1: \begin{abstract}

2:   Evaluating treatment effect heterogeneity widely informs treatment

3:   decision making. At the moment, much emphasis is placed on the

4:   estimation of the conditional average treatment effect via flexible

5:   machine learning algorithms. While these methods enjoy some

6:   theoretical appeal in terms of consistency and convergence rates,

7:   they generally perform poorly in terms of uncertainty

8:   quantification. This is troubling since assessing risk is crucial

9:   for reliable decision-making in sensitive and uncertain

10:   environments. In this work, we propose a conformal inference-based

11:   approach that can produce reliable interval estimates for

12:   counterfactuals and individual treatment effects under the potential

13:   outcome framework. For completely randomized or stratified randomized experiments with perfect

14:   compliance, the intervals have guaranteed average coverage in finite

15:   samples regardless of the unknown data generating mechanism. For

16:   randomized experiments with ignorable compliance and general

17:   observational studies obeying the strong ignorability assumption,

18:   the intervals satisfy a doubly robust property which states the

19:   following: the average coverage is approximately controlled if

20:   either the propensity score or the conditional quantiles of

21:   potential outcomes can be estimated accurately.  Numerical studies

22:   on both synthetic and real datasets empirically demonstrate that

23:   existing methods suffer from a significant coverage deficit even in

24:   % deceptively 

25:   simple models. In contrast, our methods achieve the

26:   desired coverage with reasonably short intervals. %\allison{I think

27:    % the abstract stinks.}

28: \end{abstract}

29: