Zhang Yao presents R2P, a method that discovers patient subgroups with different treatment responses while measuring uncertainty. It uses conformal prediction to build confidence intervals around treatment effects, then splits patients into groups only when those intervals don't overlap, avoiding false discoveries from random variation.
Chapters
Hi, my name is Yao. I will present our paper Robust Recursive Partitioning for Heterogeneous Treatment Effects with Uncertainty Quantification. This work is published in Europe this year. In this paper, we developed a heterogeneous treatment effect analysis method to find subgroups of individuals who have similar covariance and treatment response. As shown in the figures, the colors indicate which groups, which subgroups an individual belongs to. The subgroups on the left are interpretable in terms of individual covariates X0 and X1. The subgroups on the right are highly overlapped and non-interpretable. Our methods aim to find subgroups, interpretable subgroups, while allowing for any black box machine learning model to be used for treatment effect estimation. The identification of the subgroups improves the interpretation of the treatment effects. Our method also quantifies the uncertainties in the subgroup discovery process. All these advantages together make our methods possible to develop more effective treatments and to improve the design of further experiments based on the data collected from randomized trials and observational studies. In the previous HTE analysis methods, subgroups are identified by maximizing the heterogeneous heterogeneities across subgroups. They often rely on a specific CAT estimator, conditional average treatment effect estimator, decision tree. Suppose the true treatment effect is generated by a Gaussian distribution with zero mean with standard deviation 0.1. Clearly, there are no subgroups, as shown in the figure. The previous methods may find subgroups if there are groups that have different sample means due to sampling error. Our method will reject this subgroup discovery by constructing confidence intervals for the treatment effects. If the confidence intervals cross subgroups overlap, we will reject the partition. Our method starts by constructing a conditional average treatment effect estimator for the entire covariance space. Here we can use any kind of CATE estimator. For example, we can estimate treated response curve by regression model μ₁ and the control response curve by another regression model μ₀. Then we take the difference between μ₁ and μ₀ as the CAD estimator. Then we measure the effect homogeneity within the subgroups by the concentration of the CAD within the subgroup. Specifically, we develop the confidence interval for the CAD estimator by using a conformal prediction method. That the confidence interval constructed by conformal prediction satisfies the frequentist coverage guarantees in finite samples. We now develop the confidence criterion for partitioning covariance space into subgroups recursively. In the criterion, we minimize the absolute deviation of the CAT confidence interval S_X, we also minimize the expected confidence intervals with W_X because the narrower confidence intervals can also help to reduce the deviation. As shown in the figure, the criterion splits the covariance space X into two subgroups. The criterion improves both homogeneities within the subgroup and heterogeneities across subgroups. After splitting, we construct confidence intervals using the samples within each subgroup. Then the confidence intervals satisfy the coverage guarantees within each subgroup. The coverage guarantees of the confidence intervals in our methods can help to avoid false discoveries of subgroups. If the confidence intervals across subgroups have large overlaps, the constructed subgroups are not well identified. Conversely, if the confidence intervals have little and no overlap across subgroups, subgroups are well identified. Finally, we evaluate our method R2P by comparing it to other state-of-the-art heterogeneous treatment effect estimation methods. For this, we mainly use variance, cost, and within-subgroups to evaluate the heterogeneity and homogeneity, respectively. From the tables, we can see our method R2P achieves the best performances in all the performance metrics while discovering subgroups with minimum overlap to avoid false discoveries, as shown in the figure below. Overall, in this paper, we propose a robust recursive partitioning method for estimating heterogeneous treatment effects. Each subgroup discovered by our method is interpretable based on the tree structures in the recursive partitioning process. Our method can be used with any black-box machine learning models and enable them to produce trustworthy and interpretable treatment estimates in practice. Thank you for watching. If you are interested in the papers, take a look at the papers and software on our website.
