A Quick Response To A Remark About Traffic Stop Racial Bias and Predictive Policing
In a recent article of mine I urged caution for municipalities considering adoption of predictive policing strategies and advocated that a proper policy framework be in place prior to any implementation. A point belabored in that article was how predictive policing programs could be vulnerable to a “feedback loop” and amplify biases present in policing. If the predictive algorithm used in this type of policing strategy is fed tainted or skewed data, then the algorithm will direct law enforcement officers to police in an inequitable way. After the officer records a crime, that record is added to the data set that the algorithm “learns” from, thereby reinforcing the bias.
The Electronic Frontier Foundation eloquently summarized the systematic problem as one which “… amplifies racial disparities in policing by relying on patterns in police record keeping, not patterns in actual crime”.
My concern about biased policing is broad and I included how this could potentially result in racial, economic, political or religious disparities. I primarily focused on racial bias within the piece as that was the most pertinent issue referenced by critics of the Los Angeles PredPol program (which served as a case study of predictive policing) and racial bias is an obvious way policing problems could manifest.
A reader took issue with my statement, “For example, black drivers are generally stopped by police at a higher rate than white drivers…” and countered that,
…it is the CAR (sic) that is stopped because the police TARGET (sic), among other things, erratic driving, expired licence (sic), false plates, road law infringements.
I felt that the comment brought up a couple important points that warranted a response.
To paraphrase the objection, confounding variables like driver behavior explains the racial disparities in traffic stops and not police racial bias. While the larger argument still holds irrespective of if there is racial bias in traffic stops or not, how do we know if there actually is racial bias? This line of reasoning brings up a methodological question.
First, the semantic emphasis on “car” seems mostly extraneous. Of course a car stopped during a traffic stop implies that a driver is stopped as well. An implication of making the car the sole focus would be that a law enforcement officer is only considering driver behavior and/or the car sui generis.
Second, setting aside the issue of driver behavior as a confounder, the differentiation between “car” and “driver” is not particularly relevant in this context. A few statements to this point:
- We do not know that a given police officer conducting a traffic stop is unaware of the racial identity of the driver. If the police officer knows the race of the driver, that is highly relevant information for the question of racial bias.
- Even if the police officer does not have specific knowledge of the driver’s racial identity, the officer could be either explicitly or implicitly inferring the racial identity by stereotyping the visual appearance of the car. That is, certain makes and models of cars and identifiers like bumper stickers, underglow, interior accessories, spoilers, hood scoops, rims, wraps, aftermarket parts etc. can be more often associated with particular demographics than others.
- A specific racial identity group could populate a particular geographic area more heavily. E.g a police officer could simply guess that a driver is black because he is in “the black part of town”.
Third, in regards to the issue of driver behavior, the citation which was hyperlinked to the statement explicitly considered confounding variables and concluded that a portion of the racial disparity is still attributable to racial bias. The Stanford Open Policing Project’s technical paper explains in depth how certain measures of discrimination are imperfect, namely how older measures may not be comparing similar “pools” of minorities. However, the threshold statistical test used for the aforementioned paper controls for confounders including driver behavior, age and gender.
Fourth, per my original statement, I said, “For example, black drivers are generally stopped by police at a higher rate than white drivers…”, a claim that is fully substantiated by the reference when looking at stop rates per capita for the two populations. Even if a given traffic stop resulting in a statistical racial disparity is itself not attributable to racial bias, the disparity is important given the factors of predictive policing. An overrepresentation in the historic data of a population could result in racially biased policing, even without demographic information because of the nature of predictive policing.
If we knew the exact quantified amount of racial bias, an adjustment function could hypothetically weight the data to correct selection bias. However, “under the hood”, one of the biggest predictive policing programs is essentially built on a relatively simple moving average and does not include demographic information to alter weights.