Abstract
Scholars have identified many social-psychological factors correlated with support for Donald Trump; however, attempts at modeling these factors tend to suffer from omitted variable bias on the one hand, or multicollinearity on the other. Both issues obscure inferences. Using two nationally representative surveys, we demonstrate the perils of including or failing to include many of these factors in models of Trump support. We then reconceptualize the social-psychological sources of Trump support as components of a broader “profile” of factors that explains Trump support in 2018 and vote choice in 2016, as well as attitudes about issues connected to Trump. Moreover, this profile—an amalgamation of attitudes about, for example, racial groups, immigrants, and political correctness—rivals partisanship and ideology as predictors of Trump support and is negatively related to support for mainstream Republican candidates. Our analyses suggest that Trump benefited from activating dimensions of public opinion that transcend traditional party cleavages.
Why did Americans vote for and support Donald Trump? Trump’s status as a non-traditional, inexperienced outsider has made the answer to this question considerably more elusive than similar questions asked about other presidents. Even intra-party dynamics appear different, with many researchers embracing the idea that Trump supporters differ in important ways from non-Trump supporting Republicans (Barber & Pope, 2019; Blum & Parker, 2019; Reny et al., 2019). These differences have elevated the accounting of Trump support to a “key social-science challenge” (Federico & de Zavala, 2018, p. 110).
What explanations does the literature offer? On the one hand, the classical social-psychological antecedents of political behavior (i.e., partisan and ideological identities) are consistently strong predictors of Trump support (i.e., positive feelings toward him and vote choice) (Bartels, 2018). On the other hand, an expanding literature has uncovered numerous other deep-seated social-psychological factors that additionally predict Trump support. These factors range from group orientations, such as racial prejudice and sexism, to more general postures toward political power and culture, such as anti-political correctness attitudes and conspiracy thinking (e.g., Schaffner et al., 2018). While any one of these factors is unlikely to supplant the role of partisan and ideological orientations on its own, they may still contribute to a more complete accounting of Trump support, as others have demonstrated.
Even though researchers now have a better grasp of which social-psychological orientations might lead to support for Trump, this literature has developed in a piecemeal fashion, leaving questions about how to best model Trump support. This can partially be attributed to Trump’s tactics catching researchers off guard and without the necessary survey instruments to test explanations against one another. Consequently, many studies fail to include potentially confounding explanations of Trump support, leading to omitted variable bias. Yet, other studies find that inclusion of appropriate controls can result in unstable estimates of the effects of key social-psychological predictors (e.g., Grossmann & Thaler, 2018), due to multicollinearity. Regardless of the statistical problem, our understanding of the attitudinal antecedents of Trump support are being obscured.
In this study, we use unique data from the 2018 Cooperative Congressional Election Study (CCES) and publicly available data from the 2016 American National Election Study (ANES) to demonstrate these problems. After observing high correlations between constructs such as racial resentment, sexism, and xenophobia, we demonstrate the perils of both including the many hypothesized predictors of Trump support in a single model, and of failing to include many of such predictors. 1 Undertaking a series of model robustness checks, we find that inferences about the substantive impact—including size, statistical significance, and sign—of the many predictors of Trump support are highly contingent on which of those predictors are included in the model. To avoid such problems, we model Trump support as a constellation, or “profile,” of attitudes and orientations that are empirically related. The profile we construct either matches or outperforms partisanship and ideology in predicting Trump vote choice, general support for Trump, and a host of attitudes about several of Trump’s core issues (e.g., building the U.S.–Mexico border wall). Importantly, this profile does not merely capture partisanship or ideology in another way; indeed, it is negatively correlated with support for establishment Republicans running in the 2016 primaries.
Our findings have several methodological and theoretical implications for the study of vote choice. Given that social-psychological orientations toward social, racial, and political “others” are, as the burgeoning literature on social identity suggests (e.g., Mason, 2018), frequently related to each other, perhaps such orientations should not be treated separately. Our analyses empirically demonstrate the inferential pitfalls of separately modeling constructs that have become tightly interwoven over time. As polarization and sorting have made it more difficult to disentangle inherently overlapping orientations, social scientific theories about which orientations matter and when, and how to empirically decipher the unique effects of such orientations, have not subsequently adjusted. Our results also demonstrate that significant correlations between left-right predispositions and choices, like partisanship and vote choice, and various social-psychological orientations, like xenophobia, imply little about the precise nature of the partisan divide. Indeed, we find that even though the average Trump supporter registers high levels of racial resentment and xenophobia in the absolute, they also register low or middling levels of sexism and conspiracy thinking; likewise, Clinton supporters exhibit non-trivial levels of anti-political correctness attitudes, even though there are significant differences between Trump and Clinton supporters when it comes to these orientations. Thus, our analyses highlight blind spots in how survey questions designed to estimate orientations like sexism and xenophobia are interpreted and applied to models of vote choice.
The Many Predictors of Support for Donald Trump
Since 2015, scholars across social scientific disciplines have attempted to explain why Americans support Donald Trump. Traditional explanations, such as partisanship and ideology, account for considerable variance in attitudes toward Trump, just as they have for most presidents since the advent of modern polling (Bartels, 2018). Even so, these factors cannot explain why some key, historically Democratic districts supported Trump, nor can they account for intra-party differences in choosing Trump over his Republican competitors. A complete accounting of the sources of Trump’s support requires more than the traditional model of vote choice has to offer. In attempting to decipher precisely what additional factors may play a role in Trump support, researchers have, for the most part, made inferences about which elements of Trump’s personality and communication style might resonate with the mass public. This exercise has resulted in many explanations for Trump support, spanning the sociodemographic, psychological, and cultural.
To begin, examinations of sociodemographic explanations have garnered only weak support. Education, for example, is negatively associated with Trump support, but it is unclear whether something about a lack of educational attainment itself makes Trump more appealing or lower educational attainment merely serves as a proxy for other orientations (Silver, 2016). Others saw a theme of working class vulnerability in Trump’s campaign communications (Bucci, 2017; Morgan & Lee, 2018), but a wide range of analyses employing different models, data, and assumptions suggest that—at best—economics is an inconsistent predictor of support (e.g., Green & McElwee, 2018; Mutz, 2018; Ogorzalek et al., 2020; Silver, 2016).
Of course, it is easy to find examples of racism, sexism, xenophobia, conspiracism, and authoritarianism in Trump’s rhetoric (e.g., Finley & Esposito, 2019; Jamieson & Taussig, 2017; Oliver & Rahn, 2016; Sanchez, 2018); thus, these orientations have constituted fruitful avenues for explaining Trump support. Many researchers have found a connection between Trump support and various racial orientations—from white identity to racial prejudice (Abramowitz & McCoy, 2019; Craig et al., 2018; Donovan & Redlawsk, 2018; Engelhardt, 2019; Green & McElwee, 2018; Jardina, 2019; Lajevardi & Oskooii, 2018; Lopez Bunyasi, 2019; Schaffner et al., 2018; Sides et al., 2019). Immigration being a key element of Trump’s platform, xenophobic tendencies also correlate with Trump support (Hooghe & Dassonneville, 2018; Manza & Crowley, 2018; Wright & Esses, 2019). The same can be said about attitudes toward women, from hostile to ambivalent sexism (Bracic et al., 2019; Cassese & Holman, 2019; Deckman & Cassese, 2019; Frasure-Yokley, 2018; Schaffner et al., 2018; Setzler & Yanus, 2018). Importantly, the literature also finds that attitudes about racial groups, immigration, and gender each played a greater role in explaining vote choice in 2016 than in previous elections, suggesting that Trump activated these orientations in a way that previous Republican presidential candidates had not (Donovan & Redlawsk, 2018; Reny et al., 2019; Sides et al., 2017; Valentino et al., 2018, but see Al-Gharbi, 2018).
Finally, several broad orientations toward political power and culture can be found among both Trump’s rhetoric and the psychology of his supporters. Negative attitudes toward political correctness—a posture regarding “appropriate” interactions with members of various groups—are correlated with Trump support (Conway et al., 2017). Trump is also famous for his engagement with populist and conspiratorial ideas (Oliver & Rahn, 2016), garnering him labels such as “conspiracy theorist in chief” (Cillizza, 2017). Indeed, both populist (Carmines et al., 2016) and conspiratorial (Cassino, 2016) views are positively related to support for Trump. Finally, Trump’s “strong-man” image elevated authoritarian views in comparison to previous elections (Knuckey & Hassan, 2019), all but ensuring that various conceptualizations of authoritarianism are associated with Trump support (Hetherington & Weiler, 2018; Ludeke et al., 2018; MacWilliams, 2016; Womick et al., 2018).
Our understanding of the antecedents of Trump support now includes numerous predictive attitudes and orientations, and multiple permutations of each of those. Yet, it is still unclear how these explanations should be modeled. It is evident from the literature that many of the social-psychological factors hypothesized to explain Trump support are highly correlated. Indeed, Grossmann and Thaler (2018, pp. 768–770) find that as models are altered to include or exclude particular explanatory factors (e.g., racial resentment), estimates associated with other explanatory factors subsequently change quite radically. For example, while several studies find that authoritarianism is a significant predictor of Trump support (e.g., MacWilliams, 2016), Grossmann and Thaler (2018) find that the inclusion of controls for other attitudinal variables results in a non-statistically significant coefficient estimate for authoritarianism. This should not be taken as an indication that authoritarianism is not an important predictor of Trump support, but rather, as Grossmann and Thaler (2018, p. 768) suggest, that some attitudinal factors theoretically and empirically overlap, share tendencies that are “rooted” in each other. Numerous papers similarly show that the inclusion or exclusion of theoretical predictors of Trump support can substantially change model estimates (Blair, 2017, p. 16, Knuckey & Hassan, 2019, p. 11, Obschonka et al., 2018, p. 292). 2 This often leaves researchers omitting certain variables, therefore succumbing to the inferential problems caused by omitted variable bias.
We, therefore, contend that it is more efficient to conceptualize the explanatory factors explored above as elements of a broader profile of orientations, rather than as individual attitudinal explanations to be tested against each other. This reconceptualization of the sources of Trump support will, as we demonstrate, ameliorate the problems associated with omitted variable bias and multicollinearity. It will also provide the bedrock of a more unified theory of the sources of Trump support—one built not on any one explanation, but a constellation of attitudes and orientations toward groups and culture. We do not claim to have developed the “best” or “true” profile of attitudes and orientations—new explanations continue to be offered, and models are inherently false oversimplifications of reality, in any case. With that said, our strategy captures the basic contours of the extant literature and provides researchers with a strategy for avoiding statistical pitfalls.
Data and Analytical Strategy
In order to simultaneously examine the many factors posited to explain support for Donald Trump, we measure several of these on a single survey. We fielded a module on the 2018 CCES that included indicators of attitudes regarding racial minorities, women, immigrants, political correctness, and conspiracy thinking (Table 1). We do not suggest that these are the only factors driving Trump support, but rather that they represent the basic contours of the emerging literature regarding Trump support. The survey was administered to 1,000 respondents during October 2018. We also employ the 2016 ANES, which includes similar survey instruments, as well as authoritarianism, with two particular goals in mind: 1) to test the power of our analytical strategy in a different temporal context, and 2) to examine the effects of these factors prior to the 2016 election, on support for other candidates in the 2016 Republican primaries.
Question wording for all items employed below. 2018 CCES.
Our analytical strategy unfolds in four steps. First, we employ the CCES data to examine the interrelation between the Trump support factors, with an emphasis on the stability of observed effects across model specifications. Here, we are interested in understanding the extent to which inferences about the effect of any given predictor of Trump support are contingent on the inclusion or exclusion of other variables in the model. Second, we present a strategy for conceptualizing and measuring the Trump support factors which results in an empirical estimate of a “Trump profile”—or, amalgamation of attitudes associated with Trump support. Third, we demonstrate that this profile is a better predictor of Trump support and attitudes associated with Trump than any individual factor alone, all factors separately applied, or partisanship and ideology. Finally, we estimate the Trump profile using a close approximation of variables from the 2016 ANES, confirming the robustness of previous findings and showcasing the ability of the profile to discriminate between support for Trump and support for other Republican candidates.
Empirical Analysis
Each of the factors posited to explain Trump support, except partisanship and ideology (for which we use the standard measures), are measured via multiple-item scales. This allows us to reduce measurement error and generate sharper estimates. This also means that when we compare the effects of these variables, differences are less likely to be due to measurement error, but to true differences in the (controlled) effect of these variables, assuming that the models are reasonably correct. Information about the items composing each scale, statistical reliability, and the proportion of shared inter-item variance accounted for by the first factor of an exploratory factor analysis appear in Table 2. Each of the scales is statistically reliable (i.e., high alpha) and squarely unidimensional (i.e., high proportion of variance explained by first factor). Thus, these variables will be on a roughly equal playing field in terms of the influence of measurement error when comparing their effects in models below (Westfall & Yarkoni, 2016).
Characteristics and psychometric properties of Trump support variable scales. 2018 CCES.
We begin by examining the bivariate relationships between the hypothesized predictors. Table 3 contains correlations between the factors, as well as partisanship and ideology. Most of the correlations are quite large and statistically significant at the p < .05 level. The exception is conspiracy thinking, which is consistent with previous studies (Miller, 2020; Uscinski et al., 2016). Racial resentment, sexism, anti-immigrant attitudes, anti-PC attitudes, and traditional political orientations are all highly correlated, ranging from 0.489 to 0.725.
Correlations between Predictors of Trump Support. 2018 CCES.
Cell entries are Pearson correlation coefficients. * denotes p < .05 with respect to a two-tailed test.
From the magnitude of the intercorrelations in Table 3, one can already imagine potential difficulties in disentangling the effects of each of these constructs. Multicollinearity may affect inferences by increasing standard errors, leading to inaccurate tests of statistical significance. It can also increase the sensitivity of estimates to model specification. Relatedly, omitting variables—especially ones that are highly related to others in the model—causes issues of a different sort. Models omitting too many variables will be mis-specified and incapable of adjudicating between the various explanations for Trump support. Further, omitting variables that are highly correlated with others included in the model may produce biased estimates, and the likelihood of such a scenario increases as the magnitude of the intercorrelations between predictors increases.
To demonstrate the effects of both multicollinearity and omitted variable bias, we undertake a series of model robustness checks (Young & Holsteen, 2017). These are designed to reveal the (in)stability of estimates across model specifications and the effects of multicollinearity and omitted variable bias. Based on the prevailing literature, we decide on a core set of theoretical and control variables including partisanship, ideological self-identification, racial resentment, anti-immigrant attitudes, sexism, anti-PC attitudes, conspiracy thinking, income, educational attainment, age, frequency of attendance of religious services, and dummy variables for self-identification as black, white, or Hispanic, gender, and residence in the South. Then, we specify regression models with all possible combinations of these variables. This results in 215, or 32,768, model specifications. Finally, retained estimates associated with each variable of interest are used to construct those variables’ “modeling distributions”—where variability in coefficient estimates is attributed to different model specifications, rather than sampling error.
We are particularly interested in the shape of the empirical modeling distributions, including variance in the magnitude, sign, and statistical significance of the effect of a given variable across models. Relatively high instability according to the former criteria would suggest a large effect of some combination of multicollinearity and omitted variable bias. Substantively, this scenario would suggest that previous attempts to investigate explanations for Trump support may be inaccurate—or, at least, incomparable—unless all of such predictors are included in the model. In other words, investigations of the effects of our variables of interest may be misleading if controls for other explanatory factors are excluded from models.
Figure 1 contains the empirical modeling distributions for each of the explanatory factors of interest from logistic regression models of Trump vote choice. There are some systematic characteristics of the set of distributions worth considering. First, most distributions are heavily skewed (e.g., racial resentment, sexism); some are even bimodal (e.g., conspiracy thinking, ideological self-identifications). More importantly, estimates from the full model—depicted by vertical dashed lines—rarely represent either the mean or median estimate (i.e., the center of the empirical modeling distributions). This signifies that omitted variable bias is present—removal of a subset of predictors dramatically alters the magnitude of the estimated coefficients. Finally, three hypothesized predictors of Trump support—conspiracy thinking, racial resentment, and anti-political correctness attitudes—are not statistically significant (as denoted by red dashed lines) in the full model, despite non-zero coefficients in most models.
Density plots of coefficient estimates from model robustness analyses. Dashed lines represent estimate from full model with all controls. Dashed lines presented in red signify a statistically non-significant estimate from the full model. 2018 CCES.
Per the discussion of Table 2, fluctuation in the magnitude, direction, and statistical significance of estimated effects is likely not due to measurement error. Rather, it appears that our inferences about the substantive impact of the predictors of Trump support are highly contingent on which of those predictors are included in the model. Of course, diagnosing the problem says nothing of how fix it. Including all hypothesized support variables in the model—an improvement over the situation that others have found themselves in—still leaves us with the problem of multicollinearity. Note that the effects of racial resentment, conspiracy thinking, and political correctness attitudes are not statistically significant in the full model. However, we should not take this as a sign that Trump failed to activate racist, conspiratorial, or anti-change orientations; indeed, these are some of the most popular explanations for Trump support. While conservative estimates are generally more desirable than the overestimations created by omitted variable bias, such conservative estimates also leave us unable to adjudicate between the various reasonable, theoretically-informed accounts of the psychological foundations of Trump support.
Creating a Profile
The solution to the remaining statistical problem lies in our understanding of the explanatory factors under consideration. Though researchers have proposed individual explanations for Trump’s popularity, few would dispute the interrelation between most of those explanatory factors. Given that racial resentment, sexism, anti-immigrant attitudes, anti-PC attitudes, and conspiracy thinking are correlated, the observed attitudes used to capture these psychological orientations might be amenable to conceptualization as components of a broader “profile” of related, relevant attitudes. In other words, the observed correlations between these individual sets of attitudes—and the omitted variable bias and multicollinearity following from those correlations—is a consequence of a substantive relationship between the attitudes.
Conceptualization of a general orientation—or profile—that unifies the attitudes associated with support for Trump is congruent with both the speculations of previous scholarship and our analyses. It is theoretically parsimonious and, empirically, this reconceptualization of Trump support circumvents some of the problems associated with multicollinearity, resulting in a more powerful model. To understand the utility of this strategy, consider the emotion “extraversion.” While such a discrete emotion surely exists, it is likely composed of subdimensions, the aggregation of which we label “extraversion.” For example, warmth, gregariousness, and excitement-seeking are all elements of extraversion, though not everyone we consider to be extraverted exhibits the same levels of each of these characteristics.
Thus, attitudes used to estimate extraversion can be modeled as the simultaneous product of one general “extraversion” dimension, as well as several more specific subdimensions (e.g., warmth) along which people may vary. This is precisely what we aim to do in modeling the social-psychological foundations of Trump support.
Specifically, we employ a bifactor model of the covariances between the observed indicators of racial resentment, sexism, anti-immigrant attitudes, anti-political correctness attitudes, and conspiracy thinking. A path diagram of the model appears in Supplemental Appendix 1 (Figure A1). The specific explanatory factors related to Trump support are analogous to the specific dimensions of extraversion, and a general “profile” is analogous to general extraversion. Observed covariances between the individual indicators of racial resentment, sexism, anti-immigrant attitudes, anti-political correctness attitudes, and conspiracy thinking, are, theoretically, the causal product—to varying degrees—of both specific constructs of the same name and a general profile.
Note that this model does not imply that racism, sexism, etc. are downstream by products of this profile, which would be better operationalized via a hierarchical model (e.g., a “second order” model). Rather, our hypothesis is that the specific observed attitudes—individual survey items—that we typically sum together and label racism, sexism, etc. can be amalgamated into a broader profile of attitudes that spans these constructs. This is a subtle, but crucial distinction for both our empirical strategy and theory. If this model fits the data well, we will have achieved a more parsimonious and theoretically-powerful account of Trump support, as well as circumvented the remaining statistical issue of high and consequential multicollinearity.
Estimates from the bifactor model appear in Supplemental Appendix 1 (Table A1). Several characteristics of the model output suggest excellent fit to the data. First, all but one factor loading (of 38 such estimated loadings) is statistically significant across specific factors, and all observed indicators significantly load on the Trump profile. Moreover, all fit statistics suggest excellent model fit. The root mean squared error of approximation (RMSEA) is at the recommended 0.05 cutoff for “excellent” model fit (Kline, 2015), and both the Comparative Fit Index (0.971) and Tucker-Lewis Index (0.963) are above the recommended 0.95 rule of thumb (Hu & Bentler, 1999).
We next consider concurrent validity—the extent to which the Trump profile 3 is capable of distinguishing between groups it should theoretically be able to distinguish between. In Figure 2, we plot the distribution of the Trump profile by gender, race, educational attainment, and household income. The patterns we observe are consistent with previous polling and analyses of Trump’s electoral base. For example, men and whites are higher along the profile than women or non-whites. Likewise, those low in educational attainment and household income are higher along the profile than those with higher levels of education and household income, despite the bimodality of the higher income group.
Distribution of the Trump profile by gender, race, educational attainment, and household income. 2018 CCES.
We also note that the profile is not merely a substitute for partisanship or ideology. We plot the distribution of the Trump profile by partisanship and ideology in Figure 3. Even though we expect the Trump profile to be related to partisan and ideological identities, we should anticipate neither that all Republicans/conservatives exhibit consistently high levels of the various ingredients of the profile, nor that all Democrats/liberals exhibit low levels of the ingredient. Rather, we expect that a non-trivial proportion of conservatives/Republicans will be positioned low on the Trump profile, while a non-trivial proportion of liberals/Democrats will be positioned middling or high. Moderates and independents should be oriented at all locations along the latent continuum. The findings in Figure 3 meet these expectations. For example, 21% of Democrats lie in the upper half of the scale, and over 36% of Independents are in the lower half. Thus, the Trump profile is related to partisanship and ideology, but far from determinative of, or determined by, them.
Distribution of the Trump profile, by self-identification as a liberal/Democrat (blue, dotted), conservative/Republican (red, dashed), or moderate/Independent (green, solid). 2018 CCES.
Finally, we explore what the constellation of attitudes looks like for the average Trump and Clinton voters in an effort to unpack what precisely the Trump profile factor captures. In Figure 4, we plot the mean level of racial resentment, anti-immigrant attitudes, sexism, anti-PC attitudes, and conspiracy thinking for the average factor scores for Trump (0.76) and Clinton (0.33) voters. 4 Two important patterns emerge. First, Trump voters exhibit statistically significantly (p < .05) higher levels of each orientation than Clinton voters, on average. This showcases that these orientations do move together to promote support for Trump. Second, there is considerable variability in the level of these orientations displayed by the average Trump voter. While racial resentment, anti-immigrant, and anti-PC attitudes are quite prevalent, there is barely a difference in the level of conspiracy thinking between Trump and Clinton voters (which is at middling levels) and even Trump voters exhibit low levels on sexism, in the absolute (i.e., average levels below the 0.5 neutral midpoint). Moreover, even the average Clinton voter exhibits middling levels of anti-PC attitudes. Thus, differences between Trump and Clinton voters, or even Democrats and Republicans, does not necessarily imply that one group is high in the absolute; rather, we can only infer relative levels from correlations and regression coefficients. We extensively discuss the implications of this conclusion in the penultimate empirical section of the paper. For now, we emphasize that the Trump profile performs nicely in both separately classifying Trump and Clinton voters, and doing so using a constellation of overlapping orientations.
Profile of average Trump and Clinton voters according to Trump profile factor. The dashed line represents neutrality (on average) for each orientation. 2018 CCES.
Predicting Support for Trump and Related Sentiments
We now demonstrate the predictive power of the Trump profile. We do so with respect to both specific measures of Trump support (self-reports of voting for, and feelings toward, Trump) and issues promoted by Trump: climate change denial, skepticism about Trump/Russia collusion, and distrust of the news media. Strong and significant effects of the Trump profile variable will not only provide a final piece of predictive validity for the measurement strategy but demonstrate the statistical and substantive utility of reconceptualizing the many explanations for Trump support as components of a single, broader profile.
First, we regress Donald Trump feeling thermometer scores (0–100) and (retrospective) Trump vote choice 5 on the Trump profile, as well as partisanship, ideology, and a host of controls for retrospective evaluations of the national economy, income, religiosity, educational attainment, age, race/ethnicity, gender, and residence in the South. 6 Ideology, partisanship, and the Trump profile are all coded such that larger values denote more conservative, Republican, and Trump-sympathetic orientations. Full model results appear in Table 4, and model-based predictions over the range of the three explanatory variables of interest appear in Figure 5. 7
Regression Models to Explain Trump Support. 2018 CCES.
Note. OLS coefficients in column 1, logistic in column 2. Standard errors in parentheses.
p < .05. **p < .01. *** p < .001.
Predicted Trump feeling thermometer scores and probability of Trump vote choice, across the range of partisanship, ideological self-identifications, and the Trump profile, controlling for other factors. Dashed lines represent 95% confidence intervals. 2018 CCES.
The coefficients on the explanatory variables of interest are statistically significant and are larger than those associated with any of the other control variables in the model, including education or income. Of course, it makes good sense that both partisanship and ideology are important factors in explaining support for Trump. However, both ideology and, to a lesser extent, partisanship exhibit less predictive power than the Trump profile. Take, for instance, the predicted probability of a vote for Trump over Clinton or another candidate in panel A of Figure 5. For the strongest Democratic identifiers, this probability is about 0.30; for the strongest Republicans, about 0.60. However, for the Trump profile, those low on the scale voted for Trump with a probability of 0.13, and those very high with a probability of 0.70. A similar trend holds for Trump feeling thermometer scores in panel B, though the effects of partisanship and ideology more closely approximate that of the Trump profile.
The Trump profile also provides substantively and statistically significant predictive power in explaining attitudes about issues that Trump has regularly broached. In a final test of our analytic strategy, we regressed attitudes regarding skepticism about Russian interference in the 2016 presidential election, denial of climate change, and distrust of the media on the Trump profile, partisanship, ideology, and the control variables. Full model estimates appear Table 5, and model predictions over the range of the Trump profile and partisanship appear in Figure 6.
Regression Models to Explain Attitudes about Issues and Ideas Espoused by Trump. 2018 CCES.
Note. OLS coefficients w/standard errors in parentheses.
p < .05. **p < .01. ***p < .001.
Predicted distrust of media, skepticism of anthropogenic climate change, and disbelief in Russian collusion in the 2016 presidential election across range of the Trump profile and partisanship, controlling for other factors. Dashed lines represent 95% confidence intervals. 2018 CCES.
Here again, we observe substantively large effects of the Trump profile in explaining attitudes associated with Trump’s espoused stances on key issues. Those low on the Trump profile strongly agree that “climate change is real and caused by manmade carbon emissions,” agree that “the Russians colluded to rig the 2016 presidential election,” and disagree that “much of the mainstream news is deliberately slanted to mislead us.” Those high on the Trump profile exhibit the opposite attitudes. The Trump profile is also more predictive of these attitudes than are either partisanship (pictured) or ideology.
Trump Profile or Conservative Profile?
Finally, we consider both the robustness of the profile approach, as well as the discriminatory power of the Trump profile when it comes to other Republican candidates. In order to demonstrate the robustness of our profile-based approach to measuring Trump support, we replicate as closely as possible the previous analyses using the 2016 ANES, which included many of the variables utilized above. Of those, most are either direct replications or substantively identical (see Supplemental Appendix 1). Additionally, the 2016 ANES included items designed to measure authoritarianism, an oft-cited reason for Trump support (e.g., MacWilliams, 2016) not available on the CCES, though it does not include a measure of conspiracy thinking. Moreover, the 2016 ANES includes measures of attitudes about other issues that Trump has claimed ownership over, as well as self-reported Republican primary voting. The latter information is useful for establishing that the Trump profile is specific to Trump and not merely a substitute for traditional conservative principles or “mainstream” Republicanism.
The bifactor model again fits the data well. 8 As with the 2018 CCES data, all fit statistics meet their respective rules of thumb and all indicators load statistically significantly on the Trump profile factor. Moreover, we observe similar relationships between the Trump profile and measures of support for Trump and Trump-related issues. 9 Figure 7 depicts the relative influence of partisanship, ideology, and the Trump profile on feelings toward Trump and Trump vote choice in the general election. Although the effects of the Trump profile and partisanship are slightly more comparable than we observed using the 2018 CCES, the effects of the Trump profile rival those of partisanship and prove greater than those of ideology.
Predicted Trump feeling thermometer scores and probability of Trump vote choice, across the range of partisanship, ideological self-identifications, and the Trump profile, controlling for other factors. Dashed lines represent 95% confidence intervals. 2016 ANES data.
Figure 8 shows that the Trump profile is, again, highly predictive of attitudes about Trump-related issues, even compared to partisanship, holding other factors like ideology and sociodemographic characteristics constant. In each case, the Trump profile provides more explanatory power than either partisanship (pictured) or ideological self-identifications. This is most apparent when it comes to attitudes about “the wall” and childhood vaccinations, the latter of which is not statistically related to either partisanship or ideology.
Predicted skepticism of anthropogenic climate change, perceived danger of childhood vaccinations, and supportive attitudes about building a wall along the U.S.-Mexican border across range of Trump profile and partisanship, controlling for other factors. Dashed lines represent 95% confidence intervals. 2016 ANES data.
Next, we consider the discriminatory power of the Trump profile. If this profile of attitudes is unique to Trump supporters, it should be more strongly related to voting for Trump than any other 2016 Republican primary candidate. To test this proposition, we estimate a multinomial logistic regression where the dependent variable captures voting for Trump, Ted Cruz, John Kasich, or Marco Rubio in the 2016 primary elections. The predicted probability of casting a vote for each of the candidates across the range of the Trump profile, holding constant the strength of partisan and ideological identities, retrospective evaluations of the economy, and sociodemographic characteristics, is depicted in Figure 9. 10
Predicted probability of voting for each of four Republican candidates in the 2016 U.S. presidential primaries, controlling for other factors. Respondents are Republicans. Dashed lines represent 95% confidence intervals. 2016 ANES data.
Even among self-identified Republicans voting in the primary, the Trump profile is strongly positively related to voting for Trump. 11 We find a weak positive relationship between the Trump profile and voting for Cruz, as we might expect given his platform and rhetoric. However, such is not the case when it comes to the two mainstream candidates: Rubio and Kasich. For both of these mainstream Republican candidates, we observe a negative relationship between the Trump profile and candidate vote choice.
We do not take these patterns to suggest that particular candidates, or “mainstream” candidates more broadly, are completely unassociated with normatively undesirable characteristics, like racism and sexism—the mass public is heterogeneous and these orientations will always be operational within one’s political base, to some extent. Rather, these patterns showcase how effective Trump was at activating particular orientations, showcasing their relevance to the vote, and drawing individuals who exhibit such orientations in his direction, away from more qualified opponents. Trump did not need to make people racist, sexist, xenophobic, and the like if could simply activate what was already there in a way his opponents were either incapable of or unwilling to. That said, even though the profile we devised is composed of orientations theoretically tied to Trump, we can imagine such a strategy applying to other candidates (e.g., there may be constellations of attitudes that uniquely relate to Rubio or Kasich support, even though that was not our focus here).
A Cautionary Note on Interpretation
We end our empirical investigation with a cautionary note about the interpretation of survey questions, which we believe provides further support for our strategy of modeling many psychological predictors of Trump support as a profile of factors. As an illustration of what exactly variation along the Trump profile means—or could mean—consider the individual attitudinal profiles presented in Table 6. We randomly sampled two individuals from the 2018 CCES dataset: one at, or lower than, the 25th percentile along the 0 to 1 Trump profile, one at or higher than the 75th percentile. The goal was to blindly acquire relatively extreme individuals in both directions, so that their individual attitudinal profiles may be examined.
Examples of Individual Attitudinal Profiles for Relatively High and Low Values on the Trump Profile Factor.
Note. SA = strongly agree; A = agree; N/N = neither/nor; D = disagree; SD = strongly disagree.
Case identification numbers: 416172031, 416179346.
The first individual sampled scored a 0.29 on the profile factor. They registered as a strong Democrat and extremely liberal—the most extreme positions on the two most important political orientations in political science. The second individual, a Republican Party “leaner” who is slightly conservative, scored a 0.74. Despite a wide gap on the profile factor and scores in the first and fourth quartile, neither respondent provides particularly extreme responses to the individual questions used to tap the social-psychological orientations. Of the three extreme responses provided by the conservative respondent, one deals with perceived “deservingness” of blacks, another with a rather tame attitude (on its face, at least) about immigration being “slowed down,” and a third—strong disagreement that a woman’s place is “in the home.” This respondent appears more supportive of equal rights for women and less conspiratorial than the more liberal respondent who is low on the profile factor.
These are only two examples of many, but this analysis showcases a critical point. Even though others have observed strong associations between Trump support and the individual psychological factors posited by the literature discussed above, this says nothing of the absolute levels of those attitudes—only that some are systematically higher than others. The familiar claims that all Trump supporters are racist, sexist, xenophobic conspiracy theorists do not best characterize the psychology of the average Trump supporter, at least as far as survey questions can tell us. Indeed, a small proportion of respondents are consistently extreme in the uncharitable direction, and no more than 8.5% of respondents are extremely negative across all survey items composing any of the racial resentment, sexism, anti-immigration, anti-PC, or conspiracy thinking scales. Finally, no one in our sample is consistently extreme across all of these factors.
We do not wish imply that Trump does not find support among people who are racist, sexist, and xenophobic in the absolute sense. There are extremists of this sort among Trump supporters, or any group, and they should not be ignored. That said, the charge of social scientists is to decipher broad patterns and model structure. To this end—and also considering goals such as parsimony and prediction—our judgement is that Trump support is best modeled using a profile of the many posited psychological sources of that support. This strategy allows for variability across the psychological factors in question, while circumventing the problems of multicollinearity and omitted variable bias demonstrated above.
Conclusion
A robust literature demonstrates that numerous attitudes and orientations, beyond usual suspects like partisanship and ideology, are related to support for Donald Trump. Elucidation of these specific factors has expanded the scope of the literature and provided a better accounting of Trump support, an important political phenomenon. Unfortunately, most singular surveys are unable to capture all of these factors, and the inclusion or exclusion of any number of these factors in explanatory models can drastically affect substantive inferences.
To circumvent the problems associated with omitted variable bias and multicollinearity in modeling Trump support, both of which we empirically demonstrated, we generated a broader profile of attitudes related to the explanatory factors others have highlighted. This profile is highly predictive of voting for and feelings toward Trump, as well as attitudes associated with Trump, even controlling for partisanship, ideology, and other factors. The Trump profile does not, however, measure Republicanism or conservatism in another way: it is negatively associated with support for other Republican candidates in the 2016 primary elections.
We do not suggest that we have created the “true” Trump profile. There are other factors (i.e., variants of racism or sexism) that we could not include and the profile we generated could be improved upon. As we demonstrate, having several of the factors works well, even when some factors are exchanged for others (i.e., one of our surveys contained conspiracy thinking and the other authoritarianism). Both theory and empirics should drive these choices. Nonetheless, this strategy provides a path forward for understanding Trump support and modeling it.
Our approach is useful, first, because it frees scholars from having to accept the consequences of either multicollinearity or omitted variable bias in their models of Trump support. Using our measurement strategy, both problems are avoided. Second, our strategy is likely more congruent with reality. Even though previous work makes a strong case for specific explanatory factors, it is unlikely that any given factor on its own explains all Trump supporters. By generating a profile of attitudes Trump is suspected of activating, we are more likely to account for the myriad combinations of attitudes that make Trump an attractive candidate. While some strong Trump supporters may exhibit high levels of racial resentment, they may be low in sexism and middling in conspiracy thinking. Our strategy allows for such a profile of attitudes.
We also believe this strategy can be fruitfully applied in modeling support for other political candidates. As research into the social identity-based foundations of public opinion is increasingly showing (e.g., Kane et al., Forthcoming; Mason & Wronski, 2018), it is the alignment and overlap of identities that is most responsible for contemporary political strife. We have no reason to believe that Trump is unique in activating an interrelated constellation of attitudes—indeed, this is what all candidates attempt to accomplish. We recommend that studies of vote choice more seriously consider—both theoretically and empirically—profiles of overlapping identities and attitudes as explanations. Simple, additive approaches to modeling political choices, whereby single predictors are successively added to the model, have increasingly become deficient as identity sorting and polarization have unfolded. We also want to leave open the possibility that some voters were not attracted to Trump due to racism, sexism, and xenophobia, but as a reaction to Clinton’s opposing rhetoric on these fronts (Al-Gharbi, 2018). In other words, it is important to remember that Trump’s campaign did not exist in a vacuum—the context in which his rhetoric resonated matters.
More broadly, our findings suggest that much of what political scientists have learned about political behavior in the last 100 years is contingent on mainstream parties supporting mainstream candidates who stick to mainstream party platforms. Facing a crowded primary field, it was entrepreneurial for Trump to activate existing attitudes among the mass public that other candidates avoided (Sides et al., 2018). Unfortunately for an increasingly uncivil political culture marred by polarization and sorting, his tactics proved effective. Recent elections have seen other candidates mimicking Trump’s language and policy stances, in the U.S. and abroad. Strategic politicians do not have behave like Trump in order to activate in people what Trump has managed to activate, or to find his successes.
Supplemental Material
sj-pdf-1-apr-10.1177_1532673X211022188 – Supplemental material for On Modeling the Social-Psychological Foundations of Support for Donald Trump
Supplemental material, sj-pdf-1-apr-10.1177_1532673X211022188 for On Modeling the Social-Psychological Foundations of Support for Donald Trump by Adam M. Enders and Joseph E. Uscinski in American Politics Research
Supplemental Material
sj-pdf-2-apr-10.1177_1532673X211022188 – Supplemental material for On Modeling the Social-Psychological Foundations of Support for Donald Trump
Supplemental material, sj-pdf-2-apr-10.1177_1532673X211022188 for On Modeling the Social-Psychological Foundations of Support for Donald Trump by Adam M. Enders and Joseph E. Uscinski in American Politics Research
Footnotes
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) received no financial support for the research, authorship, and/or publication of this article.
Supplemental Material
Supplemental material for this article is available online.
Notes
Author Biographies
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
