1: \begin{abstract}

2: Optimal treatment regimes are personalized policies 

3: for making a treatment decision 

4: based on subject characteristics, 

5: with the policy chosen to maximize some value. 

6: It is common to aim to 

7: maximize the mean outcome in the population, 

8: via a regime assigning treatment only to 

9: those whose mean outcome 

10: is higher under treatment versus control. 

11: However, the mean can be an unstable measure of centrality, 

12: resulting in imprecise statistical procedures, 

13: as well as unfair decisions that 

14: can be overly influenced by a small fraction of subjects. 

15: In this work, we propose a new median optimal treatment regime that instead treats individuals whose \emph{conditional median} is higher under treatment.

16: This ensures that optimal decisions 

17: for individuals from the same group are not overly influenced 

18: either by (i) a small fraction of the group (unlike the mean criterion), 

19: or (ii) unrelated subjects 

20: from different groups (unlike marginal median/quantile criteria). 

21: We introduce a new measure of value, the Average Conditional Median Effect (ACME), which summarizes across-group median treatment outcomes of a policy, and 

22: which the optimal median treatment regime maximizes. 

23: After developing key motivating examples that distinguish 

24: median optimal treatment regimes from mean and marginal median optimal treatment regimes, 

25: we give a nonparametric efficiency bound 

26: for estimating the ACME of a policy, 

27: and propose a new doubly robust-style estimator 

28: that achieves the efficiency bound under weak conditions.

29: Finite-sample properties of the estimator are explored via numerical simulations

30: and the proposed algorithm is illustrated 

31: using data from a randomized clinical trial in patients with HIV.

32: 

33: \end{abstract}

34: