key: cord-0630817-681q9uwo authors: Onookome-Okome, Tuviere; Gorondensky, Jonah; Rose, Eric; Sauer, Jeffery; Lum, Kristian; Moodie, Erica EM title: Characterizing patterns in police stops by race in Minneapolis from 2016-2021 date: 2022-05-12 journal: nan DOI: nan sha: 7ead61d7382d9dbcb53a98d124a74f1177912d0f doc_id: 630817 cord_uid: 681q9uwo The murder of George Floyd centered Minneapolis, Minnesota, in conversations on racial injustice in the US. We leverage open data from the Minneapolis Police Department to analyze individual, geographic, and temporal patterns in more than 170,000 police stops since 2016. We evaluate person and vehicle searches at the individual level by race using generalized estimating equations with neighborhood clustering, directly addressing neighborhood differences in police activity. Minneapolis exhibits clear patterns of disproportionate policing by race, wherein Black people are searched at higher rates compared to White people. Temporal visualizations indicate that police stops declined following the murder of George Floyd. This analysis provides contemporary evidence on the state of policing for a major metropolitan area in the United States. Instances of police brutality have, once again, surged to the forefront of public consciousness. In particular, the murder of George Floyd on May 25, 2020, by on-duty members of the Minneapolis Police prompted widespread protests, public discourse, and calls for the abolition of the police in disproportionately the victims of police brutality in the United States, and police brutality is often excused or ignored by due to pervasive racism (Chaney & Robertson, 2013 ). An additional and important dynamic is that negative police interactions, including police brutality, are rarely punished, leading some legal scholars to argue that 'police brutality has been effectively decriminalized in this country [the United States]' (Davis, 1994) . Although instances of accountability for police brutality exist, several recognized obstacles from both within the institution of police, usually recognized as the 'Blue Wall of Silence' or 'Blue Code of Silence', and society at large, such as the response of 'Blue Lives Matter' to 'Black Lives Matter', pose serious obstacles to addressing negative consequences of racial bias in police actions (Cooper, 2020; Skolnick, 2002; Solomon et al., 2021) . There have been some studies examining policing and racial bias in the greater Minneapolis area. Examples including research by Gorsuch and Rho (2019) , Ritter (2017) , Ritter and Bael (2009) , Wexler (2020) , as well as related work on racial bias in the Minnesota justice system by Gorsuch & Rho, 2019; Wexler, 2020) . Gorsuch and Rho, using the same publicly available MPD stop data we describe later in the paper, found that Indigenous women were searched and arrested at a rate of more than double that of women belonging to any other racial group (Gorsuch & Rho, 2019) . 2 Moreover, Gorsuch and Rho use descriptive mapping to highlight the concentration of policing in neighbourhoods with a high Indigenous population (Gorsuch & Rho, 2019) . Ritter (2017) investigated the role of race in MPD stop and search data from 2002, finding evidence of the effect of race on police stop decisions, but not police on search decisions (Ritter, 2017) . Earlier work by Ritter and Bael (2009) used the same data from 2002 and found evidence that Black drivers were stopped 5.5% less often at night compared to during the day, supporting the "veil of darkness" theory which states that racial profiling is less likely when visibility is obscured (Ritter & Bael, 2009 ). Investigating race visibility and police stops, Wexler (2020) demonstrated similar findings to Ritter and Bael (2009) , although to a lesser extent (Wexler, 2020) . While studies on policing and racial bias in Minnesota may be limited in number, numerous studies on policing and racial bias have been published from locales across the United States over the past several decades. Early work by Harris demonstrated that Black individuals, especially Black individuals of driving age, were more than twice as likely to get a ticket compared to White drivers across several major police departments (Akron, Toledo, Dayton, and Columbus/Franklin County) in Ohio (Harris, 1999) . Nationwide audits of police-public contact from the same time period found that although there were far fewer Black drivers compared to white drivers in the United States, a greater proportion of Black drivers were stopped (12.3% of Black drivers compared to 10.4% of White drivers) (Langan et al., 2001) . Later studies on police stops in Boston showed that Non-white individuals are stopped and arrested more often than White individuals, even after adjusting for potentially confounding social and criminological factors (Fagan et al., 2015 (Fagan et al., , 2016 . In New York City, individuals of African and Hispanic descent tend to be stopped approximately twice as often as Whites for suspected violent crimes and weapons offenses . Gelman et al. also found evidence of high stop rates for minorities in predominantly White precincts (as well as high stop rates for Whites in predominantly minority precincts) and argue that such "racial incongruity" stops indicate the police act on "out of place" suspicions (Gould & Mastrofski, 2004; Alpert et al., 2005; . This latter finding aligns with research by Roh and Robinson (2009) who demonstrate the utility of analyzing stop data at both the micro-(e.g., individual) and macro-(e.g., areal units) level when investigating racial disparities in traffic stops (Roh & Robinson, 2009 ). Another study situated in Durham, North Carolina, demonstrated a positive association between daytime stops and an individual being a Black male even after accounting for officer, rather than neighborhood, factors (Taniguchi et al., 2017) . Buerger and Farrell provide a summary of the notable litigation surrounding racial profiling in the late 1990s and early 2000s, arguing that narrowing legal definitions versus widening public definitions limited system-wide solutions, even in the face of increasingly substantiated demonstration of racial disparities in stop rates across the United States (Buerger & Farrell, 2002) . 3 Studies with a focus outside the United States have also found racial bias in policing. The notable and early 'Black and Blue' study of police stops in London, England, showed that Black individuals were 2.8 times more likely to be stopped compared to their representation in the 3 Readers interested in more extensive literature reviews on racial bias and disparities in traffic stops should look to both Roh and Robinson (see section 'Literature Review') and Gorsuch and Rho 2019 (see section 'Background -Disparities in Police Stops and Their Effects') (Gorsuch & Rho, 2019; Roh & Robinson, 2009). underlying population, even as Black and White individuals had equally mild reactions during police contact (Norris et al., 1992) . A systematic review of international, empirical studies examining racial prejudice and police stops found that Black men were the most commonly stopped racial group amongst studies situated in the United States, England, Wales, and Netherlands (Ara et al., 2021) . Many of the extant studies on racial bias in policing have made use of limited access data. The expansion of open data in the last decade has led to new avenues for research into police interactions. Recently, Pierson et al. (2020) collated data on approximately 100 million traffic stops from across the United States, providing robust evidence on several factors related to racialized policing including support for the "veil of darkness" theory, excess stops of Black and Hispanic drivers compared to White drivers, and policy changes like the legalization of recreational marijuana failing to address racial disparities in police stops (Pierson et al., 2020) .Smaller-scale studies also using open data have found similar effects. A study using nearly a decade of publicly available traffic-stop records from Louisville, Kentucky found that Black drivers were more likely to be stopped even when compared with other stops of similar characteristics (Vito et al., 2020) . Another analysis using open data from the Washington, D.C., metropolitan police department found that Black individuals comprised nearly 90% of all stops even though the Black population within the jurisdiction was only 51% of the population, and that Black individuals had a disproportionately higher arrest rate compared to the White population in an area (Fielding-Miller et al., 2016) . Open data are one of the many emergent tools -alongside others including but not limited to public video recording, police body cameras, apps that automatically report potential incidents of misconduct, social media, and more -that researchers, policy makers, and activists have identified as tools that may help hold police accountable (S. E. Walker & Archbold, 2018) . Given the recent events in Minneapolis and the calls from community-based organizations to evaluate police interaction, this study set out to quantitatively assess the relationship between race and police interaction in Minneapolis, Minnesota with a particular focus on the period around the murder of George Floyd. This study makes use of a large open dataset provided by the Minneapolis Police Department that includes more than 170,000 police stops, representing a major opportunity that has been largely unavailable in previous studies of police interaction. In addition, we implement a recent computational advance in generalized estimating equations to handle large datasets with varying cluster sizes, which is important given differences in police activity by neighbourhood. Using these data, we provide highly detailed descriptive information on police stops that occurred in Minneapolis from November 2016, to March 2020, both at the individualand the Census Tract-level, illustrating how police stops disproportionately impact those who are BIPOC. We then quantitatively assess the extent to which race is associated with vehicle and persons searches among those who are stopped, finding further evidence of racial bias in police interactions. Lastly, we highlight sudden and drastic temporal variation in police interaction following the murder of George Floyd. MPD stop data were collected from the Open Minneapolis data portal A few variables, variable transformations, and data cleaning steps require elaboration. Time of stop was converted from a continuous variable to a discrete variable indicating if the stop took place during dusk or sundown using the suncalc package in R (Thieurmel & Elmarhraoui, 2019) . Additionally, reported gender contained four categories: male, female, gender nonconforming, and unknown. Although gender non-conforming was reported in very low numbers, we opted to retain it as a category rather than collapse groups (Ansara & Hegarty, 2014) . We removed stops in which the reason for the stop was truancy or curfew violations, as these reasons were reported in less than 0.01% of all stops. Lastly, we categorized individuals reporting East African for race as Black. The East African race category is available in the police stop data due to the sizable population of Somali and Somali Americans living in Minneapolis. Several descriptive data analysis techniques and visualizations were used to characterize the distribution of police stops in Minneapolis. Racial demographics for the city of Minneapolis were extracted from the US Census. The racial distribution of the city population was compared to the racial distribution of police stops to assess whether police stops by race are proportionally distributed. We then provide summary statistics of the police stop data along several dimensions including pre-stop information on suspect race communicated by dispatch, reported race, time of day, recorded problem, and reported gender. We describe the data using both descriptors of race (pre-and during stop) as reported race reflects post-event information, whereas initial information on race is reported by dispatch when available. Following the descriptive tabulations, we generate several data visualizations. To begin, we graph the absolute number of police stops by race and the percentage of police stops resulting in a person or vehicle search by race. These initial visualizations consider all police stops across the study period without regard to possible temporal or geographic trends. We further characterized the data using time series graphs and thematic maps. The time series graphs describe the racial distribution of police interactions by race for each month-year across the study period. The thematic maps show police stop rates across Minneapolis at the Census Tract level. Specifically, these maps show police stop rates for White and Black race groups for each complete year in the study period (e.g., 2017, 2018, and 2019). These additional visualizations are used to describe temporal or geographic patterns present in the police stop data. We sought to evaluate the relationship between race and being searched by the MPD while accounting for potential confounding factors and to characterize any changes in MPD practices in the wake of George Floyd's murder and subsequent protests. To account for the correlated nature of police stops within a given geographical area, generalized estimating equations (GEE) with a logistic link were employed separately for both outcomes (person search and vehicle search) (Liang & Zeger, 1986) to estimate how police stop behaviours vary among Black people compared to White people across Minneapolis. Two different models were fit for each outcome type. The first involved specifying univariate models that modeled search (the outcome) as a function of race, whether the stop such that (⋅) denotes the logit link function. We will denote the conditional variance of by ( ) and therefore ( ) = ( | ; ) = (1 − ). We assume an exchangeable correlation structure, so we have that ( , ) = . Let = ( 1 , … , ) denote the vector of outcomes for the stops in neighborhood . The correlation matrix of is then given by where is an × identity matrix and is an × matrix of ones. The covariance matrix of which we will denote by ( , ) is given by ( The GEE for the multivariate model then had the same general form as the univariate model with only the marginal mean model changed. Given the number of associations considered, we opted to control the false discovery rate at 0.05 using the Benjamini-Hochberg procedure (Benjamini & Hochberg, 1995) . Information on the race of the individual stopped by police is centrally important to the present percentages across the rest of the categories. Interestingly, Figure 1 Panel B shows that when the race is recorded "unknown" there are relatively few searches. A time-series plot of MPD stops by race is shown in Figure 2 . Hawthorne has some of the highest person search rates, there is a striking difference in the magnitude of the rate between Black and White people. Univariate and multivariate logistic regression using generalized estimating equations (GEEs) were used to model the relationship between race and police searches to account for clustering at the neighbourhood level. In sensitivity analyses, we study the impact of misclassification of individuals' race as Unknown under different assumptions regarding the true, unrecorded race. Figure 4 shows the estimated odds ratios and 95% confidence intervals for the estimated association between Black and a person/vehicle search using different proportions of White-Black random racial replacement. Under the extreme, and implausible, assumption that all those classified as Unknown were in fact White, some of the estimated effects in univariate models were found not to differ significantly from the null hypothesis of no impact of race (Figure 4a , 4b, and 4d). In contrast, under all settings considered for the true race of those classified as Unknown, all the adjusted models support the conclusion that Black individuals have higher odds of person and vehicle searches both before and after the murder of George Floyd. These findings indicate the robustness of our initial results. The results of the present analysis allow for several observations regarding race and policing in Minneapolis. Firstly, we can answer our original question of whether race is associated with being searched by police in Minneapolis. This question has been addressed many times in both lay press and academic research, and our analysis contributes to the ongoing dialogue by using open-source policing data and methods that can account for confounding and correlated data. Crude risks calculated both in the press and in this paper suggest strongly that Black people are more likely to be searched by the police than White people. Our multivariate analyses allow us to take this statement further. Our results tell us that Black people are statistically significantly more likely to have both their person or vehicle searched in Minneapolis when stopped by police as compared to White people even after accounting for several potential confounds including time of day, the problem and the reason for the stop, before and after the murder of George Floyd, and accounting for clustering within neighbourhood and more. Though the effect size slightly dampens between the crude and adjusted models, Black people clearly have increased odds of being searched. Temporal visualization of police stops by race revealed several interesting findings. We observed a sharp decrease in the total number of police stops in the month immediately following the killing of George Floyd. This finding is closely related to the disputed 'Ferguson Effect' (Davey & Smith, 2015; Gross & Mann, 2017; Mac Donald, 2015; Pyrooz et al., 2016; Rosenfeld et al., 2015) , which suggests that consistent, year-over-year declines in the rate of the violent crime since the 1990s began to reverse in the mid-2010s following massive public demonstrations over the 2014 shooting of Michael Brown in Ferguson, Missouri (Mac Donald, 2015) . The implied mechanism underlying the Ferguson Effect is that police disengage from regular behavior following public outcry over their actions (e.g., media coverage of excessive use of force), ultimately resulting in more opportunity for crime (Mac Donald, 2015) . Thus, the Ferguson Effect has two key components: (1) a decrease in police activity and (2) an increase in crime. Empirical analyses examining recent crime rates are generally limited due to a lack of data (Rosenfeld, 2016; Rosenfeld & Fox, 2019) . A large sample of monthly criminal offense data from 81 large US cities found no evidence to support change in overall, violent, and property crime rates, even if select cities experienced increases in homicide in 2014 (Pyrooz et al., 2016) . Similarly, a study using data from 53 large US cities covering the years 2010 to 2015 found no relationship between declines in arrests and homicide rates (Rosenfeld & Wallman, 2019) . However, other studies have substantiated a decrease in police activity (also called 'de-policing') following instances of major public outcry (Shjarback et al., 2017) . For example, a study of police agencies across Missouri found that agencies carried out nearly 70,000 fewer traffic stops in the year following Michael Brown's shooting, and that this effect was especially pronounced in agencies with a higher proportion of African-American residents (Shjarback et al., 2017) . The suggested reasons behind de-policing are varied, including discouragement among police managers (Nix & Wolfe, 2018) , police officer concerns over public and judicial repercussions (Oliver, 2017) , employment turnover following public outcry (Mourtgos et al., 2021) , and other explanations situated in critical cultural theory (Cooper, 2003) . In the case of Minneapolis, our analysis provides quantitative evidence that the proportion of Black people stopped was dramatically reduced while the proportion of 'Unknown' race greatly increased and that this change is correlated to the killing of George Floyd, as policing behaviors in months leading up to the killing remained stable even as COVID-19 measures were implemented across Minnesota. It has been suggested in other areas of the United States that law enforcement agencies purposefully misidentify the race of stopped individuals (Webster, 2021) , although academic research on this is limited (Alwd 6th Ed. Though we were originally interested in Black people in particular, the results highlight the need to also look at the experience of Indigenous people interacting with police in Minneapolis. We found that the odds of being searched was drastically higher among Indigenous people as compared to White people, with an effect size higher than that estimated for Black people. This is consistent with recent research examining the policing of Indigenous peoples in Minneapolis (Gorsuch & Rho, 2019) . A particularly interesting finding was that, among vehicle searches, Indigenous people had higher odds of being searched after accounting for covariates, whereas across all previous models indicated that effect sizes decreased after adjustment. There are several possible explanations for this pattern. This could be due to imprecision in the model, as there were relatively few Indigenous people in the dataset. However, the confidence intervals are narrow, and as such it is also possible that these estimates are in fact accurate. Should this be the case, a finding such as this one could likely be explained by a systematic reason wherein Indigenous people are less likely to have their car searched due to a control variable included in our multivariate analysis. At this stage we do not know exactly why this might be the case, or which variable may be responsible, and further research into the topic should be done to understand the phenomenon. While this study has many advantages, it is also important to mention its limitations. Certain variables considered in the regression analyses had missing data, specifically the pre-race information provided by dispatch. However, Black people still make up the largest proportion of stops by pre-race information, and these proportions are exacerbated when the looking at determined race. Thus, we argue that the amount of missing data is more likely explained by the difficulty of dispatchers providing pre-race information before stops occur. In addition, stops for truancy and curfew violations were removed from the datasets because there were not enough data points. Although, we removed these observations from the final analysis, these observations of this information prevents us from calculating and analyzing outcomes-based tests such as 'hit rates', a commonly used metric in research on police behavior (Persico & Todd, 2006; Shjarback et al., 2017) . Hit rates describe the proportion of searches that lead to additional legal action. While providing another perspective of analysis, hit rates do not address differences in police behavior by race and are nested under a broader discussion on the difficulty identifying the appropriate baseline for the distribution of stopped individuals (S. Walker, 2008) . In addition, the dataset does not contain any variables that have been shown to impact police-citizen interaction during stops, such as characteristics of the police officer, the type of police patrol, information on other individuals involved in the stop, or actions leading up to or taken during the stop (National Highway Traffic Safety Administration (NHTSA), 2020; Sykes & Clark, 1975; Warren et al., 2006) . Our study shows that certain populations, specifically Table 3 . GEE Models: Estimated odds ratios and 95% confidence intervals for association between race and person/vehicle search among individuals who were stopped by the police in Minneapolis, MN before/after the murder of George Floyd compared to White people. Multivariate models control for gender, pre-race coding, problem, reason, whether the stop was after sunset, latitude, and longitude. * indicates odds ratios that are significantly different than 1 with the false discovery rate controlled at 0.05 using the Benjamini-Hochberg procedure. Table 4 . GEE Models: Estimated odds ratios and 95% confidence intervals for association between race and person/vehicle search among individuals who were stopped by the police in Minneapolis, MN after the murder of George Floyd compared to White people. Multivariate models control for gender, pre-race coding, problem, reason, whether the stop was after sunset, latitude, and longitude. * indicates odds ratios that are significantly different than 1 with the false discovery rate controlled at 0.05 using the Benjamini-Hochberg procedure. Police suspicion and discretionary decision making during citizen stops Bad Predictions: How Civil Rights Violations Impact Police Data, Predictive Policing Systems, and Justice Massacre in Minnesota: The Dakota War of 1862, the Most Violent Ethnic Conflict in American History Methodologies of misgendering: Recommendations for reducing cisgenderism in psychological research Racial Prejudice and Police Stops: A Systematic Review of the Empirical Literature Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing The Evidence of Racial Profiling: Interpreting Documented and Unofficial Sources Racism and Police Brutality in America Police Stop Data. Opn Minneapolis Data Portal Blacks face higher risk of drug arrests in White neighborhoods An analysis of the New York City police department's "stop-and-frisk" policy in the context of claims of racial bias Of beauty, sex, and power: Statistical challenges in estimating small effects * Police Stops and Searches of Indigenous People in Minneapolis: The Roles of Race, Place, and Gender Suspect Searches: Assessing Police Behavior Under the U.S. Constitution*. Criminology