Thread by @MeganTStevenson

When you give an algorithmic prediction of future offending to a judge to use in felony sentencing, she may use it in unexpected ways...

Tweet summary of my new paper with @jenniferdoleac

"Algorithmic Risk Assessment in the Hands of Humans"

papers.ssrn.com/sol3/papers.cf…

1/

First we show that judges use risk assessment. Defendants who score right above the risk-score cutoff that separates "high risk" from "low risk" defendants have a substantially higher probability of incarceration and sentence length than those who score right below. 2/

But the net impacts are weird. The nonviolent risk assessment was designed to keep people out of prison: to divert the lowest risk defendants towards probation or short jail stays. But there was no net decline in incarceration or sentences for nonviolent offenders. 3/

(We use a difference-in-differences specification here, comparing outcomes before/after RA adoption for defendants who were/weren't eligible for RA.) 4/

Instead, prison beds were reshuffled so that more high-risk-score people were incarcerated and fewer low-risk-score people were incarcerated. Theoretically, incapacitating those at highest risk of reoffending should have led to a decline in recidivism. This didn't happen. 5/

Why no efficiency gains? I.e., why were we not able to reduce incarceration and/or recidivism by making smarter decisions with the use of algorithms, as was predicted in a recent QJE article? 6/

academic.oup.com/qje/article-ab…

The answer, at least in part, has to do with judicial discretion. While judges used the risk tool, they also deviated from it's sentencing recommendations quite often. In particular, they were lenient with young offenders despite their high risk scores. 7/

Age is one of the most important predictors of criminal activity. In Virginia's risk assessment, you get more points for being 23 than for having 5 or more prior incarcerations. Most other risk tools are similar. 8/

papers.ssrn.com/sol3/papers.cf…

If judges had fully followed the sentencing recommendations associated with the algorithm, there would have been a HUGE increase in sentences for young defendants, relative to older ones. But they didn't. (Trip-diffs with young/old, before/after, eligible/ineligible.) 9/

Judges used their discretion to grant leniency to young offenders, despite their high risk scores. There was still a relative increase in sentences for the young, but this was nowhere near as it would have been if there had been no discretion. 10/

Is this a mistake? Probably not. Judges are probably being lenient with young people because of lower perceived culpability, consistent with longstanding patterns in criminal justice. 11/

If a reluctance to incarcerate young people generalizes to other settings, then prior research may have overestimated the extent to which judges make prediction errors. And, if one of RA's most important inputs is effectively off-limits, then its expected benefits are lower. 12/

How about risk assessment's impact on racial disparities? This is a little more complicated. Even though black defendants have substantially higher risk scores than white defendants, judges sentence in a racially disparate manner even without risk assessment. 13/

The impact on racial disparities depends both on what judges do in the absence of risk assessment, and how they they interpret the risk score for black people. Theory makes no clear predictions. 14/

Statewide we find that risk assessment had no impact on racial disparities. However we find that racial disparities increased in the subset of courts that appeared to use RA most. And simulations show that full compliance with the algorithm would disadvantaged blacks slightly.15/

Everything I've been talking about so far has to do with a risk assessment that is only used on nonviolent offenders. But Virginia also has a sex offender risk assessment, designed to only increase sentences for those at the highest risk of reoffending. 16/

Even though the sex offender RA assessment is only designed to lengthen sentences, its use actually led to DECLINE in both the probability of incarceration and the length of the sentence. (Again, DD with ineligible defs as control.) 17/

Even though judges increased sentences for those with high risk scores, they lowered sentences even more for those with low risk scores, resulting in a net decline. We think this is due to a `Willie Horton' effect. 18/

Judges are averse to releasing people because they are afraid that they will be blamed if that person goes on to reoffend. But the risk score provides a `second opinion', a political shield that facilitates more lenient sentencing for those with low risk scores.19/

Last couple of things: the failure to find efficiency gains isn't because Virginia's risk assessment just sucks. We built an alternative risk score using random forest and the two perform comparatively. 20/

And finally, use of risk assessment fades over time, suggesting that judges did not find it useful. 21/

In sum: risk assessment can bring unexpected results when man and machine interact. Predictions about the impacts of risk assessment -- both positive and negative -- were wrong because they didn't take human incentives into account. 22/22