The Effects of Highly-Publicized Police Use-of-Force on Policing, Trust, and Crime: Evidence from Ferguson

(Very Preliminary. Please do not cite without permission.) Cheng Cheng† The University of Mississippi Wei Long† Tulane University

July 2017

Abstract Controversial police use-of-force has frequently become widely publicized due in part to the viral dissemination of information. These incidents, which often cause public outrage, civil unrest, and extensive scrutiny of the police, have potentially far-reaching effects on policing, trust, and crime. This paper examines the impact of a widely publicized 2014 fatal police shooting in Ferguson, Missouri, in which a white police officer shot and killed an unarmed black teenager. To identify effects, we exploit the shooting’s plausibly exogenous timing and its disproportionate impact on predominantly black communities using a novel panel dataset. Results indicate that the Ferguson shooting had instantaneous negative effects on arrests and self-initiated inspections. In the longer run, we find evidence that the shooting changed more policing behaviors (reducing police use-offorce, arrests, and self-initiated inspections, as well as improving 911 emergency call response time), increased crime reporting, and caused surging violent crimes.

Keywords: police use-of-force; policing; trust; crime

_________________ †Cheng

Cheng: Department of Economics, The University of Mississippi, University, Mississippi 38677 (Email: [email protected]). Wei Long: Department of Economics, Tulane University, New Orleans, Louisiana 70118 (Email: [email protected]).

1

1. Introduction Technology constantly changes the way information is spread. Recent technological improvements have fueled the dissemination of newsworthy events, with internet-based media reporting and social networks sharing instantaneously and extensively. Controversial police useof-force, in particular, has frequently become widely publicized due in part to the viral dissemination of information. Recent fatal police shootings of unarmed African American men in the U.S. – including Eric Garner in New York City (July 2014), Michael Brown in Ferguson, Missouri (August 2014), and Freddie Gray in Baltimore, Maryland (April 2015) – soon made national news headlines and drew international attention (Desmond, Papachristos, and Kirk 2016). These high-profile incidents, which often cause public outrage, civil unrest, and extensive scrutiny of the police, have potentially far-reaching effects on policing, community-police trust, and crime. For example, James Comey, the Federal Bureau of Investigation (FBI) Director, suspected that highly publicized police shootings led to less aggressive policing and contributed to an alarming spike in violent crimes (homicides in particular) in many large U.S. cities (Comey 2015, Lichtblau 2016). However, these exists little evidence. In this paper we present a detailed empirical analysis of the policing, trust, and crime effects of highly publicized police use-of-force, by exploiting the 2014 fatal shooting of Michael Brown in Ferguson, Missouri. This controversial incident immediately became the focus of world attention amid widespread public rage, extreme mainstream and social media scrutiny, and the Department of Justice (DOJ) investigations, which provides us with an ideal setting. Crucial to our analysis, we assemble a novel dataset on policing behaviors (police use-of-force, arrest, response to 911 emergency calls, and self-initiated activity), community trust in police (crime reporting through 911 calls), and crime (violent and property crimes) in St. Louis during 2013 –

2

2015. In examining the police shooting’s short-run effects, we utilize the plausibly exogenous time variation arising from the shooting that took the country by storm right away in August, 2014. Specifically, we compare outcome changes immediately before and after the shooting with a regression discontinuity (RD) design. Next, we explore the shooting’s medium- and long-run effects, by additionally exploiting its disproportionate impact on predominantly black communities using a difference-in-differences (DD) strategy. In doing so, we examine outcome changes over time in predominantly black communities, relative to similar changes in predominantly white communities. Importantly, we find strong graphical and statistical evidence that supports the DD strategy’s common trend assumption. The RD estimates indicate strong short-run response from policing officers, who made 40 percent fewer arrests and conducted 50 percent fewer self-initiated inspections immediately after the shooting. Meanwhile, we find no significant discontinuities in other policing practices, community-police trust, and crime at the time of shooting. In contrast, the DD estimates show that the shooting affected all three types of outcomes in the longer run. Frist, we find evidence that the shooting had broader effects on policing than in the short run. In addition to reducing arrests (17 percent) and self-initiated activities (22.8 percent), the shooting also led police use-of-force to drop by 42.7 percent in the medium run and improved response to 911 emergency calls by 6.3 percent. These estimates, combined with the short-run estimates, paint a clearer picture of how the shooting-induced pressure, scrutiny, and investigations affected policing behaviors through the opposing de-policing effect (causing officers to withdraw from proactive policing practices) and monitoring effect (incentivizing officers to raise work efforts in reactive policing practices). On the one hand, we can only estimate the net effect of the de-policing and monitoring effects on arrest and find that the former effect dominated. On the other hand, examining self-initiated

3

activities (proactive practices that are exclusively subject to the de-policing effect) and 911 call responses (reactive practices that are largely driven by the monitoring effect) allows us to separate these two effects. As expected, we find evidence of de-policing in proactive practices and improved performance in reactive practices: police officers pulled back from self-initiated activities in order to avoid controversial confrontations and became more active in responding to 911 calls under extensive monitoring. Additionally, we find that the declines in arrests and selfinitiated activities were larger for junior officers, which is consistent with the perception that junior officers receive less support than senior officers. Second, our results indicate that the shooting led to a (net) increase in community trust in police. Specifically, residents reported 10.8 percent more crimes after the shooting; the result stays robust after accounting for contemporaneous crime effects. This finding provides evidence that the police were regaining community trust, even though the shooting might have exacerbated the existing disbelief in police at the same time. Finally, we show that the shooting resulted in a surge in violent crimes. The estimated increases in homicides, robberies, and aggravated assaults are remarkable 63.2 percent, 16.9 percent, and 19.5 percent, respectively. This result is in line with the evidence of systematic withdrawal from proactive policing, which greatly reduced police presence and undermined crime prevention. Our estimates are robust to a number of sensitivity checks. First, our longer-run results are not sensitive to controlling for a rich set of demographic and socioeconomic factors. This implies that predominantly black and white communities were comparable in changes in other aspects, which is consistent with the DD identifying assumption. Second, the RD and DD estimates stay robust in various different specifications. For example, the RD estimates from linear, quadratic, and nonparametric specifications are almost identical. In addition, an alternative DD specification – in which we compare outcome changes among communities with high and low black population

4

shares – gives similarly significant policing, trust, and crime effects. Third, population-weighted estimates are very similar to the OLS estimates. This study contributes to several literatures. First, the main contribution is to inform the vigorous national debate about how highly publicized police use-of-force against unarmed African American civilians – which has become increasingly prevalent in recent years – affects policing, community relations, and public safety. Particularly, our analysis presents the one of the first empirical evidence of the so-called (and vaguely defined) “Ferguson effect” – rising violent crime rates following controversial and publicized police use-of-force against minority citizens. More importantly, we explore the policing and trust mechanisms that explain the post-shooting crime spikes. Second, our work complements a large literature on policing effects (Anwar and Fang 2006, Donohue III and Levitt 2001, Mas 2006), by providing a particularly detailed analysis of various policing practices. Specifically, this study joins a small but growing literature on how high-profile random shocks (e.g., excessive uses of police force, scandals, and budget cuts) change policing (DeAngelo and Hansen 2014, Heaton 2010, Shi 2009). Third, our findings contribute to the broad literature on the economics of crime (Becker 1968), by documenting how both police and criminals responded to perceived changes in incentives.

2. Background 2.1. The Highly Publicized Ferguson Shooting On August 8, 2014, an unarmed black teenager (Michael Brown) was shot and killed by a white police officer (Darren Wilson) in Ferguson, Missouri, a suburb of St. Louis.1 Thanks to social media, this fatal shooting – which prompted emotionally-charged and, in many cases,

1

See New York Times, for example, for a detailed account of the shooting: https://www.nytimes.com/interactive/2014/08/13/us/ferguson-missouri-town-under-siege-after-police-shooting.html. 5

violent protests starting from August 10 – soon became an international incident and drew extensive media coverage. 2 On August 14, President Barack Obama addressed the nation regarding the Ferguson shooting, by urging “peace and calm” and calling for an “open and transparent” investigation into Brown’s death (Welker and McClam 2014). On the same day, Governor Jay Nixon announced that the Missouri Highway Patrol would take over the security in Ferguson from the St. Louis County Police, which led to the first non-violent night of demonstrations in days (Peters 2014). On August 18, President Obama dispatched Attorney General Eric Holder to Ferguson to monitor the situation, making him the first high-level U.S. official to visit the St. Louis suburb after more than a week of public outrage at the fatal shooting (Kaplan 2014). On November 24, a grand jury decided not to indict officer Wilson for any crimes related to the death of Brown (NPR 2014). On March 4, 2015, seven months after the incident, the Department of Justice (DOJ) concluded that there was no evidence to support federal civilrights charges against Wilson (Department of Justice 2015b). Meanwhile, the DOJ’s other investigation uncovered a pattern and practice of misconducts by the Ferguson Police Department (FPD), which created deep distrust and resentment between the FPD and significant portions of Ferguson’s residents, especially African Americans (Department of Justice 2015c). Importantly, the DOJ further laid out enforceable recommendations and measures, in order to end the malpractice and restore the lost trust in police.3

2

As an example, the hashtag #Ferguson was used on Twitter 11.6 million with retweets and 1.9 million without retweets during August 9 – 25 (Grinberg 2014). 3 The recommendations include: changing policing and court practices so that they are based on public safety instead of revenue; improving training and oversight; changing practices to reduce bias, and; ending an overreliance on arrest warrants as a means of collecting fines. 6

2.2. Conceptual Framework In Figure 1, we provide a conceptual framework to illustrate how the publicized Ferguson police shooting affected policing, community trust in police, and crime. Specifically, we focus on how the shooting-induced enormous public outrage and criticism, intense mainstream and social media scrutiny, and attention-grabbing civil-rights investigations could change the behaviors of police officers, residents, and criminals. There are two opposing channels through which the Ferguson shooting could potentially affect policing behaviors. On the one hand, this publicized fatal shooting directly and significantly increased the expected cost of police use-of-force against black suspects, as police officers increasingly fear that such confrontations, legally justified or not, could easily be negatively stereotyped by the media and turned into the next high-profile and controversial incidents. This could lead to firings, lawsuits, and even indictments. As a result, one would expect less police use-of-force against offenders, especially in predominantly black communities. In addition, one would also expect police officers to pull back from proactive enforcement activities – the so-called “de-policing” effect (Rosenfeld 2016), which many argue has resulted in recent increases in violent crimes (homicides in particular) in large U.S. cities.4 On the other hand, it is also likely that the heightened external pressure and scrutiny of police practices following the fatal shooting created sufficient monitoring that prevented police officers from shirking their law enforcement responsibilities. Particularly, it is reasonable to expect that the monitoring effect is larger on the monitored reactive policing (responding to crime reporting) than on the discretionary proactive policing (preventing crime activities through self-initiating inspections).

4

Given these, we

For example, James Comey, the FBI Director, had this suspicion after talking with law enforcement, elected officials, academics, and community members throughout the country (Comey 2015). As he put it: “I don’t know whether this explains it entirely, but I do have a strong sense that some part of the explanation is a chill wind blowing through American law enforcement over the last year. And that wind is surely changing behavior.” 7

investigate the effects on four major dimensions of policing behaviors: use-of-force, arrest, responding to 911 emergency calls (largely reactive policing), and self-initiated activity (purely proactive policing). Next, we turn to community trust in police, which could also be affected by the shooting in two different ways. First, the Ferguson shooting could exacerbate the deep-rooted “legal cynicism” – the disbelief in the competence, legitimacy, and responsiveness of the criminal justice system (Kirk and Papachristos 2011) – that is thought to pervade many poor and minority communities (Desmond, Papachristos, and Kirk 2016). Consequently, the shooting would further alienate residents from the police in predominantly black communities. Meanwhile, the shootinginduced scrutiny and investigations would also pressure the police to rebuild the frayed relationship with black communities and restore the lost trust, in forms of, for example, engaging residents in conversations and holding public meetings (Associated Press 2014, 2015). 5 To measure which of these two potential impacts dominates, we measure effects on trust by examining crime reporting through 911 calls, which is considered more reliable in capturing citizens’ attitudes toward and interactions with the police than interviews or surveys (Baumer 2002, Desmond, Papachristos, and Kirk 2016, Rosenfeld, Jacobs, and Wright 2003). Finally, we discuss how the shooting could ultimately impact crime through the channels of policing and trust. The policing channel is straightforward and has been well documented: reactive and proactive policing reduces crimes through incapacitation (locking up criminals) (Berthelon and Kruger 2011, Jacob and Lefgren 2003, Levitt 1996) and deterrence (increasing police presence) (Di Tella and Schargrodsky 2004, Draca, Machin, and Witt 2011, Klick and

The highly publicized Ferguson shooting also prompted the launch of the Justice Department’s National Initiative for Building Community Trust and Justice in September, 2014, a nationwide program designed to enhance procedural justice, reduce bias, and combat distrust and hostility between law enforcement and the communities they serve (Department of justice 2014). 5

8

Tabarrok 2005). In comparison, the potentially negative trust-crime causal link is two-fold and less empirically explored in economics. First, lower community-police trust decreases the cost of committing crimes and hence results in more crimes. This is because the dampened trust discourages residents from reporting crimes to the police, of which rational criminals could easily take advantage. Second, the distrust in police could also escalate otherwise non-violent disputes or property crimes into violent crimes. For example, when people “do not trust the police to act on their behalf and to treat them fairly and with respect, they lose confidence in the formal apparatus of social control and become more likely to take into their own hands” informally and often violently (Rosenfeld 2016). However, given the aforementioned ambiguous net effects on policing and trust, the net effect of the Ferguson shooting on crime is similarly ambiguous ex ante. Combined together, the Ferguson shooting should not only affect policing, community trust in police, and crime, but also have larger impacts in predominantly black communities, where police officers, residents, and criminals are expected to respond more to the incident.

3. Empirical Strategy Following the conceptual framework, we identify the policing, trust, and crime effects of the Ferguson police shooting by comparing changes in these outcomes before and after the shooting in St. Louis. Specifically, we adopt different identification strategies to estimate the shooting’s effects in the short, medium, and long run. For the short-run effects, we use a sharp regression discontinuity (RD) design to compare outcome changes right before and right after the shooting, with time being the running variable. In doing so, we estimate the following model for the period 2013 – 2015: (1) 𝑂𝑢𝑡𝑐𝑜𝑚𝑒𝑖𝑤 = 𝛽0 + 𝛽1 𝑃𝑜𝑠𝑡𝑤 + 𝛽2 𝑓(𝑊𝑒𝑒𝑘𝑤 − 85) + 𝐗 𝑖𝑦 𝜸 + 𝑇𝑟𝑎𝑐𝑡𝑖 + 𝑀𝑜𝑛𝑡ℎ𝑚 + 𝑌𝑒𝑎𝑟𝑦 + 𝜀𝑖𝑤 ,

9

where 𝑂𝑢𝑡𝑐𝑜𝑚𝑒𝑖𝑤 is one of the outcome measures in tract 𝑖 and week 𝑤. 𝑃𝑜𝑠𝑡𝑤 is equal to 0 in weeks before the Ferguson shooting and 1 otherwise. 𝑊𝑒𝑒𝑘𝑤 (the running variable) represents the wth week. 𝑓(∙) is a flexible parametric or nonparametric function of (𝑊𝑒𝑒𝑘𝑤 − 85), the “distance” relative to the cutoff week. We treat the 85th week in our sample period – when the first round of protests occurred immediately following the Ferguson shooting (August 10 – 13, 2014) – as the first week of the post-shooting period.6 𝐗 𝑖𝑦 is vector of time-varying covariates for tract 𝑖 in year 𝑦, including demographic and socioeconomic controls. 𝑇𝑟𝑎𝑐𝑡𝑖 is the Census tract fixed effects. 𝑀𝑜𝑛𝑡ℎ𝑚 and 𝑌𝑒𝑎𝑟𝑦 are month and year fixed effects, respectively. 𝜀𝑖𝑤 is the random error term. We cluster the standard errors by week and also use the data-driven ImbensKalyanaraman (IK) optimal bandwidth (Imbens and Kalyanaraman 2012). The parameter of interest is 𝛽1, which measure the instantaneous effects of the Ferguson shooting. The identifying assumption of our RD requires other determinants of the outcomes (e.g., time-varying tract covariates) have no discontinuous changes at the time of the shooting. While such weekly data are not available for us to perform an empirical test, the identifying assumption does seem quite plausible in this context, as it is hard to believe other potential factors (e.g., demographics, residents’ educational attainment, and local economic status) would significantly change immediately after the shooting. In evaluating the longer-run effects, we turn to a difference-in-differences (DD) strategy and examine outcome changes in St. Louis’ predominantly black communities (treatment group), relative to similar changes in predominantly white communities (control group).

The

corresponding identifying assumption requires that the outcome measures in the treatment and control tracts should have trended similarly in the absence of the Ferguson shooting. The DD

6

The week of August 10 – 16. 10

strategy has a stronger assumption than the RD design. However, it allows us to move beyond the immediate shooting effects in the short run. Formally, we estimate the following quarterly panel data model: (2) 𝑂𝑢𝑡𝑐𝑜𝑚𝑒𝑖𝑞 = 𝛽0 + 𝛽1 𝑃𝑜𝑠𝑡𝑞 × 𝐈(%𝐵𝑙𝑎𝑐𝑘𝑖 ≥ 50%) + 𝐗 𝑖𝑦 𝜸 + 𝑇𝑟𝑎𝑐𝑡𝑖 + 𝑄𝑢𝑎𝑟𝑡𝑒𝑟𝑞 + 𝜀𝑖𝑞 , where 𝑞 indexes quarter. 𝑃𝑜𝑠𝑡𝑞 is equal to 0 in quarters before the Ferguson shooting, 0.5 in the quarter when the shooting occurred (i.e., 2014Q3), and 1 in other post-shooting quarters. 7 %𝐵𝑙𝑎𝑐𝑘𝑖 is tract 𝑖’s average black population share during 2013 – 2015. 𝐈(∙) is the indicator function, where predominantly black communities are defined to be the high %𝐵𝑙𝑎𝑐𝑘 tracts (%𝐵𝑙𝑎𝑐𝑘 ≥ 50% ) and predominantly white communities the low %𝐵𝑙𝑎𝑐𝑘 tracts (%𝐵𝑙𝑎𝑐𝑘 < 50%). 𝐗 𝑖𝑦 is vector of time-varying covariates for tract 𝑖 in year 𝑦, including demographic and

socioeconomic controls. 𝑇𝑟𝑎𝑐𝑡𝑖 is the Census tract fixed effects, which accounts for timeinvariant tract-specific unobservables and therefore captures pre-existing differences between tracts with high and low %𝐵𝑙𝑎𝑐𝑘. 𝑄𝑢𝑎𝑟𝑡𝑒𝑟𝑞 is the year-by-quarter fixed effects and controls for effects of the general time trend that are common to all tracts. 𝛽1 hence measures the average effect of the Ferguson shooting. Following Bertrand, Duflo, and Mullainathan (2004), in this DD specification we cluster standard errors at the tract level to account for potential within-tract serial error correlation. Furthermore, we explore the medium- and long-run effects by decomposing 𝛽1 with the following model: (3) 𝑂𝑢𝑡𝑐𝑜𝑚𝑒𝑖𝑞 = 𝛽0 + 𝛼1 𝑃𝑜𝑠𝑡2014𝑞 × 𝐈(%𝐵𝑙𝑎𝑐𝑘𝑖 ≥ 50%) + 𝛼2 𝑃𝑜𝑠𝑡2015𝑞 × 𝐈(%𝐵𝑙𝑎𝑐𝑘𝑖 ≥ 50%) + 𝐗 𝑖𝑦 𝜸 + 𝑇𝑟𝑎𝑐𝑡𝑖 + 𝑄𝑢𝑎𝑟𝑡𝑒𝑟𝑞 + 𝜀𝑖𝑞 ,

7

Approximately half of 2014Q3 belongs to the post-shooting period. 11

where the dummy variable 𝑃𝑜𝑠𝑡2014 is equaled to 1 for the post-shooting period in 2014 and the dummy 𝑃𝑜𝑠𝑡2015 is equaled to 1 for the year of 2015. Thus, we use 𝛼1 to measure the mediumrun effects (during more than four months following the shooting) and 𝛼2 the long-run effects (during 2015). Finally, we also address the validity of the common trend assumption – which is crucial in interpreting our DD estimates as causal – in two major ways. First, we directly study the outcome trends for the treatment and control groups before the Ferguson shooting and examine if there was any divergence. Second, we investigate if other factors could systematically change between the predominantly black and white communities and confound our estimated effects, by testing if the DD estimates are robust to including time-varying tract characteristics (𝐗 𝑖𝑦 ).

4. Data Our analysis relies on Census tract-level measures of policing conduct, community trust in police, and crime in St. Louis. To construct these outcome measures, we obtained from the St. Louis Police Department detailed case-level micro data on police use-of-force, arrest, 911 emergency calls, self-initiated activity, and Uniform Crime Reporting (UCR) crime. Importantly, these data include key geocoding and time information that allows us to aggregate the data at the tract-quarter level during 2013 – 2015 as required by our identification strategy. We use four measures to capture policing behaviors. The first is the direct measure of police use-of-force, for which we know if each use-of-force case involves the use of guns, tasers, or no weapon (e.g., hands and feet). The second is the number of arrests, a common measure of police officers’ routine behaviors. Our data further allow us to distinguish between arrests related

12

to serious crimes (Part I offenses) and arrests related to less severe crimes (Part II offences).8 Third, we use 911 call response time to measure the reactive policing behavior – responding to 911 emergency calls; response time is computed as the minutes elapsed between dispatch request and arrival to the scene. We also measure response time by crime relation and emergency level. First, we compute response time for crime and non-crime related 911 calls separately. Second, we calculate response time for “high emergency” 911 calls (e.g., cases involving shooting, alarms, and traffic accidents) that require all police officers to respond “immediately” and for “low emergency” calls that only require response “as time permits.” Fourth, we have data on police officers’ self-initiated activities, which directly measure proactive policing practices. When analyzing self-initiated activities by category, we focus on 11 major self-initiated activities with regular frequency during our sample period, including building check, business interview, occupied car check, unoccupied car check, directed patrol, foot patrol, investigation, pedestrian check, problem solving, traffic violation, and truck inspection. Notably, the St. Louis Police Department is one of the few U.S. Police Departments that keeps good track of self-initiated activities and can provide quality data; this invaluable data source greatly benefits our study. Finally, the policing data contain information on officer service year and offender race, which allow us to examine the shooting effects in greater detail. Our measure of community trust in police is crime reporting through 911 calls. Although it is common to use surveys or interviews to capture individuals’ attitudes, it is believed to be less reliable when it comes to eliciting citizens’ opinions toward the police (Rosenfeld, Jacobs, and

8

The Federal Bureau of Investigation divides offenses into two groups. Part I offenses refer to the most severe crimes (e.g., homicide, rape, robbery, aggravated assault, burglary, larceny, and auto theft) that occur with regularity across the U.S. Other offenses are categorized as Part II offenses, such as forgery, fraud, embezzlement, vandalism, prostitution, gambling, and driving under the influence. Details can be found at: https://www2.fbi.gov/ucr/cius_04/appendices/appendix_02.html. 13

Wright 2003). Measuring trust this way could be even more problematic in predominantly black communities where the legal cynicism is deep-rooted. For example, the “legal cynic may report crime just as regularly as the citizen who views the criminal justice system with reverence” (Desmond, Papachristos, and Kirk 2016). In addition, we also count 911 calls that were not related to criminal activities, which can be used to test if the shooting changed emergency reporting in general. The UCR crime data provide us with information on violent crimes (homicide, robbery, aggravated assault) and property crimes (burglary, larceny, and auto theft), which satisfy the uniform crime reporting criteria set by the FBI.9 These crimes account for the majority of the FBI’s Part I offenses that are widely used to measure the level and scope of crimes occurring throughout the U.S.10 Our analysis also hinges on measuring average black population shares at the Census tract level (%𝐵𝑙𝑎𝑐𝑘𝑖 ), which we use to identify predominantly black communities and the comparison group (predominantly white communities). To do so, we collected corresponding data from the 2013 – 2015 American Community Survey (ACS). Figure 2 visualizes the St. Louis Census tracts by black population share, which shows that predominantly black communities are concentrated in the northern area. The ACS also provides us with data on time-varying tract-level covariates that capture annual changes in demographic and socioeconomic factors, including white population share, percentage of males, percentage of population aged 25 – 64, percentage of population with a high school diploma, percentage of population with a bachelor’s degree, poverty rate, and median household income (in 2015 dollars). While there definitely existed other tract-

9

We do not consider rape in this study, because many rape cases lack geocoding information due to confidentiality reasons. As a result, we cannot aggregate rape data at the desired tract-quarter level. 10 https://www2.fbi.gov/ucr/cius_04/appendices/appendix_02.html 14

level unobservable factors that could affect the outcomes, the main purpose of using these observable covariates is to empirically test the validity of the DD’s common trend assumption. To the extent that the DD estimates are robust to the inclusion of these control variables, it provides evidence to support the identifying assumption. Table 1 presents the summary statistics. In order to meaningfully analyze the outcomes across tracts, we calculate the rates of use-of-force, arrests, self-initiated activities, and crimes by adjusting these outcomes for every 1,000 tract population. Table A1 in the Online Appendix breaks down the summary statistics for the 53 predominantly black communities and the 53 predominantly white communities, which shows the disparities between the treatment and control groups in both outcome measures and tract covariates.

For example, predominantly black

communities had higher poverty rates, lower household median incomes, and a population less college-educated. Highly correlated with these were higher levels of use-of-force rate, selfinitiated activity rate, 911 call response time, 911 crime reporting, and violent crime rates. These correlations highlight the importance of adopting a credible research design to identify the causal effects of the Ferguson shooting. Particularly, it is important to note that the level differences in tract attributes – which can easily confound a cross-sectional analysis – do not necessarily pose a threat to our research design. This is because our RD design only requires no instantaneous discontinuities in tract attributes at the time of shooting and our DD strategy requires comparable changes in these variables between predominantly black and white communities.

15

5. Main Results 5.1 Short-Run Effects We begin by showing the Ferguson shooting’s short-run effects. Figure 3 presents the weekly averages of the four policing measures (use-of-force, arrest, 911 call response time, selfinitiated activity), community-police trust, violent crimes (homicide, robbery, and aggravated assault) and property crimes (burglary, larceny, and theft) for the 84 weeks before the shooting and the 72 weeks afterwards. Each graph includes the nonparametric-fitting curves and their 95% confidence intervals on both sides of the cutoff. The graphical evidence indicates significant decreases in arrests and self-initiated activities at the time of shooting, implying immediate police response. However, there appeared to be no instantaneous effects on trust and crime. Table 9 reports the corresponding RD estimates, where we use linear, quadratic and nonparametric fits, as well as using IK optimal bandwidths.11 We additionally control for contemporaneous weekly crime rates for regressions of arrest and 911 crime reporting (Column 3 and Column 7, respectively), which could be susceptible to crime effects. Consistent with Figure 3, we only find large and significant immediate reductions in arrests and self-initiated activities; the estimates are quite robust to different specifications. The estimates in Column 3 indicate that arrest rate significantly fell by 0.54 immediately following the shooting, implying that the de-policing effect outweighed the monitoring effect. Compared to the average pre-shooting arrest rate (1.36), this represents a sizable 39.7 percent drop. Along similar lines, the significant and negative estimates for self-initiated activities (Column 5) reinforce the existence of police de-policing after the Ferguson shooting. According to the estimates, self-initiated activities – the proactive policing behaviors that are exposed only to the shooting’s de-policing effect and not subject to the

For nonparametric regressions, we use the Stata estimating procedure “rd” by Nichols (2011), with the choice of a triangle kernel. 11

16

monitoring effect – dropped by approximately 50 percent. Evidently, police officers’ immediate response was to strategically pull back in order to avoid controversial confrontations. Given the significant short-run effects on arrests and self-initiated activities, we next explore the effects by arrest and self-initiated activity category. Table 3 reports arrest estimates by offense severity and by arrestee race. The first two columns indicate that the shooting had large and negative short-run effects on arrests of both Part I and Part II offenses, with the larger effect on the latter (45.1 – 46.5 percent) than on the former (20.6 – 24.9 percent). This is consistent with the fact that Part II offenses are relatively less severe crimes, for which police officers could be subject to less monitoring – which would strengthen the net de-policing effect – when making arrests after the shooting. In Columns 3 and 4, we find that the immediate plunge in arrests was driven by arrests of black arrestees (48 percent), while the effect on arrests of white offenders was quantitatively small (0.9 – 3.9 percent) and statistically insignificant. Thus, as one would expect, the Ferguson police shooting – in which a black teenager was fatally shot followed by extensive scrutiny and investigations – led to immediate police response by arresting fewer African Americans, out of fear of involving in another highly publicized incident. In Table 4, we look into the effects on major self-initiated activity categories. The estimates – which are negative and highly significant in 9 out of 11 columns – indicate a strong pattern of systematic de-policing in proactive inspections, as evidenced by the large reductions in building check (30 percent), occupied car check (55 percent), unoccupied car check (99 percent), directed patrol (43 percent), foot patrol (72 percent), investigation (37 percent), pedestrian check (61 percent), problem solving (81 percent), and traffic violation (56 percent).12

12

Although being insignificant, the estimates for business interview and truck inspection are still negative, representing changes of 18 percent and 3 percent, respectively. 17

Finally, we examine whether police officers in predominantly black communities responded more when making fewer arrests and conducting fewer self-initiated inspections than those in predominantly white communities, as discussed in the conceptual framework. Table 5 provides supporting evidence, showing that the decreases in arrest rate and self-initiated activity rate are larger in black communities – 0.75 (45 percent) and 10.46 (51 percent), respectively – than the corresponding decreases – 0.34 (32 percent) and 6.62 (50 percent) – in white communities. In short, we find police officers were very responsive to the highly publicized Ferguson shooting in the short run, by dramatically reducing arrests and self-initiated activities. However, there is no evidence of significant short-run changes in trust and crime.

5.2. Medium- and Long-Run Effects Although we only find significant effects on two types of policing behaviors in the immediate short run, it is possible that other policing behaviors, as well as community-police trust and crime, may not be affected instantaneously. We now examine the Ferguson shooting’s effects in the longer run – a six-quarter post-shooting period – by comparing outcome changes before and after the shooting between predominantly black and white communities.

5.2.1. Police Use-of-Force Figure 4 provides a graphical overview of use-of-force rates for treatment and control tracts before and after the Ferguson shooting – where the solid line refers to predominantly black communities and the dashed line refers to predominantly white communities – which sets the stage for the regression analysis that follows. This graph clearly shows that the use-of-force trends in treatment and control tracts tracked each other closely during the six quarters before the Ferguson

18

shooting. This evidence of parallel trends is exactly what is required by our DD identifying assumption and highlights that white communities serve as a good control group for black communities. In addition, immediately after the Ferguson shooting, treatment tracts experienced a large relative drop in use-of-force that lasted for half a year (the last two quarters in 2014). During the following year of 2015, the use-of-force difference between treatment and control tracts bounced back and was similar to the pre-shooting level. We report formal regression estimates in Table 6. Panel 1 presents the average effect estimates based on Equation (2).

Columns 1 corresponds to the most parsimonious DD

specification, in which we only include Census tract and year-by-quarter fixed effects. The small and insignificant estimate indicates that the Ferguson shooting on average did not affect use-offorce. Panel 2 presents the medium- and long-run estimates based on Equation (3). Consistent with Figure 3, the last two estimates in Column 1 show that the insignificant average estimate masks the shooting’s large and negative medium-run effect. Specifically, the significant estimate indicates that use-of-force in predominantly black communities dropped by 0.32 during the postshooting period in 2014, compared to predominantly white communities. Relative to the preshooting mean use-of-force rate in treatment tracts (0.75), this estimate translates into a considerable 42.7 percent drop.

In comparison, the long-run estimate (0.08) is small and

insignificant. Column 2 reports our preferred estimates, in which we additionally control for tractlevel time-varying covariates, including white population share, percentage of males, percentage of population aged 25 – 64, percentage of population with a high school diploma, percentage of population with a bachelor’s degree, poverty rate, and median household income. Importantly, including these controls does not change the DD estimates. This implies that tract attributes in

19

treatment and control tracts trended similarly and lends further support to the DD identifying assumption. Combined together, we find convincing graphical and regression evidence that the Ferguson shooting reduced use-of-force in the medium run. This result shows that the police were very responsive to the rising expected cost of use-of-force, driven by the fear of getting entangled in a similarly publicized and controversial use-of-force incident. In Columns 3 through 5, we look into the use-of-force effect by weapon use. Not surprisingly, gun use – the type of use-of-force in the Ferguson incident – experienced the largest medium-run drop (75 percent), as opposed to the insignificant 22.2 percent decline in taser use and the significant 52.3 percent decrease in the use of bare hands and feet. In the long run, use-of-force appeared to shift toward taser use, which is less lethal than using firearms but still more effective than not using any weapons at all. As shown by the significant estimate in Column 4, use-of-force with tasers increased by 33.3 percent in 2015. In Columns 6 and 7, we further show that the shooting led to less use-of-force against both black and white offenders in 2014.

5.2.2. Arrest Table 7 reports the effect on arrests. The first three columns contain the main estimates as we progressively include time and tract fixed effects, tract covariates, and contemporaneous crime rates. The significantly negative and robust estimates show that the shooting’s negative impact on arrests – driven by the de-policing effect – persisted in the longer run. Specifically, the preferred estimates in Column 3 indicate that the Ferguson shooting caused a 17 percent decline in arrests, with a slightly larger effect in the medium run (18.5 percent) than in the long run (13.2 percent). This dynamic pattern is in line with the graphical evidence in Figure 5, which further shows that

20

there were no divergent arrest trends prior to the shooting. Our finding is similar to Shi (2009), which exploits the publicized 2001 Cincinnati riot – following a white officer shot dead an unarmed black adolescent – and finds substantial drops in arrests. Next, we examine the shooting’s differential arrest effects. Estimates in Columns 4 and 5 show a similar pattern as in the short run: a larger negative effect on arrests of the less severe crimes. Specifically, the small and insignificant estimates in Column 4 indicate that the shooting had no longer-run impact on arrests of Part I offenses, despite its large and significant short-run effect. This is likely due to that the monitoring effect became stronger over time that offset the de-policing effect when it comes to addressing the most important and salient public safety issues. Meanwhile, Column 5 estimates imply that arrests of Part II offenses dropped substantially by 23 percent. Furthermore, Columns 6 and 7 suggest that the police substituted away from arresting blacks toward arresting whites, implying that they kept avoiding potentially controversial confrontations in the longer run. The corresponding estimates indicate that the shooting led arrests of black arrestees to drop by 23.3 percent in 2014 and by 16.1 percent in 2015, while arrests of white arrestees increased by 38.1 percent and 25.6 percent (insignificant) accordingly.

5.2.3. 911 Call Response Time Table 8 reports results on 911 call response time, our measure of reactive policing. The preferred estimates in Column 2 combined with Figure 6 show that police officers responded to 911 emergency calls faster in the post-shooting period, with the effect driven by the long-run effect in 2015. Specifically, response time on average improved by 0.84 minute, or a 6.3 percent decrease compared to the pre-shooting mean of 13.27 minutes. This result implies that, when it comes to responding to 911 emergency calls, the Ferguson shooting generated a larger monitoring effect

21

than its de-policing effect. In Columns 3 and 4, we find suggestive evidence that the response time improvement for crime reporting (9.1 percent) was three times more than that for non-crime reporting (2.5 percent), although none of the estimates are precisely estimated. In the last two columns, we break down the analysis by 911 call emergency level, which allows us to directly examine the shooting effects on purely reactive responses and discretionary responses. Column 5 estimates show that response to “high emergency” 911 calls – situations that require immediate response – improved by 6.7 percent. In contrast, response time for “low emergency” 911 calls – for which response is discretionary depending on time constraint – rose by more than five minutes, or 11.6 percent, as evidenced by Column 6. These results provide evidence that the Ferguson shooting improved reactive policing (through the dominating monitoring effect) and led to less aggressive proactive policing (through the stronger de-policing effect).

5.2.4. Self-Initiated Activity In this section, we focus on the shooting’s impact on proactive policing. Table 9 and Figure 7 present strong evidence of the expected de-policing. Estimates in Column 2 show that police officers conducted 22.8 percent fewer self-initiated inspections in response to the shooting in the longer run; the effects were similar in 2014 (24.2 percent drop) and in 2015 (20.8 percent drop). Columns 3 through 13 further show that the shooting led to significant reductions in most selfinitiated activity categories, with the negative estimates ranging between 11.2 percent and 88.3 percent. Combined with the short-run estimates, we find that the Ferguson shooting had a longlasting negative impact proactive policing, causing a sharp decrease in police presence.

22

5.2.5. Community Trust in Police Table 10, along with Figure 8, contains results on community trust in police, which we measure by the number of 911 crime reporting calls made by residents. The preferred estimates in Column 3 indicate a significant average increase of 10.8 percent in crime reporting following the Ferguson shooting: a 5.3 percent increase in 2014 and an 11.1 percent increase in 2015. It is worth pointing out that these estimates stay robust when we control for contemporaneous crime rates, as one might worry that the increase in crime reporting was driven by surging crimes. This finding suggests that the shooting-induced scrutiny and investigations turned out to restore community trust in police more than the shooting exacerbated legal cynicism in predominantly black communities. Additionally, we ask if some unobserved trends in 911 emergency reporting could confound our estimated effect on community-police trust, by investigating the shooting’s effect on 911 non-crime reporting. The largely small and insignificant estimates in the last three columns provide little evidence of that.

5.2.6. Crime This section reports crime effects, which reflect the net effects of the Ferguson shooting through the channels of policing and trust. Figures 9 and 10 and Table 11 provide strong evidence that the Ferguson shooting increased violent crimes in the longer run. Specifically, the preferred estimates in Columns 2, 4, and 6 imply that the significant relative increases in in homicides, robberies, and aggravated assaults in predominantly black communities were sizable 63.2 percent, 16.9 percent, and 19.5 percent, respectively. These results are consistent with the strong evidence of de-policing in self-initiated activities, especially when arrests of Part I offenses experienced no meaningful changes. As police officers withdrew from proactive policing – which is vital in crime

23

prevention – out of fears of firing and other punishment after the Ferguson shooting, police presence dropped precipitously that emboldened criminals and increased crimes. Meanwhile, we do not find similar increases in property crimes. In fact, the changes in property crimes – burglary, larceny, and auto theft – are in general statistically insignificant except for the negative effects on burglary in 2015 (a 14.6 percent drop) and auto theft in 2014 (a 15.9 percent drop), as shown by estimates in Columns 7 through 12. One possible explanation is that the shooting caused worse legal cynicism for some residents – who would rather resort to violence instead of the police in otherwise non-violent confrontations with criminals – even though on average crime reporting increased.

5.2.7. Differential Effects This section examines whether junior and senior police officers had differential responses to the Ferguson shooting. Compared to their senior counterparts, junior officers are expected to face a higher risk when receiving scrutiny and being investigated following their publicized useof-force, because they fear that the police department would provide less support due to their relative lack of track record (Shi 2009). We test this hypothesis using an officer-level DD strategy, which compares the changes in use-of-force, arrests, 911 call response time, and self-initiated activities – the four outcomes for which we can aggregate data at the officer level – before and after the shooting between junior and senior officers. Specifically, we estimate the following equations and cluster standard errors at the officer level: (4-1) 𝑂𝑢𝑡𝑐𝑜𝑚𝑒𝑖𝑡 = 𝛽0 + 𝛽1 𝑃𝑜𝑠𝑡𝑡 × 𝐽𝑢𝑛𝑖𝑜𝑟𝑂𝑓𝑓𝑖𝑐𝑒𝑟𝑖 + 𝑂𝑓𝑓𝑖𝑐𝑒𝑟𝑖 + 𝑄𝑢𝑎𝑟𝑡𝑒𝑟𝑡 + 𝜀𝑖𝑡 , (4-2) 𝑂𝑢𝑡𝑐𝑜𝑚𝑒𝑖𝑡 = 𝛽0 + 𝛼1 𝑃𝑜𝑠𝑡2014𝑡 × 𝐽𝑢𝑛𝑖𝑜𝑟𝑂𝑓𝑓𝑖𝑐𝑒𝑟𝑖 + 𝛼2 𝑃𝑜𝑠𝑡2014𝑡 × 𝐽𝑢𝑛𝑖𝑜𝑟𝑂𝑓𝑓𝑖𝑐𝑒𝑟𝑖 + 𝑂𝑓𝑓𝑖𝑐𝑒𝑟𝑖 + 𝑄𝑢𝑎𝑟𝑡𝑒𝑟𝑡 + 𝜀𝑖𝑡 ,

24

where 𝐽𝑢𝑛𝑖𝑜𝑟𝑂𝑓𝑓𝑖𝑐𝑒𝑟𝑖 is a dummy variable that equals one if the officer’s service year with the St. Louis Police Department was equaled to or smaller than ten years and equals zero if the service year was more than ten years.13 𝑂𝑓𝑓𝑖𝑐𝑒𝑟𝑖 is the officer fixed effects that controls for officer-level time-invariant unobservables. Table 13 reports estimates of 𝛽1 (Panel 1), as well as 𝛼1 and 𝛼2 (Panel 2). The significant estimates in Columns 2 and 4 confirm the expected larger de-policing effect on junior officers: junior officers made nearly 9 percent fewer arrests and cut back 18.3 percent self-initiated inspections after the Ferguson shooting compared to senior officers.

5.2.8. Additional Checks This section presents several checks of the estimated longer-run effects. First, we turn to an alternative DD specification that exploits more variation in %𝐵𝑙𝑎𝑐𝑘 , by replacing 𝐈(%𝐵𝑙𝑎𝑐𝑘𝑖 ≥ 50%) with %𝐵𝑙𝑎𝑐𝑘 in Equations (2) and (3). Table 12 reports corresponding estimates, whose signs and statistical significance are basically the same as their counterparts in Tables 6 through 11. In other words, the alternative DD specification tells a similar story regarding the effects of the Ferguson shooting in the post-shooting period: predominantly black communities experienced less in use-of-force (in the medium run), fewer arrests, an improvement in responding to 911 calls, de-policing in proactive policing, increased confidence in police through crime reporting, and spikes in violent crimes, compared to predominantly white communities. Specifically, for every 10 percentage point increase in %𝐵𝑙𝑎𝑐𝑘, the significant relative changes in use-of-force, arrests, 911 response time, self-initiated activities, crime reporting, homicides, robberies, and aggravated assaults equal -7.4 percent, -4 percent, -1.2 percent, -4.4 percent, 2 percent, 18.2 percent, 3.2 percent, and 4.4 percent, respectively. 13

To ensure that the police officers used for this analysis had enough experience, we only consider those who had at least five years of service. 25

Tables 14-1 and 14-2 presents additional robustness checks. In each table, Panel 1 includes the preferred longer-run estimates from Tables 6 through 11. In Panel 2, we address the concern that the third quarter of 2014 – the first post-shooting quarter in the main analysis – is partially treated, by simply dropping it out of the sample. In Panel 3, we choose an alternative control group – tracts with less than 20% black population shares – considering that that the current control group (tracts with %𝐵𝑙𝑎𝑐𝑘 < 50%) might be contaminated with some tracts consisting of reasonably large black population shares. Along similar lines, we further restrict our attention to treatment tracts that have at least 80% black population shares in Panel 4. Finally, we re-estimate the effects in Panel 5, by using weighted least squares (WLS) with tract population as the weight. The estimates stay quite robust.

5.3. Discussion Rising violent crime rates – surging homicide rates in particular – in St. Louis and many other major U.S. cities following the Ferguson shooting (Rosenfeld 2015, 2016) have led to the birth of the so-called “Ferguson effect”, a term coined by Sam Dotson, Chief of the St. Louis Metropolitan Police Department (Gold 2015). Our analysis supports this claim by documenting the increases in violent crimes occurring in St. Louis’ predominantly black communities relative to predominantly white communities in the period after the Ferguson shooting. During our sample period, homicide, robbery, and aggravated assault rates in predominantly black communities increased by 0.16, 0.6, and 0.93. Thus, our preferred violent crime estimates in Table 10 imply that the Ferguson shooting was responsible for 75 percent, 50 percent, and 83 percent of the increases in homicides, robberies, and aggravated assault, respectively, during 2014 – 2015. As

26

to whether the Ferguson effect might have become a nationwide phenomenon, our view is that it requires much more empirical evidence from other cities.14 In addition to quantifying the Ferguson effect, our analysis also examines the behavioral responses of police officers, residents, and criminals, the mechanism through which the Ferguson shooting affected crime. In particular, we empirically demonstrate the existence of the muchhypothesized and talked-about de-policing – the dominant interpretation of the Ferguson effect (Rosenfeld 2016). Our results clearly show that police officers responded to the highly publicized Ferguson shooting by pulling back from law enforcement activities through making fewer arrests and self-initiated inspections. Importantly, our findings are consistent with changes in police attitudes and experiences following fatal encounters between the police and blacks, according to a recent Pew Research Center survey (Morin, Parker, Stepler, and Mercer 2017). Specifically, the survey shows that, following such high-profile incidents, 93 percent of the officers become more concerned about their safety, 76 percent are more reluctant to use force when it is appropriate, and 72 percent become less willing to stop and question people who seem suspicious. Meanwhile, it is also important to point out that de-policing did not appear to occur in policing practices that are dominated by the shooting’s monitoring effect, which can hold officers accountable and counteract the de-policing effect. For example, the shooting did not change arrests of Part I offenses and led to faster response to 911 emergency calls in the longer run. Finally, our analysis suggests that the Ferguson shooting’s policing and trust effects differed in the medium and long run. The graphical and regression evidence indicates more use-

14

A recent criminology study (Pyrooz, Decker, Wolfe, and Shjarback 2016) examines whether there was a nationwide Ferguson effect using crime data from 81 large U.S. cities. By comparing changes in crime trends 12 months before and after the Ferguson shooting, the authors conclude that there is no nationwide Ferguson effect on crime rates. However, there is evidence of increases in robberies in the U.S. and homicides in several cities (including St. Louis) following the shooting. 27

of-force, arrests, crime reporting, as well as faster response to 911 calls, in 2015 than in 2014. These disparities coincided with the DOJ’s conclusion of two Ferguson shooting investigations in early 2015. First, the investigation into Officer Darren Wilson’s fatal shooting of Michael Brown found “no credible evidence that Wilson willfully shot Brown as he was attempting to surrender or was otherwise not posing a threat”, leading the DOJ to recommend no civil-rights charges be brought against Wilson (Department of Justice 2015b). As a result, it lowered the potentially high expected cost of law enforcement in controversial situations facing the police in the aftermath of the Ferguson shooting. Second, the investigation of the Ferguson Police Department revealed that the FPD caused long-standing distrust and hostility from local residents (mainly African Americans) through a variety of malpractices and misconducts, which the DOJ vowed to end with a court-enforceable remedial process, including “involvement from community stakeholders as well as independent oversight” (Department of Justice 2015c). As Attorney General Eric Holder promised it, the efforts to reform law enforcement practices and establish public safety would continue in Ferguson and surrounding municipalities in order to protect and serve all members of the community (Department of Justice 2015a). This investigation, therefore, is expected to further restore the damaged community trust in police. Combined together, these two investigation altered incentives for individual police officers and public perceptions of the police in 2015, consistent with the long-run behavioral changes we have documented.

6. Conclusion With today’s information technology capable of virally disseminating information, controversial police use-of-force frequently becomes widely publicized.

Such high-profile

incidents typically arouse enormous public outrage, lead to violent protests, and draw extensive

28

scrutiny of the police, which have potentially important effects. This study empirically examines the effects on policing, community-police trust, and crime of the highly publicized 2014 police shooting in Ferguson. Our short-run analysis finds large discontinuous reductions arrests and selfinitiated inspections, showing immediate police responses at the time of shooting. In the longer run, we find that predominantly black communities saw less police use-of-force, fewer arrests, faster 911 emergency call response, limited proactive policing, more crime reporting, and surging violent crimes after the shooting, compared to comparison communities. These results indicate that highly publicized police use-of-force can significantly change law enforcement and have profound impacts on public safety and social stability.

29

Reference Anwar, Shamena and Hanming Fang (2006). "An Alternative Test of Racial Prejudice in Motor Vehicle Searches: Theory and Evidence." The American Economic Review 96(1): 127-151. Associated Press (2014). "In Wake of Ferguson, Police Try to Build Trust." Accessed June, 2017. http://www.pottsmerc.com/general-news/20140827/in-wake-of-ferguson-police-try-to-build-trust ------ (2015). "Police Present Initiative Seeks to Restore Trust." Accessed June, 2017. http://www.legalnews.com/detroit/1415460 Baumer, Eric P. (2002). "Neighborhood Disadvantage and Police Notification by Victims of Violence." Criminology 40(3): 579-616. Becker, Gary S. (1968). "Crime and Punishment: An Economic Approach." Journal of Political economy 76(2): 169-217. Berthelon, Matias E. and Diana I. Kruger (2011). "Risky Behavior among Youth: Incapacitation Effects of School on Adolescent Motherhood and Crime in Chile." Journal of Public Economics 95(1): 41-53. Bertrand, Marianne, Esther Duflo, and Sendhil Mullainathan (2004). "How Much Should We Trust Differences-in-Differences Estimates?" Quarterly Journal of Economics 119(1): 249-275. Comey, James B. (2015). "Law Enforcement and the Communities We Serve: Bending the Lines toward Safety and Justice." Speech at University of Chicago Law School. Accessed June, 2017. https://www.fbi.gov/news/speeches/law-enforcement-and-the-communities-we-serve-bendingthe-lines-toward-safety-and-justice DeAngelo, Gregory and Benjamin Hansen (2014). "Life and Death in the Fast Lane: Police Enforcement and Traffic Fatalities." American Economic Journal: Economic Policy 6(2): 231257.

30

Department of justice (2014). "Justice Department Announces National Effort to Build Trust between Law Enforcement and the Communities They Serve." Accessed June, 2017. https://www.justice.gov/opa/pr/justice-department-announces-national-effort-build-trustbetween-law-enforcement-and ------ (2015a). "Attorney General Holder Delivers Update on Investigations in Ferguson, Missouri." Accessed June, 2017. https://www.justice.gov/opa/speech/attorney-general-holderdelivers-update-investigations-ferguson-missouri ------ (2015b). "Department of Justice Report Regarding the Criminal Investigation into the Shooting Death of Michael Brown by Ferguson, Missouri Police Officer Darren Wilson." Accessed

June,

2017.

https://www.justice.gov/sites/default/files/opa/press-

releases/attachments/2015/03/04/doj_report_on_shooting_of_michael_brown_1.pdf ------ (2015c). "Investigation of the Ferguson Police Department." Accessed June, 2017. https://www.justice.gov/sites/default/files/opa/pressreleases/attachments/2015/03/04/ferguson_police_department_report_1.pdf Desmond, Matthew, Andrew V. Papachristos, and David S. Kirk (2016). "Police Violence and Citizen Crime Reporting in the Black Community." American Sociological Review 81(5): 857876. Di Tella, Rafael and Ernesto Schargrodsky (2004). "Do Police Reduce Crime? Estimates Using the Allocation of Police Forces after a Terrorist Attack." The American Economic Review 94(1): 115-133. Donohue III, John J. and Steven D. Levitt (2001). "The Impact of Race on Policing and Arrests." The Journal of Law and Economics 44(2): 367-394.

31

Draca, Mirko, Stephen Machin, and Robert Witt (2011). "Panic on the Streets of London: Police, Crime, and the July 2005 Terror Attacks." The American Economic Review 101(5): 21572181. Gold, Ashley (2015). "Why Has the Murder Rate in Some Us Cities Suddenly Spiked?". Accessed March, 2017. http://www.bbc.com/news/world-us-canada-32995911 Grinberg, Emanuella (2014). "What #Ferguson Stands for Besides Michael Brown and Darren Wilson." Accessed March, 2017. http://www.cnn.com/2014/11/19/us/ferguson-social-mediainjustice Heaton, Paul (2010). "Understanding the Effects of Antiprofiling Policies." The Journal of Law and Economics 53(1): 29-64. Imbens, Guido and Karthik Kalyanaraman (2012). "Optimal Bandwidth Choice for the Regression Discontinuity Estimator." The Review of Economic Studies 79(3): 933-959. Jacob, Brian A. and Lars Lefgren (2003). "Are Idle Hands the Devil's Workshop? Incapacitation, Concentration, and Juvenile Crime." The American Economic Review 93(5): 15601577. Kaplan, Rebecca (2014). "Obama Dispatches Eric Holder to Ferguson in Wake of Shooting." Accessed June, 2017. http://www.cbsnews.com/news/obama-dispatches-eric-holder-to-fergusonin-wake-of-shooting/ Kirk, David S. and Andrew V. Papachristos (2011). "Cultural Mechanisms and the Persistence of Neighborhood Violence." American Journal of Sociology 116(4): 1190-1233. Klick, Jonathan and Alexander Tabarrok (2005). "Using Terror Alert Levels to Estimate the Effect of Police on Crime." Journal of Law and Economics 48(1): 267-279.

32

Levitt, Steven D. (1996). "The Effect of Prison Population Size on Crime Rates: Evidence from Prison Overcrowding Litigation." The Quarterly Journal of Economics 111(2): 319-351. Lichtblau, Eric (2016). "F.B.I. Director Says 'Viral Video Effect' Blunts Police Work." Accessed July,

2017.

https://www.nytimes.com/2016/05/12/us/comey-ferguson-effect-police-videos-

fbi.html Mas, Alexandre (2006). "Pay, Reference Points, and Police Performance." The Quarterly Journal of Economics 121(3): 783-821. Morin, Rich , Kim Parker, Renee Stepler, and Andrew Mercer (2017). "Behind the Badge." Pew Research Center. Nichols, Austin (2011). "Rd 2.0: Revised Stata Module for Regression Discontinuity Estimation." http://ideas.repec.org/c/boc/bocode/s456888.html NPR (2014). "Ferguson Jury: No Charges for Officer in Michael Brown's Death." Accessed June, 2017.

http://www.npr.org/sections/thetwo-way/2014/11/24/366370100/grand-jury-reaches-

decision-in-michael-brown-case Peters, Mark (2014). "Peace Descends on Ferguson as Highway Patrol Take over Security." Accessed June, 2017. https://www.wsj.com/articles/protests-over-teens-shooting-again-turnviolent-in-st-louis-suburb-1407988466 Pyrooz, David C., Scott H. Decker, Scott E. Wolfe, and John A. Shjarback (2016). "Was There a Ferguson Effect on Crime Rates in Large Us Cities?" Journal of criminal justice 46: 1-8. Rosenfeld, Richard (2015). "Was There a “Ferguson Effect” on Crime in St. Louis." The Sentencing Project. ------ (2016). "Documenting and Explaining the 2015 Homicide Rise: Research Directions." National Institute of Justice.

33

Rosenfeld, Richard, Bruce A. Jacobs, and Richard Wright (2003). "Snitching and the Code of the Street." The British Journal of Criminology 43(2): 291-309. Shi, Lan (2009). "The Limit of Oversight in Policing: Evidence from the 2001 Cincinnati Riot." Journal of Public Economics 93(1): 99-113. Welker, Kristen and Erin McClam (2014). "'No Excuse': Obama Expresses Concern About Violence in Missouri." Accessed June, 2017. http://www.nbcnews.com/storyline/michael-brownshooting/no-excuse-obama-expresses-concern-about-violence-missouri-n180666

34

Figures and Tables Figure 1. Conceptual Framework

35

Figure 2. Tract-Level Black Population Shares in St. Louis

36

Figure 3. Weekly Outcome Measures in Treatment and Control Tracts Before and After the Ferguson Shooting Arrest

-84

2 1.5 1

-84

72

0

911 Call Response Time

Self-Initiated Activity

72

20 15 10 0

5

12

14

16

Average SIAs per 1,000 Population per Tract

18

Week

10

Response Time

.5

0 Week

25

0

.02

.04

.06

.08

Average Arrests per 1,000 Population per Tract

.1

Use of Force

-84

0

-84

72

Week

0 Week

37

72

.02

.03

.04

-84 0 .05

.1

.15

.2

.25

Average Crimes Per 1,000 Population per Tract

0

.01

2

3

4

5

6

Average 911 Calls per 1,000 Population per Tract

911 Crime Reporting

-84 Week 0

Week 72

38

72

Homicide Robbery

-84 Week 0 72

.6

.8 1

1.2

-84 0 .3

.4

.5

0

.2

-84

.1

Average Crimes Per 1,000 Population per Tract

.4

.2

.3

.4

.5

.1

.2

.3

.4

.5

Average Crimes Per 1,000 Population per Tract

.1

Aggravated Assault Burglary

Week 72 -84 Week

Week 72

39 0

Larceny

Auto Theft

-84 Week 0

72

72

Figure 4. Police Use-of-Force in Treatment and Control Tracts Before and After the Ferguson Shooting

.2

.4

.6

.8

1

Use of Force

-6

-5

-4

-3

-2

-1

0 1 Quarter

High %Black Tracts

2

3

4

5

6

Low %Black Tracts

Figure 5. Arrests in Treatment and Control Tracts Before and After the Ferguson Shooting

5

10

15

20

25

Arrest

-6

-5

-4

-3

-2

-1

0 1 Quarter

High %Black Tracts

40

2

3

4

5

Low %Black Tracts

6

Figure 6. 911 Emergency Call Response Time in Treatment and Control Tracts Before and After the Ferguson Shooting

12

13

14

Average Response Time (mins)

15

911 Call Response Time

-6

-5

-4

-3

-2

-1

0 1 Quarter

High %Black Tracts

2

3

4

5

6

Low %Black Tracts

Figure 7. Self-Initiated Activities in Treatment and Control Tracts Before and After the Ferguson Shooting

100

150

200

250

300

Self-Initiated Activity

-6

-5

-4

-3

-2

-1

0 1 Quarter

High %Black Tracts

41

2

3

4

5

Low %Black Tracts

6

Figure 8. 911 Crime Reporting in Treatment and Control Tracts Before and After the Ferguson Shooting

30

40

50

60

70

80

911 Crime Reporting

-6

-5

-4

-3

-2

-1

0 1 Quarter

High %Black Tracts

42

2

3

4

5

Low %Black Tracts

6

Figure 9. Violent Crimes in Treatment and Control Tracts Before and After the Ferguson Shooting

.5 .4 .3 .2 .1 0

Average Crimes per 1,000 Populaiton

Homicide

-6

-5

-4

-3

-2

-1

0 1 Quarter

High %Black Tracts

2

3

4

5

6

Low %Black Tracts

2.5 2 1.5 1 .5

Average Crimes per 1,000 Populaiton

Robbery

-6

-5

-4

-3

-2

-1

0 1 Quarter

High %Black Tracts

2

3

4

5

6

Low %Black Tracts

6 5 4 3 2 1

Average Crimes per 1,000 Populaiton

Aggravated Assault

-6

-5

-4

-3

-2

-1

0 1 Quarter

High %Black Tracts

43

2

3

4

5

Low %Black Tracts

6

Figure10. Property Crimes in Treatment and Control Tracts Before and After the Ferguson Shooting

1

2

3

4

5

Average Crimes per 1,000 Populaiton

6

Burglary

-6

-5

-4

-3

-2

-1

0 1 Quarter

High %Black Tracts

2

3

4

5

6

Low %Black Tracts

13 12 11 10 9 8

Average Crimes per 1,000 Populaiton

Larceny

-6

-5

-4

-3

-2

-1

0 1 Quarter

High %Black Tracts

2

3

4

5

6

Low %Black Tracts

3.5 3 2.5 2 1.5

Average Crimes per 1,000 Populaiton

4

Auto Theft

-6

-5

-4

-3

-2

-1

0 1 Quarter

High %Black Tracts

44

2

3

4

5

Low %Black Tracts

6

Table 1. Summary Statistics Variables Mean Panel 1. Outcomes (Tract-Year-Quarter Level) Use of Force (per 1,000 population) UOF 0.53 Gun 0.02 Taser 0.19 No Weapon 0.31 Black Offender 0.44 White Offender 0.09 Arrests (per 1,000 population) Arrest 15.49 Part I Offense 4.09 Part II Offense 11.39 Black Offender 12.67 White Offender 2.76 911 Call Response Time (mins) Response Time 13.40 Crime Case 16.68 Non-Crime Case 14.09 Emergency Level: High 10.94 Emergency Level: Low 44.31 Self-Initiated Activities (per 1,000 population) SIA 177.75 Building Check 37.48 Business Interview 4.08 Occupied Car Check 28.75 Unoccupied Car Check 4.55 Directed Patrol 22.26 Foot Patrol 23.95 Investigation 8.82 Pedestrian Check 16.04 Problem Solving 7.31 Traffic Violation 17.29 Truck Inspection 4.74 911 Calls (per 1,000 population) Crime Reporting 54.57 Non-Crime Reporting 189.96 Crimes (per 1,000 population) Homicide 0.15 Robbery 1.46 Aggravated Assault 2.99 Burglary 3.59 Larceny 10.54 Auto Theft 2.83 Panel 2. Tract Characteristics (Tract-Year Level) Population 3,004.82 % Black 53.21 % White 40.90 % Male 48.13 % Aged 25-64 56.33 % High School 25.71 % College 15.37 Poverty Rate 29.25 Median Household Income ($) 33,995.21

45

S.D.

Observations

0.69 0.12 0.35 0.48 0.64 0.23

1,272 1,272 1,272 1,272 1,272 1,272

22.99 9.49 14.99 19.47 4.62

1,272 1,272 1,272 1,272 1,272

3.81 12.96 4.19 3.36 25.91

1,272 1,272 1,272 1,272 1,272

139.52 41.06 5.55 28.78 6.41 28.76 33.18 7.18 21.49 10.82 20.71 14.61

1,272 1,272 1,272 1,272 1,272 1,272 1,272 1,272 1,272 1,272 1,272 1,272

31.24 97.80

1,272 1,272

0.35 1.49 3.03 2.62 8.13 1.79

1,272 1,272 1,272 1,272 1,272 1,272

1,235.53 36.50 33.78 5.27 10.38 9.31 9.94 15.31 14,194.43

318 318 318 318 318 318 318 318 318

Table 2. Short-Run Effects Policing

1

2

3

0.000 (0.013)

-0.539*** (0.121)

-0.547*** (0.124)

6996

5300

0.002 (0.013)

Arrest

Use-of-Force

Panel 1. Linear Fit Post Observations Panel 2. Quadratic Fit Post Observations Panel 3. Nonparametric Fit Post

Trust 911 Call Response Time 4

Crime

911 Crime Reporting

Self-Initiated Activity

Homicide

Robbery

Aggravated Assault

Burglary

Larceny

Auto Theft

5

6

7

8

9

10

11

12

13

1.891 (1.385)

-8.124*** (0.656)

-0.065 (0.156)

-0.072 (0.131)

0.001 (0.008)

0.011 (0.021)

-0.004 (0.058)

0.070 (0.051)

-0.075 (0.080)

0.011 (0.036)

5300

4876

2756

5724

5724

11024

4240

7844

5512

5936

7632

-0.537*** (0.138)

-0.547*** (0.141)

0.191 (1.389)

-9.154*** (0.851)

0.138 (0.169)

0.136 (0.148)

0.006 (0.009)

0.028 (0.023)

-0.004 (0.059)

0.078 (0.051)

-0.082 (0.080)

0.015 (0.035)

6996

5300

5300

4876

2756

5724

5724

11024

4240

7844

5512

5936

7632

0.001 (0.013)

-0.540*** (0.126)

-0.546*** (0.127)

1.369 (1.381)

-8.346*** (0.682)

0.003 (0.159)

-0.005 (0.135)

0.003 (0.010)

0.016 (0.021)

-0.009 (0.058)

0.073 (0.051)

-0.077 (0.080)

0.015 (0.037)

Observations 16642 16642 16642 16642 16642 16642 16642 16642 16642 16642 16642 16642 16642 Pre-Shooting Mean 0.04 1.36 1.36 12.88 16.89 3.94 3.94 0.01 0.10 0.21 0.27 0.80 0.21 IK Optimal Bandwidth Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Month and Year Fixed Effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Census Tract Fixed Effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Contemporaneous Crime Rates Yes Yes Notes: Each column in each panel represents a separate regression. The unit of observation is tract-year-week. We cluster robust standard errors at the week level and use Imbens-Kalyanaraman optimal bandwidths. Contemporaneous crime rates include homicide, robbery, aggravated assault, burglary, larceny, and auto theft rates. * Significant at the 10% level ** Significant at the 5% level *** Significant at the 1% level

46

Table 3. Short-Run Effects by Arrest Category Arrest Offense Severity Part I Part II 1 2 Panel 1. Linear Fit Post Observations Panel 2. Quadratic Fit Post Observations Panel 3. Nonparametric Fit Post

Arrestee Race Black White 3 4

-0.068*** (0.019)

-0.479*** (0.053)

-0.543*** (0.060)

-0.009 (0.012)

5406

5406

5406

5406

-0.082*** (0.021)

-0.464*** (0.065)

-0.549*** (0.066)

-0.002 (0.018)

5406

5406

5406

5406

-0.074*** (0.020)

-0.472*** (0.055)

-0.546*** (0.061)

-0.005 (0.013)

Observations 16642 16642 16642 16642 Pre-Shooting Mean 0.33 1.03 1.13 0.23 IK Optimal Bandwidth Yes Yes Yes Yes Month and Year Fixed Effects Yes Yes Yes Yes Census Tract Fixed Effects Yes Yes Yes Yes Contemporaneous Crime Rates Yes Yes Yes Yes Notes: Each column in each panel represents a separate regression. The unit of observation is tract-year-week. We cluster robust standard errors at the week level and use ImbensKalyanaraman optimal bandwidths. Contemporaneous crime rates include homicide, robbery, aggravated assault, burglary, larceny, and auto theft rates. * Significant at the 10% level ** Significant at the 5% level *** Significant at the 1% level 47

Table 4. Short-Run Effects by Self-Initiated Activity Category Building Check Panel 1. Linear Fit Post Observations Panel 2. Quadratic Fit Post Observations Panel 3. Nonparametric Fit Post

1

Business Interview 2

-1.10** (0.46)

Occupied Unoccupied Car Check Car Check

Foot Patrol

Investigation

Pedestrian Check

Problem Solving

Traffic Violation

Truck Inspection

6

7

8

9

10

11

3

4

Directed Patrol 5

-0.08 (0.06)

-1.42*** (0.23)

-0.43*** (0.04)

-0.97*** (0.14)

-1.42*** (0.21)

-0.25*** (0.05)

-0.96*** (0.17)

-0.54*** (0.11)

-0.88*** (0.20)

-0.02 (0.13)

2862

2862

2862

2862

2862

2862

2862

2862

2862

2862

2862

-1.20** (0.43)

-0.07 (0.06)

-1.69*** (0.27)

-0.61*** (0.06)

-1.13*** (0.17)

-1.51*** (0.18)

-0.30*** (0.05)

-1.09*** (0.19)

-0.55*** (0.12)

-0.66** (0.26)

-0.00 (0.13)

2862

2862

2862

2862

2862

2862

2862

2862

2862

2862

2862

-1.12*** (0.42)

-0.07 (0.06)

-1.48*** (0.21)

-0.47*** (0.04)

-1.03*** (0.14)

-1.43*** (0.18)

-0.26*** (0.04)

-0.99*** (0.16)

-0.54*** (0.10)

-0.80*** (0.19)

-0.01 (0.12)

Observations 16642 16642 16642 16642 Pre-Shooting Mean 3.74 0.39 2.80 0.51 IK Optimal Bandwidth Yes Yes Yes Yes Month and Year Fixed Effects Yes Yes Yes Yes Census Tract Fixed Effects Yes Yes Yes Yes Notes: Each column in each panel represents a separate regression. The unit of observation Kalyanaraman optimal bandwidths. * Significant at the 10% level ** Significant at the 5% level *** Significant at the 1% level

48

16642 16642 16642 16642 16642 16642 16642 2.40 2.01 0.73 1.65 0.67 1.39 0.37 Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes is tract-year-week. We cluster robust standard errors at the week level and use Imbens-

Table 5. Short-Run Effects on Arrests and Self-Initiated Activities in Predominantly Black and White Communities Arrest Black Communities White Communities 1 2 Panel 1. Linear Fit Post Observations Panel 2. Quadratic Fit Post Observations Panel 3. Nonparametric Fit Post

Self-Initiated Activity Black Communities White Communities 3 4

-0.762*** (0.185)

-0.323** (0.159)

-9.760*** (0.969)

-6.488*** (0.838)

2650

2650

2650

2650

-0.729*** (0.209)

-0.357* (0.185)

-11.470*** (1.411)

-6.837*** (0.860)

2650

2650

2650

2650

-0.757*** (0.194)

-0.334** (0.164)

-10.157*** (1.022)

-6.535*** (0.840)

Observations 8321 8321 8321 8321 Pre-Shooting Mean 1.67 1.06 20.54 13.25 IK Optimal Bandwidth Yes Yes Yes Yes Month and Year Fixed Effects Yes Yes Yes Yes Census Tract Fixed Effects Yes Yes Yes Yes Contemporaneous Crime Rates Yes Yes Notes: Each column in each panel represents a separate regression. The unit of observation is tract-year-week. We cluster robust standard errors at the week level and use Imbens-Kalyanaraman optimal bandwidths. Contemporaneous crime rates include homicide, robbery, aggravated assault, burglary, larceny, and auto theft rates. * Significant at the 10% level ** Significant at the 5% level *** Significant at the 1% level

49

Table 6. Longer-Run Effects on Police Use-of-Force

All Panel 1. Post ×I(%Black ≥ 0.5) Panel 2. Post2014 ×I(%Black ≥ 0.5) Post2015 ×I(%Black ≥ 0.5)

Use of Force Weapon Use Taser No Weapon 4 5

Offender Race Black White 6 7

1

2

Gun 3

-0.03 (0.06)

-0.03 (0.06)

0.01 (0.01)

0.05 (0.03)

-0.09* (0.05)

-0.02 (0.05)

-0.02 (0.02)

-0.32*** (0.07) 0.08 (0.07)

-0.32*** (0.07) 0.08 (0.07)

-0.03** (0.01) 0.01 (0.01)

-0.06 (0.04) 0.09** (0.04)

-0.23*** (0.06) -0.02 (0.05)

-0.25*** (0.06) 0.09 (0.07)

-0.07* (0.03) -0.01 (0.02)

Pre-Shooting Mean (Treatment Tracts) 0.75 0.75 0.04 0.27 0.44 0.71 0.04 Observations 1272 1272 1272 1272 1272 1272 1272 Year-by-Quarter Fixed Effects Yes Yes Yes Yes Yes Yes Yes Census Tract Fixed Effects Yes Yes Yes Yes Yes Yes Yes Time-Varying Tract Covariates Yes Yes Yes Yes Yes Yes Notes: Each column in each panel represents a separate regression. The unit of observation is tract-year-quarter. Robust standard errors are clustered at the tract level. Tract covariates include white population share, percentage of males, percentage of population aged 25 – 64, percentage of population with a high school diploma, percentage of population with a bachelor’s degree, poverty rate, and median household income. * Significant at the 10% level ** Significant at the 5% level *** Significant at the 1% level

50

Table 7. Longer-Run Effects on Arrests

All Panel 1. Post ×I(%Black ≥ 0.5) Panel 2. Post2014 ×I(%Black ≥ 0.5) Post2015 ×I(%Black ≥ 0.5)

Arrest Offense Severity Part I Part II 4 5

Arrestee Race Black White 6 7

1

2

3

-3.31** (1.57)

-3.61** (1.43)

-3.70** (1.47)

0.23 (0.50)

-3.93*** (1.15)

-4.24*** (1.25)

0.56* (0.31)

-3.60*** (1.31) -2.58* (1.50)

-3.70*** (1.27) -2.87** (1.34)

-4.02*** (1.22) -2.87** (1.41)

-0.23 (0.37) 0.42 (0.52)

-3.80*** (1.02) -3.30*** (1.08)

-4.65*** (1.07) -3.31*** (1.19)

0.67** (0.27) 0.45 (0.31)

Pre-Shooting Mean (Treatment Tracts) 21.74 21.74 21.74 4.68 17.06 19.94 1.76 Observations 1272 1272 1272 1272 1272 1272 1272 Year-by-Quarter Fixed Effects Yes Yes Yes Yes Yes Yes Yes Census Tract Fixed Effects Yes Yes Yes Yes Yes Yes Yes Time-Varying Tract Covariates Yes Yes Yes Yes Yes Yes Contemporaneous Crime Rates Yes Yes Yes Yes Yes Notes: Each column in each panel represents a separate regression. The unit of observation is tract-year-quarter. Robust standard errors are clustered at the tract level. Tract covariates include white population share, percentage of males, percentage of population aged 25 – 64, percentage of population with a high school diploma, percentage of population with a bachelor’s degree, poverty rate, and median household income. Contemporaneous crime rates include homicide, robbery, aggravated assault, burglary, larceny, and auto theft rates. * Significant at the 10% level ** Significant at the 5% level *** Significant at the 1% level

51

Table 8. Longer-Run Effects on 911 Emergency Call Response Time Response Time Crime Non-Crime Case Case 3 4

All Panel 1. Post ×I(%Black ≥ 0.5) Panel 2. Post2014 ×I(%Black ≥ 0.5) Post2015 ×I(%Black ≥ 0.5)

Emergency Level High Low 5 6

1

2

-0.88* (0.45)

-0.84* (0.44)

-1.61 (1.63)

-0.32 (0.56)

-0.75** (0.37)

5.12* (2.81)

-0.40 (0.52) -0.93* (0.47)

-0.36 (0.51) -0.89* (0.45)

-0.83 (2.06) -1.53 (1.73)

-0.08 (0.54) -0.46 (0.57)

-0.40 (0.46) -0.78* (0.40)

0.96 (3.44) 6.52** (2.85)

Pre-Shooting Mean (Treatment Tracts) 13.27 13.27 17.77 12.98 11.12 44.28 Observations 1272 1272 1272 1272 1272 1272 Year-by-Quarter Fixed Effects Yes Yes Yes Yes Yes Yes Census Tract Fixed Effects Yes Yes Yes Yes Yes Yes Time-Varying Tract Covariates Yes Yes Yes Yes Yes Notes: Each column in each panel represents a separate regression. The unit of observation is tract-year-quarter. Robust standard errors are clustered at the tract level. Tract covariates include white population share, percentage of males, percentage of population aged 25 – 64, percentage of population with a high school diploma, percentage of population with a bachelor’s degree, poverty rate, and median household income. * Significant at the 10% level ** Significant at the 5% level *** Significant at the 1% level

52

Table 9. Longer-Run Effects on Self-Initiated Activities

1

2

-66.81*** (15.90)

-62.22*** (15.35)

-9.75 (6.19)

0.21 (0.72)

-16.91*** (2.85)

-2.82*** (0.84)

-10.85*** (3.95)

3.30 (3.27)

-1.87** (0.74)

-14.25*** (2.09)

-4.02*** (1.17)

0.86 (2.83)

-5.53** (2.51)

-69.14*** (16.11) -62.04*** (15.64)

-66.22*** (15.66) -56.68*** (15.29)

-6.14 (6.41) -11.15* (6.17)

0.49 (0.68) 0.17 (0.74)

-20.15*** (3.15) -14.66*** (2.69)

-3.41*** (0.96) -2.74*** (0.82)

-10.46** (4.77) -10.51** (4.09)

-2.13 (2.37) 5.33 (3.65)

-3.04*** (0.59) -1.48* (0.79)

-13.14*** (2.08) -13.63*** (1.96)

-4.86*** (1.14) -3.22*** (1.17)

0.49 (2.34) 1.87 (3.20)

-2.99** (1.16) -6.21** (2.95)

All Panel 1. Post ×I(%Black ≥ 0.5) Panel 2. Post2014 ×I(%Black ≥ 0.5) Post2015 ×I(%Black ≥ 0.5)

Self-Initiated Activity Unoccupied Directed Foot Car Check Patrol Patrol 6 7 8

Building Check 3

Business Interview 4

Occupied Car Check 5

9

Pedestrian Check 10

Problem Solving 11

Traffic Violation 12

Truck Inspection 13

Investigation

Pre-Shooting Mean (Treatment Tracts) 273.04 273.04 50.42 4.70 53.34 8.35 40.47 26.59 13.20 31.08 13.93 20.09 7.03 Observations 1272 1272 1272 1272 1272 1272 1272 1272 1272 1272 1272 1272 1272 Year-by-Quarter Fixed Effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Census Tract Fixed Effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Time-Varying Tract Covariates Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Notes: Each column in each panel represents a separate regression. The unit of observation is tract-year-quarter. Robust standard errors are clustered at the tract level. Tract covariates include white population share, percentage of males, percentage of population aged 25 – 64, percentage of population with a high school diploma, percentage of population with a bachelor’s degree, poverty rate, and median household income. * Significant at the 10% level ** Significant at the 5% level *** Significant at the 1% level

53

Table 10. Longer-Run Effects on 911 Emergency Calls 911 Crime Reporting 1 2 3 Panel 1. Post ×I(%Black ≥ 0.5) Panel 2. Post2014 ×I(%Black ≥ 0.5) Post2015 ×I(%Black ≥ 0.5)

8.44*** (1.86) 3.44* (1.99) 9.36*** (2.03)

7.52*** (1.72) 3.44* (1.83) 8.25*** (1.86)

911 Non-Crime Reporting 4 5 6

7.07*** (1.21)

8.50 (6.34)

3.95 (5.45)

1.64 (4.77)

3.45*** (1.25) 7.26*** (1.26)

-1.35 (3.81) 13.21* (7.46)

-2.38 (3.49) 8.07 (6.56)

-4.45 (3.43) 5.29 (5.58)

Pre-Shooting Mean (Treatment Tracts) 65.30 65.30 65.30 211.74 211.74 211.74 Observations 1272 1272 1272 1272 1272 1272 Year-by-Quarter Fixed Effects Yes Yes Yes Yes Yes Yes Census Tract Fixed Effects Yes Yes Yes Yes Yes Yes Time-Varying Tract Covariates Yes Yes Yes Yes Contemporaneous Crime Rates Yes Yes Notes: Each column in each panel represents a separate regression. The unit of observation is tract-yearquarter. Robust standard errors are clustered at the tract level. Tract covariates include white population share, percentage of males, percentage of population aged 25 – 64, percentage of population with a high school diploma, percentage of population with a bachelor’s degree, poverty rate, and median household income. Contemporaneous crime rates include homicide, robbery, aggravated assault, burglary, larceny, and auto theft rates. * Significant at the 10% level ** Significant at the 5% level *** Significant at the 1% level

54

Table 11. Longer-Run Effects on Crimes

Homicide Panel 1. Post ×I(%Black ≥ 0.5) Panel 2. Post2014 ×I(%Black ≥ 0.5) Post2015 ×I(%Black ≥ 0.5)

Violent Crime Robbery 3 4

Aggravated Assault 5 6

Burglary 7 8

Property Crime Larceny 9 10

Auto Theft 11 12

1

2

0.14*** (0.04)

0.12*** (0.04)

0.29** (0.12)

0.30*** (0.12)

0.87*** 0.83*** (0.24) (0.22)

-0.42 (0.26)

-0.55** (0.25)

-0.08 (0.53)

-0.25 (0.50)

-0.01 (0.18)

0.12** (0.05) 0.14*** (0.04)

0.11** (0.04) 0.12*** (0.04)

0.19 (0.15) 0.29** (0.13)

0.22 (0.15) 0.30** (0.13)

1.01*** 1.09*** (0.30) (0.31) 0.85*** 0.77*** (0.26) (0.24)

0.21 (0.37) -0.53* (0.27)

0.16 (0.34) -0.69** (0.27)

-0.88 (0.74) 0.31 (0.50)

-0.92 (0.70) 0.14 (0.47)

-0.55** -0.53** (0.25) (0.25) 0.16 0.13 (0.18) (0.18)

-0.04 (0.17)

Pre-Shooting Mean (Treatment Tracts) 0.19 0.19 1.78 1.78 4.25 4.25 4.72 4.72 10.46 10.46 3.34 3.34 Observations 1272 1272 1272 1272 1272 1272 1272 1272 1272 1272 1272 1272 Year-by-Quarter Fixed Effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Census Tract Fixed Effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Tract-Level Time-Varying Controls Yes Yes Yes Yes Yes Yes Notes: Each column in each panel represents a separate regression. The unit of observation is tract-year-quarter. Robust standard errors are clustered at the tract level. Tract covariates include white population share, percentage of males, percentage of population aged 25 – 64, percentage of population with a high school diploma, percentage of population with a bachelor’s degree, poverty rate, and median household income. * Significant at the 10% level ** Significant at the 5% level *** Significant at the 1% level

55

Table 12. Alternative Difference-in-Differences Specification

UOF 1 Panel 1. Post ×%Black Panel 2. Post2014 ×I(%Black ≥ 0.5) Post2015 ×I(%Black ≥ 0.5)

Policing 911 Call Arrest Response Time 2 3

Trust

Crime

SIA

911 Crime Reporting

Homicide

Robbery

Aggravated Assault

Burglary

Larceny

Auto Theft

4

5

6

7

8

9

10

11

-0.000 (0.001)

-0.071*** (0.014)

-0.015** (0.006)

-0.986*** (0.204)

0.101*** (0.015)

0.002*** (0.001)

0.004** (0.002)

0.012*** (0.003)

-0.011*** (0.003)

-0.001 (0.007)

-0.001 (0.002)

-0.004*** (0.001) 0.002 (0.001)

-0.069*** (0.013) -0.059*** (0.013)

-0.008 (0.007) -0.017** (0.007)

-1.058*** (0.205) -0.902*** (0.202)

0.049*** (0.017) 0.105*** (0.016)

0.002*** (0.001) 0.002*** (0.001)

0.002 (0.002) 0.003* (0.002)

0.016*** (0.004) 0.011*** (0.003)

0.002 (0.005) -0.014*** (0.003)

-0.009 (0.008) 0.003 (0.006)

-0.007** (0.003) 0.001 (0.002)

Pre-Shooting Mean 0.54 17.79 12.97 223.80 51.19 0.11 1.25 2.73 3.45 10.33 2.78 Observations 1272 1272 1272 1272 1272 1272 1272 1272 1272 1272 1272 Year-by-Quarter Fixed Effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Census Tract Fixed Effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Time-Varying Tract Covariates Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Contemporaneous Crime Rates Yes Yes Notes: Each column in each panel represents a separate regression. The unit of observation is tract-year-quarter. Robust standard errors are clustered at the tract level. Tract covariates include white population share, percentage of males, percentage of population aged 25 – 64, percentage of population with a high school diploma, percentage of population with a bachelor’s degree, poverty rate, and median household income. Contemporaneous crime rates include homicide, robbery, aggravated assault, burglary, larceny, and auto theft rates. * Significant at the 10% level ** Significant at the 5% level *** Significant at the 1% level

56

Table 13. Differential Effects by Officer Seniority

Panel 1. Post ×I(Junior Officer) Panel 2. Post2014 × I(Junior Officer) Post2015 × I(Junior Officer)

UOF

Arrest

1

2

911 Call Response Time 3

0.00 (0.02)

-0.68* (0.37)

0.32 (0.42)

-15.53*** (4.01)

-0.03 (0.03) 0.02 (0.03)

-1.50*** (0.27) -0.29 (0.40)

0.74 (0.59) 0.23 (0.45)

-16.91*** (3.71) -14.30*** (4.14)

SIA 4

Pre-Shooting Mean (Junior Officers) 0.28 7.58 10.89 84.96 Observations 8640 8640 8059 8640 Year-by-Quarter Fixed Effects Yes Yes Yes Yes Officer Fixed Effects Yes Yes Yes Yes Notes: Each column in each panel represents a separate regression. The unit of observation is officer-year-quarter. Robust standard errors are clustered at the officer level. * Significant at the 10% level ** Significant at the 5% level *** Significant at the 1% level

57

Table 14-1. Robustness Checks UOF

Arrest

911 Call Response Time

SIA

911 Crime Reporting

Post ×I(%Black ≥ 0.5)

-0.03 (0.06)

-3.70** (1.47)

-0.84* (0.44)

-62.22*** (15.35)

7.07*** (1.21)

Observations

1272

1272

1272

1272

1272

Post ×I(%Black ≥ 0.5)

-0.04 (0.06)

-3.77** (1.49)

-0.80* (0.44)

-61.37*** (15.30)

7.18*** (1.21)

Observations

1166

1166

1166

1166

Post ×I(%Black ≥ 0.5)

-0.04 (0.06)

-4.95*** (1.02)

-1.32** (0.58)

-68.44*** (17.03)

972

972

972

0.02 (0.06)

-5.50*** (1.23)

-1.44** (0.65)

804

804

804

804

-0.07 (0.05)

-3.01** (1.47)

-1.05** (0.42)

-45.63*** (14.06)

Observations Post ×I(%Black ≥ 0.5) Observations Post ×I(%Black ≥ 0.5)

Homicide

Robbery

Panel 1. Baseline 0.12*** 0.30*** (0.04) (0.12) 1272 1272 Panel 2. Dropping 2014Q3 0.12*** 0.29** (0.04) (0.11)

Aggravated Assault

Burglary

Larceny

Auto Theft

0.83*** (0.22)

-0.55** (0.25)

-0.25 (0.50)

-0.04 (0.17)

1272

1272

1272

1272

0.80*** (0.22)

-0.52** (0.25)

-0.27 (0.49)

-0.02 (0.17)

1166

1166

1166

-0.59** (0.27)

0.11 (0.58)

0.08 (0.19)

972

972

0.13 (0.67)

-0.05 (0.20)

804

804

804

-0.44 (0.28)

-0.50 (0.44)

-0.14 (0.17)

1166 1166 1166 1166 Panel 3. Control Group (%Black ≤ 20%) 6.72*** 0.13*** 0.45*** 0.98*** (1.36) (0.04) (0.12) (0.23)

972 972 972 972 972 972 Panel 4. Treatment Group (%Black ≥ 80%), Control Group (%Black ≤ 20%) -77.96*** 7.53*** 0.19*** 0.39*** 1.12*** -0.84*** (18.34) (1.19) (0.05) (0.13) (0.26) (0.27) 804 804 804 804 Panel 5. WLS (Weight: Tract Population) 7.06*** 0.09** 0.27** 0.82*** (1.07) (0.03) (0.12) (0.18)

Observations 1272 1272 1272 1272 1272 1272 1272 1272 1272 1272 1272 Year-by-Quarter Fixed Effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Census Tract Fixed Effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Time-Varying Tract Covariates Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Contemporaneous Crime Rates Yes Yes Notes: Each column in each panel represents a separate regression. The unit of observation is tract-year-quarter. Robust standard errors are clustered at the tract level. Tract covariates include white population share, percentage of males, percentage of population aged 25 – 64, percentage of population with a high school diploma, percentage of population with a bachelor’s degree, poverty rate, and median household income. Contemporaneous crime rates include homicide, robbery, aggravated assault, burglary, larceny, and auto theft rates. * Significant at the 10% level ** Significant at the 5% level *** Significant at the 1% level

58

Table 14-2. Robustness Checks

Post2014 ×I(%Black ≥ 0.5) Post2015 ×I(%Black ≥ 0.5) Observations Post2014 ×I(%Black ≥ 0.5) Post2015 ×I(%Black ≥ 0.5) Observations Post ×I(%Black ≥ 0.5) Post ×I(%Black ≥ 0.5) Observations Post ×I(%Black ≥ 0.5) Post ×I(%Black ≥ 0.5) Observations Post ×I(%Black ≥ 0.5) Post ×I(%Black ≥ 0.5)

UOF

Arrest

911 Call Response Time

SIA

911 Crime Reporting

-0.32*** (0.07) 0.08 (0.07)

-4.02*** (1.22) -2.87** (1.41)

-0.36 (0.51) -0.89* (0.45)

-66.22*** (15.66) -56.68*** (15.29)

3.45*** (1.25) 7.26*** (1.26)

1272

1272

1272

1272

1272

-0.45*** (0.08) 0.08 (0.07)

-6.86*** (1.90) -2.94** (1.44)

-0.65 (0.69) -0.84* (0.45)

-81.16*** (18.03) -56.02*** (15.31)

6.61*** (1.76) 7.34*** (1.26)

1166

1166

1166

1166

-0.33*** (0.08) 0.07 (0.07)

-5.15*** (1.07) -3.84*** (0.93)

-0.74 (0.64) -1.42** (0.62)

-74.40*** (16.84) -60.95*** (17.21)

972

972

972

-0.31*** (0.09) 0.16* (0.08)

-5.36*** (1.18) -4.33*** (1.13)

-0.89 (0.70) -1.57** (0.69)

804

804

804

804

804

-0.29*** (0.06) 0.03 (0.06)

-3.81*** (1.23) -2.12 (1.43)

-0.53 (0.46) -1.15** (0.45)

-54.29*** (13.23) -39.15*** (14.37)

3.48*** (1.09) 7.28*** (1.11)

Homicide

Robbery

Panel 1. Baseline 0.11** 0.22 (0.04) (0.15) 0.12*** 0.30** (0.04) (0.13) 1272 1272 Panel 2. Dropping 2014Q3 0.11* 0.33* (0.06) (0.18) 0.12*** 0.28** (0.04) (0.12)

Aggravated Assault

Burglary

Larceny

Auto Theft

1.09*** (0.31) 0.77*** (0.24)

0.16 (0.34) -0.69** (0.27)

-0.92 (0.70) 0.14 (0.47)

-0.53** (0.25) 0.13 (0.18)

1272

1272

1272

1272

0.98** (0.41) 0.75*** (0.24)

-0.07 (0.37) -0.65** (0.27)

-1.71* (0.90) 0.12 (0.47)

-0.66** (0.31) 0.15 (0.18)

1166

1166

1166

0.35 (0.33) -0.80*** (0.29)

-0.38 (0.73) 0.40 (0.56)

0.15 (0.22) 0.12 (0.21)

972

972

-0.27 (0.87) 0.38 (0.60)

-0.18 (0.21) 0.08 (0.24)

804

804

804

-0.03 (0.31) -0.49* (0.29)

-1.02 (0.62) -0.14 (0.43)

-0.59** (0.24) 0.05 (0.17)

1166 1166 1166 1166 Panel 3. Control Group (%Black ≤ 20%) 3.46** 0.14*** 0.37** 1.28*** (1.35) (0.05) (0.15) (0.33) 6.97*** 0.11*** 0.43*** 0.87*** (1.42) (0.04) (0.13) (0.25)

972 972 972 972 972 972 Panel 4. Treatment Group (%Black ≥ 80%), Control Group (%Black ≤ 20%) -82.86*** 3.66** 0.19*** 0.24 1.41*** 0.24 (18.46) (1.39) (0.06) (0.17) (0.40) (0.39) -70.50*** 7.94*** 0.17*** 0.38** 1.02*** -1.16*** (18.53) (1.27) (0.05) (0.15) (0.26) (0.28) 804 804 804 Panel 5. WLS (Weight: Tract Population) 0.08* 0.29** 1.05*** (0.04) (0.14) (0.30) 0.09** 0.23* 0.80*** (0.04) (0.12) (0.20)

Observations 1272 1272 1272 1272 1272 1272 1272 1272 1272 1272 1272 Year-by-Quarter Fixed Effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Census Tract Fixed Effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Time-Varying Tract Covariates Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Contemporaneous Crime Rates Yes Yes Notes: Each column in each panel represents a separate regression. The unit of observation is tract-year-quarter. Robust standard errors are clustered at the tract level. Tract covariates include white population share, percentage of males, percentage of population aged 25 – 64, percentage of population with a high school diploma, percentage of population with a bachelor’s degree, poverty rate, and median household income. Contemporaneous crime rates include homicide, robbery, aggravated assault, burglary, larceny, and auto theft rates. * Significant at the 10% level ** Significant at the 5% level *** Significant at the 1% level

59

Online Appendix Table A1. Summary Statistics for Treatment and Control Tracts

Panel 1. Outcomes (per 1,000 population) UOF Arrest 911 Call Rsponse Time SIA 911 Cirme Reporting Homicide Robbery Aggravated Assault Burglary Larceny Auto Theft Observations Panel 2. Tract Covariates Population % Black % White % Male % Aged 25-64 % High School Diploma % College Degree Poverty Rate Median Household Income ($) Observations

60

Black Communities

White Communities

1

2

0.72 (0.76) 18.71 (19.67) 13.50 (4.17) 210.89 (151.34) 70.52 (26.47) 0.27 (0.44) 2.05 (1.63) 4.74 (3.22) 4.80 (2.79) 10.65 (6.51) 3.38 (1.71)

0.33 (0.54) 12.26 (25.51) 13.29 (3.41) 144.61 (117.71) 38.62 (27.26) 0.04 (0.14) 0.86 (1.06) 1.25 (1.39) 2.39 (1.75) 10.43 (9.49) 2.29 (1.70)

636

636

2,547.03 (1,107.24) 86.60 (15.69) 10.73 (13.27) 46.58 (5.59) 50.55 (7.16) 30.79 (7.02) 8.49 (5.19) 38.74 (12.49) 24,621.16 (8,617.98)

3,462.60 (1,189.51) 19.82 (13.53) 71.07 (16.75) 49.68 (4.43) 62.11 (9.89) 20.63 (8.53) 22.24 (8.73) 19.76 (11.53) 43,369.26 (12,373.82)

159

159