Trump Gained among Low-SES and Conservative Latinos

Grimmer, Marble, and Tanigawa-Lau (Reference Grimmer, Marble and Tanigawa-Lau2022) contend that while models focused on changes in vote choice across elections can identify shifts in candidate support, assessing how these shifts are translated into vote totals requires a different approach. A bloc’s contribution to election outcomes depends on three components—turnout, vote choice, and composition. Simply knowing if a voting bloc became more likely to vote for a candidate between elections is insufficient for knowing if that bloc produced a net increase in that candidate’s vote total. As Grimmer, Marble, and Tanigawa-Lau (Reference Grimmer, Marble and Tanigawa-Lau2022) show, one can estimate this “net votes” quantity within a given voting bloc x using the following equation:

(1) $$ \begin{array}{rl}\begin{array}{r}\mathrm{Diff}\ \mathrm{Net}(x)={\mathrm{Net}\ \mathrm{Trump}}_{2020}(x)-{\mathrm{Net}\ \mathrm{Trump}}_{2016}(x)=\\ {}({\displaystyle \underset{\mathrm{Vote}\ \mathrm{choice}\ \mathrm{difference}\ (2020)}{\underbrace{p{(\mathrm{Trump}\hskip0.15em \mid \hskip0.15em \mathrm{turnout} = 1,x)}_{2020}\hskip0.15em -\hskip0.3em p{(\mathrm{Biden}\hskip0.15em \mid \hskip0.15em \mathrm{turnout} = 1,x)}_{2020})}}}\overset{\mathrm{Turnout}\ \mathrm{rate}\ (2020)}{\overbrace{p{(\mathrm{turnout} = 1\hskip0.15em \mid \hskip0.15em x)}_{2020}}}{\displaystyle \underset{\mathrm{Composition}\ (2020)}{\underbrace{p{(x)}_{2020}}}}-& \\ {}{\displaystyle \underset{\mathrm{Vote}\ \mathrm{choice}\ \mathrm{difference}\ (2016)}{\underbrace{(p{(\mathrm{Trump}\hskip0.15em \mid \hskip0.15em \mathrm{turnout} = 1,x)}_{2016}\hskip0.15em -\hskip0.15em p{(\mathrm{Clinton}\mid \hskip0.15em \mathrm{turnout} = 1\hskip0.15em ,x)}_{2016})}}}\overset{\mathrm{Turnout}\ \mathrm{rate}\ (2016)}{\overbrace{p{(\mathrm{turnout} = 1\hskip0.15em \mid \hskip0.15em x)}_{2016}}}{\displaystyle \underset{\mathrm{Composition}\ (2016).}{\underbrace{p{(x)}_{2016}}}}\end{array}& \end{array} $$

This equation clarifies the necessary components for calculating if a candidate gained votes from a bloc over time. The first component captures the percentage point difference in vote choice between Trump and his Democratic competitor within voting bloc x, the second component is x’s turnout rate, and finally, the third component is the share of the Latino electorate in voting bloc x. (See Appendix A.2 of the Supplementary Material for further explanation of this decomposition method, and Appendix A.3 of the Supplementary Material for an explanation of the utility of this decomposition for our understanding.)

To better understand the role of different Latino voting blocs in 2020, we estimate Latino-specific survey weights using entropy balancing (Hainmueller Reference Hainmueller2012) and apply the Grimmer—Marble–Tanigawa–Lau (GMTL) decomposition to key political and demographic subgroups using data from the 2016 and 2020 Cooperative Election Study (CES). The principal advantages of the CES are the size of its Latino sample ( $ {N}_{2016} $ = 7,495; $ {N}_{2020} $ = 6,978) and the inclusion of turnout and voter validation data. Given that we aim to make inferences about Latinos, we use entropy balancing to estimate Latino-specific weights using data from the 2016 and 2019 American Community Survey (ACS).Footnote ⁵ We assess if Trump observed increases in net votes from 2016 to 2020 among Latino subgroups based on age, sex, income, education, ancestry, generational status, partisanship, ideology, religion, crime policy attitudes, immigration attitudes, and social media usage. Given inconsistent survey items across CES surveys, item response theory (IRT) was used to place respondents on the same latent scale through the use of common items present in both 2016 and 2020 (see “Additional Study Details” documentation on Dataverse for question wording and scale construction details; Fraga, Velez, and West Reference Fraga, Velez and West2024).

Figure 1 presents estimates of net vote Trump increases from 2016 to 2020 with bootstrapped 95% confidence intervals.Footnote ⁶ Positive estimates indicate shifts in relative Trump support from 2016 to 2020, whereas negative estimates indicate shifts toward Biden. Observable shifts in votes for Trump from 2016 to 2020 were mostly contained within partisan, religious, ideological, and attitudinal voting blocs, such as Catholic, restrictionist, pro-police, partisan, and ideologically conservative Latinos. That is, indicators for alignment with the Republican party most strongly predict Latino vote shifts. This suggests that shifts are sustainable and not necessarily specific to Trump or 2020. We also explicitly examine election-specific effects, such as reactions to COVID-19 and BLM protests and do not find support for these alternative explanations of Latino vote shift (see Appendices A.4 and A.5 of the Supplementary Material). Our social media analysis presented in the first panel in the third row of Figure 1 also suggests misinformation may have had muted effects on net votes.Footnote ⁷

Figure 1. Increase in Net Republican Votes from 2016 to 2020 among Different Latino Subgroups

Note: Positive scores indicate that Trump gained votes within a subgroup across elections. Negative scores indicate that Biden gained votes (relative to Clinton) across elections. Estimates are calculated for Latinos only.

Patterns for demographic voting blocs were smaller, with considerable uncertainty in the estimates. Net vote increases of 2 percentage points (pp) were observed among the least educated and lowest income quartile. Those with a college degree provided Biden with a net vote increase of approximately 1pp. This is consistent with a shift in the electorate in general, and thus again indicates the rightward shift among Latinos may be sticky. We observe suggestive evidence of a shift toward Trump among first-generation Latinos (i.e., American-born children of immigrants) (p = 0.10). Shifts toward Trump according to age, sex, social media use, ancestry, or geographic region are less discernible.

In Figure 2, we display the different components of the “net votes” estimand for the voting blocs that had statistically discernible shifts (see “Additional Study Details” documentation on Dataverse for estimates). The figure displays percentage point changes from 2016 to 2020 with respect to the different components: turnout rate, Trump support, and group size. The closer the estimate of a component is to zero, the less likely it is to be an explanation for increases in net votes. Positive (negative) estimates for the turnout rate indicate that the voting bloc increased (decreased) its turnout from 2016 to 2020, positive (negative) estimates for the subgroup Republican support measure indicate that the subgroup increased (decreased) its Trump support from 2016 to 2020, and positive (negative) estimates for the composition measure indicate that a group grew (shrunk) as a share of the electorate.

Figure 2. Decomposition of Net Votes into Changes in Turnout Rate, Composition, and Vote Choice from 2016 to 2020 for Voting Blocs Demonstrating Discernible Shifts toward Trump

Note: Estimates are calculated for Latinos only.

Shifts in Trump vote choice from 2016 to 2020 help explain net vote increases among voting blocs defined by conservative crime policy attitudes, generational status, and ideology. Turnout increases from 2016 to 2020 appear to be responsible for the net vote increases for Biden among those with a college degree. In other cases, a combination of changes in turnout, Trump vote, and/or composition is responsible for the observable shifts within voting blocs. For example, those scoring at the lower end of immigration restrictionism had a higher turnout rate in 2020 than 2016 and increased their support for Biden, but became a smaller proportion of the electorate. Those scoring at the upper end of the scale became more numerous, increased turnout, and increased Trump support in 2020 over 2016. Relative increases in Trump support among first-generation Latinos can be explained by a mixture of increased turnout and increased Trump vote choice, whereas shifts among low-income voters can mostly be attributed to vote choice. Gains among Catholics can be explained by changes in turnout and vote share, whereas gains among atheists/agnostics can mostly be explained by increases in turnout.

We find changes in vote choice among low SES and conservative Latino voting blocs generated increases in net votes for Trump. This dovetails with trends among white voters seen since 2016, and suggests that this rightward 2020 shift among Latinos may stick. In contrast, mobilization among voters with stable voting patterns who were already opposed to Trump (e.g., self-identified Democrats, liberals, and college-educated voters) contributed to decreases in his vote totals. Our estimates are consistent with ideological sorting, rather than an increase in the share of conservative Latinos. While a threat-mobilization process may have driven opposition to Trump among those already predisposed (e.g., Democrats, college-educated Latinos) (Gutierrez et al. Reference Gutierrez, Ocampo, Barreto and Segura2019; Ocampo, Garcia-Rios, and Gutierrez Reference Ocampo, Garcia-Rios and Gutierrez2021; Pantoja and Segura Reference Pantoja and Segura2003), this effect did not extend to more conservative or less politically active Latinos.

Shift to Trump was Geographically Broad/Concentrated in Areas with Low-SES, Newly Activated Latinos

To further probe Latino voter shifts, we rely on official records of turnout and election results at the sub-county level. We identify the population of 2020 voters using individual-level voter file records from TargetSmart, a vendor that compiles voter registration and vote history data in each state, geocoding registrants’ addresses and using a combination of given name, surname, and geographic context to model individual race/ethnicity. We aggregate the number of voters in 2020 by voting behavior in the 2018, 2016, and 2014 elections, along with sums of the modeled probabilities of voter race/ethnicity, to the Census tract level. We reaggregate precinct-level 2016 and 2020 election results produced by the Voting and Election Science Team (2018; 2020) to the 2010 Census tract level, merging the resultant election results with the voter file-derived turnout totals.

Figure 3 indicates the increase or decrease in Trump’s two-party 2020 vote share as compared to the 2016 election at the Census-tract level. There are broad gains in Trump vote share in neighborhoods with substantial numbers of Latino voters. Here, the trend is monotonically rising from 25% Latino onward, with an 80% Latino tract seeing a roughly 15pp increase in Trump two-party vote share between 2016 and 2020. In Appendix A.7 of the Supplementary Material, we show that these gains can be observed even outside Florida and Texas.

Figure 3. Census Tract-Level Trump 2020 Two-Party Vote Share Minus Trump 2016 Vote Share as a Function of Percent Latino Voters

Note: Blue line represents a loess smoother. Red dashed line indicates 0 change in Trump two-party vote between 2016 and 2020.

Which factors predict increased Trump vote share in the voter file data? Table 1 presents estimates from a weighted linear regression on the Census-tract level election and turnout data.Footnote ⁸ The dependent variable is the 2020 Trump share of the two-party vote, and % Latino is the modeled share of tract voters in 2020 who were Latino. Model 1 indicates a linear decrease in Trump support as the share of Latino voters increases. That is, at baseline, Democrats perform better than Republicans in heavily Latino areas. However, once we control for Trump’s vote share in 2016 at the tract level (Model 2), the relationship reverses and the percent of voters who are Latino in a Census tract positively predicts Trump gains in 2020. Model 3 adds a voter file-derived variable related to previous voting history: the percent of Latino voters in the tract who could have voted prior to the 2020 election,Footnote ⁹ but have no recorded vote history before Trump’s re-election campaign. After controlling for 2016 Trump vote share, and the overall percent Latino voter in the tract, the percent of Latinos who were first-time voters in 2020 significantly predicts an increase in Trump vote share. In conjunction with Figure A.5 in the Supplementary Material, this implies a newly activated group of Latino voters produced some of Trump’s raw gains.

Table 1. Census Tract-Level Regressions Predicting Trump 2020 Two-Party Vote Share

Note: Unit of observation is the 2020 Census tract. Estimates derived using a weighted least-squares model. * p < 0.05, ** p < 0.01, *** p < 0.001.

Leveraging the Census tract-level aggregated data, Models 4–6 of Table 1 add estimates from the Census American Community Survey 2016–2020 5-Year data. These additional variables proxy for individual-level attributes discussed in the net-votes analysis. In line with the GMTL decomposition, median household income for Latino-led households in the tract is associated with a significant decrease in Trump support. Similarly, as percent Latino non-college increases, Trump’s vote share also increases. Both of these corroborate the story that lower-SES Latinos were a source of increased 2020 Trump support, even after accounting for his performance in the same Census tracts in 2016.

Model 5 of Table 1 attempts to capture generational dynamics. Recall that in the GMTL decomposition, there was a large, though imprecisely estimated, boost in Trump support among individuals who indicated that they were the children of immigrants. The Census Bureau does not ask about generational status directly. Instead, we use three variables in an attempt to establish how personal proximity to the immigration experience predicts an increase in Trump support: percent immigrant, which uses the total Latino population in the tract as the denominator; percent of Latino immigrants who are naturalized, an interaction between these variables that should expose the independent effect of the Latino immigrant voting-eligible population; and the percent of native-born Latinos who report that they speak English less than “very well.” This final measure speaks to the size of the less “acculturated” U.S. or Puerto Rican-born Latino population within a Census Tract.

The results again provide evidence of shifting loyalties among Latinos proximate to the immigrant experience. In places with more immigrants and a larger share of potential immigrant voters (the interaction term), Trump support 2016–2020 increased significantly, implying that immigrants were a source of Trump gains. Yet, the independent effects of the Latino immigrant population, Latino immigrant naturalization rate, and the percent of the native-born population that is limited English proficient tell a different story. Places with many noncitizen Latino immigrants did not see an increase in Trump support in 2020. Tracts with a high naturalization rate, but few Latino immigrants overall, saw relatively lower levels of Trump support in 2020 compared to 2016. A larger share of U.S or Puerto Rican-born Latinos who exhibit limited English proficiency predicts an increase in Trump support. Model 6 demonstrates that these heterogeneous estimates persist after controlling for education and income. These results offer tentative evidence that, all else equal, places with a less acculturated and lower-SES Latino population were disproportionately likely to shift toward Trump in 2020. We thus confirm some of the demographic correlates of increased Latino support found in the GMTL decomposition.