A subsequent discussion (November, 1999) with Rob Schapire revealed that the boosting results reported in these papers were not optimal due to an implementational deviation from the commonly used resampling procedure. Rob reported that, using his implementation, boosting results on the Statlog datasets closely match those of the subspace forests. For another comparison between bagging and boosting on the Irvine datasets, please see [J.R. Quinlan, Bagging, Boosting, and C4.5, Proc. 13th National Conf. on AI, Portland, Oregon, Aug 4-8, 1996, 725-730].