Learning From Multiple Analogies: An Information Theoretic Approach to Predicting Criminal Recidivism.

If recidivism is defined as re-arrest within a finite period following release from prison, then the three kinds of outcomes typically available to researchers include: (i) whether or not the individual was rearrested within the follow-up period; (ii) how many times the individual was rearrested; and (iii) what was the duration from release to first (or subsequent) re-arrest. Since these outcomes are all different manifestations of the same underlying stochastic process, they provide us multiple analogies from which to recover information about it. This paper develops a semi-parametric approach for utilizing information in these, and several other related outcomes, to predict criminal recidivism. Accuracy is assessed in both out-of-sample and off-the-support prediction problems. Implications of the modeling strategy for actuarial risk assessment work are discussed. ..PAT.-Unpublished Manuscript [ABSTRACT FROM AUTHOR]