Police Killings

Incident-level data on police killings come from the Los Angeles Times Homicide Database, which tracks all known officer-involved killings in Los Angeles county and includes 292 incidents between the 2002 and 2010 general elections.Footnote ⁵ For each killing, the data include the name, age, and race of the deceased, as well as the exact address and date of the event. We supplement this with information on whether the incident was reported by local newspapers.Footnote ⁶ For roughly 85% of killings, we were also able to determine whether a weapon was recovered from the deceased. This information was hand-coded from Los Angeles County District Attorney reports as well as police reports and other sources.

Note that these contextual measures may provide an incomplete picture of the surrounding events. Often officers acted under faulty information. For example, in one incident, police killed a man who was reported to have a gun but who was actually holding a water hose nozzle. In other cases, killings were precipitated by seemingly innocuous encounters that quickly escalated—such as when a man lunged for an officer’s gun after he was stopped for riding a bicycle on the sidewalk. Nonetheless, weapon information has the benefit of being objectively verifiable and can be found in all available incident reports.

Panel A of Table 1 provides a summary of the police killings data. Fifty-two percent of deceased individuals were Hispanic, 29% were Black, 16% were white, and 3% were Asian.Footnote ⁷ Relative to their population shares, Black (Hispanic) individuals are roughly six (two) times more likely to be killed by police than whites. The vast majority of individuals (96%) were male and the average age was 30 years.

Table 1. Summary Statistics

Panel A provides the summary statistics of the police killings data. Sample means are reported for all police killings in column 1. Columns 2 and 3 report means separately for killings of Blacks and Hispanics, and killings of whites and Asians, respectively. We were unable to obtain race for one individual. Newspaper mentions come from a search of each incident by deceased name in six local newspapers. “Any” indicates mention in any article. “Total” refers to the mean number of articles mentioning the incident. “Median” is the median number of articles, conditional on at least one mention. “Unarmed” refers to suspects that did not have a weapon, “gun” refers to suspects with firearms (including BB guns and replicas), “knife/other” refers to suspects with any other type of weapon, and “unknown” refers to incidents where weapon type was not obtainable from District Attorney reports and other sources.

Panel B provides the summary statistics of the neighborhood characteristics and voter registration information. Column 1 reports sample means for all Census blocks in the analysis sample. Columns 2 and 3 limit the sample to Census blocks in Census block groups that experienced a police killing during the sample period. Column 2 examines the exact Census blocks where police killings occurred, whereas column 3 examines all other Census blocks in the same Census block groups. Registration data are from the 2002 general election and include total registration and vote counts as well as the share of registrations by race, age, party affiliation, and new registrant status (i.e., those who registered less than 4 years prior). Voter turnout is defined by $ \frac{\mathrm{Votes}}{\mathrm{Registrations}} $ and Reg per Pop 18+ is given by $ \frac{\mathrm{Registrations}}{\mathrm{Pop}\ 18+} $ . The reason there are 285 killing blocks versus 292 killings is because some incidents involved multiple deaths. In total, there were 287 distinct incidents, two of which occurred in the same block on separate dates.

Consistent with national statistics, 47% of those killed were armed with a firearm (including BB guns and replicas), whereas 22% possessed some other type of weapon. This includes items like knives and pipes as well as individuals who attempted to hit someone with a vehicle. Fifteen percent of individuals were completely unarmed. We were unable to find contextual information for the remaining 16% of incidents. Notably, the vast majority of killings received little or no media coverage. Only 18% of sample killings were ever mentioned in any of the six local newspapers. Conditional on coverage, the median number of articles is 2. The most mentions of any incident was 28, far below the level of media attention garnered by recent high profile police killings.

Examining contextual factors separately by race, Black and Hispanic individuals killed by police were younger on average than white and Asian individuals (29 vs. 36 years old, respectively) and more likely to possess a firearm (53% vs. 24%). However, rates of media coverage are similar between groups (19% vs. 16%). Regardless of circumstance, involved officers were never prosecuted. The District Attorney did not pursue criminal charges against police following any of the 292 sample killings. This is consistent with national statistics, which find that criminal charges are filed against police in fewer than half a percentage of all officer-involved shootings.

Voter Registration, Turnout, and Preferences

Police killings are geocoded to Census blocks and merged to voting information from the California Statewide Database. The database contains information on the number of individuals registered to vote and the number of ballots cast at the 2010 Census block-level for each general election from 2002 to 2010.Footnote ⁸ The advantage of these data relative to standard voter registration files is that they capture registration and voting at the date of each election, which allows us to more precisely measure effects.Footnote ⁹ This is particularly important when considering geographies as small as Census blocks, which average less than 0.06 square miles and fewer than 100 adults. In addition to total registration and vote counts, disaggregated counts by ethnicity (i.e., Hispanic and Asian), party affiliation (i.e., Democrat, Republican, and Independent/other), age, and duration of registration are also available. We combine these data with block-level demographic information on the voting age population from the 2000 and 2010 Censuses.

In order to examine the impact of police killings on voter preferences, we leverage voting data from two referenda that proposed changes to the severity of criminal sentencing laws. The first—Proposition 66 in 2004—would have limited California’s “three strikes” law to apply only to violent and serious felonies. The second—Proposition 5 in 2008—would have enacted numerous measures to reduce criminal penalties for drug offenses, including the reduction of marijuana misdemeanors to infractions, and the expansion of drug treatment and rehabilitation programs. Block-level estimates of the share of ballots cast for and against each proposition come from the California Statewide Database.Footnote ¹⁰

Analysis Sample

Since vote counts are only available until 2010, our main analysis focuses on the 2002–2010 general elections. To improve precision, we restrict the sample to blocks with five or more residents of voting age in the 2000 and 2010 Censuses. In robustness analyses, we find similar results under alternative sample restrictions and when extending the registration analysis to include more recent elections.Footnote ¹¹

Panel B of Table 1 provides a summary of the voter registration data. Relative to the full sample (column 1), residents in Census blocks where a police killing occurred (column 2) are more likely to be Black, Hispanic, Democratic, and young (18–34 years old), and have lower registration and voter turnout rates. Notably, however, Census blocks with police killings are demographically and politically similar to other Census blocks in the same Census block groups (column 3). Registered voters in the former are roughly equally likely to be Black (27% vs. 26%), Democratic (62% vs. 61%), or young (32% vs. 30%) as registered voters in the latter. The two sets of areas also experienced similar voter turnout (36% vs. 39%) and registration rates (54% vs. 58%) in 2002. Indeed, the main difference between these areas is population size (190 vs. 100 adult residents). In the next section, we explain how our research design leverages the similarity of Census blocks in the same block group to estimate causal effects.

Exposure to Police Killings

This article is interested in understanding the causal impact of police violence on civic engagement. Police shootings, however, are not randomly distributed across neighborhoods. Instead, they are concentrated in low-income neighborhoods with higher shares of minority residents. Thus, regressing voter turnout rates on local measures of police violence would likely be biased by meaningful geographic differences in political preferences and participation. Even controlling for area fixed effects would not resolve such biases, if changes in local crime or law enforcement correlate with the timing and location of police killings.

To address these concerns, we adopt a difference-in-differences approach similar to Ang et al. (Reference Ang, Bencsik, Bruhn and Derenoncourt2021) that exploits hyperlocal, within-neighborhood differences in exposure to police killings. At its core, our empirical strategy compares changes in civic behavior before and after police killings in the Census blocks where killings occurred (treatment blocks) to changes in civic behavior among other blocks in the same Census block group (control blocks). This strategy exploits the fact that the vast majority of sample police killings received little or no media coverage, such that any effects are likely to be highly geographically concentrated.Footnote ¹²

Our approach accounts for any time-invariant differences between treatment and control blocks as well as any time-varying factors that affect the entire Census block group similarly. However, time-varying factors that disproportionately affect treatment blocks relative to nearby control blocks would nonetheless bias our estimates. Thus, causal identification rests on the standard “parallel trends” assumption—absent police killings, treatment and control blocks in the same Census block group would have experienced similar trends in civic behavior. While this assumption is impossible to prove in the counterfactual, we provide supporting evidence by showing parallel trends in the elections leading up to police killings (i.e., see the discussion of Figure 2 under the Main Results section).

The validity of our strategy is further aided by two factors. First, as discussed earlier and demonstrated in Panel B of Table 1, Census blocks with police killings are strikingly similar to other blocks in the same Census block group in terms of racial, political, and demographic composition. Treatment and control blocks also experienced similar rates of civic engagement historically. Combined with their geographic proximity, these cross-sectional similarities provide greater confidence that treatment and control areas would have experienced similar trends in voting and registration in the counterfactual.

Second, police killings are quite rare and difficult to predict. Over three hundred thousand arrests and nearly 60,000 violent crimes occur in Los Angeles each year, compared with fewer than 50 officer-involved killings. Furthermore, many police killings are entirely unaccompanied by violent crime, as only a quarter of events involved armed suspects who fired at others. Thus, while crime rates and policing intensity may differ across neighborhoods, the exact timing and location of officer-involved shootings within those neighborhoods is plausibly exogenous. Indeed, as we demonstrate later, we find little evidence of differential changes in local crime or policing either before or after police killings (i.e., see the subsection Potential Threats to Identification).

Graphical Evidence

To corroborate our empirical strategy and demonstrate the hyperlocal impact of police killings, we first examine how geographic proximity to a killing mediates changes in voter registration. To do so, we construct the following distance metric measuring each Census block’s distance to a police killing: for each police killing, we estimate the minimum radius needed for a circle centered on the incident location to cover at least 75% of a block’s area.Footnote ¹³ Using this metric, we run the following generalized difference-in-differences regression:

(1) $$ \begin{array}{rl}{y}_{b,t}={\delta}_b+{\delta}_{n,t}+{\displaystyle \sum_d}{\alpha}_dDistanc{e}_d+{\delta}_{PO{P}_b\times ELE{C}_t}+{\epsilon}_{b,t}.& \end{array} $$

Here, $ {y}_{b,t} $ is the number of registered voters in block b at election t. $ {\delta}_b $ are Census block fixed effects. $ {\delta}_{n,t} $ are neighborhood-by-election fixed effects, where neighborhoods are defined as Census block groups, which average about 11 blocks and 0.61 square miles. The inclusion of these fixed effects allows for election-specific behavior to differ at the neighborhood level and effectively limits comparisons to treatment and control blocks in the same Census block group. Because block-level population counts are only available from the decennial Census, we include interactions between election fixed effects and deciles of estimated voting age population in 2002 ( $ {\delta}_{PO{P}_b\times ELE{C}_t} $ ) to account for the possibility of differential population growth between blocks. $ {\sum}_dDistanc{e}_d $ are a set of mutually exclusive treatment indicators that track a block’s distance to the nearest police killing that occurred prior to election t. Footnote ¹⁴ We partition shootings first by whether or not the block was directly exposed to a killing (i.e., occurred in the block), and then into 0.1-mile bins up to 2 miles from the shooting. Drawing on Bertrand, Duflo, and Mullainathan (Reference Bertrand, Duflo and Mullainathan2004), standard errors are clustered within each of the sample’s 6,400 Census block groups.

Figure 1 plots $ {\alpha}_d $ coefficients under our main specification (red diamonds). We also estimate an alternative specification replacing the neighborhood-by-election fixed effects with election fixed effects (blue dots). The coefficients represent the average difference between the pre-post change in registrations experienced by blocks with distance d and blocks in the omitted group (i.e., blocks in the “zero-impact” region between 0.7 and 0.8 miles from a killing).Footnote ¹⁵ The graphical evidence suggests that impacts are indeed hyperlocal and that residents living in nearby blocks may serve as a valid control for residents in Census blocks with a police killing. Across specifications, Census blocks with a police killing experience a significant increase in registrations of roughly five counts (6% of the pre-killing mean). Consistent with the underpublicized nature of police killings, effects fall off dramatically with spatial distance, with near-zero estimates for blocks only 0.1 miles from where a killing occurred. Figure 1 also provides further corroboration of the exogeneity of police killings. For some unobserved confound to generate these effects, it would have to coincide with the exact dates and locations of nearly three hundred police killings spread across several years and thousands of square miles.

Figure 1. Effects by Distance from Police Killing

Note: This figure shows coefficients and 95% confidence intervals from the estimation of Equation 1 on registrations. Unit of observation is the Census block election. The sample spans the 2002–2010 general elections. Indicators are mutually exclusive and track a Census block’s minimum distance to police killings that occurred prior to a given election. “In CB” refers to blocks where killings occurred. Other blocks are partitioned into 0.1-mile bins. Blue dots include election fixed effects. Red diamonds include Census block group by election fixed effects. Standard errors are clustered at the Census block group level. Full regression results are included in columns 1 and 2 of Table A.1 in the Supplementary Material.

Figure 2. Effects on Civic Engagement

Note: This figures shows coefficients and 95% confidence intervals from the estimation of Equation 2 on registrations (pre-killing mean $ =81.6 $ ) and votes (pre-killing mean $ =42.9 $ ). Unit of observation is the Census block election. Standard errors are clustered by Census block group. The sample spans the 2002–2010 general elections and treatment is defined by blocks where a police killing occurred during the sample period. The red vertical line marks when the police killing occurred. Full regression results are included in column 1 of Table A.3 in the Supplementary Material.

Primary Specification

To estimate effects on civic engagement, we next employ a dynamic difference-in-differences (i.e., event study) model. Drawing on the distance analysis, treatment is defined as Census blocks that experienced a police killing and neighborhood is defined at the Census block group level. We estimate the following base equation on the block-level panel data:

(2) $$ \begin{array}{rl}{y}_{b,t}={\delta}_b+{\delta}_{n,t}+{\displaystyle \sum_{\tau \ne -1}}{\beta}_{\tau }Shoo{t}_{t,\tau }+{\delta}_{PO{P}_b\times ELE{C}_t}+{\epsilon}_{b,t}.& \end{array} $$

This is essentially analogous to Equation 1 except that we replace the set of treatment distance indicators ( $ \sum {\alpha}_dDistanc{e}_d $ ) with a set of time to treatment indicators ( $ \sum {\beta}_{\tau }Shoo{t}_{t,\tau } $ ), fixing treatment to the first killing that occurred in a Census block between the 2002 and 2010 general elections.Footnote ¹⁶ The coefficients of interest ( $ {\beta}_{\tau } $ ) represent the differential change between relative time $ \tau $ and the last period prior to the police killing in the incident block relative to that same change over time among other blocks in the same Census block group.

Main Results

Turning to our primary results, Panel A of Figure 2 examines the impact of police killings on local registration counts. The omitted period is the last election prior to a killing and the sample spans the 2002 to 2010 general elections. In the elections prior to a killing, we find strong evidence of parallel trends between treatment and control blocks. Treatment coefficients for $ \tau <0 $ are near zero and statistically insignificant, both individually and jointly ( $ F=0.33,p=0.806 $ ). These findings reinforce the plausible exogeneity of police killings and provide support for parallel trends in the counterfactual. Following police killings, registration increases significantly among nearby citizens. Treated blocks gain, on average, about 2.5 additional registrants in the election immediately following the killing and about 4.5 registrants within 4 years. Given that treated blocks contain an average of 80 registrants prior to treatment, these effects represent a meaningful increase of 3%–5%. The stability of point estimates 4–8 years after exposure suggests that effects on registration are persistent over time. As corroboration, Figure A.2 in the Supplementary Material expands the sample to include police killings and elections through 2016 and finds significant effects on registration more than a decade after a killing.Footnote ¹⁸

Panel B presents analogous results for vote counts. We again find little evidence of differential trends in ballots cast in the lead-up to police killings. The pretreatment coefficients are individually and jointly insignificant ( $ F=1.09,p=0.350 $ ). After killings, we find a significant, if short-lived, increase of approximately two votes (5% of the pre-killing mean).

Table 2 presents a series of robustness checks under alternative specifications. For concision, column 1 estimates our base model with a single posttreatment indicator, representing the average treatment effect across future elections. We find that, on average, police killings lead to 3.6 more registrations and 1.7 more votes per election in treated blocks. To account for potential confounds due to local crime, column 2 controls for the number of homicides in a block in the 2 years preceding each election. Column 3 includes interactions between quintiles of minority population share in 2000 and election fixed effects to allow for differential voting trends among minority neighborhoods, which may be more likely to experience police killings. Given that younger individuals may have aged into voting eligibility during the sample period, column 4 includes interactions between quintiles of adolescent population (i.e., ages 10–17) in 2000 and election fixed effects. To account for differential population growth, column 5 includes interactions between percent population change from 2000 to 2010 and election fixed effects. In contrast, column 6 excludes all population controls from the regression. To demonstrate robustness to sample selection, column 7 drops the single Census block that experienced multiple distinct police killing incidents over the sample period, whereas column 8 expands the sample to include all Census blocks, even those with less than five adults in 2000 or 2010. Alternatively, column 9 restricts the sample to Census blocks with at least 10 registered voters in 2002.

Table 2. Effects on Civic Engagement: Alternative Specifications

Note: This table shows results from the estimation of Equation 2 on registrations and votes, replacing time to treatment indicators with a single posttreatment indicator. Column 1 examines our preferred specification. Column 2 controls for the number of homicides in a block in the 2years preceding each election, yielding $ {\beta}_{homicide} $ of 0.320 ( $ p=0.079 $ ) for registrations and 0.215 ( $ p=0.079 $ ) for votes. Column 3 includes interactions between minority share quintiles in 2000 and election fixed effects. Column 4 includes interactions between quintiles of population aged 10–17 in 2000 and election fixed effects. Column 5 includes interactions between percent population change (from 2000 to 2010) and election fixed effects. Column 6 removes all population controls (i.e., interactions between population decile and election fixed effects). Column 7 drops the single Census block that experienced more than one police killing during the sample period. Column 8 includes all Census blocks, even those with less than five people over the age of 18 in 2000 or 2010. Column 9 restricts the sample to Census blocks with 10 or more registered voters in 2002. Unit of observation is the Census block election. The sample period spans the 2002–2010 general elections. Standard errors are clustered by Census block group. Mean registrations and votes are reported for treatment blocks in the election prior to the killing. *** $ p<0.01 $ , ** $ p<0.05 $ , and * $ p<0.10 $ .

We find similar results across all specifications, with significant and positive effects on voter registration and turnout. These findings provide evidence of the robust causal relationship between police killings and local political participation. In particular, we find that individuals are mobilized to register and vote by extreme acts of police violence.

Heterogeneity

Voter and Deceased Race

Given large demographic and partisan differences in views of law enforcement, we examine heterogeneous responses to police violence. To explore how effects differ across voter race, we make use of vote and registration counts by ethnicity provided by the California Statewide Database. The database predicts Hispanic ethnicity from voter surname using the Census Bureau’s Passel-Word list and Asian ethnicity using the surname dictionary of Lauderdale and Kestenbaum (Reference Lauderdale and Kestenbaum2000).Footnote ¹⁹ From these measures, we generate estimates of Black (white) vote and registration counts using the following formula:

(3) $$ VoteBl{k}_{b,t}={\displaystyle \begin{array}{l}\left(VoteTo{t}_{b,t}-VoteHis{p}_{b,t}-VoteAs{n}_{b,t}\right)\\ {}\hskip-6em \times \left(\frac{\%Bl{k}_{b,2010}\times \%VoteBl{k}_t}{\left(\%Bl{k}_{b,2010}\times \%VoteBl{k}_t\right)+\left(\%Wh{t}_{b,2010}\times \%VoteWh{t}_t\right)+\left(\%Ot{h}_{b,2010}\times \%VoteOt{h}_t\right)}\right),\end{array}} $$

where $ VoteTo{t}_{b,t} $ , $ VoteHis{p}_{b,t} $ , and $ VoteAs{n}_{b,t} $ are the number of total votes, Hispanic votes, and Asian votes in block b at election t, and $ \%Bl{k}_{b,2010} $ , $ \%Wh{t}_{b,2010} $ , and $ \%Ot{h}_{b,2010} $ are the share of residents over age 18 who are Black, white, and other race from the 2010 Census. To account for racial differences in voter turnout rates, we weight by each racial group’s statewide turnout rate in election t as estimated by the CPS Voting and Registration Supplement ( $ \%VoteBl{k}_t $ , $ \%VoteWh{t}_t $ , and $ \%VoteOt{h}_t) $ .Footnote ²⁰ Essentially, we weight non-Hispanic, non-Asian votes and registrations in a given block election by each racial group’s predicted vote and registration share relative to the other remaining racial groups.Footnote ²¹

We then estimate our simplified version of Equation 2 on predicted counts by race. As shown in Panel A of Figure 3, a striking pattern emerges. Police killings lead to large increases in Black and Hispanic participation. On average, each police killing increases Black registrations by 1.8 and Black votes by 1.0. These estimates are highly significant and represent a 9%–11% increase over the pre-killing mean (20.2 registrations and 9.2 votes). We find similar, if proportionally smaller, responses among Hispanics with increases of 1.4 registrations (5% of mean) and 1.0 votes (6% of mean). In contrast, we find no significant impact on white and Asian participation, with point estimates representing less than 2% of the pre-killing mean.

Figure 3. Heterogeneous Effects

Note: The panels show coefficients and 95% confidence intervals from the estimation of Equation 2 on registrations and votes, collapsing the time to treatment indicators into a single posttreatment indicator. Panels A and C–E examine the effects separately for each voter group (i.e., by race, age, registration length, and party affiliation). Panels B and F include distinct posttreatment indicators corresponding to each incident type (i.e., by deceased race in Panel B and by proximity to deceased home in Panel F). Standard errors are clustered by Census block group. Unit of observation is the Census block election. The sample spans the 2002–2010 general elections and treatment is defined by blocks where police killings occurred during the sample period. Full regression results are included in Table A.5 in the Supplementary Material.

Further disaggregating by race of the deceased in Panel B, we find suggestive evidence that racial concordance between the voter and the deceased may lead to larger effects. Among Black and Hispanic voters, point estimates for registrations (votes) are roughly 200% (75%) larger for killings of Black and Hispanic individuals than for killings of white and Asian individuals. Similarly, for white and Asian voters, we find positive, if noisy, estimates for killings of white and Asian individuals, but near zero estimates for killings of different-raced individuals.

That effects are primarily concentrated among Black and Hispanic citizens is consistent with a host of evidence documenting large racial differences in perceptions of law enforcement. Researchers have found that race is the single strongest predictor of trust in police (Taylor et al. Reference Taylor, Turner, Esbensen and Winfree2001) and that Black and Hispanic individuals are far more likely than others to believe that use of force is excessive, unjustified, or a serious social concern (AP-NORC 2015; Davis, Whyde, and Langton Reference Davis, Whyde and Langton2018; Weitzer and Tuch Reference Weitzer and Tuch2002).

Family and Household Responses

To interrogate the possibility that effects are driven by family and household members of the deceased, we obtained full-count voter registration files from 2004 to 2010 and searched for new registrants who lived in the Census block of a killing and bore the same last name as the deceased. Across 204 killings that occurred in this window, we identified 25 total matches or roughly 0.12 new surname-match registrants per killing. While this is obviously an imperfect measure of kinship networks, it suggests that increased participation among family members of the deceased cannot account for the significant effects on voter participation, which average 3.5 new registrants in the surrounding block.

As further corroboration, we examine whether effects differ depending on whether a police killing occurred near the home of the deceased. While we do not observe the exact home address of all individuals killed by police, we are able to infer whether a killing occurred near the deceased’s residence based on district attorney incident report descriptions.Footnote ²³ If increased participation was driven primarily by the deceased’s household, we would expect police killings that occurred far from the deceased’s home to have little impact on registration rates in the block of the incident. However, as shown in Panel F of Figure 3, we find significant effects regardless of incident location. This suggests that less proximal individuals, such as witnesses, neighbors, and other local community members—as opposed to family or household members—likely account for much of the mobilization effects.

Role of Media

Given the recent proliferation of viral footage capturing acts of police violence, we interrogate the role of media in community responses to these events. Specifically, we examine how spatial spillovers vary by media coverage. To do so, we estimate separate distance gradients for killings that were and were not mentioned in local newspapers using Equation 1.

As shown in Figure 4, we find suggestive evidence of wider spread effects following police killings that were covered in the media than those that were not. While effects for unmentioned killings are contained to the immediate block, we find statistically significant effects extending to blocks nearly 0.50 miles away for killings mentioned in the media. While media coverage may simply reflect (rather than influence) community awareness or perceptions of an incident, these patterns may nonetheless help to explain the discrepancy between the national responses to recent high-profile police killings and the highly localized effects we observe here.

Figure 4. Effects by Media Coverage

Note: This figure shows coefficients and 95% confidence intervals from the estimation of Equation 1 on registrations, separately for killings that were and were not reported in local newspapers. Unit of observation is the Census block election. The sample spans the 2002–2010 general elections. Indicators are mutually exclusive and track a Census block’s minimum distance to police killings that occurred prior to a given election.“In CB” refers to blocks where killings occurred. Other blocks are partitioned into 0.1-mile bins. The Census block group by election fixed effects and Census block group-clustered standard errors are included. Full regression results are included in columns 3 and 4 of Table A.1 in the Supplementary Material.