We have evidence of racial discrepancies in police use of force. But some studies miss the extent of the problem by using data that's already biased.
fivethirtyeight.com/features/why-s…
@jazzmyth's awesome chart helps me explain this paper by @dean_c_knox, @conjugateprior and @jonmummolo:
fivethirtyeight.com/features/why-s…
@jazzmyth's awesome chart helps me explain this paper by @dean_c_knox, @conjugateprior and @jonmummolo:
In short, as the paper explains, looking at use of force just among the set of people police have *stopped* isn't enough to let you correctly estimate racial discrepancies. If there's bias in who gets stopped in the first place, that confounds your estimate.
A simple example: Even assuming use of force among stopped people is equal across race (which is unrealistic), bias in stops means that your denominator is wrong. More Black people have been stopped without cause, so "equal" treatment is actually evidence of discrimination:
But the system doesn't start at the police encounter. Police are not equally deployed across neighborhoods. As @AlliPatter said, more exposure to police makes experiencing police violence more likely. The bias “just compounds all the way down the line,” @dean_c_knox told me.
I wrote the article before a preprint came out yesterday that questions the Knox, Lowe and Mummolo paper. It's caused a reasonable amount of debate on causal inference Twitter (e.g. & ).
From my read, while the authors identify an assumption - subset ignorability - which solves the treatment and mediator ignorability problems, KLM show (as I try to illustrate in my article!) is that subset ignorability isn't a realistic assumption to make in this context.
Other useful context:
and
and
