Summary.  Using criminal population conviction histories of recent offenders, prediction mod els are developed that predict three types of criminal recidivism: general recidivism, violent recidivism and sexual recidivism. The research question is whether prediction techniques from modern statistics, data mining and machine learning provide an improvement in predictive performance over classical statistical methods, namely logistic regression and linear discrim inant analysis. These models are compared on a large selection of performance measures. Results indicate that classical methods do equally well as or better than their modern counterparts. The predictive performance of the various techniques differs only slightly for general and violent recidivism, whereas differences are larger for sexual recidivism. For the general and violent recidivism data we present the results of logistic regression and for sexual recidivism of linear discriminant analysis.

This study uses longitudinal official record data on adult offenders in The Netherlands (n=4,246) to compare recidivism after community service to that after short-term imprisonment. To account for possible bias due to selection of offenders into these types of sanctions, we control for a large set of confounding variables using a combined method of 'matching by variable' and 'propensity score matching'. Our findings demonstrate that offenders recidivate significantly less after having performed community service compared to after having been imprisoned. This finding holds for both the short-and long-term. Furthermore, using the Rosenbaum bounds method, we show that the results are robust for hidden bias.

In sparse tables for categorical data well-known goodness-of-t statistics are not chi-square distributed. A consequence is that model selection becomes a problem. It has been suggested that a way out of this problem is the use of the parametric bootstrap. In this paper, the parametric bootstrap goodness-of-t test is studied by means of an extensive simulation study; the Type I error rates and power of this test are studied under several conditions of sparseness. In the presence of sparseness, models were used that were likely to violate the regularity conditions. Besides bootstrapping the goodness-of-t usually used (full information statistics), corrected versions of these statistics and a limited information statistic are bootstrapped. These bootstrap tests were also compared to an asymptotic test using limited information. Results indicate that bootstrapping the usual statistics fails because these tests are too liberal, and that bootstrapping or asymptotically testing the limited information statistic works better with respect to Type I error and outperforms the other statistics by far in terms of statistical power. The properties of all tests are illustrated using categorical Markov models.

No abstract