Fig. 4

Bias of AUC estimates obtained by different strategies based on bootstrapping. Bias is shown for one varying data set characteristic in each panel (a, b number of covariates p; c, d sample size n; e, f degree of missingness miss; g true effect auc), while keeping all remaining characteristics constant: sample size (n=200), p=10 covariates, 25 % missing values, missing at random (MAR), balanced outcome class distribution (frac=0.5), uncorrelated covariates (ρ=0). Results are shown for absence (theoretical auc=0.5; a, c, e, g) and presence (theoretical auc=0.66; b, d, f, g) of a moderate true effect of the covariates on the outcome