Fig. 1

An illustration of risk and benefit calibration figures with performance metrics of simulated data. We sampled (n = 3,600) from a simulated trial super population (100,000) with 12 binary risk predictors with 6 true treatment interactions [13]. Panel A depicts observed outcome versus predicted outcome by local regression (blue line, displayed between 0 and 0.5) and quantiles of predicted outcome (black dots), with the E-statistics, cross-entropy, Brier score, and C-index. Panel B depicts the calibration for benefit in groups with 95% confidence intervals, with the C-for-benefit. Panel C depicts observed versus predicted pairwise treatment effect by local regression (blue line, displayed between -0.2 and 0.3) and quantiles of predicted pairwise treatment effect (black dots), with the newly proposed metrics