Public Opinion Quarterly, Vol. 73, No. 4, Winter 2009, pp.

679703

THE NATURE OF POLITICAL IDEOLOGY IN THE CONTEMPORARY ELECTORATE

SHAWN TREIER D. SUNSHINE HILLYGUS

Abstract Given the increasingly polarized nature of American politics, renewed attention has been focused on the ideological nature of the mass public. Using Bayesian Item Response Theory (IRT), we examine the contemporary contours of policy attitudes as they relate to ideological identity and we consider the implications for the way scholars conceptualize, measure, and use political ideology in empirical research. Although political rhetoric today is clearly organized by a single ideological dimension, we nd that the belief systems of the mass public remain multidimensional, with many in the electorate holding liberal preferences on one dimension and conservative preferences on another. These crosspressured individuals tend to self-identify as moderate (or say Dont Know) in response to the standard liberal-conservative scale, thereby jeopardizing the validity of this commonly used measure. Our analysis further shows that failing to account for the multidimensional nature of ideological preferences can produce inaccurate predictions about the voting behavior of the American public. There appears to be a consensus among scholars and political observers that U.S. political elites have grown more polarized in recent decades. Democrats and Republicans in Congress more consistently oppose each other on legislation (McCarty, Poole, and Rosenthal 2006), the party platforms are more ideologically extreme (Layman 1999), and issue activists are more committed to one political party or the other (Stone 1991). In contemporary American

SHAWN TREIER

is with the University of Minnesota, 1414 Social Sciences Building, 267 19th Ave S, Minneapolis, MN 55455, USA. D . SUNSHINE HILLYGUS is with Duke University, 409 Perkins Library, Box 90204, Durham, NC 27708-0204, USA. The authors would like to thank Chris Federico, Paul Goren, Dean Lacy, Andrew Martin, Caroline Tolbert, Claudine Gay, Dan Carpenter, Kevin Quinn, Eric Schickler, the participants of the University of Minnesota American Politics Pro-seminar, three anonymous reviewers, and editor Jamie Druckman for helpful feedback and suggestions. Address correspondence to D. Sunshine Hillygus; e-mail: [email protected].

doi:10.1093/poq/nfp067 Advance Access publication December 4, 2009 C The Author 2009. Published by Oxford University Press on behalf of the American Association for Public Opinion Research. All rights reserved. For permissions, please e-mail: [email protected]

680

Treier and Hillygus

politics, Republican politicians consistently line up on the conservative side of an issue while Democratic politicians consistently line up on the liberal side, across different policy domains. With just the liberal or conservative label, then, we can quite accurately predict a politicians stance on policy issues as disparate as taxes, health care, or abortion. Put another way, the belief systems of political elites in the United States today are captured with a single dimension of ideology. Can the policy preferences of the American public be similarly characterized? Although this question has been the subject of considerable research over the years (e.g., Marcus, Tabb, and Sullivan 1974; Conover and Feldman 1984; Jacoby 1991), it takes on a new prominence given the recent polarization debate. Some scholars have argued that the sharpening of policy differences between political elites in recent decades has increased ideological identication and polarization in the public as well (Abramowitz and Saunders 1998). In contrast, others argue that the majority of the public remain moderate on most policy issues even as elected representatives have grown further apart (Dimaggio, Evans, and Bryson 1996; Fiorina 2004). Morris Fiorina concludes that the great mass of American people are for the most part moderate in their views and tolerant in their manner. . . it is not voters who have polarized, but the candidates they are asked to choose between (Fiorina 2004, pp. 8, 49). Given the increasing salience of political ideology in American politics, it seems important to examine how ideology is conceptualized by the public relative to how it is operationalized and measured by researchers. Both sides of the polarization debate seem to assume that the ideological labels people use are a meaningful representation of their public policy preferencesan assumption once challenged by early public opinion research that concluded the public was incapable of ideological thinking (Converse 1964). The generalizations that scholars make about the behavior, attitudes, or thinking of the American electorate could be wholly inaccurate if the liberal-conservative continuum so often used in empirical analysis is an inadequate measure of policy preferences. In this article, we examine the contemporary contours of policy attitudes as they relate to ideological identity. And we consider the implications for the way scholars conceptualize, measure, and use political ideology in theories and models of political behavior. Although political rhetoric today is clearly organized by a single ideological dimension, we nd that the belief systems of the mass public are multidimensional. Using Bayesian Item Response Theory (IRT), a methodological approach with unique advantages over previous estimates of ideological preferences, we show that many in the electorate hold liberal preferences on one dimension and conservative preferences on another. These cross-pressured individuals tend to self-identify as moderate (or say Dont Know) in response to the standard liberal-conservative scale, raising questions about the validity of this commonly used measure and undermining characterizations of the American public as either policy centrist or ideologically innocent.

The Nature of Political Ideology

681

40% 35% 30% 25% 20% 15% 10% 5% 0% 1972 1976 1980
Slightly Liberal

1984

1988
Dont Know

1992
Slightly Conservative

1996

2000

Liberal

Moderate

Conservative

Figure 1. Political Ideology, American National Election Study Cumulative File.

Measuring Ideology
In characterizing the ideological preferences of the American public, scholars typically rely on a survey question asking respondents to place themselves on a liberal-conservative continuum. The question is typically of the following sort: When it comes to politics, do you usually think of yourself as extremely liberal, liberal, slightly liberal, moderate or middle of the road, slightly conservative, conservative, extremely conservative, or havent you thought much about this?1 As Fiorina (2004) and others have pointed out, only a small percentage of Americans consider themselves to be extreme ideologues, either liberal or conservative. As shown in gure 1 the plurality of Americans characterize themselves as ideologically moderate or say Dont Know.2 Just as Americans like to consider themselves middle class, individuals like to think of themselves as ideologically moderate. The trend line does suggest that Americans today are better able to place themselves (as evidenced by the decline
1. This is the standard question asked since 1972 in the National Election Study (and often replicated in other surveys). 2. The extreme categories are collapsed with the respective conservative and liberal categories. In 2000, for instance, 2 percent of respondents called themselves extremely liberal and 3 percent extremely conservative.

682

Treier and Hillygus

in DK responses), but we also see a corresponding increase in the percentage identifying as moderate. Does this imply that the public is necessarily centrist in their policy preferences? Or is this a reection of the ideological innocence of the electorate? In other words, what exactly do the ideological labels mean? Classic public opinion research treated the high levels of DK and moderate responses as evidence that the public lacked the political sophistication to think ideologically (Converse 1964). This perspective was bolstered by research showing disconnects between self-identied ideology and policy preferences (Robinson and Fleishman 1988), response instability in ideology questions across time (Kerlinger 1984), and sensitivity to question wording (Schuman and Presser 1981). Levitin and Miller (1979, p. 768) concluded that when people describe themselves as having an ideological position, they also seem to be saying something about their positions on the parties quite apart from their issue or policy stands. Others found that ideological self-placement was rooted in symbolic considerations, group afliations, and parental socialization rather than political issues (Conover and Feldman 1981). In contrast, recent research concludes that the polarization of the broader political environment has helped to clarify the meaning of ideological labels for the general public (Levine, Carmines, and Huckfeldt 1997; Abramowitz and Saunders 1998). When political elites are more ideologically consistent, politics becomes packaged on an ideological basis (Hinich and Munger 1997). Thus, as candidates have polarized and have campaigned on explicitly ideological rhetoric, ideological labels have become increasingly salient for the voters, thereby allowing individuals to better sort themselves into the appropriate ideological category (Levendusky 2009). George W. Bush, for instance, campaigned on a message of compassionate conservatism in 2000, and although Al Gore did not label himself a liberal, the media frequently did, and he clearly associated himself with the liberal New Deal groups like labor unions and civil rights organizations. Not surprisingly, then, analysis of the 2000 electorate nds that citizens were able to place candidates and parties quite accurately along the liberal-conservative continuum. In fact, they were not conned to the most knowledgeable strata, as has been the case in most prior election years (Jacoby 2004, p. 118). Thus, as Knight (1999, p. 62) explains: while a generation ago direct self-labeling tended to be dismissed (e.g., Free and Cantril 1967), direct measures of liberal-conservatism are now the dominant means of assessing individual ideology in political science. Yet, even as the public might be more aware of what it means to be a conservative or liberal, it does not mean that the liberal-conservative scale adequately captures their policy preferences. The standard ideology survey question assumes that individuals can be placed on a single (liberal-conservative) dimension but, as the political world has polarized, there is also increasing differentiation between social and economic issues (Shafer and Claggett 1995; Inglehart 1997; Layman and Carsey 2002; Carmines and Ensley 2004). And

The Nature of Political Ideology

683

the distinctiveness of these two dimensions of ideology might have consequences for our ability to interpret responses to the standard ideology question (Kerlinger 1984). If ideological preferences are multidimensional, it means that responses to the unidimensional ideology question, especially the moderate and DK categories, likely capture not only those who are centrist but also those who are crosspressured between policy domains. For someone with a liberal position on one policy dimension and a conservative one on another, the liberal and conservative labels are simply inadequate descriptors of political beliefs. As such, when asked their political ideology on a one-dimensional scale, these individuals should be more likely to say DK or to select the middle category. Research on attitudinal ambivalence has shown that individuals who are torn between competing considerations are more likely to skip the survey question or to select the middling category (Alvarez and Brehm 1995). We expect a similar pattern for the ideologically cross-pressured.3 On the other hand, even if we conceptually recognize the existence of multiple dimensions, a single dimension will remain adequate if ideology is able to predict preferences across a variety of different issue domains, as appears to be the case for political elites (McCarty, Poole, and Rosenthal 2006). In the analysis that follows, we examine the extent to which the onedimensional ideology question captures the policy preferences of the American public in todays polarized environment by examining the contours of policy attitudes as they relate to ideological identication. We then scrutinize any disconnects between policy preferences and ideological identication to determine if they are the result of inadequacies of the survey respondents or inadequacies of the survey question (Achen 1975).

Data and Methods

To evaluate the relationship between policy attitudes and ideological identity, we estimate latent measures of economic and social policy preferences using Bayesian item response theory (IRT). The Bayesian IRT model offers a number of methodological advantages to alternative methods, such as an additive scale of issues (Heath, Evans, and Martin 1994; Abramowitz and Saunders 1998) or factor analysis (Layman and Carsey 2002; Ansolabehere, Rodden, and Snyder 2008). An additive scale, although easy to compute, assumes that every issue contributes equally to the underlying preference dimension. The IRT measure, like factor analysis, does not require such an assumption. For instance, if social preferences are more strongly related to abortion attitudes than to environmental policy attitudes, this difference will be captured in the IRT discrimination parameters. But in contrast to conventional factor analysis, the IRT measure
3. Using the 1984 NES, William Jacoby similarly nds that attitudinal consistency is related to ideological identication (Jacoby 1991).

684

Treier and Hillygus

does not assume a multivariate normal distribution for all observed variables. When this is not an appropriate approximation (e.g., dichotomous or ordinal variables), conventional factor analysis can produce biased preference estimates (Kaplan 2004).4 The IRT model directly models the appropriate distribution of the observed indicators, whether nominal, binary, ordinal, or continuous (or any mixture of types). Finally, with the Bayesian IRT model, the latent measures (or factor scores) are estimated directly and simultaneously with the discrimination parameters rather than as postestimation by-products of the covariance structure, as is the case with conventional factor analysis. Consequently, these traits are subject to inference just like any other model parameter, so we can calculate the uncertainty estimates for the latent measures. It is a simple fact that all latent concepts are necessarily measured with error, but alternative methods require the assumption that the resulting estimate is the true value. In contrast, we can quantify if we do a better or worse job of estimating someones placement on the ideological dimensions. And we can then take into account this uncertainty when we use these latent measures as independent variables in subsequent empirical models. In estimating our latent policy dimensions, we rely on 23 questions from the 2000 American National Election Study (survey details and question wording reported in the appendix). Although these questions do not exhaust the universe of policies that might be related to an individuals general belief system, they offer a wide range of politically relevant issues. Five of the issue questions had a split sample design, in which respondents received either a scale or branching question format, so each respondent was asked just 18 different policy questions. Because the ideology questions are split between a branching and scale format in the preelection survey, this survey also offers the opportunity to evaluate different approaches to measuring ideological identication. We model individual issue responses as a function of the unobserved preference dimension via an ordinal item-response model (Treier and Jackman 2008). Given the large number of parameters in the model and the difculty in estimating the parameters jointly using classical methods of maximum likelihood, we work in a Bayesian setting, using Markov Chain Monte Carlo (MCMC) methods to explore the joint posterior density of the model parameters (for a survey of these methods and their applicability to researchers see Jackman (2000, 2004); Gill (2008)). We implement this MCMC scheme using WinBUGS (Lunn et al. 2000).5 We use diffuse normal priors for the discrimination parameters with mean zero and variance 1,000 and standard normal priors for
4. Alternative approaches are available (e.g., factor analysis on polychoric correlation matrix) but these too have limitations and it is more common for researchers to simply ignore this assumption. 5. We let the algorithm run for 100,000 iterations as burn-in, moving away from the start values such that subsequent iterations represent samples from the joint posterior density. Estimates and inferences are based on 500,000 iterations, thinned by 100, in order to produce 5,000 approximately independent draws from the posterior density.

The Nature of Political Ideology

685

the preferences. The cutpoints are parameterized as the sum of parameters dened as in Treier and Jackman (2008, p. 215) and are assigned priors which ensure the ordinality of the cutpoints. One attractive feature of the Bayesian approach is that it simplies treatment of missing data, which are quite high in the measures usedboth because of the question wording experiments and because of item nonresponse. For instance, for individual issue questions we nd that as many as 14 percent of respondents refused to answer or gave a DK response. In total, just 42 percent of respondents answered all issue questions that were asked of them. With a Bayesian approach, an individuals latent ideology scores are estimated with the data available for that individual, and those estimates are simply less precise for those with less data. Critically, that uncertainty can then be accounted for in subsequent statistical models. In contrast, classical factor analysis would require a correction to the swiss cheese data structure, by either collapsing the different question formats, using listwise deletion, or in some way imputing data to ll in the holes. Estimation of the IRT model requires a number of restrictions for identication. For a one-dimensional model, the location and scale are established by normalizing the mean to zero and the variance to one. For more than one dimension, similar to a conrmatory factor analysis, the parameters can be identied with appropriate restrictions on the discrimination parameters. At minimum, this requires identifying a representative item for each dimension, for which the discrimination parameter is set equal to 1, and restricting at least one of these items to load only on that dimension (for details, see Aguilar and West 2000). To allow for a more complex and realistic latent measure, we impose additional restrictions, as discussed below, and conduct robustness checks to ensure the restrictions do not unduly impact the results.

Dimensions of Ideology
Is one dimension of ideology sufcient to capture the policy preferences of the American public? We compare alternative estimations of latent ideology in table 1. Reported are the estimated discrimination parameters, which tap the extent to which each issue explains variation in the latent scores. If a policy item does not help us distinguish among respondents with different preferences on each dimension, the discrimination parameter will be indistinguishable from zero. The wide variation we see in the size of the discrimination parameters highlights the advantage of the IRT measure over a simple additive scale, which would have assumed that all issues loaded equally on the liberal-conservative continuum. Starting with the unidimensional measure reported in the rst column, we see that the individual social issues do not load as highly on the latent scale compared to the economic issues. We are hesitant to conclude that this means

686

Treier and Hillygus

Table 1. Discrimination Parameters for One and Two-Dimensional Models

Independent 1D Aid to poor spending 1.49 Government services (branching) 1.59 Guaranteed jobs (branching) 1.33 Health insurance (branching) 1.23 Public school spending 1.28 Welfare spending 1.06 Guaranteed jobs (scale) 0.99 Social security spending 0.88 Government services (scale) 1.49 Health insurance (scale) 0.93 Tax cut from surplus 0.40 Afrmative action 0.92 Environment (scale) 1.13 Gun control 0.97 Environment (branching) 0.92 Death penalty 0.46 Abortion, partial-birth 0.44 Abortion, parental consent 0.39 Abortion 0.45 Womens role (scale) 0.65 Womens role (branching) 0.43 Gays in military 0.69 Gay adoption 0.85 DIC 65,842.6 Correlation 1.89 0.24 1.67 0.00 1.60 0.18 1.37 0.10 1.24 0.26 1.15 0.01 1.13 0.09 1.10 0.31 0.00 1.00 0.91 0.08 0.25 0.50 0.94 0.12 0.99 0.44 0.85 0.44 0.81 0.37 0.42 0.19 0.21 0.76 0.09 1.12 0.00 1.00 0.41 1.06 0.11 1.29 0.45 1.58 0.77 2.43 63,839.2 Correlated

Economic Social Economic Social 1.80 0.00 1.70 0.00 1.53 0.00 1.36 0.00 1.27 0.00 1.17 0.00 1.12 0.00 1.03 0.00 1.00 0.00 0.92 0.00 0.29 0.00 0.91 0.11 0.87 0.41 0.73 0.41 0.72 0.35 0.38 0.15 0.00 0.72 0.00 0.96 0.00 1.00 0.00 1.09 0.00 1.14 0.00 1.52 0.00 2.51 64,010.7 0.30

that social issues are necessarily less important, however. For one, since we included more economic issue questions than social issue questions, it is not terribly surprising that our underlying latent dimension is more economicbased. More importantly, the fact that the social issues all load rather poorly on a single dimension offers an initial indication that these issues might form a distinct dimension. Certainly, scholars have long argued that political issues fall along at least two different dimensions (Shafer and Claggett 1995).6
6. Others have identied additional dimensions on foreign policy and race (Carmines and Stimson 1989). As a robustness check, we estimated a third dimension dened in reference to afrmative action attitudes (reported in online appendix on POQ website). This dimension had some odd characteristicssmall and even unexpectedly negative loadings on some issueslikely reecting the small number of questions available. And the largest loadings for this dimension (e.g., welfare spending) had even larger loadings on other dimensions, indicating they were more closely related to another dimension. More importantly for our purposes, the structure of the economic and social

The Nature of Political Ideology

687

Conrming this expectation, we nd that two dimensions of ideology (middle columns in table) better summarize the publics policy preferences, as evidenced by the improvement in the Bayesian deviance information criterion (DIC), the goodness of t measure. The smaller the value of DIC, the better the model. We also see that the items which loaded weakly on the one-dimensional model are the strongest items on the second (social) dimension.7 This two-dimensional model was estimated with the loosest restrictions possiblean unconstrained model where all but two indicators were allowed to load on either dimension.8 The economic dimension is dened by the question concerning government spending and services (scale format), while the item on abortion is associated exclusively with the social dimension. Although this specication makes clear that a two-dimensional model is preferable to a one dimensional one, it seems unrealistic to assume that the economic and social dimensions are orthogonal. To relax this assumption, allowing preferences on these two dimensions to be correlated, we have to implement additional restrictions. To estimate these correlated dimensions, we restrict many of the items to load on only one dimension or the other. Those items that did not clearly load better on one dimension or the other in the independent models were allowed to load on both dimensions.9 The resulting discrimination parameters are reported in the nal columns of the table.10 In addition to providing a clearer structure which identies these dimensions as economic and social preferences, these preferences are now allowed to correlate, which they do at a moderate level of 0.30. Reassuringly, we nd the individual discrimination parameters look nearly identical with either specication. In the remainder of the analysis, we rely on this correlated two-dimensional latent measure to evaluate the self-reported measures of ideology.11
dimensions are unchanged by the inclusion of a third dimension. The correlation between economic dimensions with and without the third dimension is 0.979 (0.954 for social dimensions). 7. Indeed, the respondents preferences in the 1D model match well with the economic preferences in the 2D model (0.962 correlation), but less so with social preferences (0.580). 8. An alternative identifying restriction is to x individual respondents on the latent dimension. We obtain similar results when normalizing by xing the positions of three respondents: one extremely liberal on both dimensions [3, 3], one extremely conservative on both [3, 3], and one respondent extreme on both, but cross-pressured [3, 3]. The resulting latent scores correlate with those reported at .995 on the economic dimension and .995 on the social dimension. 9. As a robustness check, we have also estimated a model where every item loaded only on one dimension, with nearly identical results. The resulting latent scores correlate with those reported at .995 on the economic dimension and .993 on the social dimension. 10. We nd no evidence that the estimates do not converge. For each set of estimates, 5.5 percent or less of the parameters fail Gewekes diagnostic, an amount, under standard levels of statistical signicance, one would expect to randomly occur if all of the parameters converged to the stationary distribution. 11. A comparison of an additive scale created using the issue items available for each respondent nds a correlation of .82 for the economic dimension and .67 on the social dimension. The lower correlation for the social dimension no doubt reects the greater variability found in the discrimination parameters, again afrming the advantages of the IRT model. Unfortunately, we

688

Treier and Hillygus

2 Economic Preferences 0.95 0.41 0 0.47 0.92 1 0.31 0.06 0.62 0.26

3
Extremel Liberaly Slightly Liberal Slightly Extremely Conservative Conservative Conservative

Liberal

Moderate

Dont Know

2 1.08 Social Preferences 1 0.57 0.3 0 0.89 1 1.32 0.64 0.12 0.32

3
Extremel Liberaly Liberal Slightly Liberal Moderate Slightly Extremely Conservative Conservative Conservative Dont Know

Figure 2. Latent Policy Preferences by Self-Reported Ideology.

Comparing Ideology Measures

We start by separately comparing each latent dimension to the ideological self-placement scale in the postelection survey. In gure 2, we graph the corresponding box plot of estimated latent scores for each response category.12 The gure shows a clear relationship between policy attitudes and ideological
cannot create a directly comparable estimate using factor analysis given the pattern of missing data and the discrete nature of some of the variables. 12. In a box plot graph, the box indicates the interquartile range (marking the lower quartile, median, and upperquartile), the whiskers show the range of the data (1.5 times the interquartile range), and circles indicate outliers.

The Nature of Political Ideology

689

self-placementwith self-identied liberals having more liberal policy preferences on social and economic issues than self-identied conservatives and vice-versa. At the same time, the graph highlights the considerable variability in the policy preferences of each group, especially among self-identied moderates and those answering dont know. Indeed, the interquartile range (the box) of the moderate respondents overlaps the median value of all but the extreme ideologues, offering the rst indication that the moderate label, in particular, covers a wide range of policy attitudes. The box plot graphs offer other interesting comparisons as well. First, relative to self-identied liberals (of any intensity), self-identied conservatives in the sample have more diverse policy preferences on both dimensions, as evidenced by the longer whiskers. This no doubt reects the negative connotations associated with the liberal label that has been prevalent in American politics in recent decades. Differences in the social desirability of the ideological labels mean that more people are willing to identify as conservative even though they dont necessarily hold the policy preferences associated with the label (Miller 1992). Comparing across dimensions also nds that self-identied liberals and conservatives are more polarized (i.e. further from 0) on the social dimension than on the economic dimension. Second, we see that the intervals between categories are not equal, as is assumed by the 7-point self-placement scale. In particular, there is a much larger gap in the median between those who call themselves extremely liberal (conservative), and those who call themselves liberal (conservative) relative to other gaps on the scale. In contrast to the ideological self-placement measure in which ideological labels can mean different things to different people our latent measure of policy preferences places all respondents on the same underlying policy scale and provides ner distinctions between individuals. Finally, we see that those who say dont know have a preference distribution quite similar to that of self-identied moderates. Self-reported ideology questions often have rather high levels of DK responses, leaving scholars unsure how to handle the missing data problem. The similarities we see here offer some empirical justication to simply recoding DK to be moderate (rather than omitting them from the analysis), as is common in some research. To be clear, though, the reason the two groups look so similar is that self-identied moderates, like the dont knows, have an equally diverse set of policy preferences. Undoubtedly, both the DK and moderate responses are selected for a variety of reasons, only some of which reect centrism (for self-identied moderates) or uncertainty (for DK). The latent measures offer a benchmark for evaluating the various measures of ideology available in the 2000 American National Election Study. Comparing the correlations between each of the ideology questions and the latent measures, shown in table 2, nds that the postelection scale measure outperforms the other two. It is perhaps not surprising that a postelection measure outperforms a preelection measure since people might be better able to

690

Treier and Hillygus

Table 2. Comparison of Latent Measures and Self-reported Ideology

Latent economic Correlation Preelection branching ideology Preelection scale ideology Postelection scale ideology
NOTE.Polyserial correlations are reported.

Latent social Correlation .490 .523 .547

.425 .478 .486

select an appropriate ideological label because of campaign information. But we also nd that the preelection scale outperforms the preelection branching question. This result offers some challenge to previous research that concluded that branching measures were preferable to scale measures (Aldrich et al. 1982), but is consistent with other analysis of the 2000 ANES data (Aldrich, Grifn, and McKay 2002). For the remainder of the analysis evaluating ideological identication, we rely on this standard postelection 7-point scale. Turning to a comparison of the individual-level relationships between the social and economic dimensions, we plot in gure 3 each respondents social and economic scores for self-reported conservatives, liberals, moderates, and dont knows.13 For those who call themselves liberal or conservative, we see that there is a clear relationship between the latent social and economic dimensions, with respondents clustered in the corresponding conservative or liberal quadrants of the graphs. Thus, the more conservative an individuals preferences on social issues, the more conservative we expect her to be on economic issues, offering evidence of ideological constraint among this subset of the electorate. In contrast, there is a much weaker relationship between economic and social issue preferences for moderates and DK respondents. For these respondents we see a diffuse cloud of data points and a smaller Beta coefcient.14 This indicates that individuals who self-identify as moderates are not simply neutral or moderate across political issues, but are often cross-pressured between the economic and social dimension (located in quadrants 2 and 4). Some 44 percent of moderates and 47 percent of DK respondents are in these cross-pressured quadrants. If we dene someone as holding centrist policy positions if they fall within the middle tercile of both the economic and social latent dimensions, we nd that just 17 percent of self-identied moderates have centrist

13. The end categories of the post-election 7-point scale are collapsed so liberals and conservatives include the extreme ideologues. Not shown are the slightly categories; not surprisingly, the observed relationships for this group fall in between that of ideologues and moderates. 14. The reported Betas are the regression coefcients for the economic dimension regressed onto the social dimension.

The Nature of Political Ideology

691

^ = 0.37
3
10.9%

Liberal

Moderate
^ = 0.21
3
8.4% 20.7%

24.9%

Social Dimension

65.3%

15.3%

31.6%

22.8%

Conservative
^ = 0.32
3

Dont Know
^ = 0.21
3
62.1% 34.7%

15.3%

26.1%

Social Dimension

8.9%

13.7%

27.1%

12.1%

Economic Dimension

Economic Dimension

Figure 3. Economic versus Cultural Dimension of Ideology by Self-Reported Ideology. Regression coefcients and percent of observations reported for each category. preferences.15 Even expanding the denition of centrist to include respondents with preferences in the 25th to 75th percentile of both dimensions, we still nd that only 35 percent of self-identied moderates hold centrist positions on both policy dimensions. We also nd slight asymmetries between liberal and conservative identiers that again seem to reect differences in the social desirability of the two
15. Using this same threshold, we nd that 59 percent are cross-pressured, either centrist on one dimension and extreme on the other or extreme on both in opposing directions, another 8 percent are misclassied conservatives, and 15 percent are misclassied liberals.

692

Treier and Hillygus

labels. Self-identied liberals have a stronger relationship between the economic and social dimensions on average than do the self-identied conservatives. And slightly more conservatives fall outside their corresponding quadrant (38 percent compared to 35 percent). This pattern is consistent across the various thresholds for the latent dimensions16 , even as the majority of selfidentied liberals and conservatives hold policy preferences consistent with their ideological labels across both dimensions. Very few liberals or conservatives are completely misclassied, holding opposing preferences on both dimensionsfewer than 9 percent based on the quadrant classication and fewer than 3 percent based on the tercile classication. In sum, this descriptive analysis indicates that, even in todays polarized environment, the commonly asked survey question about ideological self-placement is inadequate for capturing the complex belief systems of a sizable portion of the American public. As political commentator Jim Hightower puts it: Most of us are mavericks, political muttseach one of us a heady and sometimes hot mix of liberalisms, conservatisms and radicalisms (Hightower 1997, p. 235). These political mutts often identify as political moderates, making it difcult to interpret the standard measure of ideology.

Predicting Use of Ideological Labels

The descriptive analysis above suggests that self-identied moderates are made up of at least two very different kinds of peoplepolicy centrists and the ideologically cross-pressured. The key alternative perspective, however, is that the observed variation in economic and social preferences simply reects a lack of political sophistication. To account for this possibility, we estimate a logit model with the appropriate controls predicting an individuals decision to select the Moderate, Ideologue, or DK categories. To avoid arbitrary thresholds, we measure cross-pressured preferences as the product of the economic and social dimensions (multiplied by 1 so that values range from least to most cross-pressured), so that the measure captures the ideological distance between dimensions. We include in the model a measure of political knowledge to account for the possibility that moderate identication simply reects political ignorance, as hypothesized by Converse (1964).17 We also control for the potential symbolic aspects of ideological labels with a strength of partisanship measure (partisan versus independent) and the socialization and group identity aspects of ideological labels with demographic controls for race, gender, and age. Finally, we control for policy extremism on the individual economic and social dimensions.
16. For instance, 10 percent of self-identied liberals have preferences in the conservative tercile on at least one dimension, while 12 percent of conservatives have preferences in the liberal tercile on at least one dimension. 17. Following Bartels (1996), we use the interviewer assessment of political knowledge.

The Nature of Political Ideology

693

The model results are estimated simultaneously with the IRT measures to take into account the measurement error involved in the estimation of our latent measures.18 Thus, if we have a lot of uncertainty in our estimate of the economic or social preferences of an individualperhaps because of skipped questions or a random response patternthat uncertainty is carried over into the subsequent model estimates. Simply including the latent measure without accounting for the uncertainty in the estimate of that measurethe approach typically used with factor scores from a traditional factor analysiscould result in biased coefcient estimates. In this respect, our approach offers additional reassurance that we are controlling for political knowledge by accounting for the uncertainty in the latent measures (since some individuals are measured with more error than others). The ndings are reported in table 3. Reported are the coefcient estimates and the 90 percent highest posterior density (HPD) intervals. This offers a gauge of signicance since we can say there is a 90 percent probability that the coefcient lies within the interval.19 Looking rst at some of our political controls, we see that the decision to identify as a political moderate is related to many of the characteristics hypothesized in previous research. Partisans are less likely than independents to identify as moderates, and the more politically knowledgeable are less likely than the politically ignorant to call themselves moderate (and more likely to call themselves moderate than to say Dont Know). Even with all of these controls, though, being cross-pressured has a sizable and signicant effect on use of the moderate label. There is a 93 percent probability that the cross-pressure coefcient is greater than zero for the Moderate versus Ideologue comparison.20 To illustrate the substantive impact of policy cross-pressures, we graph in gure 4 a contour plot of predicted probabilities to show the variation in the probability of identifying as a moderate across various levels of economic and social preferences. We see that those most cross-pressured (the darker areas in the image) have a higher elevation or probability of calling themselves moderate.21 Individuals who are the least cross-pressured (most ideologically extreme on both dimensions) have just a 10 percent chance of calling themselves ideological moderate, compared to a 25 percent chance among those with

18. The simultaneous estimation is implemented by imposing a standard multinomial logit model for self-placement as an ideologue, moderate, or DK alongside the measurement model (with parameters ), assuming vague normal priors on the logit coefcients (labeled ). The MCMC chain then updates the estimates of , given the previous estimates of , and is updated given the previous estimates of . 19. By contrast, with standard (frequentist) condence intervals, we know only that 90 percent of the time the interval contains the true value; there is no indication of how likely a set of values are. 20. There is an 89 percent probability that the coefcient is greater than zero for the DK v. Ideologue comparison. 21. Estimates are made holding variables at their mean category or mode for indicator variables.

Table 3. Predicting Ideological Self-Placement

All respondents Dont Know versus Ideologue Moderate versus Dont Know Moderate versus Ideologue Dont Know versus Ideologue Dont Know versus Moderate Knowledgeable respondents

694

Moderate versus Ideologue

Intercept

Age

Female

Nonwhite

South

Income

Partisan

0.88 [0.12, 1.63] 0.006 [0.002, 0.015] 0.08 [0.36, 0.19] 0.18 [0.24, 0.58] 0.15 [0.46, 0.13] 0.023 [0.06, 0.016] 0.94 [1.22, 0.68]

0.85 [1.88, 0.27] 0.019 [0.008, 0.03] 0.34 [0.03, 0.70] 0.94 [0.49, 1.39] 0.48 [0.12, 0.85] 0.22 [0.29, 0.15] 0.49 [0.86, 0.10]

1.73 [0.64, 2.93] 0.013 [0.024, 0.001] 0.42 [0.80, 0.027] 0.76 [1.25, 0.25] 0.63 [1.02, 0.23] 0.19 [0.12, 0.27] 0.44 [0.86, 0.04]

Political knowledge

Economic extremism

Social extremism

Cross-pressure

1.29 [0.61, 2.01] 0.0074 [0.0004, 0.015] 0.14 [0.38, 0.10] 0.0015 [0.34, 0.40] 0.11 [0.36, 0.14] 0.017 [0.05,0.017] 0.88 [1.11, 0.61] 0.17 [0.29, 0.05] 0.81 [1.20, 0.44] 0.60 [1.02, 0.22] 0.23 [0.03, 0.49] 1.00 [1.41, 0.53] 0.62 [1.05, 0.16] 0.19 [0.01, 0.50]

0.81 [1.63, 0.09] 0.009 [0.018, 0.0003] 0.31 [0.63, 0.005] 0.90 [1.31, 0.48] 0.39 [0.71, 0.055] 0.09 [0.03, 0.14] 0.30 [0.61, 0.03] 0.85 [0.67, 1.01] 0.60 [1.11, 0.10] 0.54 [1.00, 0.03] 0.07 [0.22, 0.38]

0.71 [1.29, 0.08] 0.18 [0.77, 0.44] 0.071 [0.45, 0.31]

N Correctly predicted

Treier and Hillygus

Naive (by mode) Proportional Reduction of error

2.09 [1.27, 2.89] 0.017 [0.009, 0.026] 0.17 [0.12, 0.45] 0.92 [0.56, 1.28] 0.28 [0.01, 0.57] 0.11 [0.15, 0.05] 0.57 [0.87, 0.29] 1.02 [1.18, 0.87] 0.21 [0.61, 0.20] 0.06 [0.42, 0.32] 0.16 [0.05, 0.36] 1,298 0.618 [0.605, 0.632] 0.556 0.111 [0.087, 0.136] 0.29 [0.81, 0.19] 0.44 [0.95, 0.08] 0.26 [0.02, 0.57] 1,008 0.645 [0.630, 0.622] 0.626 0.031 [0.0063, 0.057]

NOTE.Entries are coefcients, 90% Highest Posterior Density (HPD) intervals in brackets.

The Nature of Political Ideology

695

Conservative

1.5

0.

0.4
0.45

0.1

5
0.2

1.0

0.5

0.2
0.3 5

Social Preferences

0.5

0.6
00

0.3

0.5

0.3
0.2
0.1

0.3

0.55

0.2

1.0
Liberal

1.5

5
1.0 0.5 0.0

0.4
0.5 1.0

0.3
1.5
Conservative

1.5
Liberal

Economic Preferences

Least

Most

Shading code for "CrossPressure" level curves are probabilities

Figure 4. Predicted Probability of Identifying as a Moderate. the greatest ideological distance between their economic and social policy preferences (97.5th percentile of cross-pressures measure).22 As an additional robustness check, we have estimated the model only for the politically knowledgeable respondents (rather than simply controlling for them in the models), and nd nearly identical results. Thus, even accounting for symbolic considerations, political knowledge, and policy centrism, individuals with divergent economic and social preferences are more likely to call themselves moderate than to use a liberal or conservative label. This nding offers an important corrective to recent characterizations of the American public as being either centrist or polarized. The reality is that the belief systems of the mass public are more complexand the label of ideological moderate so often used to describe the American public represents a diverse set of policy attitudes.

Implications
The nding that the moderate label masks differences between policy centrists and cross-pressured respondents has consequences for our theories and models
22. Examining changes in the predicted probabilities across varying levels of values for all three variables still nds that the individuals are most likely to identify as a moderate when they are centrist on both dimensions or when they are cross-pressured between them.

696

Treier and Hillygus

of political behavior. Ideologically moderate voters, after all, are often at the heart of theories of electoral democracy. The classic median voter theorem, for instance, predicts that moderates are the pivotal voters in the election, inducing politicians to advocate centrist public policies (Downs 1957). Moderates are also considered the all-important swing voters willing to change candidate support across or within elections and, as such, hold the balance of power in national elections (Converse 1964). Following the 2006 midterm election, for example, one analyst concluded: the outcome of this electionand others in our recent historywas determined by the shifting sentiments of Independents and moderates.23 There is a clear tendency for scholars and journalists to treat political moderates as a homogeneous group, painting broad strokes about their attitudes and behavior. Yet, our analysis documents important heterogeneity in the policy preferences of moderates that could well inuence conclusions about their voting behavior. We evaluate this possibility empirically by comparing predictions across vote choice models that either include or exclude the latent ideology scores, while controlling for ideological self-placement.24 As we see in gure 5, we can get widely different predictions about the behavior of self-identied moderates. The graph plots the estimated error in individual predictions if we treat moderates as homogeneous in their policy preferences. Across different combinations of economic and social preferences, we nd that an individuals predicted probability of voting for Bush can be off by as much as 38 percentage points. In contrast, predictions are off by no more than 7 percentage points for liberals (not statistically signicant) and 13 percentage points for conservatives (rarely changing predicted candidate choice). Consequently, while the standard selfplacement scale may be adequate in summarizing the preferences of liberals and conservatives and predicting their political behavior, it performs quite poorly for self-identied moderates. Beyond the example above, the complex contours of policy attitudes that we have identied may have a variety of consequences for studies of electoral dynamics. A more complete understanding of the structure and meaning of ideological identication could well change our expectations and conclusions about split ticket voting, political engagement, and campaign effects. Moreover, recognizing the distinction between policy centrists and the ideologically cross-pressured might affect our basic theories about candidate behavior, since centrists and cross-pressured respondents may be responsive to very different campaign strategies (Hillygus and Shields 2008).

23. Andrew Kohut, The Real Message of the Midterms, Pew Center Report November 14, 2006. 24. The logit model is again estimated simultaneously with IRT measures and includes party identication, age, gender, race, income, and an indicator for the South. Predicted probabilities are calculated holding all other variables at their means or modes. The full set of coefcients are available in an online appendix on the POQ website. The results do not change when the liberal conservative scale is replaced by dummy variables.

The Nature of Political Ideology

697

Social 25% Median 75%

0.4

Economic 25% Median 75%

0.3

0.3

0.2 Difference in Predicted Probability Difference in Predicted Probability 1.5 1.0 0.5 0.0 0.5 Economic Policy Preferences 1.0

0.2

0.1

0.1

0.0

0.0

0.1

0.1

0.2

0.2

0.3

0.3

1.5

1.0 0.5 0.0 0.5 Social Policy Preferences

1.0

Figure 5. Error in the Predicted Probability of a Bush Vote.

Discussion
Given the current polarized nature of American politics, renewed attention has been focused on the ideological preferences of the mass public. Yet, the way we conceptualize and measure those preferences shapes our conclusions about their distribution and inuence. Our analysis documents the multidimensional nature of policy preferences in the American electorate, and nds a noteworthy number in the public are liberal on one dimension and conservative on another. Because these cross-pressured individuals tend to call themselves moderate (or say DK), it undermines interpretation of the standard 7-point ideological identication scale so often used in political research. Thus, even as scholars nd that ideological labels are more meaningful than ever before, those labels are accurate representations of policy preferences only for those self-identifying as a liberal or conservative. We are by no means the rst to acknowledge that ideological self-placement is a awed measure because of mismatches between ideological identication and policy preferences. But in contrast to early research, we cannot attribute the disconnect between self-reported ideology and issue attitudes to the lack of political knowledge alone. Rather, many people are coherent along the economic and social dimension separately, but are simply cross-pressured between them. Our results show that failing to account for the multidimensional nature of ideological preferences can produce inaccurate predictions of voting behavior

698

Treier and Hillygus

for the plurality of Americans who do not call themselves liberal or conservative. As such, we recommend that future research use distinct measures of social and economic preferences in empirical models of mass behavior. Since the standard approach to measuring ideology has been to ask about identication along a unidimensional scale, it would be particularly fruitful for scholars to explore the potential for including direct measures of preferences across multiple ideological dimensions as an alternative to creating issue-based measures used here (see Hooghe, Marks, and Wilson 2002 for example with European elites). Until a set of valid and reliable survey questions are identied, researchers should include a large and diverse set of policy items on their survey questionnaires so that issue-based scales can be created. To be sure, our ndings do not imply that the ideological self-placement measure should never be used. Scholars have long noted the symbolic importance of liberal and conservative labels (Stimson 2004), and our results suggest that these labels are meaningful representations of policy preferences for self-identied liberals and conservatives. However, researchers should at least operationalize ideological self-placement as a series of dummy variables in their empirical models since the measure cannot be assumed to be an ordinal scale with political moderates in the middle. And even then, this approach cannot distinguish between so-called moderates who are centrist and those who are cross-pressured, making it inadequate for any theory or model that is dependent on a measure of policy centrism. Beyond these practical implications, these ndings are relevant to the ongoing polarization debate. On one side are those who say that political moderates have either followed political elites to the ideological extremes or, frustrated by the polarized environment, have dropped out of the political process altogether (Abramowitz and Saunders 1998, 2005; Layman and Carsey 2002). On the other side are those who contend that the majority of Americans have remained ideologically centrist even as political elites have grown more polarized (Dimaggio, Evans, and Bryson 1996; Fiorina 2004). It turns out that neither portrait of the American moderate is entirely accurate. Our ndings make clear that the American public is not as ideologically extreme as often portrayed, but nor are they truly centrist. And the heterogeneous political complexion of the American public has consequences for the way we measure political ideology and the way we use it in our theories and models of political behavior.

Appendix
2000 AMERICAN NATIONAL ELECTION STUDY The 2000 ANES was conducted by the Center for Political Studies of the Institute for Social Research. The preelection survey was conducted from September 5 to November 6, and the postelection re-interview ran from

The Nature of Political Ideology

699

November 8 to December 18. The study population was all U.S. citizens of voting age living in the forty-eight contiguous states. The sampling design was a dual frame sample that included both a traditional area probability sampling using face-to-face (FTF) interviews (1,006 pre respondents) as well as a RDD stratied equal probability sample interviewed by telephone (801 prerespondents). The response rates, calculated as the ratio of completed interviews to the total number of potential respondents, were 64.8 percent for FTF and 57.2 percent for phone for preelection and 57.2 percent for FTF and 85.8 percent for phone for postelection. More details about the methodology are available at http://www.electionstudies.org/studypages/2000prepost/2000prepost.htm.
QUESTION WORDING

All variables were recoded to run from liberal to conservative. Question wording for model controls are available in an online appendix on the POQ website.

Ideological

self-placement measures: Liberal-Conservative Scale (V001368/V000440/VCF0803) [FTF]: Where would you place yourself on this scale, or havent you thought much about this? Scale: (1) extremely liberal, (2) liberal, (3) slightly liberal, (4) moderate; middle of the road, (5) slightly conservative, (6) conservative, (7) extremely conservative. [phone]: When it comes to politics, do you usually think of yourself as extremely liberal, liberal, slightly liberal, moderate or middle of the road, slightly conservative, conservative, extremely conservative, or havent you thought much about this?; Liberal-Conservative Branching Measure (V000446) (FTF and phone): When it comes to politics, do you usually think of yourself as a liberal, a conservative, a moderate, or havent you thought much about this? If you had to choose, would you consider yourself a liberal or a conservative? Would you call yourself a strong liberal or a not very strong liberal? Would you call yourself a strong conservative or a not very strong conservative?

Issue questions: V000748: Do you think gay or lesbian couples, in other

words, homosexual couples, should be legally permitted to adopt children?; V000545: [FTF] Where would you place yourself on this scale, or havent you thought much about this? (17 scale) 1 govt should provide many fewer services, 7 govt should provide many more services; V000549: [phone] Which is closer to the way you feel or havent you thought much about this? Should the government reduce/increase services and spending a great deal or (reduce/increase services and spending) only some.; V000609: [FTF] Where would you place yourself on this scale, or havent you thought much about this? 17 scale, 1 govt insurance plan, 7 private insurance plan; V000610: [phone] Which is closer to the way you feel or havent you thought much about this? do you feel strongly or not strongly that there should be a government insurance plan?/do you feel strongly or not strongly that individuals

700

Treier and Hillygus

should pay through private insurance plan?; V000615: [FTF] Where would you place yourself on this scale, or havent you thought much about this? 17 scale: 1. govt should see to jobs and standard of living 7. govt should let each person get ahead on own.; V000619: [phone] Which is closer to the way you feel or havent you thought much about this? Do you feel strongly that the government should see to it that every person has a job and a good standard of living, or not so strongly? Do you feel strongly that the government should just let each person get ahead on their own, or not so strongly?; V000674a: [standard version] Some people think that if a company has a history of discriminating against blacks when making hiring decisions, then they should be required to have an afrmative action program that gives blacks preference in hiring. What do you think? Should companies that have discriminated against blacks have to have an afrmative action program? [EXPERIMENTAL VERSION] Some people think that if a company has a history of discriminating against blacks when making hiring decisions, then they should be required to have an afrmative action program that gives blacks preference in hiring. What do you think? Should companies that have discriminated against blacks have to have an afrmative action program or should companies not have to have an afrmative action program? [BOTH VERSIONS] Do you feel strongly or not strongly (that they should not have to have afrmative action)?; V000676: Should federal spending on welfare programs be increased, decreased, or kept about the same?; V000680: Should federal spending on aid to poor people be increased, decreased, or kept about the same?; V000681: Should federal spending on social security be increased, decreased, or kept about the same?; V000683: Should federal spending on public schools be increased, decreased, or kept about the same?; V000690: Some people have proposed that most of the expected federal budget surplus should be used to cut taxes. Do you approve or disapprove of this proposal? Do you approve of this proposal strongly or not strongly? Do you disapprove of this proposal strongly or not strongly?; V000694: [FTF] There has been some discussion about abortion during recent years. Which one of the opinions on this page best agrees with your view? You can just tell me the number of the opinion you choose. [PHONE] There has been some discussion about abortion during recent years. I am going to read you a short list of opinions. Please tell me which one of the opinions best agrees with your view? You can just tell me the number of the opinion you choose. options: (1) by law, abortion should never be permitted. (2) the law should permit abortion only in case of rape, incest or when the womans life is in danger. (3) the law should permit abortion for reasons other than rape, incest, or danger to the womans life, but only after the need for the abortion has been clearly established. (4) by law, a woman should always be able to obtain an abortion as a matter of personal choice.; V000702: Would you favor or oppose a law in your state that would require a teenage girl under age 18 to receive her parents permission before she could obtain an abortion? Strongly or not strongly?; V000705: There has been discussion recently about a proposed law to ban

The Nature of Political Ideology

701

certain types of late-term abortions, sometimes called partial birth abortions. Do you favor or oppose a law that would make these types of abortions illegal? Do you strongly or not strongly favor/oppose a law that would that would make these types of abortions illegal?; V000727: Do you think homosexuals should be allowed to serve in the United States Armed Forces or dont you think so? Do you feel strongly or not strongly that homosexuals should be allowed to serve? Do you feel strongly or not strongly that homosexuals should not be allowed to serve?; V000731: Do you think the federal government should make it more difcult for people to buy a gun than it is now, make it easier for people to buy a gun, or keep these rules about the same as they are now? A lot easier/more difcult or somewhat easier/more difcult?; V000752: Do you favor or oppose the death penalty for persons convicted of murder? (Do you favor/oppose the death penalty for persons convicted of murder) strongly or not strongly?; V000755: [FTF standard] Where would you place yourself on this scale, or havent you thought much about this? [FTF experimental] where would you place yourself on this scale? 17 scale: 1. women and men should have equal roles, 7. a womans place is in the home.; V000759: [phone version 1] Which is closer to the way you feel, or havent you thought much about this? [phone version 2] which is closer to the way you feel? [both phone versions] Do you feel strongly or not strongly that men and women should have equal roles? Do you feel strongly or not strongly that a womans place is in the home?; V000771: [FTF] Where would you place yourself on this scale, or havent you thought much about this? 17 scale: 1. tougher regulations on business needed to protect environment, 7. regulations to protect environment already too much a burden on business.; V000775: [Phone] Which is closer to the way you feel, or havent you thought much about this? Do we need to toughen regulations to protect the environment a lot, or just somewhat? Are regulations to protect the environment way too much of a burden on business or just somewhat of a burden?

Supplementary Data
Supplementary data are available online at http://poq.oxfordjournals.org/

References
Abramowitz, Alan I., and Kyle L. Saunders. 1998. Ideological Realignment in the U.S. Electorate. Journal of Politics 60(3):63452. Abramowitz, Alan, and Kyle Saunders. 2005. Why Cant We All Just Get Along? The Reality of a Polarized America. The Forum: A Journal of Applied Research in Contemporary Politics 3(2):124. Achen, Christopher H. 1975. Mass Political Attitudes and the Survey Response. American Political Science Review 69(4):121831.

702

Treier and Hillygus

Aguilar, Omar, and Mike West. 2000. Bayesian Dynamic Factor Models and Portfolio Allocation. Journal of Business & Economic Statistics 18(3):33857. Aldrich, John, John Grifn, and Amy McKay. 2002. Apples and Oranges: Comparing Forum and Format on the National Election Studies. Duke University. Aldrich, John, Richard G. Niemi, George Rabinowitz, and David W. Rohde. 1982. The Measurement of Public Opinion about Public Policy: A Report on Some New Issue Question Formats. American Journal of Political Science 26(2):391414. Alvarez, R. Michael, and John Brehm. 1995. American Ambivalence towards Abortion Policy: Development of a Heteroskedastic Probit Model of Competing Values. American Journal of Political Science 39(4):105582. Ansolabehere, Stephen, Jonathan Rodden, and James Snyder. 2008. The Strength of Issues: Using Multiple Measures to Gauge Preference Stability, Ideological Constraint, and Issue Voting. American Political Science Review 102(2):21532. Bartels, Larry. 1996. Uninformed Voters: Information Effects in Presidential Elections. American Journal of Political Science 40(1):194230. Carmines, Edward, and Michael Ensley. 2004. Strengthening and Weakening Mass Partisanship: Issue Preferences and Partisan Attitudes in an Increasingly Polarized Party System. Midwest Political Science Association, Chicago. Carmines, Edward G., and James A. Stimson. 1989. Issue Evolution: Race and the Transformation of American Politics. Princeton, NJ: Princeton University Press. Conover, Pamela Johnston, and Stanley Feldman. 1981. The Origins and Meaning of Liberal/ Conservative Self-Identications. American Journal of Political Science 25(4):61745. Conover, Pamela Johnston, and Stanley Feldman. 1984. How People Organize the Political World: A Schematic Model. American Journal of Political Science 28(1):95126. Converse, Philip E. 1964. The Nature of Belief Systems in Mass Publics. In Ideology and Discontent, ed. David E. Apter. New York: The Free Press. Dimaggio, Paul, John Evans, and Bethany Bryson. 1996. Have Americans Social Attitudes Become More Polarized? American Journal of Sociology 102(3):690755. Downs, Anthony. 1957. An Economic Theory of Democracy. New York: Harper Collins. Fiorina, Morris. 2004. Culture War? The Myth of a Polarized America. New York: Pearson Longman. Free, Lloyd, and Hadley Cantril. 1967. The Political Beliefs of Americans: A Study of Public Opinion. New Brunswick, NJ: Rutgers University Press. Gill, Jeff. 2008. Bayesian Methods. 2nd ed. New York: Chapman and Hall/CRC Press. Heath, Anthony, Geoffrey Evans, and Jean Martin. 1994. The Measurement of Core Beliefs and Value. British Journal of Political Science 24(1):11532. Hightower, Jim. 1997. Theres Nothing in the Middle of the Road but Yellow Stripes and Dead Armadillos. New York: HarperCollins. Hillygus, D. Sunshine, and Todd Shields. 2008. The Persuadable Voter. Princeton, NJ: Princeton University Press. Hinich, Melvin J., and Michael C. Munger. 1997. Analytical Politics. Cambridge: Cambridge University Press. Hooghe, L., G. Marks, and C. J. Wilson. 2002. Does Left/Right Structure Party Positions on European Integration. Comparative Political Studies 35(8):96589. Inglehart, Ronald. 1997. Modernization and Postmodernization: Cultural, Economic, and Political Change in 43 Societies. Princeton, NJ: Princeton University Press. Jackman, Simon. 2000. Estimation and Inference via Bayesian Simulation: An Introduction to Markov Chain Monte Carlo. American Journal of Political Science 44(2):375404. Jackman, Simon. 2004. Bayesian Analysis for Political Research. Annual Review of Political Science 7:483505. Jacoby, William. 1991. Ideological Identication and Issue Attitudes. American Journal of Political Science 35(1):178205.

The Nature of Political Ideology

703

Jacoby, William. 2004. Ideology in the 2000 Elections: A Study in Ambivalence. In Models of Voting in Presidential Elections, eds. Herbert Weisberg, and Clyde Wilcox. Stanford: Stanford University Press. Kaplan, David. 2004. The Sage Handbook of Quantitative Methodology for the Social Sciences. Newbury Park, CA: Sage. Kerlinger, Fred. 1984. Liberalism and Conservatism: The Nature and Structure of Social Attitudes. NJ: Hillsdale. Knight, Kathleen. 1999. Liberalism and Conservatism. In Measures of Political Attitudes, eds. J. P. Robinson, L. S. Wrightsman, and P. R. Shaver. San Diego, CA: Academic. Layman, Geoffrey. 1999. Culture Wars in the American Party System: Religious and Cultural Change Among Partisan Activists Since 1972. American Politics Research 27(1):89121. Layman, Geoffrey, and Tom Carsey. 2002. Party Polarization and Conict Extension in the American Electorate. American Journal of Political Science 46(4):786802. Levendusky, Matt. 2009. The Partisan Sort: How Liberals Became Democrats and Conservatives Became Republicans. University of Chicago Press. Levine, Jeffrey, Edward Carmines, and Robert Huckfeldt. 1997. The Rise of Ideology in the Post-New Deal Party System, 19721992. American Politics Research 25(1):1934. Levitin, Teresa E., and Warren E. Miller. 1979. Ideological Interpretations of Presidential Elections. American Political Science Review 73(3):75171. Lunn, David J., Andrew Thomas, Nicky Best, and David J. Spiegelhalter. 2000. WinBUGSA Bayesian Modelling Framework: Concepts, Structure, and Extensibility. Statistics and Computing 10(4):32537. Marcus, George E., David Tabb, and John L. Sullivan. 1974. The Application of Individual Differences Scaling to the Measurement of Political Ideologies. American Journal of Political Science 18(2):40520. McCarty, Nolan, Keith Poole, and Howard Rosenthal. 2006. Polarized America: The Dance of Ideology and Unequal Riches. MIT Press. Miller, Alan S. 1992. Are Self-Proclaimed Conservatives Really Conservative? Trends in Attitudes and Self-Identication among the Young. Social Forces 71(1):195210. Robinson, John, and John Fleishman. 1988. Ideological Identication: Trends and Interpretations of the Liberal-Conservative Balance. Public Opinion Quarterly 52(1):13445. Schuman, Howard, and Stanley Presser. 1981. Questions and Answers in Attitude Surveys. New York: Academic. Shafer, Byron E., and William J. M. Claggett. 1995. The Two Majorities: The Issue Context of American Politics. Johns Hopkins University Press. Stimson, James. 2004. Tides of Consent. New York: Cambridge. Stone, Walter. 1991. On Party Switching among Presidential Activists: What do We Know? American Journal of Political Science 35(3):598607. Treier, Shawn and Simon Jackman. 2008. Democracy as a Latent Variable. American Journal of Political Science 52(1): 201217.