heterogeneous treatment effect estimation

This study addresses the degradation of conditional average treatment effect (CATE) estimation generalizability when extrapolating randomized controlled trial (RCT) results to the source population, due to selection bias and high-dimensional covariates. It provides the first systematic evaluation of four strategies for mitigating selection bias within the causal forest framework. Results show that directly incorporating selection variables yields theoretically unbiased estimates but incurs substantial variance inflation; in contrast, inverse probability weighting (IPW)-based correction achieves lower bias and well-controlled variance across most simulation settings, demonstrating superior robustness. The study proposes a practical IPW–causal forest integration framework that delivers CATE estimates with both theoretical validity and empirical robustness for RCT external validity. This work fills a critical gap by offering the first comprehensive assessment of selection bias correction methods under high-dimensional settings.

Traditional cluster randomized trials typically estimate only the average treatment effect, overlooking heterogeneity at both individual and cluster levels. This work proposes a unified mixed-effects machine learning framework that integrates individual- and cluster-level covariates to estimate conditional average treatment effects while marginalizing over unobserved cluster heterogeneity. For the first time, it systematically combines methods such as Bayesian additive regression trees, multilevel Bayesian causal forests, mixed-effects random forests, mixed-effects gradient boosting, and generalized additive mixed models, incorporating cluster-specific random intercepts to account for within-cluster dependence. The approach is validated through diverse simulation studies and an application to a cluster randomized trial on hypertension management in Ghana, with accompanying open-source code and practical implementation guidelines provided.

Conventional conditional average treatment effect (CATE)-based methods struggle to identify treatment effect heterogeneity when effect modifiers are unobserved or subject to severe measurement error. Method: We propose a variance-comparison inference framework that does not require fully observed covariates. Leveraging the variance difference of potential outcomes as a novel causal identification anchor, we construct a doubly robust and asymptotically linear nonparametric estimator, integrating causal machine learning with a variance-sensitive testing paradigm. Contribution/Results: We establish theoretical consistency and asymptotic normality under weak regularity conditions. In a reanalysis of a randomized controlled trial, our method detects statistically significant heterogeneity in therapeutic hypothermia efficacy. The approach demonstrates robustness across diverse data-generating mechanisms and overcomes the strong reliance of CATE-based methods on high-fidelity covariate measurement.

Reliably evaluating the goodness-of-fit of conditional average treatment effect (CATE) estimates derived from observational data remains a critical challenge for applying causal inference in policy and personalized decision-making. This work proposes the CAFE framework, which introduces the first validation approach directly targeting CATE estimation—rather than the full outcome model—by leveraging auxiliary randomized controlled trial (RCT) data. CAFE stratifies the covariate space using propensity scores and conducts hypothesis tests based on group-level treatment effect comparisons. To enhance sensitivity to local model misspecification, it incorporates a maximal test statistic and employs a two-stage procedure to detect potential unmeasured confounding. The framework is compatible with both parametric models and flexible machine learning methods such as causal forests. Extensive experiments demonstrate that CAFE effectively identifies CATE model mismatches, offering a reliable assessment when both RCT and observational data are available.

This paper addresses the practical violation of conditional unconfoundedness—a core assumption in causal machine learning—by systematically analyzing the bias mechanisms of causal forests and X-learners under unmeasured confounding. It introduces negative control outcomes (NCOs) as a novel, pragmatic diagnostic tool for detecting unobserved confounding, marking the first application of NCOs for this purpose in causal ML. Through extensive simulation studies across diverse confounding structures, sample sizes, and degrees of treatment effect heterogeneity, the study demonstrates that unconfoundedness violations induce spurious heterogeneity detection. Although NCOs do not strictly satisfy theoretical identification conditions, they robustly flag subgroups severely affected by unmeasured confounding, thereby substantially improving the credibility of individualized treatment effect estimation. This work advances the integration of NCOs from sensitivity analysis into standard causal modeling pipelines, providing methodological foundations for robust causal machine learning.

This study addresses the challenge of flexibly modeling heterogeneous treatment effects while preserving the validity of randomization-based inference. The authors propose a model-assisted randomization test that avoids sample splitting by estimating the unsigned conditional average treatment effect (CATE) through the residual covariance structure, retaining the original treatment assignment for inference, and optimizing sign assignments to best fit the observed outcomes. This approach represents the first seamless integration of flexible CATE modeling with randomization testing, achieving strict Type I error control while substantially improving statistical power. Moreover, it enables the identification of heterogeneous subgroups exhibiting distinct treatment responses. Empirical evaluations demonstrate that the method outperforms conventional covariate adjustment and sample-splitting strategies in both power and precision.

This work proposes a unified framework to test the homogeneity of conditional average treatment effects (CATE) across multiple experimental and observational studies and to assess the sensitivity of effect estimates to unobserved confounding. Built upon double machine learning, the approach accommodates high-dimensional covariates and is applicable to locally identified settings such as randomized controlled trials, instrumental variables, and difference-in-differences designs. It is the first framework to jointly integrate CATE homogeneity testing with robustness evaluation against unmeasured confounding, thereby enabling data-driven judgments about the validity of extrapolation. Simulations demonstrate favorable finite-sample performance, and an application to the International Stroke Trial (IST) provides empirical evidence supporting both the generalizability of causal effects and the plausibility of identification assumptions.

Traditional average treatment effects fail to capture individual heterogeneity, and under high-dimensional settings, the sublevel set structure of the conditional average treatment effect (CATE) function is complex, lacking a concise global measure of heterogeneity. This work formalizes the probability curve of CATE sublevel sets as a target parameter for the first time, revealing its non-pathwise differentiability. By integrating Grenander-type monotone estimation with debiased machine learning techniques, the authors develop a nonparametric inference framework. The proposed estimator demonstrates strong finite-sample performance and is applied empirically to randomized trial data on diabetes medication, effectively uncovering heterogeneous treatment effects across subpopulations.

This study addresses the critical need for accurate estimation of the conditional average treatment effect (CATE) for specific events in survival analysis under competing risks and right censoring, which is essential for personalized medicine. The authors propose a meta-learner–based framework in a binary treatment setting, defining CATE as the absolute risk difference at a fixed time point. They systematically evaluate six meta-learner strategies that combine either Cox regression or random survival forests to model event-specific risks, paired with elastic net or random forest models to directly estimate CATE. Comprehensive simulations encompassing diverse risk structures, treatment effect heterogeneity, treatment assignment mechanisms, and censoring levels are conducted to assess performance. Based on empirical findings, the study offers practical modeling recommendations and releases the open-source R package crsurvlearners to facilitate broader application.

This study addresses a systematic bias—termed “group bias”—that arises in heterogeneous treatment effect (HTE) modeling when aggregating individual conditional average treatment effects (CATE) to estimate group-level average treatment effects (GATE). The authors formally define this bias for the first time and develop a unified statistical framework that enables asymptotically normal bias measurement and hypothesis testing. They further propose a shrinkage-based bias correction method, yielding a computationally tractable closed-form solution for optimal adjustment under minimal assumptions. The approach efficiently detects and corrects group bias with high accuracy. Empirical evaluation on large-scale digital platform experiments demonstrates substantial bias reduction and improved decision-making performance in personalized intervention strategies, particularly in profit-maximization settings where it enhances both policy accuracy and real-world returns.