Thursday, August 13, 2015

Detecting Discrimination In Policing (Or, The Dangers of Counterfactual Causal Thinking…)

Guest Blogger

Issa Kohler-Hausmann

For the Symposium: Deconstructing Ferguson One Year Later

How do we know when policing is discriminatory? Statistical evidence is one way to show discrimination, especially in legal challenges to police practices when the smoking gun of invidious intent is hard to find. Plaintiffs seek to demonstrate that observed racial or ethnic disparities in policing outcomes (stops, misdemeanor arrest, police murders, etc.) are because of race and not something else. But what does it mean to show discrimination because of race?

Discrimination is, to borrow from Bernard Williams, a “thick ethical concept” that simultaneously describes and evaluates. Thick ethnical concepts are both “world-guided”—in the sense that facts about the world and how it is organized determine the applicability of the concept—and “action guiding”—in the sense that if one has properly applied the term to a situation it provides a moral reason to pursue a course of action (or desist from a course of action).

These philosophical abstractions come crashing to the ground with bloody urgency around the question of discrimination in policing. When we set out to detect discrimination in a particular field of social activity, we ought to be attentive to the fact that methodological debates about how to identify the phenomenon are not only debates about what the phenomenon is but also necessarily debates about what is right—about the “good,” “fair,” or “proper” way of conducing business in that field.

We can explore these intertwined conceptual, legal and political questions by looking at how police discrimination is alleged and measured by in two arenas: social science studying police outcomes and legal challenges to police tactics. I am particularly interested in how those two arenas interact and mutually inform a particular concept of discrimination through the use of quantitative methods to prove discrimination in police cases. In both of these arenas—and especially when they interact in the form of expert statistical evidence—we see the emergence of a particular concept of discrimination that I call the counterfactual causal notion of race discrimination. The counterfactual causal notion of race discrimination conceptualizes race as a treatment that triggers an outcome. Briefly, the causal effect of treatment t on unit u is defined by comparing some measure of interests on u, Y(u), between a world in which u experienced the treatment (t=1) and the world in which it did not, (t=0): Yt=1(u) – Yt=0(u). For purpose of my discussion, assume the unit is a space (such as a police precinct) and the treatment is a continuous variable (the percent of the space’s population that are Black or Hispanic individuals). Discrimination, on this account, is detected when all other facts about the spatial unit are “controlled for” and there is still some persistent racial discrepancy in observed police outcomes.

Why is this important? It is important because the way courts consume quantitative evidence too easily glosses over what are essentially conceptual and normative questions and make them appear to be technical or methodological questions. What we are arguing about when we argue about whether or not statistical evidence provides proof of discrimination is precisely what we mean by the term discrimination. Similarly, what we are arguing about when we argue about what ought to be “controlled for” in a quantitative exercise to detect discrimination is what are the fair and authorized bases for making decisions and allocating resources.

I will illustrate these arguments with data from an ongoing project with Cong Peng about policing of misdemeanor crimes in New York City. I argue that interpreting the coefficient on race or ethnic composition from a regression model of police practices as an isolated “treatment” effect often entails abstracting from the very social forces that produce race as a salient social category in the first place (There is a long debate about whether race can be thought of as a “cause” in the counterfactual framework. See for example, Zuberi, Sen & Wasow, Holland, Zuberi & Bonilla-Silva). It can make race appear to be an essential, obvious trait that somehow mechanically produces outcomes by virtue of physical attributes as opposed to through entrenched social systems of stratification and signification.

Starting in the early 1990s misdemeanor arrests exploded as part of a policing strategy, often referred to as Broken Windows or quality-of-life policing. Figure 1, graphing the number of misdemeanor arrest events by the race or ethnicity of the arrestee from 1990 to 2014, shows that these policies have been highly concentrated on Black and Latino populations. In 2014, 46% of misdemeanor arrests were of Black individuals, 34% of Hispanic, 14% (non-Hispanic) white and 6% of “other.”

            Misdemeanor arrests are highly spatially concentrated in precincts that are primarily Black or Hispanic. Figure 2 shows the population adjusted density of misdemeanor arrests by precinct. Of the top 20, only two precincts have more than 30% (non-Hispanic) white population.[1]

Figure 2

The most common response to this is that police resources are allocated in precincts that have higher serious felony crime. Such was the defense of the City of New York and the NYPD in the Floyd case challenging the stop, question and frisk program (SQF) on 4th and 14th Amendment grounds. The City argued that if they could show there was no “effect of race” on the level of stops in certain spaces (precincts or census blocks) after controlling for a number of other variables—meaning no little stars after the % Black or Hispanic variable indicating statistical significance at conventional levels—then this was evidence that the SQF program was not discriminatory. And what were those other variables that needed to be controlled for in order to get at the true effect of race? To drastically simplify a complicated debate, those other variables included the social and economic conditions of the space, patrol strength, and some measure of criminal activity. It was taken for granted by both parties that a true measure of the “effect of race” was had only after controlling for such things.

My point is not to criticize quantitative methodology or legal strategy in this or any other case. My point is a conceptual one: In presenting quantitative evidence to courts, it is a mistake to talk as if we have gotten at the true effect of race by modeling it as a counterfactual treatment. In fact, I contend that to do so cements an already predominate and problematic understanding about race in public and legal discourse: one that is distressingly dehistoricized and desocialized.

Consider that in New York City, as in so many US cities, Black and Hispanic population in a precinct and measures of socioeconomic disadvantage have been highly correlated over extended periods of time. Figures 3 through 5 show the persistence of select measures over four decennial censuses, 1980 to 2010. Or consider the fact that the mean violent felony complaint rate in a high Black or Hispanic population precincts (75% or more) is 2.6 times that that in high white population precincts. In fact, there is not a single high white population precinct (75% or more white) that has a violent felony complaint rate at the mean value of high Black or Hispanic precincts.

Figure 3

Figure 4

Figure 5

And that is precisely the problem that discrimination doctrine should be seeking to remedy. As others have pointed out more eloquently than I, the story of state action in minority neighborhoods is one of dual state failure: failure to secure the conditions of safety and flourishing and then the use of intensive and aggressive police tactics to address the predictable social disorder and violence that occurs as a result (See for example, Coates, Miller, Fagan & Mears). The story about why there are high concentrations of poverty, unemployment and violence in predominantly Black and Hispanic urban neighborhoods is the story about the making and remaking of race in America, one that the state has been deeply implicated in from the start.  

Imagine that we can make the little stars of statistical significance on the variable measuring minority composition in a model of misdemeanor arrests go away by including other variables about precincts such level of schooling, unemployment, poverty, health, violence, etc. What exactly have we isolated? The number of people in a given space with designated physical attributes? Questioning this is what it means to be a constructivist about race and ethnic categories because these are not obvious natural human divisions but social understandings that come to have meaning and import from the historic ordering of social and economic affairs. Said another way, if “% Black” or “% Hispanic” were no longer highly correlated with measures of socioeconomic disadvantage then the very meaning of the categories would be different—they would not have the same “effect” in the world as they currently do.

If a court finds that statistical evidence fails to prove discrimination because the correlation between race and a policing outcome does not survive controlling for the historical covariates of race, then we should all be clear—including those learned jurists who hold the power to make such findings—that such a claim does not go to the “world-guided” aspect of the term discrimination (facts about how the world operates) but rather about the “action-guiding” aspect of it (claims about what we ought to do cognizable of those facts). It would be false to say that a court denied a discrimination claim because quantitative evidence isolated a true measure of whether or not race caused the outcome by controlling a host of factors, factors that have historically produced race as a salient social category in the first place. It would be true, sadly, to say that a court held that in so far as race produces outcomes through the mechanisms of high unemployment, poverty, or violence in these spaces, well, the law does not call upon the state to act upon those facts.

Any doctrine that thus absolves the state of its twin failures is a misguided one indeed.

Issa Kohler-Hausmann is Associate Professor of Law at Yale Law School. You can reach her by e-mail at

[1] Those precincts are Midtown North and Midtown South, both with very low residential population and high concentrations of transportation, shopping and tourism centers.

Older Posts
Newer Posts