Figure 3.

Coverage of 95% two-sided confidence intervals for the impact of treating the optimal subgroup versus the average confidence interval width (standardized so that the maximum width in a given sample size-data generating distribution pair is 1). The stabilized one-step confidence intervals have nominal coverage for all null distributions, and have coverage approaching nominal for all alternative distributions. The non-oracle CV-TMLE achieves near nominal coverage for all alternative distributions, though it has below nominal coverage of approximately 90% for all of the null distributions.