91:883-904. Delli Carpini, Michael X. And Scott Keeter. 1996. What Americans Know About Politics and Why it Matters. New haven CT: Yale University Press. Dempster, A. P. 1983. “Comments on Morris (1983).” Journal of the American Statistical Association. 78:57-58. Dempster, A.P. and A. Monajemi. 1976. “An Algorithmic Approach to Estimating Variance,” Research Report S-42, Harvard University, Department of Statistics. Erikson, Robert, Michael MacKuen, and James A. Stimson. forthcoming. Macropolity. New York: Cambridge University Press. Gelman, Andrew, John B. Carlin, Hal S. Stern and Donald B. Rubin. 1995. Bayesian Data Analysis. London: Chapman & Hall. Gelman, Andrew and Gary King. 1993. “Why are American Presidential Elections Campaign polls so Variable When Votes are so Predictable.” British Journal of Political Science. 23:409-451. 21 Harville, D.A. 1974. “Bayesian Inference for Variance Components Using only Error Constraints.” Biometrika. 61: 383-385. Jackman, Simon. 1999. “Estimation and Inference via Bayesian Simulation: An Introduction to Markov Chain Monte Carlo.” Working Paper. Johnston, Richard, André Blais, Henry E. Brady, and Jean Crête. 1992. Letting the People Decide: Dynamics of a Canadian Election. Stanford, CA: Stanford University Press. Key, V.O. 1966. The Responsible Electorate. Cambridge: Harvard University Press. Lazardsfeld, Paul, Bernard Berelson, and Hazel Gaudet. 1944. The People’s Choice. New York: Columbia University Press. Luskin, Robert C. 1987. “Measuring Political Sophistication.” American Journal of Political Science. 31: 856-899. Mebane, Walter R. Jr. and Jonathan Wand. 1997. Markov Chain Models for Rolling Cross-section Data: How Campaign Events and Political Awareness Affect Vote Intentions and Partisanship in the United States and Canada." Paper presented at the annual meeting of the Midwest Political Science Association. Patterson, H. D. and R. Thompson. 1971. “Recovery of Interblock Information When Block Sizes are Unequal.” Biometrika. 58: 545-554. Petty, Richard E. and Jon A. Krosnick. 1995. Attitude Strength; Antecedents and Consequences. Mahweh, NJ: Lawrence Earlbaum Associates Seltzer, Michael, H., Wing Hung Wong, and Anthony S. Bryk. 1996. “Bayesian Analysis in Applications of Hierarchical Models: Issues and Methods.” Journal of Educational and Behavioral Statistics. 21:131-167. Sniderman, Paul M., Richard A. Brody, and Philip E. Tetlock. 1991. Reasoning and Choice: Explorations in Political Psychology. New York: Cambridge University Press. Stimson, James A. 1975. “Belief Systems: Constraint, Complexity, and the 1972 Election.” American Journal of Political Science, 19:393-417. Western, Bruce. “Causal Heterogeneity in Comparative Research: A Bayesian Hierarchical Modeling Approach.” American Journal of Political Science. 42:1233-1260. Wong, George Y. and William M. Mason. 1985. “The Hierarchical Logistic Regression Model for Multilevel Analysis.” Journal of the American Statistical Association. 80:513-524. 22 Zaller. John R. 1992. The Nature and Origins of Mass Opinion. 1992. New York: Cambridge University Press. 23