|
Abstract:
|
Variable importance measures for random forests are receiving increasing attention as a screening tool in high dimensional classification tasks, e.g. in statistical genomics. However, we show and illustrate in simulation studies that suboptimal predictor variables may be artificially preferred by these measures when predictor variables vary in their scale level or their number of categories. The two statistical mechanisms underlying this deficiency are biased variable selection in the individual classification trees on one hand, and effects induced by bootstrap sampling with replacement on the other hand. An alternative implementation of random forests and bagging, providing unbiased variable selection, is presented. When this method is applied with subsampling without replacement, the resulting variable importance measure can be used reliably as a screening tool in any data situation.
|
