Third simulation on linear regression.

Comparison of AR with Stepwise (SW), SCAD and multi-step adaptive LASSO (ALASSO) in case of linear regression for two scenarios (p = 50 and p = 500). For the first scenario we consider only BIC, for the second high-dimensional scenario both BIC and mBIC. To assess the quality of classification we report the average over 200 simulation runs of the following quantities: Power, number of false positives (FP), false discovery rate (FDR) and number of misclassifications (Mis). The next two statistics quantify for each procedure the performance in terms of minimizing the criterion: the mean over the squared differences with the criterion of the best method (mse) and the percentage of simulation runs in which a procedure gave the model with the smallest criterion (best). Finally we report the average computational time for one replication (time). For each quantity the best method is printed in bold.