Racial Discrepancies in Police Shootings

When the government decided not to use force to remove armed antigovernment protesters in Oregon, many people asked if the police would’ve responded violently had the protesters not been white. It’s a question, of course, that has been asked increasingly often of late: do the police show racial bias in whom they shoot? That is, given a white American and a black American behaving identically, are they are more likely to shoot the black American?

Prior to this year, there was no good data on police shootings. If it seems amazing that the American government wasn’t keeping proper track of how many of its citizens its police shot, I find this even more so: when independent organizations like the Guardian and the Washington Post began tracking police shootings, they found that police shot about 1,000 people in 2015 -- a count more than double that recorded by official FBI statistics.

This new data makes it clear that white, black, and Hispanic Americans shot by police are behaving very differently.

To be clear: most people shot by police are armed and are attacking. But there are substantial racial discrepancies. Black and Hispanic Americans shot by police are younger than white Americans shot by police, more likely to be unarmed, and more likely to not be attacking [1]. White Americans are more likely to be mentally ill.

	White Americans Shot By Police	Black Americans Shot By Police	Hispanic Americans Shot By Police
Average Age	40	32	33
Was Unarmed	7%	15%	11%
Was Not Attacking	19%	29%	40%
Showed Signs of Mental Illness	32%	15%	20%

These disparities are large, statistically significant, and and consistent with previous analyses -- black Americans shot by police are more than twice as likely as white Americans shot by police to be unarmed, and more than 50% more likely to not be attacking.

When I mapped shootings to zipcodes, I found that black Americans are also disproportionately shot by police in poor zipcodes: poorer than those in which white Americans are shot by police, and poorer than zipcodes in which the average black American lives [2]. Police shootings overall occur in zipcodes which are poorer and blacker than that in which the average American lives. (White Americans are also shot by police in poorer zipcodes than those in which the average white American lives; I found no large disparity for Hispanic Americans. These findings are broadly consistent with FiveThirtyEight’s analysis of a smaller dataset; I looked at a few other things which I describe in this footnote [3].)

Now, one obvious explanation for the fact that white Americans are less likely to be unarmed when shot is that the police show racial bias in whom they shoot: given a white American and a black American behaving identically, they are more likely to shoot the black American. But there are other explanations as well. One is that white Americans are interacting with police differently: to take an extreme hypothetical, if every white American interacting with the police was carrying a gun, we would expect every white American shot by police to be carrying a gun as well.

To check whether white Americans were more likely to be armed when interacting with the police, I looked at FBI data on crimes committed in the United States in 2013 from the Uniform Crime Report (UCR). I checked whether white Americans were more likely to carry weapons when committing the same crimes, focusing on white and black Americans because I could not find good data on ethnicity in the UCR. I checked both whether the perpetrator was carrying a gun and whether they were carrying any weapon at all, and I examined more than a dozen of the most common crimes where perpetrators frequently carried weapons [4]. And I actually found the opposite: black Americans were in general more likely to carry weapons when committing the same crimes [5]. Maybe black Americans are just stopped by police doing less severe crimes, where they’re less likely to carry weapons? I could not find evidence for this either in the UCR.  So it’s not that all black Americans who interact with police are more likely to be unarmed; only those shot by police are.

I want to be clear: this analysis doesn’t prove the police are more likely to shoot a black American than a white American behaving identically. In an ideal world, we’d have data on every police encounter and we could look at whether race was significantly correlated with whether someone was shot when we controlled for other factors. But instead we have one dataset of people shot by police (compiled by non-government organizations) and one dataset of some police encounters (compiled by the government, and lacking many minor police encounters like traffic stops) which has completely different features. I can’t think of a way to match up these datasets neatly enough to rule out all plausible alternate explanations for the data. Maybe black Americans have more unarmed encounters with police which are too minor to be recorded in the UCR but might still produce police shootings; white Americans are in general more likely to own guns. Or maybe there’s a correlation with a third variable, like location: maybe black Americans tend to live in areas where police are more likely to shoot unarmed people of any race. (Neither of these alternate explanations is reassuring, of course, since they still imply that systemic racial disparities are a major factor in police shootings, but they do change our understanding of why they occur.)  All these explanations are very hard to rule out without a single complete dataset; let me know if you have ideas or other approaches.

But I also want to be straightforward with you about what I believe. Yes, if I had to bet, I’d bet that police bias plays a role in explaining these results. I believe this for three reasons. The first is that it’s a clean explanation for data which is otherwise somewhat difficult to explain. The second is that we know that bias exists at many other levels of the criminal justice system, and it would be weird if it suddenly disappeared when an officer was pointing a gun. The third is that experiments from social psychology do find that police officers are quicker to shoot black suspects in simulations -- though, importantly, they’re less biased than civilians are. (I liked this summary of relevant social science research.)

At the very least, this data points out a flaw in the argument, made by MacArthur Fellow Sendhil Mullainathan, that “eliminating the biases of all police officers would do little to materially reduce the total number of African American killings”. He bases this on the fact that while black Americans are much more likely to be shot by police than the general population, they are also much more likely to be in police encounters than the general population, and their odds of being shot given that they are in a police encounter are roughly the same. The problem is that this argument doesn’t control for the nature of the police encounters. For example, if black Americans were only stopped by the police when jaywalking, and white Americans were only stopped by the police when brandishing large knives, it would be very weird if both groups were shot at equal rates. The new data on police shootings shows us that this objection is more than hypothetical: white and black Americans shot by police are in fact behaving differently, and it is entirely possible that police bias plays a role. (I contacted Mullainathan for comment and will update this if he replies.)

In closing: we need better data -- a single, complete dataset of police shootings that includes data on all police encounters. To quote the director of the FBI, the lack of data is “embarrassing and ridiculous”. It is not the obligation of the Washington Post or the Guardian to cobble together flawed datasets on police shootings; it is the obligation of the state that annually shoots 1,000 of its own citizens in ways that show disturbing correlations with race.

The code for this analysis is available on GitHub. I am a computer scientist with no background in criminology and this data is complex, so I welcome your suggestions, corrections, or extensions via email. Thanks to Simone Landon, Leah Pierson, Nat Roth, Shengwu Li, and Seth Stephens-Davidowitz for helpful thoughts.

Notes:

[1]  This uses the Washington Post’s data. All these differences are statistically significant (p < .002) using logistic regression regardless of whether I coded race as a categorical variable or used “is_white” as a boolean variable. I examined only black, white, and Hispanic Americans because they were the only large groups in the data.

[2] I used the Guardian’s data, mapped it to zipcode using Google’s API, and compared to zipcode data from the American Community Survey. I also wanted to control for crime rate by sampling from the UCR rather than the overall population distribution, but unfortunately, so far as I can tell, the UCR does not include zipcode information for each crime; it’s only down to the police department level, and one department may include many zipcodes.

[3] I looked at a number of other zipcode features as well. Black Americans were shot in zipcodes that were about 33% black, about the same as the zipcode in which the average black American lives; white Americans were shot in zipcodes that were about 8% black, about the same as the zipcode in which the average white American lives; Hispanic Americans were shot in zipcodes which were about 50% Hispanic, slightly more than zipcodes in which the average Hispanic American lives (46%). (These numbers also speak to the extraordinary racial and ethnic segregation in living patterns.) I also examined whether black Americans tended to be shot in zipcodes with greater racial income inequality, but found no difference. My analysis disagrees with FiveThirtyEight’s on one point: they do not find that black Americans are shot in poorer neighborhoods than those in which black Americans live. I also looked at whether victims of police shootings were more likely to be unarmed when shot in poorer or blacker neighborhoods; I did not find a significant effect.

[4]  I compared white and black Americans because I could not find good data in the UCR on Hispanic origin.

[5] This remained true when I ran a multiple regression controlling for the age, race, sex, and type of crime simultaneously, and was true regardless of whether I looked at all weapons or only at guns.