Fear and Deprivation in Trump’s America: A Regional Analysis of Voting Behavior in the 2016 and 2020 U.S. Presidential Elections

Regional Level and Spatial Analysis Considerations

The main unit of analysis is the county level, which we supplement with analyses at the Core-Based Statistical Areas (CBSA) level for reasons we describe below. Each CBSA is a metropolitan area composed of one or more counties: an urban core and its surrounding commuting territory.

Spatial analyses require unique analytical considerations. First, these analyses require appropriate geographic control variables. We control for population density because voters in rural regions with low population density tend to vote for conservative candidates. We also control for state-fixed effects by including state dummies that account for variance shared between regions in the same state, including any omitted confounding variables at the state level. It is particularly important to account for state-fixed effects in election research because election laws and procedures that differ from state to state could influence the results. Second, we must anticipate potential violation of the statistical assumption that error terms will be uncorrelated because neighboring regions are non-independent (Griffith, 1987). We address this concern (i.e., spatial autocorrelation) by conducting additional analyses that add spatial lags to our models using Stata’s “spregress” command (StataCorp., 2017).⁴ Third, we conduct analyses at both the county and CBSA levels to address the potential modifiable area unit problem (MAUP), wherein one’s findings depend on the choice of the geographical unit (Ebert et al., 2022; Openshaw & Taylor, 1979).

Personality Data

Personality data come from the Gosling-Potter Internet Personality Project’s most recent dataset, collected from 2003 to 2015 from participants who voluntarily completed personality surveys on the website www.outofservice.com in exchange for feedback about their personality. The data collection was declared exempt from informed consent by the approval of the Institutional Review Board at the University of Texas at Austin because there were no significant risks to participants (IRB #2004–10-0073). A partial list of papers that have used Gosling-Potter Internet Personality Project data can be found at http://www.thebigfiveproject.com/published-papers.

We aggregated the individual-level personality scores (N = 3,167,041) to the county (N = 2,083) and CBSA (N = 923) levels based on where participants reported living. Only counties and CBSAs with at least 100 participants per region were included in our analyses. As the independent variable in our analyses, we focus on neuroticism (α = 0.79) as compared to and controlling for the other personality traits most strongly related to voting behavior in prior research: conscientiousness (α = 0.80) and openness (α = 0.74). These personality traits were measured using the 44-item Big Five Inventory (BFI-44; John & Srivastava, 1999). We address the (non-)representativeness of the sample by conducting robustness checks in which we weight individual respondents by age and gender to make the personality dataset more representative of the U.S. population.

Voting Data

The 2016 U.S. general election data come from open data sources (Github, 2017; OpenDataSoft, 2016), and the 2020 general election results are from David Leip’s Atlas of U.S. Presidential Elections (Leip, 2020). The 2016 U.S. primary election data are from Bucknell University’s Digital Commons (Pirmann, Sherwood, & Tevebaugh, 2016).

For our dependent variable, we focus on three measures of voting behavior. The first measure is Trump’s simple two-party vote share in 2016 and 2020. Vote share was calculated by taking the raw votes for Trump in each region as a proportion of the combined votes for Trump and the Democratic candidate in each region during the 2016 and 2020 U.S. presidential elections. This measure ignores third-party votes. The second measure is Trump’s gains beyond the Republican presidential candidate in the previous election. This measure captures the degree to which regions shifted their vote share to Trump in 2016 from Romney’s vote share in 2012 (“2016 Trump gains”) and to Trump in 2020 from his own initial vote share in 2016 (“2020 Trump gains”). For example, if Trump won 40% of the vote in 2016 and 45% in 2020, his gains from 2016 to 2020 would be 5%-points. Obviously, the size of Trump’s gain corresponds to the size of the Democrats’ loss. Note that the manuscript reports that Trump’s mean two-party vote share is much greater than 50% in both elections, even though he did not win the popular vote in either election. This is because the mean vote share takes the mean across counties rather than individuals, and there are more rural counties (where Trump’s vote share is very high) than urban counties.

The third category of voting behavior captures the vote share of other 2016 Democratic and Republican primary candidates’, which we compare to Trump’s 2016 primary vote share. Analyzing primary election results allows us to investigate the alleged common appeal of Trump and left-wing populist Bernie Sanders as well as whether Trump’s voter base was similar to that of more traditional Republican candidates. We focused on the primary candidates who were arguably Trump’s biggest competition: Bernie Sanders on the Democratic side and John Kasich, Ted Cruz, and Marco Rubio on the Republican side. Although there were other Republican primary candidates in 2016, none received more than 3% of the popular vote.

Sanders’s primary vote share was calculated as the percentage of votes cast for him out of all primary votes cast for candidates of his party in a given region. Put differently, Sanders’s vote share was calculated as the percentage of votes cast for Sanders out of all votes cast for both Bernie Sanders and Hillary Clinton in the Democratic primary because all other 2016 Democratic primary candidates withdrew before or shortly after the primary season began. The vote share for each primary Republican candidates was calculated as the percentage of votes cast for that candidate out of the total number of votes for all four candidates combined (Trump, Kasich, Rubio, and Cruz). We did not consider votes for Bernie Sanders in the 2020 Democratic primary elections because the political landscape had shifted from 2016 such that there were multiple Democratic candidates in 2020 (e.g., Elizabeth Warren) who were running on what might be considered left-wing populist platforms. We did not consider votes for other Republican primary candidates in 2020 because, as sitting president, Donald Trump was the de facto Republican nominee.

Economic, Demographic, and Health Data

Table 1 includes 18 variables broadly capture regional economic, demographic, and health-related conditions in 2016 and 2020.⁵ We factor analyzed these variables to derive factor scores for each region, which were saved and used as variables in our regression analyses. We elaborate on the rationale and procedure for our factor analytic approach below. Enders and Uscinski (2021) also describe a detailed rationale for examining a broad “profile” of factors when modeling Trump support.

Using factor scores as indicators of economic, demographic, and health conditions allows us to address several concerns. Every variable has strengths and weaknesses when it comes to representing a construct of interest. Per capita income may be a good indicator of people’s absolute buying power in a given region, whereas unemployment may be a good indicator of the health of the region’s job market. Both variables provide important information about a region’s economic conditions. Relying on a diversity of sources and measures also helps diminish the influence of limitations, errors, or biases in any individual source’s data. To avoid arbitrarily choosing one variable over another, one could include all relevant variables that might plausibly be related to regional economics, demographics, and health in regression analyses. However, this method presents challenges, such as multicollinearity in analyses that include highly correlated variables, a large number of coefficients to interpret, and ambiguity around how to interpret theoretically related variables that yield contradictory results. The factor analytic approach allows us to avoid these issues while characterizing regions’ general economic, demographic, and health-related conditions.

Therefore, we conducted exploratory factor analyses for 2016 and 2020 at both regional levels. Before conducting each factor analysis, we consulted a scree plot which suggested that three factors had an eigen value > 1 (Kaiser, 1960) and might parsimoniously reflect our structural constructs of interest: economics, demographics, and health. We then conducted a factor analysis specifying three factors, maximum likelihood estimation, median imputation (to account for missing data unbiased by extreme scores), oblimin rotation (to allow for correlated factors), and the tenBerge method (to estimate correlation-preserving factor scores) using the “fa” function in the “psych” R package (Revelle, 2021, Version 1.9.12.31) and the “GPA rotation” R package (Bernaards & Jennrich, 2005, Version 2014.11.1). Results of the 2016 and 2020 county-level factor analyses are reported in Table 2.

Regions that scored high on the first factor, which we call economic deprivation, had low rates of college education, low income, low levels of migration and income inequality, lacked Internet access, had a high share of agriculture/mining industries, and high obesity. These economically deprived regions tend to be rural, as indicated by the factor’s negative correlation (r = -.17) with population density. Importantly, regions that scored high on this factor were not the very poorest regions, since variables like poverty and unemployment did not load strongly on this factor. Instead, regions that scored strongly on the first factor are rural regions that probably used to drive the American economy (e.g., with manufacturing and agriculture) but are no longer doing as well (Eriksson et al., 2021; Low, 2021). Employment share in agriculture and manufacturing have been shrinking in number and value for several decades, leading to socioeconomic distress in counties in the Rust Belt and the Midwest. Therefore, we describe this factor using the word “deprivation,” which is synonymous with “loss,” to reflect these regions’ loss of economic dominance.

Regions that scored high on the second factor, which we call ethnic diversity, had more ethnic diversity, high school dropouts, teen births, poverty, and uninsured people. This factor most likely captures the Black and Hispanic ethnic makeup of a region because these populations account for about twice as many high school dropouts, teen births, uninsured people, and people in poverty than do White and Asian populations, according to the National Center for Education Statistics (NCES, 2019), the Centers for Disease Control (CDC, 2019), and the Kaiser Family Foundation (KFF, 2019a, 2019b). The fact that this factor includes uninsured and high school dropouts shows that communities of color disproportionately lack access to healthcare and education. Similarly, the fact that, in 2016, poverty and unemployment load strongly on this factor shows how racialized lack of economic opportunity is in the U.S. This factor was uncorrelated with population density, suggesting that it captures ethnic diversity in both urban and rural settings.

Regions that scored high on the third factor, which we call health disadvantage, had higher smoking rates, more reported poor physical health days, and more babies born underweight. Using the most stringent significance threshold, this factor was unrelated to population density, suggesting that health disadvantage can be found in both urban and rural regions. The fact that this factor includes poverty (which also loaded strongly on the ethnic diversity factor in 2016) and obesity (which also loaded strongly on the economic deprivation factor) suggests economic, demographic, and health conditions are all intertwined. One benefit of taking a data-driven inductive approach to factor analysis is that it reveals these kinds of interdependencies.

Indeed, correlations between the factors show that regions that were economically deprived also tended to have greater ethnic diversity and worse health. In 2016, the economic deprivation and ethnic diversity factors were positively correlated (r = .18), the economic deprivation and health disadvantage factors were positively correlated (r = .48), and the ethnic diversity and health disadvantage factors were positively correlated (r = .44). In 2020, the economic deprivation and ethnic diversity factors were correlated (r = .10), the economic and health factors were correlated (r = .45), and the ethnic diversity and health disadvantage factors were correlated (r = .42).

Group-Based Dominance Data

The ethnic diversity factor provides a good sense of the objective demographic conditions of a region, but it does not reflect how people perceive or feel about minorities and other historically marginalized groups. Negative subjective attitudes towards minorities were considered core to the original conceptualization of the authoritarian personality (Adorno et al., 1950) and have more recently been shown to be strong predictors of Trump support (Bartels, 2018; Mason, Wronski, & Kane, 2021) and anti-democratic tendencies (Bartels, 2020). Similarly, some research has found that sexism against Hillary, which was not accounted for in our structural factors, was related to Trump voting (Ratliff et al., 2019). Accounting for regional differences in such attitudes may be as important as accounting for the actual demographic makeup of a region because it is possible that two counties have similar demographics, but one of these counties has less bias (e.g., due to higher levels of intergroup contact). More generally, and as previously mentioned, threats to one's group are not experienced in absolute terms but rather in relation to one’s standing relative to other groups. In support of this idea, prior work has demonstrated that psychological factors that capture a preference for one’s own group to dominate or aggress against other groups (e.g., social dominance orientation, right-wing authoritarianism) predict individuals’ support for Trump (Van Assche, Dhont, & Pettigrew, 2019; Womick et al. 2019).

Therefore, we benchmark the main effects of our focal psychological and structural variables against variables that capture regions’ tendency to experience several forms of group-based dominance: social dominance orientation (SDO, N = 690,692, α = .82–.87)⁶, right-wing authoritarianism (RWA, N = 732,347, α = .80–.92), implicit anti-Black bias (race IAT, N = 3,114,109), and implicit gender bias (gender-career IAT, N = 1,039,163). All four group-based dominance variables were accessed from Project Implicit’s public datasets, specifically the Race IAT dataset (available at https://osf.io/52qxl/) and Gender-Career IAT dataset (available at https://osf.io/abxq7/) (Xu, Nosek, & Greenwald, 2014). Implicit race and gender bias was assessed by the D measure from the implicit attitudes test (IAT), which captures the speed with which people associate group categories (i.e., race and gender categories) with other words (e.g., “good” vs. “bad”; “career” vs. “family”). RWA was measured from 2007 to 2020 with items such as “Our country will be destroyed someday if we do not smash the perversions eating away at our moral fiber and traditional beliefs” and “Some of the worst people in our country nowadays are those who do not respect our flag, our leaders, and the normal way things are supposed to be done.” SDO was measured from 2006 to 2020 with items like “It’s OK if some groups have more of a chance in life than others” and “If certain groups stayed in their place, we would have fewer problems.” Note that, unlike our focal psychological and structural variables, measures of SDO and RWA contain political content, which should make them strong predictors of political behavior like voting. Therefore, accounting for these measures of group-based dominance is a particularly conservative test of other psychological and structural variables’ explanatory power.

Moreover, personality traits that were not included in our previous regression models—extraversion and agreeableness—may also play a role in people’s tendency to aggress against others. People who are willing to denigrate women, minorities, and other groups are likely to be assertive (a facet of extraversion) and lacking in compassion (a facet of agreeableness). Indeed, recent research has linked both agreeableness and extraversion to support for Trump and other populists at the individual level (Bakker, Matthijs, & Gijs, 2016a; Bakker, Schumacher, & Rooduijn, 2021; Bakker, Rooduijn, & Schumacher, 2016b; Fortunato et al., 2018). Thus, we also account for the main effects of extraversion (α = .87) and agreeableness (α = .81), as measured in the Gosling-Potter Internet Personality Project by the BFI-44.

We aggregated individual scores on the group-based dominance and personality variables to the county level. We dropped counties that included less than 50 individual observations rather than those with less than 100 individual observations (as was the case in our previous analyses) to increase our total sample of counties. Even so, our final sample size for regressions included group-based dominance variables was about half (N = 1,096) of that used in our focal analyses.

Main Effects of Structural Factors and Regional Personality on Trump Voting

Models 1, 2, and 3 evaluated the extent to which regional differences in the regional covariates, personality traits, and structural factors explained voting behavior, respectively. Table 4 presents Trump votes and gains at the county level in 2016. Table 5 presents Trump votes and gains at the county level in 2020.

Model 1 revealed that state-fixed effects and population density explain roughly 20–40% of total variance in votes for Trump in 2016 and 2020. Model 2 showed that contextual factors explain an additional 30–40% of variance in Trump’s vote share and gains. More specifically, the economic deprivation factor (2016 votes: b = 11.36, p < .001; 2016 gains: b = 3.69, p < .001; 2020 votes: b = 12.42, p < .001; 2020 gains: b = 0.69, p < .001) explained voting such that a one standard deviation increase in economic deprivation led to a 3.69%-point gain for Trump in 2016 compared to Mitt Romney in 2012, an 11.36%-point greater voting share for Trump in 2016 compared to Hillary Clinton in 2016, a 0.69%-point gain for Trump in 2020 compared to his own performance in 2016, and a 12.42%-point greater voting share for Trump in 2020 compared to Biden in 2020. Regarding the ethnic diversity factor, more ethnically diverse counties were less likely to vote for Trump in 2016 and 2020 and were also the counties where Trump showed losses over Romney (2016 votes: b = -5.67, p < 0.001; 2020 votes: b = -2.79, p < .001; 2016 gains: b = -1.12, p < .001). Unexpectedly however, Trump showed gains in ethnically diverse counties from 2016 to 2020 (b = 1.16, p < .001). The health disadvantage factor also showed a mixed relationship between Trump’s vote share versus his gains over previous years. Counties with worse health conditions were less likely to vote for Trump in 2016 and 2020 as compared to the Democratic candidate in those years (b = -5.74, p < .001; b = -6.16, p < .001), but Trump showed gains in these counties compared to the Republican candidate in the previous election (b = 0.75, p < .01; b = 0.42, p < .01).

In Model 3, we substituted the three structural factors with three personality traits: neuroticism, conscientiousness, and openness. Compared to the baseline model 1, the traits explained an additional 10–20% of variance in votes for Trump. Our focal trait—regional neuroticism—was positively associated with Trump votes and gains in both elections (2016 votes: b = 4.27, p < .001; 2020 votes: b = 4.73, p < .001; 2016 gains: b = 1.75, p < .001; 2020 gains: b = 0.46, p < .001). Consistent with prior work, openness was negatively associated with Trump votes and gains in both elections. Conscientiousness showed weaker relations with voting, positively predicting Trump gains but not Trump votes.

In Model 4, we jointly include the personality traits and structural factors. Most of the results from the Models 2 and 3 stay unchanged except that the neuroticism trait no longer significantly predicts Trump gains in 2020. Additionally, openness and conscientiousness do not consistently predict Trump gains in 2016 and 2020 in this model.

Interactive Effects of Structural Factors and Regional Personality on Trump Voting

Models 5 and 6 investigated interactive effects of the personality traits with each of the three structural factors: economic deprivation, ethnic diversity, and health disadvantage. Model 5 included only the neuroticism interactions with the three structural factors, whereas Model 6 included interaction between all three traits (neuroticism, openness, and conscientiousness) and structural factors (economic deprivation, ethnic diversity, and health disadvantage). Therefore, from the latter model we only discuss the openness and conscientiousness interactions.

The interaction between neuroticism and the economic deprivation factor negatively predicted Trump’s vote share in 2016 but positively predicted gains in both elections (2016 votes: b = -0.7, p < .05; 2016 gains: b = 0.38, p < .001; 2020 votes: b = -0.57, p > .05; 2020 gains: b = .28, p < 0.01). Neuroticism had weaker predictive effect on Trump 2016 votes in economically disadvantaged regions (at +1 SD; 2016 votes: b = 0.88, 95% CI [-0.09, 1.85], p > .05) than in economically advantaged regions (at -1 SD; 2016 votes: b = 2.29, 95% CI [1.10, 3.48], p < .001). For 2020 the interaction between neuroticism and economic deprivation was insignificant. Conversely, neuroticism was a stronger predictor of Trump gains in economically disadvantaged regions (at +1 SD; 2016 gains: b = 0.84, 95% CI [0.61, 1.08], p < .001; 2020 gains: b = 0.31, 95% CI [0.13, 0.49], p < .001) than in economically advantaged regions (at -1 SD; 2016 gains: b = 0.08, 95% CI [-0.31, 0.47], p > .05; 2020 Trump gains: b = -0.24, 95% CI [-0.53, 0.05], p > 0.05).

The interaction between neuroticism and the ethnic diversity factor positively predicted Trump votes in 2016 and 2020 (2016 votes: b = 1.21, p < .001; 2020 votes: b = 0.75, p < 0.05), but not Trump gains in both elections. Neuroticism was a stronger predictor of Trump votes in ethnically diverse regions (at +1 SD; 2016 votes: b = 2.79, 95% CI [1.82, 3.77], p < .001; 2020 votes: b = 2.40, 95% CI [1.35, 3.45], p < .001) than in less ethnically diverse regions (at -1 SD; 2016 votes: b = 0.37, 95% CI [-0.82, 1.57], p > 0.05; 2020 Trump votes: b = 0.89, 95% CI [-0.11, 1.91], p > 0.05).

Similarly, the interactions between neuroticism and the health disadvantage factor positively predicted Trump votes in 2016 and 2020 (2016 votes: b = 1.35, p < .001; 2020 votes: b = 1.71, p < 0.05), but not Trump gains in both elections. Neuroticism was a stronger predictor of Trump votes in regions high on health disadvantage (at +1 SD; 2016 votes: b = 2.93, 95% CI [2.01, 3.85], p < .001; 2020 votes: b = 3.35, 95% CI [2.51, 4.20], p < .001) than in regions low on health disadvantage (at -1 SD; 2016 votes: b = 0.24, 95% CI [-0.89, 1.37], p > .05; 2020 Trump votes: b = -0.06, 95% CI [-1.25, 1.14], p > .05]).

Taken together, the interactions show the expected pattern: neuroticism tended to exacerbate the strength of the relationship between structural factors and Trump voting. However, size of the interaction coefficients demonstrate that the interaction effects were weak, not explaining much variance in voting beyond main effects of personality and structural factors. Moreover, the interaction effects were not particularly robust across dependent variables (vote share and gains) or after accounting for robustness checks (as summarized in Table 9).

Therefore, we focus our interpretation of the results on the large and robust main effects of neuroticism and structural factors on voting. Openness and conscientiousness also interacted with structural factors on votes for Trump, consistent with interactionist theories. We do not interpret these interactions because they also were not particularly robust or consistent, and we did not have specific predictions regarding interactions with other personality traits.

Group-Based Dominance Effects on Trump Voting

As previously mentioned, variables that capture a tendency towards group-based dominance like right-wing authoritarianism (RWA), social dominance orientation (SDO) as well as implicit race and gender bias have been shown to be strong predictors of voting for Trump. Similarly, people high in extraversion and low in agreeableness may be particularly likely to engage in aggressive behavior. To test these alternative explanations, we regressed Trump vote shares and gains on these variables in Table 6. Model 1 includes RWA, SDO, and the implicit race and gender bias in a model with state fixed effects and population density. Model 2 adds the two remaining Big Five traits, extraversion and agreeableness. Model 3 adds the remaining main effects from our focal analyses (neuroticism, openness, conscientiousness, and structural factors). We do not include interaction effects because our previous analyses show that these were not robust.

Across the three models, implicit anti-Black bias positively predicted Trump vote share in 2016, his gains in 2016, and his vote share in 2020. However, this relationship did not hold for Trump’s gains in 2020, when anti-Black bias was either unrelated or negatively related to Trump performance. This result is consistent with the previous finding showing that Trump made gains in 2020 over his performance in 2016 in more ethnically diverse regions. Right-wing authoritarianism also consistently predicted Trump’s vote share in 2016 and 2020. Unexpectedly, however, the relationship between right-wing authoritarianism and Trump’s vote gains in 2016 and 2020 became negative in Model 3, which includes neuroticism and our structural factors. Contrary to past research, SDO and implicit gender bias did not consistently predict Trump voting.

Perhaps most importantly, the main effects of neuroticism, economic deprivation, ethnic diversity, and health disadvantage largely held when accounting for variables capturing group-based dominance. Moreover, the effect size of our focal variables were comparable to or larger than the effect size of the group-based dominance variables.

Analyses of the 2016 Primary Elections: Comparisons to Sanders, Kasich, Cruz, and Rubio

To investigate whether Trump’s voter base is similar to that of other primary candidates, we analyzed Trump’s vote share in the 2016 primary elections, comparing the characteristics of regions that voted for Trump to those that voted for Bernie Sanders, John Kasich, Ted Cruz, and Marco Rubio. Results of these analyses are depicted in Table 7 and Table 8. They suggest some similarities but also notable differences between Trump’s voter base and those of his competitors.⁷

Beginning with comparisons between Sanders and Trump, more neurotic counties were indeed more likely to vote in 2016 for both Sanders (b = 1.04, p < .01) and Trump in the primaries (b = 1.11, p < .001). The same is true for the ethnic diversity factor where both Sanders (Model 4: b = -3.94, p < .001) and Trump in the general election received more votes in less ethnically diverse counties. However, neither economic disadvantage nor health conditions showed a clear relationship with Sanders’s vote share, in contrast to the strong relationship of these factors with Trump in general election and partly in primary election voting. The maps in Figure 1 further show that the spatial distribution of neuroticism and economic disadvantage (marked in dark blue) overlap fairly substantially with regions that voted for Trump but overlap to a much lesser extent with regions that voted for Sanders.

Counties that voted for Sanders were also lower on conscientiousness (b = -1.40, p < .001), whereas counties that voted for Trump in 2016 tended to be lower on openness – a pattern not found in the Sander’s primary election votes. Neuroticism also did not interact significantly with any of the structural factors on Sanders’s vote share. Moreover, despite the significant main effects mentioned above, it is worth noting that most of the variance in voting for Bernie Sanders in the 2016 Democratic primaries—66.8%—was accounted for by geographical factors: population density and state fixed effects. Taken together, regions that voted for Sanders were similar in some ways and different in others to regions that voted for Trump.

Click to enlarge

Figure 1

Regional Distribution of Neuroticism, Economic Deprivation, and Voting (Counties)

Note. As described in the data section, counties with fewer than N = 100 respondents were dropped from the analysis and appear white on the maps. Trump 2016 gains = Gains in the two-party Republican vote share between the 2012 and 2016 elections; Trump 2020 gains = Gains in the two-party Republican vote share between the 2016 and 2020 elections. We include Sanders vote share in the 2016 primary elections as a point of comparison. In the maps depicting voting patterns, darker shades of blue represent greater vote shares and gains. In the Neuroticism x Economic Factor map (top left), dark blue represents counties that scored both above the median on neuroticism and above the median on economic deprivation; light blue represents counties that scored below the median on both neuroticism and economic deprivation; medium blue represents counties with either a combination of low neuroticism and high economic deprivation or high neuroticism and low economic deprivation.

It is also interesting to note the major differences between the Republican candidates in the primaries. Not a single independent variable consistently predicted voting for Republican primary candidates. Indeed, there is no overlap between the profile of regions that voted for Trump and those that voted for Rubio and Kasich. In fact, counties that voted for Kasich and Rubio tended to have less economic deprivation. However, the profile of counties that voted for Cruz is somewhat similar to those that voted for Trump in the general election, characterized by high neuroticism, low openness, high economic deprivation, and low ethnic diversity.

Interestingly, there are some differences between the determinants of Trump’s 2016 general election voting share and his primary voting share. While neuroticism and economic deprivation are significant predictors in both elections, neither ethnic diversity, health disadvantage, nor openness predicted Trump’s primary results. One reason for the discrepancies between primary and general election results could be that Trump’s campaign changed somewhat over time to appeal to additional voting groups. A related interpretation is that the regions that contributed to Trump’s nomination are not exactly the same as those that contributed to his general election win.

Summary of Findings with Robustness Checks

As noted in the Methods, spatial analyses require unique considerations such as including appropriate controls like state fixed effects, as we included in the models reported above. In addition, we conducted robustness checks to account for spatial autocorrelation (by running analyses with spatial lags on the independent variables using the normalized inverse of distances between regions), the potential MAUP problem (by conducting analyses at the CBSA level), and non-representativeness of the personality data (by weighting respondents by age and gender). Full results of these analyses are reported in the Supplementary Materials and summarized in Table 9.

As is evident from the table, the main effects of neuroticism and the structural factors on voting for Trump were consistent across nearly all robustness checks and dependent variables. Higher neuroticism, more economic disadvantage, and lower ethnic diversity predicted Trump voting. Worse health predicted more Trump gains but fewer Trump votes in 2016 and 2020. Of the 80 main effects of neuroticism and structural factors depicted in Table 9, only three robustness checks failed: Ethnic diversity did not predict Trump’s vote share in 2016 and 2020 at the CBSA level, and neuroticism did predict Trump gains in 2020 in the spatial autocorrelation setting. We also note that neuroticism did not predict Trump gains in 2020, and greater ethnic diversity predicted Trump’s gains in 2020, two findings which were consistent across robustness checks. In contrast to the robustness of the main effects, our focal interaction effects between neuroticism and the structural factors on Trump voting were less consistent and robust.

History Does Not Repeat but It Rhymes

Our findings suggest that Trump voting can be conceptualized as authoritarian voting and is explained by an array of regional factors that vary psychologically and structurally (Jost et al., 2003). Intellectually, this approach stems from the work of scholars from the Frankfurt School who sought to understand the roots of authoritarianism in the 20^th century. Our findings are particularly consonant with Frankfurt scholar Erich Fromm’s work. In his book Escape from Freedom (1941), which he published after leaving Nazi Germany and a decade before Adorno and colleagues published The Authoritarian Personality, Fromm “bridges the gap between economics and psychology and shows how no theory which invokes only man’s way of earning a living or man’s human nature alone is sufficient,” according to a review by Margaret Mead (Fromm, 1956). Like Fromm—whose book was titled ‘Fear’ of Freedom in the UK—we emphasize the importance of negative emotions and structural conditions in explaining people’s political choices

The work of scholars like Fromm, who sought to explain the rise of Nazism in the 20th century, has striking parallels in the modern-day context. Trump’s base may be different from the voter base of his contemporary political rivals. Still, his voter base is uncannily similar in some respects to the voter base of the historical figure who instigated the modern study of authoritarianism—Adolf Hitler. In particular, regions that voted for Trump were not those experiencing the greatest economic hardships, just as people who voted for Hitler were not those hardest hit by the Great Depression (King et al., 2008). Instead, both leaders found support among those who have suffered economically but still have further to fall. In Trump’s case, these regions have a high share of agriculture and manufacturing industries—regions that formerly drove the American economy but are struggling in the modern technology-based economy (as reflected by their low levels of income, college attainment, internet access, and migration). Importantly, these economically deprived regions are not those with the highest poverty or unemployment. Similarly, quantitative archival analyses of German voting records show that “those who were hurt by the economy but were at little risk of unemployment…constituted the groups that gave the most disproportionate support to the Nazis,” whereas “those who were unemployed or at high risk of becoming unemployed gave disproportionate support to the Communists or, to a lesser extent, to the Social Democrats” (King et al., 2008, pp. 952–953). That the working poor comprised much of the voter base of both Trump and Hitler is consistent with “last-place aversion” theories arguing that the second to last rung in the social ladder is a more psychological threatening position to occupy than the last rung (Gidron & Hall, 2017; Kuziemko et al., 2014).

Our factor analytic method painted a broad portrait of the threatening economic conditions that explain Trump’s vote share and his performance gains over time. But one variable from the economic deprivation factor stood out: college degree attainment. An examination of zero-order correlations showed that lack of college attainment correlated more strongly than any other variable with Trump’s vote share in 2016 and 2020, Trump’s gains in 2016 over Romney’s performance in 2012, and Trump’s gains in 2020 over his previous performance in 2016. Therefore, boosting rates of college attainment by increasing college enrollment or retention (e.g., Talaifar et al., 2021) might be an important way to mitigate the threat of authoritarianism in the U.S. (Scott, 2022). This strategy is likely to be particularly effective because higher education improves other economic outcomes (Chetty et al., 2020) and racial attitudes (Wodtke, 2012, 2018).

Intertwined Economic and Racial Roots of Trump Voting

As might be evident from the historical example described in the previous section, economic deprivation in Trump-supporting regions must be understood in the context of white people’s perceptions that minorities’ status is increasing and the antipathy that can result from such perceptions (Craig & Richeson, 2014; Wetts & Willer, 2018). Trump voting was not only disproportionally concentrated in regions that were high on economic deprivation but also in regions high on anti-Black bias and low ethnic diversity. That said, Trump voting was not higher in regions that had higher gender implicit bias or a more general preference for social inequality, as measured by social dominance orientation⁸. Together, these results suggest that purely economic accounts of Trump voting are insufficient and that antipathy toward specific ethnic minority groups is associated with Trump voting.

Importantly, our method was not designed to determine the relative influence economic deprivation versus racial resentment, as these factors are so intimately intertwined (McGovern, Kirkland, & Blake, 2021). For instance, White people interpret race relations as a zero-sum game, such that their economic hardships are viewed as minorities’ gains (Norton & Sommers, 2011). Another example of how economics and race relations are intertwined comes from research showing that contact with ethnic minorities may improve Whites’ attitudes towards minorities and decrease their Trump support, but only in the context of economic prosperity (Knowles & Tropp, 2018). Therefore, claiming that one structural factor is more important than another is overly simplistic. Instead, a strength of our work is its synthesis of economic, racial, and health factors, showing that threatening circumstances of many kinds jointly predict voting behavior. Indeed, to reflect the fact that economic, demographic, and health-related conditions are enmeshed in the real world, all our variables were allowed to load on all three factors, and factors were allowed to correlate with each other.

Our analyses show that more ethnically diverse regions preferred Democratic presidential candidates Clinton in 2016 and Biden in 2020, and that ethnically diverse regions were less likely to vote for Trump than they were to vote for Mitt Romney in 2012. However, Trump performance improved in ethnically diverse regions, and anti-Black implicit bias did not positively predict Trump’s gains in 2020 the way it did in 2016. Further work is needed to understand why Trump’s performance improved in ethnically diverse regions. One possible explanation is that the Republican effort in 2020 to paint Democratic rival Joe Biden as a socialist might have been effective in provoking fear among some Hispanic community members who may have had experience with socialist regimes in their countries of origin (e.g., Cuba, Venezuela). Another interpretation is more mundane – Trump may have already been so popular in ethnically homogenous regions that the only places where he could improve his performance were ethnically diverse.

Regarding regions high on health disadvantage, Democratic presidential candidates in 2016 and 2020 outperformed Trump in these regions overall. This finding is consistent with theories arguing that people vote for the leaders that they believe can best address the given threat at hand (Brandt & Bakker, 2022; Eadeh & Chang, 2020), since Democrats have typically been those to expand access to healthcare. But once again, Trump seems to be improving his performance in regions with poor health. He performed better in regions high on health disadvantage in 2020 than he did in 2016, and health disadvantage predicted Trump’s vote gains in 2016 over Romney’s performance in 2012.

Psychological Roots of Trump Support

Prior research shows an inconsistent relationship between neuroticism and right-wing voting behavior. Only recently has this relationship been robustly empirically documented (Obschonka et al., 2018). The recency and inconsistency of this link suggests that neuroticism may only be related to right-wing voting under certain threat-activating circumstances. We do not find strong evidence for this moderation hypothesis. Trump voting does not seem to be the result of regional neuroticism exacerbating the effects of threatening structural conditions, or of structural conditions activating the latent threat-sensitivity of regions high on neuroticism. Instead, we find that regional neuroticism and threatening structural conditions contribute in an additive rather than interactive way to Trump’s vote share and gains. This additive effect may seem surprising, but classic theories arguing that behavior is a function of both psychological and environmental factors did not specify exactly how psychological and environmental factors will combine to predict behavior (Kihlstrom, 2013; Lewin, 1951).

One promising avenue for future research would be to examine interactions at different levels of analysis. For instance, regional neuroticism may interact with threatening national conditions to predict voting behavior. Alternatively, individuals’ neuroticism may interact with threatening regional conditions to predict voting behavior. Importantly, individual voting behavior might have different antecedents than regional voting behavior. From the present research, we can only conclude that psychological and structural characteristics of one’s county and CBSA—not individual neuroticism or personal perceptions of threat—predict regional voting behavior. Drawing conclusions about individuals from this paper would be committing the ecological fallacy (Robinson, 1950). Moreover, it is possible that neuroticism has an opposite relationship to political behavior at the individual and regional levels. An individual’s own neuroticism may predict less voting for Trump (Fortunato et al., 2018; Samek, 2017) while living in a region with many other people experiencing high levels of negative emotionality (independent of one’s own experience) may predict more voting for Trump.

When it comes to other personality traits, openness and conscientiousness have, in past research, been most strongly linked to political ideology at the individual and regional levels (e.g., Garretsen et al., 2018; Gerber et al., 2010). Lack of openness to experience at the regional level was indeed robustly related to Trump’s vote share in 2016 and 2020, similar to prior research. But lack of openness was not consistently related to Trump’s gains over Romney, gains over his own performance, or voting for any primary candidate other than Cruz. Similarly, lack of conscientiousness at the regional level did negatively predict voting for Sanders, as expected. However, regional levels of conscientiousness were not consistently related to voting for Trump or other primary Republican candidates. We do not have a ready explanation for these findings, other than the fact that the strength of structural factors—explaining 25–40% of variance in voting—may overwhelm the ability of any single variable to explain even more regional variance in voting. That said, regional neuroticism consistently predicted Trump voting beyond the large effects of these structural variables.

Limitations

Although we attempted to control for a wide variety of potential confounds, the most obvious limitation of this work is the lack of evidence for a causal role for psychological and structural factors on Trump voting. Indeed, recent research suggests that personality traits and political preferences can influence each other bidirectionally (Bakker, Lelkes, & Malka, 2021; Luttig, 2021). Moreover, one recent study showed that exposure to authoritarian speeches and messages causally increases meaning in life (Womick et al., 2021) but actually lowers positive affect and increases negative affect. If this is the case, then authoritarianism may support people’s existential needs even as it worsens their emotional experience or, over the long-term, their neuroticism.

Moreover, the factors we studied are highly stable over time at the regional level, suggesting that the roots of Trump support have been long in the making (as was also the case for the roots of Nazi support; De Bromhead, Eichengreen, & O’Rourke, 2013). Such stable structural variables cannot be easily manipulated by experimenters. Even if feasible, manipulating psychological and structural factors in the lab or the field to influence voting behavior would likely be unethical. One possible solution would be to leverage quasi-natural experiments. For example, economists have made causal inferences about the influence of trade with China on voting by exploiting the quasi-random regional distribution of import competition (e.g., Autor et al., 2020).

Neuroticism and threatening economic, demographic, and health conditions are not the only psychological and structural variables that may explain Trump support. With regard to psychological variables, we emphasized the role of “hot” emotional factors, but “cold” cognitive factors like cognitive ability and rigidity have also played a role in recent voting behavior (Choma & Hanoch, 2017; Ganzach, Hanoch, & Choma, 2019; Zmigrod, Rentfrow, & Robbins, 2018). Moreover, there are other personality traits (e.g., the Dark Triad) and measures of group-based dominance (e.g., explicit hostile and benevolent sexism) that we did not examine but which have been linked to Trump support (Glick, 2019; Yalch, 2021). Regarding structural factors, we emphasized the role of fairly enduring structural factors rather than more immediate features of the electoral context, such as media coverage (Reuning & Dietrich, 2019) or the emergence of the COVID-19 pandemic (Baccini, Brodeur, & Weymouth, 2021; Clarke, Stewart, & Ho, 2021), which some researchers have argued accounted for Trump’s win in 2016 and loss in 2020, respectively. We also did not consider the role of political polarization or threats from the outparty, foreign actors, or climate change, which may also increase authoritarianism (Fritsche et al., 2012; Graham & Svolik, 2020; Stanley & Wilson, 2019). One especially important avenue for future research is to understand how social media platforms and misinformation may amplify or manufacture the salience of all of these threats while allowing people to avoid information that challenges their political views (Ashokkumar et al., 2020).

One limitation of our analytic approach is that it ignores third party votes, which might have an influential effect on election outcomes. Furthermore, our focus was on predicting voting for Trump, but vote choice is just one component of what determines election outcomes (Grimmer & Marble, 2019). Voter turnout is also important for determining the winner of an election. Turnout in the 2020 election was much higher than in 2016 for both parties, but Democrats increased turn out more than Republicans did. Future research could investigate whether neuroticism and threatening structural variables also predict voter turnout. Evidence that anger is a critical negative emotion driving turnout (Lambert, Eadeh, & Hanson, 2019) suggests that neuroticism, which is associated with anger (Maciantowicz & Zajenkowski, 2020), may also be related to greater voter turnout. However, existing research on neuroticism seems to show that neuroticism is actually related to lower voter turnout and other forms of political participation (Gerber et al., 2011b).

Conclusion

Around the world voters in democratic regimes are electing leaders who “routinely ignore constitutional limits on their power” (Zakaria, 1997, p. 22). The election of Trump, who flouted an array of democratic norms, was perhaps the most conspicuous example of this trend. Even after losing the 2020 election, Trump maintained a strong hold on the Republican party, with some polls estimating that more than 70% of Republicans thought he should run for president again (Rakich & Wilkes, 2021). Here, we proposed that researchers must consider both psychological and structural factors to gain a holistic understanding of the roots of Trump’s appeal. We find that the prevalence of neuroticism in a region and its threatening economic, demographic, and health conditions both predict voting for Trump in 2016 and 2020. It remains to be seen whether other authoritarian leaders will emerge to capitalize on these same psychological and structural dynamics in the U.S. and elsewhere.