Political Ideology and Issue Importance

Introduction

Researchers have long understood that the salience of issue attitudes varies considerably both across and within individuals (Howe and Krosnick 2017). While these differences have been leveraged to explain a number of discrete outcomes, ranging from attitude accessibility (Krosnick 1988) to candidate evaluation (Peterson 2004), we know far less about how variations in attitude salience affect composite measurements of public opinion such as latent measures of political ideology. In particular, because differences in issue salience constitute a form of differential survey item functioning, it may be important to account for issue salience in order to produce accurate estimates of political ideology.

The resurgent debate regarding the impact of ideology on Americans’ political behavior further motivates our consideration of issue salience in latent measurement (Fowler 2020). Many in the field view ideology as having little, if any, effect on citizens’ political behavior; indeed, recent studies find that ideology is only a weak predictor of vote choice and other preferences (e.g. Kinder and Kalmoe 2017). But such conclusions rest on measurement approaches that do not incorporate individual-level variations of issue salience into resulting estimates of ideology. As such, past research might understate the impact of political ideology on vote choice, representation, and other political outcomes of interest.

In this paper, we employ several different models for estimating latent ideology, each of which uses respondents’ rating of issue importance. Each approach incorporates information about the preferences and salience of each issue for a respondent, but incorporates salience information in vastly different ways. One approach uses salience as a weight in the estimation of latent ideology, another uses salience weighting to recode responses in a manner akin to rating scales, and a third employs salience weights to select meaningful responses for latent trait estimation. We assess the value-added from salience weighting across these diverse approaches using two survey data sets: (1) a subset of respondents to the 2016 Cooperative Congressional Election Study (CCES) and (2) a sample of respondents recruited by the online panelist recruitment firm Lucid. Throughout, we find that the salience-weighted ideological scales do not produce significantly better predictions of political outcomes of interest, providing evidence that accounting for issue salience does not improve commonly used scaling approaches for measuring ideology.

Accounting for Issue Importance in Political Ideology

Public opinion scholars have long recognized that while citizens hold multiple attitudes at the same time, not all attitudes hold equal weight across citizens. Some might care most deeply about their Second Amendment rights and care little about tax policy, while another might care deeply about tax policy while remaining relatively agnostic on the Second Amendment. Scholars often distinguish between objective or contextual salience—where an issue is receiving a great deal of attention from important actors in the political system—and subjective or individual salience—where an issue is particularly important to some people but less important to others (Lecheler, de Vreese, and Slothuus 2009). While these two dimensions are surely interrelated, we explore whether it is individual-level issue importance that may be an important missing piece from how individuals’ ideologies are currently estimated.

It is clear that individuals express far more concern for certain issues than others, relying on their most salient attitudes for decision-making (Howe and Krosnick 2017; Krosnick, Berent, and Boninger 1994). This concept—which we refer to as both attitude salience and issue importance (Ansolabehere and Puy 2018; Bernstein 1995; Boninger, Krosnick, and Berent 1995)—explains key features of political behavior, including political participation (Holbrook et al. 2016), information processing (Ciuk and Yost 2016), political learning (Nadeau, Niemi, and Amato 1995), attitude accessibility (Krosnick 1988), and attitude stability (Krosnick 1990). ¹ Of more debate is whether attitude salience is a significant predictor of vote choice. Peterson (2004) finds evidence that voters do rely on such attitudes in candidate evaluation; however, Leeper and Robison (2020) uncover little support for the idea that people are more likely to engage in “issue voting” on issues that they rate as being more important to them. Likewise, Lau, Andersen, and Redlawsk (2008) find that accounting for issue importance makes little difference for whether an individual is found to have voted correctly.

While the differential influence of issue attitudes has informed our understanding of many discrete outcomes, it remains less clear how variation in issue salience might affect the estimation of composite measures of public opinion, most notably the latent scaling of political ideology. Researchers often summarize a citizen’s political ideology by scaling a number of issue attitudes together with a latent variable technique (see, e.g., Laver 2014), producing an estimate that represents a citizen’s “ideal point” of issue preferences along a single dimension (Ansolabehere, Rodden, and Snyder 2008; Berinsky 2017; Erikson, MacKuen, and Stimson 2002; Jacoby 1995). ² In this paper, we explore variation in issue salience as a particular form of differential item functioning (DIF), which is characterized by items that exhibit different functioning along the dimension of latent ideology for different subgroups of respondents. In our area of interest, take for example an item that operates differently for men and women of similar ideological perspectives. Variation in issue salience across respondents constitutes another potential form of DIF, and one that if addressed could hypothetically improve latent estimates of political ideology. Critically, it is also one here that we do not detect or estimate, but rather is one that we know by virtue of survey items explicitly asking respondents to report how important the underlying topic of the question is to them.

Therefore, methodologically speaking, the incorporation of issue salience potentially provides a simple but important intervention for latent variable approaches. In standard latent variable models, particular items may be more strongly associated with (or discriminating of) the underlying dimension, thus contributing more to estimates of the latent trait among all respondents (Bollen 2002). Allowing some issues to contribute more information to the estimation of the population’s ideology is one way to account for objective or contextual salience. Issues that are attracting more attention from politicians and the news media may naturally provide more information about the public’s ideology. However, because standard models specify the operation of items at the global level, such approaches fail to address individual-level differences in issue importance. If some members of the public care quite deeply about the issue of gun control, then their views on that issue are likely to be more strongly driven by their underlying ideology. By contrast, when an issue is not particularly important to a group of people, their views on that issue may be less reflective of their ideology.

Existing work on individual-level issue importance provides some evidence for why we might expect this dynamic. For example, Lecheler, de Vreese, and Slothuus (2009) find that issue framing has no effect on attitudes on issues that are very important to individuals, but that the effects are much stronger on less important issues (see also Arceneaux 2008). Given that an important aspect of ideological thinking is issue constraint (Converse 1964), it is significant that people are more resistant to changing their views on issues that they identify as important to them. Additionally, Hill and Huber (2019) find that people express more confidence in the issue positions they take on issues that are more important to them. In short, these research studies suggest that when an issue is important to an individual, it is likely to provide consistent, clear, and confident information about the person’s ideology; by contrast, positions on less important issues are ephemeral, taken with less confidence, and therefore provide less clear signals about ideology. Thus, by masking this important variation between individuals, it is possible that existing approaches might produce less precise estimates of political ideology.

Data

We rely on two separate surveys for our analysis. First, we use the 2016 CCES common content file. ⁵ We use the 2016 CCES specifically because the survey questionnaire originally included a lengthy set of items which asked 13,244 individuals to rate how important different issues were to them. Respondents were asked to rate the importance of each issue area on a five-point scale ranging from “No Importance at All” to “Very High Importance.” While respondents were asked about the importance of fifteen different issue areas, we ultimately had accompanying issue questions for just eleven of these areas. Those issue areas are gun control, abortion, taxes, immigration, defense spending, the environment, jobs, crime, national security, health care, and gay marriage. The full details for the questions used from the CCES are available in the Supplemental Information.

Second, we fielded our own survey in 2018 on a sample of 1,508 respondents recruited using Lucid. Here, we used a different, novel approach to measuring issue importance. We did so in order to address potential concerns that simply rating the importance of issues was not particularly informative because respondents did not need to make trade-offs when providing these ratings (i.e., a CCES respondent could rank every issue as having “Very High Importance”). Thus, our 2018 survey first asked respondents to provide their positions on a set of nine issue questions across three areas (immigration, gun control, and abortion). Following that task, respondents were able to allocate one hundred points among each of their nine positions to indicate how important each position was to them. Details on this instrument are available in the Supplemental Information.

Standard

Consider first the standard IRT model for estimating ideology. Each respondent i ∈ { 1 , … , n } chooses whether or not to support a policy j ∈ { 1 , … , m } . If the respondent supports the policy, y i j = 1 , and if not, y i j = 0 . The probability of supporting the policy is considered a logit model of the following form:

P ( y i j = 1 ) = l o g i t − 1 ( α j + β j * θ i ) ,

where α j is the standard difficulty parameter, β j is the standard discrimination parameter, and θ i is the latent trait. ⁶

Weighted

In the standard model formulation, the difficulty parameter captures the point at which an item is equally likely to fall into either response category, while the discrimination parameter captures how well an item differentiates between the classifications. That is, both are relatively central from the perspective of educational testing, and indeed, the very names are evidence of the point. In the context of estimating ideology, however, neither captures variation in the differential weightings that citizens might have across different categories of questions.

Therefore, in our remaining three models, we explicitly incorporate the respondent rating of policy importance. For our weighted approach, we reformulate the basic model to incorporate the importance scales as follows. For each response y i j , we identify the relevant importance scale x i j . The model is then,

P ( y i j = 1 ) = l o g i t − 1 ( α j + θ i ( β j + x i j ) ) .

Here, the importance scales x i j are multiplicative on the latent ideology parameter θ i , magnifying the effect of the latent ideology parameter where x i j is large and conversely minimizing the effect where x i j is small. The approach, therefore, explicitly incorporates the importance scales as a weighting within the model estimation itself.

Scale

Our third approach builds from the well-established area of rating scale models (see, e.g., Andrich 1978). Rating scale models offer a generalization of binary response IRT models that feature just two response categories (here, support or oppose). When we know more than the binary response (e.g., strongly support, weakly support, neutral, weakly oppose, or strongly oppose), the additional information can aid estimation of the underlying latent variable. Here, by treating the importance scales as indicative of how strongly one feels about their support or opposition of a policy, we rescale the responses to a space that is more indicative of the respondent’s relative preferences on the issue.

Therefore, for this model, we weight the responses rather than the estimates themselves. To do so, we incorporate the importance scales by recoding the responses y to {−1, 1}, then multiplying the responses by the importance scales. The rating scale model follows Andrich (1978) and Furr (2019) and is estimated as,

P ( y i j = 1 , y > 0 ) = e x p ∑ s = 1 y ( θ i − α j − K s ) 1 + ∑ k = 1 m e x p ∑ s = 1 k ( β j θ i − α j − K s ) ,

P ( y i j = 1 , y = 0 ) = 1 1 + ∑ k = 1 m e x p ∑ s = 1 k ( β j θ i − α j − K s ) .

In the rating scale model, m is the total number of steps, and K s represents the difficulty K for step s . Each of the remaining parameters remains identical to the previous discussions.

It bears reiterating that, for this approach, we explicitly define the extremity of positions by rescaling the responses themselves. This is a relatively heavy-handed approach to incorporating importance scales as it fundamentally reshapes and redefines the observations that are explained by the latent concept.

Moderated

Finally, our fourth approach instead uses importance scales to choose the set of responses to employ in estimating latent ideology. The underlying intuition here is that only those items which a respondent actively cares about are indicative of latent ideology. This model, then, comes closest to the attitude salience perspective outlined above; citizens are concerned with one or a few sets of relevant policies, and those preferences are the guiding lights for their decision-making. The model incorporates this thinking by relying only on those attitudes which are—by virtue of the importance scales—identified as important by the respondent.

To wit, we take as our launching point effort-moderated item response models (e.g., Wise and DeMars 2006), and estimate a model that effectively removes the influence of responses for which the respondent rated the importance of the policy less than the median importance rating for that policy among all survey respondents.

The form of the moderated model is,

P ( y i j = 1 ) = l o g i t − 1 ( b i j * ( α j + β j * θ i ) + ( 1 − b i j ) * 0 . 5 ) .

Here, b i j is an indicator, taking the value of one where the respondent rates the policy area of the question at or above the median value of importance for that respondent, and the value of zero when the respondent rates the policy area below the median importance rating for that respondent. Where the respondent rates the policy as important, their responses default to the standard item response model as above, whereas if the respondent rates the policy below the median (i.e., b i j = 0 ), the model defaults to uninformative guessing. ⁷

We use the median to address differences across issue areas in the average rating of issue importance. The median is particularly appropriate here, as it best allows us to identify relative ratings of importance within a category while addressing trade-offs in data selection. Specifically, a shift in either direction would carry costs; requiring even higher ratings of importance would decrease the amount of data we have with which to estimate the latent dimension—particularly pernicious where researchers do not have a robust surplus of respondents and questions, the usual setting for estimates of citizen ideology—while setting a lower threshold for inclusion would decrease the signal one can gain from incorporating the rating of importance.

In all, our three models incorporating respondent importance scales take methodologically diverse approaches that reflect the theoretical diversity of ways in which importance could come to influence the estimation of citizen ideology. In the weighted model, we weight the estimates themselves as part of the estimation procedure. In the scale model, we weight the responses, rescaling that responses on policies identified as important by the respondent are treated as more extreme than policies that are not identified as important by the respondent. Finally, the moderated model weights the responses, holding that only responses identified as important are useful to estimation of the latent dimension.

Results with CCES

We begin with a comparison of standard and salience-weighted ideal point estimates derived from the CCES data. In Figure 1, we provide a variety of metrics by which to compare the approaches. On the diagonal, we plot the distributions of estimates of ideology for each measure, divided by the self-identified party identification of the respondent. ⁸ Each approach partitions Republican and Democrat respondents as would be expected, and likewise for “Other” respondents.

OPEN IN VIEWER

Figure 1.

Comparison of latent ideology measures (CCES data).

We turn then to understanding the differences in individual estimates. Under the diagonal, we include comparative scatterplots of estimates for each model combinatorial. Throughout the six plots, it is strikingly evident that the measures are highly correlated. In fact, Pearson’s correlation coefficients for each of the combinatorials are never less than .95 overall. Above the diagonal, we include the correlation coefficients for each of the model comparisons. Digging deeper, we do find some minimal evidence of differences if one examines correlations among subgroups of respondents based on party identification. While each of the weighting approaches is highly correlated with other weighting approaches, the lowest correlations appear between estimates among Democratic respondents according to the rating scale approach. This is most evident in the second column of Figure 1, where the Democratic respondents (blue dots) are more dispersed than those in any other plot. On initial inspection, then, we see little evidence to suggest incorporating importance scales has critical ramifications for the estimation of ideology, though there is some suggestion one particular approach—rating scale weighting—may offer very slight changes to estimates among Democratic respondents.

One of our primary areas of interest is in whether the incorporation of salience might suggest differences in estimates of ideological polarization. To explore whether approaches incorporating issue salience better capture polarization in the ideological space of American voters, for each distribution of ideal points, we calculate the bimodality coefficient (see, e.g., Lelkes 2016). This measure captures the extent to which a distribution is unimodal or bimodal, and is estimated as a function of the distribution’s size, skewness, and excess kurtosis. Larger values of the bimodality coefficient indicate more bimodal data. Here, we observe values of 0.44 for the standard model, 0.38 for the weighted model, 0.44 for the moderated model, and 0.46 for the scale model. Though we observe some minimal variance in the estimates across approaches, the magnitude of the differences is substantively minor, and approaches incorporating issue salience do not systematically overestimate or underestimate polarization relative to the standard approach. In all, we find little evidence to suggest issue salience shifts our understanding of ideological polarization in the electorate.

While these descriptive results suggest our salience-weighted approach offers little improvement in the estimation of latent ideology, we can better evaluate the merits of the approaches by assessing their performance on prediction tasks. Thus, we turn now to analyzing whether the slight difference yields meaningful improvements on two predictive tasks. The first task we considered was predicting responses to a question from the postelection wave of the survey which asks respondents to state their preference for raising taxes versus cutting spending on a zero to one hundred scale. ⁹ In Figure 2, we plot the relevant estimate of ideology (x-axis) against deficit preferences (y-axis) for each of the four models. Under each plot, we include Pearson’s correlation coefficient, and the R 2 from a linear model where deficit preference is the dependent variable and the relevant ideological estimate is the only predictor variable.

OPEN IN VIEWER

Figure 2.

Relationship between ideal point estimates and deficit preferences.

Across both metrics, we find that the standard approach slightly outperforms each of the weighting approaches. Of the group of weighting approaches, the weighted approach matches but does not exceed the performance of the standard approach, with ρ = 0 . 66 for both the weighted and standard approaches, and R 2 = . 43 for each as well. Though the weighted approach matches the results of the standard approach, we find it does not yield improvements in the estimation of ideology at least for predicting deficit preferences, and other approaches (scale and moderated) both underperform relative to the standard approach.

As a second evaluative method, we use our measures to predict electoral choices. ¹⁰ Since general election vote choice is largely determined by partisanship, we turn to the presidential primaries, where voters might rely more on ideology to choose between candidates. For example, the 2016 Democratic nomination featured socialist Bernie Sanders competing against the more moderate Hillary Clinton. On the Republican side, the top three vote-getters included one of the most conservative senators (Ted Cruz), the more moderate John Kasich, and the ideologically enigmatic Donald Trump. Despite some clear ideological distinctions between the candidates, traditional measures of ideology were less powerful in predicting primary vote choices in 2016 than they had been in previous election cycles (Sides, Tesler, and Vavreck 2018). While much of the limited influence of ideology has been attributed to the unusually intense focus on identity issues, we also consider the possibility that the traditional ideology scales may be less useful in capturing intraparty ideological divisions.

In the 2016 CCES, respondents were asked if they voted in the 2016 primary and then, if they had voted, which candidate they supported. To assess the relative performance of the latent ideology measures, we employ each to predict primary vote choice. We create two subsets of the data: those individuals who report voting for a Democratic candidate and those individuals who report voting for a Republican candidate. For each separate model, we limit response options to the major candidates and an “other party candidate” category. Given the multiple categorical outcomes for both Democratic and Republican respondents, we estimate multinomial logit models of vote choice. The approach is a more conservative approach than ordered logit in case of violations of the proportional odds (or parallel regressions) assumption of ordinal logit.

We begin with Democratic vote choice. Our sample is limited to 4,721 respondents who identified a Democratic candidate as their primary vote choice. Here, the standard and salience-weighted models perform almost universally identically. Importantly, most models disproportionately assign votes to Hilary Clinton; in fact, the standard model assigns all respondents to Clinton because of the association of higher probability of Clinton votes across the entire spectrum of estimates of ideology, while the rating and moderated models predict nearly all respondents as Clinton votes. This is evident in Figure 3, where we plot the predicted probability of voting for each of the three potential outcomes (Clinton, Sanders, or Other) across the majority of the range of measures of ideology, with the x-axis limited to the mean of the ideological measure plus and minus two standard deviations.

OPEN IN VIEWER

Figure 3.

Predicted probabilities from multinomial logit of primary vote choice among self-identified Democrats.

Relative to the other models, the weighted model attaches slightly higher probabilities for very liberal Democrats to vote for Bernie Sanders. The result is that—unlike in other specifications—365 respondents are classified as Sanders voters. Of these, 234 are accurately classified as Sanders voters, but 131 are actually Clinton voters. Thus, the incorporation of importance provides only a marginal benefit, with very minimal improvement in predictive accuracy (2.2%) in one of the three alternative specifications.

Turning next to Republican respondents (Figure 4), our sample is limited to 4,289 respondents who report a primary vote choice of a Republican candidate. Here, the standard model actually performs slightly better than each of the alternative approaches, correctly predicting an additional four respondents as compared to the weighted model, an additional nineteen as compared to the rating scale model, and an additional sixteen as compared to the moderated model. Though the overall accuracy is again similar, there are differences in the priority given to different candidates. Relative to the other models, the weighted model predicts the most Trump voters (thus its success relative to other approaches), and fewer Kasich voters. On the other hand, the moderated model predicts the most Cruz voters, and the fewest Kasich voters. None of the models predicts a Rubio voter. Throughout, individual increases in predicting votes for candidates who ultimately lost (i.e., Cruz and Kasich) lead to improvements in the prediction of voters within those categories, but at the cost of inaccurately predicting an increased number of Trump voters.

OPEN IN VIEWER

Figure 4.

Predicted probabilities from multinomial logit of primary vote choice among self-identified Republicans.

Stepping back, we see that as was the case with the prediction of deficit reduction preferences, incorporating issue salience into measures of ideology offers little consistent improvement in the prediction of primary vote choice. Though estimates diverge in potentially interesting ways, they hold nearly identical predictive power in this series of predictive tasks. Incorporating salience in approaches to estimating ideology is a laborious and costly one for such marginal changes in the prediction of attitudinal and behavioral outcomes.

Results with Lucid

To check the robustness of the above, in our Lucid survey, we employ a novel approach for capturing respondent importance weighting. This approach instead treats importance as compositional, forcing respondents to allocate among nine responses a finite amount of “points” which roughly corresponds to the probability of their preferred position becoming policy. Thus, we explicitly force respondents to consider trade-offs in signaling an issue’s importance. We again estimate standard and salience-weighted ideal point models following the approach outlined above. Models are identical to the above, except we rescale the compositional measure to a one-to-five Likert scale in order to make the rating scale model approach tractable. We plot the results in Figure 5.

OPEN IN VIEWER

Figure 5.

Comparison of latent ideology measures (Lucid data).

Despite the different approach to measuring importance, the ideal point estimates remain highly correlated, with Pearson’s correlation coefficients of .93, .99, and .94 between the standard and weighted, rating scale, and moderated approaches, respectively. Likewise, the distribution of ideal point densities, plotted in Figure 2, closely resembles Figure 1, as in each case, the distributions are quite similar between estimation approaches. Unlike before, we do not however see the within-subgroup differences, as Democrats and Republicans are approximately equally correlated across each of the comparisons.

We again estimate the bimodality coefficient to explore whether the measures suggest different rates of partisan ideological polarization. Recall that larger values of the bimodality coefficient indicate more bimodal data. Here, we find values of 0.40 for the standard model, 0.41 for the weighted model, 0.41 for the moderated model, and 0.51 for the rating scale model. Overall, we find little variance, with the notable exception of the rating scale model; yet while the bimodality coefficient for the rating scale model is relatively more distant from other model coefficients, the extent is much different than observed in the CCES case, and again does little provide consistent evidence of improvement.

Still, to see if the marginal differences in estimates from our approaches yield improvements in our ability to estimate ideology or to better capture ideological polarization, we again evaluate the value-added for the salience-weighted approaches by turning to predictive tasks. Here, survey respondents were asked three items which asked them to rate their preferences on a five-point scale for (1) tariffs on imported steel, (2) increased federal funding for charter schools, and (3) mandatory minimum prison sentences of five years for drug offenses. For each of the questions, we estimate separate multinomial logit models using our standard and weighted ideal point estimates.

We plot predicted probabilities for each of the analyses in Figure 6. As the figure makes clear, there are no large differences across any of the approaches, and there is no systematic improvement from using importance weighting across any of the approaches as compared to the standard approach. Starting with the top row, which features support for tariffs on imported steel, the standard approach offers a small (0.1%) improvement above the rating scale approach, but both the weighted and moderated approaches perform slightly worse (another 0.1%). In the second row, support for increased federal funding, each of the approaches incorporating importance performs better than the standard approach (0.1%–0.3%). Finally, in the third row, we see that the weighted model does better (1.0%) at predicting support for mandatory minimum sentences of five years for drug offenses, while the moderated approach performs much worse (5.3%) compared to the standard approach.

OPEN IN VIEWER

Figure 6.

Predicted probabilities from ordered logit of policy preferences.

As before, we find little evidence to suggest the incorporation of issue importance substantially and consistently improves our ability to capture ideological preferences. Throughout the series of predictive tasks and across two different surveys and approaches to measuring importance, we find no systematic evidence that incorporating issue importance improves our ability to measure the ideological preferences of survey respondents in one dimension.

Discussion and Conclusion

In this paper, we compared a standard latent trait modeling approach to estimating ideology to a series of approaches for estimating latent ideology in a way that incorporates issue salience as a weighting strategy. Specifically, we sought to account for individual-level differences in issue salience as a means of weighting how certain items contribute to the estimation of an underlying ideology. We compared the utility of these strategies to one another and to standard IRT models in explaining ideological polarization in the electorate, both in descriptive terms and as indicated by relevant behavioral outcomes. Our approach builds upon previous work that accounts for the differential functioning of survey items in the estimation of ideology (e.g. Hare et al. 2015), but does so by weighting how items contribute to the estimate rather than standardizing them.

While our salience-weighted approaches produce some minimal differences, the latent measures produced from these disparate approaches are highly correlated with a latent measure built purely on scaling issue positions. Additionally, in a series of prediction tasks—prediction of deficit reduction preferences, primary vote choice, and policy preferences—we find no systematic evidence of improvements over the standard IRT model. We find this to be true with both standard importance weighting survey questions and a new compositional measure of importance derived from a survey instrument requiring respondents to weight their importance preferences. Across surveys and predictive tasks, we find little evidence to suggest that accounting for issue importance offers substantive improvements in our ability to estimate ideology. In this way, our findings are consistent with recent work by Leeper and Robison (2020), who find that individuals are not more likely to engage in issue voting when it comes to issues that they rate as more important to them. That personal issue salience does not send a stronger signal about one’s underlying ideology fits with a pattern whereby people are not significantly more likely to use important issues (versus unimportant ones) to make decisions between competing candidates.

Overall, our results have important theoretical and empirical implications. In the former case, researchers should consider exploring the theoretical basis of issue importance items. Does the fact that accounting for issue salience adds little, if any, information to the scaling of issues reflect something about the current polarized nature of politics in the United States, or does it say something more broadly applicable about how citizens formulate and hold their issue positions? One possibility seems particularly plausible to us. Americans may choose which party to support based on a small number of issues that they care deeply about; however, they then adapt their positions on other (less important) issues in order to coincide with that party’s platform (Barber and Pope 2019). In this way, issue importance may fail to provide us with much additional information about ideology simply because the positions that most people take on less salient issues are highly consistent with their views on issues they care most about. Issue importance clearly matters in this process—but not in a way that would be obvious from scaling a large number of salient and nonsalient issue items.

In terms of empirical implications, our work makes clear that scaling issue preference items offers a robust indicator of ideology, regardless of how salient those issues are to respondents. Looking ahead, one important future avenue for scholars to explore is how this issue salience manifests in models that treat ideology as multidimensional rather than unidimensional. In all, though, the conclusions that have been drawn by scholars related to the role of ideology in predicting vote choices and representation would not appear to be impacted by models that account for personal issue salience, nor do we find systematic evidence of more polarization among Americans when we make such accommodations.

Supplemental Material