Variable selection performance of multivariate tree boosting compared to MANOVA, the Lasso, multivariate CART, and bagged multivariate CART. The performance (higher is better) is shown for a given nonlinear effect (ex, x2, x3, x is the linear model) for a range of effect sizes. Higher AUCs indicate better performance: an AUC of 1 is perfect and an AUC of .5 corresponds to variable selection no better than chance. For effects that are not linear (x2, x3), boosting dominates all other methods followed by bagged multivariate CART. For ex and x (linear effect), boosting does not perform significantly worse than MANOVA or the Lasso.