Fig. 3.

Median values (short horizontal lines) and scatter plots showing variability in (A) tuning parameter estimate ($\text{λ}$) and (B) predictive performance of the developed model in the large validation data, for various methods across varying sample sizes for model development; for each sample size, 500 data sets were simulated as described in Section 2.3, and for each data set, a model was developed for each method with (A) tuning parameter estimated, and then (B) predictive performance tested. In (B), the long horizontal lines are the large sample performance values. CITL = calibration-in-the-large; Horizontal spread within each sample size grouping is just random jitter to aid display; c-index is not shown for heuristic or bootstrap shrinkage, as these methods do not change the ranking of predictions, and thus the c-index is the same as maximum likelihood estimation.