NYC's Compstat & Predictive Policing - Transformed American Policing
The Story - The Impact - How Prevention Became the Driving Force

The black box of justice: How secret algorithms have changed policing

Crime prevention efforts increasingly depend on data analysis about neighborhoods. But there’s a lot we don’t know—and vast opportunities for bias.

The story of predictive policing begins in the 1990s with a process developed by the New York Police Department. Today New York is one of the safest big cities in America. In 2018, 289 people were murdered in the five boroughs. The city’s murder rate—3.31 per 100,000 people—was the lowest measured in 50 years

In 1990, it was a different city: 2,245 people were murdered, a rate of around 31 per 100,000 (the city’s population increased markedly in the intervening 28 years). Here’s what the New York Times said about its hometown at the end of 1990: “The streets already resemble a New Calcutta, bristling with beggars. . . . Crime, the fear of it as much as the fact, adds overtones of a New Beirut. . . . And now the tide of wealth and taxes that helped the city make these streets bearable has ebbed. . . . Safe streets are fundamental; going out on them is the simplest expression of the social contract; a city that cannot maintain its side of that contract will choke.” To stop the choking, the city knew it had to get crime under control, but the police didn’t have the right information.

In 1993, New York elected its first Republican mayor in almost 30 years—an ambitious former federal prosecutor named Rudy Giuliani. It may seem hard to believe now, but back then Rudy seemed to have at least a modicum of political nous. He ran a law-and-order campaign, and soon after taking office appointed Bill Bratton, formerly Boston’s police commissioner, then head of New York City’s Transit Police, to head the NYPD.

Bratton soon ran into a problem: he found that his new department had no focus on preventing crime. At the time, that was not unusual. Police did not have crystal balls. They saw their job as responding to crimes, and to do that, the crimes had to have happened already. They were judged by how quickly they responded, how many arrests they made, and how many crimes they solved.

Police did not have access to real-time crime data. And, as Lou Anemone, then the NYPD’s highest-ranking uniformed officer, explained in a 2013 report, “The dispatchers at headquarters, who were the lowest-ranking people in the department, controlled field operations, so we were just running around answering 911 calls. There was no free time for officers to focus on crime prevention.”

So the department began using computers to crunch statistics. For the first time, crime data became available in near-real time. The department also began calling regular meetings, where commanding officers grilled captains and lieutenants, asking what they were doing to combat crime in their precincts. The department named this practice—the agglomeration and mapping of real-time crime data, as well as the legendarily terrifying meetings—Compstat. Its progenitor, Jack Maple, said it stood for “computer statistics or comparative statistics—no one can really be sure which.”

Maple’s invention rested on four main principles: accurate and timely intelligence, rapid deployment, effective tactics, and relentless follow-up and assessment. It sounds simple, even obvious: of course police should try to prevent as well as respond to crime; and of course, to do this effectively, they will need as much data as possible. Neither of those ideas were obvious at the time.

At around the time that Compstat was put in place, crime began falling. I do not intend to analyze, litigate, or even hypothesize about the precise causal relationships of Compstat to falling crime. With apologies to Dorothy Parker, eternity isn’t two people and a ham; it’s two criminologists arguing over the causes of the late twentieth-century crime drop. Perhaps the key was altered police practices. Perhaps it was changing demographics. Perhaps it was regulations that got the lead out of household paints and gasoline. Perhaps it was some combination of environmental, political, and demographic factors. Determining the correct answer is, mercifully, beyond the scope of both this book and my time on this planet.

Still, the fact is that Compstat transformed American policing (in much the same way as, and not long before, data-driven approaches transformed baseball). Other departments adopted it. New York has maintained and tweaked it. Today, a major-metro police department that considers only response and not prevention, and that purports to fight crime without data and accountability, is all but unthinkable. Predictive algorithms seem to be the natural outgrowth of the Compstat-driven approach: perfectly suited to departments concerned about preventing crime, not just responding to it.
How such algorithms make their predictions is not clear. Some firms say that publicly revealing the precise factors and weights that determine their predictions will let criminals game the system, but that hardly passes the smell test: a guy isn’t going to decide to snatch wallets on Thirty-Fourth Street today because he knows his local police department uses XYZ Safety Program, and their algorithm currently forecasts high crime—and hence recommends increased police presence—on Thirty-Eighth Street.

The algorithms are proprietary, and keeping them secret is a matter of commercial advantage. There is nothing inherently wrong with that—Coca-Cola keeps its formula secret, too. And, as I said earlier, there is nothing inherently wrong with using algorithms. But, as Phillip Atiba Goff of the New York University’s Center for Policing Equity said to me, “Algorithms only do what we tell them to do.” So what are we telling them to do?

Jeff Brantingham, an anthropologist at the University of California, Los Angeles, who cofounded PredPol, told me he wanted to understand “crime patterns, hot spots, and how they’re going to change on a shift-by-shift or even moment-to-moment basis.” The common understanding of the geography of street crime—that it happens more often in this neighborhood than that one—may have some truth in the long run, but has limited utility for police shift commanders, who need to decide where to tell their patrol officers to spend the next eight hours. Neighborhoods are big places; telling police to just go to one is not helpful.

So PredPol focuses on smaller areas—those 150-by-150-meter blocks of territory. And to determine its predictions, it uses three data points: crime type, crime location, and crime date and time. They use, as Brantingham told me, “no arrest data, no information about suspects or victims, or even what does the street look like, or neighborhood demographics. . . . Just a focus on where and when crime is likely to occur. . . . We are effectively assigning probabilities to locations on the landscape over a period of time.” PredPol does not predict all crimes; instead, it forecasts only “Part 1 Crimes”: murder, aggravated assault, burglary, robbery, theft, and car theft.

PredPol is not the only predictive-policing program. Others use “risk-terrain modeling,” which includes information on geographical and environmental features linked to increased risks of crime—ATMs in areas with poor lighting, for instance, or clusters of liquor stores and gas stations near high concentrations of vacant properties. Other models include time of day and weather patterns (murders happen less frequently in cold weather).

All of these programs have to be “trained” on historical police data before they can forecast future crimes. For instance, using the examples above, programs treat poorly lit ATMs as a risk factor for future crimes because so many past crimes have occurred near them. But the type of historical data used to train them matters immensely.

THE BIAS OF PRESENCE

Training algorithms on public-nuisance crimes—such as vagrancy, loitering, or public intoxication—increases the risk of racial bias. Why? Because these crimes generally depend on police presence. People call the police when their homes are broken into; they rarely call the police when they see someone drinking from an open container of alcohol, or standing on a street corner. Those crimes often depend on a police officer being present to observe them, and then deciding to enforce the relevant laws. Police presence tends to be heaviest in poor, heavily minority communities. (Jill Leovy’s masterful book Ghettoside: A True Story of Murder in America is especially perceptive on the simultaneous over- and underpolicing of poor, nonwhite neighborhoods: citizens often feel that police crack down too heavily on nuisance crimes, but care too little about major crimes.)

Predictive-policing models that want to avoid introducing racial bias will also not train their algorithms on drug crimes, for similar reasons. In 2016, more than three-fourths of drug-related arrests were for simple possession—a crime heavily dependent on police interaction. According to the Drug Policy Alliance, a coalition that advocates for sensible drug laws, prosecutors are twice as likely to seek a mandatory-minimum sentence for a black defendant as for a white one charged with the same crime. I could go on for a few hundred more pages, but you get the idea: America enforces its drug laws in a racist manner, and an algorithm trained on racism will perpetuate it.

Police forces use algorithms for things other than patrol allocation, too. Chicago’s police department used one to create a Strate- gic Subject List (SSL), also called a Heat List, consisting of people deemed likely to be involved in a shooting incident, either as victim or perpetrator. This differs from the predictive-policing programs discussed above in one crucial way: it focuses on individuals rather than geography

Much about the list is shrouded in secrecy. The precise algorithm is not publicly available, and it was repeatedly tweaked after it was first introduced in a pilot program in 2013. In 2017, after losing a long legal fight with the Chicago Sun-Times, the police department released a trove of arrest data and one version of the list online. It used eight attributes to score people with criminal records from 0 (low risk) to 500 (extremely high risk). Scores were recalculated regularly—at one point (and perhaps still) daily.

Those attributes included the number of times being shot or be- ing the victim of battery or aggravated assault; the number of times arrested on gun charges for violent offenses, narcotics, or gang affiliation; age when most recently arrested; and “trend in recent criminal activity.” The algorithm does not use individuals’ race or sex. It also does not use geography (i.e., the suspect’s address), which in America often acts as a proxy for race.

Both Jeff Asher, a crime-data analyst writing in the New York Times, and Upturn, a research and advocacy group, tried to reverse-engineer the algorithm and emerged with similar results. They determined that age was a crucial determinant of a person’s SSL score, which is unsurprising—multiple studies have shown that people tend to age out of violent crime.

Shortly before these studies were published, a spokesman for the Chicago Police Department said, “Individuals really only come on our radar with scores of 250 and above.” But, according to Upturn, as of August 1, 2016, there were 280,000 people on the list with scores over 250—far more than a police department with 13,500 officers can reasonably keep on its radar. More alarmingly, Upturn found that over 127,524 people on the list had never been shot or arrested. How they wound up on the list is unclear.

Police have said the list is simply a tool, and that it doesn’t drive enforcement decisions, but police have regularly touted the arrests of people on the SSL. The algorithm’s opacity makes it unclear how someone gets on the SSL; more worryingly, it is also unclear how or whether someone ever gets off the list. And the SSL uses arrests, not convictions, which means some people may find themselves on the list for crimes they did not commit.

An analysis by reporters Yana Kunichoff and Patrick Sier published in Chicago magazine found that just 3.5 percent of the people on the SSL in the dataset released by the CPD (which covered four years of arrests, from July 31, 2012, to August 1, 2016) had previously been involved in a shooting, either as victim or perpetrator. The factors most commonly shared by those on the list were gang affiliation and a narcotics arrest sometime in the previous four years.

Advocates say it is far too easy for police to put someone into a gang-affiliation database, and that getting into that database, which is 95 percent black or Latino, reflects policing patterns—their heavy presence in the mostly black and Latino South and West Sides of the city—more than the danger posed by those who end up on it. The above analysis also found that most black men in Chicago between the ages of twenty and twenty-nine had an SSL score, compared with just 23 percent of Hispanic men and 6 percent of white men.

Perhaps mindful of these sorts of criticisms, the city quietly moth- balled the SSL in late 2019—and, according to the Chicago Tribune, finally stopped using it in early 2020.

Article originally published on fastcompany.com
 

Commissioner Bratton spoke to the LP community at the Daily's
'Live in NYC at the NRF Big Show' event in 2016.