Partisan Stability During Turbulent Times: Evidence from Three American Panel Surveys

Recent years have witnessed growing partisan division. The phenomenon of affective polarization, whereby partisans increasingly dislike their partisan rivals, seems to have accelerated after 2000 (Iyengar, 2022). According to Mason (2018), partisans themselves have become more “sorted” insofar as Democrats increasingly identify as liberal and Republicans increasingly identify as conservative, a pattern she argues amplifies these groups’ emotional investment in politics. At the same time, opposing partisans have become more geographically (Brown & Enos, 2021) and socially (Webster et al., 2022) segregated, a pattern that may facilitate the transmission and expression of partisan attachments. Whether these underlying trends in mass partisanship reflect elite cues or create incentives for partisan invective among party leaders, public officials of both parties have increasingly expressed their contempt for their partisan opponents (Fowler et al., 2016; Parker, 2020). The current political environment, described as “turbulent” by contemporary historians (Zelizer, 2022), makes the party politics of the Reagan era seem genteel.

At the same time, the parties themselves have changed. The advent of Donald Trump brought about profound changes in the Republican Party’s platform (Rubin, 2016), and his pugnacious nativism and disdain for political convention attracted new support among non-college educated Whites (Vavreck, 2020). Reflecting on the changes he brought about, Trump commented that “The party is a much bigger party now, and it’s like really a party for the working men and women in this country, in addition to plenty of others...In the true sense, it has been changed” (Bennett, 2018). Trump’s transformation of the GOP has become increasingly prominent in academic commentary. Gary Jacobson, for example, argues that “By projecting a sharp, highly divisive image of who and what the Republican Party stands for and, equally important, who and what it stands against, the Trump presidency is poised to have a durable impact on party images and identities, especially among younger citizens entering the electorate for the first time” (Jacobson, 2018). On the Democratic side, the Progressive wing’s policy proposals and leaders rose to new prominence (Raphael and Todd, 2016) with the presidential primary bid of Bernie Sanders in 2016 and the meteoric rise of Alexandria Ocasio-Cortez in 2018. As Zingher (2018) notes, “In the aggregate, white citizens have viewed the Democratic Party as moving progressively further and further away from their own position. An increasing majority of whites hold the perception that the Democratic Party has followed the shifting median on the economic dimension to the left” (p. 870).

The changing issue stances and social imagery of the parties raise the question of whether events of recent years have led voters to rethink their fit with the parties. Has the tumult of contemporary party politics hastened the pace with which party identities change, as disaffected issue voters gravitate away from their former party? Or, conversely, have the new features of party politics contributed to more stable party ties that are buttressed by more ardent emotions and perceived social distance between warring partisan camps?

This paper revisits the decades-old research literature that uses multi-wave panel surveys to assess the extent to which respondents’ party attachments change over time. Studies of this kind date back to The American Voter (Campbell et al., 1960) but gained methodological sophistication in the wake of Asher (1974) and Achen (1975), which called attention to the ways in which response error may exaggerate the actual rate of partisan change. One corrective has been to use instrumental variables regression or similar maximum likelihood estimators to distinguish observed change from true change in latent party identification (Palmquist & Green, 1992). For example, Green and Palmquist (1994) apply measurement error models to nine multi-wave panel studies from the United States, and Schickler and Green (1997) do so for eight panel studies from outside the US. Their results, which show party identification to be highly stable but by no means immutable, provide a benchmark for comparison. However, modeling approaches of this kind are not without their critics, and other approaches have gained prominence as well. One such approach is to create additive indices based on multiple measures of an underlying factor and to assess the over-time correlations of these more reliable indices (Ansolabehere et al., 2008). More recent works, such as Tucker et al. (2019), offer modeling innovations of their own, capitalizing on the fact that internet panels often feature many more waves of interviews than traditional face-to-face surveys. To our knowledge, however, no one has attempted to pull together an assortment of contemporary panel surveys in order to assess whether methodologically robust conclusions may be drawn about the pace of partisan change in the current partisan environment.

Our essay is organized as follows. We begin by reviewing leading theories about the conditions under which party attachments change over time, connecting each to scholarly debates about the political implications of demographic, technological, and institutional change. By our reading, recent trends have ambiguous implications for partisan stability, as some theories suggest the potential for instability due to changes in party positions on salient issues, while others suggest the potential for stability on the grounds that the parties are more socially distinctive than before. Next, we introduce three multi-wave panel surveys that collectively cover the period 2011 to 2020. Before attempting to model patterns of partisan change, we provide both quantitative and graphical descriptions of how stated party attachments evolve over successive interviews. Measurement models are introduced and applied to the three panel surveys, and these results are cross-validated using other approaches, namely, the creation of multi-item indices and the use of regression to estimate the over-time trajectories of individual respondents. A consistent picture emerges from all of the approaches. Perhaps surprisingly, party attachments over the past decade have changed very gradually, and party identification in contemporary panel studies appears to be at least as stable as it was in the American National Election Studies panel surveys dating back to the 1950s. We conclude by spelling out the implications of these findings for long-term changes in partisanship. Although substantial changes in party identification are rare over the course of a single election campaign or presidential term, meaningful change is common over the span of a voter’s lifetime.

Although party attachments in the United States have long been characterized as unusually stable compared to other political attitudes (Converse, 1964), the importance of party identification as an explanatory construct has led generations of public opinion scholars to offer hypotheses about the conditions under which party attachments change. Propositions about why change may be afoot in the current era trace their roots to longstanding arguments that fall roughly into three theoretical categories.

The first concerns the “spatial” proximity between the voter and the parties’ positions on leading issues. This perspective (Franklin and Jackson, 1983; Jackson, 1975; Kollman & Jackson, 2021) contends that when new issues become prominent or when parties change their platforms, voters gravitate toward the party that is more congenial to their policy views. This theory directs our attention to two sources of change, the repositioning of the parties vis-à-vis issues and the ways that technological change may have drawn public attention to specific issues and where the parties stand on them.

There are good reasons to suppose that issue-based evaluations of the parties have grown increasingly important over time. The two major parties present far more ideologically distinctive “brands” than in decades past (Butler & Powell, 2014), and the ranks of prominent elected officials feature a dwindling number of conservative Democrats or liberal Republicans (Fiorina, 2017). Although public opinion scholars disagree about the extent to which the public has sorted itself into ideologically distinctive partisan camps (Abramowitz, 2018; Levendusky, 2009), as opposed to updating their issue stances to fall into line with the positions staked out by party leaders (Barber & Pope, 2019; Lenz, 2012), it seems clear that the correlation between party and self-described liberalism/conservatism has climbed over time (Bafumi & Shapiro, 2009). At the same time, the parties have staked out increasingly divergent positions on issues that were largely orthogonal to party prior to the 1990s, such as gun rights (Joslyn et al., 2017) and immigration (Fennelly et al., 2015), and the same may be said of high-salience moral issues that are sufficiently resonant with the public to upend their party attachments (Goren & Chapp, 2017). As noted above, the Trump era is noteworthy for its changes in Republican Party stances, and much the same could be said for the increasingly visible Progressive wing of the Democratic Party. For authors such as Kollman and Jackson (2021, Chap. 7), such shifts in issue stances have the potential to profoundly alter the balance of party identification.

On the other hand, public opinion scholars have long expressed skepticism about whether the public is sufficiently knowledgeable and concerned about issues to use spatial proximity when evaluating candidates (Stokes, 1963) or parties (Converse, 1964). Although this critique continues to find support in contemporary surveys (Kinder & Kalmoe, 2017), another school of thought contends that a growing segment of the American public does care about issues (Bullock, 2011; Carsey & Layman, 2006; Mummolo et al., 2021), and their ability to connect their issue stances to their party evaluations has been aided by the advent of ideologically polarized television networks (DellaVigna & Kaplan, 2007), social media (Cacciatore et al., 2016), and permissive campaign finance rules (Brooks & Murov, 2012) that together have greatly increased the volume of communication that voters receive about issues. On balance, the issue proximity hypothesis seems to imply that party affiliations should have changed at a faster-than-usual pace between 2011 to 2020.

Another prominent theory has to do with the social imagery of the two parties. What kinds of people come to mind when one thinks about Democrats or Republicans? The prodigious literature on social stereotypes and group identities (Green et al., 2002; Huddy et al., 2015) also finds expression in two competing hypotheses, one that emphasizes the potential for change and another, the hardening of existing party attachments. The former focuses on White identity politics amid Whites’ growing concern about ethnic change (Craig & Richeson, 2014; Jardina, 2019; Major et al., 2018). White ethnic appeals may be seen as a force that pulled Whites in the direction of the GOP during this period, especially after Obama’s election notified White voters that the Democratic Party is preferred by and supportive of Black Americans (Tesler & Sears, 2010). However, recent experimental studies call into question whether interventions that prime White ethnic concerns in fact make them more likely to embrace the Republican Party (McCarthy, 2022). By the same token, the widely-publicized criticisms of undocumented immigrants from Mexico by Trump and prominent GOP figures might have led to increasing Democratic attachment among Hispanics, although a series of experimental tests by Hopkins et al. (2020) suggest that such effects on partisan attachments are surprisingly weak. Perhaps driven by the normalization of “uncivil discourse” (Mutz, 2015) between politicians from warring camps, recent years have seen a surge in “affective polarization” by which partisans increasingly dislike one another (Druckman & Jeremy, 2022). Coupled with increasing residential and social segregation by party, the tenor of current politics may have helped to harden existing partisan social divisions. On balance, the social imagery hypothesis seems to suggest a slower-than-usual pace of partisan change during the 2011–2020 period.

Finally, performance evaluations have long figured prominently as explanations of individual-level (Fiorina, 1981) and aggregate-level (MacKuen et al., 1989) partisan change. The economic swings of the past decade, combined with a media landscape that accentuates economic resentment (Soroka, 2014), have led many to speculate about the extent to which disaffection with the one or both parties reflects the way in which they are blamed for the loss of jobs, stagnant wages, and rising personal debt. This line of argument is also used to explain the growing regional partisan divide between the more economically vibrant coasts and the country’s interior (Hopkins, 2017). The public’s favorable assessment of the national economy under Trump (Small & Eisinger, 2020) would have been expected to attract new Republican partisans, at least until the COVID-19 epidemic led to an abrupt economic contraction. However, Trump’s relatively low approval ratings throughout his term offset the partisan gains that might ordinarily follow from an economic surge. In the end, the performance records of neither Obama nor Trump are sufficiently distinctive to imply faster-than-usual rates of partisan change.

Taken together, these three theoretical perspectives offer competing empirical predictions. The character and pace of what Carmines and Stimson (1989) termed “issue evolution” militates in favor of a quickened pace of partisan change; on the other hand, the widening social divide between the two parties arguably makes partisans more resistant to change, while the vicissitudes of presidential performance seem to imply no clear prediction. We therefore consider the core hypothesis to be two-sided, given the strong theoretical reasons to suspect issue-driven change or affect-driven stability.

To our knowledge, there are three publicly available multi-wave panel surveys of American adults spanning the Obama to Trump administrations that repeatedly measured party identification: the Institute for the Study of Citizens and Politics (ISCAP) survey,¹ Views of the Electorate Research Survey conducted by the Democracy Fund Voter Study Group (VSG),² and The American Panel Survey (TAPS).³ The three panels started and finished at somewhat different times but collectively spanned November 2011 to October 2020.

Each of the three panel surveys utilized sampling procedures designed to recruit a nationally representative set of respondents. ISCAP panelists were recruited offline via address-based sampling or random-digit dialing, and panel demographics closely mirrored the adult population of the United States. VSG respondents were recruited according to a stratified sample that matched US population benchmarks. TAPS respondents were recruited via address-based sampling and drew upon a number of publicly available datasets, including US Census files, the White Pages, and credit agency data, to sample in proportion to US demographic benchmarks. In addition, all three panels employed survey weights after each survey wave to improve sample representativeness.

Each panel survey repeatedly measured party identification on a seven-point scale (PID-7) using the American National Election Study (ANES) wording, with slight variations.⁴ The main differences across panels have to do with the number of times respondents were interviewed. The ISCAP panel measured PID-7 on nine occasions,⁵ VSG six, and TAPS twenty-four.⁶ For compactness and comparability, many of the analyses below will draw from exactly six waves for each panel. For VSG, this represents the entire panel. For ISCAP and TAPS, this represents the first and last wave and four waves selected at regular intervals in between. (Analyses of the full ISCAP and TAPS panels are provided in the Online Appendix.) See Table 1 for the survey dates for each of the six waves selected for analysis from each panel; for the full ISCAP and TAPS panel wave dates, see Online Appendix Tables A1 and A2, respectively.

The panels also varied greatly in the number of respondents they interviewed; the right-hand column of Table 1 presents the number of respondents who completed the PID-7 survey items in each wave. VSG recruited considerably more respondents (often nearing 8000) at each survey wave than ISCAP or TAPS, which each averaged between one and two thousand respondents per wave. Due to a combination of budget limitations and respondent fatigue, only 365 ISCAP respondents, 4013 VSG respondents, and 445 TAPS respondents provided PID-7 survey responses for each of their panel’s waves (we refer to these respondents as “complete cases”). We consider the implications of panel attrition below. The rates of attrition undermine the justification for using sample weights; for simplicity, we present unweighted results, although as shown in Online Appendix Table A16 and Online Appendix Fig. A1, weighted estimates tend to be similar.

Although raw over-time correlations suggest that party identification is stable by comparison to most other expressions of political opinions, these correlations still suggest that party identification is subject to change. For example, a correlation of 0.95 between two successive survey waves implies that \(0.95^2=0.9025\) or 90% of the variance in \(wave_t\) is predicted by \(wave_{t-1}\). On the other hand, a correlation of 0.98 between two successive survey waves implies that 96% of the variance in \(wave_t\) is predicted by \(wave_{t-1}\), which suggests a much slower pace of change. In other words, when correlations are high, subtle differences in the magnitude of these correlations make a difference for the characterization of partisan stability. For this reason, methodological choices about whether and how to “disattenuate” observed correlations that are arguably distorted by measurement error can profoundly affect substantive conclusions about the pace of partisan change.

Corrections for measurement error require modeling assumptions. When evaluating these models, it is important to attend both to the plausibility of the assumptions and the statistical consequences of violating them. In this section, we review the assumptions of the most widely used method of correcting for measurement error, which is rooted in instrumental variables (IV) estimation. An alternative approach (Green & Schickler, 1993; Ansolabehere et al., 2008) is to build multi-item scales, which are deemed sufficiently reliable to require no further statistical adjustment. We review the assumptions of both approaches. In the next section, we show that they produce similar results.

We begin by considering how models that are prone to errors-in-variables bias compare to instrumental variables models. Table 5 reports the results of OLS regressions in which PID-7 for each wave is regressed on the same measure in the preceding wave.¹² Across the three panels, OLS slope estimates range from 0.889 to 0.992, and the corresponding \(R^2\) values range from 0.79 to 0.98. The implication is that 2% to 20% of the variance in partisanship is “fresh” from wave to wave. Instrumental variables regression reports larger slopes, ranging from 0.93 to 1.01; the reported \(R^2\) values, however, are essentially unchanged because IV conventionally reports the squared correlation between the dependent variable and the observed regressor times the estimated slope. Since OLS maximizes the \(R^2\), the IV estimates slightly reduce \(R^2\). In order to disattenuate the \(R^2\) in a manner that accounts for measurement error in both the independent and dependent variables, we apply the Wiley–Wiley estimator for measurement error in waves 2 and 3 of a four-wave panel, as explained in footnote 11. This correction alters the implied rate of partisan change. Focusing on results for the second wave through the fifth wave out of the six selected, we find that the average \(R^2\) rises from 0.894 before correction for measurement error to 0.953 after correction for measurement error.¹³

Tables 6, 7, and 8 repeat this analysis, focusing on temporally overlapping sets of survey waves across the three panels. For example, Table 6 showcases the results from four survey waves fielded nearly contemporaneously in both the ISCAP and TAPS panels. Despite the somewhat disparate \(R^2\) resulting from the estimated relationship between 2012 and 2015 PID-7 responses before correcting for measurement error (0.882 versus 0.830), ISCAP and TAPS panelists present much more similar—and higher—\(R^2\) after the Wiley–Wiley correction is administered (0.948 versus 0.933, respectively). Similarly, the comparison between VSG and TAPS respondents in Table 7 shows converging \(R^2\); approximately a four percentage point difference prior to correction (0.824 versus 0.783) was reduced to less than one after both estimates of \(R^2\) increased (0.874 versus 0.881, respectively).

To summarize the results, Fig. 4 shows how the passage of time between waves (the horizontal axis, on a log scale) diminishes the disattenuated \(R^2\) with which party identification at one point in time predicts party identification in a subsequent interview (the vertical axis, also on a log scale). The solid line shows the fitted regression function using the estimates from the three panel studies, while the dashed line shows the fitted regression function for the estimates of six nationally representative panel surveys fielded between 1956 and 1992, which Green and Palmquist (1994) previously used to characterize the annual rate at which \(R^2\) declines.¹⁴ Although the older data sets generate a more dispersed set of estimates and a somewhat steeper sloping line, an F-test reveals no statistically significant difference between the two lines. If anything, the slightly flatter solid line suggests that the rate of change was even more gradual during the 2011–2020 period.

Although much has changed in American politics during the decades since the path-breaking 1956–1960 ANES panel survey, the stability of party identification during the Obama-Trump era looks very much as it did during the Eisenhower Administration or, for that matter, during the eras encompassing Vietnam, Watergate, Stagflation, the Iran Hostage Crisis, the Reagan Ascendancy, the Iran-Contra Scandal, and the Persian Gulf War (Green and Palmquist, 1994).

During the 2011–2020 period, raw correlations between party identification scores measured in successive waves of panel interviews tend to be quite high by comparison to most political attitudes. They are higher still when corrections are made for measurement error. Whether these corrections derive from instrumental variables regression or down-to-earth approaches such as index creation, disattenuated correlations imply that party identification changes at a glacial pace.

The same picture emerges from other ways of describing partisan change statistically. Individual-level response variation is relatively rare across panel waves, a pattern affirmed by other recent studies of multi-wave panel studies, most notably Tucker et al. (2019), who analyze twenty-waves of the TAPS panel from 2011 to 2016. They find that shocks at the individual level dissipate quickly; a shock that moves party identification 0.21 scale points in one wave has an effect of just 0.04 scale points 4 months later and just 0.01 eight months later. When we track individual-level partisan trajectories using all three panel datasets, we too find that a small portion of the public experiences durable change, even in turbulent political times. Nor do we see evidence of aggregate party change, whether we track panel respondents over time or examine independent cross-sectional surveys conducted by the Gallup Poll.

Looking back at the dominant theoretical perspectives that are used to explain change or stability in party identification, it seems that our results underscore the importance of deepening social divides. Our initial hypothesis was two-sided in the sense that the stabilizing effects of growing affective polarization and residential segregation could have been overshadowed by the destabilizing effects of changing party issue positions, the emergence of new issues that divide the parties, and new communication technologies that accentuate those divisions. The fact that party identification seems at least as stable now as it did when the parties were less ideologically distinctive and mercurial vindicates a central argument in Campbell et al. (1960), namely, that party attachment is not primarily driven by ideological affinity. We are quick to concede, however, that this conclusion is not rooted in a direct test of individual-level responsiveness to perceived party stances, a test that presents a host of methodological challenges when using non-experimental panel surveys (Lenz, 2012; Green and Palmquist, 1990).

Although stability over time remains a key feature of American party attachments, we conclude by calling attention to the crucial distinction between slow change and none at all. For those who study elections using cross-sectional survey data, the results presented here are reassuring insofar as they suggest that the pace at which partisanship changes is too slow to be consequential during a given election season. At the same time, the caricature of party identification as an “unmoved mover” creates a host of empirical anomalies that become apparent when researchers track partisan attachments over decades and find substantively large and sustained movements (cf. Kollman and Jackson 2021, Chapter 4). To be empirically sustainable, theoretical accounts must explain why party attachments resist change as well as why meaningful changes do occur over voters’ lifetimes.