Prediction modelling for population conviction data

Abstract

In this thesis, the possibilities of using prediction models for judicial penal case data are investigated. The development and refinement of a risk taxation scale based on these data is discussed. When false positives are weighted equally severe as false negatives, 70% can be classified correctly. Moreover, the improvement in prediction is investigated by adding criminal case file information to the scale. The performance over time and over place is also assessed. Furthermore, an attempt to improve the predictive performance of this and other scales is done using machine learning and modern statistical techniques. This is done both for data having a binary outcome variable (i.e. recidivism yes/no after four years) as for models having a survival outcome (i.e. censored recidivism data). Finally, the effect of a penal measure for very frequent adult offenders is estimated using a combination of propensity score matching and multiple imputation. A combination of propensity score matching, difference-in-difference and multiple imputation is also considered.