Dialogue in U.S. Senate Campaigns? An Examination of Issue Discussion in Candidate Television Advertising Noah Kaplan Assistant Professor of Political Science University of Houston Hoffman Hall (PGH) Room 447 Houston TX 77204-3472 (713) 743-3907 nkaplan@uh.edu David Park Assistant Professor of Political Science Washington University Campus Box 1063 One Brookings Drive St. Louis MO 63130-4899 (314) 935-5810 dpark@artsci.wustl.edu Travis N. Ridout Assistant Professor of Political Science Washington State University Department of Political Science PO Box 644880 Pullman WA 99164-4880 (509) 335-2264 tnridout@wsu.edu Please address all correspondence to the third author. Studies have found that electoral campaigns in the U.S. do not facilitate dialogue. That is, competing candidates rarely discuss the same issues. Indeed, Simon s (2002) formal model predicts that dialogue should not occur at all under normal circumstances. We extend Simon s model by incorporating candidate uncertainty regarding the median voter s preferred position. This modified model predicts that candidates may adopt a mixed strategy for discussing an issue, thus producing some probability that dialogue will occur. This suggests that the greater the uncertainty regarding which candidate is closer to the median voter, the greater the expected level of dialogue. We also develop a new measure of dialogue based upon television advertising in U.S. Senate campaigns during the 1998-2002 period. This measure indicates that there is substantial variation in the extent of dialogue across campaigns. Finally, we predict variation in dialogue across campaigns based on several issue-specific and campaign-specific factors.
Dialogue in U.S. Senate Campaigns? An Examination of Issue Discussion in Candidate Television Advertising 1. Introduction Recent scholarship suggests that on the campaign trail, competing candidates avoid engaging in dialogue. Instead, they talk past each other, emphasizing different issues. Petrocik (1996), for instance, finds that the 1980 presidential candidates spoke overwhelmingly about issues owned by their respective parties. Only four percent of newspaper stories generated by the Republican Ronald Reagan and his surrogates concerned Democratic issues. Similarly, Simon s (2002) analysis of U.S. Senate campaigns concludes that only ten to fifteen percent of campaign activity involves dialogue with one s opponent. Such findings are troublesome to democratic theorists who emphasize the importance of discussion and deliberation in helping voters to reach informed decisions at the ballot box. Dialogue (or discourse) occurs when two competing candidates talk about the same issues on the campaign trail, and it is an essential prerequisite of informed democratic choice. Kelley (1960), for instance, suggests that a campaign should expose the grounds on which candidates disagree and the differences between the candidates differences of personality, interest, affiliation, policy commitment, and all others that may affect performance in office (p. 14). Yet Kelley s standard is unattainable if competing candidates discuss completely different issues. Indeed, the contemporary campaign described by Petrocik and Simon seems consistent with Fishkin s (1991) description of non-deliberation: an endpoint where an alternative is not contrasted effectively with its rivals, where arguments are not answered, and where the decision 1
makers have little competence or factual background to evaluate the proposals offered to them (p. 38). If dialogue is essential for democratic choice, then why might instances of dialogue be so rare on the campaign trail? Extending Petrocik s logic, Simon suggests that it is in the best interest of competing candidates for each to emphasize a different set of issues, rather than engaging in dialogue with each other. Voters cast ballots on the basis of issues salient to them (RePass 1971). Because of this, the role of the campaign is to make certain issues salient to voters, as opposed to changing their views on the issues, a much more difficult task (Petrocik 1996). By emphasizing issues on which a candidate or his party has an advantage among the electorate, the candidate tries to make the electorate vote on that basis. Sticking to one s own issues is advantageous to both candidates. For example, if a Republican candidate talks about crime, an issue on which the vast majority of voters view that party s approach favorably, then a Democratic opponent who also discusses crime only increases the salience of the issue in the minds of voters. The Democrat will be better off to ignore crime and emphasize an issue on which her party is viewed favorably, such as Medicare. Moreover, political messages may be most effective when they play to existing party stereotypes (Ansolabehere and Iyengar 1994). At a psychological level, people may be more willing to accept messages that are consistent with their previous beliefs about party issue positions, and they may find such messages more credible, thus increasing their persuasiveness. In The Winning Message (2002), Adam Simon presents a formal model that predicts that rational politicians, those with the goal of election foremost in mind, do not 2
engage in dialogue with each other. Only in three instances does Simon expect to see dialogue when the mass media create it, during critical elections when one issue is overridingly salient among the public, and when, for whatever reason, the candidates behave irrationally. Very rarely, then, according to Simon s logic, should one observe any dialogue in a campaign. Yet this model omits several important considerations that might entice rational candidates to engage in dialogue. We focus on one in particular the uncertainty that candidates have about voter positions. In Section 2 we outline Simon s model and extend it to incorporate candidate uncertainty regarding the median voter s position. We also outline other explanations for dialogue. In Section 3 we discuss the data and the statistical model we use to analyze the data. In Section 4, we present results from a quantitative analysis predicting the extent of dialogue in the U.S. Senate campaigns of 1998, 2000, and 2002. Section 5 discusses these results and concludes. 2. Theories of Dialogue 2A. Simon s Model Simon s model is a non-cooperative, static game with complete information (see Simon, Appendix A for a formal presentation of the model). In this game, there are three players: candidate A, candidate B, and the median voter. Each issue dimension is a single dimension, which is independent of all other issue dimensions. Each player has a policy ideal point for each issue dimension. These ideal points are exogenous to the game (set by nature in the first play of the game) and cannot be changed. All players have the standard quadratic loss function as policy diverges from the players ideal points. 3
Candidates wish to win the election. Candidates have a finite budget, which they use to discuss issues. Additional discussion has a decreasing marginal benefit (i.e., there is a decreasing marginal effect of talking on the issue s salience). The candidates action set is to allocate resources across issues the second play of the game. The candidate who wins the election can implement his or her most preferred policy (the final play of the game). For each dimension, the median voter prefers whichever candidate is closest to him or her. The median voter s action set is to vote for either candidate A or candidate B the third play of the game. The median voter s decision rule is to sum preferences across all salient dimensions. However, each dimension is not weighted equally. Rather, each dimension is weighted based upon the extent to which the issue is salient. The salience of an issue depends upon its initial value as determined by nature and the extent to which candidates discussed the issue during the campaign (thus the decision rule is based upon a weighted average across issues, with the issues most discussed by the candidates given the greatest weight by the median voter). In other words, salience of an issue is a monotonically increasing function of the candidates discussion of that issue. Simon specifies that the two candidates are fully informed about their respective ideal points, their budget allocations, and the distributions of the median voter s ideal points. Assuming a finite budget, Simon concludes that a candidate should allocate resources to those issues for which the candidate has an advantage (by advantaged, Simon simply means that the distance between the candidate s ideal point and the expected ideal point of the median voter is less than the distance between the opponent s ideal point and the expected ideal point of the median voter). In other words, each 4
candidate has a dominant strategy: do not allocate any resources to issues upon which the candidate is disadvantaged because it can only help the opponent win the election. 2B. The Model With Uncertainty In effect, for each issue, candidates are one of two types of players advantaged or disadvantaged (and for each issue, one candidate must be advantaged and one must be disadvantaged). One reason that Simon s model has a pure strategy Nash equilibrium is because Simon assumes that the candidates know their own types (and hence their opponents types) with certainty. What happens if we permit the candidates to be uncertain about the location of the median voter s ideal point? Doing so means that the candidates are uncertain about their type. What strategy should a candidate adopt if she thinks there is a probability, p, that she is advantaged on any one issue and that her opponent is advantaged with probability 1-p? Or, to use equivalent language, what is the optimal strategy for a candidate when she thinks there is a probability, p, that nature determined a state of the world in which she is the advantaged candidate for an issue, and that there is some probability, 1-p, that nature selected a state of the world in which her opponent is the advantaged candidate for that issue? As Harsanyi demonstrated in his classic series of papers, uncertainty can be introduced to a static game with complete information that has a pure strategy Nash equilibrium by specifying uncertainty regarding the types of the players. This type of game is defined as a static Bayesian game, and Harsanyi showed that a mixed strategy is optimal for the players in such a game (i.e., that a Bayesian Nash equilibrium exists). A mixed strategy for a candidate on any one issue is to randomly allocate resources to discussing that issue with probability p. So, for example, if candidate A thought that she 5
was the advantaged candidate with p=0.6 for a particular issue, then candidate A should adopt a decision rule to allocate resources to that issue such that if a random number generator (which generated integers from 0 to 9) produced an integer of less than 6, she would allocate resources to that issue. Otherwise, she would not allocate resources to that issue. Once Simon s model has been modified to incorporate uncertainty regarding the types of the candidates (advantaged or disadvantaged), the model no longer predicts zero dialogue in a campaign. Rather, the model has two equilibria: the pure Nash equilibrium of no dialogue as discerned by Simon, and a mixed strategy equilibrium in which dialogue occurs. 1 Of course, the extent of dialogue in the mixed strategy equilibrium depends upon both the candidate s estimate of p for any one issue, and upon the number of issues for which a candidate is uncertain regarding her type. This model predicts no dialogue only in that subset of cases in which the candidates know with certainty their types (and have a shared knowledge of their types) for all issues. By assuming some uncertainty, not only does the model permit dialogue, but it also predicts that the greater the uncertainty, the greater the probability of dialogue. This is because the probability of dialogue for any one issue increases as the probability of being the advantaged player approaches 0.5, and the probability of dialogue increases as the number of issues for which a candidate is uncertain of her type increases. Put more evocatively, the more uncertain the candidates are about the positions of the median 1 As long as there is some possibility p that the voter is closer to candidate A on issue 1 and closer to candidate B on issue 2, and some possibility (1-p) that the voter is closer to candidate B on issue 1 and closer to candidate A on issue 2, then there is a mixed strategy equilibrium. This is similar to the logic for portfolio diversification across two stocks given that there is some probability that stock A will out- perform stock B, and there is some probability (1-p) that stock B will out-perform stock A in a specified period. 6
voter, the more uncertain they are about whether they are advantaged or not. Thus, the more uncertain the candidates are about the ideal points of the median voter, the more likely one would observe dialogue. Candidates might be uncertain about the most preferred issue policies of the median voter and how close their own positions are to those positions for a variety of reasons. For example, candidates might be uncertain about the public s preferences due to the inherent limits of polling. 2 Though polling provides candidates with information regarding the issue preferences of the public, polling results are always subject to uncertainty and multiple interpretations. The infeasibility of polling the population means that polls based upon a random sample of the population contain a margin of error. Furthermore, public opinion on an issue may depend on how that issue is framed. Extensive work using survey data has shown that slight changes in question wording can significantly alter the distribution of public opinion on an issue (Zaller and Feldman 1992). So although two candidates might disagree about the best approach to an issue, each can present his or her approach in a manner that appeals to the majority of voters. Take, as an example, the issue of school vouchers. A majority of Americans might believe that government should help inner-city parents send their children to good private schools and simultaneously believe that government should not give up on the public schools. Thus, both pro-voucher and anti-voucher candidates might do well to advertise on the issue, depending on the effectiveness of their frames. 2 Polling is just one source of information regarding the policy preferences of a candidate s constituency. Candidates have extensive networks of informants (Fenno 1978). Though such networks are an important source of information for candidates and politicians, candidates understand their limitations and constantly attempt to supplement such information via polls, discussions with other candidates and/or politicians, and the mass media. 7
Also, public opinion can change over time. Thus polling results can become outdated, and even misleading to the extent that results from past polls are assumed to reflect the current preferences of the public. Finally, polling is expensive, and the more issues a poll investigates, the more expensive the poll. Thus, finite fiscal resources may prevent candidates from being able to use polls to gauge public opinion on issues at any one time, much less over time. 2C. Alternative Explanations of Dialogue 3 Because candidates face uncertainty about where the median voter lies on an issue, we argue that there is the potential for considerable issue discourse at least in some campaigns. But the extent of the dialogue is likely to vary depending on the circumstances of the campaign. Scholars have identified several factors that may help to predict how much discourse occurs in a particular campaign. Simon controls for three issue characteristics: whether the issue is consensual or valenced, whether an issue is owned by a party, and the extent to which an issue is critical (i.e., the salience of an issue). Candidates may be more like to engage in dialogue on salient, or critical issues. As Kahn and Kenney explain: When issues such as the state of the economy, or the rising costs of health care, or the high incidence of crime receive extensive media attention, they become more salient to citizens. In these situations, candidates feel compelled to discuss these topics because voters view these issues as especially pressing. [For example, m]edia coverage about the economy 3 Alternative is not meant to imply contradictory. Alternative factors may well be complementary, empirically and theoretically. 8
was so prevalent in New Hampshire throughout 1992 that any effort to avoid the issue would have been a strategic mistake (1999, 18-19). Candidates also may be more likely to engage in discourse with their opponents when the issue is not owned by either party. Issue ownership occurs when a party develops a reputation for policy and program interests, produced by a history of attention, initiative, and innovation toward these problems, which leads voters to believe that one of the parties (and its candidates) is more sincere and committed to doing something about them (Petrocik, p. 826). Thus, a Republican candidate is unlikely to talk much about the environment, an issue on which the electorate views Democrats as best able to handle, and a Democratic candidate is unlikely to speak much about fighting crime, an issue that citizens perceive Republicans as best able to handle. This argument suggests that it is the issues on which no party has clear ownership where one will likely see the most dialogue. And, as Simon notes, candidates may be more likely to engage in dialogue on those issues that everyone agrees are laudable goals. Stokes makes a distinction between position issues and valence issues. Position issues (or non-consensual issues) refer to issues that involve advocacy of government actions from a set of alternatives over which a distribution of voter preferences is defined (p. 170). These are issues on which parties take different positions. The availability of abortion or whether restrictions should be placed on gun ownership are classic examples of non-consensual issues. Valence issues, or what we will term consensual issues, by contrast, involve the linking of the parties 9
with some condition that is positively or negatively valued by the electorate (p. 170). 4 Government integrity is an example of such an issue all candidates, regardless of party, advocate and want to be associated with government integrity. Unable to identify specific policy solutions, the voters are left with the question of trust in competing parties regarding these issues. Both types of issues are commonly observed in nearly all election campaigns. Thus we expect to observe greater dialogue regarding consensual issues than valence issues. While the extent of dialogue may vary across issue areas, it may also vary across campaigns. Somewhat surprisingly, the only campaign-specific characteristic that Simon included in his model was campaign competitiveness. We hypothesize that the following campaign characteristics might affect dialogue also: whether it s for open seat, the relative richness of the campaign, the size of the electorate, and the extent to which the candidates go negative. It has long been recognized that incumbents have important electoral advantages over their opponents, such as greater fund raising capabilities, greater media coverage and greater name recognition among the electorate. Incumbents tend to maximize these benefits by minimizing activities that provide their opponents with greater visibility. For example, incumbents are less likely than incumbents to find debates beneficial. Thus, we might expect to find candidates in open seats engaging in more dialogue than those races with incumbents. The richness of the information environment may have an impact on how much dialogue occurs in a campaign. Richness can be thought of in a couple of different ways. 4 The distinction between a consensual and non-consensual issue is often not clear-cut. Take, for example, the economy. On the one hand, all politicians like to be associated with a growing economy, but the parties do offer different proposals for reaching that end. 10
The richness of the information environment can be conceptualized as the extent to which the candidates are issue oriented. Quite simply, the more issue messages that the candidates disseminate, the greater the potential to engage in dialogue. Alternatively, richness can be thought of more literally in terms of fiscal resources. Because resources are finite, the candidates cannot talk about everything they might want to in their campaign communications. Thus under-funded challengers may be less likely to engage in dialogue with their opponents than well-funded challengers. Lee and Oppenheimer (1999) suggest that the political heterogeneity of a state s population may influence senate campaigns along a variety of dimensions (p. 83-122). They review the various arguments for measures for political homogeneity, and argue persuasively that state size is the best measure of heterogeneity, given the imperfections of all such available measures. Extending this logic suggests that senate candidates will have a harder time ducking issues in smaller, more homogeneous states than larger, more heterogeneous states. Finally, a campaign may also feature more dialogue when one of the candidates goes negative. A cardinal rule of American politics is that candidates must respond to attacks by an opponent (Lau et al. 1999). Thus, when a candidate accuses an opponent of having a bad position on an issue, the opponent may be tempted to respond on that issue, thus introducing dialogue into the campaign. So if a Republican runs an advertisement claiming her Democratic opponent wants to raise your taxes, the Democrat may respond with an advertisement pledging not to raise taxes. 3. Model and Data 3A. The Model 11
OLS regression is the standard quantitative technique for analyses with a continuous dependent variable (OLS results are reported in the first column of Table 6). However, OLS assumes that cases are independently and identically drawn (i.i.d.). A number of factors can lead to the violation of the i.i.d. assumption. For example, data which is longitudinal often violates the i.i.d. assumption because observations are usually related to one another (e.g., a respondent s answer at time t+1 is not independent of that respondent s answer at time t). We suspect that the data used in this analysis violate the i.i.d. assumption in two ways. First, there is the possibility of heterogeneity within a campaign across issues (i.e., that the issues discussed within a campaign are not independent of one another). Alternatively, there may be heterogeneity within an issue across campaigns. The latter is equivalent to saying that dialogue on an issue at the national level influences the dialogue of that issue at the campaign level. One standard approach for testing and correcting for heterogeneity is to estimate a random effects model. A random effects model estimates a parameter for within unit variance. A value of this parameter significantly different from 0 indicates that the homogeneity (i.i.d.) assumption is violated. The random effects model adjusts the structure of the variance-covariance matrix and by doing so provides superior estimates of the coefficients S.E.s based upon the extent of the within-unit variance relative to the between-unit variance. We run random effects models to test for various types of heterogeneity. A random effects model is the most basic specification of the more general multilevel model. As noted in regard to the random effects model, multilevel modeling allows us to correctly estimate the variation in the estimated parameters given the 12
clustering of observations (i.e., heterogeneity), and it provides a good compromise between two classic alternatives in the estimation of the model coefficients. One alternative to a multilevel model simply pools all of the cases, assuming their independence. But this approach results in standard errors that are biased downward, which lends too much confidence to inferences. A second alternative to a multilevel model analyzes each group separately. But this approach ignores information about the highest level and prevents observations in different groups from borrowing strength from each other. Thus, a multilevel correctly inflates the standard errors because of clustering but does not ignore data in estimating model coefficients. Another advantage of multilevel modeling is that we can directly combine different data sources. In our example, we have levels of dialogue at the campaign level and predictors at the issue (national) and campaign (local) levels. One possible approach to estimation would be to run a linear regression model with an indicator (dummy variable) for each issue. But in classic regression it is not possible to include issue-level indicators as well as issue-level predictors since they would be collinear. To avoid this problem, one could fit the model with issue indicators but without issue-level predictors and then fit a second model. But because this approach relies on the classic regression estimates of the coefficients for the issue-level indicators, they may produce biased and inefficient estimates. For our data, one can think of the issue-level predictors as national in character (so we have an N of 43), but the campaign-level predictors as local (or state) in character (N of 65). We can write a multilevel model as a linking of local regressions in each 13
group (i.e., issue) as in (1). Within each group j, a regression is performed on the local predictors X, with a constant term α that is indexed by group: y = + X + N i n (1) 2 i α j( i) ( β ) i εi, εi ~ (0, σ y ), for = 1,...,. Here, the matrix X of local predictors does not include a constant term, since this is represented by α. The local errors ε i represent variation among the levels of dialogue within campaigns that is not explained by the campaign level predictors. The second state of the model describes the α parameters given group-level predictors W: 2 α j = ( Wγ ) j + η j, η j ~ N(0, σ α ), for j = 1,..., J. (2) W is a vector of issue indicators and γ j is a set of coefficients to which we fit a 2 multilevel model (3). The variance parameter or hyperparameter, σ α, is given uniform non-informative prior distributions (4). γ = (3) 2 j N(0, σα ) 2 σ ~ uniform(0,1000) (4) α The group-level errors represent the variation among issues that is not explained by the issue- and campaign-level predictors. The two models (1) and (2) must be estimated simultaneously. It is natural, but not correct, to first use linear regression to estimate the parameters in (1), and then use the estimated α j s to estimate the parameters γ, σ α in the regression (2). The computational challenge of multilevel modeling is to estimate the equations together. Therefore, we employ Markov Chain Monte Carlo (MCMC) simulation via WinBUGS (see Appendix A to see the full model in WinBUGS). 14
3B. The Data In seeking to tap dialogue in campaigns, we examine the candidates television advertising and the issues mentioned therein. We believe that this approach offers the truest expression of how candidates choose to express themselves to voters given resource constraints. Simon, by contrast, examines newspaper coverage of U.S. Senate campaigns in a state s leading newspaper to tap dialogue. His focus here is on candidate-initiated stories, those written about situations in which candidates, rather than reporters, raise issues. While this focus ameliorates some of our concerns about the use of newspaper stories to measure dialogue, examining newspapers still has some drawbacks. For one, newspaper coverage may reflect the priorities of individual journalists. Although Simon may limit his analysis to candidate-initiated stories, one wonders how many stories candidates tried to initiate, but that journalists did not deem newsworthy. Moreover, whether what a candidate wants to talk about makes it into the newspaper may depend on the size of the newspaper s staff, the size of the news hole and the priorities of editors. Finally, journalistic norms of fairness may lead to an overstatement of the amount of dialogue taking place. In other words, if one candidate talks about an issue, the journalist may seek out the opposing candidate s comments on the issue. The data on the content of candidate advertising come from the Wisconsin Advertising Project, which compiled commercial advertising tracking data during the 1998, 2000 and 2002 elections and coded each advertisement (see Ridout (2003) for a more detailed description of the data collection and assessments of its quality). A team of coders evaluated each advertisement on several attributes, including the issues 15
mentioned. Coders were allowed to mark up to four issues per advertisement. Coders could choose from approximately 50 different issue categories each year. Almost all categories were consistent across years, though a few that reflected topical concerns, such as the Clinton-Lewinsky scandal, were dropped, and others, such as terrorism, were added. We examined general election advertising in U.S. Senate races in the 75 largest media markets in 1998 and 2000, and in the 100 largest markets in 2002. Only a few Senate races were not covered by one of these markets those in Alaska, Hawaii, Montana, North Dakota, South Dakota, Vermont and Wyoming. 5 This resulted in a total of 65 campaigns (in 38 states) and 982 instances of dialogue. 6 Although the dataset is quite comprehensive, we do want to note one limitation: We are missing data on fourteen Senate races because these races occurred in states that did not overlap with one of the top seventy-five media markets in the country. These fourteen races were not missing at random; they were from the states with the smallest populations. This could potentially bias the coefficient on one of our predictor variables, a state s voting age population. Intuitively, though, we would expect that these missing cases would bias the absolute value of the coefficient downward, since we are missing extreme values on the left hand side of the distribution of states voting age populations. We also excluded from our dataset all U.S. Senate races that had incumbents without major party opponents or opponents who were sacrificial lambs. We define a 5 In addition, candidates in the following Senate races aired no advertising in the markets tracked: the 2002 Delaware, 1998 Idaho, 2000 and 2002 Massachusetts, 2000 Ohio and 2002 Virginia races. 6 Cases were dropped if the following conditions were met: (1) Democratic or Republican mentions were missing; or (2) Democratic and Republican mentions were zero; or (3) The issue was coded Other or None; or (4) There were no Democratic or Republican challengers 16
sacrificial lamb as a major party candidate who did not run any television campaign advertisements. This is readily justifiable since the model, as specified, requires the presence of two major party candidates. We also excluded advertisements run by parties and interest groups. Theoretically, these are independent expenditures beyond the control of candidate decision-making. If one s goal is to analyze the extent to which candidate decision-making itself results in discourse, as is our goal, then these exclusions are not problematic. 7 If, however, one s purpose is to examine the effect of discourse on the citizens who receive messages about the candidates, then one might want to expand the analysis beyond paid television advertisements to include news media coverage of the candidates as well. 8 The Dependent Variable Dialogue Dialogue is measured on a scale ranging from 0 to 1, with 0 indicating no dialogue on an issue and 1 indicating the maximum possible dialogue. This measure provides the relative amount of dialogue that occurs on a given issue by campaign. The specific calculation of the measure is shown in Equation 1. Dialogue = 1 Total Democratic Mentions - Total Republican Mentions (1) Total Democratic Mentions + Total Republican Mentions 7 We already noted that the media probably under-report candidate discussion of consensual issues due to the media s need to focus on dramatic and conflict laden stories. For this reason, the media also focus much more attention on the more competitive races than on those races viewed as less competitive. Newspapers tend to support incumbents disproportionately, and this editorial page tendency is correlated with a newspaper s tendency to provide more news reportage of incumbents (Kahn and Kenney, 2002a). 8 On average, ads aired by sponsor broke down as follows: 81.4% were sponsored by the campaigns, 14.6% were sponsored by the parties, 2.8% were sponsored by interest groups, and the sponsor could not be identified for the remaining 1.2% of ads aired. 17
For example, in the 2000 New York senate campaign between Democrat Hillary Clinton and Republican Rick Lazio, Clinton mentioned the economy 4,107 times and Lazio 2,971. Therefore, dialogue is measured as: = 1 4107-2971 4107 + 2971 = 0.84 There was much dialogue, then between these two candidates on the economy. Table 1 presents the 43 issue dimensions examined in the paper and four corresponding descriptive measures. In addition to the total number of issue mentions by Democratic and Republican candidates, the table includes a measure of dialogue on the issue, averaged across individual campaigns, and a measure of overall dialogue, calculated by considering all campaigns together. The table is sorted by average dialogue, which reveals that the candidates performances in office, Social Security, education and health care were the issues on which the candidates most often went toe-to-toe. [Insert Table 1 Here] One can also examine how dialogue varies across campaigns. Table 2 displays the average amount of dialogue in a campaign across those issues that at least one candidate mentioned. Our measure of dialogue is quite high in several campaigns.49 in the 2002 Missouri campaign,.38 in the 2000 Nebraska campaign, and.40 in the 1998 Nevada campaign. This indicates that when one candidate mentioned an issue, his or her opponent often mentioned that same issue and to a considerable degree. By contrast, the dialogue measure is 0 in six other campaigns, indicating that the candidates in the race never mentioned the same issue. 18
[Insert Table 2 Here] Independent Variables A rigorous test of the model would require independent variables which captured the extent of uncertainty regarding which candidate was advantaged on the various issues. Unfortunately, no such measure exists at the issue level. Consequently, the key independent variable is a measure of uncertainty at the aggregate level: margin of victory. We expect margin of victory to have a negative relationship with dialogue, since the greater the margin of victory, the less uncertainty there was prior to the election regarding which candidate was advantaged. Margin of victory is expressed in terms of the difference in the percentage of the total vote received by the winning candidate and his or her major-party opponent. The average margin of victory was 16 percent, with a standard deviation of 12. The obvious disadvantage of this aggregate measure of uncertainty is its ex post facto characteristic. Fortunately, CQ provides observers of congressional races with a characterization of each race s competitiveness prior to election day. This ex ante measure of aggregate uncertainty ranges from 0 to 3, with zero indicating an uncompetitive race and 3 indicating the most competitive races. This variable has a mean of 1.5 and a standard deviation of 1.2. Since a higher value indicates a more competitive race, we expect CQ s ranking to be positively related to dialogue. Control Variables at the Campaign Level Since the dependent variable is a function of the number of times an issue is mentioned at the campaign level, it is critical to control for the total number issue mentions during a campaign. We would expect the total number of times an issue was 19
mentioned in a campaign to have a positive relationship with the extent of dialogue on that issue. The variable total mentions is expressed in thousands of mentions. Similarly, we expect total spending on a race to have a positive relation to dialogue (indeed, total mentions at the campaign level and total spending had a Pearson s correlation coefficient of 0.61). Total spending is expressed in millions of dollars. The average spent on these senate campaigns was $12 million with a standard deviation of 14 million. 9 Following Simon s lead, we also controlled for whether the race featured an open seat. Open seats are coded 1; races with incumbents are coded 0. Out of the 65 senate campaigns, 12 were open seat contests and 53 included an incumbent. Descriptive statistics for these variables are presented in Table 3. [Insert Table 3 Here] Another aspect of the campaign that we include in the model is the negativity of the campaign, which we expect will have a positive effect on the amount of dialogue present. Negativity is measured as the natural log of the proportion of a campaign s total ad airings that are coded negative by coders. As Table 4 shows, the proportion of a campaign s total ad airings that were negative varies dramatically across states and years, from over.55 in the 2000 Ashcroft-Carnahan race in Missouri and the 1998 Hollings- Inglis race in South Carolina, to 0 in twelve other campaigns. 10 Table 3 provides summary statistics of the variable. 9 Total money spent sums the expenditures of the Republican and Democratic candidates. However, we had no prior commitment to how expenditures might influence dialogue. So, for example, we investigated the effect of expenditures by the challenger alone, and we investigated the effect of the difference in expenditures between the incumbent and the challenger. The operationalization of campaign expenditures had no effect on the regression results. 10 We thought it possible that a number of variables might have decreasing marginal effects on dialogue. Consequently, we investigated a number of non-linear relationships with dialogue for all of the continuous independent variables. Percent of negative ads was the only independent 20
[Insert Table 4 Here] The final campaign-specific variables included in the model are the voting-age population of the state, and the year of the campaign. The former is operationalized as the natural log of the state s voting age population in millions; Table 3 provides summary statistics. The latter control is included to reflect the possibility that year specific factors such as a concurrent presidential contest could change the dynamic of the senate races. Control Variables at the Issue Level Consensual issues, as discussed in Section 2C, are more or less goals, and so they are not generally subject for partisan dispute (Simon 2002:134). Non-consensual issues, by contrast, concern topics on which parties put forth and advocate different positions. Table 5 lists the twelve issues deemed consensual in character. Consensual issues were coded 1, the remaining issues were coded 0. Table 5 also lists the eleven issues determined to be owned by one of the two parties (owned issues were coded 1, all remaining issues were coded 0). We applied Simon s criteria to our count of issue mentions to determine which issues were owned by a party. Specifically, those issue which had a minimum to 2,000 total mentions and for which the ratio of total Democratic mentions to total Republican mentions was either greater than two or less one-half were defined as owned by a party (see Table 1 for the total number of Republican and Democratic mentions by issue). [Insert Table 5 Here] One caveat with regard to our measure of ownership is that it is potentially endogenous to our measure of dialogue. More specifically, we define ownership as variable for which we found a significant difference in the results. Values of 0 were recoded to 0.01 prior to the natural log transformation in order to prevent a loss in the number of cases. 21
occurring when one party dominates discussion of an issue across all campaigns, but that variable is used to predict dialogue, which is based upon the ratio of candidate mentions of an issue in a single race. Given the large number of races we examine, however, the correlations between dialogue in an individual race and dialogue across all races is relatively low (Pearson s correlation coefficient of -0.16). The salience of the issue in the general public should also be positively related to campaign dialogue. Issue salience is gauged by pooling National Election Studies (NES) 11 survey data from 1996, 1998 and 2000. Respondents were asked: What do you think are the most important problems facing this country? The NES coded their openended responses into almost 150 categories, which we subsequently collapsed into our more streamlined system of categorization. The salience variable, then, is the total number of mentions of each issue, weighted by year, and rescaled to range from 0 to 1. Summary statistics for this variable are reported in Table 3. 4. Results Table 6 reports results for a multilevel model estimated using OLS, random effects, and a multilevel model using MCMC simulation (columns 1, 2 and 3, respectively). 12 [Insert Table 6 Here] Table 6, column 1 reports the results for a standard OLS model. However, inferences are suspect due to the possibility of heterogeneity between campaigns within 11 These materials are based on work supported by the National Science Foundation under Grant Nos.: SBR-9707741, SBR-9317631, SES-9209410, SES-9009379, SES-8808361, SES-8341310, SES-8207580, and SOC77-08885. Any opinions, findings and conclusions or recommendations expressed in these materials are those of the author(s) and do not necessarily reflect those of the National Science Foundation, Warren E. Miller, and the National Election Studies. 12 Results are based on two chains each running for 5000 iterations, with a 5000-iteration burn-in. 22
issues, and between issues within campaigns. Consequently, we ran a random effects model twice. The first time we estimated the random effects model, we specified dialogue by campaign to be the unit of analysis and checked to see if there was heterogeneity within campaigns across issues. The result was an estimate of within-unit variance of 0, with less than a.001 probability of being different from 0 based upon a 2 test. In other words, somewhat to our surprise, dialogue within campaigns across issues seems consistent with the i.i.d. assumption (hence we do not report the estimates from this model). We next estimated the random effects model specifying dialogue by issue to be the unit of analysis, checking to see if there was heterogeneity within issues across campaigns. The result was an estimate of within unit variance of 0.044, which was highly significant based upon a 2 test (a 2 value of 6.38, p<.006). Since the random effects model corrects for the violation of the i.i.d. assumption, the results of this second random effects model are reported in Table 6, column 2. Without yet going into a detailed discussion of the results from OLS and random effects models, it is interesting to compare the coefficients and S.E.s of the two models for the issue-level predictors (the campaign-level coefficients and their S.E.s vary little between the two models, which is to be expected given that there does not appear to be heterogeneity within campaigns across issues). The coefficients on the Consensual Issue, Issue Salience and Owned Issue variables are all smaller in the random effects model. Furthermore, the standard errors for these three coefficients are all greater in the random effects model. Consequently, any inference from the OLS model is likely to be biased in the direction of increasing the probability of accepting a false positive. And, indeed, 23
Issue Salience would be deemed statistically significant at the p<0.05 level based upon the OLS results, but would not based upon the random effects model. However, even the random effects model assumes that campaign-specific predictors and issue-specific predictors are comparable units of analysis. Consequently, the random effects model uses an N of 982 rather than an N of 43 when calculating the S.E.s of the issue-specific parameters. This obviously biases the results in the direction of increasing the probability of accepting a false positive. The multilevel model avoids this by estimating the two regression equations simultaneously as specified in Section 3A above. We ran the model with 20,000 iterations after a burn-in of 19,000. All parameters converged fully based upon the Gelman-Rubin diagnostics. The results of the full, multilevel model are reported in Table 6, column 3. From the perspective of the theoretical model incorporating uncertainty, the key variable is Margin of Victory. The coefficient on this variable is statistically significant and in the negative direction. A one percent increase in the winner s margin of victory is associated on average with a 0.4 percentage point decrease in dialogue, controlling for all other factors. In other words, on average, the more lopsided a race s outcome, the less dialogue occurred during the race. Among the 65 Senate races included in this analysis, the mean margin of victory was 15.5 percent. Thus, the expected difference in dialogue between the average race and a neck-to-neck one was 6 percentage points. A one standard deviation increase in margin (11.6) would be associated with a decrease in dialogue of 4.6 percentage points. The greatest margin of victory was 41 percent, so the 24
variable s maximum effect is 16 percentage points (.004 * 41). Keep in mind that the mean level of dialogue was 22 percent. 13 Only one independent variable had a considerably larger substantive effect than margin of victory: the total number of issue mentions. The coefficient is statistically significant and in the predicted direction. An increase of 1000 mentions is associated on average with an 8.7 percentage point increase in dialogue, controlling for all other factors. Thus a standard deviation increase in Total Mentions of 1,650 mentions is associated with an increase in dialogue of a little over 14.3 percentage points. We attempted to assess empirically the hypotheses that the homogeneity of a state s population and the extent to which a race s campaigns went negative affected dialogue. The natural log of voting age population (in millions) had a statistically significant relationship with dialogue in the predicted direction. On average, controlling for all other variables in the model, dialogue decreases by almost 5.6 percentage points for every unit increase in the natural log of state voting age population. We can get a sense of the substantive effect that population size had on dialogue by noting the difference between the variable s minimum value, its mean and maximum values (approximately 1.5 and 3.8 respectively): hence campaigns in the nation s smallest state 13 This finding is consistent with Simon s empirical finding that competitiveness has a statistically significant and positive relationship to dialogue. However, Simon is agnostic on the predicted direction of the relationship between competitiveness and dialogue. Simon outlines two alternative explanations for a relationship between competitiveness and dialogue (neither of which are related to his formal model) one that predicts a negative relationship and one that predicts a positive relationship. The theory that predicts a negative relationship specifies that a candidate with little or no chance of winning might behave irrationally by talking about issues for which the opponent is advantaged. Third party candidates are the extreme example of this behavior. Alternatively, efforts to directly persuade voters may lead candidates to engage in dialogue. (Recall that in Simon s model talking only affects the salience of issues and is not meant to directly persuade voters to vote for a particular candidate.) 25
would be expected to have approximately 8.4 percent more dialogue than in the mean state, and 21.3 percent more dialogue than in the nation s largest state. The natural log of percent negative advertising also was statistically significant and in the expected direction (positive). A one unit increase in the log of percent negative advertising was associated with an increase of 1.7 percentage points in dialogue on average, controlling for all other variables in the model. Again, we can get a sense of the substantive effect that the tone of the campaign had on dialogue by noting the difference between the variable s minimum and maximum values (approximately 4 respectively): hence the most negative campaigns would be expected to have approximately 6.8 percentage points more dialogue on average than campaigns with the least mud-slinging. Total Spending and Open Race were not statistically significant and had relatively small coefficients. Upon reflection, it is not very surprising that Total Spending is not significant, since Total Spending itself is largely a function of the State Voting Age Population and a race s perceived competitiveness (i.e., Margin of Victory). The same can be said for Open Race in regard to Margin of Victory. It should be noted that the direction of the relationships between Total Spending and Open Race with dialogue were in the expected directions: positive and negative, respectively. We also controlled for year of the campaign. A likelihood ratio test based upon the results from the random effects model reported in Table 6, column 2 indicate that the two year dummies were jointly significant (a 2 value of 6.83, p<.033). 14 Senate campaigns in the off-years featured substantively more dialogue than those during 2000. 14 We use the likelihood ratio test based upon the maximum likelihood estimation of the random effects model because we do not think that Bayes factors provide a convincing basis upon which to distinguish between models estimated using MCMC techniques. 26