Local demographic changes and US presidential voting, 2012 to 2016

Significance

In recent years, advanced industrial democracies have grown more ethnically and racially diverse. This increasing diversity has the potential to reshape voting behavior in those countries, in part because majority groups may react by shifting support toward anti-immigration candidates and parties. This paper considers whether local demographic changes in the United States were associated with pro-Republican shifts between 2012 and 2016, when the Republican presidential candidate was especially outspoken in opposition to immigration. By showing that demographic changes were not associated with shifts toward the Republican, this research indicates that local demographic changes are not on their own increasing support for anti-immigration candidates.

Abstract

Immigration and demographic change have become highly salient in American politics, partly because of the 2016 campaign of Donald Trump. Previous research indicates that local influxes of immigrants or unfamiliar ethnic groups can generate threatened responses, but has either focused on nonelectoral outcomes or analyzed elections in large geographic units, such as counties. Here, we examine whether demographic changes at low levels of aggregation were associated with vote shifts toward an anti-immigration presidential candidate between 2012 and 2016. To do so, we compile a precinct-level dataset of election results and demographic measures for almost 32,000 precincts in the states of Florida, Georgia, Michigan, Nevada, Ohio, Pennsylvania, and Washington. We employ regression analyses varying model specifications and measures of demographic change. Our estimates uncover little evidence that influxes of Hispanics or noncitizen immigrants benefited Trump relative to past Republicans, instead consistently showing that such changes were associated with shifts to Trump’s opponent.

How is increasing ethnic and racial diversity reshaping the electoral politics of advanced industrial democracies? Recent elections in the United States, the United Kingdom, France, Italy, and elsewhere have brought this question to the foreground, as candidates and parties have found success while amplifying concerns about immigration and demographic change (1–3). Some scholars contend that growing ethnic and racial diversity has the potential to upend traditional political divisions over economic issues by realigning voting patterns on the basis of ethnicity, nativity, nationalism, and education (4–6).

At first glance, Donald Trump’s unexpected 2016 victory seems consistent with this trend: His support was related to his outspoken opposition to immigration (7). Even so, the hypothesis that increasing ethnic and racial diversity fuels support for Trump and other populist, anti-immigration candidates is difficult to test empirically. While advanced industrial democracies have grown more ethnically and racially diverse in recent decades, they have also experienced other large social and economic changes, such as greater exposure to international trade and declining economic prospects for the less educated. These changes provide alternative explanations for the success of populist and anti-immigration politicians. Unfortunately, 1-time shifts in overall national election results provide little leverage to disentangle multiple simultaneous causes.

Instead, scholars interested in the effects of changing demographics and ethnicity have sometimes considered local-level variation. Because the United States is a large and diverse country, some localities have seen substantial influxes of immigrants and/or associated pan-ethnic groups, while many others have not. Studying responses to local demographic changes thus provides substantially increased statistical power with which to address 1 specific set of hypotheses about demographic change and voters’ lived experiences in their communities.

Local demographic changes are critical in certain theories of anti-immigration attitudes. Local population changes are better measures of the local, community-level experiences that individuals have in their everyday lives (see especially refs. 8–13). Indeed, prior work on the United States finds that local demographic changes are associated with a range of outcomes, including anti-immigration attitudes, hate crimes, increased voter turnout, and opposition to antidiscrimination laws and local bond measures (refs. 13–20, with ref. 20 a meta-analysis spanning developed democracies). Research on the United Kingdom has found that support for Brexit, the UK Independence Party, and reducing immigration are higher in localities that have low immigrant shares but recent demographic changes (21–23), with related research in continental Europe (24, 25).

To date, though, there has been less research on the overall impact of local demographic changes on American partisan election outcomes. The importance of partisan attachments to voting, combined with growing elite polarization, may limit the capacity of local immigration concerns to shape elections. In addition, despite its disparate local impacts, immigration may be a symbolic, nationalized issue whose effects do not depend on local experiences. While prior research has focused overwhelmingly on the negative reactions of native-born Americans in receiving communities, it is also possible that the average citizen may react positively, even while some subgroups react negatively.* Proximity may also lead to positive intergroup contact, especially at low levels of aggregation (26). Moreover, while prior work has focused on short-term responses to changing demographics, the long-term effects are less clear as natives adapt to their changing communities (see especially ref. 17).

Existing research on demographic changes and electoral outcomes has been inconclusive. When analyzing federal election outcomes between 1990 and 2010, ref. 27 reports that county-level increases in low-skilled immigrants are associated with pro-GOP shifts, while increases in high-skilled immigrants are associated with the reverse. Similarly, ref. 13 identifies a county-level association between the percentage change in the Hispanic population and shifts to the GOP from 2012 to 2016. But ref. 28 does not find a similar relationship in survey data, and ref. 29 uses survey data to show that the relationship between local demographic change and Trump favorability among Republicans was time-dependent. The latter studies share a common design in taking advantage of the presence of an anti-immigrant candidate, Trump, to understand microfoundations of who or where support moves in response to anti-immigrant rhetoric.

Prior research on US election outcomes has overwhelmingly employed county-level measures, perhaps because counties have fixed boundaries and readily available data. But a county is a large aggregate, particularly in more populated places. Moreover, county-level changes are unlikely to capture the hyperlocal community experiences that some theories of immigrant threat suggest are critical. While counties may offer effective tests of threats stemming from labor-market competition or media-market coverage, they are likely too large to measure the more experiential mechanisms through which local contexts may operate (e.g., refs. 10, 12, and 30).

Here, we move analysis to a lower level of aggregation that may more closely approximate neighborhoods as envisioned by theories of threat operating through local experience.^† We combine precinct-level election returns and tract-level Census data to generate almost 32,000 precinct-level observations of electoral changes from 7 states: Florida, Georgia, Nevada, Michigan, Ohio, Pennsylvania, and Washington. Four of these states were electorally crucial in 2016 and closely contested—Florida, Nevada, Michigan, and Pennsylvania—with the remaining 3 also modestly competitive. Four states flipped from supporting the Democrat in 2012 to Trump in 2016, perhaps making them informative about the general pattern of change in party performance. The states varied demographically and geographically. They included some of the northeastern and midwestern battlegrounds that allowed Trump to win the electoral college despite losing the national vote (Michigan, Ohio, and Pennsylvania), as well as competitive southern states (Florida and Georgia), states with sizable Hispanic populations (Florida and Nevada), and western states (Nevada and Washington). These states include more than 77 million residents, making them home to nearly one-quarter of the US population. Their demographic diversity roughly mirrors that of the nation as a whole, although none of these states have aggressive contemporary anti-immigration policy efforts.

In our analysis, we focus primarily on the relationship between changing party vote shares from 2012 to 2016 and changes in the local Hispanic population. To an important extent, Hispanics have become the public face of contemporary immigration (31). We then specify a wide range of regression models in which we examine the conditional associations between changes in the Hispanic population and changes in presidential voting between 2012 and 2016. We also consider the noncitizen foreign-born population as an alternative measure of local demographic change. This robustness check proves valuable, as it demonstrates that our results are not driven by the voting patterns of the newcomers themselves because noncitizens are ineligible to vote. To be sure, any results could be driven by the idiosyncrasies of the 2012 candidates as surely as those of the 2016 candidates. But in interpreting our findings, we rely on prior research, such as ref. 7 emphasizing Trump’s strident anti-immigration position in 2016 as both unusual and salient.

Across specifications, time intervals, and measures, we consistently find that increasing local ethnic diversity and immigrant populations were not associated with shifts toward the anti-immigration candidate. To the contrary, we find that localities with these characteristics shifted toward his opponent, the pro-immigration Democrat Hillary Clinton. To the extent that local demographic changes caused threatened responses, these responses do not appear to have, on balance, benefitted the anti-immigration candidate.

Data and Measurement

We present the full details on the construction of the dataset in SI Appendix, section A and summarize key elements here. Our goal is to isolate the conditional association between demographic changes and election-to-election shifts in partisan support in precincts. To do so, we combine precinct-level election returns with tract-level Census data. The median precinct in our dataset has a 2016 population of 4,623, compared to a median county population in the United States of 25,839. Precinct-level measurement provides substantial increases in statistical power and is likely to more accurately measure residents’ local experiences.

The data-acquisition and -preparation work involved in generating precinct-level measures is substantial, explaining why our analyses focus on only 7 states. We first collected precinct election returns from each state for the 2012 and 2016 elections. We next identified precincts that had fixed boundaries over the 4 years to avoid incomparable geographies. We then merged tract-level demographic and economic measures from the 2000 decennial Census and several American Community Surveys (ACSs) with our precinct-level election returns. Census tracts do not perfectly overlap with precincts, so we used the set of registered voters’ addresses in each precinct to allocate tract demographics proportionally to precinct registration. SI Appendix, Table S1 presents summary statistics. In total, our data represent 28.9 million votes cast in 2016. Data, code, and materials for reproducing all results in this paper are available at https://doi.org/10.7910/DVN/J5GCZQ (32).

Measuring Demographic Change.

Prior research provides valuable guidance on estimating the effect of demographic changes on voting (e.g., ref. 20), but still leaves key questions unanswered. One is the appropriate measure of demographic change. For example, in some instances, the relevant measure of demographic change might be the proportional increase in the Hispanic population; even a small number of Hispanics might be influential if they represent a sudden increase from a low baseline (e.g., ref. 13). In other instances, the relevant measure might instead be the increase in the Hispanic share of the population, or even the number of new Hispanic residents. In any case, effects may be nonlinear, as especially large changes might generate disproportionate levels of threat. Estimating the effects of demographic changes also requires researchers to make other choices, including the relevant time period and the geographic scope in which to measure demographic changes.

Given that prior work has not settled on a single, definitive measure, we measured influxes of Hispanics (and, later, noncitizen immigrants) using multiple empirical approaches. As we were analyzing electoral change from 2012 to 2016, we first measured change in the proportion of the total population in the precinct that is Hispanic—and, in alternate specifications, noncitizen foreign-born—from 2011 to 2016. Increasing values of this fraction indicate that the Hispanic share of the local population has increased relative to the non-Hispanic population. Second, we considered proportional change in the Hispanic population, which measures population growth as a fraction of the group’s original population. In this measure, the size of an influx is weighted by the inverse of the baseline population, so, for example, an increase of 100 Hispanic residents is a larger shift if the baseline were 200 than 500.

Researchers studying the electoral impacts of demographic changes must also choose the window of time over which to measure those changes. Prior research on demographic changes in the United States typically used 10-year windows, but did so for reasons of convenience: Until the last decade, the best available measures of local demographics were from the decennial Census. The ACS now provides within-Census estimates at low levels of aggregation. Here, we coupled ACS data and Census data to construct measures of demographic changes for 2000 to 2016 and 2011 to 2016. These windows capture 2 theoretically distinct characterizations of immigrant threat, either of which is plausible. The 1st is the idea that cumulative, long-term changes in local demographics may create citizen unease that was activated by Trump’s candidacy. The 2nd represents a characterization in which more recent changes are most salient for individuals’ perceptions of their local communities. Our goal is to provide readers with a variety of measures that one might map to the theoretical construct of interest so that our results are not dependent on specific choices about measurement. We consider different geographic scopes by expanding the Census tracts used to calculate demographics in SI Appendix, section E.

Results: Change in Republican Vote Share and Change in Hispanic Population.

In Fig. 1, we examine how changes in Hispanic populations correlate with increases in Republican precinct-level vote share between 2012 and 2016. We plot change in the Republican share of the 2-party vote from 2012 to 2016 (positive values indicate pro-Republican shifts) against 4 different measures of change in the Hispanic population on the x axis. The 1st frame measures changing population as the change in the Hispanic proportion of the overall population from 2011 to 2016, the 2nd as the same change from 2000 to 2016, and the 3rd and 4th as proportional changes in the Hispanic population for each period.

Fig. 1.

Change in Republican vote share, 2012 to 2016, and change in Hispanic population. Note: Points are random samples of 2,000 precincts. Loess lines are generated from all observations. Points are shaded corresponding to density, with darker colors indicating more precincts.

In contrast to demographic change driving voters toward Trump, the figure shows a negative relationship between increasing Hispanic populations and heightened Republican support. This association holds for either the between-election time period of 2011 to 2016 or the longer time period of 2000 to 2016. Proportional changes in the 3rd and 4th frames both show a flat relationship between proportional change and change in Republican support.^‡The slope in Fig. 1, Lower Right is positive for proportional changes greater than 1. However, the corresponding regression models illustrate that this result disappears when one accounts for the base-rate Hispanic in the precinct. In other words, the apparent positive relationship is driven by failing to account for initial levels.

In Table 1, we present multiple least-squares regression estimates of these relationships. The columns present our 4 measures of local context and different sets of control variables to probe robustness to measures and specifications. Columns with additional Census controls (indicated by the row “Additional Census Controls” at the bottom) include measures of 7 other changes that may be associated with influxes of Hispanic residents and/or shifts in voting: population proportion poor, unemployed, and employed in manufacturing, change in overall population, change in average rent, change in rent as proportion of household income, and change in proportion owner-occupied housing valued at less than $150,000. The time interval used for each control variable is the same as that for the measure of Hispanic or immigrant context in the column. Columns with controls for levels in the base year (indicated by the row “Control for levels”) include the proportion of Hispanic, poor, unemployed, employed in manufacturing, Black, and with a bachelor’s degree or higher, as well as population density, average rent, rent as a proportion of household income, and the proportion of housing valued at less than $150,000. We also include county fixed effects to account for time-invariant features of counties, in which precincts nest. Finally, the row “Republican Vote Share” indicates whether or not we control for 2012 Republican presidential vote share in the precinct, entered as indicators by decile.

Table 1.

Change in Republican vote share 2012 to 2016 and change in Hispanic population, various time intervals

Variables	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
Change in prop. Hispanic, 2011 to 2016	−0.040** (0.01)	−0.071** (0.01)	−0.077** (0.01)
Prop. Hispanic 2011		−0.13** (0.00)	−0.15** (0.00)	−0.15** (0.00)
Prop. change in prop. Hispanic, 2011 to 2016				−0.0041** (0.00)
Change in prop. Hispanic, 2000 to 2016					−0.077** (0.01)	−0.047** (0.01)	−0.085** (0.01)
Prop. Hispanic 2000						−0.13** (0.00)	−0.14** (0.00)	−0.15** (0.00)
Prop. change in prop. Hispanic, 2000 to 2016								−0.0055** (0.00)
Observations	31,949	31,352	31,352	31,352	31,949	31,949	31,949	31,949
R-squared	0.001	0.658	0.704	0.704	0.004	0.649	0.689	0.687
Control for levels	No	Yes	Yes	Yes	No	Yes	Yes	Yes
County fixed effects	No	Yes	Yes	Yes	No	Yes	Yes	Yes
Additional Census controls	No	Yes	Yes	Yes	No	Yes	Yes	Yes
Republican share 2012	No	No	Yes	Yes	No	No	Yes	Yes

Robust standard errors are in parentheses. *P < 0.05; **P < 0.01. Precinct-level analysis; weighted to number of votes 2012; proportional changes top and bottom coded at 1 and −1. Note: Dependent variable is change in GOP vote share, 2012 to 2016. Prop., proportion.

Across specifications, time intervals, and measures, the results consistently show that increases in the Hispanic population are associated with shifts toward the pro-immigration candidate Clinton in 2016. Our first measure is change in Hispanic population share from 2011 to 2016. The coefficient in the 1st column indicates that a 1-SD increase in this measure (0.039) corresponds to a 0.16-percentage-point increase in Clinton’s vote share. A 1-SD increase in the change in Hispanic population share from 2000 to 2016 (0.055) corresponds to 0.5 percentage points for Clinton per the column 7 specification with all controls. The coefficient estimates for proportional changes (columns 4 and 8) present similar relationships, and the CIs for all estimates exclude the positive values that would indicate threatened responses. The evidence in Table 1 suggests that increases in the Hispanic population were associated with shifts to Clinton in 2016.

In Fig. 2, we present coefficient estimates from the specifications in Table 1 along with a set of additional specifications indicated in the figure’s note. Across specifications, time intervals, and measures of demographic threat, CIs in only 1 specification—without controlling for base rates—cross 0 into positive values. The anti-immigrant candidate does not appear to have benefited from recent or longer-term local demographic or immigrant population changes.

Fig. 2.

Variation in magnitude of coefficient relating change in Hispanic population to change in Republican vote share by model specification and time interval. Note: The figure demonstrates that in no specification or time interval does change in Hispanic population benefit Republican presidential vote. Each point is the coefficient estimate from that model with lines showing 95% CIs. Proportional changes are divided by 10 to scale with changes in levels. Model numbers on the y axis correspond to varying model specifications. See SI Appendix, section K for details of each.

In SI Appendix, section F, we reproduce Table 1 and Fig. 2 using the noncitizen foreign-born population to measure immigrant threat.^§ The results are consistent with those for the Hispanic population: Irrespective of measurement choices, increases in the noncitizen foreign-born population correspond to increasing Democratic vote share when controlling for the base rate. Noncitizens are ineligible to vote, so these results diminish the possibility that the overall pattern is driven by changes in the local electorate.

Heterogeneity by Population Density and Trade Exposure.

One theory of demographic change is that its effects are stronger in rural areas and in areas negatively exposed to international trade. In SI Appendix, section C, we consider both of these possibilities by looking at subsets of precincts separated by density and trade exposure (year 2000 exposure from ref. 33). We find very limited heterogeneity. Even in low-density or high-trade-exposure places, increasing Hispanic population benefitted Clinton.

Robustness to Economic Disadvantage, Homogeneous Precincts, Nonlinearity, Political Geography, Scope of Geographic Context, and State Subsets.

In SI Appendix, sections B through I, we provide additional tests of the robustness of our findings to potential omitted variable bias or model misspecification. We show the following: The results hold even in more economically disadvantaged precincts (SI Appendix, Table S2); measuring demographic change at geographic scopes larger than the precinct’s Census tracts (by including tracts within 1, 5, and 10 miles of the precinct’s Census tracts when calculating demographic composition and change) does not alter our findings (SI Appendix, Table S6); limiting analysis to nondiverse precincts does not change the negative relationship between demographic change and movement toward Trump (SI Appendix, Table S8); splitting the sample into deciles of 2012 Republican presidential vote produces the largest correlations in the most Republican precincts, inveighing against the result being driven by Hispanics moving exclusively to Democratic strongholds (SI Appendix, Table S9); allowing a nonlinear relationship continues to produce a negative or flat relationship (SI Appendix, Table S10); and the negative or flat relationship is consistent within each of our 7 states (SI Appendix, Tables S11 through S17).

Conclusion

Recently, extensive evidence has connected demographic changes to attitudinal or behavioral shifts in developed democracies. In 2015 and 2016, presidential candidate Trump heightened the salience of demographic changes in the United States as he made opposition to immigration a central pillar of his candidacy. Some prior scholarship suggests that it’s precisely under these conditions—local demographic changes coupled with salient national rhetoric—that attitude changes are likely (e.g., ref. 15). Influxes of people from different ethnic or racial backgrounds are thought to induce divisive local contestation over communities’ identities. Yet, if anything, our evidence suggests that local demographic changes are consistently associated with reduced support for Trump. Across 7 states, including 4 battleground states where the campaign was most intense, almost 32,000 precincts, and many measures of demographic change, there is little evidence that precinct-level demographic changes are associated with vote swings toward the anti-immigration candidate.

There are at least 3 explanations for this unexpected finding. First, it may be that the electoral benefits for pro-immigration candidates in places with demographic changes are larger than the electoral benefits to anti-immigration candidates. In this account, places that become more Hispanic become more Democratic because the more conservative voting behavior of long-time residents is outweighed by new or existing voters. The evidence above that influxes of noncitizen foreign-born residents are also associated with pro-Democratic shifts suggests that compositional changes in the electorate are unlikely to explain this result, as such immigrants are ineligible to vote. Second, in SI Appendix, section H, we show that even in the most Republican precincts, the top decile where mean 2012 Republican vote share was 75%, increases in the Hispanic population correspond to benefits for the pro-immigration candidate in 2016.

Another possibility is that threatened reactions to demographic changes may diminish over time. In this view, exposure to Hispanics or noncitizen immigrants may lead to some initial animosity, but such negative reactions are short-lived (17, 23). More generally, while there is little doubt that certain groups of native Whites found Trump’s anti-immigrant rhetoric appealing (7), this is different from claiming that such appeals were more persuasive for the average voter in places undergoing demographic change or that Trump’s victory depended on them. While Trump’s rhetoric may have activated some supporters, we cannot thus conclude that he gained more votes than he lost.

A 3rd possibility generating our results is omitted variable bias. It could be that changing demographics do engender threat, but also that the process that drives Hispanics to certain places may be correlated with factors that predict vote choice such that the effects of threat are overwhelmed by those of selection bias. Certainly, our statistical models attempt to control for these factors in various ways, and there is no consistent pattern indicating that more fully saturated models show more threatened responses. Nonetheless, both our estimates and those in prior research on election outcomes rely on the assumption that all else is conditionally equal across precincts.^¶

Finally, the stability of party cleavages and the US 2-party system may limit the capacity of local changes to influence voting behavior. It is possible that local demographic changes influence Americans’ immigration attitudes without materially influencing general election vote choice.

These results do not rule out a link between demographic change and support for populist, anti-immigration candidates like Trump. If the precinct is the appropriate level at which to measure hyperlocal mechanisms of threat, other mechanisms may operate over broader geographic units, such as the labor market, media market, or even the nation as a whole. It is quite possible, for instance, that immigration is a nationalized political issue. However, if the effect of immigration and demographic change operate principally through perceptions about nation-level changes, existing theories of local demographic threats would require revision. Citizens’ perceptions of the national context is a notably different theoretical mechanism than lived local experience.