Herbert F. Weisberg Steven P. Nawara

Transcription

1 HOW SOPHISTICATION AFFECTED THE 2000 PRESIDENTIAL VOTE: TRADITIONAL SOPHISTICATION MEASURES VERSUS CONCEPTUALIZATION* Herbert F. Weisberg Steven P. Nawara The Ohio State University The definitive version of this article can be found in the December 2010 issue of Political Behavior (Volume 32, Issue 4, pp ) and can be accessed at *We appreciate the helpful comments of Paul Goren and Howard Lavine.

2 1 Abstract: This paper models the 2000 Presidential vote using voter sophistication as a source of heterogeneity. Three measures of sophistication are employed: education, knowledge, and the levels of conceptualization. Interacting them with vote predictors shows little meaningful variation. However, removing the assumption of ordinality from the levels of conceptualization uncovers considerable heterogeneity in the importance of the vote predictors in explaining the vote. Thus, different sophistication measures should not be treated as equivalent, nor combined as if they are equivalent. Few of the issue and candidate components are relevant to those with a less sophisticated understanding of politics. The opposite partisan attachments of the two most sophisticated groups suggest that sophistication s impact on the vote can be confounded by partisanship.

3 2 HOW SOPHISTICATION AFFECTED THE 2000 PRESIDENTIAL VOTE: TRADITIONAL SOPHISTICATION MEASURES VERSUS CONCEPTUALIZATION How does political sophistication affect voting behavior? As Lavine and Gschwend (2006) summarize, the prevailing literature contends that the capacity for ideological thought conditions how citizens perceive and evaluate objects in the political world. Whereas sophisticates are attuned to the abstract liberal/conservative character of political debate, nonsophisticates respond to political stimuli using simpler and more proximal (i.e., object-specific) considerations. Studies also routinely find that models predict the vote better for sophisticates (e.g., Stimson 1975; Macdonald, Rabinowitz, and Listhaug 1995). These generalizations, however, do not show how different forces impact the voting of people with different levels of sophistication in actual election settings. In an analysis of the 2000 U.S. presidential election, this paper shows that different matters influence voting at different sophistication levels depending on how sophistication is measured. As Sniderman, Brody, and Tetlock (1991, 21) state, political sophistication not only affects reasoning about political choices in its own right, but in addition affects how other factors affect reasoning about political choices. The initial studies of sophistication effects (Stimson 1975; Knight 1985) demonstrated that, as to be expected, liberal-conservative ideology has a greater effect at higher levels of sophistication. Later, Lavine and Gschwend (2006) showed that relative candidate evaluation (measured by the difference in thermometer ratings given to the two nominees) depends more on issues for ideological voters and more on perceptions of candidate preference for non-ideological voters. However, Pierce (1993) found that the effects of candidate traits on candidate preference did not vary much by sophistication, and Goren (2004) showed that people at different levels of sophistication structure their beliefs equivalently in the issue domains he studied. It is important to note that most of these studies looked at candidate

4 3 preference (Knight 1985; Lavine and Gschwend 2006; Pierce 1993) or belief structure (Goren 2004) rather than vote choice. Similarly, studies of the vote disagree as to whether sophistication matters: Sniderman, Glaser, and Griffin (1990) found that policy voting is more prevalent among the better educated, while Rahn, et al. (1990) determined that candidate assessment was similar across levels of sophistication, and, their vote model obtained similar coefficients for affect and partisanship in affecting the vote regardless of differences in political expertise. These disparate findings suggest that there is not a simple relationship between sophistication and voting and that the relationship will vary depending on what measures are employed. This paper uses American National Election Study (ANES) data to compare the importance of different factors in voting decisions in the 2000 U.S. presidential election according to the person s political sophistication. We find that education and political knowledge minimally impact the considerations that people brought to bear in voting, but that there were indeed important differences between sophisticates and non-sophisticates when using a more active measure of sophistication. We also find partisan differences between two groups of political sophisticates, with different effects on voting, suggesting that there are multiple routes to political sophistication that have different political implications. Measuring Sophistication Luskin (1987; see also Luskin 1990) differentiates several types of measures of political sophistication: consistency, abstraction, information holding, and composite measures. The several studies cited above differ in which version of sophistication they examine. For example, Sniderman, Glaser, and Griffin (1990) used education, whereas Rahn et al. (1990) used a composite measure of education, political knowledge, and interest in politics, which could explain some of the differences in their results. Knight (1985) and Pierce (1993) employed The

5 4 American Voter s levels of conceptualization (which we discuss in more detail below), and Goren used political knowledge. Stimson (1975) combined education with political knowledge, while Lavine and Gschwend (2006) measured the extent of ideological thinking by combining self-identification as a liberal or conservative, feeling close to that group, and identification with the corresponding party. While composite measures generally have advantages, it is easier to determine the sources of different findings when singular measures are employed. We shall employ three separate measures of sophistication in this paper. The first, education, is generally considered to tap abstract concept formation. However, education means different things depending on what the person studied there can be vast differences between social science majors, for example, and engineers or fine arts students in how relevant their education is to their conceptualization of politics. Many highly educated people are truly interested in politics and have highly sophisticated understandings of politics, but many people with low levels of education also follow politics closely and some people with high levels of education have no interest in politics. 1 The second, objective political knowledge, is included to capture another distinct aspect of sophistication, a factual understanding of the political realm. Presumably, advanced degrees and a distinguished educational background can only take an individual so far in understanding politics. Zaller defined political awareness as the extent to which an individual pays attention to politics and understands what he or she has encountered (1992, 21). At a certain point, one must actively acquire information and be able to use it to make informed decisions. Presumably, an objective knowledge scale would address this concept. The measurement of political knowledge is an admittedly difficult task. The five-point scale proposed by Delli Carpini and Keeter (1993) 1 For this paper, we divided educational experience into three categories: those with high school degrees or less, those with some college, and those with at least a bachelor s degree.

6 5 is the most popular knowledge measure, but too few of those questions were included in the 2000 ANES to use that scale. Instead, we employ the scale of objective knowledge questions previously used by Baum (2005) in his study of the 2000 election. 2 The third sophistication measure to be used in this paper is the levels of conceptualization (Campbell, Converse, Miller, and Stokes 1960; Converse 1964, Lewis-Beck, et al. 2008), which is directly based on how analytically the person discusses politics in response to survey questions. Respondents are classified into four main categories based on their answers as to what they like and dislike about the major parties and their presidential nominees. The highest level (Level A, Ideology ) consists of those who think about politics in ideological terms, whether they actually use terms like liberal and conservative or give answers that use those concepts without mentioning the terms per se. Level B ( Group Benefits ) involves talking about politics in group terms, which can be seen as near-ideology since so much ideology is really group conflict. Examples of Level B responses include when individuals discuss the parties and candidates in terms of how they relate to unions, African-Americans, or the upper class. Level C ( Nature of the Times ) either focuses on a single issue, without making reference to any overarching concepts, or just speaks in terms of conditions being good or bad. A prototypical Level C ( Nature of the Times ) respondent would indicate a preference for an incumbent during good economic times. Finally, the bottom level (Level D, No Issue Content ) makes no issue references, either talking only about personality aspects of the candidates or making no comments at all about any of the parties or candidates. This level includes those who justified their opinions based solely on partisanship. For reference, Lewis-Beck et al. s (2008) analysis 2 This ten-point additive scale includes the identification of Trent Lott, Tony Blair, William Rehnquist, and Janet Reno, the identification of the major presidential and vice presidential candidates home states (with multiple choice options), and knowledge of which party controlled both the House and Senate before the election (mean = 3.8, std. dev = 2.6).

7 6 placed 20 percent of the 2000 ANES sample in Level A, 28 percent in Level B, 28 percent in Level C, and 24 percent in Level D. Eric Smith (1980; 1989) challenged the levels of conceptualization measure after he found considerable variation among individuals in their level across different waves of a panel survey. He attributed this movement to massive amounts of measurement error, thus putting the reliability and validity of the levels in doubt. Smith concluded that the ideologies measured by the levels of conceptualization do not particularly involve elaborate belief systems. However, Luskin (1987, 879) argued that the levels are very stable the changes in them over time are simply measurement error uncorrelated with other variables. Nie, Verba, and Petrocik (1981) pointed out that Smith s findings could be due to his reliance on the master codes rather than determining a respondent s level of conceptualization by analyzing the interview transcripts of open-ended questions as The American Voter did. Hagner and Pierce (1982) found that the characteristics of those residing within each level do not change across time, which attests to their validity. Cassel (1984) used alternative tests of validity based on cross-sectional correlations between the levels and theoretically-related variables and concluded that the levels of conceptualization are valid measures of ideological thinking. Thus, the preponderance of the literature accepts the levels as a valid measure. There are two approaches to coding the levels of conceptualization. The National Election Studies presidential year datasets code the answers to the party and candidate likes-anddislikes questions into a large number of discrete master codes. Some studies classify these codes into the levels of conceptualization. That is much easier than categorizing the person s actual answers into the levels of conceptualization, but it is also less valid. The intensive coding required to compute the levels of conceptualization from interview transcripts is why the levels

8 7 of conceptualization are not available for many recent elections. The 2000 election between Bush and Gore, however, is an exception. Following the approach of The American Voter and Knight (1985), Jacoby, Duff, and Pyle (2008) categorized respondents actual answers directly for The American Voter Revisited (Lewis-Beck, Jacoby, Norpoth, and Weisberg 2008), and we use their coding in this study. In this paper, the use of education and knowledge, along with the levels of conceptualization, provides an examination of whether results are robust across separate sophistication measures. It is clear that the levels of conceptualization are theoretically and practically distinct from other measures of sophistication such as education, political interest, and political knowledge. Correlations between the levels and these measures were modest in the 2000 ANES survey, ranging from 0.2 to 0.33, indicating that the levels should not be considered a surrogate for these variables. Instead, the levels should be thought of as an active use measure of sophistication (Converse 1964) that captures how the complexity of one s belief systems are publicly expressed in conversation; that is, how one s sophistication is put into practice. Some may consider this aspect of the measure as a weakness, however. If individuals absorb campaign rhetoric and merely regurgitate it back during the open-ended portion of the interview regardless of their true sophistication, the measure would be invalid. Fortunately, this is not the case. Zaller (1992) argues that the more sophisticated will be more likely to receive campaign messages, but also will be less likely to accept them. If this is true, those in the higher levels of conceptualization will be expressing true indications of ideological thinking while those in the lower levels would be unlikely to verbally parrot campaign rhetoric due to a lack of message reception.

9 8 Some previous studies have treated the levels as a fully ordered set of categories while others treat them as nominal categories. We examine both possibilities, but only find evidence for the latter. We consider Levels C and D to be non-sophisticates based on their narrower understanding of politics and Levels A and B to be sophisticates due to their broader political thought. 3 Furthermore, the top two levels reflect different approaches to ideological thinking. Ideologues think about politics in an abstract manner, while those in the group benefits category instead think about politics in more concrete group conflict terms. Group benefits may be thought of as providing people who don t have college degrees a mode of sophisticated thinking about politics even if they have not had the opportunity to develop their abstraction abilities (Hagner and Pierce 1982). Additionally, Pierce and Hagner (1982) have found that these top two levels have different partisan composition, with Republicans being more likely to be placed in Level A and Democrats more likely to be in Level B, which suggests that different factors might affect voting decisions for these two groups. The Components of the Vote Decision Model In our analysis of how sophistication affects voting decisions, we use the components of the vote decision model employed in The American Voter (1960; see also Stokes, Campbell, and Miller 1958; Stokes 1966; Lewis-Beck et al. 2008). This model classifies people s answers to the open-ended party and candidate likes and dislikes questions into six basic categories: comments about the Republican candidate, the Democratic candidate, domestic policy, foreign policy, social groups, and parties as managers of government. We use the coding of these questions by Weisberg and McAdams (2009) for the 2000 NES survey. To measure each component, the number of pro-republican (and anti-democratic) comments related to that topic 3 This is in contrast to Knight s (1985) finding that Levels B and C are qualitatively different from Level A on the one hand and Level D on the other in the extent to which ideological concerns are important to their vote.

10 9 (e.g. foreign policy) that the person makes in response to the open-ended questions is added up, as is the number of pro-democratic (and anti-republican) comments on that same topic, and the difference is computed. The results show net how pro-republican the person is on each of the six components. Presidential vote is then regressed on these six components using logistic regression, with a control for the standard seven-point party identification scale. The regression coefficients show how much each component affected the vote decision. The means on each component indicate how much that component favored the Republicans across the electorate. Therefore, to gauge each component s impact in the electorate, we can calculate the difference between the predicted probability of voting for Bush when the components are at their neutral point and when one component is at its mean. 4 Previous studies (Stokes 1966; Kagay and Caldeira 1980; Pomper 1975; A. Miller and Miller 1976; Popkin 1976; A. Miller and Wattenberg 1981; Lewis-Beck et al. 2008) have used the vote components model to measure the relative importance of different factors in election outcomes, but they have looked at voting across the entire electorate without taking into account its heterogeneity. It is unrealistic, however, to assume the weights for different factors are the same for everyone. As Rivers (1988, 753) argues, when models of the average decision maker are applied to heterogeneous populations, estimates requiring a homogeneity assumption will not, in general, accurately describe average behavior. Therefore, it is important to look for differences in voter decision-making across different degrees of political sophistication. Different types of people, particularly those at different degrees of sophistication, should be expected to emphasize different components. First, gathering and processing information on political issues requires an increased ability to think about politics in the abstract, so it goes to 4 As Rahn, Krosnick, and Breuning (1994) show, responses to the open-ended likes and dislikes questions may be rationalizations of prior evaluations of the candidates, which suggests that any differences found between voters of degrees of sophistication would indicate differences in how they are able to rationalize those evaluations.

11 10 follow that those in the higher levels of education, knowledge, and conceptualization are more likely to rely on issues when voting. Therefore, following Lavine and Gschwend (2006), we would expect that people who conceptualize politics in ideological terms are more likely to use the two issue components than are other voters. Secondly, since individuals placed in Level B ( Group Benefits ) characterize their political ideology based on which social groups benefit or are hurt by particular candidates or parties, we expect those in Level B to differ from the other levels with regards to the social group component. As a result, people who conceptualize politics in group terms are expected to be more likely to emphasize the social group component. 5 Third, we expect the votes of those with less education, knowledge, and ideological sophistication to be affected mainly by their opinions of the candidates themselves because such opinions are easier to form than are issue opinions. Consequently, again following Lavine and Gschwend (2006), people at the bottom level are expected to be more likely to use the candidate components. Finally, in accord with Stimson (1975), the votes of the low sophistication group are expected to be explained less by the component model because they are less likely to employ the full range of components in their voting decisions. We test these hypotheses in the following sections. We will first analyze how the importance of the six components varies across sophistication levels, using education, knowledge, and conceptualization as measures. After we show the limited impact of education and sophistication measures, we then turn to the levels of conceptualization for expanded analysis. Removing the assumption of the levels ordinality, we uncover rich heterogeneity 5 While the levels and the vote components are coded off the same open-ended comments, there is not necessarily a relationship between group benefits voters and the social group component. Group benefits voters can also mention domestic policies and the candidates in their open-ended comments, and those components could be found more important in the statistical analysis than the social group component. Conversely, people who comment about the social groups can also make ideological comments, so that the social group component could be found to be more important for ideological voters than for level B.

12 11 across the levels. We follow the analysis of The American Voter, elaborating on the importance of the components on the vote by examining how far each component moved the overall election results within each level of conceptualization. Therefore, we will not only look at the direction that each component pulled the overall vote, but how far it did as well. Our findings indicate that while voters of varying conceptualization rely on widely differing vote components, stratification by education or knowledge provides little additional insight. Importantly, the same vote component can have opposite directional effects on the overall vote across the levels of conceptualization. Therefore, not only are vote influences affected by heterogeneity, but the indicator of sophistication used in the analysis can also greatly affect the results as well. How Increased Sophistication Affects Voting Considerations We begin by regressing the 2000 major-party presidential vote on the six vote components derived from the open-ended responses, with party identification controlled, while also interacting individuals sophistication with each component. 6 The resulting coefficients provide insight into how important each factor is in contributing to the vote. Table 1 displays the results of these logistic regressions, with a vote for Bush as the dependent variable, which explains the negative coefficients for the Democratic candidate component. [Table 1 about here] Not surprisingly, party identification is correlated with the vote in all three models. Attitudes toward Al Gore are also related to the vote in all three, though attitudes toward George Bush do not have an independent effect in any of these models. Domestic policy is the only other 6 In analyzing the 2000 ANES data, we use the post-election weights, since the dependent variable is a post-election measure. These weights are the product of the household non-response adjustment factor, the within-household selection weight (number of eligible persons), and a post-stratification adjustment factor by age and education: v000002a. We use weights in the regression since we do not include demographic characteristics as control variables.

13 12 component consistently related to the vote, while foreign policy and social groups are significant in some models but not others. People scoring higher on the knowledge measure were more likely to vote against Bush, while education and conceptualization level do not affect the vote. The more crucial question is whether the interactions of sophistication with the components are significant. Model 1 shows that people with diverse amounts of education do not vary in how much they emphasize the different vote components (Boehmke 2008). This finding is contrary to Sniderman, Glaser, and Griffin (1990), who found issue importance to vary across educational levels; using education fails to reveal variation in the vote component model for those of varying sophistication levels. Model 2 replaces education as the indicator of sophistication with political knowledge. The effects of three components vary across knowledge levels: comments regarding political parties as managers of government become more important at higher knowledge levels, while comments about the Democratic candidate and social groups become less important as knowledge rises. While one would expect that people with greater political knowledge would be more likely to vote on the basis of policy concerns, the interactions between knowledge and the policy components are not significant. Thus, knowledge affects the importance of some components to voting, but not the issue components that political knowledge should be expected to affect the most. Finally, when the respondents level of conceptualization is used as the sophistication measure in Model 3, none of the interactions between the components and the levels of sophistication are statistically significant, making the levels of conceptualization s performance very similar to education s.

14 13 Table 1, taken as a whole, is perhaps more interesting for what it does not show as for what it does show. The impact of sophistication on the vote component analysis appears to be less than anticipated. Aside from three statistically significant relationships found in Model 2, there appears to be no effect of sophistication on the vote components importance in predicting the vote. Most importantly, from a normative standpoint, policy-related concerns of lowsophisticates are not any less related to the vote than those concerns of high-sophisticates, certainly a counter-intuitive finding. How Conceptualization Affects Voting Considerations All together, these results are so bland that it is appropriate to check whether these measures of sophistication are being properly analyzed, especially the implicit assumption that the levels of conceptualization are an ordered measure. The preceding regression analysis treated the levels as if there was an order to them, with people in Level A (Ideology) being presumed to be more sophisticated than those in Level B (Group Benefits), and so on. However, removal of this assumption may yield more interesting results. Smith (1980, 1989) and Cassel (1984) essentially treat the levels of conceptualization as ordinal, but generally authors (e.g., Hagner and Pierce 1982; Pierce and Hagner 1982; Pierce 1993) treat them as unordered nominal categories following the pattern of The American Voter (Campbell, et al. 1960). For instance, one might hypothesize that those in Levels A and B are largely equal in their objective levels of sophistication, while approaching politics through different means, with Ideologues (Level A) relying on abstraction and those in Group Benefits (Level B) looking at the tangible effects of candidates and policies on social groups. Additionally, ordinality implies a monotonicity of vote component effects, while in actuality, one component may be very important to those in a particular level of conceptualization, but not to the others. This would cause the component

15 14 interactions in Table 1 to obscure heterogeneity across the levels. Therefore, to investigate the possibility of nonordinality within the levels of conceptualization, we now treat the levels as separate nominal categories. Table 2 presents the logistic regressions of the vote component model, broken up by the levels of conceptualization. Separating the sample by levels of conceptualization shows how applying the model to the entire electorate masks the great deal of heterogeneity within the electorate, supporting our hypothesis that the vote will be affected by different factors for those of varying degrees of sophistication. When all levels are combined (Column 1), all of the components are statistically significant, however, when the levels are analyzed individually, a great deal of variation is uncovered. As expected, most components are statistically significant among the two highest levels of conceptualization and both policy components are significant within Level A (Ideology). However, only two components are significantly related to the vote in Level C (Nature of the Times), with domestic policy being significant for those voters because of the importance of current economic conditions, but not foreign policy since foreign issues did not play a major role in the 2000 presidential campaign. Neither policy-related component is significant in Level D (No Issue Content). Indeed, the percentage of correctly predicted votes drops off dramatically in Level D, which suggests that the least-sophisticated voters cast their ballots in a manner different than the population as a whole, perhaps due to increased randomness in their behavior and in their responses to survey questions. 7 [Table 2 about here] Analysis of the individual vote components across the levels provides further insight. The Democratic candidate component is significant for all levels, but comments regarding President 7 Adding controls on such demographics as gender, age, education, and race has minimal effect on the equations in Table 2, with there being no consistent pattern to the occasional significance of one demographic variable for one conceptualization level.

16 15 Bush are significantly related to the vote only at Levels A and B. Domestic policy had significant effects on the vote for every level except Level D. The significant overall relationship between the foreign policy component and the vote is completely driven by Level A respondents; in fact few respondents made foreign policy-related comments, particularly those at the lower levels. 8 There is a clear nonmonotonicity of effects for the remaining two components. The social group component and the parties as managers of government are only significant among those in Level B. 9 Interestingly, group-related comments were also not related to the vote within Level A, indicating that these abstract thinkers do not vote on the basis of social groups. Furthermore, there was significant variation in the partisan direction of the respondents comments regarding social groups, an important phenomenon which will be described below. Aggregate Vote Analysis Ideally, one would like to go beyond the importance of the components in predicting individual votes and examine the impact of each component on the aggregate vote totals. As Campbell et al. (1960) explain, in order for a component to influence the election, it is not enough for many people to care about it; the component must also favor one candidate over the other. The American Voter used Table 2 s linear regression equivalent and multiplied its coefficients by each component s mean to gauge the aggregate effect each component had on the overall vote. Since our analysis employs logistic regression, we use the equations in Table 2 to estimate the aggregate component effects as changes in predicted probability for both the overall vote and within each level of conceptualization. For the following analysis, each component is set at zero, as if the respondents net comments did not favor one side over another, and party 8 Just one of the 157 respondents in Level D made a (single) comment about foreign policy, so that component lacks sufficient variance to be included in the regression equation. 9 Comments regarding the parties as managers of government are generally highly partisan in nature, so their potential effect is likely absorbed by the control on party identification.

17 16 identification is held constant at pure Independent. We then compare the predicted probability of voting for Bush with all the components set at zero to the predicted probability after varying one component to its sample mean, thus providing us with an estimate of the directional impact of that component for pure Independents. 10 Whereas the logistic regression coefficients of Table 2 estimate the effect each component had on the predicted vote of individuals, the predicted probability differences of Table 3 estimate the aggregate directional effect of each component. This distinction is important because a component that is highly predictive of an individual s vote would not have a sizable aggregate effect if the electorate s opinions on that component were evenly divided. The component means are reported in parentheses within Table 3. Note that the component means vary significantly across the levels for all six components. The sharpest differences are between those who conceptualize politics in terms of Ideology, who favor the Republicans on most of the six components, and those who instead view politics through the lens of Group Benefits, who favor the Democrats on most of the components. This advances Pierce and Hagner s (1982) finding that Republicans are more likely to be placed in Level A and Democrats more likely to be in Level B, showing that Level A and Level B respondents differ in partisan usage. The means for the Nature of the Times voters and the No Issue Content voters are generally smaller in magnitude than those for Levels A and B, showing that the Level C and D voters are more balanced in their views of the two parties. [Table 3 About Here] 10 We appreciate the advice of Luke Keele, who suggested this approach for estimating the impact of the individual components on the vote within the logit framework. It is analogous to Stokes b*xbar approach for regression in that it shows zero impact when the coefficient is zero and when the component mean is at the neutral zero point. The predicted probabilities were calculated using CLARIFY: Software for Interpreting and Presenting Statistical Results, Version 2.0 (Tomz, Wittenberg, and King 2001; King, Tomz, and Wittenberg, 2000).

18 17 The numbers above the means in Table 3 show these directional effects for each component within each level of conceptualization, as well as for the entire sample. Positive percentages indicate the component's benefit to the Republican candidate, while negative percentages show its benefit to the Democrat. 11 Just a glance at Table 3 shows the very disparate impact the vote components have on Republican vote share conditional upon voter sophistication, cautioning us against the treatment of the voters as uniform as well as against the treatment of the levels of conceptualization as ordinal. The clearest effect found in Table 3 is that the components of those in Level B are often radically different than those in other levels in that they strongly favor Gore. As a result, the Group Benefits level voted overwhelmingly in Gore s favor (75 percent) while the other three levels favored Bush, a major difference that would be missed by most conventional vote analyses. In particular, Levels A and B have very different partisan dynamics, showing that Level B voters do not fall between Levels A and C in an ordinal fashion. The differences between Levels A and B are especially large on the candidate components. Opinions regarding Governor Bush shifted the predicted vote in his favor 5.2 percentage points among the ideologues, but reduced his vote by 6.4 points among the group benefits voters. And the opposite pattern is evident as regards the effect of Al Gore s candidacy: opinions about him shifted the predicted vote of ideologues against him by 5.8 percentage points while increasing the vote for him by 11.1 percentage points among group benefits voters. There is an even sharper difference between Levels A and B with respect to the domestic issues component, with a 26.7 percentage point difference in its impact for the two sets of voters. 11 Because pure Independents are least constrained by partisanship, these estimates of impact should be regarded as higher than would be obtained for other partisan categories. While we choose to analyze these predicted probability changes for Independents, some may prefer to drop party identification from the analysis to consider the components without having to specify partisanship when calculating the predicted probabilities. The appendix provides these estimates from a logistic regression analysis in which partisanship was not controlled.

19 18 Levels A and B may conceptualize politics in broader terms than the rest of the electorate, but they see politics quite differently from one another. Earlier we showed that the impact of the issue components is statistically insignificant among the least-sophisticated. Here, we find that the partisan impact of the issue components is concentrated among Levels A and B. The magnitude of the directional effects of domestic policy attitudes on the vote drops off sharply for Levels C and D, while foreign policy s directional effect is only evident among those in Level A. 12 The parties as managers of government component only moved the vote slightly, with the only statistically significant effect being a slight loss for Bush among Level B voters. The group component hurt Bush s vote among all voters, but this relationship is driven mainly by the 23 point shift of the vote against him within the Group Benefits level (B). Ideologues also made social group comments, but those comments had little net impact on their votes. Looking at the different levels in turn provides important perspectives on the differences between them. The votes of Level A and B respondents are impacted by domestic policy and by the combined effects of the two candidate components. Foreign policy also has an impact for Level A, but a small impact, corresponding to the fact that this election did not turn on foreign policy issues. Not surprisingly, social groups have the greatest impact on the votes (-23.3) for Level B, but domestic policy and the combined impact of candidates were also substantial. The real contrast is between the several factors that impact the vote at Levels A and B compared to the few factors with an impact at Levels C and D. Negative reactions to Al Gore increased the probability of Level C voting Republican by 4.1 points, while domestic policy decreased the probability of them voting Republican by 1.3. The only significant component for 12 The differences between domestic and foreign policy components also reflect the number of policy comments made by respondents. Domestic policy was the subject of the most comments, with the average person giving 3.8 responses, while the average number of foreign policy responses was just 0.26.

20 19 Level D is reactions to the Democratic candidate, but its impact on their vote is truly negligible (0.1) because the aggregate numbers of comments about Gore were generally more balanced among respondents in Level D than in other levels, cancelling out their effects on the vote. Attitudes toward the Republican candidate effectively helped Bush, though the effect is not significant. Thus, individuals of low sophistication rely primarily on their perceptions of the candidates when casting a ballot, but their net effect is relatively mild. Summary and Conclusions Luskin (1987) describes political sophistication as an application of the broader concept of cognitive complexity. People would be judged to be politically sophisticated if their political belief systems are large, wide-ranging, and highly constrained (860). Sniderman, Brody, and Tetlock (1991, 21) describe political sophistication as a bundle concept one that packs together related, if distinguishable, properties including a tendency to pay close attention to politics, to have ready at hand banks of information about it, to understand multiple arguments for and against particular issue positions, and to recognize interrelations among those arguments. Researchers typically either have chosen one measure from this bundle because of its convenience, as when Sniderman, Brody, and Tetlock (1991, 21) justify their use of education as a proxy because it covaries with other aspects of sophistication but also because it is both well measured and invariably measured, or they have combined multiple measures, assuming that a composite measure presumably because of the standard argument that a composite measure has better reliability than a single measure would. However our tests show different measures can have very different effects, so that combining them would lead to a jumble in which the true causal forces would not be clear.

21 20 Education does not have any effect on the 2000 presidential vote, and does not condition the effects of other vote predictors. Countless studies have shown education to perform well as a proxy for sophistication, but apparently it is not related to politics enough to affect voting directly. This result is reminiscent of Luskin s (1990) finding that education was not directly related to sophistication when intelligence and other relevant variables are controlled. Political knowledge has a direct effect and also conditions how some predictors affect voting, but it does not affect policy voting, which is what it would be most expected to affect. The levels of conceptualization do not have an effect and do not condition the effects of other vote predictors when treated as an ordered variable, but treating them as nominal categories through separate regressions shows that vote determinants do vary according to how one conceptualizes politics and this is why the levels of conceptualization are a particularly useful measure: they show that people who think about politics in different ways will emphasize different forces in their voting decisions. Thus, our evidence does suggest that the concept of sophistication can be encapsulated in a variety of different forms. Measures like education and knowledge capture bookish aspects of sophistication, which are the result of individuals life experiences, but such measures do not adequately capture a complete notion of practical sophistication. The levels of conceptualization, however, act as a measure of individuals sophistication when they are called to actively express political opinions. Our results show that this aspect of sophistication is truly distinct from the more commonly-used measures of education and knowledge; this suggests that it is important to employ active-use measures of sophistication. Accordingly, this paper will hopefully encourage others to code the levels of conceptualization in future voting studies, or perhaps even inspire the creation of an active-use sophistication variable that requires less

22 21 rigorous work. These results also imply that the common practice of combining different measures of sophistication could blur the very different effects of different measures. Our results confirm Lavine and Gschwend s (2006) generalization that sophisticates evaluate the political world in more abstract terms, while non-sophisticates employ simpler considerations. In this study, this means that the sophisticates utilize a full range of candidate and issue evaluations while the least-sophisticated rely at most on the candidate evaluations. This conclusion differs markedly from Rahn, et al. (1990), which found that differences in political expertise mattered little in their vote models, where they measured expertise in terms of education, political knowledge, and interest in politics. Contrary to Sniderman, Glaser, and Griffin (1990), we do not find important differences when we measure sophistication by education, but we find significant differences with regards to the effects of different vote determinants when we instead employ the active use measure of how people talk about politics, the levels of conceptualization. The sharp partisan difference between Levels A and B, as well as their partisan composition as compared to Levels C and D, suggests an important more general point: analyses of sophistication s impact on voting can easily be confounded by partisan differences. If, for example, people with higher levels of education and/or knowledge tend to be more Republican than people with lower levels, then finding differences in voting determinants between sophistication levels may reflect those partisan differences as much as sophistication differences. Traditionally there has been a difference in partisanship between those with college degrees and those with at most high school degrees, but this difference is diminishing, which might explain the lack of effect of education on vote determinants in Table 1.

23 22 The candidate and issue factors significantly affect the vote of people at the two highest levels of conceptualization, but we find that the dynamics of those two levels differ dramatically. Some analysts over the years have regarded viewing politics in group-benefits terms as a form of ideology. After all, Karl Marx, who was the epitome of an ideologue, viewed politics in terms of group conflict. But if thinking about politics in group-benefits terms is ideological, it has very different partisan implications than thinking about politics as more directly ideological. On all six components, ideologues were more pro-republican than group benefits voters. Indeed, on most components, the ideologues were pro-republican and the group benefits respondents were pro-democrat. As a result, when these two sophistication levels are combined, the Republican advantage among ideologues is pretty much neutralized by the Democratic advantage among group benefits voters. The special affinity of group benefits voters with the Democrats fits with that party often being considered to be a coalition of different social groups. By contrast, people who discuss politics in ideological terms are more Republican (Devine 2010), which fits with the Republican voting slant of Level A voters. The American Voter describes the group-benefits level as ideology by proxy, lacking the abstract reasoning of ideologues while partak[ing] of ideology by endorsing a [group] leadership that has ideology (Campbell, et al., 1960, p. 220). Yet we see here that these two paths to ideology can lead to different political results. Pro-Republican conservatism predominated in 2000 among the more abstract ideologue types, while pro-democratic group identification predominated among the less abstract group-benefits types. Social psychologists might not be willing to consider the latter to be experts, but their view of politics is richer than those who view politics only in terms either of single issues or candidate conflict. Further, we see that the votes of Group Benefits respondents are not exclusively driven by the social groups

24 23 they mention; the candidates and domestic policy also affect their voting. Lau (1991) claimed that the levels of conceptualization differ from schemata in that a person can have multiple schemas while that person is categorized into a single conceptualization level. Similarly, we see here that, properly studied, multiple factors can influence the votes of people in a single conceptualization category, especially for the more sophisticated half of the electorate. Why was the 2000 election so close that the disputed votes in one state could decide the Electoral College? If we looked at the total results, we would focus on how the social group factor helped Al Gore while all the other factors helped George Bush. If instead we look at the different segments of the electorate, the story changes: Bush s gain on domestic policy and the candidate factor among those in Ideology was fairly well neutralized by Gore s advantage on social groups, domestic policy, and the candidate factor among Group Benefits voters, while the candidate factors pulled the Nature of the Times and No Issue Content voters slightly toward the Bush side. The net advantage was to Gore because of the Group Benefits voter, which gave him the edge in popular vote, but not enough to ensure his Electoral College victory when the other levels ran in Bush s favor and when a recount was stopped in midstream. Our results are based on the 2000 election, but we would expect fairly similar results for other elections. The partisan differences we find between Level A and B voters elaborate on the differences found by Pierce and Hagner (1982). The lack of effect of education suggests its diminished role as compared to the results of Sniderman, Glaser, and Griffin (1990) for The other large difference we would anticipate has to do with the role of foreign issues. Foreign issues were remarkably irrelevant in the 2000 election, but they have been much more important in some other presidential elections (Aldrich, Sullivan, and Borgida 1989). We therefore look forward to being able to do a comparable analysis for the 2004 election, to test whether foreign

25 24 issues can affect the votes of the less sophisticated part of the electorate during wartime. Yet we expect our basic results will hold: different measures of sophistication are not equivalent, and, depending on how sophistication is measured, voters at different levels of sophistication naturally vote on the basis of different factors. This adds a complication to our understanding of electoral behavior, but it is an appropriate complication.

26 25 References Aldrich, John H., Sullivan, John L., and Borgida, Eugene. (1989). Foreign affairs and issue voting: Do presidential candidates waltz before a blind audience? American Political Science Review 83: Baum, Matthew A. (2005). Talking the vote: Why presidential candidates hit the talk show circuit. American Journal of Political Science 49(2): Boehmke, Frederick J. (2008). GRINTER: A STATA utility to graph interaction effects after regression models. (Version 1.5). Campbell, Angus, Converse, Philip E., Miller, Warren E., and Stokes, Donald E. (1960). The American voter. New York: Wiley. Cassel, Carol A. (1984). Issues in measurement: The levels of conceptualization index of ideological sophistication. American Journal of Political Science 28(2): Converse, Philip E. (1964). The nature of belief systems in mass publics. In David Apter (ed.), Ideology and discontent. New York: Free Press, Delli Carpini, Michael X. and Keeter, Scott. (1993). Measuring political knowledge: Putting first things first. American Journal of Political Science 37(4): Devine, Christopher J. (2010). Why liberals don t call themselves liberals: The effects of elite communication on ideological comprehension and identification. Paper presented at the Annual Meeting of the Midwest Political Science Association, Chicago, IL. Goren, Paul. (2004). Political sophistication and policy reasoning. American Journal of Political Science 48(3): Hagner, Paul R. and Pierce, John C. (1982). Correlative characteristics of the levels of conceptualization. The Journal of Politics 44(3):