Interpreting Police Union Donations

We study the consequences of electing a mayor who is backed by a police union, defined as a candidate who receives a campaign contribution from a police union. Throughout the article, we treat police union donations as a strong signal of organizational support—functionally similar to an endorsement—given that unions donate to only one of the top two candidates in 93% of races. At the same time, we emphasize the analytical value of focusing on donations rather than formal endorsements. Donations are typically less visible to the broader electorate and may operate as what we term an inward-facing endorsement: a signal directed toward the candidate and the police workforce rather than toward voters. This form of support may be especially consequential in contests where Democratic candidates—whose electoral coalitions may be skeptical of police unions—quietly accept union contributions while avoiding the potential backlash associated with a public endorsement. This dynamic aligns with findings from Gaudette (Reference Gaudette2025), who shows that voters often interpret police union endorsements as signaling a Republican affiliation, likely due to the partisan composition of police forces themselves (Ba et al. Reference Ba, Ge, Kaplan, Knox, Komisarchik, Lanzalotto and Mariman2025; Goncalves and Tuttle Reference Goncalves and Tuttle2024).

Police unions pursue two distinct objectives through their political engagement. The first is influence-seeking: donations intended to cultivate access to, or leverage over, the eventual officeholder, regardless of prior policy alignment. The second is preference alignment: support for candidates who are ex ante more closely aligned with the union’s policy priorities, with the aim of increasing their electoral success. As in much of the literature on interest-group political activity—where donations are taken as given and identification derives from which candidate ultimately wins rather than from exogenous variation in contributions—our treatment should be understood as a compound one that may reflect both channels (Fowler, Garro, and Spenkuch Reference Fowler, Garro and Spenkuch2020; Gulzar, Ladino, and Riaño Reference Gulzar, Ladino and Riaño2025; Harding et al. Reference Harding, Prem, Ruiz and Vargas2024; Ruiz Reference Ruiz2017).

We attempt to empirically distinguish between these mechanisms by examining donation patterns that are more consistent with influence-seeking and those more indicative of preference alignment. This distinction is consequential because the two channels yield different empirical interpretations and policy implications. When interpreted through an influence-seeking lens, our estimates speak directly to campaign finance policy, shedding light on how outcomes might differ had the mayor not received police union donations. Interpreted through a preference-alignment lens, our results instead capture the consequences of electing a candidate whose policy positions already align with police union preferences. Both interpretations are substantively important, but they imply distinct mechanisms through which police unions may shape local governance and, in turn, different policy implications.

Police Spending

A natural starting point for theorizing about police budgets is to consider the institutional interests of police departments. All else equal, police prefer higher levels of public spending on policing, whether through increased compensation or expanded staffing. Realizing these preferences requires bargaining with municipal governments. To enhance their leverage in this bargaining process, police can organize collectively, most commonly through unions (Anzia and Moe Reference Anzia and Moe2015). Indeed, protective service workers—among them police officers—rank among the most highly unionized occupational groups in the United States (Bureau of Labor Statistics 2025), with police unionization rates around 80% in some states (Goncalves Reference Goncalves2021). While unionization facilitates direct negotiation over compensation and working conditions, police unions can also exert influence through the political arena (Anzia Reference Anzia2022b; DeLord and York Reference DeLord and York2017; Kaplan and Naidu Reference Kaplan and Naidu2025). In particular, they can shape the electoral environment in ways that enhance their ability to secure favorable policy outcomes. As described in the previous section, two complementary strategies are available. Under influence-seeking, unions provide endorsements or campaign contributions to cultivate access to candidates who, once in office, may deliver increased police funding regardless of their prior positions over police funding. Under preference alignment, unions strategically support candidates who are ex ante more closely aligned with their preferences for greater police resources.

What should we expect when a police-union-backed candidate is elected? The theoretical logic is straightforward: police unions prefer higher spending on law enforcement, while local governments face binding budget constraints. The election of a union-backed mayor increases the likelihood that the preferences of the police will be represented more directly in the budgeting process. As a result, we expect total police spending to increase following the election of a candidate supported by the police union.

Disaggregating total police spending into its two main components—compensation and employment—leads to less clear-cut expectations. On the one hand, police officers have a clear interest in higher compensation, and unions are generally responsive to these demands. This responsiveness is reinforced by the competitive landscape of police union representation: unlike more centralized labor organizations, local police unions are fragmented across affiliations, such as AFSCME, SEIU, the Fraternal Order of Police, and the International Union of Police Associations (Kearney and Mareschal Reference Kearney and Mareschal2017). On the other hand, unions also have an institutional incentive to expand police employment. A larger force increases potential membership, thereby enhancing the union’s political and organizational power. Ideally, unions would pursue both higher wages and greater employment, but they must operate within the fiscal limits imposed by municipal budgets. As Anzia and Moe (Reference Anzia and Moe2015) point out, unions often confront a trade-off between maximizing benefits for current members and expanding their organizational influence. The extent to which they prioritize one goal over the other is ultimately an empirical question. However, because compensation is typically renegotiated only at discrete intervals, tied to the expiration of collective bargaining agreements, we expect the more immediate and flexible margin of adjustment to be employment. Thus, we anticipate stronger and more observable effects on police staffing levels than on compensation in the short run.

Police unions, like other public sector unions, prioritize personnel policies as their core area of influence (DiSalvo Reference DiSalvo2022). While this includes wages, a substantial and often underappreciated component involves fringe benefits—such as health insurance, pensions, and other post-employment compensation. These forms of compensation are less visible to voters than wages and thus attract less political scrutiny (Anzia and Moe Reference Anzia and Moe2015; Hunter and Rankin Reference Hunter and Rankin1988). Public sector unions are therefore incentivized to emphasize benefits over wages in bargaining, exploiting what Bawn et al. (Reference Bawn, Cohen, Karol, Masket, Noel and Zaller2012) term an “electoral blind spot.” This logic suggests that the most pronounced policy effects of electing union-backed candidates are likely to emerge in the domain of benefits. Although our data do not allow us to directly test this proposition, we will present suggestive evidence consistent with this expectation.

Police Performance

To theorize the potential impact of a police-union-backed mayor on police performance, we draw on the literature on ideological alignment between politicians and bureaucrats (here, the mayor as the political principal and the police as the bureaucratic agents). This literature emphasizes that alignment can shape bureaucratic behavior, but it does not deliver unambiguous predictions about performance: the direction of the effect depends on the mechanisms through which alignment operates.

Police performance depends on two core inputs. First, performance is positively related to the resources available to the police department, such as personnel, equipment, and funding. Second, performance depends on the effort exerted by police officers in carrying out their duties. The election of a police-union-backed mayor can therefore affect police performance through its impact on police resources (a resource effect) and on officer effort (an effort effect).

If a union-backed mayor allocates greater resources to the police department, performance may improve through a resource effect: holding officer effort constant, increased funding and staffing can enhance operational capacity and effectiveness. This channel generates a positive prediction for police performance. An alternative version of the resource effect, however, arises when police unions seek not additional resources per se but specific rules governing their allocation. This may blunt the performance gains associated with higher spending. A prominent example is the negotiation of seniority-based assignment protections, which give more experienced officers priority over desirable deployments and shifts, while concentrating newer officers in the most challenging neighborhoods or assignments (Ba et al. Reference Ba, Bayer, Rim, Rivera and Sidibé2021; DiSalvo Reference DiSalvo2020; Rushin Reference Rushin2017). Under this alternative version of the resource effect, increased resources may fail to translate into improved performance.

The effect of a union-backed mayor on officer effort, however, is theoretically ambiguous. The effort effect may be positive or negative, depending on the underlying mechanism. We highlight three mechanisms that yield distinct predictions.

First, police unions may use political support—such as campaign donations or endorsements—as a substitute for job performance. Under this mechanism, the election of a union-backed mayor weakens officers’ incentives to exert effort, leading to lower performance. If favorable outcomes (e.g., higher compensation or expanded staffing) can be secured through political loyalty rather than demonstrated effectiveness, officers have less reason to signal their value through performance. This logic echoes political science models of pay-to-play and interest-group capture, in which material political support is exchanged for favorable treatment even in the absence of strong performance. While no existing study directly examines this dynamic for bureaucrats, a growing empirical literature on non-bureaucratic public service providers shows that political donations can reduce accountability pressures, making performance less central as a source of leverage (e.g., Colonnelli, Prem, and Teso Reference Colonnelli, Prem and Teso2020; Fazekas, Ferrali, and Wachs Reference Fazekas, Ferrali and Wachs2023; Gulzar, Ladino, and Riaño Reference Gulzar, Ladino and Riaño2025; Harding et al. Reference Harding, Prem, Ruiz and Vargas2024; McCarty and Sumner Reference McCarty and Sumner2024; Ruiz Reference Ruiz2017). Related evidence in Foy (Reference Foy2024) shows that teachers perform better following union decertification, suggesting that union power can dampen individual effort incentives. Like campaign contributions, unionization can strengthen bureaucrats’ bargaining power vis-à-vis elected officials, potentially substituting political influence for performance. This mechanism is also consistent with theories of clientelism, which emphasize that material political support, such as endorsements or donations, can serve as a substitute for programmatic effectiveness in winning political favor (Hicken Reference Hicken2011; Kitschelt and Wilkinson Reference Kitschelt and Wilkinson2007).

At the same time, two alternative mechanisms predict an increase in officer effort following the election of a union-backed mayor. First, ideological alignment may generate a morale effect: officers may exert greater effort when governed by a mayor who shares their values or vision for policing (Besley and Ghatak Reference Besley and Ghatak2005; Spenkuch, Teso, and Guo Reference Spenkuch, Teso and Xu2023). Second, police may engage in strategic effort, increasing performance to bolster the electoral prospects of a politically aligned mayor. As Moe (Reference Moe2012) observes, this channel is largely absent from formal models of bureaucratic behavior, likely because bureaucratic performance is often not directly observable to voters. Police departments constitute an important exception. Police officers are the largest group of municipal employees (Bee and Moulton Reference Bee and Moulton2015), and their highly visible, street-level activities make their performance electorally salient. Emerging empirical evidence supports the plausibility of this channel; for example, Wirsching (Reference Wirsching2026) documents pervasive strategic shirking and heightened effort within the NYPD aimed at diminishing or improving the electoral prospects of city councilors according to their support for police budgets.

In sum, the effect of electing a police-union-financed mayor on police performance depends on how such an election shapes both the resources allocated to the police department and the effort exerted by police personnel. The net effect is therefore an empirical question, determined by the relative strength of these competing mechanisms in the institutional and political context we study.

Local Campaign Finance Data

A persistent challenge in the study of campaign finance in U.S. local elections has been the lack of comprehensive and standardized data on political donations. A central contribution of this article is the construction of a new dataset covering a large number of mayoral elections across the United States between 2001 and 2019. We focus on elections in which at least one of the top two candidates—the winner or the runner-up—received campaign contributions. For these elections, we collect information on donations received by the two leading candidates in the two years preceding each election.Footnote ⁸

Our dataset draws on two primary sources. First, we compile data from campaign finance databases maintained by seven states: Connecticut, Iowa, Illinois, Oregon, South Carolina, Virginia, and Washington. These databases include detailed records of monetary contributions, in-kind contributions, and loans made to political committees supporting candidates in various local elections. We match political committees to the candidates they supported and extract donation data for the top two candidates in mayoral contests. The “Data” section of the Supplementary Material provides detailed information on the data collection procedures and sources for each of these seven states. Second, we incorporate data from Yu (Reference Yu2022), which contains hand-collected records of mayoral campaign contributions in 36 additional states. This dataset was assembled from original campaign disclosure reports housed in local government offices, municipal archives, and boards of elections. By combining these two sources, we construct a comprehensive dataset covering 2,303 mayoral elections in 1,073 cities.

For the purpose of this article, we exclude 34 cities that do not have municipal police departments. Our final sample comprises 2,249 elections across 1,039 cities in 43 states, involving 3,251 unique candidates and 709,687 individual donations.

Table 1 provides a breakdown of the dataset by state. An important strength of our data is its broad coverage across municipalities of varying sizes. The smallest jurisdiction in the sample is Pawleys Island, South Carolina, with a population of 106 as of the 2017 election; the largest is New York City. Figure 1 illustrates the distribution of covered elections by population size, highlighting the dataset’s capacity to capture electoral dynamics in both small and large urban contexts.

Table 1.

Summary of the Contributions Data for Each State

Table 1. Long description

The table contains 9 columns: State, Elections, Cities, Candidates, Donations, Earliest year, Latest year, Minimum population, and Maximum population.

Key data points include:

* C A: 151 elections, 89 cities, 259 candidates, 157,039 donations, spanning 2001 to 2018, with populations from 5,133 to 3,975,788.

* I L: 752 elections, 262 cities, 934 candidates, 103,549 donations, spanning 2001 to 2019, with populations from 562 to 2,819,117.

* N Y: 50 elections, 25 cities, 82 candidates, 46,851 donations, spanning 2007 to 2017, with populations from 6,732 to 8,396,091.

* S C: 307 elections, 147 cities, 416 candidates, 29,027 donations, spanning 2007 to 2019, with the smallest minimum population of 106.

* W A: 141 elections, 57 cities, 215 candidates, 64,951 donations, spanning 2001 to 2019.

Other states listed with smaller datasets include A K, A R, A Z, C O, C T, F L, G A, I A, I D, I N, K S, K Y, L A, M A, M D, M E, M I, M N, M O, M S, N C, N D, N E, N H, N J, N M, N V, O H, O K, O R, P A, R I, S D, T N, T X, U T, V A, and W I.

The final row for All states totals 2,249 elections, 1,039 cities, 3,251 candidates, and 709,687 donations across the entire period from 2001 to 2019.

Note: Number of elections, cities, candidates, and donations covered by our local campaign finance dataset, together with the earliest and latest election year and minimum and maximum population for each state.

Figure 1.

Number of Elections in the Contributions Data, by Population

Note: The figure shows the number of elections covered by our local campaign finance dataset in each of the population brackets on the $ x $ -axis. The population of a city is measured in the year of the election.

For each election, we observe the candidate’s name, the election date, the name of each contributor (whether an individual or an organization), and the amount donated. As described in the following sections, we further enrich this dataset by merging it with additional sources of information.

Identifying Donations by Police Unions

To identify campaign contributions made by police unions, we implement a two-step process combining automated keyword detection with manual verification.

First, we apply a string-matching algorithm to the names of contributors recorded in the campaign finance data. This automated step uses a set of predefined keywords—“Police,” “Detectives,” “Sergeants,” “Lieutenants,” and “Captains”—to flag contributions plausibly originating from police-affiliated organizations. The goal of this step is to capture a broad set of potential police union donors while minimizing false negatives. We then manually review the flagged entries to remove contributions erroneously attributed to police unions.

Second, we conduct a comprehensive manual review of all remaining contributor names to identify additional police union donations potentially missed by the keyword filter. This step allows us to account for naming irregularities, acronyms, and idiosyncratic spellings that may obscure police union affiliations.Footnote ⁹ To guide this process, we conduct targeted searches of known police unions across the cities and states represented in our dataset, which informed additional keywords and the identification of acronyms.

By combining systematic detection with human oversight informed by local context and naming conventions, this two-step approach reduces both Type I and Type II errors in classification. As a result, the final set of identified police union donations achieves a high degree of accuracy and reliability, which is critical for the validity of our empirical analysis.

Elections, Candidates’ and Institutional Characteristics

We augment the campaign finance data with information on election outcomes and candidate characteristics. Specifically, for each election, we collect data on: (i) the vote share and total number of votes received by the top two candidates, (ii) the candidate’s partisan affiliation, (iii) the incumbency status of each candidate, and (iv) municipal characteristics.

Election results are assembled from a combination of sources. We first incorporate data from Ferreira and Gyourko (Reference Ferreira and Gyourko2009), Jerch, Kahn, and Li (Reference Jerch, Kahn and Li2017), and de Benedictis-Kessner et al. (Reference de Benedictis-Kessner, Da In Lee, Velez and Warshaw2023). For cities not included in these datasets, we collect results directly from county and city election offices, often via official websites. When these are unavailable, we consult newspaper archives to retrieve historical election outcomes.Footnote ¹⁰

Identifying candidates’ partisan affiliations presents a particular challenge in nonpartisan elections, where formal party labels do not appear on the ballot. In order to identify the partisan leaning of the candidates in these elections, we rely on a combination of sources, including local newspaper coverage, candidate websites, and official documents. We also examine whether candidates had previously contested partisan elections or received formal party endorsements within a proximate time window. Full details on this coding process are provided in the “Data” section of the Supplementary Material.

We classify approximately 26% of the candidates in our sample as Democrats, 20% as Republicans, 12% as independents, and 9% as affiliated with other parties. The remaining 33% could not be reliably classified due to insufficient public information on their partisan alignment.

Finally, we obtain information on the form of government from the International City and County Management Association’s (ICMA) Municipal Form of Government Surveys (ICMA 2018). Since not all cities are surveyed in every wave, we assign to each election the form of government recorded in the most recent survey year in which the city was included.Footnote ¹¹

Municipal Budgets, Municipal Public Employment, and Crime Outcomes

We further enrich our dataset with city-year-level variables measuring the size and compensation of police departments, and public safety outcomes.

First, we obtain data on police department staffing from the Law Enforcement Officers Killed and Assaulted (LEOKA) dataset. This source provides annual figures on the number of personnel employed by each law enforcement agency, disaggregated into sworn officers and civilian employees.Footnote ¹²

Second, we use data from the FBI’s Uniform Crime Reporting (UCR) Program to measure crime incidence and clearance rates. Specifically, we calculate the number of reported index crimes per capita—separately for violent and property crimes—as well as the share of these offenses cleared by the police. In addition, we include the number of arrests for non-index crimes, also expressed as a share of the municipality’s population.Footnote ¹³ Data from the Uniform Crime Reports (UCR) are widely used in studies of crime in U.S. cities. A well-known limitation, however, is that these data are self-reported by local police agencies to the FBI. As Cook and Fortunato (Reference Cook and Fortunato2023) show, reporting error in the UCR is systematic rather than random and varies with state legislative capacity. Our empirical strategy mitigates this concern by relying on within-state comparisons, which absorb cross-state differences in reporting institutions. Nevertheless, we remain cautious in interpreting estimates that rely on UCR outcomes, given the potential for residual measurement error.Footnote ¹⁴

Third, we draw on the U.S. Census Bureau’s Annual Survey of Public Employment and Payroll (ASPEP) (U.S. Census Bureau 2025) to capture compensation levels in police departments. The ASPEP provides data on the total number of police personnel and the aggregate gross monthly wage bill, with separate figures for sworn officers and civilian employees. The wage bill includes all gross salaries, wages, fees, commissions, and overtime paid to employees, as well as incentive payments that are paid at regular pay intervals. It excludes the employer share of fringe benefits. We compute the average wage by dividing total payroll by the number of employees within each category. Due to the survey’s structure, however, wage data are only available for most cities at five-year intervals, resulting in lower temporal resolution relative to the other sources.Footnote ¹⁵

Two-Way Fixed Effects Estimation

A central concern in estimating the causal effect of police-union-backed mayors is endogeneity: cities where police unions support winning mayoral candidates may differ systematically from other cities in ways that are correlated with policing outcomes. For example, cities with stronger or more politically active police unions may also exhibit larger police forces or distinct crime trends, regardless of who is elected mayor.

To address these concerns, we exploit within-city variation across multiple mayoral election cycles. Many cities in our dataset are observed over time, and some experience changes in whether the winning mayor received police union donations. This allows us to examine changes in outcomes within the same city, before and after the election of a police-backed mayor. We estimate the following equation:

⁽¹⁾ $$ \begin{array}{rl}{y}_{ct}={\alpha}_c+{\gamma}_{ts(c)}+{\delta}_{\tau s(c)}+\beta PoliceMayo{r}_{ct}+{X}_{ct}\theta +{\epsilon}_{ct},& \end{array} $$

where $ {y}_{ct} $ is the outcome of interest for city $ c $ , measured in year $ t $ . The key explanatory variable, $ PoliceMayo{r}_{ct} $ , is an indicator equal to one if the mayor in office received campaign contributions from a police union in the most recent election.Footnote ²³ We include city fixed effects ( $ {\alpha}_c $ ) to absorb any time-invariant city characteristic, and State-by-calendar-year fixed effects ( $ {\gamma}_{ts(c)} $ ), which account for time-varying shocks common to all cities within the same state. In addition, we include State-by-term-year fixed effects ( $ {\delta}_{\tau s(c)} $ ), which control for the fact that cities hold elections in different years and may be at different points in their mayoral term cycles. Finally, $ {X}_{ct} $ includes controls for mayoral partisanship and the mayor’s margin of victory, as both characteristics are associated with the likelihood of police union support, as shown above.Footnote ²⁴ Standard errors are clustered at the city level to account for serial correlation over time within the same city.

Stacked Event Study Design

As emphasized by a recent literature on difference-in-differences with staggered adoption (e.g., Borusyak, Jaravel, and Spiess Reference Borusyak, Jaravel and Spiess2024; Cengiz et al. Reference Cengiz, Dube, Lindner and Zipperer2019; De Chaisemartin and d’Haultfoeuille Reference De Chaisemartin and d’Haultfoeuille2020; Goodman-Bacon Reference Goodman-Bacon2021), the standard two-way fixed effects estimator may yield biased estimates when treatment timing varies across units and treatment effects are heterogeneous. Specifically, part of the identifying variation comes from comparisons between “already-treated” and “not-yet-treated” units, which can introduce bias if treatment effects evolve over time.

To address this issue, we implement a version of the stacked event study design proposed by Cengiz et al. (Reference Cengiz, Dube, Lindner and Zipperer2019). We identify as treated units those cities that switch from having a mayor not financed by police unions in election cycle $ \tau -1 $ to one financed by police unions in cycle $ \tau $ . For each treated city $ k $ , we construct a control group consisting of cities in the same state whose mayors never received police union donations during the same period.Footnote ²⁵ Each treated–control group pair defines an “event,” indexed by $ k $ , and we stack all such events into a common dataset. We then compare pre–post changes in outcomes in each treated city $ k $ to pre–post changes in outcomes, over the same time window, in the cities that serve as its control group. Specifically, this amounts to estimating the following equation:

(2) $$ \begin{array}{rl}{y}_{ctk}={\alpha}_{ck}+{\gamma}_{tk}+\beta Treate{d}_{ck}\times Pos{t}_{tk}+{\epsilon}_{ctk},& \end{array} $$

where $ {y}_{ctk} $ is the outcome of interest for city $ c $ , in year $ t $ , within event $ k $ . City-event fixed effects, $ {\alpha}_{ck} $ , absorb time-invariant differences between treated and control cities, and year-event fixed effects, $ {\gamma}_{tk} $ , absorb time-specific shocks common to all cities in the event. The indicator variable $ Treate{d}_{ck} $ indicates treated cities, and the indicator $ Pos{t}_{tk} $ equals one in the years following the treatment (i.e., in the years in which the mayor supported by a police union is in office). The coefficient $ \beta $ captures the average treatment effect of electing a police-union-backed mayor. Standard errors are clustered at the city level.

We also estimate an augmented version of Equation 2 that includes controls for mayoral partisanship and margin of victory, each interacted with the post-treatment indicator.

To trace the dynamic evolution of the treatment effect and assess potential differential pre-trends between treated and control cities, we also estimate a nonparametric event study specification:

(3) $$ \begin{array}{rl}{y}_{ctk}={\alpha}_{ck}+{\gamma}_{tk}+{\displaystyle \sum_{s=-3}^4}{\beta}_sTreate{d}_{ck}\times {D}^{sk}+{\epsilon}_{ctk},& \end{array} $$

where $ {D}^{sk} $ is a set of relative event-time dummies (ranging from three years before to four years after the election, with the year of the election as the omitted category). The coefficients $ {\beta}_s $ capture the differential evolution of the outcome in treated versus control cities in the years before and after treatment.

The sample size varies by outcome due to differences in data availability across years and cities. For example, the sample includes 22 treated cities and 110 control cities for arrest rates (non-index crimes), and up to 41 treated cities and 270 control cities for police employment outcomes.