Please read Scott Greenfield’s post “Numbers Don’t Lie (But People Do)”

Yesterday, I posted “Like the ostrich that buries its head in the sand, Mr. Holder is wrong about data-driven sentencing.” Today, Scott Greenfield at Simple Justice has posted a rejoinder entitled “Numbers don’t lie (But People Do). Scott makes excellent points about reliance at sentencing on social science data if used to predict future violence. Read Scott’s post.

I add the following more as a clarification than anything else:

I am not terribly wound up by the idea that I am under pending reforms supposed to reduce the number of offenders in federal prison by trying to determine at sentencing which are likely to be violent when released and which are not likely to be violent upon release. Indeed, if you read the Rorschach test of sentencing, that is 18 U.S.C. § 3553(a), you will struggle to find a directive that I ought to look at prison overcrowding and the cost to the federal fisc. That “little” problem aside, I am not at all sure that I am competent to read, understand and apply the relevant social science data that would allow me to rationally determine the risk of future violence. But if Attorney General Holder wants federal judges to be seriously sensitive to the issue of future violence and prison overcrowding he is, in my opinion, being obtuse or disingenuous when he suggests that we ignore mounting social science data that rely upon “immutable characteristics” and other factors (like socio-economic information*) that make the rest of us twitchy about issues of race, gender, age and poverty. The Attorney General should not be allowed to have his cake and eat it too by suggesting we ignore the uncomfortable.


*Philadelphia is highly segregated. Yet the successful prediction instrument, with a 66% accuracy rate, developed in the City of Brotherly Love uses zip codes as a scoring factor. See Nancy Ritter, Office of Justice Programs, National Institute of Justice, Predicting Recidivism Risk: New Tool in Philadelphia Shows Great Promise (February, 2013).

3 responses

  1. I am not at all sure that I am competent to read, understand and apply the relevant social science data that would allow me to rationally determine the risk of future violence.

    This also assumes that the relevant social science data, properly understood, can rationally determine the risk of future violence. Even if you are very good at this sort of thing, our level of accurate statistical inference on this is woeful.

    For example, that study you link, while interesting, is also tainted by some poor scientific practices common to criminal justice related studies. The 66% accuracy cited was based on an arbitrary constraint: the number of probationeers who could be fit into a program for high-risk offenders in the Philadelphia probation system. (see pg. 20). The size of the “high risk” class (and indeed all the classes) was then not set by empirics, but by the Philadelphia parole department budget.

    Further, there is a huge confounding variable problem: the high risk probationeers were given intensive intervention as part of the program. That means we don’t know if the model accurately predicted their recidivism, because we did something special to them based on the results of the model.

    These are common issues in criminal justice studies. The Philadelphia parole department aren’t going to spend a ton on running two parallel programs to do a randomized controlled trial. Nor are they going to let a researcher come in and say “you need to do an intensive program that will double your costs because we predict 40% of your offenders are high risk.”

    But the fact that these issues are common also tells you that most of the studies you see are strongly influenced by biases resultant from the need for support (logistical or otherwise) from the institutions which run the criminal justice system.

  2. One thing no one seems to be talking about is the utility of many of the evidence based practices on supervised release issues. For example, if companionship is really one of the strong predictors of recidivism,it is something that the sentencing court can directly attack by imposing Associational and Geographic Restrictions as special conditions of supervised release.

    If you think about Olsen’s factors from the supervised release perspective, they can lead you to conclude that the special conditions that are imposed at sentencing may be just as important as the sentence itself.

  3. John,

    I agree. Empirical information can do much to inform the selection of supervised release conditions. They can also be used during supervised release. For example, they can be used to determine who will require high intensity supervision and that helps allocate resources more effectively.

    All the best.


%d bloggers like this: