Tuesday, September 20, 2016

More on police shootings

Over the summer, a paper by Roland Fryer got a lot of attention.  He summarized his findings:  "there are racial differences--sometimes quite large--in police use of force, even after accounting for a large set of controls ... Yet, on the most extreme use of force--offi cer-involved shootings--we are unable to detect any racial di fferences...."  You could restate this by saying that there is more anti-black bias in non-lethal force than in lethal force, and it's not clear if there is any bias (in either direction) in the use of lethal force.  The difference between lethal and non-lethal force was surprising to me--I figured that if there was bias, it would be more pronounced for the more extreme use of force.  I thought his paper was convincing on that point, partly because of the evidence he presented and partly because of a simple comparison to the data on fatal shootings by police that I've written about before.   Blacks comprise 27% of those fatally shot by the police.  This is considerably higher than their share of the total population, but not relative to other forms of negative involvement with the criminal justice system.  For example, blacks make up 39% of those arrested for violent crime.

The difficulty is in figuring out whether 27% is more, less, or about the same as what it would be if police shootings took place without regard to race--that is, if a white person and a black person in the same situation faced the same risk of being shot--which is why Fryer said "we are unable to detect any" racial differences rather than "there are no" or even "there appear to be no."

Although the data Fryer used has a lot more detail, the data I used also has some advantages:  it covers the whole nation and has more cases.  It includes a variable for whether the person who was killed was attacking a police officer, "other," or "unknown" and one for what kind of weapon, if any, they had.  I combined those into a new variable with three values:  people who were not attacking ("other" or "unknown") and unarmed (or "undetermined"), people who were armed but not attacking, and people who were attacking.  I'll call them low, medium, and high levels of apparent threat.  The breakdown of people killed by apparent threat:

            Black   Hispanic   White
low          35%      22%       39%
medium       26%      20%       48%
high         26%      16%       55%

There are statistically significant racial differences--the share of blacks and Hispanics is highest for the lowest threat level. You could also put it in terms of the chance that a person will be killed by the police when they are unarmed and not attacking:  blacks have about six times the risk of non-Hispanic whites, and Hispanics have about three times the risk.

The limitation of this comparison (and the ones Fryer did) is that we don't know the number of people who were in a comparable situation but were not fatally shot.  So it's possible that blacks and Hispanics were just less likely to be in the low-threat relative to the high threat situations.  That doesn't seem likely to me--the low-threat situations can include a wide variety of circumstances (e. g., bystanders who were killed by accident), so it seems the racial distribution of people in them in them should be closer to that in the general population.  It's also possible that police are unbiased in low-threat situations but less likely to kill blacks and Hispanics in high-threat situations.  However, the most plausible interpretation seems to be that there is some anti-black bias in fatal police shootings.

PS:  There were a total of 1,499 fatal shootings in the 18 months covered by the data: 133 low threat, 418 medium, and 948 high.


  1. I did a state-level (+ DC) analysis of police killings per million population.

    Most powerful variables:
    % black and its square
    violent crime rate
    incarceration rate
    Dummy for D.C.

    1. Interesting. What was the exact relationship to % black? Just looking at the rates, I didn't think that there would be much of one.

  2. I get (pctblk)*-0.516 and (pctblk2)*0.009

    I fit this using a robust regression though not much different from OLS estimates of:

    (pctblk)*-0.648 + (pctblk2)*0.012

    all net of violent crime rate, incarceration rate, and D.C. dummy variable