Thirdly, other aspects of classification methods are important but have not been considered in our study, for example issues related to the transportability of the constructed prediction rules. By transportability, we mean the possibility for interested researchers to apply a prediction rule presented in the literature to their own data [9, 10]. With respect to transportability, LR is clearly superior to RF, since it is sufficient to know the fitted values of the regression coefficient to apply a LR-based prediction rule. LR also has the major advantage that it yields interpretable prediction rules: it does not only aim at predicting but also at explaining, an important distinction that is extensively discussed elsewhere [1] and related to the “two cultures” of statistical modelling described by Leo Breiman [41]. These important aspects are not taken into account in our study, which deliberately focuses on prediction accuracy.
