Armed with Technology: The Effects on Fatal Shootings of Civilians by the Police Min-Seok Pang and Paul A. Pavlou Fox School of Business, Temple University minspang@temple.edu, pavlou@temple.edu July 14, 2016 Abstract The police in the United States shot and killed 986 civilians in 2015. Deaths of civilians by the police in recent years have led to protests and disruptions in several large cities, such as New York, Chicago, and Baltimore. In this study, we investigate how the use of technology by the police affects use of lethal force on civilians. Drawing upon signal detection theory, we propose a simple, stylized model on a police officer’s decision to shoot. This model posits how the use of technology for intelligence access (e.g., statistical analyses of crime data) and evidence gathering (e.g., wearable body cameras) affects use of deadly force on civilians by the police. Empirical investigations with a large-scale dataset on fatal shootings revealed both encouraging and surprising findings. We found that both the use of smartphones and the statistical analyses of crime data are associated with a decrease in deadly shootings. In contrast, the use of wearable body cameras is related to an increase in the deaths of civilians by the police, contrary to an intuitive expectation that the adoption of body cameras would prevent deadly shootings. Interestingly, we also found that the observed effect of technology use is more pronounced for African Americans or Hispanics than Whites or Asians and for armed suspects than unarmed ones. We contribute to the literature by demonstrating the farreaching role of technology use in novel contexts, specifically in highly risky, violent environments. Keywords: Fatal Shootings, Use of Force, Intelligence, Wearable Body Cameras, Police, Signal Detection Theory -1- 1. Introduction On October 20, 2014, police officers responded to reports of a break-in at a Southside Chicago neighborhood. In pursuit of the suspect, one of the responding officers shot and killed a 17-year-old African American male who was reportedly carrying a knife. The video camera in a nearby patrol vehicle recorded this incident, and the victim’s family members requested to make the recording public in a lawsuit against the City of Chicago. The city had been refusing to do so, at the time that the mayor was running for a re-election. However, a judge ordered the city to release the video, which revealed that the officer shot at the victim 16 times. The release of the video erupted the African American community in Chicago, protesting over police brutality and demanding resignation of the Mayor of Chicago. The officer was charged for a first-degree murder, and the police commissioner was dismissed as well. The mayor also announced several reforms to his police department, including the expanded use of Tasers and wearable cameras (The New York Times 2016a). The relationship between the police and the public in many U.S. cities is more acrimonious than ever before. According to the Washington Post, 986 civilians were shot and killed by on-duty police officers in 2015 in the U.S. (The Washington Post 2016a). Several high-profile use-of-force incidents by police officers have sparked fierce protests and community upheavals in several cities such as Chicago, Baltimore, St. Paul, Baton Rouge, and Ferguson. This led to intense scrutiny from the media and the public on police practices and policies on the use of force. At the same time, however, citing a recent rise in violent crime rates, some law enforcement officials including the Director of the Federal Bureau of Investigation (FBI) have raised a concern that a so-called “Ferguson effect” makes police officers reluctant to enforce laws aggressively (CNN 2015), undermining the effectiveness of policing. Some officials even attributed a recent rise in crime rates to the Ferguson effect (The New York Time 2016b). These remarks have renewed an ongoing debate on how to strike a balance between enforcing laws for public safety and protecting the rights and lives of citizens (Hahn and Jeffries 2003, Dantzker 2010). -2- Policymakers and law enforcement officials are considering the use of digital technologies as a means to resolve this dilemma by improving policing capabilities (Manning 2010, Pang and Pavlou 2016). For instance, data analytics has become an essential tool for the police to gain necessary intelligence for crime solving (Chan 2004, Skogan and Frydl 2004, Bachner 2013). It is also a pillar for predictive policing, an initiative in which the police predict and deter crimes before they take place (Pearsall 2010, Haberman and Ratcliffe 2012). Police officers in the field are increasingly equipped with a range of advanced technologies such as mobile devices, gunshot detection systems, and license plate readers (Ariel et al. 2015, Government Technology 2015, The Baltimore Sun 2016). Also, video cameras mounted on police vehicles (dash-cams) or worn by officers (wearable body-cams) are expected to improve the accountability and transparency of policing by providing evidence of encounters with citizens (U.K. Home Office 2007, Harris 2010, Harvard Law Review 2015). To the best of our knowledge, however, we do not have sufficient understanding of the effectiveness of these technologies, particularly in terms of the use of lethal force on civilians. Police officers have every reason to avoid killing civilians, whether or not they are criminal. Whenever an officer-involved death occurs, a police department or an external law enforcement agency (e.g. a state police or a prosecutor’s office) has to spend considerable costs and manpower to investigate the incident. The police officers themselves have to undergo an intense internal-affairs investigation and close scrutiny from the media. Whether or not the shooting is deemed justified, the municipality could face a lawsuit from the family of the deceased, which could take several years to conclude in the courts or have to pay a substantial compensation to settle. After the shooting-death incident mentioned above, the City of Chicago agreed to pay the victim’s family $5 million in a settlement (The New York Times 2016a). The City of Cleveland also recently settled with the family of a 12-year-old shooting victim for $6 million (The Washington Post 2016b). Besides, a civilian death by the police deteriorates the relationship of the police with the public, thus damaging the effectiveness of law enforcement (Thurman and Reisig 1996, Skolnick and Barley 1998). Against this backdrop, we investigate how police technology use for -3- intelligence analyses and access (e.g., smartphones) and evidence gathering (e.g. wearable body cameras) affects the use of lethal force and deaths of civilians (Table 1). Intelligence Analyses Intelligence Access Evidence Gathering Table 1. Police Technology Use Records Computerized records of criminal incidents - Narrative descriptions of offenses - State statutes or municipal offense codes - Victim characteristics - Suspect characteristics - Offense location - Offense date and time Statistical Analyses If a police department conducts research or statistical analyses using computerized records Access Information that patrol officers have direct access from the field - Motor vehicle records - Driver license records - Criminal history - Outstanding warrants - Protection orders - History at address Smartphone If police officers in the department use smartphones Camera mounted on patrol vehicles (dash cams) Wearable body cameras (body cams) To theorize the role of technology use, we draw upon signal detection theory (Green and Swets 1988, Wicknes 2002, MacMillan 2002, Correll et al. 2002, 2014) to propose a simple, stylized model for a police officer’s decision to pull the trigger. We model that when deploying deadly force, the officer takes two factors into consideration – (i) a risk that a suspect poses an imminent, life-threatening danger to bystanders and/or the officer herself and (ii) a perceived risk that she would be held accountable for the death. Based on this model, we derive how technology use influences fatal shootings by police officers. First, technology use for intelligence analyses and access help reduce the ambiguity in the degree of violence of the suspect perceived by an officer. Second, the use of evidence gathering technologies, such as wearable body cameras, is likely to help the officer justify her shooting, making her less reluctant to deploy fatal force. -4- We test these predictions with a novel dataset that integrates multiple data sources. The unit of analysis in this study is a local police department in the U.S. We collected the records of civilian deaths by police shootings in 2015 from the Washington Post. In addition to these data, we acquired additional police homicide records in 2013-2014 from killedbypolice.net and the FBI. We also obtained information on police technology use from the Law Enforcement Management and Administration Survey (LEMAS) published by the U.S. Department of Justice. We built a large-scale dataset of 2,652 local police departments across the U.S. The dependent variable is the number of civilians shot and killed by officers in 2015. As our identification strategy, we adopted a spatial-autoregressive model to account for unobserved local heterogeneity (Arraiz et al. 2010, Drukker et al. 2013). We controlled for a variety of indicators for crime rates, police operations, and demographic and socioeconomic conditions. Our empirical analysis produced several interesting findings. First, we found that in police departments that conduct statistical analyses of digitized crime data, there are 2.15% fewer fatal shootings, substantiating our theoretical prediction that criminal intelligence can prevent police officers from using lethal force. Similarly, the use of smartphones by officers for intelligence access is related to 2.72% fewer deadly shootings. We obtained similar results from the alternative data from killedbypolice.net and the FBI. Surprisingly, we found that the use of wearable video cameras is associated with a 3.64% increase in shooting-deaths of civilians by the police. We explain that video recordings collected during a violent encounter with a civilian can be used in favor of a police officer as evidence that justifies the shooting. Aware of this evidence, the officer may become less reluctant to engage in the use of deadly force. We conducted more in-depth analyses with incident circumstances (e.g. whether a subject was armed) and demographics of victims (e.g. race, age), and we obtained more intriguing findings. Notably, the above-mentioned effect of technology use on fatal shootings is more pronounced for (a) African American or Hispanic victims than Whites or Asians and (b) for armed suspects than unarmed civilians. We contribute to the Information Systems (IS) literature by showing the significant impacts of technology use in highly uncertain and violent encounters. To the best of our knowledge, this study is the -5- first to theorize how technology use shapes human behaviors in a risky, life-or-death setting. To do so, we adopted a new interdisciplinary approach that integrates theories from multiple disciplines – IS, criminology, and psychology. We also contribute to the literature by uncovering the nuanced effects of technology use that vary by race, gender, and age. By doing so, this study extends the literature on the societal impacts of IS (e.g. Chan and Ghose 2016, Greenwood and Wattal 2016, Jha et al. 2016, Venkatesh et al. 2016) and uniquely contribute by venturing out from the comfort zone in the business sector and tacking one of the most contentious societal challenges in the U.S and other countries. This study offers crucial implications for policymakers and practitioners in law enforcement. In response to nationwide attention on the police use of lethal force, a number of police departments are considering increased use of wearable body cameras, hoping that this approach will ultimately reduce deaths of civilians by the police (Ariel et al. 2015). We provide empirical evidence that demonstrates otherwise; the use of body cameras by officers is associated with more deaths of civilians. Our research informs policymakers in law enforcement that the use of technology for intelligence analyses and access can be more effective in preventing civilian deaths, particularly for African Americans and Hispanics. 2. Prior Work In the IS literature, there is a dearth of research on crimes and violence. Among the few studies, Chan et al. (2016) found that the introduction of broadband Internet to a locality was associated with an increase in hate crimes. Greenwood and Wattal (2016) found that the entry of a ride-sharing service led to a decrease in alcohol-involved vehicular homicides. Based on the IT productivity literature, Garicano and Heaton (2010) expected to find positive effects of IT use on police performance measured by crime occurrence and clearance rates, but they only found insignificant effects. Hekim et al. (2013) also did not find a significant impact of IT use on crime clearance rates. A few prior studies in the criminology and public administration literatures have examined how technology use influences police practices, policies, and outcomes. For example, Nunn (2001) found that more computer use in police departments was -6- associated with an increase in operational expenditures and a decrease in the size of the police force. In a qualitative case study, Chan (2001) found that IT use was associated with improved transparency in interactions with citizens and increased use of data by police officers. To the best of our knowledge, our study is the first to investigate the impact of technology use on fatal shootings by the police. The criminology literature has been interested in analyzing police officers’ decisions to shoot. Prior studies extensively drew upon signal detection theory from psychology to model this decision (e.g. Correll et al. 2002, 2006, 2007a, Kenworth et al. 2011, Ma and Correll 2011, Akinola and Mendes 2012, Sandler et al. 2012, Sim et al. 2013, Ma et al. 2013). These studies conducted experiments with videogames that simulated a confrontation with a criminal suspect (e.g. Correll et al. 2002). In a 2-by-2 manipulation, a suspect was shown to be either Black or White and either armed with a gun or unarmed. Student participants were asked to press a button to shoot if the target appeared to be armed. The results demonstrated substantial racial bias by the participants in shooting decisions. Specifically, they shot unarmed Black targets more frequently than unarmed Whites (false alarms) and missed to shoot armed White targets more frequently than armed Blacks (misses) (Correll et al. 2002). Subsequent studies produced substantiating but more nuanced findings (e.g. Correll et al. 2006, 2007b, Ma and Correll 2011, Kenworthy et al. 2011, Sadler et al. 2012, Ma et al. 2013). For instance, Correll et al. (2007a) and Sim et al. (2013) showed that actual police officers demonstrated racial bias in shooting decisions to a lesser extent than untrained civilian participants. The criminology literature has also examined which factor influences police use-of-force incidents using archival datasets (e.g. Jacobs and O'Brien 1998, Alpert and MacDonald 2001, Smith 2003, 2004, McElvain and Kposowa 2008). For instance, Jacobs and O'Brien (1998) showed that killings by police officers in large U.S. cities are affected by crime rates (murder), divorce rates, and the racial makeup of population. McElvain and Kposowa (2008) found that the characteristics of police force such as gender, age, and educational attainment are significantly related to the deaths of criminals by the police. Our study differs from prior literature in that ours is the first to investigate the impact of technology use by the police on civilian deaths. In addition, prior studies use police homicide data from the FBI, which is -7- considered to be incomplete, as we explain below. In contrast, this study uses a more comprehensive civilian-death dataset from independent sources (the Washington Post and killedbypolice.net). In sum, our theoretical model draws upon and integrates research streams from three distinct disciplines – the impact of technology on law enforcement, the behavioral model of police use of force, and the antecedents of crimes and violence. 3. Theoretical Model On April 23, 2015, the Long Beach, California, Police Department responded to a report of trespassers in a vacant apartment. During an encounter with suspects, one officer fatally shot and killed a 19-year-old Hispanic male. According to the officer’s account, while turning toward him, the suspect was bending his knees and extending his arm to grab what appeared to be a gun. In fear of his life, the officer fired at the suspect three times. A subsequent investigation, however, found no weapon on the suspect (Long Beach Press-Telegram 2015). In order to model such a confrontation between a police officer and a potentially violent suspect in a highly uncertain, life-threatening circumstance, we introduce a simple, stylized model that depicts the officer’s decision to deploy a lethal weapon. According to signal detection theory, the officer’s decision can be modelled as a yes-or-no decision under uncertainty (Correll et al. 2002, 2006, 2007a). Such decisions include diagnosing a tumor by a doctor, predicting an earthquake by a seismologist, calling whether a ball is a strike by a baseball umpire, or deciding whether a person is guilty or not by a juror (MacMillan 2002). 3.1. The Model Facing with a potentially violent suspect, a police officer must detect whether or not he is armed with lethal weapon that could pose life-threatening danger to others and decide whether to overpower him with deadly force. In such an uncertain and risky situation, she should make this decision within a split second. If the suspect is indeed armed and she fails to neutralize him, she herself or other bystanders -8- could be in grave danger. A risk to the officer, however, is that he could be unarmed, and mistakenly killing an unarmed suspect could result in punishment or criminal liability against the officer. This is a major, life-or-death, decision that needs to be made under extremely stressful conditions. Figure 1. Shooting Decision by Signal Detection Theory (a) Higher Uncertainty (b) Lower Uncertainty -9- Signal detection theory explains that when confronted with the suspect, the officer receives an ambiguous sign on threats and has to discern whether it is a noise (e.g. the absence of threat) and a correct signal (the presence of threat). For instance, in the example of Long Beach shooting introduced above, the officer had to determine whether the suspect was trying to grab a gun (signal) or something else that is non-threatening (noise), and it turned out that the officer made an incorrect decision that the suspect was armed. Following signal detection theory, both perceived noise (N) and perceived signal (S) are assumed to be stochastic; specifically, in line with the theory, we assume that N and S follow a normal distribution. 𝑁~N(0, σ) and 𝑆~N(1, σ) (1) In this distribution, the variance (σ) represents the uncertainty the officer faces in distinguishing noise and signal. As shown in Figures 1-(a) and 1-(b), when σ is lower, it becomes easier for the officer to determine whether the suspect is life-threatening or not. Signal detection theory puts forth that as shown in Figure 1, the officer decides to pull a trigger when the cue she perceives is stronger than a certain criterion λ. It results in one of the four outcomes as in Table 2. Table 2. The Four Possible Outcomes of a Shooting Decision Noise (Non-Threatening) Signal (Threatening) Not Shooting Shooting Correct Reject 𝑃(𝑁 < 𝜆) Miss 𝑃(𝑆 < 𝜆) False Alarm 𝑃(𝑁 > 𝜆) Correct Hit 𝑃(𝑆 > 𝜆) In deciding to pull the trigger, the officer takes into consideration a tradeoff between (i) immediate dangers to lives of bystanders and herself and (ii) risks of becoming liable for her use of lethal force. Thus, the officer’s interest in this situation is to minimize both a miss, which could lead to deaths of the innocent and herself, and a false alarm, which could result in a death of the unarmed suspect that the officer could be held responsible for. The dilemma to the officer is that too low a criterion λ carries a high possibility of a miss, while too high a λ carries a high likelihood of a false alarm. We assume that the officer chooses the optimal λ that minimizes the following risk function. - 10 - min 𝑅 = 𝑃(Miss) + 𝑤 𝑃(False Alarm) = 𝑃(𝑆 < 𝜆) + 𝑤 𝑃(𝑁 > 𝜆) 𝜆 (2) In other words, the officer minimizes the weighted sum of the odds of a miss and a false alarm. We can reasonably assume 0 < 𝑤 < 1, since it is more important for the officer to protect the lives of the innocent. The first order condition is given by 𝜕𝑅 𝜕𝜆 Solving 𝜕𝑅 𝜕𝜆 = 1−𝜆 1 − 2 2𝜎 (𝑒 √2𝜋𝜎 − 𝑤𝑒 − 𝜆 2𝜎2 ). (3) = 0 produces 1 𝜆∗ = 2 + 𝜎 2 log 𝑤. (4) The second-order condition confirms that λ* minimizes R. 𝜕2 𝑅 𝜕𝜆2 𝜆=𝜆∗ = 1+4𝜎2 (log 𝑤)2 √𝑤 − 8𝜎2 𝑒 √2𝜋𝜎 3 >0 (5) 3.2. The Impact of Technology Use for Intelligence Analyses and Access We posit that the use of technology for intelligence analyses and access reduces perceived ambiguity (σ), and we examine how a decrease in σ affects the officer’s decision to shoot. Lemma 1. (The Effect of Uncertainty on the Shooting Criterion) 𝜕𝜆∗ 𝜕𝜎 = 2σ log 𝑤 < 0 since 0 < 𝑤 < 1. Lemma 1 indicates that the more uncertain the cues (σ) are, the lower criterion (λ*) the officer uses to shoot. Figures 1-(a) and 1-(b) illustrate this effect. When the signal and the noise are more uncertain, there is a greater risk of a miss (failing to shoot an armed and dangerous suspect), causing the officer to shoot more liberally with a lower λ*. On the other hand, when the officer becomes more certain in discerning between perceived signal and noise, she uses a higher criterion and deploys force more conservatively. - 11 - Figure 2. The Impact of Uncertainty on the Likelihood of Shooting (a) False Alarm (b) Correct Hit Lemma 2. (The Effect of Uncertainty on the Likelihood of Shooting) (a) (False Alarm) (b) (Correct Hit) 𝜕 𝑃(𝑁 𝜕𝜎 𝜕 𝑃(𝑆 𝜕𝜎 ∗) >𝜆 ∗) >𝜆 = = 1−2𝜎2 log 𝑤 − 𝑒 2√2𝜋𝜎 2 (1+2𝜎2 log 𝑤) 8𝜎2 1+2𝜎2 log 𝑤 − − 2 2𝜋𝜎2 𝑒 √ 2 > 0. 2 (1−2𝜎2 log 𝑤) 2 8𝜎 > 0 if and only if 𝜎 > 1 √−2 log 𝑤 or 𝜆∗ < 0. Lemma 2-(a) shows that a lower uncertainty (σ) monotonically leads to a less likelihood of shooting due to a false alarm (Figure 2-a). On the other hand, Lemma 2-(b) shows that a decrease in σ is associated with fewer correct hits only when σ is high, as illustrated in Figure 2-(b). When uncertainty (σ) is sufficiently low, on the other hand, the officer can easily distinguish between the absence (noise) and the presence (signal) of threats, and thus, a decrease in σ leads to more correct hits. We theorize that the use of technology for intelligence analyses and access reduces the ambiguity (σ) that a police officer faces when confronted with a violent suspect. For instance, with intelligence access technologies, such as smartphones or in-vehicle laptops, the officer is able to obtain actionable intelligence about the potential offender, such as criminal history, behavioral characteristics, types of weapons, or his modus operandi, real-time and on the fly. Intelligence access technologies can also help the officer better determine whether the suspect poses a life-threatening danger to her or others. Moreover, statistical analyses of digitized crime data provide police officers with a range of credible intelligence, such as crime prediction, behavioral patterns of organized criminals, or movement of weapons and drugs. Such technologies provide intelligence at a granular level, such as neighborhoods or - 12 - blocks, enabling the police to conduct more targeted policing (Pearsall 2010, Haberman and Ratcliffe 2012). Accordingly, if intelligence suggests that the suspect is unlikely to be dangerous, the officer would refrain from deploying lethal force against him. With intelligence that the suspect is likely to be armed and dangerous, the officer can coordinate with fellow officers to overwhelm him with backup support and prevent him from resisting violently. This leads us to propose: Proposition 1. (The Impact of Technology Use for Intelligence Analyses and Access) (i) The use of intelligence analyses and access technology is associated with a reduction in the shootingdeaths of unarmed targets (Lemma 2-a). (ii) If technology use for intelligence analyses and access reduces the uncertainty (σ) moderately, it would lead to a decrease in the shooting-deaths of armed targets (Lemma 2-b). (iii) But if technology use for intelligence analyses and access further reduces σ substantially, it would lead to an increase in the fatal shootings to armed targets (Lemma 2-b). A rationale for Proposition 1-(iii) is that when σ is sufficiently low, police officers would be able to easily distinguish threatening targets from non-threatening civilians. In the following sections, we will empirically examine to what extent technology use for intelligence affects deaths of civilians. 3.3. The Impact of Evidence Gathering Technology Here, we put forth that the use of technology for evidence gathering decreases the weight (w) in the risk function (Eq. 2). We explain why this would be the case below. Lemma 3. (The Impact of Weight on the Shooting Criterion) 𝜕𝜆∗ 𝜕𝑤 Lemma 4. (The Impact of Weight on the Likelihood of Shooting) (a) (Correct Hit) 𝜕 𝑃(𝑆 𝜕𝑤 𝜕 ∗) >𝜆 = 𝜎 − − 𝑤 2𝜋 𝑒 √ (b) (False Alarm) 𝜕𝑤 𝑃(𝑁 > 𝜆∗ ) = − 𝑤 (1−2𝜎2 log 𝑤) 8𝜎2 𝜎 − 𝑒 √2𝜋 2 <0 2 (1+2𝜎2 log 𝑤) 8𝜎2 - 13 - <0 = 𝜎2 𝑤 >0 Lemma 3 suggests that when a police officer weighs the risk of a false alarm more importantly (i.e. w is high), she uses a higher criterion (λ*), pulling the trigger more conservatively. According to Lemma 4, this higher λ* always leads to lower chances of shooting both threatening and non-threatening suspects (correct hits and false alarms, respectively). Lemma 4 suggests that the use of evidence gathering technologies, such as video cameras mounted on patrol vehicles (dash-cams) or worn by officers (body-cams), increases fatal shootings by the police. With video cameras, the officer would believe that the video recordings are likely to help justify her use of force (Harris 2010), reducing the perceived risk of false alarms (w). For instance, in Scott v. Harris (550 US 372), the U.S. Supreme Court ruled in an 8-1 decision that the use of deadly force by Sheriff’s Deputy Harris in Georgia was justified, and the Justices used the video recording from the defendant’s patrol vehicle as key evidence in their decisions (The U.S. Supreme Court 2007), creating a national precedent for use of force litigations with video recordings. Kahan et al. (2009) conducted an experiment with this case and found that 74% of the participants who watched the same video agreed that the deputy appropriately used force to the subject who posed danger to the public. Ariel et al. (2015) found in a randomized field experiment that the use of wearable body cameras is associated with a 92% reduction in the number of use-of-force complaints per officer from citizens. Jennings et al. (2015) also obtained a similar finding and also found that the majority of the officers who worn body cameras perceived the cameras to be helpful in evidence collection and recollection of events. Consequently, in her risk calculation, the officer will put more emphasis on preventing misses (i.e. a lower λ*), which according to Lemma 4, lead to more use of deadly force against the criminal target. Proposition 2. The use of evidence gathering technology is associated with an increase in shootingdeaths of civilian targets. In sum, our stylized model based on signal detection theory provides the following propositions. First, the use of technology for intelligence analyses and access can reduce fatal shootings by police officers if it moderately reduces ambiguity (σ) that the officer faces in violent encounters. Second, with a - 14 - substantial decrease in σ, intelligence technology use can lead to an increase in fatal shootings. Third, the model predicts that the use of evidence gathering technologies is associated with an increase in the use of deadly force. We will empirically test these predictions in the sections to follow. 4. Research Methodology 4.1. Data and Measures We conducted empirical analyses with a dataset that combines information from a variety of public sources. The unit of analysis is local police departments in the U.S. (cities, townships or counties). The dependent variable is the number of civilians shot and killed by on-duty police officers in 2015. This data is collected from the Washington Post police-involved shooting database, which compiles the entire civilian shooting-death records in 2015 from news reports, social media, online sources, and other public records. The Washington Post obtained additional information from its own follow-up investigations and by filing Freedom of Information Act requests to law enforcement agencies. This dataset provides detailed, comprehensive information of shooting incidents such as a victim’s gender, race, age, and mental status and types of weapon that the victim used, all of which we use in our analyses below. The dataset does not include records of civilian deaths by off-duty police officers and nonshooting deaths (e.g. civilian deaths by vehicle accidents involved with patrol cars). We excluded incidents by federal and state law enforcement officials (e.g. the FBI, the Drug Enforcement Agency, or a state highway patrol) and specialized police agencies (e.g. park, transit, or school police). Carefully reading each shooting-death report, we identified to which department the police officer who killed a civilian belongs.1 1 For instance, suppose that a police officer in City A chased a suspect who fled to a nearby jurisdiction (City B) and ended up killing him in City B. We coded that the suspected is killed by the officer from City A, not City B. In a case that a suspect is killed by officers from multiple agencies, we conducted an extensive search to find out which officer killed the suspect, and if it is unclear, we excluded the case. - 15 - For additional analyses with homicides by the police in 2013 and 2014, we collected data from the two sources – (i) a Web site called killedbypolice.net and (ii) the Uniform Crime Reports (UCR) database2 maintained by the FBI. Like the Washington Post, killedbypolice.net uses news and Internet sources to build a compilation of civilian deaths by police officers. Nonetheless, we deem the data from killedbypolice.net to be less reliable, because the site is completely anonymous and does not disclose about who operates the site and builds the database, from which source it collects information, and which methodology it uses. The UCR Supplementary Homicide Reports from the FBI includes the number of “felons” killed by police officers (Smith 2004, McElvain and Kposowa 2008). However, this data source is considered to be incomplete, because it only includes the deaths of those who the police determined to be “felons”; in other words, it only reports incidents of police-caused homicides the department itself deemed justified (Five Thirty Eight 2014). Indeed, the UCR reports that the number of civilians killed by police in 2011-2013 is as many as 450 per year; on the other hand, the Washington Post reports that 986 citizens were killed by police in 2015 (The Washington Post 2016a). Consequently, we consider the Washington Post database to be more comprehensive, reliable, and objective than the two other sources. We obtained police IT data and other information on police operations from the Law Enforcement Management and Administrative Statistics (LEMAS) collected by the Bureau of Justice Statistics (the BJS) under the U.S. Department of Justice (DOJ). In 2013, the BJS conducted a survey with a random sample of 2,800 state and local law enforcement agencies. This dataset includes information on police personnel, operations, equipment, policies, and technology use. Combing the data from the Washington Post and the LEMAS, we built a cross-sectional dataset of 2,657 local police departments. Table 3 describes the profiles of police departments in the sample. Our sample covers both large and small cities and both urban and rural areas across the U.S. According to ttests, the departments in the sample are not significantly different from others with respect to population 2 At the time of our data collection in March 2016, the UCR database for 2015 was not available. - 16 - or crime rates. They are also nationally representative in terms of median income, racial make-up, and other demographics. Table 3. Profiles of Police Departments Population Number of Departments > 1,000,000 40 500,000 – 1,000,000 76 200,000 – 500,000 173 100,000 – 200,000 259 50,000 – 100,000 396 20,000 – 50,000 527 10,000 – 20,000 380 < 10,000 806 Urban/Rural Number of Departments Urban 1) 1,611 Rural 1,046 Share of White Population Number of Departments > 90% 841 70% - 90% 1,104 50% - 70% 462 < 50% 250 Median Household Income Number of Departments > $70,000 372 $50,000 - $70,000 740 $30,000 - $50,000 1,349 < $30,000 196 1) Located in Metropolitan Statistical Areas with population over 200,000 Tables 4A describes the main variables. Technology use for intelligence analyses and access are measured as follows: Records is the number of categories in computerized data that a police department maintains (e.g. offense descriptions, suspect and victim characteristics, and offense locations). Statistical Analyses is a dummy variable that is equal to one if the police department conducted research or statistical analyses using computerized records of criminal incidents in 2013. Access is the number of data categories that patrol officers in the field have direct electronic access to (e.g. criminal histories, driver license records, and outstanding warrants). Smartphone is a dummy variable that is equal to one if officers use smartphones for data collection and access. For the use of evidence gathering technology, we used two variables – Dash Camera and Body Camera – which are equal to one if the police department uses - 17 - video cameras installed on patrol vehicles and worn by officers, respectively. Table 5 provides the descriptive statistics. Table 4A. Variable Definitions and Data Sources (Main Variables) Variable Definition Data Sources Dependent Variable Civilian Killed Number of civilians shot and killed by police department The Washington in 2015 Post Independent Variables – Police IT Use Records # of computerized crime data records categories Law Enforcement Statistical Analyses 1 if department conducts research or statistical analysis Management and with computerized crime data Administration Access # of crime data categories accessible by patrol officers Survey (LEMAS) Smartphone 1 if officers are equipped with smartphones from the BJS Dash Camera 1 if patrol vehicles are equipped with video cameras Body Camera 1 if patrol officers are equipped with video cameras Control Variables – Crime Occurrence and Clearance Crime Occurrence Log (# of crimes known to police in 2013-2014) Uniform Crime Crime Clearance Share of crimes cleared to crimes known in 2013-2014 Reports (UCR) Officer Assaults Log(# of officers killed and assaulted in the line of duty from the FBI in 2013-2014) Control Variables – Basic Locality Information Population Log (population) UCR Miles Log (square-miles covered by department) MSA Core City 1 = core city of metropolitan area; 0 = otherwise Control Variables – Police Operation Operational Budget Log (operational budget ($) per capita) LEMAS Education Educational requirements for new officers (1 = no Requirements requirement, 4 = high school or higher) White Officer Share of White officers to total officers with arrest powers Female Officer Share of female officers Weapon # of types of sidearm (e.g. 10mm) allowed in duty Special Units # of special units dedicated to certain crimes Community Policing # of community policing training hours - 18 - Table 4B. Variable Definitions (Demographic and Socio-Economic Conditions) Variable Definition Data Sources - American Community Survey from the U.S. Census Bureau Control Variables – Demographic Information Male Share of male population White Share of White-only population Young Share of young (15-24) population High School Share of population with a high school degree or higher Control Variables – Economic Conditions Income Median household income ($ thousand) Inequality Gini coefficient for income inequality Vacant Homes Number of vacant homes per capita Control Variables – Social Conditions Moved Share of population who moved within one year Female Household Head Share of female household heads (a single mother or a single grandmother) Two Parent Household Share of households with two parents 4.2. Identification Strategy We took several approaches to address potential endogeneity concerns. First, our estimation controls for state and MSA (metropolitan statistical area) fixed-effects to account for potential unobserved heterogeneity in terms of local economic conditions or regional law enforcement practices. Second, regional unobserved heterogeneity could be correlated with each other among neighboring localities. One possible source is spatial contagion of crimes and violence. As criminals are unlikely to restrict their behaviors into a single local area or jurisdiction, unobserved factors in one municipality are likely to be correlated with those in neighboring municipalities. In addition, unobserved economic and societal circumstances or policing practices could be similar among jurisdictions in one region. In order to account for this possibility, we adopted generalized spatial two-stage least-square (GS2SLS) estimators for spatial-autoregressive residuals (Kelejian and Prucha 2010, Arraiz et al. 2010, Drukker et al. 2013). To do so, we estimated the following specification. - 19 - Variable Civilian Killed Records Statistical Analyses Access Smartphone Dash Camera Body Camera Crime Occurrence Crime Clearance Officer Assaults Population Miles MSA Core City Operational Budget Education Requirements White Officer Female Officer Weapon Special Units Community Policing Male White Young High School Income Inequality Vacant Homes Moved Female Household Head Two Parent Household Table 5. Descriptive Statistics (N = 2,657) Mean Std. Dev. Minimum (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) (16) (17) (18) (19) (20) (21) (22) (23) (24) (25) (26) (27) (28) (29) (30) 0.2149 6.9289 0.6692 4.3963 0.6282 0.7166 0.2744 6.7237 0.3050 1.0349 10.1265 3.3963 0.1656 4.8754 1.2303 0.8253 0.0930 9.4919 3.3007 1.3429 0.4889 0.7794 0.1401 0.2026 51.3251 0.4403 0.0671 0.1678 0.0542 0.1497 𝐲 = 𝐗β + 𝐮 and 𝐮 = 𝜌𝐌𝐮 + 𝛜 0.9031 1.7311 0.4706 2.1841 0.4834 0.4507 0.4463 2.3912 0.1862 1.6426 1.6995 2.2579 0.3718 1.4601 0.7020 0.2381 0.0775 2.1031 3.8007 1.2651 0.0263 0.1819 0.0542 0.0637 20.1283 0.0481 0.1093 0.0626 0.0276 0.0454 Maximum 0 0 0 0 0 0 0 0 0 0 5.2781 -1.8420 0 0 0 0 0 0 0 0 0.3458 0.0017 0.0145 0.0140 15.1020 0.2762 0 0.0251 0 0.0106 21 8 1 6 1 1 1 12.8498 1.1906 7.5914 16.1144 9.9061 1 8.9258 4 1 0.6250 12 14 3 0.8163 1 0.7850 0.4842 207.2220 0.6311 3.0694 0.5833 0.2758 0.4375 (5) In this model, M is a spatial-weighting matrix in which non-diagonal elements are inverses of the Euclidean distance between two municipalities and the diagonal elements are zero. In other words, the correlation in unobserved factor (u) between two localities is inversely related to the distance between the two. This is a reasonable assumption; it is well-known in the public economics literature that in policy implementation and delivery of public services, neighboring local governments imitate each other (Besley and Case 1995, Figlio et al. 1999, Baicker 2005). Hence, jurisdiction-specific unobserved heterogeneity that may affect police use of force is likely to be spatially correlated. While our dataset is cross-sectional, - 20 - this GS2SLS approach enables us to tease out unobserved heterogeneity (u) from residuals by taking advantage of the distance among the jurisdictions. Monte Carlo simulations show that estimates from Equation 5 are consistent and exhibit very little bias even in small samples (Kelejian and Prucha 2010). This proposed method has widely been used by many recent studies in urban and regional economics (e.g. Holly et al. 2011, Breustedt and Habermann 2011, Van Duijn and Rouwendal 2013). Third, in order to control for observable heterogeneity of localities, the model includes several control variables that may affect police use of lethal force (Tables 4A and 4B). We controlled for the number of occurrences in violent and property crimes and the crime clearance rate (percentage of crimes cleared) in 2013 and 2014.3 The more frequently crimes occur or the more vigorously the police attempt to clear crimes, the more frequently they are likely to use deadly force (Alpert and MacDonald 2001, Jacobs and Carmichael 2002, McElvain and Kposowa 2008). We also controlled for the number of police officers killed and assaulted in the line of duty, since police officers are more likely to use lethal force when they are in danger. The data for crimes and officer deaths/assaults were obtained from the UCR. Other control variables also include basic locality information such as population, geographic size, and whether the department covers the core city in an MSA. Our estimation also controlled for several indicators for police operations and personnel that the criminology literature suggests to affect crimes and violence (Jacobs and O’Brian 1998, Adams et al. 2005, Kaminski 2008, McElvain and Kposowa 2008, Fridell et al. 2009, Kaminski and Stucky 2009, Cordner 2014). These variables include the size of operational budget, educational requirements for officers, the share of white and female officers, the types of weapons authorized to use, the number of specialized units, and community-oriented policing trainings. The data for these measures were obtained from the LEMAS (Table 4A). We also controlled for demographic, economic, and societal indicators in localities, utilizing data from the American Community Survey by the U.S. Census Bureau (Table 4B). The demographic indicators are the share of male, White, and young (15-34) and level of educational 3 Crime occurrence and clearance data in 2015 from the UCR is not available at the time of our data collection. - 21 - attainment. We use these variables to measure the share of most crime-prone population (e.g. Miethe et al. 1991, Baumer et al. 1998, Baller et al. 2006, Kaminski 2008, Kent 2010, Garicano and Heaton 2010). The economic measures are median household income, income inequality, and share of vacant homes, following the criminology literature that economic distress is significantly related to crimes and violence (e.g. Baumer 1994, Baumer et al. 1998, Batton and Wilson 2006, Garicano and Heaton 2010). Finally, social disorganization theory suggests that social instability is strongly related to crimes and violence (e,g, Land et al. 1990, Baumer et al. 1998, Baumer 1994, Baller et al. 2006). To measure social instability, we used societal indicators for population movement and household characteristics. Fourth, reverse causality could be a concern in this study; the use of video cameras could be motivated by frequent occurrences of police use-of-force incidents. In the following section, we present a series of robustness checks such as falsification tests and Heckman (1989) estimations for self-selection, in which we address endogeneity problems due to reverse causality. To do so, we utilize the data on criminal deaths by the police prior 2013 from the UCR. Fifth, measurement error in the dependent variable is unlikely to be a problem. We believe that the Washington Post provides the most comprehensive records of fatal shootings by the police. Even though there could be unreported incidents in use of deadly force, such cases are believed to be rare (Jacobs and O'Brien 1998), since it is very difficult and politically risky for the police to cover up officerinvolved homicide cases. Measurement errors in the technology use variables could be a concern, in that we regressed fatal shootings in 2015 on technology use in 2013. It is possible that technology use in 2015 could be different from 2013. To alleviate this concern, we will present estimations with alternative data sources such as killedbypolice.net and the UCR for civilian deaths in 2013 and 2014. - 22 - Dependent Variable Method Data Source Year Records Statistical Analyses Access Smartphone Dash Camera Body Camera Crime Occurrence Crime Clearance Officer Assaults Population Miles MSA Core City Operation Budget Education Req White Officer Female Officer Weapon Special Units Community Poli Male White Young High School Income Inequality Vacant Homes Moved Female Head Two Parent ρ Controls Wald χ2 Table 6. The Baseline Estimation Results Log(Civilian Killed + 1) Spatial Autoregressive Regression The Washington The Washington Post killedbypolice.net & killedbypolice.net Post 2015 2013-2014 2013-2015 (1) (2) (3) -0.0023 (0.0027) -0.0009 -0.0256***(0.0098) 0.0010 (0.0022) -0.0021 -0.0249** (0.0107) -0.0240* -0.0060 (0.0142) 0.0069 0.0358***(0.0138) -0.0018 0.0049 -0.0507** (0.0039) (0.0032) -0.0026 -0.0380***(0.0125) -0.0375** (0.0027) -0.0013 (0.0129) UCR 2013-2014 (4) (0.0038) -0.0026 (0.0147) (0.0032) -0.0031 -0.0404***(0.0149) (0.0028) -0.0363***(0.0099) (0.0023) -0.0016 (0.0112) (0.0163) -0.0045 (0.0190) -0.0048 (0.0148) (0.0156) (0.0179) (0.0144) 0.0126***(0.0038) 0.0173 0.0156***(0.0050) 0.0117 0.0079***(0.0028) (0.0255) -0.0857***(0.0317) -0.1082***(0.0371) -0.0864***(0.0255) 0.0482***(0.0089) 0.0657***(0.0112) 0.0815***(0.0121) 0.0664***(0.0113) 0.0382***(0.0080) 0.0563***(0.0102) 0.0739***(0.0117) 0.0283***(0.0078) 0.0027 0.0015 0.0035 (0.0033) (0.0043) (0.0050) -0.0075** (0.0034) 0.1620***(0.0255) 0.1820***(0.0308) 0.2698***(0.0348) 0.1866***(0.0287) 0.0102***(0.0029) 0.0131***(0.0037) 0.0176***(0.0042) 0.0075***(0.0029) 0.0078 (0.0126) 0.0049 (0.0097) (0.0254) (0.0651) 0.0139 (0.0090) -0.0001 (0.0108) -0.0079 (0.0253) -0.0230 (0.0336) -0.0246 (0.0393) -0.0475* -0.0418 (0.0532) (0.0728) (0.0825) 0.0485 0.0960 0.0126 -0.0113***(0.0025) -0.0100***(0.0031) -0.0127***(0.0036) -0.0093***(0.0027) 0.0104***(0.0018) 0.0123***(0.0023) 0.0162***(0.0026) 0.0090***(0.0020) 0.0026 0.0035 0.0004 0.0101** (0.0045) (0.0041) (0.0049) (0.0057) -0.1551 (0.1539) -0.1573 (0.1704) -0.2295 (0.2166) -0.1109 (0.1541) 0.0395 (0.0453) -0.0073 (0.0628) -0.2640** -0.0096 -0.0011** (0.0508) 0.0206 0.0305 (0.0447) (0.1244) -0.2396* (0.1393) -0.3503** (0.1686) -0.1932 (0.1240) (0.1274) 0.1511 (0.1528) 0.1132 (0.1854) -0.1056 (0.1268) (0.0004) -0.0001 (0.0005) -0.0007 (0.0006) -0.0010** (0.0005) 0.1683 (0.1204) 0.3472** (0.1499) 0.2851 (0.1740) 0.4662***(0.1353) 0.0285 (0.0318) 0.0214 0.0305 (0.0584) 0.0615** (0.0298) -0.1092 0.0785 -0.0137 (0.0476) (0.0998) -0.1372 (0.1337) -0.2028 (0.1504) -0.4155***(0.1106) (0.2031) (0.2602) (0.3060) 0.1602 (0.1556) -0.2647 -0.7704***(0.2369) 0.2626 (0.1773) -0.2969 -0.5307***(0.1994) 0.0367 (0.2232) (0.2161) -0.0127 (0.1529) -0.6212***(0.2157) State, MSA State, MSA State, MSA 6.5×106*** 2.0×107*** 1.3×107*** * p < 0.1, **p < 0.05, ***p < 0.01; N = 2,657; Robust standard errors are in parentheses. - 23 - -0.7844***(0.2644) State, MSA 1.3×107*** 5. Results 5.1. The Baseline Estimation Result Table 6 presents the baseline estimates with spatial autoregressive regressions. In all columns, ρ is statistically significant, indicating that unobserved heterogeneity is significantly correlated across adjacent municipalities. It suggests that our use of spatial autoregressive model is appropriate. The estimation with the Washington Post data in 2015 (Column 1) shows that the coefficients of Statistical Analyses and Smartphone are negative and statistically significant. It is predicted that in police departments that conduct statistical analyses with digital crime information, the number of fatal shootings by officers is 2.53%4 fewer than in others. Smartphone use by officers is also associated with a 2.46% reduction in the shooting-deaths of civilians. This is consistent with our theoretical prediction (Proposition 1) that crime intelligence obtained by statistical analyses and access to the intelligence via smartphones lead to a decrease in fatal shootings by the police. In contrast, the number of crime records (Records) and the categories of data accessible by patrol officers (Access) are not significantly related to the incidents of civilian deaths. Table 6, Column 1 also shows that the coefficient of Body Camera is positive and statistically significant. When officers wear video cameras on duty, they are 3.64% more likely to kill a suspect. Video cameras installed on patrol vehicles (Dash Camera) are not significantly associated with shootings of civilians by officers. This finding supports Proposition 2 that the use of evidence gathering technology (specifically wearable video cameras) reduces an officer’s perceived risk of becoming liable for her use of deadly force. Expecting that a wearable video camera would provide evidence to justify the use of force, the officer becomes less reluctant to deploy deadly weapons. Table 6, Columns 2-4 provide the consistent estimates with alternative data sources for police killings. In Column 2 with the killedbypolice.net data for 2013-2014, statistical analyses of crime data 4 1 – e-0.0256 = 0.0257 - 24 - and smartphone use by officers are associated with fewer deaths of civilians by the police. We found a similar finding from the estimation with a combine dataset of the Washington Post and killedbypolice.net in 2013-2015 (Column 3). The magnitude in the effects of statistical analyses and smartphone use is larger in Column 3 than in Column 1. Column 3 shows that smartphone usage is related to a 4% reduction in shooting-deaths. The coefficient of body cameras in Column 3 is again positive, albeit insignificant. Column 4 shows the estimation with the UCR dataset, that reports deaths of those who the police designate as felons. The coefficient of statistical analyses is also negative and significant. Table 6 shows that the impact of wearable video cameras use in 2013 is significant only in 2015 (Column 1). It appears that there is a learning effect in the use of wearable video cameras. It could take a while for police officers to realize how helpful evidence from body cameras can be in justifying the use of lethal force. Anecdotal evidence suggests that some police officers who were initially skeptical of the effectiveness of body cameras have later embraced them after they learned how the use of cameras could lead to a reduction in the use-of-force complaints from citizens (Ariel et al. 2015) and how video recordings are used in use-of-force cases as evidence to exonerate officers involved (The New York Times 2013, NBC News 2016). - 25 - Table 7. Estimation by Victim Characteristics Dependent Variable Method Data Source Subgroup Log(Civilian Killed + 1) Spatial Autoregressive Regression The Washington Post (2015) Armed with Gun (1) (3) (4) Black and Hispanic (5) Dash Camera Body Camera Crime Occurrence Crime Clearance Officer Assaults Population Miles MSA Core City ρ Controls Wald χ2 - 26 - -0.0023*** (0.0019)* -0.0227*** (0.0068)* 0.0004*** (0.0015)* -0.0169*** (0.0080)* -0.0004*** (0.0109)* 0.0368*** (0.0108)* 0.0046*** (0.0023)* -0.0674*** (0.0183)* 0.0345*** (0.0076)* 0.0199*** (0.0061)* 0.0020*** (0.0024)* 0.1135*** (0.0204)* -0.9707*** (0.3005) (6) -0.0013*** (0.0026)* -0.0235*** (0.0096)* 0.0009*** (0.0021)* -0.0250*** (0.0105)* -0.0044*** (0.0139)* 0.0393*** (0.0134)* 0.0042*** (0.0039)* -0.0468*** (0.0253)* 0.0457*** (0.0088)* 0.0395*** (0.0080)* 0.0014*** (0.0033)* 0.1631*** (0.0249)* -0.6783*** (0.2281)* Female (7) -0.0012*** (0.0009)* -0.0041*** (0.0023)* 0.0000*** (0.0005)* 0.0001*** (0.0029)* 0.0001*** (0.0041)* 0.0009*** (0.0041)* 0.0010*** (0.0007)* -0.0123*** (0.0066)* 0.0069*** (0.0022)* -0.0010*** (0.0016)* 0.0014*** (0.0008)* 0.0059*** (0.0072)* -1.4195** (0.6305) Age ≤ 31 (8) Age > 31 (9) State, MSA 1.7×107*** Smartphone -0.0014*** (0.0019)* -0.0115*** (0.0075)* 0.0008*** (0.0016)* -0.0083*** (0.0077)* -0.0009*** (0.0102)* 0.0067*** (0.0098)* 0.0000*** (0.0029)* 0.0145*** (0.0200)* 0.0239*** (0.0060)* 0.0268*** (0.0055)* -0.0003*** (0.0024)* 0.0717*** (0.0174)* -0.6928** (0.2998) Male State, MSA State, MSA State, MSA State, MSA State, MSA State, MSA State, MSA State, MSA 1.4×107*** 8.7×105*** 1.9×107*** 7.3×106*** 6.0×106*** 8.4×106*** 7.7×106*** 5.4×105*** * p < 0.1, **p < 0.05, ***p < 0.01; N = 2,657; Robust standard errors are in parentheses.; Other control variables are omitted for brevity. Access 0.0010*** (0.0007)* -0.0090*** (0.0036)* 0.0001*** (0.0007)* 0.0000*** (0.0037)* 0.0017*** (0.0053)* 0.0024*** (0.0049)* 0.0011*** (0.0012)* -0.0041*** (0.0093)* 0.0071*** (0.0036)* 0.0041*** (0.0029)* 0.0027*** (0.0011)* 0.0326*** (0.0104)* -2.1115*** (0.4449)* White and Asian -0.0015*** (0.0022)* -0.0219*** (0.0077)* -0.0017*** (0.0017)* -0.0083*** (0.0084)* 0.0009*** (0.0112)* 0.0181*** (0.0110)* -0.0002*** (0.0030)* -0.0032*** (0.0207)* 0.0314*** (0.0070)* 0.0284*** (0.0063)* 0.0016*** (0.0026)* 0.0784*** (0.0196)* -0.6898*** (0.2377)* Statistical Analyses -0.0007*** (0.0015)* -0.0123*** (0.0060)* 0.0000*** (0.0012)* -0.0066*** (0.0063)* -0.0089*** (0.0094)* 0.0187*** (0.0084)* -0.0008*** (0.0030)* -0.0274*** (0.0136)* 0.0231*** (0.0056)* 0.0198*** (0.0056)* -0.0023*** (0.0019)* 0.0696*** (0.0165)* -1.4049*** (0.3519)* Unarmed -0.0015*** (0.0016)* -0.0144*** (0.0064)* 0.0028*** (0.0014)* -0.0181*** (0.0077)* -0.0002*** (0.0101)* 0.0307*** (0.0096)* 0.0047*** (0.0025)* -0.0625*** (0.0174)* 0.0318*** (0.0067)* 0.0200*** (0.0055)* -0.0005*** (0.0021)* 0.1101*** (0.0186)* -1.5276*** (0.4749)* Records -0.0021*** (0.0021)* -0.0201*** (0.0077)* 0.0005*** (0.0017)* -0.0193*** (0.0087)* 0.0098*** (0.0108)* 0.0305*** (0.0111)* 0.0047*** (0.0026)* -0.0358*** (0.0212)* 0.0350*** (0.0071)* 0.0246*** (0.0058)* 0.0015*** (0.0026)* 0.0989*** (0.0196)* -0.5806*** (0.2015)* Armed with Other Weapons (2) 5.2. Additional Analyses with Victim Characteristics The Washington Post database provides several characteristics of shooting victims such as race, age, gender, and whether or not they were armed. We dug deeper with this information to find how the impact of technology use varies by victim characteristics and incident circumstances. Table 7, Columns 13 show that the impact of statistical analyses, smartphone, and wearable video cameras is more pronounced when the victims are armed with guns (Column 1) than otherwise (Columns 2-3).5 Compared to deaths of unarmed civilians (Column 3), the effect of technology use is stronger for civilians with lesslethal weapons such as knives or hatchets (Column 2). In other words, the more threatening a civilian target is, the stronger effect the technology use has on fatal shootings by the police. According to Proposition 1-(ii), it seems that the use of technologies for intelligence reduces the ambiguity (σ) that an officer faces during a dangerous encounter, but not to an extent that the officer fully discerns between noise and signal without much risk so easily (Proposition 1-iii, 𝜎 < 1 √−2 log 𝑤 ). This result illustrates that with statistical analyses and smartphone use, police officers are able to obtain realtime intelligence on potential suspects (e.g. how violent they would be or what kind of weapons they would use) and take appropriate measures to subdue them without lethal force. Consistently with Lemma 4-(a), with an expectation that video recordings would be used as substantiating evidence, police officers become less hesitant to deploy lethal force when they perceive the presence of deadly threats from suspects, leading to more correct hits (Column 1). While Lemma 4-(b) also predicts that the use of video cameras can also increase the likelihood of shootings to unarmed citizens (false alarms), its effect is found be positive but insignificant (Table 7, Column 3). Table 7, Columns 4-9 present intriguing and potentially disturbing findings. The coefficients of statistical analyses, smartphone, and video cameras are more negative and significant when shooting 5 The difference in the coefficients of smartphones and body cameras between Columns 1 and 3 is statistically significant (p < 0.05), but not for statistical analyses. - 27 - victims are African Americans or Hispanics (Column 5),6 male (Column 6), and younger than 31 years old (Column 8).7 Surprisingly, the impact of all three technologies is found to be insignificant for White and Asian deaths (Column 4), even though these two groups constitute 51.4% of the shooting victims by the police in 2015. Table A1 in Appendix A presents a similar result from the estimation that combines the data from the Washington Post and killedbypolice.net for 2013-2015, which corresponds to Table 6, Column 3. Note that while the coefficient of body cameras is insignificant in Table 6, Column 3, it becomes significant for African Americans and Hispanics (Table A1, Column 2). Prior literature in police shootings puts forth the existence of a racial prejudice in which African Americans and Hispanics are perceived to be more violent than Whites and Asians (e.g. Correll et al. 2007a, Salder et al. 2012). Correll et al. (2014) argued that people “interpret ambiguous behavior as more violent when the actor is Black rather than White” (p. 202). Correll et al. (2006) conducted neurological experiments and found that the participants perceive more threats from African Americans than from Whites, as measured by their eventrelated brain potentials. According to our theoretical model, when confronted with an African American or a Hispanic suspect, a police officer may perceive more uncertainty (σ) in distinguishing between signal and noise. Correll et al. (2002) found that the experiment participants use a lower criterion (λ) to shoot Black targets than Whites targets. Lemma 1 shows that the higher uncertainty (σ) the police officer faces, the lower 𝜕𝜆∗ criterion (λ) she uses in shooting ( 𝜕𝜎 < 0). From Lemma 1, 𝜕 2 𝜆∗ 𝜕𝜎 2 = 2 log 𝑤 < 0, indicating that the impact of uncertainty reduction on the criterion (λ) is greater when σ is higher. This explains that the uncertainty reduction effect from statistical analyses and smartphone use is stronger when the officer is confronted by a suspect who has stereotypical characteristics associated with violence (non-White, male, 𝜕 2 𝜆∗ and youth). Likewise, 𝜕𝜎𝜕𝑤 = 2σ w > 0, indicating that the positive impact of evidence gathering 6 The difference in the coefficient of body cameras between Columns 4 and 5 is statistically significant (p < 0.05), but not for statistical analyses and smartphones. 7 The median age of the civilians killed by the police in 2015 is 31. - 28 - technology use (wearable video cameras) on shooting criterion (Lemma 3) is magnified under higher uncertainty. This explains the higher coefficients of video cameras in Table 7, Columns 5, 6, and 8. 5.3. Robustness Tests We have conducted several robustness checks as follows. First, with respect to unobserved heterogeneity, we conducted falsification tests to rule out potential alternative explanations for police use of force. Specifically, the impact of technology use on fatal shootings by the police could be compounded by unobserved general trends that reduce crimes and violence. To address this concern, we ran our estimations with alternative dependent variables – the number of murders and manslaughters reported to police in 2013 and 2014, excluding homicides by the police. As the results in Table 8 show, the coefficients of technology use are estimated to be insignificant (Columns 1-2). In addition, while it is unlikely that the deaths of criminals in 2007-20118 would directly cause fatal shootings in 2015, we nonetheless included this variable as additional control variable. Table 9, Column 1 provides a very similar result to our baseline estimate in Table 6, Column 1. While we noted in the previous section that the UCR only reports the deaths of civilians who the police designate as felons, too many deaths of offenders in 2007-2011 could still motivate police departments to adopt technologies in 2012, particularly video cameras, and to begin using them in 2013. Table 9, Column 1 shows that controlling for the number of criminal deaths by the police prior 2013, our primary findings remain consistent. 8 We use different time frames for this analysis, such as 2009-2011 or 2010-2012, but did not obtain different results. - 29 - Dependent Variable Method Data Source Year Records Statistical Analyses Access Smartphone Dash Camera Body Camera Crime Occurrence Crime Clearance Officer Assaults Population Miles MSA Core City Operational Budget Education Req White Officer Female Officer Weapon Special Units Community Policing Male White Young High School Income Inequality Vacant Homes Moved Female Household Head Two Parent Household ρ Controls Wald χ2 Table 8. Falsification Tests Log (Murder + Log (Criminal Death Log (Murder + 1) Manslaughter + 1) + 1) Spatial Autoregressive Regression UCR 2013-2014 2013-2014 2007-2011 (1) (2) (3) 0.0104 0.0322 -0.0019 -0.0082 0.0206 0.0429 (0.0081) (0.0288) (0.0066) (0.0274) (0.0335) (0.0308) -0.3234***(0.0754) 0.1796***(0.0181) 0.2405***(0.0203) -0.0061 (0.0090) 0.7343***(0.0559) 0.0341***(0.0085) -0.0152 (0.0237) -0.1741** (0.0680) 0.1529 (0.1707) -0.0099 (0.0066) 0.0341***(0.0046) -0.0179* (0.0104) 1.0115** (0.4251) -0.9701***(0.1215) -0.0215 (0.3216) 0.2833 (0.3614) -0.0051***(0.0014) 0.6265* (0.3374) 0.3198***(0.1144) -0.1951 (0.3045) 3.4142***(0.6480) 0.1708 (0.4011) 0.0097 0.0393 -0.0002 -0.0089 0.0255 0.0441 (0.0081) (0.0290) (0.0066) (0.0274) (0.0337) (0.0309) -0.3154***(0.0761) 0.1782***(0.0180) 0.2446***(0.0204) -0.0046 (0.0092) 0.7362***(0.0553) 0.0369***(0.0085) -0.0128 (0.0239) -0.1760** (0.0685) 0.1581 (0.1713) -0.0089 (0.0066) 0.0342***(0.0046) -0.0166 (0.0104) 0.9613** (0.4248) -0.9682***(0.1219) 0.0263 (0.3226) 0.3030 (0.3626) -0.0052***(0.0014) 0.5850* (0.3355) 0.3201***(0.1156) -0.1666 (0.3045) 3.4481***(0.6453) 0.2244 (0.4027) -1.4676***(0.4437) -1.2241***(0.3746) State, MSA State, MSA 1.1×107*** 8.9×107*** * ** *** p < 0.1, p < 0.05, p < 0.01; N = 2,657; Robust standard errors are in parentheses. - 30 - -0.0005 (0.0037) -0.0336** (0.0140) -0.0043 (0.0035) 0.0028 (0.0158) -0.0094 (0.0195) 0.0249 (0.0193) 0.0049 (0.0043) -0.0994** (0.0435) 0.1038***(0.0128) 0.0471***(0.0098) -0.0009 (0.0044) 0.2711***(0.0389) 0.0051 (0.0041) 0.0030 (0.0134) -0.0774** (0.0365) 0.1225 (0.0879) -0.0175***(0.0037) 0.0137***(0.0026) 0.0129** (0.0064) 0.0351 (0.2138) -0.0353 (0.0703) -0.3374* (0.1733) -0.1807 (0.1843) -0.0013** (0.0007) 0.6528***(0.1891) 0.1145** (0.0475) -0.4480***(0.1480) 0.1177 (0.3338) -0.4622** (0.2079) -1.3422***(0.3670) State, MSA 7.1×106*** Table 9. Estimations to Address Endogeneity Concerns Dependent Variable Log(Civilian Killed + 1) Method Spatial Autoregressive Regression Data Source The Washington Post (2015) (1) (2) (3) Records -0.0021 (0.0025) -0.0016 (0.0027) -0.0027 (0.0027) Statistical Analyses -0.0199** (0.0090) -0.0216** (0.0097) -0.0235** (0.0097) Access 0.0021 (0.0020) 0.0010 (0.0021) 0.0011 (0.0021) ** ** Smartphone -0.0243 (0.0102) -0.0240 (0.0107) -0.0234** (0.0108) Dash Camera -0.0040 (0.0132) -0.0019 (0.0143) -0.0049 (0.0141) 0.0300** (0.0125) Body Camera 0.0346** (0.0137) 0.0358***(0.0137) Criminal Deaths in 07-11 0.2239***(0.0286) Crime Occurrence 0.0013 (0.0038) 0.0262***(0.0051) 0.0089** (0.0040) ** Crime Clearance -0.0162 (0.0243) -0.0607 (0.0259) -0.1528***(0.0338) Officer Assaults 0.0265***(0.0082) 0.0492***(0.0088) 0.0564***(0.0095) Population 0.0283***(0.0075) 0.0397***(0.0081) 0.0410***(0.0082) Miles 0.0029 (0.0032) 0.0026 (0.0033) 0.0021 (0.0033) *** *** MSA Core City 0.0997 (0.0252) 0.1559 (0.0253) 0.1547***(0.0252) *** Operation Budget 0.0087 (0.0027) 0.0108***(0.0028) 0.0103***(0.0028) Education Req 0.0128 (0.0088) 0.0135 (0.0089) 0.0150* (0.0091) White Officer 0.0073 (0.0243) -0.0036 (0.0251) -0.0084 (0.0254) Female Officer -0.0741 (0.0501) -0.0344 (0.0531) -0.0623 (0.0533) *** *** Weapon -0.0073 (0.0023) -0.0106 (0.0025) -0.0111***(0.0025) Special Units 0.0075***(0.0017) 0.0098***(0.0018) 0.0099***(0.0018) Community Policing -0.0004 (0.0038) 0.0022 (0.0040) 0.0015 (0.0040) Male -0.1579 (0.1379) -0.1355 (0.1549) -0.1520 (0.1539) White 0.0451 (0.0431) 0.0365 (0.0452) 0.0417 (0.0454) * ** Young -0.1856 (0.1122) -0.2763 (0.1243) -0.2521** (0.1226) High School 0.0267 (0.1180) -0.0065 (0.1264) -0.0198 (0.1279) Income -0.0007** (0.0004) -0.0011***(0.0004) -0.0012***(0.0004) Inequality 0.0519 (0.1060) 0.1803 (0.1198) 0.1734 (0.1193) Vacant Homes 0.0079 (0.0292) 0.0256 (0.0325) 0.0301 (0.0319) Moved -0.0093 (0.0936) -0.0750 (0.0992) -0.1086 (0.1001) Female Head 0.0471 (0.1932) 0.1254 (0.2017) 0.1173 (0.2038) Two Parent 0.0845 (0.1448) 0.0029 (0.1554) -0.0152 (0.1550) Inverse Mills Ratio 0.1657***(0.0289) for Statistical Analyses Inverse Mills Ratio -0.4755***(0.1056) for Body Camera ρ -0.822***(0.2571) -0.7714***(0.2359) -0.7606***(0.2347) State, MSA State, MSA State, MSA Controls 2 6*** 6*** 7.8×10 7.5×10 5.2×107*** Wald χ * p < 0.1, **p < 0.05, ***p < 0.01; N = 2,657; Robust standard errors are in parentheses. - 31 - Second, in regard to reverse causality, we also conducted another falsification test, in which we regressed the number of criminal offenders killed by officers in 2007-2011, which was collected from the UCR, on the explanatory variables (Table 8, Column 3). Table 8, Column 3 shows that such a reverse casualty concern is unlikely to be serious for the use of smartphones and wearable body cameras. Indeed, the correlation between criminal deaths in 2007-2011 and body-camera use in 2013 is -0.006. However, the negative and significant coefficient of statistical analyses in Table 8, Column 3 may indicate the presence of reverse causality. While the primary motivation for technology use in law enforcement is to reduce crime rates and clear more crimes (Manning 2001, Nunn 2001, Garicano and Heaton 2010), it might be the case that police departments whose officers kill more offenders in 2007-2011 are more likely to conduct statistical analyses in 2013 to prevent civilian deaths. To further address this endogeneity concern, we followed the approach of Heckman (1979) and Shaver (1998) by obtaining inverse Mills ratios for statistical analyses and body cameras. We first ran Probit regressions in which each of the two variables was regressed to the number of criminals killed, crime occurrences, crime clearance rates, and the number of officers killed and assaulted in 2007-2011 (Table A2 in Appendix). Table A2, Column 2 indicates that the number of criminals killed in 2007-2011 do not have a significant influence on the use of wearable body cameras in 2013. Then we added the inverse Mills ratios from the Probit regressions to the spatial-autocorrelation model (Eq. 5) as an additional control variable. As Table 9, Columns 2-3 demonstrate, while the coefficients of the ratios are significant, the overall results are consistent with our primary estimations in Table 6, Column 1. Third, in terms of measurement error, the baseline analyses only consider civilian deaths by fatal shootings, but there are other ways in which the police can kill civilians, such as Tasers or vehicle accidents. According to the Guardian, 20 people in the U.S. were killed in 2015 by police officers with Tasers (The Guardian 2015), which cause electronic shocks and muscle contractions to a target. We did - 32 - not include them in our main analyses since it is not typically considered to be a lethal weapon9 (Governing 2015). Unlike Tasers, guns are used by the police with an explicit aim to subdue violent targets by killing them. Nevertheless, to demonstrate the robustness of our analysis, we tested the impact of technology use on civilian deaths caused by both lethal (guns) and non-lethal (Taser) weapons, with additional data from the Guardian and killedbypolice.net. This estimation presented in Table A3 produces very similar findings to those in Table 6. Fourth, with respect to alternative empirical specifications, since the dependent variable is in fact a count variable (the number of civilians killed), one could argue that non-linear models such as Poisson or negative binomial regressions would be more appropriate than linear models. However, Greene (2002) pointed out inconsistency in non-linear models that include fixed-effects. In addition, the spatial autoregressive model allows us to account for unobserved heterogeneity correlated with peer jurisdictions. Nonetheless, we estimated the model with Poisson regressions as shown in Table A4 to demonstrate the robustness of our findings. Column 1 with the Washington Post data shows that the effect of wearable video camera use is positive and significant. Columns 2 and 3 with killedbypolice.net data for 2013 and 2014 show that the use of smartphones is still negatively associated with deadly shootings by the police. It could be the case that deadly shooting incidents by the police are not independent with each other, leading us to run a negative binomial regression as in Table A4, Column 4. This regression produced a similar result to Column 1. Finally, we addressed any other potential unobserved heterogeneity concerns as follows. While we controlled for a variety of factors that drive technology use and fatal shootings by the police, we used alternative control variables to further rule out alternative effects. First, one might argue that violent crimes (murders, manslaughters, robberies, and assaults) are more likely to be related to deadly shootings to civilians than other crimes. In Table A5, Column 1, we show the estimation with violent crime 9 The U.S. Department of Defense categorizes a Taser as a non-lethal weapon. http://jnlwp.defense.gov/CurrentNonLethalWeapons/X26Taser.aspx, accessed on May 16, 2016. - 33 - occurrence and clearance rates as alternative control variables, instead of total crimes that include property crimes (burglaries, larcenies, and auto thefts). The main result does not change. In our baseline estimation (Table 6), we did not include the size of police force as a control variable, as it is highly correlated with the amount of operational budget. However, frequent fatal shootings could be driven by a large size of police officers. In Table A5, Column 2, we controlled for the number of sworn full-time officers per capita. The estimate remains consistent. Finally, in Table A5, Column 3, we control for poverty rate (share of population below a poverty line), instead of the median household income. This estimation also provides similar results. 6. Discussion and Conclusion 6.1. Key Findings This study addresses one of the most controversial societal challenges in the U.S – fatal shootings to civilians by the police. The police are mandated to protect the lives of all citizens, be they criminals or otherwise. At the same time, however, they should often make a critical, high-stake decision in an uncertain and violent setting – whether or not to shoot and kill a civilian to save others and themselves. This decision carries a considerable risk; failing to shoot could put innocent people and the officers themselves in danger, but there is always a possibility that the target is innocent or unarmed, an outcome that the officers could be held responsible for. In this study, we ask the following question that no prior IS study has shed light on: how does technology use by the police affect a decision to deploy deadly force? Our analysis with the dataset that combines multiple data sources from both governments (the FBI and the DOJ) and media (the Washington Post) produced several interesting findings. We obtained an encouraging finding that the use of technology for intelligence (statistical analyses and smartphones) is significantly associated with fewer fatal shootings by the police. It illustrates that technology use for intelligence analyses and access helps police officers make more informed decisions in using lethal force. We obtained similar results from the analyses with alternative data on civilian deaths (killedbypolice.net - 34 - and the UCR), demonstrating the robustness of our findings. We found, however, that the use of wearable body cameras is associated with an increase in shooting-deaths of civilians. This contradicts the expectation of many law enforcement officials and policymakers that video cameras would reduce incidents of use of deadly force. We explain this surprising finding in a sense that police officers may believe that the video footages that record encounters with violent civilians would help justify their use of force and exonerate themselves in case of criminal conviction. Based on signal detection theory, we explain that this expectation leads them to put more emphasis on the lives of the innocent in their shooting decisions, becoming less reluctant to use lethal force against suspects. We also find that the impact of technology use – statistical analyses, smartphones, and body cameras – is more pronounced for suspects who are armed with more lethal weapons. In addition, the relationship between technology use and fatal shootings is stronger for civilians who are minorities (African American or Hispanic), male, or younger than 31 years old. Surprisingly, the technology impact is found to be insignificant for Whites or Asians, who make up of 51% of the shooting victims in 2015. According to our analytical model, it seems that technology use influences an officer’s decision to fire more profoundly under greater uncertainty. 6.2. Contribution to the Literature We contribute to the evolving literature on the societal impacts of IS (Chan and Ghose 2016, Greenwood and Wattal 2016, Jha et al. 2016, Venkatesh et al. 2016) by developing a simple, stylized model that depicts a police officer’s behavior in a life-threatening encounter. To our knowledge, this is one of the first attempts in the IS literature to examine the effect of technology in uncertain, turbulent, and life-threatening circumstances. To do so, we adopted a new interdisciplinary approach integrating IS, criminology, and psychology literatures. This study differentiates itself from prior IS research by discovering the significant effect of ethnicity in technology use and its impacts. By studying one of the most politically charged societal problems, we extend the boundary of the IS literature toward the public policy arena, an effort to expand the relevance and influence of IS research. - 35 - 6.3. Implications for Public Policy This study delivers both encouraging and cautionary messages to law enforcement officials and policymakers. As mentioned in the Chicago shooting case in the Introduction, the adoption of wearable video cameras is one of the most widely-adopted policies to reduce deadly incidents. It is believed that the use of video cameras can restrain both excessive use of force by the police and violent resists by civilians (Harris 2010, Ariel et al. 2015). This is why the U.S. Congress is considering the Police CAMERA Act,10 which would authorize the federal government to provide monetary grants to local governments for the purchase of video cameras. Our study shows that this expectation could be misguided. While it is a silver lining that the use of body cameras does not have a significant impact on deaths of unarmed civilians (Table 7, Column 3), it is concerning to law enforcement that it is more significantly related to the fatal shootings of African American and Hispanic victims than Whites and Asians (Table 7, Columns 4 and 5). This paper demonstrates that technology use for intelligence analyses and real-time intelligence access by officers can be much more effective in reducing police-caused homicides. Accordingly, we advise law enforcement officials that the use of video cameras be accompanied with technology for intelligence access and more extensive training for police officers that would help them overcome cognitive racial bias (Correll et al. 2007, 2014). 6.4. Limitations and Future Research There are a few limitations in our study. Since the LEMAS in 2013 is the most recent data available for police IT use, we were able to build only a cross-sectional dataset. Still, we overcame this limitation by building a large-scale dataset with more than 2,600 departments and devising a sound estimation and identification strategy. In addition, the LEMAS survey does not provide in-depth, granular CAMERA stands for “Creating Accountability by Making Effective Recording Available.” The U.S. Senate bill S.877 https://www.congress.gov/bill/114th-congress/senate-bill/877 and House Bill H.R.1680 https://www.congress.gov/bill/114th-congress/house-bill/1680 , accessed on May 16, 2016. 10 - 36 - information on how the police utilize technology. For instance, we were not able to obtain information on, for example, what kind of intelligence police departments obtain from statistical analyses of crime data or for what specific activities they utilize the intelligence. Also, due to data limitation, we were not able to examine how technology use affects warranted and unwarranted shootings differently. This is because it is very challenging to unequivocally judge whether use of deadly force is justified, due to limited evidence and conflicting accounts. Also, the determination of warranty is inherently subjective and depends heavily on judgement of investigators, prosecutors, or judges, each of which can have different interpretations of evidence and laws. In many cases, the results of internal investigations by the police or prosecutors are not available to the public. We believe that our estimation with more objective information on whether the victims were armed or unarmed (Table 7, Columns 1-3) sheds some light on this matter. Law enforcement provides IS researchers with abundant opportunities for future research. It would be interesting to study how police departments can utilize digital technologies in improving the relationship between the community. The criminology literature stresses that a close relationship with the public is imperative for effective law enforcement (Mastrofski et al. 1995, Skolnick and Barley 1998, Kerley and Benson 2000, Adams et al. 2005). The Internet has become an essential tool for the police to communicate and engage with citizens they serve (Government Technology 2012, 2013). Future research can look at how this effort helps the police maintain a symbiotic relationship with the community and affects their response to police use of force incidents. It would be worthwhile to investigate what factors influence the adoption and use of law enforcement technologies. In particular, since the political nature of law enforcement and its impact, we expect that political circumstances in municipalities have a significant impact on technology use in the police. To the best of our knowledge, the IS literature has paid little attention on political impacts on IS (Pang 2016). Future research can also examine how social media affects police practices. In particular, there is a growing concern among law enforcement officials on a socalled “viral video effect” or “YouTube effect” (The Washington Post 2015, Reuter 2016), in which police officers become less reluctant to deal with violence in a fear that their behaviors are being - 37 - videotaped and could be viral on the Internet. Is this really the case? We believe it would be an intriguing IS research question. 6.5. Concluding Remarks Whenever a civilian is killed by the police, be it justified or not, it causes emotional distress and tension, damaging the relationship between the police and the community. Yet, it is often inevitable for police officers to deploy deadly force in order to protect others and themselves. Our study suggests that technology use for intelligence analyses and access could be an effective means to prevent the loss of lives. This is particularly the case for unarmed suspects who may not pose an imminent threat as well as for African Americans and Hispanics, who believe that they have unfairly suffered from police brutality. This is an encouraging result, given that the U.S. history is with century-long struggles and fights for the civil rights, which many people believe are still on-going (e.g. Hall 2005, CBS News 2014, USA Today 2014). While there have been substantial advances in the civil right movement for the last several decades, police use of force is still a politically contentious issue that is significant to all political spectrums. This research provides important insights to the on-going discourses in public safety and civil rights and influence formulation and implementation of the public safety policies in technology adoption. References Adams, R.E., Rohe, W.M., and Arcury, T.A. (2005) “Awareness of Community-Oriented Policing and Neighborhood Perceptions in Five Small to Midsize Cities,” Journal of Criminal Justice (33:1) pp. 43-54. Alpert, G.P. and MacDonald, J.M. (2001) “Police Use of Force: An Analysis of Organizational Characteristics,” Justice Quarterly (18:2) pp. 393-409. Akinola, M. and Mendes, W.B. (2012) “Stress-Induced Cortisol Facilitates Threat-Related Decision Making Among Police Officers,” Behavioral Neuroscience (126:1) pp. 167-174. Ariel, B., Farrar, W.A., and Sutherland, A. (2015) “The Effect of Police Body-Worn Cameras on Use of Force and Citizens’ Complaints against the Police: A Randomized Controlled Trial,” Journal of Quantitative Criminology (31:3) pp. 509-353. Arraiz, I., Drukker, D.M., Kelejian, H.H., and Prucha, I.R. (2010) “A Spatial Cliff-Ord-Type Model with Heteroskedastic Innovations: Small and Large Sample Results,” Journal of Regional Sciences (50:2) pp. 592-614. Baicker, K (2005). “The Spillover Effects of State Spending,” Journal of Public Economics (89:2-3) pp. 529-544. - 38 - Baller, R.D., Anselin, L., Messner, S.F., Deane, G., and Hawkins, D.F. (2001) “Structural Covariates of U.S. County Homicide Rates: Incorporating Spatial Effects,” Criminology (39:3) pp. 561-590. Baumer, E. (1994) “Poverty, Crack, and Crime: A Cross-City Analysis,” Journal of Research in Crime and Delinquency (31:3) pp. 311-327. Baumer, E., Lauritsen, J.L., Rosenfeld, R., and Wright, R. (1998) “The Influence of Crack Cocaine on Robbery, Burglary, and Homicide Rates: A Cross-City, Longitudinal Analysis,” Journal of Research in Crime and Delinquency (35:3) pp. 316-340. Batton, C. and Wilson, S. (2006) “Police Murders: An Examination of Historical Trends in the Killing of Law Enforcement Officers in the United States, 1947 to 1998,” Homicide Studies (10:2) pp. 79-97. Besley, T. and Case, A. (1995) "Incumbent Behavior: Vote-Seeking, Tax-Setting, and Yardstick Competition," American Economic Review (85:1) pp. 25-45. Breustedt, G. and Habermann, H. (2011) “The Incidence of EU Per-Hectare Payments on Farmland Rental Rates: A Spatial Econometric Analysis of German Farm-Level Data,” Journal of Agricultural Economics (62:1) pp. 225-243. CBS News (2014) “50 Years after Civil Rights Act, Americans See Progress on Race,” available at: http://www.cbsnews.com/news/50-years-after-civil-rights-act-americans-see-progress-on-race/, accessed on May 16, 2016. Chan, J.B.L. (2004) “The Technological Game: How Information Technology Is Transforming Police Practice,” Criminal Justice (1:2) pp.139-159. Chan, J., Ghose, A., and Seamans, R. (2016) “The Internet and Racial Hate Crimes: Offline Spillovers from Online Access,” MIS Quarterly (40:2) pp. 381-403. CNN (2015) “FBI Chief Tries to Deal with the ‘Ferguson Effect,’” available at: http://www.cnn.com/2015/10/26/politics/fbi-comey-crime-police/, accessed on May 10, 2016. Cordner, G. (2014) “Community Policing,” In The Oxford Handbook of Police and Policing, Reisig, M.D. and Kane, R.J. (eds.), Oxford University Press, pp. 148-171. Correll, J., Hudson, S.M., Guillermo, S., and Ma, D.S. (2014) “The Police Officer’s Dilemma: A Decade of Research on Racial Bias in the Decision to Shoot,” Social and Personality Psychology Compass (8:5) pp. 201-213. Correll, J., Park, B., Judd, C.M., and Wittenbrink, B. (2002) “The Police Officer’s Dilemma: Using Ethnicity to Disambiguate Potentially Threatening Individuals,” Journal of Personality and Social Psychology (83:6) pp. 1314-1329. Correll, J., Park, B., Judd, C.M., and Wittenbrink, B. (2007a) “The Influence of Stereotypes on Decisions to Shoot,” European Journal of Social Psychology (37:6) pp. 1102-1117. Correll, J., Park, B., Judd, C.M., Wittenbrink, B., Sadler M.S., and Keesee, T (2007b) “Across the Thin Blue Line: Police Officers and Racial Bias in the Decision to Shoot,” Interpersonal Relations and Group Processes (92:6) pp. 1006–1023. Correll, J., Urland, G.R., and Ito, T.A. (2006) “Event-Related Potentials and the Decision to Shoot: The Role of Threat Perception and Cognitive Control,” Journal of Experimental Social Psychology (42:1) pp. 120-128. Dantzker, M.L. (2010) Understanding Today’s Police, Boulder, CO: Lynne Rienner Publishers. Drukker, D.M., Prucha, I.R., and Raciborski, R. (2013) “Maximum Likelihood and Generalized Spatial Two-Stage Least-squares Estimators for a Spatial-Autoregressive Model with Spatial-Autoregressive Disturbances,” The Stata Journal (13:2) pp. 221-241. Figlio, D.N., Kolpin, V.W., and Reid, W.E. (1999) “Do States Play Welfare Games?,” Journal of Urban Economics (46:3) pp. 437-454. - 39 - Five Thirty Eight (2014) “Nobody Knows How Many Americans the Police Kill Each Year,” available at: http://fivethirtyeight.com/features/how-many-americans-the-police-kill-each-year/, accessed on May 11, 2016. Fridell, L., Faggiani, D., Taylor, B., Brito, C.S., and Kubu, B. (2009) “The Impact of Agency Context, Policies, and Practices on Violence against Police,” Journal of Criminal Justice (37:6) pp. 542-552. Garicano, L. and Heaton, P. (2010) “Information Technology, Organization, and Productivity in the Public Sector – Evidence from Police Departments,” Journal of Labor Economics (28:1) pp. 167201. Green, D.M. and Swets, J.A. (1988) Signal Detection Theory and Psychophysics, Peninsula Publishing: Newport Beach, CA. Greene, W. (2002) “The Behavior of the Fixed Effects Estimator in Nonlinear Models,” Unpublished working paper. Greenwood, B. and Wattal, S. (2016) “Show Me the Way to Go Home: An Empirical Investigation of Ride Sharing and Alcohol Related Motor Vehicle Homicide,” MIS Quarterly, forthcoming. Governing (2015) “Have Non-Lethal Weapons Reduced Deadly Police Force?,” available at: http://www.governing.com/columns/tech-talk/gov-non-lethal-weapons-police.html, accessed on May 16, 2016. Government Technology (2012) “Chicago Crowdsources Gun Violence Solutions Via Twitter,” available at http://www.govtech.com/public-safety/Houston-Website-Solicits-Anonymous-Tips-AboutGangs-.html, accessed on May 16, 2016. Government Technology (2013) “New App is Latest Evidence of Florida PD’s Aggressive Engagement Strategy,” available at http://www.govtech.com/public-safety/New-App-is-Latest-Evidence-ofFlorida-PDs-Aggressive-Engagement-Strategy.html, accessed on May 16, 2016. Government Technology (2015) “License Plate Readers Solve Crimes but Prompt Privacy Concerns,” available at: http://www.govtech.com/public-safety/License-Plate-Readers-Solve-Crimes-butPrompt-Privacy-Concerns.html, accessed on May 16, 2016. Haberman, C.P. and Ratcliffe, J.H. (2012) The Predictive Policing Challenges of Near Repeat Armed Street Robberies,” Policing (6:2) pp. 151-166. Hahn, H. and Jeffries, J.L. (2003) Urban America and Its Police, Boulder, CO: University of Colorado Press. Hall, J.D. (2005) “The Long Civil Rights Movement and the Political Uses of the Past,” The Journal of American History (91:4) pp. 1233-1263. Harris, D.A. (2010) “Picture This: Body Worn Video Devices ('Head Cams') as Tools for Ensuring Fourth Amendment Compliance by Police,” Texas Tech Law Review (43:1) pp. 357-372. Harvard Law Review (2015) “Considering Police Body Cameras,” Harvard Law Review (128:6) pp. 1794-1817. Heckman, J. (1979) “Sample Selection Bias as a Specification Error,” Econometrica (47:1) pp. 153-61. Holly, S., Pesaran, M.H., and Yamagata, T. (2011) “The Spatial and Temporal Diffusion of House Prices in the UK,” Journal of Urban Economics (69:1) pp. 2-23. Kahan, D.M., Hoffman, D.A., and Braman, D. (2009) “Whose Eyes Are You Going to Believe? Scott v. Harris and the Perils of Cognitive Illiberalism,” Harvard Law Review (122:3) pp. 837-906. Kaminski, R.J. (2008) “Assessing the County-Level Structural Covariates of Police Homicides,” Homicide Studies (12:4) pp. 350-380. Kaminski, R.J. and Stucky, T.D. (2009) “Reassessing Political Explanations for Murders of Police,” Homicide Studies (13:1) pp. 3-20. - 40 - Kelejian, H.H. and Prucha, I.R. (2010) “Specification and Estimation of Spatial Autoregressive Models with Autoregressive and Heteroskedastic Disturbances,” Journal of Econometrics (157:1) pp. 53-67. Kent, S.L. (2010) “Killings of Police in U.S. Cities Since 1980: An Examination of Environmental and Political Explanations,” Homicide Studies (14:1) pp. 3-23. Kenworthy, J.B., Barden, M.A., Diamond, S., and del Carmen, A. (2011) “Ingroup Identification as a Moderator of Racial Bias in a Shoot-No Shoot Decision Task,” Group Processes & Intergroup Relations (14:3) pp. 311-318. Kerley, K.R. and Benson, M.L. (2000) “Does Community-Oriented Policing Help Build Stronger Communities?” Police Quarterly (3:1) pp. 46-69. Jacobs, D. and Carmichael, J.T. (2002) “Subordination and Violence against State Control Agents: Testing Political Explanations for Lethal Assaults against the Police,” Social Forces (80:4) pp. 12231251. Jacobs, D. and O’Brien, R.M. (1998) “The Determinants of Deadly Force: A Structural Analysis of Police Violence,” American Journal of Sociology (103:4) pp. 837-852. Jennings, W.G., Lynch, M.D., and Fridell, L.A. (2015) “Evaluating the Impact of Police Officer BodyWorn cameras (BWCs) on Response-to-Resistance and Serious External Complaints: Evidence from the Orlando Police Department (OPD) Experience Utilizing a Randomized Controlled Experiment,” Journal of Criminal Justice (43:6) pp. 480-486. Jha, S.K., Pinsonneault, A., and Dube, L. (2016) “The Evolution of an ICT Platform-Enabled Ecosystem for Poverty Alleviation: The Case of eKutir,” MIS Quarterly (40:2) pp. 431-455. Land, K.C., McCall, P.L., and Cohen, L.E. (1990) “Structural Covariates of Homicide Rates: Are There Any Invariances across Time and Social Space?,” American Journal of Sociology (95:4) pp. 922963. Long Beach Press-Telegram (2015) “Police: No Weapon Found at Scene of Officer-Involved Shooting in Long Beach Thursday,” available at: http://www.presstelegram.com/general-news/20150424/policeno-weapon-found-at-scene-of-officer-involved-shooting-in-long-beach-thursday, accessed on May 10, 2016. Ma, D.S. and Correll, J. (2011) “Target Prototypicality Moderates Racial Bias in the Decision to Shoot,” Journal of Experimental Social Psychology (47:2) pp. 391-396. Ma, D.S., Correll, J., Wittenbrank, B., Bar-Anan, Y., Sriram, N., and Nosek, B.A. (2013) “When Fatigue Turns Deadly: The Association between Fatigue and Racial Bias in the Decision to Shoot,” Basic and Applied Social Psychology (35:6) pp. 515-524. MacMillan, N.A. (2002) “Signal Detection Theory,” in Stevens’ Handbook of Experimental Psychology, Third Edition, Pashler, H. and Wixted, J. (eds), pp. 43-90, John Wiley & Sons: New York, NY. Mastrofski, S.D., Worden, R.E., and Snipes, J.B. (1995) “Law Enforcement in a Time of Community Policing,” Criminology (33:4) pp. 539-563. McElvain, J.P. and Kposowa, A.J. (2008) “Police Officer Characteristics and the Likelihood of Using Deadly Force,” Criminal Justice and Behavior (35:4) pp. 505-521. Miethe, T.D., Hughes, M., and McDowall, D. (1991) “Social Change and Crime Rates: An Evaluation of Alternative Theoretical Approaches,” Social Forces (70:1) pp. 165-185. NBC News (2016) “Body Cameras Win Converts Among Police Officers on the Beat,” available at: http://www.nbcnews.com/news/crime-courts/body-cameras-win-converts-among-police-officersbeat-n566311, accessed on May 11, 2016. Nunn, S. (2001) “Police Information Technology: Assessing the Effects of Computerization on Urban Police Functions,” Public Administration Review (61:2) pp. 221-234. - 41 - Pang, M.S. (2016) “Politics and Information Technology Investments in the U.S. Federal Government in 2003-2016,” Fox School of Business Research Paper No. 15-048, available at: http://papers.ssrn.com/abstract=2540359. Pang, M.S. and Pavlou, P. (2016) “Does IT Use by the Police Keep the City's Finest Safer?,” Fox School of Business Research Paper No. 16-016, available at: http://papers.ssrn.com/abstract=2746194. Pearsall, B. (2010) “Predictive Policing: The Future of Law Enforcement?,” National Institute of Justice Journal (266:1) pp. 16-19. Reuter (2016) “Viral Videos of Police Brutality Lead to More Crime, FBI Director Claims,” available at: https://www.rt.com/usa/343027-fbi-viral-police-director/, accessed on May 16, 2016. Sadler, M.S., Correll, J., Park, B., and Judd, C.M. (2012) “The World Is Not Black and White: Racial Bias in the Decision to Shoot in a Multiethnic Context,” Journal of Social Issues (68:2) pp. 286-313. Shaver, J.M. (1998) “Accounting for Endogeneity When Assessing Strategy Performance: Does Entry Mode Choice Affect FDI Survival?,” Management Science (44:4) pp. 571-585. Skogan, W. and Frydl, K. (2004) Fairness and Effectiveness in Policing: The Evidence, The National Academies Press: Washington, D.C. Skolnick, J.H. and Barley, D.H. (1988) “Theme and Variation in Community Policing,” Crime and Justice (10:1) pp. 1-37. Sim, J.J., Correll, J., and Salder, M.S. (2013) “Understanding Police and Expert Performance: When Training Attenuates (vs. Exacerbates) Stereotypic Bias in the Decision to Shoot,” Personality and Social Psychology Bulletin (39:3) pp. 291-304. Smith, B.W. (2004) “Structural and Organizational Predictors of Homicide by Police,” Policing (27:4) pp. 539-557. The Baltimore Sun (2016) “When Shots are Fired,” available at: http://www.baltimoresun.com/news/opinion/editorial/bs-ed-shotspotter-20160429-story.html, accessed on May 16, 2016. The Guardian (2015) “Bolts from the Blue,” available at: http://www.theguardian.com/usnews/2015/nov/05/police-tasers-deaths-the-counted, accessed on May 16, 2016. The New York Times (2013) “Wearing a Badge, and a Video Camera”, available at: http://www.nytimes.com/2013/04/07/business/wearable-video-cameras-for-police-officers.html, accessed on May 10, 2016. The New York Times (2016a) “Chicago after Laquan McDonald,” available at: http://www.nytimes.com/2016/04/24/magazine/chicago-after-laquan-mcdonald.html, accessed on May 10, 2016. The New York Times (2016b) “Homicide Rates Jump in Many Major U.S. Cities, New Data Shows,” available at: http://www.nytimes.com/2016/05/14/us/murder-rates-cities-fbi.html, accessed on May 15, 2016. The Washington Post (2015) “’YouTube Effect’ Has Left Police Officers under Siege, Law Enforcement Officials Say,” available at: https://www.washingtonpost.com/news/postnation/wp/2015/10/08/youtube-effect-has-left-police-officers-under-siege-law-enforcement-leaderssay/, accessed on May 16, 2016. The Washington Post (2016a) “Final Tally: Police Shot and Killed 986 People in 2015,” available at: https://www.washingtonpost.com/national/final-tally-police-shot-and-killed-984-people-in2015/2016/01/05/3ec7a404-b3c5-11e5-a76a-0b5145e8679a_story.html, accessed on May 10, 2016. The Washington Post (2016b) “Cleveland Agrees to Pay $6 Million to Settle Tamir Rice Lawsuit, Won’t Admit Any Wrongdoing,” available at: https://www.washingtonpost.com/news/postnation/wp/2016/04/25/cleveland-agrees-to-pay-6-million-to-settle-tamir-rice-lawsuit/, accessed on May 10, 2016. - 42 - The U.S. Supreme Court (2007) Scott v. Harris, 500 U.S. 372, available at http://www.supremecourt.gov/opinions/06pdf/05-1631.pdf, accessed on May 11, 2016. Thurman, Q.C. and Reisig, M.D. (1996) “Community-Oriented Research in an Era of CommunityOriented Policing,” American Behavioral Scientist (39:5) pp. 570-586. U.K. Home Office (2007) Guidance for the Police Use of Body-Worn Video Devices, available at http://library.college.police.uk/docs/homeoffice/guidance-body-worn-devices.pdf, accessed on May 16, 2015. USA Today (2014) “Equality still elusive 50 years after Civil Rights Act,” available at: http://www.usatoday.com/story/news/nation/2014/01/19/civil-rights-act-progress/4641967/, accessed on May 16, 2016. Van Duijn, M. and Rouwendal, J (2013) “Cultural Heritage and the Location Choice of Dutch Households in a Residential Sorting Model,” Journal of Economic Geography (3:4) pp. 473-500. Venkatesh, V., Rai, A., Sykes, T.A., and Aljafari, R. (2016) “Combating Infant Mortality in Rural India: Evidence from a Field Study of eHealth Kiosk Implementations,” MIS Quarterly (40:2) pp. 353-380. Wickens, T.D. (2002) Elementary Signal Detection Theory, Oxford University Press: New York, NY. - 43 - Appendix – Additional Estimation Tables Table A1. Estimations by Race in 2013-2015 Dependent Variable Log (Civilian Killed + 1) Method Spatial Autoregressive Regression Data Source and Year The Washington Post and killedbypolice.net (2013-2015) Race White and Asian Black and Hispanic (1) (2) 0.0009 (0.0028) -0.0246** (0.0110) -0.0018 (0.0024) -0.0163 (0.0112) 0.0060 (0.0141) -0.0105 (0.0138) 0.0055 (0.0037) -0.0015 (0.0269) 0.0405***(0.0087) 0.0520***(0.0081) 0.0023 (0.0036) 0.1160***(0.0256) 0.0117***(0.0031) 0.0083 (0.0085) 0.0023 (0.0300) -0.0739 (0.0568) -0.0092***(0.0027) 0.0104***(0.0020) -0.0017 (0.0043) -0.1015 (0.1528) 0.1324***(0.0434) -0.1407 (0.1215) 0.1311 (0.1341) 0.0002 (0.0005) 0.0456 (0.1229) -0.0066 (0.0434) -0.1129 (0.1108) 0.0883 (0.2182) -0.3544** (0.1619) -0.8054***(0.2459) State, MSA 6.2×106*** * ** *** p < 0.1, p < 0.05, p < 0.01; N = 2,657; Robust standard errors are in parentheses. Records Statistical Analyses Access Smartphone Dash Camera Body Camera Crime Occurrence Crime Clearance Officer Assaults Population Miles MSA Core City Operational Budget Education Req White Officer Female Officer Weapon Special Units Community Policing Male White Young High School Income Inequality Vacant Homes Moved Female Household Head Two Parent Household ρ Controls Wald χ2 - 44 - -0.0023 (0.0026) -0.0239** (0.0095) -0.0003 (0.0021) -0.0337***(0.0110) -0.0060 (0.0142) 0.0338***(0.0127) 0.0101***(0.0033) -0.1226***(0.0264) 0.0532***(0.0104) 0.0281***(0.0085) 0.0017 (0.0033) 0.1696***(0.0259) 0.0051* (0.0028) 0.0026 (0.0097) -0.0236 (0.0232) 0.0975 (0.0642) -0.0076***(0.0025) 0.0105***(0.0018) 0.0007 (0.0040) -0.1401 (0.1310) -0.1268***(0.0444) -0.2875** (0.1191) -0.0148 (0.1234) -0.0011***(0.0004) 0.4594***(0.1285) 0.0587** (0.0298) -0.1471 (0.1055) 0.0703 (0.2140) 0.0186 (0.1446) -0.7785***(0.2355) State, MSA 6.0×106*** Table A2. Probit Regressions to Obtain Inverse Mill’s Ratios Dependent Variable Statistical Analyses Body Camera Method Probit Regression (1) (2) *** 0.2768 (0.1080) 0.0841 (0.0562) Criminal Deaths in 2007-2011 0.2522***(0.0149) -0.0014 (0.0133) Crime Occurrence in 2007-2011 -0.2765***(0.1662) 0.6428***(0.1617) Crime Clearance in 2007-2011 *** 0.0485 (0.0192) -0.0397** (0.0163) Officer Assaults in 2007-2011 *** -1.4611 (0.1127) -0.6339***(0.1032) Constant 0.1757 0.0075 Pseudo R2 -1390.2816 -1549.4117 Log L 592.56*** 23.45*** Wald χ2 * p < 0.1, **p < 0.05, ***p < 0.01; N = 2,657; Robust standard errors are in parentheses. - 45 - Table A3. Estimations with Deaths by Guns and Tasers Dependent Variable Log (Civilian Killed + 1) Method Spatial Autoregressive Regression The Washington Post The Washington Post Data Source The Guardian The Guardian killedbypolice.net Year 2015 2013-2015 (1) (2) -0.0022 (0.0027) -0.0248** (0.0100) 0.0012 (0.0022) -0.0243** (0.0108) -0.0057 (0.0144) 0.0325** (0.0137) 0.0036 (0.0042) -0.0567** (0.0256) 0.0527***(0.0089) 0.0411***(0.0085) 0.0020 (0.0034) 0.1724***(0.0256) 0.0104***(0.0029) 0.0116 (0.0092) -0.0048 (0.0256) -0.0556 (0.0539) -0.0112***(0.0026) 0.0106***(0.0019) 0.0014 (0.0041) -0.1370 (0.1526) 0.0358 (0.0459) -0.2722** (0.1252) 0.0027 (0.1289) -0.0010** (0.0004) 0.1885 (0.1220) 0.0345 (0.0319) -0.1119 (0.1015) 0.1462 (0.2035) -0.0174 (0.1559) -0.7379***(0.2298) State, MSA 1.4×107*** * ** *** p < 0.1, p < 0.05, p < 0.01; N = 2,657; Robust standard errors are in parentheses. Records Statistical Analyses Access Smartphone Dash Camera Body Camera Crime Occurrence Crime Clearance Officer Assaults Population Miles MSA Core City Operational Budget Education Req White Officer Female Officer Weapon Special Units Community Policing Male White Young High School Income Inequality Vacant Homes Moved Female Household Head Two Parent Household ρ Controls Wald χ2 - 46 - -0.0018 (0.0038) -0.0354** (0.0149) -0.0013 (0.0032) -0.0423***(0.0150) -0.0061 (0.0193) 0.0103 (0.0182) 0.0153***(0.0052) -0.1147***(0.0378) 0.0841***(0.0123) 0.0784***(0.0119) 0.0030 (0.0051) 0.2749***(0.0350) 0.0181***(0.0043) 0.0019 (0.0127) -0.0213 (0.0394) 0.0026 (0.0832) -0.0119***(0.0037) 0.0163***(0.0026) -0.0025 (0.0057) -0.2240 (0.2168) 0.0100 (0.0639) -0.3489** (0.1707) 0.1614 (0.1883) -0.0006 (0.0006) 0.2971* (0.1768) 0.0386 (0.0574) -0.2081 (0.1542) 0.3661 (0.3109) -0.2965 (0.2192) -0.6584***(0.2215) State, MSA 1.2×107*** Dependent Variable Table A4. Poisson and Negative Binomial Regressions Civilian Killed Method Data Source Year Records Statistical Analyses Access Smartphone Dash Camera Body Camera Crime Occurrence Crime Clearance Officer Assaults Population Miles MSA Core City Operation Budget Education Req White Officer Female Officer Weapon Special Units Community Poli Male White Young High School Income Inequality Vacant Homes Moved Female Head Two Parent Controls Pseudo L Pseudo R2 Wald χ2 Poisson The Washington Post 2015 (1) killedbypolice.net The Washington Post & killedbypolice.net 2013-2014 (2) 2013-2015 (3) Negative Binomial1) The Washington Post 2015 (4) -0.0564 (0.0486) -0.0412 (0.0399) -0.0498* (0.0296) -0.0555 (0.0400) 0.2406 (0.2106) -0.2170 (0.2058) -0.0098 (0.1564) 0.0809 (0.2112) 0.0383 (0.0461) (0.0467) (0.0327) 0.0170 (0.0366) -0.1767 (0.1264) -0.2379** (0.1116) -0.2173** (0.0850) -0.1295 (0.1073) -0.1036 (0.1387) (0.1118) 0.3215***(0.1225) 0.2605** (0.1015) 0.0346 0.0906 -0.0811 (0.1160) (0.1164) 0.5848***(0.0968) 0.0375 0.0247 0.0785 (0.0922) -0.1833 (0.0832) 0.2602** (0.1029) 0.4294***(0.1002) 0.2570* (0.1532) 0.6687 (0.5634) -0.1782 (0.5355) 0.1810 (0.4232) 0.0121 0.0621 (0.0439) (0.0369) 0.0383 (0.0297) 0.0860***(0.0229) 0.0258 (0.4656) 0.5054***(0.1290) 0.3929***(0.1051) 0.4660***(0.1073) 0.5170***(0.1481) 0.1038 (0.0724) 0.1567***(0.0544) 0.1233** (0.0511) 0.0802 0.7682***(0.1719) 0.2585* (0.1528) 0.4696***(0.1239) 0.5408***(0.1759) 0.1131 (0.0708) 0.1725** (0.0705) 0.1499***(0.0502) 0.1783* (0.0958) 0.0661 (0.0781) -0.0970 (0.0770) -0.0264 (0.0568) -0.0196 (0.0623) 0.2331 (0.3911) (0.2983) 0.1569 (0.2393) -0.0263 (0.2573) 0.1211 (0.0572) -1.5621 (1.1260) -0.3161 (1.0828) -0.7502 (0.8047) -1.1571 (0.9805) -0.0846** (0.0373) (0.0282) 0.0030 (0.0299) -0.0300 (0.0235) -0.0235 (0.0195) -0.0176 (0.0175) -0.0005 (0.0142) -0.0701 (0.0502) 0.0234 (0.0380) -0.0153 (0.0321) -0.0291 (0.0385) -4.8195 (4.6486) -2.2199 (3.4134) -3.3263 (2.7735) -1.4146 (3.7030) -0.2787 (0.7257) -0.7085 (0.6201) -0.4705 (0.4654) -0.5759 (0.4451) 0.0246 0.0397** (0.0183) -5.0911***(1.7065) -4.3951 (2.7221) -1.4522 (1.8951) -2.6401 (1.6110) -2.0787 (3.4148) 3.3034 (2.8182) (2.3234) -2.7467 -0.0227** (0.0106) 0.0154* (0.0093) -4.2908* (2.5587) 0.3582 (2.1249) -1.4797 -0.8315 (1.5469) -5.6027** (2.2873) -2.8724* (1.4698) -1.2988 1.6874 (2.1090) 2.6145 -1.1981 (5.0581) -0.8364 (2.9776) -4.4500 6.8264 (1.8868) 1.3087 0.0002 2.5089* (0.0081) (1.9289) -0.0258***(0.0069) (1.7099) -4.0430** (1.8942) (1.3685) (1.5087) 3.5026** 3.5027 (3.4847) -7.0408* (2.8482) -2.8649 (2.0945) 2.1929 (4.4183) State, MSA State, MSA -680.1627 -909.5409 -1257.2117 0.6175 0.6705 0.6855 5*** 10*** 1.4×10 2.9×10 2.4×108*** * p < 0.1, **p < 0.05, ***p < 0.01; N = 2,657; Robust standard errors are in parentheses. 1) Estimation with state and MSA fixed-effects failed to achieve convergence. (1.3649) (3.6357) (2.0759) State, MSA - 47 - -905.7210 0.3180 1115.97*** Dependent Variable Method Source Alternative Control Records Statistical Analyses Access Smartphone Dash Camera Body Camera Crime Occurrence Crime Clearance Violent Crime Occurrence Violent Crime Clearance Officer Assaults Population Miles MSA Core City Operational Budget Officer Education Req White Officer Female Officer Weapon Special Units Community Policing Male White Young High School Income Poverty Level Inequality Vacant Homes Moved Female Household Head Two Parent Household ρ Controls Wald χ2 Table A5. Alternative Control Variables Log (Civilian Killed + 1) Spatial Autoregressive Regression The Washington Post (2015) Violent Crimes Police Force Size Poverty Rate (1) (2) (3) -0.0026 (0.0027) -0.0274***(0.0099) 0.0010 (0.0022) -0.0255** (0.0107) -0.0065 (0.0142) 0.0355** (0.0138) 0.0096** (0.0045) -0.0451** (0.0192) 0.0473***(0.0089) 0.0340***(0.0082) 0.0036 (0.0033) 0.1585***(0.0257) 0.0100***(0.0029) 0.0141 (0.0091) -0.0068 (0.0252) -0.0440 (0.0531) -0.0115***(0.0025) 0.0102***(0.0018) 0.0032 (0.0041) -0.1529 (0.1542) 0.0428 (0.0454) -0.2701** (0.1244) -0.0187 (0.1277) -0.0010** (0.0004) -0.0023 -0.0229** 0.0008 -0.0234** -0.0015 0.0352** 0.0056 -0.0433* (0.0027) (0.0097) (0.0021) (0.0107) (0.0142) (0.0136) (0.0038) (0.0254) -0.0022 (0.0027) -0.0264***(0.0098) 0.0009 (0.0022) -0.0244** (0.0107) -0.0067 (0.0142) 0.0357** (0.0138) 0.0050 (0.0039) -0.0465* (0.0254) 0.0461***(0.0088) 0.0523***(0.0094) 0.0032 (0.0033) 0.1571***(0.0254) 0.0487***(0.0089) 0.0400***(0.0082) 0.0020 (0.0033) 0.1616***(0.0256) 0.0100***(0.0029) 0.0199***(0.0041) 0.0156* (0.0090) 0.0056 (0.0253) -0.0671 (0.0530) -0.0119***(0.0025) 0.0095***(0.0018) 0.0030 (0.0040) -0.1675 (0.1549) 0.0477 (0.0450) -0.2922** (0.1265) -0.0473 (0.1262) -0.0012***(0.0004) 0.1841 (0.1203) 0.1011 (0.1180) 0.0275 (0.0316) -0.1687***(0.0634) -0.1163 (0.1001) -0.1867* (0.1038) 0.0408 (0.2029) 0.0403 (0.2062) -0.0186 (0.1553) -0.1154 (0.1625) -0.7695***(0.2370) -0.8017***(0.2410) State, MSA State, MSA 9.9×106*** 6.0×106*** * p < 0.1, **p < 0.05, ***p < 0.01; N = 2,657; Robust standard errors are in parentheses. - 48 - 0.0133 (0.0090) -0.0097 (0.0253) -0.0479 (0.0531) -0.0113***(0.0025) 0.0104***(0.0018) 0.0024 (0.0041) -0.2054 (0.1503) 0.0343 (0.0459) -0.2353* (0.1257) 0.1480 (0.1101) 0.1160 (0.0905) 0.1259 (0.1242) 0.0375 (0.0320) -0.0759 (0.0989) 0.0788 (0.2139) -0.0831 (0.1537) -0.779***(0.2397) State, MSA 6.4×106***