Predictions for Nonrandom Samples and for Individuals
In previous chapters, we have considered predictions of recidivism for random samples of our release cohorts. Such random samples will differ significantly from the random samples used to estimate our models only by chance. We now turn to the much more difficult task of making predictions for groups that differ systematically from the samples used in estimation. This type of prediction has often been strongly recommended but rarely pursued in practice.1 We also consider predictions of recidivism at the individual level. Making either of these types of predictions requires a model that includes individual characteristics as explanatory variables. This is obviously the case with individual predictions, as a model that did not include individual characteristics would only predict the same distribution of time until recidivism for every individual. On the other hand, in the case of predictions for groups that are nonrandom samples of the release cohort, the purpose of having individual characteristics in the model is to correct for differences between the group in question and the rest of the cohort. For example, in making predictions of the rate of recidivism of a group of youthful offenders, the model can account for the age of the individuals (assuming that age was a variable used in the model), as well as for differences between the group and the estimation sample in the values of the other relevant explanatory variables.
KeywordsFalse Positive Rate False Negative Rate Validation Sample Recidivism Rate Nonrandom Sample
Unable to display preview. Download preview PDF.