The Impact of the 2016 Election on Mental Distress Across U.S. Counties
Schmidt, Joshua
Introduction
With the recent reelection of former president Donald Trump, the United States has once again entered a politically charged era marked by sharp divisions in public opinion and national discourse. Trump’s political style – confrontational and often inflammatory – has consistently sparked strong reactions across the ideological spectrum, with some feeling emboldened by his rhetoric and others threatened. During his first term as president, President Trump took a hardline stance on many issues, including immigration, such as but not limited to targeting those with Latin-American ethnic backgrounds. Such rhetoric has contributed to a broader climate of uncertainty and stress among Latin-American and other ethnically marginalized immigrant communities.
The political climate during and following the 2016 election had entered an inflection point from which there was no turning back. People from both sides had very different interpretations of what his election meant for the country. Some viewed him as a vehicle for exclusionary and nationalistic ideals, while others saw him as the personification of long-standing grievances across both parties. Regardless, he was an outsider, approaching politics with plenty of flare and populist rhetoric to embolden his supporters and infuriate his dissenters.
Amidst this turbulence, mental health became a salient issue. The early years of Trump’s presidency were accompanied by widespread protest, with people vocalizing their dissatisfaction with his stances on contentious issues and proposed policies. Many of these protests vocalized dissatisfaction regarding his talking points about immigrant populations within the United States, many people finding them offensive and racist. These talking points fueled growing concern and fear among immigrant communities, undoubtedly taking a negative psychological toll on their well-being. Such anxiety was visible in nearly all mediums: from people sharing their stories over social media on how Trump’s presidency has impacted their mental health to national surveys showing how rates of mental distress increased across the board.
This project seeks to analyze the rates of mental distress – which have been surveyed on a national level – but also to contextualize them politically by county. By introducing such granularity, we can better observe patterns across different demographic makeups, something that studies that have conducted similar experiments have done at the state level. By focusing on the thousands of counties that make up the United States, we can better understand how county demographics and partisanship impact mental health.
Literature Review
Political Climate and Mental Health Shifts (2016-2018)
The 2016 election and early presidency of Donald Trump were marked by divisive rhetoric and policy shifts, particularly around Latin-American immigration. Studies suggest that these events contributed to measurable increases in psychological distress, with effects varying by political affiliation and ethnic group.
Yan et al. (2021) found that adults in Clinton-voting states experienced a significant post election increase in poor mental health days, while those in Trump-voting states did not. The increase translated to over 50 million more poor mental health days nationally in December 2016 alone. Similarly, Morey et al. (2021) used a difference-in-difference (DiD) approach to show that White adults in Clinton states reported worsening mental health after the election, consistent with the idea of symbolic disempowerment.
Mental Health in Hispanic and Immigrant Communities
The Trump administration’s hardline stance on immigration had disproportionate effects on Latino communities. Krupenkin et al. (2019) analyzed Google search trends and found a sustained spike in mental health-related searches among Spanish-speaking Latinos following the 2016 election, indicating heightened anxiety. Johnson et al. (2024), using survey data from 2011-2018, show that Latino non-citizens and even U.S.-born Latinos experienced rising psychological distress after Trump took office, likely due to increased deportation threats.
Morey et al. (2021) also observed differences by language: English-speaking Latinx in Trump supporting states reported higher distress, while Spanish-speaking respondents showed mixed or reduced self-reported distress, possibly due to underreporting or coping strategies. Fox (2022) found that Latina pregnant women in California reported higher anxiety linked to concerns about Trump’s rhetoric on immigration and gender.
Data Description
The outcome analyzed in this study is the county-level frequency of mental distress, measured annually from 2014 to 2018. This variable is defined as “14 or more mentally unhealthy days during the past 30 days where the person has experienced poor mental health because of stress, depression, or problems with emotions” (https://www.cdc.gov/places/measure-definitions/health-status.html). This variable (mental_distress_pct) is from the AHRQ Social Determinants of Health (SDOH) Database, which compiles health-related indicators at the county level. The SDOH data incorporates the Behavioral Risk Factor Surveillance System (BRFSS)-based estimates provided by the County Health Rankings for each year (hrq.gov/sites/default/files/wysiwyg/sdoh/SDOH-Data-Sources-Documentation-v1-Final.pdf). This measure of frequent mental distress serves as an indicator of whether an individual is in poor mental health, which will be used to analyze whether a particular person is suffering, especially from stress, depression, or other problems with emotions. To ensure changes are captured regarding the 2016 presidential election, the range of dataset used will be from 2014-2018, as well as the years when this variable was recorded (changed later to multiple other vars, not included in this study). In the regression framework, this variable is the dependent variable, which allows for quantification of mental health outcomes at the county level and how they evolved pre and post-2016 election.
The choice of frequent mental distress as the outcome is motivated by its relevance as a public mental health indicator and its sensitivity across demographics and socio-political stressors. Because it is county-level data, we can observe a granularity that would be impossible otherwise. It provides reasonably reliable observations across counties for the five years the variable was recorded. By focusing on this metric, we can observe changes in severe or persistent mental health issues rather than transient mood fluctuations. Additionally, extreme variation and overall trends can be observed with this dataset. In the context of causal inference, treating mental distress as the outcome allows us to examine whether the 2016 election had any measurable impact on the population’s mental health after accounting for other confounding factors.
The county-level voting variables from the 2016 election come from the Harvard Dataverse 2016 County Election Returns (https://dataverse.harvard.edu/file.xhtml?fileId=5028533&version=1.1). Including such variables is critical to our analysis, as it helps us analyze the differences in partisanship in the United States for our sample (2014-2018). The variables in our model from these data include rep_share, representing the ratio of Republican votes to total votes. This variable is used in the RDD analysis to construct the threshold. The other variable we use for our analysis is Party, which is a dummy variable where 0 = Democrat and 1= Republican.
We control for county demographic composition using data from the U.S. Census County Population Totals (2010-2019), focusing on the Hispanic population share. From the census population estimates – which provide annual county resident populations by race and Hispanic origin (https://www2.census.gov/programs-surveys/popest/technical-documentation/file-layouts/2010-2019/cc-est2019-alldata.pdf) – we calculate (percent_hispanic) as the percentage of each county’s population identified as Hispanic (of any race) for each year. This continuous covariate measures ethnic composition differences across counties, allowing us to establish a baseline mental health measure for the group and the observation of its differences across counties and partisanship. When including this as a covariate in our regression, we can measure how the demographics of a county impact its mental health with consideration to the inflammatory rhetoric towards immigrants.
In addition, we include a dummy variable (hisp_20) equal to 1 for counties where Hispanics comprise over 20% of the population (and 0 otherwise). This threshold lags counties with a substantially large Hispanic population. We include (hisp_20) to allow for effect heterogeneity in the analysis – specifically, to test whether high-Hispanic counties saw different changes in mental distress after the election relative to other counties. The choice of a 20% cutoff balances identifying communities with significant Hispanic representation while retaining enough counties for statistical power. In the regression models, this variable can be interacted with post-2016 indicators to measure differential impacts. Including the continuous share and the high share dummy also helps flexibility control for non-linear demographic effects. Changes across counties are recorded for our years of interest (2014-2018).
Local economic conditions are another crucial determinant of community mental health and a potential confounder in our study. We include the annual county unemployment rate for 2014-2018 as a control variable, sourced from the USDA Economic Research Service’s county-level data compilation (https://catalog.data.gov/dataset/county-level-data-sets). The unemployment rate (unemployment_rate) is measured as the average percentage of the unemployed labor force in each county year. We obtain these data from the ERS County-Level Data Sets, which aggregate Labor Department (BLS Local Area Unemployment Statistics) figures for each county and year.
Controlling for unemployment serves two purposes. First, it accounts for economic stress: areas with worsening job markets might see rises in mental distress, confounding the effect of the election if economic changes coincide with that period. Second, it captures general socioeconomic status differences across counties (e.g., persistent poverty or industrial decline) that could bias our estimates if left unaccounted for. By including unemployment rates, we adjust for time-varying economic shocks at the county level. For example, if a particular county experienced a post-2016 economic upswing or downtown, the models will attribute changes in mental distress partly to that rather than incorrectly to the election. In a casual inference framework, unemployment is a covariate to control for confounding: it is plausibly related to both the treatment (the election’s impacts may vary with the local economy) and the outcome (mental health). As such, including this variable improves our specification by isolating the effect of interest from broad economic trends.
We further incorporate two variables from the NCHS Drug Poisoning Mortality data (2014-2018) to proxy aspects of the community health environment and geography. The first is the county drug overdose model based death rate (Model.Based.Death.Rate), expressed as annual deaths per 100,000 population from drug poisonings (overdoses). This measure coms from the National Center for Health Statistics, which produces smoothed county-level overdose mortality estimates using hierarchical Bayesian model that borrows strength across neighboring counties and years (https://www.cdc.gov/nchs/data-visualization/drug-poisoning-mortality/index.htm). We use the NCHS model-based rate (as opposed to raw death rates) because it is more stable for small counties and reduces random noise in mortality outcomes. In our context, this variable serves as a proxy for extreme negative mental health outcomes and social distress (e.g., the severity of the opioid epidemic in a country). High overdose mortality could directly contribute to community mental distress (through trauma or loss) and also correlate with sociopolitical factors. Thus, controlling for it helps ensure that our results are not confounded by underlying health crises that might have evolved during the study period. For instance, if some counties were experiencing surging opioid fatalities in 2017-2018, that could raise mental distress independently of any election effect; including the death rate accounts for this dynamic. It also captures unobserved risk factors and public health conditions at the county level, improving our model’s completeness.
The second variable from this source is the county’s urban-rural classification (Urban.Rural.Category). Each county is categorized into one of the six NCHS designations: Large Central Metro, Large Fringe Metro, Medium Metro, Small Metro, Micropolitan, or Noncore. These categories are based on the 2013 NCHS Urban-Rural Classification Scheme for Counties (https://www.cdc.gov/nchs/data-visualization/drug-poisoning-mortality/index.htm) and remain fixed over our study period. We include a factor variable for these categories as controls in the regression (with Large Central Metro as the reference group). The rationale is to adjust for geographic and urbanicity differences that could influence mental health and its response to shocks. Urban versus rural areas differ systematically in factors like healthcare access, social cohesion, economic opportunities, and political context – all of which might affect baseline mental distress levels and trends. By accounting for the urban-rural category, we effectively compare counties within the same urbanicity class when estimating the election’s impact, mitigating bias. Moreover, this classification can absorb some broad regional variation (since urban cores and remote areas often cluster spatially). Including it as a control recognizes that mental distress after 2016 might evolve differently in urban and rural America; the categorical fixed effects ensure that those differences are not mistakenly attributed to our main variables. Overall, the NCHS variables improve our dataset by capturing a critical public health outcome (drug mortality) and the fundamental urban-rural context, both of which are treated as confounding factors or controls for heterogenous baseline conditions.
Methodology and Results
Overview
To investigate whether the 2016 U.S. presidential election of Donald Trump causally impacted county-level mental distress – particularly in Hispanic and Republican-leaning counties – we employ three econometric approaches. First, we implement a fuzzy regression discontinuity design (RDD) exploiting the 50% vote-share cutoff in the 2016 election as a quasi-experimental threshold. This RDD provides a local randomization around the tipping point of a county voting Republican, allowing us to test for any sharp discontinuity in mental health outcomes attributable to a narrow Trump victory versus loss. Second, we use a difference-in-difference (DiD) framework to compare changes in mental distress before vs. after 2016 between counties with different demographic and political characteristics. The DiD approach offers a temporal comparison that can isolate the election’s overall effect and heterogeneity by county type (e.g., high versus low Hispanic population share, Republican-leaning versus others) under the parallel trends assumption. Third, we estimate a two-way fixed effects (TWFE) panel mode on multi-year county data, including fixed effects for time and county, to control for time-invariant unobserved heterogeneity and common shocks. The TWFE specification uses the full panel (pre- and post-election years) to capture dynamic effects and persistent changes in distress following 2016. Using all three methods, we strengthen causal inference by addressing separate potential biases differently and using consistent findings across the three effects to bolster confidence in our results. Below, we describe each model’s setup and assumptions, followed by the empirical results from each approach.
RDD
RDD Setup:
We first implement a fuzzy regression discontinuity design to test for any abrupt change in mental health outcomes at the threshold of a county voting for Donald Trump in 2016. The running variable is Trump’s 2016 vote share in the county, with a cutoff at 50%. This design treats counties that just barely voted for Trump versus just barely against him as quasi-randomly “treated” with a Republican victory at the local level.
Figure 1:
Because the treatment (exposure to the local symbolic victory) is not perfectly deterministic at the threshold, we use a fuzzy RDD framework with the vote-share cutoff as an instrument for treatment. The first-stage is strong: crossing the 50% threshold leads to a discrete jump in the probability that the county’s winner is Republican (nearly 0 to 100% around the cutoff). However, the RDD reveals no evidence of any discontinuous change in mental distress at this threshold. The estimated jump in distress for counties just above vs. just below 50% is essentially zero and not statistically significant. Even after adding covariates (e.g. local unemployment, overdose death rates, urban/rural status) to improve precision, the treatment effect remains near zero with confidence intervals that include zero. In short, narrowly “flipping” from a Democratic to a Republican majority in 2016 did not produce any sharp local change in mental health outcomes. This null result suggests that any mental health impact of Trump’s election was not concentrated in a way that a threshold-based design could detect.
RDD Results:
Table 1:
## Fuzzy RD estimates using local polynomial regression.
##
## Number of Obs. 2997
## BW type mserd
## Kernel Triangular
## VCE method NN
##
## Number of Obs. 556 2441
## Eff. Number of Obs. 306 487
## Order est. (p) 1 1
## Order bias (q) 2 2
## BW est. (h) 0.101 0.101
## BW bias (b) 0.155 0.155
## rho (h/b) 0.649 0.649
## Unique Obs. 555 2440
##
## First-stage estimates.
##
## =============================================================================
## Method Coef. Std. Err. z P>|z| [ 95% C.I. ]
## =============================================================================
## Conventional 0.049 0.030 1.634 0.102 [-0.010 , 0.107]
## Robust - - 5.261 0.000 [0.102 , 0.223]
## =============================================================================
##
## Treatment effect estimates.
##
## =============================================================================
## Method Coef. Std. Err. z P>|z| [ 95% C.I. ]
## =============================================================================
## Conventional 1.181 5.799 0.204 0.839 [-10.185 , 12.546]
## Robust - - -0.194 0.846 [-15.095 , 12.371]
## =============================================================================The RDD analysis reveals no evidence of a discontinuous change in mental distress at the 50% Republican vote threshold. The estimated treatment effect (i.e., the jump in distress when moving from just below to just above 50% Trump vote) is substantively small and statistically indistinguishable from zero. Without covariates, the point estimate implies a very slight increase in distress (around +1.2 percentage points for counties that just voted Trump, but the effect is not significant (p-value = 0.84) and the 95% confidence interval is wide – spanning roughly -10 to +12 percentage points. This wide interval includes zero and even large negative/positive values, indicating a high degree of uncertainty and no clear effect.
Table 2:
## Covariate-adjusted Fuzzy RD estimates using local polynomial regression.
##
## Number of Obs. 2997
## BW type mserd
## Kernel Triangular
## VCE method NN
##
## Number of Obs. 2557 440
## Eff. Number of Obs. 439 245
## Order est. (p) 1 1
## Order bias (q) 2 2
## BW est. (h) 0.205 0.205
## BW bias (b) 0.315 0.315
## rho (h/b) 0.652 0.652
## Unique Obs. 2555 439
##
## First-stage estimates.
##
## =============================================================================
## Method Coef. Std. Err. z P>|z| [ 95% C.I. ]
## =============================================================================
## Conventional 1.000 0.0007135647393840547.000 0.000 [1.000 , 1.000]
## Robust - -5960423591233177.000 0.000 [1.000 , 1.000]
## =============================================================================
##
## Treatment effect estimates.
##
## =============================================================================
## Method Coef. Std. Err. z P>|z| [ 95% C.I. ]
## =============================================================================
## Conventional -0.178 0.268 -0.666 0.505 [-0.703 , 0.346]
## Robust - - -0.468 0.640 [-0.772 , 0.475]
## =============================================================================Upon adding covariates to the RDD, the estimated discontinuity actually reverses sign (becoming a slight negative of about -0.2 percentage points) and remains statistically insignificant (p-value = 0.5). The confidence interval narrows considerably with covariate adjustment (spanning -0.7 to +0.35 percentage points) but still comfortably covers zero. In sum, we find no statistically significant discontinuity in post-election mental distress at the 50% vote-share cutoff, even after adjusting for potential confounders. These results suggest that narrowly voting for Trump versus narrowly voting for Clinton had no abrupt causal impact on a county’s mental distress level.
Figure 2:
The absence of any sharp local jump implies that any effect of the 2016 election on distress was likely more homogeneous rather than tied to the binary outcome of the county’s vote. Given the null RDD finding, we interpret that there is no measurable instant “shock” to mental health from tipping into a Trump-voting status. This lack of a localized effect motivates us to turn to alternative methods (DiD and TWFE) that can detect more gradual or widespread changes in distress over time, beyond the narrow margin of victory context.
DiD
Basic DiD Model:
We next implement a difference-in-difference design to estimate the overall post-2016 change in mental distress and how it varies with county characteristics. The basic DiD setup compares mental health outcoems in the post-election period (post-2016) to the pre-election baseline, across different counties. All counties are exposed to the national political shock in 2016, but we leverage cross-county differences (e.g., in ethnic composition) as a way to identify heterogenous effects. Formally, we estimate an OLS panel regression of the form:
\[ Y_{st} = \beta_0 + \beta_1 \cdot \text{Post}_t + \beta_2 \cdot \text{Percent_Hispanic}_{st} + \beta_3 \cdot X_{it} + \varepsilon \]
where Y_it is the mental distress rate in county _i and year _t. beta_1 Post_t is an indicator for years after 2016, capturing the overall “treatment” of the Trump era. beta_2 HispanicShare_it is a measure of the county’s Hispanic population, included to account for baseline differences in distress levels associated with demographics. beta_3 X_it represents other control variables such as economic conditions (unemployment rate) or urban/rural status, to net out confounding trends. The key identifying assumption for DiD is that, absent the 2016 election shock, mental distress trends in different types of counties would have moved in parallel. We partially address this by inspecting pre-2016 trends and including controls; however, with only a couple of pre-election years available (our sample covers 2014-2018), the parallel trends assumption is acknowledged as a maintained assumption that cannot be definitely proven. Despite this limitation, the DiD provides a straightforward estimate of the average post-2016 effect, while controlling for any static differences across counties.
Basic DiD Results:
Table 3:
| Mental Distress Pct | |
|---|---|
| + p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001 | |
| (Intercept) | 7.720*** |
| (<0.001) | |
| PartyRepublican | -0.381*** |
| (<0.001) | |
| post_2016TRUE | 1.823*** |
| (<0.001) | |
| percent_hispanic | -0.010*** |
| (<0.001) | |
| unemployment_rate | 0.403*** |
| (<0.001) | |
| Model.based.Death.Rate | 0.076*** |
| (<0.001) | |
| Urban.Rural.CategoryLarge Fringe Metro | -0.254+ |
| (0.061) | |
| Urban.Rural.CategoryMedium Metro | 0.733*** |
| (<0.001) | |
| Urban.Rural.CategoryMicropolitan | 1.329*** |
| (<0.001) | |
| Urban.Rural.CategoryNoncore | 1.324*** |
| (<0.001) | |
| Urban.Rural.CategorySmall Metro | 0.904*** |
| (<0.001) | |
| Num.Obs. | 15005 |
| R2 | 0.407 |
| R2 Adj. | 0.407 |
| AIC | 61576.9 |
| BIC | 61668.3 |
| Log.Lik. | -30776.472 |
| RMSE | 1.88 |
The baseline DiD model indicates a significant increase in mental distress in the post-2016 period. Specifically, the coefficient on the Post-2016 indicator (β₁) is about +1.82 percentage points, meaning the average county’s mental distress rate rose by roughly 1.8 percentage points after Trump’s election (relative to the pre-2017 level, holding other factors constant). This effect is statistically significant (p < 0.01) and suggests a noteworthy deterioration in mental health coincident with the advent of the Trump administration. For context, an increase of 1.82 pp is substantial given that baseline county distress rates tend to be in the teens or low twenties as a percentage – it represents a broad upward shift in psychological distress across the country. We also find that urbanicity and economic controls have important effects: for example, rural counties (micropolitan or noncore areas) exhibit significantly higher distress on average than metropolitan counties and higher local unemployment or opioid overdose rates are associated with elevated distress. After adjusting for these factors, post-2016 signals a discernible uptick of nearly two percentage points, suggesting possible increased mental distress universally.
Extended DiD with Interaction Terms:
To further probe heterogeneity in the election’s impact, we augment the model with interaction terms that create a 2 by 2 difference-in-differences structure along two dimensions: (1) high vs. low Hispanic share counties, and (2) Republican-leaning vs. other counties. We then estimate a model including the following interactions: beta_1 Post_t * Hisp_20_st and beta_2 Republican_t * Hisp_20_st . This specification allows the post-2016 trend to differ for high-Hispanic counties, and also lets Republican-leaning, high-Hispanic areas have a distinct level shift. In essence, we are comparing four groups of counties – (i) low-Hispanic, non-Republican; (ii) low-Hispanic, Republican; (iii) high-Hispanic, non-Republican; and (iv) high-Hispanic, Republican – and asking whether their mental distress evolved differently after 2016. Going forward, we are using the same controls as the basic models. The model specification is as follows:
\[ Y_{st} = \beta_0 + \beta_1 (\text{Post}_t \times \text{Hisp_20}_{st}) + \beta_2 (\text{Republican}_t \times \text{Hisp_20}_{st}) + \beta_3 X_{it} + \varepsilon \]
Interaction Model Results:
Table 4:
| Mental Distress Pct | Mental Distress Pct (with Hispanic interactions) | |
|---|---|---|
| + p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001 | ||
| (Intercept) | 7.720*** | 7.797*** |
| (<0.001) | (<0.001) | |
| PartyRepublican | -0.381*** | -0.435*** |
| (<0.001) | (<0.001) | |
| post_2016TRUE | 1.823*** | 1.835*** |
| (<0.001) | (<0.001) | |
| percent_hispanic | -0.010*** | -0.012*** |
| (<0.001) | (<0.001) | |
| unemployment_rate | 0.403*** | 0.405*** |
| (<0.001) | (<0.001) | |
| Model.based.Death.Rate | 0.076*** | 0.076*** |
| (<0.001) | (<0.001) | |
| Urban.Rural.CategoryLarge Fringe Metro | -0.254+ | -0.295* |
| (0.061) | (0.032) | |
| Urban.Rural.CategoryMedium Metro | 0.733*** | 0.692*** |
| (<0.001) | (<0.001) | |
| Urban.Rural.CategoryMicropolitan | 1.329*** | 1.278*** |
| (<0.001) | (<0.001) | |
| Urban.Rural.CategoryNoncore | 1.324*** | 1.277*** |
| (<0.001) | (<0.001) | |
| Urban.Rural.CategorySmall Metro | 0.904*** | 0.854*** |
| (<0.001) | (<0.001) | |
| hisp_20TRUE | -0.082 | |
| (0.578) | ||
| post_2016TRUE × hisp_20TRUE | -0.109 | |
| (0.264) | ||
| PartyRepublican × hisp_20TRUE | 0.315** | |
| (0.006) | ||
| Num.Obs. | 15005 | 15005 |
| R2 | 0.407 | 0.408 |
| R2 Adj. | 0.407 | 0.407 |
| AIC | 61576.9 | 61572.1 |
| BIC | 61668.3 | 61686.3 |
| Log.Lik. | -30776.472 | -30771.033 |
| RMSE | 1.88 | 1.88 |
The enriched 2×2 DiD model confirms a significant and widespread post-2016 increase in mental distress across all county groups, consistent with the baseline model. Before the election, Republican-leaning counties exhibited slightly lower distress compared to Democratic-leaning counties (approximately –0.3 to –0.4 percentage points). However, this advantage reverses in counties with substantial Hispanic populations. Specifically, the interaction between Republican-leaning counties and high Hispanic populations (≥20%) is positive, sizable, and statistically significant at the 1% level, indicating notably higher distress in these areas compared to either predominantly white Republican counties or high-Hispanic Democratic counties. This finding highlights compounded vulnerability where large Hispanic communities exist in politically conservative environments. Conversely, the interaction between post-2016 and high-Hispanic share alone is not statistically significant, suggesting no unique nationwide distress spike confined solely to Hispanic-heavy areas.
TWFE
TWFE Model Setup:
We implement a two-way fixed effects (TWFE) panel model, utilizing annual county-level data from 2014 through 2018. This model includes county-fixed effects, controlling for all stable, unobserved county-level characteristics that could influence mental health, such as historical, cultural, or structural factors. Additionally, year-fixed effects capture nationwide shocks or common events affecting all counties simultaneously, such as economic downturns or policy changes. The TWFE approach extends the DiD framework by using multiple pre- and post-election periods, allowing for observing dynamic changes over time and better control of unobserved heterogeneity. The treatment event—Donald Trump’s election—occurs simultaneously across all counties at the end of 2016. Thus, our model compares each county’s mental distress after 2016 to its pre-election baseline, net of any standard year-to-year shifts affecting all areas. Controls for concurrent socioeconomic conditions, including unemployment rates and overdose mortality rates, ensure that these local factors do not confound observed changes in mental distress. The primary assumption of parallel trends holds that, without Trump’s election, counties’ distress trends would remain consistent over time once we account for fixed differences and covariates. This setup maximizes data utilization and provides robust estimates of the election’s causal effects. The model specification is as follows:
\[ Y_{st} = \beta_0 + \beta_1 (\text{Post}_t \times \text{Hisp_20}_{st}) + \beta_2 (\text{Republican}_t \times \text{Hisp_20}_{st}) + \beta_3 X_{it} + \delta_t + \delta_s + \varepsilon \]
TWFE Results:
Table 5:
| Mental Distress Pct | Mental Distress Pct (with Hispanic interactions) | TWFE Model | |
|---|---|---|---|
| + p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001 | |||
| (Intercept) | 7.720*** | 7.797*** | 7.200*** |
| (<0.001) | (<0.001) | (<0.001) | |
| PartyRepublican | -0.381*** | -0.435*** | -0.314 |
| (<0.001) | (<0.001) | (0.116) | |
| post_2016TRUE | 1.823*** | 1.835*** | |
| (<0.001) | (<0.001) | ||
| percent_hispanic | -0.010*** | -0.012*** | -0.015 |
| (<0.001) | (<0.001) | (0.170) | |
| unemployment_rate | 0.403*** | 0.405*** | 0.561*** |
| (<0.001) | (<0.001) | (<0.001) | |
| Model.based.Death.Rate | 0.076*** | 0.076*** | 0.065*** |
| (<0.001) | (<0.001) | (<0.001) | |
| Urban.Rural.CategoryLarge Fringe Metro | -0.254+ | -0.295* | -0.376* |
| (0.061) | (0.032) | (0.021) | |
| Urban.Rural.CategoryMedium Metro | 0.733*** | 0.692*** | 0.564** |
| (<0.001) | (<0.001) | (0.005) | |
| Urban.Rural.CategoryMicropolitan | 1.329*** | 1.278*** | 1.067*** |
| (<0.001) | (<0.001) | (<0.001) | |
| Urban.Rural.CategoryNoncore | 1.324*** | 1.277*** | 1.032*** |
| (<0.001) | (<0.001) | (<0.001) | |
| Urban.Rural.CategorySmall Metro | 0.904*** | 0.854*** | 0.691** |
| (<0.001) | (<0.001) | (0.001) | |
| hisp_20TRUE | -0.082 | -0.021 | |
| (0.578) | (0.950) | ||
| post_2016TRUE × hisp_20TRUE | -0.109 | -0.161 | |
| (0.264) | (0.361) | ||
| PartyRepublican × hisp_20TRUE | 0.315** | 0.349 | |
| (0.006) | (0.251) | ||
| year = 2014 | 0.180 | ||
| (0.196) | |||
| year = 2016 | 0.530*** | ||
| (<0.001) | |||
| year = 2017 | 1.584*** | ||
| (<0.001) | |||
| year = 2018 | 4.032*** | ||
| (<0.001) | |||
| Num.Obs. | 15005 | 15005 | 15005 |
| R2 | 0.407 | 0.408 | 0.614 |
| R2 Adj. | 0.407 | 0.407 | 0.613 |
| AIC | 61576.9 | 61572.1 | 55170.1 |
| BIC | 61668.3 | 61686.3 | 55299.5 |
| Log.Lik. | -30776.472 | -30771.033 | |
| RMSE | 1.88 | 1.88 | 1.52 |
| Std.Errors | by: state_abbrev | ||
The TWFE model confirms a significant rise in mental distress following the 2016 election, peaking in 2018 (the end of our sample), which suggests a cumulative impact of political stressors over Trump’s early presidency. Republican-leaning counties, after controlling for fixed effects and socioeconomic trends, experienced higher distress compared to pre-election levels, indicating the partisan distress gap narrowed or reversed post-2016. This finding contrasts with simpler models that initially indicated slightly lower baseline distress in Republican counties. It suggests that, once accounting for unobserved heterogeneity such as rural isolation or local economic decline, Republican counties faced heightened vulnerability post-election. Importantly, we find no distinct post-2016 trend in counties with large Hispanic populations, indicating no general disproportionate distress in these communities. Instead, heightened distress among Hispanic communities emerged specifically within Republican-leaning contexts. Finally, consistent with expectations, local socioeconomic conditions—including unemployment and opioid overdose rates—remain strongly associated with distress levels, reinforcing confidence that observed mental health declines post-2016 reflect genuine political impacts.
Conclusion
The evidence indicates that Donald Trump’s 2016 election corresponded with increased county-level mental distress, but with important nuances. The regression discontinuity design (RDD) found no significant jump in distress at the electoral margin, suggesting no abrupt localized shock. In contrast, the difference-in-differences (DiD) and two-way fixed effects (TWFE) models detected a significant overall rise in mental distress after 2016, implying that any election-related impact unfolded broadly over time rather than as a discrete event.
This increase in distress was not uniform nationwide; it varied by counties’ political and demographic profiles. Our results showed that counties with large Hispanic populations that voted Republican experienced especially elevated increases in poor mental health days. This highlights the intersection between Trump’s rhetoric and demographic vulnerability. Hispanic communities in pro-Trump counties, who perhaps may have been exposed to anti-immigrant rhetoric - appear to have higher rates of poor mental health than their contemporaries.
These findings align with prior research while adding new granularity. Yan et al. (2021) documented a post-2016 surge in poor mental health days concentrated in Clinton-voting states, reflecting distress among those on the losing side. Morey et al. (2021) describe this as “symbolic disempowerment,” noting acute mental health declines among Latinx individuals in Republican states and among whites in Democrat states. Our county-level analysis extends these state-level patterns by revealing that even within Republican-voting regions, minority communities (notably high-Hispanic counties) saw larger increases in mental distress. This observation is consistent with evidence that Trump-era policies and rhetoric disproportionately harmed Latin-American mental well-being. Krupenkin et al. (2019) found significant spikes in online searches for depression and enxiety among Spanish-speaking Latinos after the election. Likewise, Johnson et al. (2024) report that Trump’s presidency heightened anxiety and depressive symptoms in Latino populations, and Fox (2022) documents adverse effects of the administration’s anti-immigrant stance on Latina women’s prenatal mental health.
The 2016 election’s mental health impact was not monolithic; it was filtered through Americans’ political and demographic context. By using county-level data, this study offers a granular perspective, showing that local characteristics shaped the magnitude of distress following Trump’s victory. These insights deepen our understanding of how major political events can permeate public health, underscoring that mental health consequences are unevenly distributed. Our findings highlight the value of examining subnational contexts to fully capture political influences on population mental health while illuminating the particular vulnerability of communities facing symbolic disempowerment or targeted rhetoric.
Appendix
Figure 1:
Figure 2:
Figure 3:
Figure 4:
Figure 5:
Figure 6:
Table 1:
## Fuzzy RD estimates using local polynomial regression.
##
## Number of Obs. 2997
## BW type mserd
## Kernel Triangular
## VCE method NN
##
## Number of Obs. 556 2441
## Eff. Number of Obs. 306 487
## Order est. (p) 1 1
## Order bias (q) 2 2
## BW est. (h) 0.101 0.101
## BW bias (b) 0.155 0.155
## rho (h/b) 0.649 0.649
## Unique Obs. 555 2440
##
## First-stage estimates.
##
## =============================================================================
## Method Coef. Std. Err. z P>|z| [ 95% C.I. ]
## =============================================================================
## Conventional 0.049 0.030 1.634 0.102 [-0.010 , 0.107]
## Robust - - 5.261 0.000 [0.102 , 0.223]
## =============================================================================
##
## Treatment effect estimates.
##
## =============================================================================
## Method Coef. Std. Err. z P>|z| [ 95% C.I. ]
## =============================================================================
## Conventional 1.181 5.799 0.204 0.839 [-10.185 , 12.546]
## Robust - - -0.194 0.846 [-15.095 , 12.371]
## =============================================================================Table 2:
## Covariate-adjusted Fuzzy RD estimates using local polynomial regression.
##
## Number of Obs. 2997
## BW type mserd
## Kernel Triangular
## VCE method NN
##
## Number of Obs. 2557 440
## Eff. Number of Obs. 439 245
## Order est. (p) 1 1
## Order bias (q) 2 2
## BW est. (h) 0.205 0.205
## BW bias (b) 0.315 0.315
## rho (h/b) 0.652 0.652
## Unique Obs. 2555 439
##
## First-stage estimates.
##
## =============================================================================
## Method Coef. Std. Err. z P>|z| [ 95% C.I. ]
## =============================================================================
## Conventional 1.000 0.0007135647393840547.000 0.000 [1.000 , 1.000]
## Robust - -5960423591233177.000 0.000 [1.000 , 1.000]
## =============================================================================
##
## Treatment effect estimates.
##
## =============================================================================
## Method Coef. Std. Err. z P>|z| [ 95% C.I. ]
## =============================================================================
## Conventional -0.178 0.268 -0.666 0.505 [-0.703 , 0.346]
## Robust - - -0.468 0.640 [-0.772 , 0.475]
## =============================================================================Table 3:
| Mental Distress Pct | |
|---|---|
| + p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001 | |
| (Intercept) | 7.720*** |
| (<0.001) | |
| PartyRepublican | -0.381*** |
| (<0.001) | |
| post_2016TRUE | 1.823*** |
| (<0.001) | |
| percent_hispanic | -0.010*** |
| (<0.001) | |
| unemployment_rate | 0.403*** |
| (<0.001) | |
| Model.based.Death.Rate | 0.076*** |
| (<0.001) | |
| Urban.Rural.CategoryLarge Fringe Metro | -0.254+ |
| (0.061) | |
| Urban.Rural.CategoryMedium Metro | 0.733*** |
| (<0.001) | |
| Urban.Rural.CategoryMicropolitan | 1.329*** |
| (<0.001) | |
| Urban.Rural.CategoryNoncore | 1.324*** |
| (<0.001) | |
| Urban.Rural.CategorySmall Metro | 0.904*** |
| (<0.001) | |
| Num.Obs. | 15005 |
| R2 | 0.407 |
| R2 Adj. | 0.407 |
| AIC | 61576.9 |
| BIC | 61668.3 |
| Log.Lik. | -30776.472 |
| RMSE | 1.88 |
Table 4:
| Mental Distress Pct | Mental Distress Pct (with Hispanic interactions) | |
|---|---|---|
| + p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001 | ||
| (Intercept) | 7.720*** | 7.797*** |
| (<0.001) | (<0.001) | |
| PartyRepublican | -0.381*** | -0.435*** |
| (<0.001) | (<0.001) | |
| post_2016TRUE | 1.823*** | 1.835*** |
| (<0.001) | (<0.001) | |
| percent_hispanic | -0.010*** | -0.012*** |
| (<0.001) | (<0.001) | |
| unemployment_rate | 0.403*** | 0.405*** |
| (<0.001) | (<0.001) | |
| Model.based.Death.Rate | 0.076*** | 0.076*** |
| (<0.001) | (<0.001) | |
| Urban.Rural.CategoryLarge Fringe Metro | -0.254+ | -0.295* |
| (0.061) | (0.032) | |
| Urban.Rural.CategoryMedium Metro | 0.733*** | 0.692*** |
| (<0.001) | (<0.001) | |
| Urban.Rural.CategoryMicropolitan | 1.329*** | 1.278*** |
| (<0.001) | (<0.001) | |
| Urban.Rural.CategoryNoncore | 1.324*** | 1.277*** |
| (<0.001) | (<0.001) | |
| Urban.Rural.CategorySmall Metro | 0.904*** | 0.854*** |
| (<0.001) | (<0.001) | |
| hisp_20TRUE | -0.082 | |
| (0.578) | ||
| post_2016TRUE × hisp_20TRUE | -0.109 | |
| (0.264) | ||
| PartyRepublican × hisp_20TRUE | 0.315** | |
| (0.006) | ||
| Num.Obs. | 15005 | 15005 |
| R2 | 0.407 | 0.408 |
| R2 Adj. | 0.407 | 0.407 |
| AIC | 61576.9 | 61572.1 |
| BIC | 61668.3 | 61686.3 |
| Log.Lik. | -30776.472 | -30771.033 |
| RMSE | 1.88 | 1.88 |
Table 5:
| Mental Distress Pct | Mental Distress Pct (with Hispanic interactions) | TWFE Model | |
|---|---|---|---|
| + p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001 | |||
| (Intercept) | 7.720*** | 7.797*** | 7.200*** |
| (<0.001) | (<0.001) | (<0.001) | |
| PartyRepublican | -0.381*** | -0.435*** | -0.314 |
| (<0.001) | (<0.001) | (0.116) | |
| post_2016TRUE | 1.823*** | 1.835*** | |
| (<0.001) | (<0.001) | ||
| percent_hispanic | -0.010*** | -0.012*** | -0.015 |
| (<0.001) | (<0.001) | (0.170) | |
| unemployment_rate | 0.403*** | 0.405*** | 0.561*** |
| (<0.001) | (<0.001) | (<0.001) | |
| Model.based.Death.Rate | 0.076*** | 0.076*** | 0.065*** |
| (<0.001) | (<0.001) | (<0.001) | |
| Urban.Rural.CategoryLarge Fringe Metro | -0.254+ | -0.295* | -0.376* |
| (0.061) | (0.032) | (0.021) | |
| Urban.Rural.CategoryMedium Metro | 0.733*** | 0.692*** | 0.564** |
| (<0.001) | (<0.001) | (0.005) | |
| Urban.Rural.CategoryMicropolitan | 1.329*** | 1.278*** | 1.067*** |
| (<0.001) | (<0.001) | (<0.001) | |
| Urban.Rural.CategoryNoncore | 1.324*** | 1.277*** | 1.032*** |
| (<0.001) | (<0.001) | (<0.001) | |
| Urban.Rural.CategorySmall Metro | 0.904*** | 0.854*** | 0.691** |
| (<0.001) | (<0.001) | (0.001) | |
| hisp_20TRUE | -0.082 | -0.021 | |
| (0.578) | (0.950) | ||
| post_2016TRUE × hisp_20TRUE | -0.109 | -0.161 | |
| (0.264) | (0.361) | ||
| PartyRepublican × hisp_20TRUE | 0.315** | 0.349 | |
| (0.006) | (0.251) | ||
| year = 2014 | 0.180 | ||
| (0.196) | |||
| year = 2016 | 0.530*** | ||
| (<0.001) | |||
| year = 2017 | 1.584*** | ||
| (<0.001) | |||
| year = 2018 | 4.032*** | ||
| (<0.001) | |||
| Num.Obs. | 15005 | 15005 | 15005 |
| R2 | 0.407 | 0.408 | 0.614 |
| R2 Adj. | 0.407 | 0.407 | 0.613 |
| AIC | 61576.9 | 61572.1 | 55170.1 |
| BIC | 61668.3 | 61686.3 | 55299.5 |
| Log.Lik. | -30776.472 | -30771.033 | |
| RMSE | 1.88 | 1.88 | 1.52 |
| Std.Errors | by: state_abbrev | ||
Table 6:
Table 7:
References
Agency for Healthcare Research and Quality. n.d. “Social Determinants of Health Database (SDOH).” https://www.ahrq.gov/sdoh/data-analytics/sdoh-data.html.
Centers for Disease Control and Prevention. n.d. “Measure Definitions: Health Status.” https://www.cdc.gov/places/measure-definitions/health-status.html.
County Health Rankings & Roadmaps. n.d. “SDOH Data Sources Documentation.” Agency for Healthcare Research and Quality. https://www.ahrq.gov/sites/default/files/wysiwyg/sdoh/SDOH-Data-Sources-Documentation-v1-Final.pdf.
Fox, Megan. 2022. “Changes in Mental Health Symptoms and Sociocultural Factors across the COVID-19 Pandemic in Mothers of Mexican Descent.” ResearchGate. https://www.researchgate.net/publication/372887745_Changes_in_mental_health_symptoms_and_sociocultural_factors_across_the_COVID-19_pandemic_in_mothers_of_Mexican_descent.
Harvard Dataverse. 2018. “County Presidential Election Returns 2000–2016.” https://dataverse.harvard.edu/file.xhtml?fileId=5028533&version=1.1.
Johnson, Rebecca C., et al. 2024. “Latino Mental Health After Trump’s Election: National Trends in Anxiety and Depression.” Proceedings of the National Academy of Sciences 121 (12): e2306554121. https://www.pnas.org/doi/10.1073/pnas.2306554121.
Krupenkin, Masha, D.J. Hopkins, and Samara Klar. 2019. “Do Threatening Political Messages Increase Latinos’ Anxiety and Depression? Evidence from Online Search Behavior.” SAGE Open 9 (1): 1–13. https://journals.sagepub.com/doi/10.1177/2158244019830865.
Morey, Brittany N., et al. 2021. “Political Partisanship, Perceptions of Safety, and Mental Health Following the 2016 U.S. Presidential Election: A Difference-in-Difference Analysis.” Social Science & Medicine 289: 114402. https://pubmed.ncbi.nlm.nih.gov/34656819/.
National Center for Health Statistics. n.d. “Drug Poisoning Mortality by County: United States.” Centers for Disease Control and Prevention. https://www.cdc.gov/nchs/data-visualization/drug-poisoning-mortality/index.htm.
U.S. Census Bureau. 2019. “County Population Totals: 2010–2019.” https://www2.census.gov/programs-surveys/popest/technical-documentation/file-layouts/2010-2019/cc-est2019-alldata.pdf.
U.S. Department of Agriculture, Economic Research Service. 2019. “County-Level Data Sets.” https://catalog.data.gov/dataset/county-level-data-sets.
Yan, Bo, et al. 2021. “Rising Mental Distress in the United States: Analysis of the Behavioral Risk Factor Surveillance System, 2011–2017.” Preventing Chronic Disease 18: E81. https://pmc.ncbi.nlm.nih.gov/articles/PMC7602772/.