The partisanship of mayors has no detectable effect on police spending, police employment, crime, or arrests

Abstract

In this paper, we examine whether mayors’ partisan affiliations lead to differences in crime and policing. We use a large new dataset on mayoral elections and three different modern causal inference research designs (a regression discontinuity design centered around close elections and two robust difference-in-differences methods) to determine the causal effect of mayoral partisanship on crime, arrests, and racial differences in arrest patterns in medium and large US cities. We find no evidence that mayoral partisanship affects police employment or expenditures, police force or leadership demographics, overall crime rates, or numbers of arrests. At the same time, we find some suggestive evidence that mayoral partisanship may modestly affect the racial composition of arrests. Overall, the results from our multimethod analyses indicate that local partisan politics has little causal impact on crime and policing.

Across hundreds of cities and three decades, there is no impact of a mayor’s partisan affiliation on crime and arrest rates.

INTRODUCTION

In the lead-up to the 2020 presidential election, President Trump blamed increases in crime on Democratic leaders, arguing that Democrat-run cities were “rampant with crime” (1). Other Republican politicians have often claimed that increases in crime in large cities are a result of the “soft-on-crime” policies of Democratic leaders in those places (2). Much of the national news narrative around elections for prosecutors and mayors has echoed these claims (3–6). Yet these arguments from politicians and pundits alike rest upon an empirical assumption that Democratic leadership leads to increases in crime in cities.

Why might local leadership influence criminal justice outcomes? Levitt (7) was one of the first to attempt to hurdle potential endogeneity concerns and establish that increased police spending and hiring can lower crime rates. Using newer empirical techniques, recent literature has continued to establish the connection between officers and crime rates (8–10). Republican mayors commonly take a “tough on crime” approach and promise that they will reduce local crime rates by hiring more officers and, often, via punitive policy (11, 12). This suggests that the election of Republican leaders may decrease local crime rates. However, political constraints may often curtail an official’s ability to use policy tools to influence local crime rates (13, 14). Thus, it is unclear whether mayoral partisanship should affect crime and arrests.

Here, we examine this question directly using the most comprehensive data available on local elections, policies, and outcomes and find no evidence that the mainstream narrative around partisanship and crime in cities is correct. We use three different methodological techniques to identify the causal impact of mayoral partisanship. Specifically, we use a regression discontinuity design (RDD) that focuses on narrowly won elections where we can identify the effect of the mayor’s party without confounding from other characteristics of cities. We augment this with two other causal inference techniques—both of which are generalized difference-in-differences methods that leverage changes in mayoral partisanship within cities—to holistically examine the causal impact of partisanship. To do so, we draw on nearly three decades worth of data on mayoral elections in nearly 400 medium and large American cities that encompass 99% of the population in the universe of cities more than 75,000 in population that elect mayors. We combine these election data with fiscal policy data, original data on police leadership, and administrative data on crime and policing. To our knowledge, no previous study has systematically assessed the impacts of mayoral partisanship on the wide set of outcomes within the domain of public safety and policing; therefore, empirical evidence on the matter is lacking.

Overall, our findings show that mayors’ partisanship has limited influence on crime and policing. Electing a Democrat rather than a Republican as mayor leads to no detectable impact on police staffing or expenditures on criminal justice, nor does it lead to changes in crime or arrest rates. Racial differences in policing are also mostly unaffected by mayoral partisanship, with a few potential exceptions. Electing a Democratic mayor rather than a Republican mayor appears to marginally decrease the Black share of individuals arrested for several types of crimes and marginally increase the Black share of law enforcement officers. Yet the first of these results is not consistently robust to alternative measures of calculating racial disparities in arrests nor the use of alternative research designs, and the latter is not verifiable using reliable data.

Our study draws on data spanning cities across the country and multiple decades. There are clear advantages to this approach with regard to statistical power and generalizability. However, this approach limits our analysis to outcomes which are available at that scale. There are increasingly rich data on more finely grained policing outcomes, including, for example, the race of individuals involved in traffic stops and the specific outcomes of those stops: whether people were searched, a citation was issued, or no action was taken (15, 16), but these data are not available for a wide enough set of cities or years to be used in our study. While our results largely document similar outcomes across Republican- and Democrat-run cities, it remains possible that there are differences in more nuanced outcomes. Our suggestive finding that Democrat-led cities have a higher share of Black officers, which has been documented to have downstream effects for racial disparities in policing (17, 18), suggests that there may be effects of local partisanship that are not detectable using the data at hand.

Background

A long line of research in political science demonstrates that parties shape the behavior of national- and state-level politicians. Republican politicians have more ideologically conservative policy positions and voting records than Democratic politicians in Congress and state legislatures (19, 20), and this polarization between parties has only increased over time (21). Recent research has suggested that local elections have grown increasingly nationalized (22, 23), and partisan effects on policy outcomes extend to the local level (24–28). In particular, the election of a Democratic mayor (compared to a Republican mayor) leads to greater municipal expenditures (24).

Policing is an important area of local politics where mayoral partisanship may matter. Republican mayors often promise that they will reduce local crime rates by increasing police spending and making criminal justice policies more punitive (11, 12). Democratic mayors and other local politicians often promise to reduce racial inequities in policing (29). Much of the national narrative around crime and policing suggests that the Republican party has stronger ownership of crime and public safety as a policy issue (30).

However, mayors have important constraints on their ability to unilaterally influence policy (31, 32). Mayors often face oversight both from below (by city councils) and above (by states). Potential policy changes are limited by budget restrictions and civil service rules. Limits on mayors’ ability to reform policing may be especially pronounced with the rise of police unions’ strength in the final decades of the 20th century (33) as well as the activity of these interest groups in local politics more broadly (34).

Moreover, pledges to increase spending on police often lie in tension with conservatives’ larger policy goal of shrinking the size of government, leading to little partisan polarization in mayors’ policy preferences in this area (35). This may simply be a policy area where politicians are unlikely to diverge along partisan lines in the policies they pursue (36). Thus, perhaps it is expected that there is mixed evidence from past research on the effect of mayoral partisanship on police expenditures (24, 31).

Yet there has been little prior work on whether mayoral partisanship affects overall crime rates or arrest patterns. Some work has shown it has no impact on rates of murder, larceny, robbery, or burglary (37). In contrast, descriptive work finds that Democratically led cities tend to have higher levels of crime—in large part because of underlying demographic patterns (37, 38).

Mayoral partisanship could also affect the racialization of policing. While Democrats and Republicans alike campaign on promises to reduce crime (39–41), Democrats tend to focus more than Republicans on reducing racial and socioeconomic inequalities in criminal justice policies (29, 42). These inequalities are widespread: Previous academic work has shown that Black drivers are more likely to be stopped than white drivers; searches of Black drivers are less likely to produce “contraband,” indicating a lower threshold for pulling them over (15, 43). Contact with police is more likely to lead to arrest for Black Americans even controlling for contextual factors (44), with substantial evidence of this phenomenon in the context of drug arrests in particular (45, 46). Black Americans are also more likely to face arrest or experience force at the hands of white officers relative to officers of color (17, 47, 48). Recent work integrates such studies with questions about the impact of partisanship. One study, for instance, documents that law enforcement officers are more likely to be Republican than the jurisdictions they police. In Florida, white Republican officers exhibit a larger racial disparity in policing than white Democratic officers—although disparities grew for both groups in the 2010s (49). Thus, we might expect that Democratic mayors could reduce racial inequalities in policing relative to Republicans.

RESULTS

In this section, we discuss the impact of mayoral partisanship on criminal justice outcomes in cities (for details on our data and research design, please see Materials and Methods below). First, we examine the impact of mayors on overall police expenditures and staffing. Next, we assess how partisanship affects the demographic composition of police forces and the police chiefs who lead those forces. Last, we examine whether mayoral partisanship affects aggregate crimes and arrests, as well as the racial composition of arrests.

Police expenditures and staffing

Republican politicians’ tough on crime rhetoric implies that they would raise police expenditures and staffing levels relative to Democrats. We evaluate this hypothesis by examining the impact of mayoral partisanship on police employment levels [using the Annual Survey of Public Employment and Payroll (ASPEP) data] and related municipal expenditures (using the Historical Database of Individual Government Finances). Figure 1 displays these results. Each point represents the estimated increase in the noted outcome when a Democrat rather than a Republican is mayor—in other words, the causal effect of electing a Democratic mayor rather than a Republican. The bars emanating from each point are 90 and 95% confidence intervals. Each symbol in the figure represents the estimate from a different research design, with the RDD results represented by stars, PanelMatch results represented by crossed circles, and FEct results represented by vertical lines (see Materials and Methods for details). Supplementary Section D also has tabular versions of all results. These tables show the estimates and associated confidence intervals, as well as the bandwidth selected by rdrobust and the effective n within that bandwidth, and the number of matched observations for PanelMatch results.

Fig. 1. The null effect of mayoral partisanship on municipal police employment and criminal justice spending.

Points indicate causal effect estimates from each of our three research designs: from the RDD estimated with rdrobust (stars); from PanelMatch (crossed circles); and FEct (vertical lines), and horizontal lines indicate 90% (thick lines) and 95% (thin lines) confidence intervals, using robust bias-corrected confidence intervals for the RDD.

The top set of results in Fig. 1 indicates that electing a Democratic mayor has little detectable impact on the per capita total number of police officers employed by a city. We replicate these analyses using RDD on two other datasets that also track the number of police officers employed by a city: the Law Enforcement Management and Administrative Statistics (LEMAS) and Law Enforcement Officers Killed and Assaulted (LEOKA) datasets. Although we are most inclined to trust the estimates from the Census Bureau’s ASPEP rather than the other two datasets, which are based on voluntary opt-in surveys of police departments, we present all three results for the sake of transparency and completeness in Supplementary Text E.

The estimated impact on the size of the police force is small and statistically indistinguishable from zero when using all three research designs (RDD, PanelMatch, and FEct). The largest of these three estimates is roughly equivalent to 2.5% of the mean of the outcome, as reported in Table 1; the smallest is ¹/₁₀ of that. The second and third sets of results display the estimates of mayoral partisanship on two categories of municipal spending: those directly on police protection as well as expenditures on corrections. In both cases, we find no discernible impact of electing a Democrat as mayor on per capita spending in these policy areas related to criminal justice. The largest estimated effect in the middle panel suggests a 1.2% increase relative to the mean of the outcome. The lack of any partisan impacts on overall policing employment or expenditures suggests that despite Republicans’ tough on crime platforms, there may not be substantial effects of mayoral partisanship on policy levers related to criminal justice.

Table 1. Summary statistics for main outcome variables, part 1.

	Total	Democratic	Republican	Party unknown
	(N = 12,436)	(N = 5757)	(N = 4143)	(N = 2536)
Total sworn officers per 100,000 capita
Mean (SD)	200 (87.9)	222 (98.0)	171 (61.1)	167 (59.6)
Median [min, max]	181 [0, 969]	202 [0, 969]	160 [0, 620]	153 [0, 403]
Missing	3841 (30.9%)	858 (14.9%)	805 (19.4%)	2178 (85.9%)
Police expenditures per capita ($)
Mean (SD)	287 (121)	330 (135)	259 (93.1)	236 (89.8)
Median [min, max]	265 [0, 1280]	307 [0, 1280]	245 [0, 985]	222 [0, 1240]
Missing	1173 (9.4%)	515 (8.9%)	496 (12.0%)	162 (6.4%)
Corrections expenditures per capita ($)
Mean (SD)	14.7 (56.0)	21.4 (70.7)	6.43 (28.4)	12.8 (49.2)
Median [min, max]	0 [0, 1300]	0 [0, 1150]	0 [0, 276]	0 [0, 1300]
Missing	1173 (9.4%)	515 (8.9%)	496 (12.0%)	162 (6.4%)
Police chief race/ethnicity
Black	678 (5.5%)	524 (9.1%)	151 (3.6%)	3 (0.1%)
Hispanic	261 (2.1%)	151 (2.6%)	93 (2.2%)	17 (0.7%)
White	4307 (34.6%)	2247 (39.0%)	1783 (43.0%)	277 (10.9%)
Missing	7190 (57.8%)	2835 (49.2%)	2116 (51.1%)	2239 (88.3%)
Police chief gender
Woman	284 (2.3%)	206 (3.6%)	75 (1.8%)	3 (0.1%)
Man	4880 (39.2%)	2674 (46.4%)	1912 (46.2%)	294 (11.6%)
Missing	7272 (58.5%)	2877 (50.0%)	2156 (52.0%)	2239 (88.3%)
Black share of police force
Mean (SD)	0.0921 (0.109)	0.124 (0.133)	0.0568 (0.0606)	0.0661 (0.0710)
Median [min, max]	0.0555 [0, 0.854]	0.0833 [0, 0.854]	0.0368 [0, 0.446]	0.0377 [0, 0.548]
Missing	1977 (15.9%)	529 (9.2%)	666 (16.1%)	782 (30.8%)
Hispanic share of police force
Mean (SD)	0.101 (0.148)	0.107 (0.153)	0.106 (0.146)	0.0741 (0.134)
Median [min, max]	0.0523 [0, 1.00]	0.0524 [0, 1.00]	0.0621 [0, 1.00]	0.0323 [0, 1.00]
Missing	1977 (15.9%)	529 (9.2%)	666 (16.1%)	782 (30.8%)
White share of police force
Mean (SD)	0.780 (0.183)	0.739 (0.194)	0.809 (0.170)	0.843 (0.145)
Median [min, max]	0.829 [0, 1.00]	0.777 [0, 1.00]	0.855 [0, 1.00]	0.875 [0, 1.00]
Missing	1975 (15.9%)	527 (9.2%)	666 (16.1%)	782 (30.8%)
Woman share of police force
Mean (SD)	0.106 (0.0489)	0.119 (0.0521)	0.0988 (0.0407)	0.0875 (0.0448)
Median [min, max]	0.101 [0, 0.904]	0.114 [0, 0.874]	0.0966 [0, 0.270]	0.0822 [0, 0.904]
Missing	821 (6.6%)	245 (4.3%)	283 (6.8%)	293 (11.6%)

Police demographics

We next examine the impact on not only the size of the police force but also the demographics of the police. In particular, one crucial tool at the disposal of mayors is (in most places) the power to appoint a police chief who plays a direct role in both police staffing and officers’ everyday practices when doing their job. Although our police chief data are limited in their time span to the 2010–2022 period, these results do represent the most thorough investigation of police chief demographics with the data available. We also examine the demographic composition of the entire police force. Police force demographics are especially relevant for two reasons: (i) Hiring and personnel decisions may be some of the more immediate levers available to a mayor (50), including indirectly through police chief selection, and (ii) a growing body of work documents how demographics of law enforcement officers affect policing outcomes—including on outcomes that we are not able to observe. Black and Hispanic officers make fewer stops and arrests than white officers and are less likely to use force than white officers, while agencies run by white officers have higher Black-to-white arrest ratios (17, 47). Women officers are less likely to search drivers during traffic stops (51) and less likely to use force than officers who are men (17).

Our analyses of police chief and police force demographics are displayed in Fig. 2 using all three of our methodological techniques. The results for police chiefs are shown in Fig. 2A. The results provide no consistent evidence that a Democratic mayor is more likely than a Republican mayor to appoint a police chief of any particular demographic group. Our estimates of the causal impact of mayoral partisanship on the probability of having a Black police chief, a Hispanic police chief, a white police chief, or a woman police chief are all statistically indistinguishable from zero. Table 1 shows that 5.5% of police chiefs we have identified are women, and 18% are Black. Our largest estimates for each group suggest large (albeit imprecise) relative effects—of nearly three-fourths of the mean in the case of both women as police chiefs and 85% of the mean in the case of Black police chiefs.

Fig. 2. The effect of mayoral partisanship on changes in the demographics of police chiefs and the police force between the election year and the several years after the election.

Points indicate causal effect estimates for police chiefs (A) and the police force (B) from each of our three research designs: from the RDD estimated with rdrobust (stars); from PanelMatch (crossed circles); and FEct (vertical lines), and horizontal lines indicate 90% (thick lines) and 95% (thin lines) confidence intervals, using robust bias-corrected confidence intervals for the RDD. The horizontal axis of (B) is cut off to maintain presentational legibility, but the upper limits of the confidence intervals on the white share of the police extends beyond the plot limits (see Supplementary Section D for full tabular results).

Our results examining the demographics of the police force as a whole (in Fig. 2B) using our three different research designs are more mixed. Our RDD indicates that the share of officers who are Black significantly increases by approximately one percentage point as a result of electing a Democrat (rather than a Republican) as mayor ($P=0.04$); relative to the sample average of 9.5% Black share of police officers, this represents an increase of roughly 9.9%. Yet the two difference-in-differences strategies do not lead to this conclusion and suggest that changing from a Republican to a Democratic mayor has no effect on the Black share of the police force. Also, for other racial groups and for gender, none of our research designs indicate that mayoral partisanship influences the demographic makeup of the police force.

Overall crime, arrests, and policing

We next examine the empirical impact of mayoral partisanship on overall crime, clearance rates, and arrests in Fig. 3. The horizontal axis reports our estimates of the causal impact of Democratic mayors (versus Republican ones) on the overall per capita levels of reported crime and clearance rates (Fig. 3A) and overall numbers of arrests (Fig. 3B).

Fig. 3. The null effect of mayoral partisanship on changes in crime and arrests between the election year and 3 years after the election.

Points indicate causal effect estimates for per capita reported crimes and clearance rate (A) and per capita arrests (B) from each of our three research designs: from the RDD estimated with rdrobust (stars); from PanelMatch (crossed circles); and FEct (vertical lines), and horizontal lines indicate 90% (thick lines) and 95% (thin lines) confidence intervals, using robust bias-corrected confidence intervals for the RDD.

Notably, none of the estimates are statistically different from zero. For example, the top point in Fig. 3A shows that electing a Democrat as mayor rather than a Republican has no detectable causal effect on overall levels of crime, and the point estimate is negative from two of our research designs. The confidence interval around this estimate, while wide, allows us to rule out increases in overall crime of more than 0.32 crimes per 100 capita. Even that would represent a small effect relative to the sample average of 6.6 total crimes per 100 capita.

Given concerns about potential underpowered analyses in RD analyses in particular (52), we also conduct post hoc power analyses (53, 54). These power calculations indicate that the probability of rejecting the null hypothesis were the true effect to be equivalent in size to 0.54 crimes per capita, or half an SD, is relatively high, at 0.87. This—alongside the fact that our two other research designs come to similar conclusions—suggests that we are unable to reject the null hypothesis of no effect on crime not due to a lack of statistical precision but instead the small size of these effects.

We also find no discernible effect of mayoral partisanship on crime when disaggregating to available categories of crimes (index, property, and violent), which we show in the next set of points in Fig. 3A. Likewise, we cannot reject the null hypothesis of no change in the clearance rate—the share of reported crimes for which an arrest was made (shown with the bottom points and lines in Fig. 3A).

In Fig. 3B, we focus on arrests. The estimated impact of Democratic mayors on total arrests per capita is also not statistically distinguishable from zero. Post hoc power calculations for the RDD analyses indicate that these are relatively precisely estimated nulls and that the probability of rejecting the null hypothesis were the true effect to be equivalent in size to half an SD in the outcome, or 0.31 arrests per 100 capita, is 0.7. We also test for impacts of the mayor’s party on specific categories of arrests. We find that the estimated impacts of partisanship on overall numbers of violent, property, drug, and other crime arrests are close to zero and are all statistically insignificant. Overall—alongside our earlier analyses—our findings indicate that mayoral partisanship is not causally associated with differences in the levels of police employment, police spending, reported crimes, or arrests.

While mayoral partisanship has no effect on overall crime or arrests, it is possible that it leads to changes in the way that police forces act in the conduct of their jobs—and specifically the racial composition of their arrests. Figure S20 in Supplementary Text F depicts our causal estimates of mayoral partisanship on the Black share of total arrests as well as the Black share of arrests for each category of crimes separately. We find suggestive evidence that electing a Democratic mayor rather than a Republican mayor marginally decreases the Black share of individuals arrested for several types of crimes. But these findings are not robust across research designs and generally fall below conventional levels of statistical significance.

DISCUSSION

Here, we examine whether political control of city governments in the US influences local policing, crime rates, and arrests. Using a large dataset of local elections and three different credible research designs, we are able to disentangle the effect of mayoral partisanship from other city-level characteristics that might affect policing and crime outcomes.

We find no detectable effects of mayoral partisanship on overall police employment and criminal justice expenditures, the demographics of the police, overall levels of crime, numbers of arrests, or the clearance rate of crimes. These results stand in stark contrast to national political rhetoric on policing, crime, and political partisanship. Candidates for political offices at the local, state, and federal level consistently raise crime as an important issue (55–57). Voters may hold politicians accountable for these types of outcomes (58–60), at least partially as a result of increased news coverage (61)—even if this coverage does not always match reality (62).

When we examine the effects of city leadership on outcomes in the area of race and policing, we again find no consistent evidence that mayoral partisanship influences the demographic composition of police forces. Previous work on police use of force suggests that the racial composition of police officers can have a strong effect on racial disparities in policing activity and violence (18, 63, 64). The demographics of the police force might therefore be one way in which local politicians exert control to reduce racial disparities in citizen-police contact. Yet our results yield no consistent evidence when it comes to both police chief and police force demographics across the thorough examination that we conduct using three different research designs.

Republican politicians in particular have claimed that Democrats’ soft-on-crime policies have led to crime increases in large cities (2). National news outlets have pointed to the outcomes of recent recall efforts and elections for prosecutors and mayors to suggest that they represent a backlash to progressive policies on crime (3–5). Yet these popular claims ignore the reality that our results make clear: Mayoral partisanship of cities does not lead to any detectable causal differences in crime or arrests. If the partisanship of leaders does not influence objective performance metrics, then voters may struggle to hold those leaders accountable along partisan lines (14). Our results suggest that this may be true for mayors and criminal justice outcomes and suggest that accountability in local politics must rely on some form of retrospection regardless of party instead.

Overall, our results indicate that politics—and in particular partisan politics—play a limited role in crime and policing. Partisanship has little causal impact on bottom-line outcomes like crime or arrest rates. There is also only limited evidence that partisanship affects the demographics of the police or the racialized nature of citizen-police interactions. Our analyses are inherently limited in their ability to uncover other mechanisms by which mayoral partisanship might influence police behavior due to a lack of broad intermediate data on mayoral influence or direct policymaking. Nor do our analyses enable us to discern the effect of other influences on policing, such as police union strength or other activity by police unions and affiliated interest groups (34, 65).

Future research could focus more on these intermediate steps in the causal chain between elections and criminal justice outcomes. It could also focus on within-party variation in policymaking. For instance, mayors might appoint different types of police chiefs (including “progressive” or “reform” chiefs) regardless of race, or they may require their police forces to undergo certain types of training to reduce racial biases in policing. Our analyses can neither assess whether these mechanisms are at play nor the bevy of mechanisms that operate together to result in racialized policing. Overall, our results help build a more complete picture of crime and policing in US cities by providing evidence that local partisanship is not a central driver of differences in crime and policing in US cities.

MATERIALS AND METHODS

To examine the policy effects of the partisan control of city governments, we use data on city mayoral elections and criminal justice policy and outcomes in medium and large cities with a population of more than 75,000 people in 2020. We then leverage three different research designs to identify the causal effect of electing mayors of different parties on crime and policing.

Data

The foundation of our analysis is administrative data on mayoral elections in medium and large cities. The election data consist of 3254 individual elections between 1990 and 2021 in 398 cities with at least 75,000 people in 2020 (66). These election data cover 99% of the target population of cities above this population threshold that hold mayoral elections (see Supplementary Text A for more details on the coverage of our election data). Our analysis requires information on each candidate’s partisanship even in officially nonpartisan elections. We therefore use data on raw election returns augmented with information about individual candidates’ partisanship from matching with a wide range of auxiliary data. These auxiliary data include information from both L2 and TargetSmart’s national voter files, as well as campaign finance–based ideology scores (67, 68), information on candidates who served in Congress or state legislatures (20, 69), and information from previous academic studies (31, 37).

We examine the impact of mayoral partisanship on a number of criminal justice outcomes. First, we use data on fiscal policy from the Historical Database of Individual Government Finances to examine the effects of mayoral partisanship on policing-related government expenditures. These data are based on a Census of Governments conducted every 5 years and the Annual Survey of Governments collected in every noncensus year. We adjusted all monetary figures into 2019 dollars based on the consumer price index. We also harness data from the Census Bureau’s ASPEP, which records both the number of employees of different types and the payroll expenditures on those employees for local governments.

To examine not just the overall size of the police force but also the composition of the police force, we draw on several sources of data. First, we gather original data on the names (and demographics) of police chiefs in the cities in our election data for the period 2010–2022.

We determined chiefs’ gender and racial backgrounds based on the media’s identification or self-identification of chiefs in news articles or via biographies on personal or departmental websites that listed race/ethnicity or membership in a group based on race/ethnicity (e.g., “member of the National Organization of Black Law Enforcement Executives”) or, in some cases, based on multiple photos of chiefs. These data on police chiefs are limited in their time span to the 2010–2022 period due to the difficulty of finding news articles on the subject before 2010, so we caution that our results using these data are by no means the definitive answer on partisan control of cities and police chiefs, but they represent an important first attempt at the question.

We also use data on police officer demographics from LEMAS, a survey of law enforcement agencies administered by the Bureau of Justice Statistics roughly every 4 to 5 years since 1987. The most recent available wave of the survey is from 2020. These surveys, for most of the survey waves, provide data on the racial demographics of officers as well as their gender. We merge these data with our election data to create outcomes based on changes in police demographics between the most recent LEMAS survey before the election and the next LEMAS survey after the election. The LEMAS data are an imperfect source of longitudinal information on police forces, given that they are not conducted yearly, so we restrict our use of this outcome variable to surveys within a reasonable time range around the election by only using baseline presurveys that were 0 to 4 years before the election and postsurveys that were 2 to 4 years after the election. We supplement this demographic information with data from the Federal Bureau of Investigation (FBI)’s LEOKA survey, which (along with information on violent police-civilian interactions) records the gender of police officers employed by a city.

Our data on crime and policing outcomes are drawn from the FBI Uniform Crime Report (UCR) data (70, 71). UCR data are compiled by the FBI based on reports from law enforcement agencies and provide annual agency-level counts of reported crime offenses, clearances, and arrests for a variety of offense types. As we study mayors, we restrict the law enforcement agencies to city police departments. Throughout most of our analysis, we normalize variables that are in levels (rather than proportions) by the city’s population.

We report data on crimes overall and in several categories: total “index crimes,” which are the eight offenses used by the FBI to produce crime indices (murder, rape, robbery, aggravated assault, burglary, larceny, motor vehicle theft, and arson), violent index crimes, and property index crimes. The latter two are subsets of the eight index offense types. Likewise, we report results on arrests overall and by category, including property crime, violent crime, drug crime, and “other” crime. The other category of arrests includes offense types for which enforcement may be particularly subject to law enforcement officer discretion: e.g., vagrancy, loitering, or drunkenness, to name a few.

Last, we leverage data on the racial composition of arrests in the UCR data. We estimate the Black share of arrests based on the number of arrestees coded as Black in the data in a given city-year divided by the total number of arrests in that city-year. We construct this both for overall arrests and each of the categories noted above. We also construct the ratio of Black-to-white arrests as an alternative measure (results shown in Supplementary Text F.1). Note that the arrest data are disaggregated by the age, race, and sex of the arrestee—but not by ethnicity (Hispanic versus non-Hispanic).

Tables 1 and 2 show descriptive statistics for our main dependent variables that we examine in the remainder of the paper, among the entire sample of city-years since 1990 (in the first column), the city-years when under Democratic mayoral control (second column), Republican mayoral control (third column), or unknown mayoral party control (fourth column). While the majority of mayors in our sample are Democrats, there are also many city-years with Republican mayors.

Table 2. Summary statistics for main outcome variables, part 2.

	Total	Democratic	Republican	Party unknown
	(N = 12,436)	(N = 5757)	(N = 4143)	(N = 2536)
All crimes per 100 capita
Mean (SD)	6.60 (3.26)	6.93 (3.38)	5.73 (2.86)	7.30 (3.27)
Median [min, max]	6.08 [0.000629, 21.5]	6.39 [0.0129, 21.5]	5.17 [0.000629, 20.3]	6.97 [0.0746, 21.1]
Missing	680 (5.5%)	264 (4.6%)	226 (5.5%)	190 (7.5%)
Violent crimes per 100 capita
Mean (SD)	0.670 (0.515)	0.783 (0.558)	0.508 (0.382)	0.678 (0.531)
Median [min, max]	0.538 [0, 4.35]	0.651 [0, 4.09]	0.424 [0, 4.35]	0.531 [0, 3.88]
Missing	683 (5.5%)	265 (4.6%)	226 (5.5%)	192 (7.6%)
Property crimes per 100 capita
Mean (SD)	4.70 (2.37)	4.79 (2.42)	4.12 (2.11)	5.42 (2.41)
Median [min, max]	4.29 [0, 16.8]	4.41 [0.00515, 16.8]	3.68 [0, 15.1]	5.12 [0.0707, 16.1]
Missing	680 (5.5%)	264 (4.6%)	226 (5.5%)	190 (7.5%)
Clearance rate
Mean (SD)	0.265 (0.110)	0.249 (0.110)	0.277 (0.110)	0.281 (0.105)
Median [min, max]	0.265 [0, 1.00]	0.248 [0, 0.630]	0.279 [0, 1.00]	0.282 [0, 0.768]
Missing	684 (5.5%)	268 (4.7%)	226 (5.5%)	190 (7.5%)
Total arrests per 100 capita
Mean (SD)	3.15 (1.77)	3.21 (1.89)	2.78 (1.42)	3.61 (1.88)
Median [min, max]	2.81 [0.0279, 19.4]	2.83 [0.0865, 15.9]	2.49 [0.0279, 19.4]	3.30 [0.257, 14.9]
Missing	2749 (22.1%)	1333 (23.2%)	847 (20.4%)	569 (22.4%)
Violent crime arrests per 100 capita
Mean (SD)	0.815 (0.515)	0.904 (0.581)	0.673 (0.391)	0.852 (0.489)
Median [min, max]	0.692 [0, 8.15]	0.753 [0.0601, 6.67]	0.609 [0, 8.15]	0.758 [0.00184, 3.74]
Missing	2749 (22.1%)	1333 (23.2%)	847 (20.4%)	569 (22.4%)
Property crime arrests per 100 capita
Mean (SD)	0.476 (0.424)	0.473 (0.403)	0.401 (0.332)	0.609 (0.554)
Median [min, max]	0.384 [0, 8.39]	0.381 [0.00844, 6.27]	0.331 [0.00559, 5.35]	0.498 [0, 8.39]
Missing	2749 (22.1%)	1333 (23.2%)	847 (20.4%)	569 (22.4%)
Drug crime arrests per 100 capita
Mean (SD)	0.593 (0.439)	0.653 (0.520)	0.544 (0.336)	0.539 (0.369)
Median [min, max]	0.496 [0, 4.89]	0.530 [0, 4.89]	0.476 [0, 2.54]	0.465 [0, 4.47]
Missing	2749 (22.1%)	1333 (23.2%)	847 (20.4%)	569 (22.4%)
Other crime arrests per 100 capita
Mean (SD)	1.26 (0.944)	1.18 (0.930)	1.17 (0.795)	1.61 (1.11)
Median [min, max]	1.04 [0.00401, 9.79]	0.953 [0.00890, 7.67]	0.971 [0.00401, 5.61]	1.35 [0.00705, 9.79]
Missing	2749 (22.1%)	1333 (23.2%)	847 (20.4%)	569 (22.4%)
Mayoral party in power
Democratic	5757 (46.3%)	5757 (100%)	0 (0%)	0 (0%)
Republican	4143 (33.3%)	0 (0%)	4143 (100%)	0 (0%)
PID unknown	2536 (20.4%)	0 (0%)	0 (0%)	2536 (100%)

Research designs

Our first causal inference technique by which we examine the effect of electing mayors of different parties is a RD design. The RDD is a strategy that has been widely used to estimate the causal effects of elected official identity on political and policy outcomes (24, 31, 37, 37, 72–75).

We view it as providing the clearest causal identification of the impact of mayoral partisanship on criminal justice outcomes. The RD design’s main limitation is relatively weak statistical power (52). As a result, we augment it with several difference-in-difference designs that examine the effect of switches in mayoral partisanship.

Our RD design exploits the fact that a sharp electoral threshold, 50% of the two-party vote share, determines which party wins mayoral elections. Cities where the mayoral election was won by a Democrat over a Republican (or vice versa) by a very narrow margin are likely to be similar to one another on a host of characteristics other than mayoral partisanship. With some assumptions, this allows us to detect the causal effect of partisanship while avoiding potential confounding from these other characteristics. The RD method therefore focuses on differences in outcomes in very close elections. In practice, the effect of electing a Democratic mayor rather than a Republican mayor is identified by restricting the sample to elections within a bandwidth around the 50% threshold in the Democrats’ vote share and estimating the “jump” in outcome variables at the threshold—or the elections closest to a tie. This design identifies a local average treatment effect at the threshold of 50% vote share. Following (76), we use a local polynomial nonparametric regression; we outline the estimator in more detail in Supplementary Text B. We implement this approach using the rdrobust package in R (77) which selects an optimal bandwidth to minimize mean squared error in the estimate and adjusts confidence intervals to account for the remaining bias from the bandwidth selection procedure.

This approach might raise concerns about the applicability of the estimates from this design to cities where there are not close mayoral elections. This concern is at least partially assuaged by the fact that, of medium and large cities more than 75,000 in population in our data, our election data cover 99% of the population in this target universe, and 89% of those cities in our election data had a close election at some point and are therefore included in our RD analyses. The coverage of our data and the subsample of cities with close elections is further described in Supplementary Text A.

The validity of the RD design depends on the assumption that only the party of the winning candidate—and not the distribution of units’ potential outcomes—changes discontinuously at the threshold (78, 79). This is often called the “continuity assumption” and involves assuming that there is no other endogenous cause of changes in outcomes that occur at the same threshold that triggers the change in treatment status—in our case, changes in partisanship of the winner (80). Results from tests in Supplementary Text B document that this assumption is likely satisfied in our setting. Consistent with a recent large-scale validation of electoral RD design studies (81), we also observe no significant discontinuities in lagged values of the running variables or outcome variables. Supplementary Text C shows these placebo results.

To increase statistical efficiency, we estimate all RDD treatment effects on changes in outcomes rather than on levels (79). Our main analyses focus on the difference between crime and policing outcomes in the election year and 3 years after the election to account for the lag in time between a politician taking office and their ability to influence policy outcomes.

The RDD—and the other designs we use in this paper—all measure the effects of the compound treatment of partisanship alongside other characteristics that politicians have in tandem with their partisanship. This technique, however, might also have the downside of potential bias due to compensating differentials (82) arising from, for instance, differences in candidate competence that occur in close elections. While the RDD does not enable us to measure the effects of partisanship disentangled from these other characteristics, we are primarily interested in the real-world effects of this bundled party treatment and not the “pure” effect of party on policy outcomes.

Second, we use two recently developed generalized difference-in-differences strategies to assess the effect of changes in mayoral partisanship on crime and policing outcomes. A basic difference-in-differences model in our setting might take the form

$$Y_{\mathit{ct}}=\mathrm{\beta }{\text{Dem}}_{\mathit{ct}}+{\mathrm{\tau }}_t+{\mathrm{\gamma }}_c+{\mathrm{ϵ}}_{\mathit{ct}}$$ (1)

where $Y_{\mathit{ct}}$ represents our outcome (e.g., per capita numbers of arrests) and ${\text{Dem}}_{\mathit{ct}}$ is an indicator variable equal to one if a city $c$ has a Democratic mayor in year $t$. Also included are city and year fixed effects (${\mathrm{\gamma }}_c$ and ${\mathrm{\tau }}_t$). This basic approach fails to account for recently highlighted biases stemming from two-way fixed effects difference-in-differences models when treatment occurs at different points in time (83), which is very much the case in our setting.

Instead, we estimate nonparametric difference-in-differences models using the PanelMatch method (84), which compares units with similar treatment histories (i.e., party control) and similar pretreatment outcomes (e.g., crimes or arrests) that are “treated” with a Democrat taking control of the mayoral office versus those that are not treated (i.e., a Republican or nonpartisan mayor takes control). Specifically, we match using Mahalanobis distance on lagged outcomes in the 3 years before treatment. This design avoids the negative weights problem in traditional two-way fixed effects methods (84). It identifies an average treatment effect on the treated (ATT) units across cities and time periods in cities that switch mayoral parties. We also conduct placebo tests for this method by looking at the effects of changes in mayoral partisanship on pretreatment outcomes in Supplementary Text C. Note that we find some non-null placebos, so readers should bear that in mind when evaluating our main results using PanelMatch. However, our main results using PanelMatch are mostly null which could ameliorate concerns of confounding. The downside of PanelMatch’s algorithm is that matching treatment and control units reduces our effective sample size and the corresponding statistical power.

To build on these analyses, we also estimate the effects of changes in partisanship within cities over time using the counter-factual fixed effects (FEct) models developed in (85) via the fect package in R. This avoids the negative-weight problem in traditional two-way fixed effects methods and corrects biases induced by treatment effect heterogeneity by not using the treated observations at the modeling stage and by imposing uniform weights on individualistic treatment effects on treated observations (85). This method is also based on a difference-in-differences approach and allows us to accommodate switches both from Republican mayors to Democratic mayors and the reverse. We examine switches from non-Democratic to Democratic control, but results examining switches to Republican control are similar and presented in Supplementary Text J. This design identifies ATT cities that switch mayoral parties. The FEct approach also enables us to demonstrate the absence of pretreatment placebo effects, which we show in Supplementary Text C, and to visualize the dynamic trajectory of the posttreatment impact of partisanship.

While each research design has its own limitations, the three designs together present holistic evidence about the effects of politics on crime and policing. While we have detailed the biases of each method in this context, we also find similar effects using a more traditional two-way fixed effects design.

Supplementary Materials

This PDF file includes:

Supplementary Text

Tables S1 to S28

Figs. S1 to S52

References