FIGURE 3.

Performance metrics and economic effects for the models. The AUROC, AUPR and F1 are standard metrics measuring the performance of a binary classifier. We also show the economic effect of assigning either a 6‐ or 12‐month donation interval to donors who are predicted to be deferred. A negative effect means savings in units of euro per donation. From the Hb prediction of the Bayesian linear mixed models, we calculate the probability of deferral based on Finnish deferral limits, while the random forest model outputs probabilities of deferrals directly. To calculate economic effects, these probabilities of deferrals need to be dichotomized into a deferral status by a cutoff value. The threshold panel shows, which cutoff for the probability of deferral, applied to a given model, gave the optimal savings (shown on the economic effect panels), where the candidates for cutoffs were 0.02, 0.04, …, 0.98. For all the panels except the threshold panel, 95% confidence intervals computed using bootstrapping are shown. Most of the savings come from avoiding female deferrals