Professional judges in traffic courts sentence many hundreds of offenders per year. Using 639 case files from archives, we compared the Matching Heuristic (MH) to compensatory, weighing algorithms (WM). We modeled and cross validated the models on different subsets of the data, and took several other methodological precautions such as allowing each model to select the optimal number of variables and ordering and weighing the variables in accordance to different logics. We did not reproduce the finding by Dhami (2003), who found the MH to be superior to a compensatory algorithm in modeling bail-granting decisions. These simulations brought out the inner logic of the two family of models, showing what combination of parameters works best. It remains remarkable that using only a fraction of the variables and combining them non-compensatorily, MH obtained nearly as good a fit as the weighing method.
