In the midst of one of the most heated weeks on the topic of police shootings in years, the Harvard economist Roland Fryer dropped a working paper that turned heads, a study of police shootings in Houston that found no evidence of racial bias in the decision to fire. He did find evidence of racial disparities almost everywhere else:

They were more likely to stop African-Americans than white citizens. In those stops, officers were more likely to draw their guns when the suspect was black, and once the weapon was drawn, they were more likely to point it at someone who was African-American. Blacks were more likely to get handcuffed, thrown against a wall, and pushed down. The racial discrepancy, controlling for circumstance, was present in nearly every situation.

But he didn't find it in shootings. In fact, he found that blacks were less likely to be shot, again controlling for circumstance. And this is a critical distinction. Fryer was looking at the odds of an officer shooting a suspect after a stop is made, not at the total number of suspects shot. But after it got picked up by the New York Times it touched off a debate, not just for its conclusions but its methodology.

It received some mild criticisms from Barnard's Rajiv Sethi, and stronger ones from the University of Minnesota's Michelle Phelps. For instance, Phelps argues that, because blacks are stopped so much more often, the "hit rate" for finding criminality when whites are stopped is going to be higher. But that obviously doesn't mean the total numbers will be higher. "Indeed, evidence from FBI reports of police shootings suggest that when the initial interaction is less serious (e.g. when the suspect has no weapon), racial disparities are the greatest," Phelps writes. "In Ross’s study, the race divide was so large that the rates of police shootings were higher for unarmed black suspects than armed white suspects."

And blacks and Latinos were far more likely to be shot by police in Houston, Harvard doctoral candidate Justin Feldman found in researching a response to Fryer's paper. Feldman in turn pointed to a 2007 study led by the University of Chicago psychology professor Joshua Correll, in collaboration with researchers at the University of Colorado and the University of Denver. It takes a very different approach: Rather than trying to reverse-engineer bias from events captured in data, they set up a lab experiment, as befitting a psychologist's investigation of the issue.

In short, they recruited police officers and civilians from the Denver area, as well as a national sample from a training seminar, to play a video game. "Subjects" in the game were either armed with guns or objects that could possibly look like guns (a wallet, a cell phone, a pop can). It bore a passing resemblance to the arcade game Area 51: points for shooting an armed target, loss of points for shooting an unarmed target or not shooting an armed target.

The world, of course, is not an A/B test. Police do not automatically shoot armed suspects. Trends towards de-escalation are an attempt to reduce the use of force, although that trend has taken hold more substantially since their study was done. Lab experiments are always a simplification of the real world, and this study narrows the situation down by forcing you to decide whether or not to shoot.

Correll and his co-authors did find evidence of bias in the use of force, but their findings also lend experimental credence to Fryer's findings. It's a compelling and nuanced study (emphasis mine):

Most important for our hypothesis, the officers also differed from the community sample in the role that a target’s race played in the placement of SDT [signal detection theory] criteria for the decision to shoot. This difference primarily affected Black targets. When the target was White, all of the samples (Denver community, Denver police, and national police) set a relatively high criterion, and none of the samples differed from one another. But when the target was Black, the community set a significantly lower (more trigger-happy) criterion than the officers.

Generally speaking, the officers were quicker to make the correct decision and more likely to make it. Which is what one would expect, or at least hope, from people with training. And the authors suggest that training may alleviate instinctual bias in use of force, but not eliminate it.

In spite of the fact that police showed minimal bias in the SDT analysis, the officers were similar to the community sample (and to literally hundreds of past participants in this paradigm) in the manifestation of robust racial bias in the speed with which they made shoot/don’t-shoot decisions. Accurate responses to targets congruent with culturally prevalent stereotypes (i.e., armed Black targets and unarmed White targets) required less time than did responses to stereotype-incongruent targets (i.e., unarmed Black targets and armed White targets).

In other words, that bias is still there, showing up in reaction times, but the results suggest that training mitigates it when it comes to the decision to shoot: "the bias evident in their latencies did not translate to the decisions they ultimately made." (Though it's worth keeping in mind that the two studies cover a limited number of cities, and forms and amounts of training can vary.) Obviously this doesn't mitigate the disparities in the raw numbers on police stops, but it does point toward the root of the problem, just as Fryer's seems to. The roots are everywhere: not just at the point of deciding whether or not to shoot, but in the many decisions and policing strategies leading up to that point.

It is both encouraging and discouraging from a policy standpoint, which tends toward concrete, inevitably modest changes. It's encouraging because training seems to do something; making better decisions than the public at large may be a low bar, given the gravity of the job, but it isn't nothing. Even if it only works on what we statistically consider the margins, those margins can mean a life.

It's discouraging because it suggests how the problem is societal rather than one inherent to policing. The authors take this line of thought further:

[O]fficers from the national sample who reported working in communities with (a) high levels of violent crime and (b) high proportions of minority residents showed particularly strong patterns of bias in their latencies. Did their experiences with minority suspects foster associations that made counterstereotypic trials particularly difficult to process?

The situation is almost certainly more complex. It is clear from the analysis of Study 1 that officers serving in heavily (more densely) populated communities also showed greater anti-Black bias in their reaction times. In combination, these variables seem to suggest that racial bias in the decision to shoot may reflect the disproportionate representation of Black people (and perhaps other ethnic minority groups) in low-income, poverty-stricken, and high-crime areas.

This is echoed in a study by UC-Davis anthropologist Cody T. Ross… except for the high-crime part. Ross found that "racial bias in police shootings is most common among police working in larger metropolitan counties with low median incomes and a sizable portion of black residents, especially when there is high financial inequality in that county," but "the results of this study provide no empirical support for the idea that racial bias in police shootings… is driven by race-specific crime rates (at least as measured by the proxies of assault- and weapons-related arrest rates in 2012)."

In short, the greatest predictors of the problem are vast structural inequities, something traced in the Denver study down to the level of individual reaction times. All this reminded me of something Ta-Nehisi Coates wrote this week:

[T]here is a strong temptation to focus on the problems of “implementation,” as opposed to building the kind of equitable society in which police force is used as sparingly as possible.

There is no shortcut out. Sanctimonious cries of nonviolence will not help. “Retraining” can only do so much. Until we move to the broader question of policy, we can expect to see Walter Scotts and Freddie Grays with some regularity.

As big, messy, and slow as the process of reforming policing in America is, compared to fighting inequality, it's a smaller problem that voters have more leverage over. Building an equitable society, on the other hand, is a vast project, and one that requires not just asking more of the police, but of every single one of us.