This table reports differences in the default prediction accuracy, measured by the
accuracy ratio (in percentage points, DAR), between the full model and the baseline
model. All model comparisons are performed on subsamples derived from tercile
splits of the reported variables to allow for non-monotonic effects (except LIMITED_
LIABILTY which is a dummy variable). The analysis is based on 25,344 private
German firms spanning the period from 2002 to 2005 (97,174 firm-year observations).
Reported are the results of clustered bootstrapping with 1000 replications
comparing the mid and upper terciles with the lower tercile.