Simulation Study

Standard BART achieved perfect sensitivity across all 12 simulation scenarios. The full Level 3 model fit in under two minutes for datasets with 2000 individuals. In the linear scenario, more flexible levels recovered the simpler truth without inventing large non-linear effects. The simulation compared spBART against standard BART under two data-generating mechanisms. In the quadratic U-shaped scenario, Level 1 failed while Levels 2 and 3 recovered the U-shaped curve. In the S-shaped scenario, Level 3 captured features that Level 2 only approximated globally. Standard BART had consistently higher predictive AUC than spBART in the reported simulations. The main simulation study used 300 datasets for each of three scenarios with 2000 individuals per dataset. spBART showed lower sensitivity at n=1,500 as dimensionality increased.