Using Regression Kernels to Forecast a Failure to Appear in Court

Forecasts of prospective criminal behavior have long been an important feature of many criminal justice decisions. In this paper, we apply a form of kernel logistic regression to forecast at an arraignment whether an individual charged with drug possession will return to court when ordered to do so. The practical goal is to help inform a magistrate’s release decision. We focus on individuals with drug possession charges because they have atypically high rates of failures to appear (FTAs). We apply a form of kernel logistic regression because recent work has shown that conventional logistic regression typically will not forecast as accurately as machine learning procedures. Our approach to kernel logistic regression, which can be seen as a hybrid of conventional logistic regression and machine learning, clearly dominates conventional logistic regression as a forecasting tool, and in some settings can be a legitimate competitor to machine learning procedures such as support vector machines, stochastic gradient boosting, and random forests. The methods applied are implemented in the R package kernReg currently available on CRAN.