Two central objectives of individualized treatment are precision and optimality. A third objective is robustness, and this talk aims to explore what could happen if we account for robustness in the decision process. The first case study is post-selection inference for effect modification, in which predictive precision is sacrificed. We will introduce a relatively straightforward method by combining Robinson’s transformation (an instance of Neyman orthogonalization/doubly robust estimation) with post-selection inference in linear models. The second case study is obtaining individualized treatment rules with unmeasured confounding. The key observation here is that the treatment rules only admit a partial order and optimality is not definite. We will introduce a method based on Rosenbaum’s sensitivity analysis and stepwise multiple testing to select and rank individualized treatment rules.