Figure 2:

Average difference in treatment effect using the ITE recommendation in place of the ATE recommendation as a function of training set size. The estimated difference in the population is shown in (a); the estimated difference in the subpopulation where treatment recommendations differ is shown in (b). The difference in treatment effect is estimated by the Vickers et al.¹³ method with a test set of 100,000 examples. 95% confidence intervals are shown calculated over 100 replications with different training sets.