Minority Report Is Coming

Crime prediction tools are here, they're good, and they're going to get much better

Punishment was never much of a deterrent and could scarcely have afforded comfort to a victim already dead. — Ed Witwer, Minority Report

Murder can be predicted and we can prevent it from happening without the need for a totalitarian state.

Arrest and victimization records for nearly 644,000 people profiled by the Chicago Police Department were recently used to predict which Chicagoans were the most at risk of getting shot. Of the 500 people with the highest predicted risk, 13% were shot within 18 months—a victimization rate 128 times higher than the Chicago average.

Predictions can get much better than this.

How is that possible?

---

Victims and Offenders

The victims of violent crime are usually people who were once criminal offenders. These two categories—criminal and victim—greatly overlap. Because bodies usually can’t be hidden, homicide provides us our best indication of the extent of this overlap. Consider this data from the following six American cities:¹

To put these data differently, assume that victim-offender overlap is 80% for homicides, that 77 million Americans have some form of criminal record out of a population of 335 million, and that the one-year implied ratio of people with criminal records to people without has held up since 1985.

If a given year’s homicide rate is 8 per 100,000, then there were 8 * 335,000,000/100,000 = 26,800 homicides, of which 0.8 * 26,800 = 21,440 were homicides of people with criminal records and 26,800 - 21,440 = 5,360 were homicides of people without criminal records. The homicide rate for people with criminal records would thus be 21,440/77,000,000 * 100,000 ≈ 27.84 per 100,000 versus a non-criminal homicide rate of 5,360/(335,000,000 - 77,000,000) * 100,000 ≈ 2.08 per 100,000. In other words, the homicide rate for criminals would be 27.84/2.08 ≈ 13.38 times greater than the homicide rate for non-criminals, and thus the homicide risk for non-criminals is right about 7.5% of the risk faced by criminals.

If we perform these calculations for each year of NIBRS homicide data (1985-2022), here’s how things look:

The difference in homicide risk for criminals and noncriminals is visually striking. But one question you might be asking is: How similar are victims and suspects in terms of the number of priors rather than simply having any priors?² The answer is “highly”. Among those with criminal records in the Portland and Washington D.C. data I cited above, the similarity in priors is striking:

These findings are not the only ones like them. In this report by the Illinois Criminal Justice Information Unit, 79.64% of the victims of firearm homicides had priors, and those with priors averaged 13.8 arrests. The ones with convictions (55.60%) averaged 3.1, and the ones who had been incarcerated (31.78%) had been incarcerated twice on average! Per the Bureau of Justice Statistics’ reporting on crime across large urban counties, “More than half of murderers (53%) had a conviction record” and “An estimated 39% of murderers and 29% of rapists had more than one prior conviction.”

The overlap between victims and offenders is substantial, replicable, and clearly matters. It matters enough that one group recently argued that it’s “among the most important basic facts of crime.”³

Criminal-offender overlap is a key component in the predictability of shootings, homicides, and other violent crime. It fits in alongside sex—most offenders and victims are male, including 77.26% of the homicide victims with a known sex in 2022—age—see my earlier article on the age-crime link—and race—most murderers and homicide victims are Black, including 55.86% of those with a known race in 2022. If we follow the inclusion-exclusion principle with just these four variables, we can discover some interesting facts about the predictability of homicides.

P(M) is the probability of a homicide victim being male, P(A) is the probability of a victim being ages 16-30, P(C) is the probability of a victim being a criminal, and P(B) is the probability of a victim being Black. P(M∩A∩C∩B) is P(M) + P(A) + P(C) + P(B) − P(M∪A∪C∪B), which is to say, P(M∩A∩C∩B) = P(M)*P(A)*P(C)*P(B), or that the probability a victim is a young Black male with a criminal record is at least the product of the probabilities that a victim is any of those things. If 80% of victims are male, 80% are 16-30, 80% are criminals, and 50% of Blacks, then at least 0.8 * 0.8 * 0.8 * 0.5 = 25.6% of homicide victims are all of those things. Given P(M∩A∩C∩B) ≤ min(P(M), P(A), P(C), P(B)), at maximum, min(0.8, 0.8, 0.8, 0.5) = 50% of homicide victims can be all of those things.

If half of the population is male, 20% of the population is ages 16-30, 23% of the population has any criminal record, and 13% of the U.S. population is Black, then at most 13% of the population could be all those things, and at minimum, 0.5 * 0.2 * 0.23 * 0.13 = 0.299% of the population could be. Most Blacks are not ages 16-30, male, and criminals; the lower-bound is closer to the truth. To calibrate better, about half of Black males have arrest by age 23. Since that’s right down the middle of the 16-30 age range, we can probably assume—safely enough—that it’s a decent characterization of the criminal record proportion among Blacks for the purposes of our calculation, and thus, to contain at least a quarter of the homicides in the U.S., we would only need to effectively surveil at minimum 0.65% of the total population.⁴

By now, I’ve demonstrated homicide is highly concentrated by noting four simple facts: homicide victims are usually criminals, usually Black, usually male, and usually young adults. With just four variables, we’re able to see that at least a quarter of the murders are done by an easily identified less than 1% subset of the population.⁵

Imagine how predictable homicide might be with 10, 20, 50, 100, or a thousand different variables. The answer is very, but only if the variables exist and can be gathered to build the model. Thankfully, more than enough can.

---

Killing Fields

In the Minority Report movie, the Washington D.C Precrime Bureau has effectively solved murder. Thanks to the precrime initiative—where killers are caught before they kill—, one of America’s stereotypically bloodiest cities hasn’t had a homicide in almost six years.

In the course of eliminating homicide from the District, the precrime initiative first knocked out premeditated murder. The reason was simple: the precog mutants used to predict who would kill whom were better able to see and help officers anticipate premeditated murders because they involved more thought, further out from the time of the crime.

In the real world, the opposite will happen: the first murders we’ll quash will be crimes of passion.