On Evidence-Based Sentencing and the Variables of Race, Age, and Social Achievement

I was reading a paper on evidence-based sentencing called “Risk in Sentencing: Constitutionally-Suspect Variables and Evidence-Based Sentencing“. And in it, the authors list (list is generated by another study), fifteen different variables with statistically significant relationships with recidivism. Here are some (on a 0.30 scale):

  • Criminal companions: z=0.21
  • Antisocial personality: z=0.18
  • Adult criminal history: z=0.17
  • Family rearing practices: z=0.14
  • Social achievement (education, marital status, employment): z=0.13
  • Race: z=0.17
  • Age: z=0.11
  • Gender: z=0.06
  • Socio-economic status of origin: z=0.05

Immediate things that pop out: race, criminal companions (who you hang out with), social achievements (education, marital status, employment), age, gender, and socio-economic status of origin. According to this study, these factors indicate some probability of recidivism. Luckily, several of these variables (such as race and age) are constitutionally barred from being taken into account in sentencing decisions. But the point I want to make is that I don’t think most of these should be a factor in determining a person’s sentence. And I think this study is a great example of why we should be careful when drawing conclusions from analyzing data. I sometimes tell this joke: “100% of divorces are caused by marriage”. It’s silly, but I think it’s relevant here. Yes, divorces begin with marriage, but if you blame marriage on divorce, you’re kind of missing some important underlying cause. Sure, young poor uneducated black people who hang out with other criminals might have an increased chance of recidivism, but is that really the underlying cause? Is it really their fault that they are young, poor, uneducated and black living in a neighborhood where everyone else is young, poor, uneducated, and black?

This is a great example of algorithms just pointing out the obvious and yet missing the larger picture. It’s like Google’s flu detector which actually might only be a winter detector. We need to think about how we construct these algorithms and how we are employing them to make decisions that might affect hundreds of thousands of people. We shouldn’t be asking “how does the race variable relate to recidivism?” There’s nothing “variable” about race. Or age. Or socio-economic status. These are the wrong questions. Instead, why don’t we ask ourselves “What can we do, to improve a person’s life, such that the color of their skin doesn’t correlate with a high recidivism rate?” I think that’s a more worthwhile pursuit.


Leave a Reply