How computers are predicting crime - and potentially impacting your future

Four months before he allegedly took part in a botched robbery that left a Spring Garden father bleeding to death in front of his young daughter, Maurice Roberts was assessed for the likelihood that he would commit just such a violent crime.

Philadelphia’s Adult Probation and Parole Department won’t say how he scored.

And the answer is hidden in the recesses of a computer algorithm that officials don’t want to talk about.

That reluctance came as a surprise to Richard Berk, a criminologist from the University of Pennsylvania who helped create the algorithm.

“The sad thing is you risk shooting yourself in the foot when you behave as if you have something to hide,” Berk said. “There’s nothing to hide.”

Charles Hoyt, head of the Probation and Parole Department, stood by the risk-assessment tool, which he said was assessed annually and “performing as expected.”

“The results so far have supported its continued use,” he said.

Still, the department’s unwillingness to release details about the internal workings of the program, which is used to manage supervision for nearly every offender under its watch for the last eight years, strikes at concerns that have been simmering as Philadelphia prepares to create a similar computer model for use in bail decisions.

Some who are watching that process closely have questioned whether the tool will be racially biased, whether the factors it weighs will be made public, and, fundamentally, whether a computer algorithm should play any role in deciding a person’s future.

The debate is sure to be rigorous, as it has been in the dozens of other jurisdictions across the country already using risk-assessment tools to help guide decisions about bail, sentencing, and parole.

The tools, like judges, are bound to make bad forecasts that could lead to the release of a suspect better kept incarcerated until trial or the over-supervision of a parolee who might then struggle to keep a job.

The question that divides the criminal justice world is whether risk-assessment tools make the imperfect process used now better or worse.

Looking into a person’s past

Judges for decades have looked into a person’s past when weighing that person’s future.

The idea of a computer making that calculation is less familiar.

Philadelphia is considering taking that step for bail decisions as part of a $3.4 million grant from the MacArthur Foundation aimed at reducing disparity in the prison system and the overall prison population by a third in three years.

Each tool is unique, but the simplest use a point system to weigh factors such as a person’s age, the seriousness of the offense, and how many times that person failed to appear in court in the past.

The model used by Philadelphia’s Probation and Parole Department, and which Berk is expected to develop for use in the city’s bail system, is known as a random-forest tool and, as the name suggests, far more complex.

Consider one tree in that forest that starts with a question and branches off into many more. How old is the person? Under 26, go left. Over 26, go right. How old was he at first arrest? Over 18, go left. Under 18, go right. Is the current offense violent? No, go left. Yes, go right.

A single risk-assessment tool can have hundreds of trees, each with about a dozen splits.  The computer runs through them all in a matter of seconds to reach its recommendation. In almost all cases, those scores merely guide judges or bail magistrates, who still make the ultimate call.

The tools are controversial, given the potential for bias in the data they use.

Race and zip code (which is tantamount to race in some of Philadelphia’s highly segregated communities) clearly carry bias.

Other factors are less clear cut.

Is it fair, for instance, to use a person’s age at first arrest, given that juveniles in high-crime, minority neighborhoods are more likely to interact with police than their peers in less patrolled neighborhoods? What about using arrests at all, rather than convictions, given how many charges are dropped before trial?

Kevin Harden Jr., chairman of the Philadelphia Bar Association’s’ criminal justice section, said he worries that risk-assessment tools will become “self-fulfilling prophecies,” baking existing inequities deeper into the criminal justice system.

Proponents argue that such tools actually reduce disparity in the system by treating everyone the same, while judges now use the facts in front of them subjectively.

City and court officials, who have shared little about the forthcoming tool, see that potential and wrote in their application for the MacArthur grant that it would “introduce objectivity to the release decision, thereby reducing jail admissions, racial and ethnic disparities, and income-based disparities.” They have said the tool will not consider race or zip code.

Cherise Fanno Burdeen, CEO of the Pretrial Justice Institute, agreed the tools can improve the current system.

“The job of the tool at this juncture is to take a decision that if left to complete subjectivity will result in biased outcomes, and has resulted in biased outcomes for decades, [and] introduce some neutralization,” Fanno Burdeen said.

But she qualifies that endorsement if there is not transparency.

“Whatever increase in predictive power you’re getting by using random-forest modeling, you have to weigh that against the offset of removing transparency,” she said. “Let’s say you get an accuracy increase of 5 percent. Is it worth it?”

The debate is an ethical minefield. Standing in its center is Berk.

He said critics hold random-forest tools to unrealistic standards when what really needs to be asked is: Do they predict future behavior more accurately than judges? They do, Berk said.

He said the tools can be even more accurate but need to use all the data available, even the controversial factors such as race and zip code. Take out known predictive factors and you will have more homicides, more robberies, more assaults, he said.

“This is where we’ve got to learn to be grownups,” Berk said. “Grownups means we’ve got to live with trade-offs.”

Weighing the trade-offs

Officials at Probation and Parole weighed those trade-offs before launching their risk-assessment tool in 2009.

The tool rates how likely an offender is to commit a crime, and of what magnitude, within two years. It does so based on the histories of thousands of other people who have been through the department and how they ultimately fared. Those ratings help determine how often a person has to check in with a parole officer.

Gabriel Roberts, a spokesman for the city’s courts, said the department “does not discuss the risk tool or its operation” with individual offenders. Several criminal defense lawyers told the Inquirer and Daily News that they have little idea how it works. Even Keir Bradford-Grey, head of the city’s public defenders’ association, said she doesn’t know what factors it weighs.

“They’re not introduced in court, so there’s no forum in which that information would be shared with us,” she said.

Much of what is publicly known about the tool comes from a 2012 report for the Department of Justice by the University of Pennsylvania team that created it. According to the researchers, at the time it weighed 12 factors.

The most influential predictors of future behavior, they found, were how many times the offender had stayed in a city prison, residential zip code, how long it had been since the person’s last serious offense, the person’s current age, and the age at the time of the first adult arrest.

The three least influential factors were the age at first juvenile offense, the number of serious crime charges on the current case, and the number of past charges for sexual offenses. The researchers said those factors were included because officials believed that they were “politically necessary” to defend the model to others in the criminal justice system. They said including them didn’t hurt the algorithm’s accuracy.

James Funt, a Philadelphia criminal defense lawyer, bristled at learning the tool considered time spent in a city prison.

“It’s an unfair factor that applies in a profoundly disproportionate way toward indigent folk,” he said.

He and others said that beyond that, people have a right to know what is in a risk-assessment tool that can impact their future.

“If our clients aren’t being told, if this is a secret in probation or courts, then it’s a secret that needs to be outed,” lawyer Paul Hetznecker said.

Hoyt, head of the Probation and Parole Department, said in a statement that the tool allows the department to manage large caseloads with limited resources. He said it primarily weighs “probationers’ and parolees’ criminal histories” and no longer considers zip code, but he declined to provide a full list of the factors it does consider. He said it does not weigh race.

He declined to provide any of the department’s annual assessments of the tool.  The Inquirer and Daily News also asked how the tool scored two specific offenders.

A court spokesman declined to release the risk-assessment score for Maurice Roberts, the man accused of being involved in this month’s slaying in Spring Garden.

The reason given: Roberts has a pending open case.

So a reporter asked about a probationer who doesn’t have an open case, Nicholas Glenn, who went on a shooting rampage last year that left a woman dead. Glenn was killed in the ensuing scuffle with police.

The spokesman again declined to provide the man’s score.

This story was prepared as part of a John Jay/H.F. Guggenheim Center on Media, Crime and Justice reporting fellowship.