Practical Guide to Honest Causal Forests for Identifying Heterogeneous Treatment Effects

Abstract

“Heterogeneous treatment effects” is a term which refers to conditional average treatment effects (i.e., CATEs) that vary across population subgroups. Epidemiologists are often interested in estimating such effects because they can help detect populations who may particularly benefit from or be harmed by a treatment. However, standard regression approaches for estimating heterogeneous effects are limited by pre-existing hypotheses, test a single effect modifier at a time, and are subject to the multiple comparisons problem. The objective of this text is to offer a practical guide to honest causal forests, an ensemble tree-based learning method which can discover as well as estimate heterogeneous treatment effects using a data-driven approach. We discuss the fundamentals of tree-based methods, describe how honest causal forests can identify and estimate heterogeneous effects, and demonstrate an implementation of this method using simulated data. Our implementation highlights the steps required to simulate datasets, build honest causal forests, and assess model performance across a variety of simulation scenarios. Overall, this paper is intended for epidemiologists and other population health researchers who lack an extensive background in machine learning yet are interested in utilizing an emerging method for identifying and estimating heterogeneous treatment effects.

A data-adaptive method for investigating effect heterogeneity with high-dimensional covariates in Mendelian randomization

“…Recently, a forest for IV regression has also been discussed [4]. However, when dealing with complex scenarios where some effect modifiers may be downstream effects of the exposure, the current causal (or IV regression) tree and forest methods may not adequately address collider bias and could produce severe estimation bias, particularly if the variables used for splitting in the tree are either colliders or mediators between the exposure and the outcome [26]. The Q tree that we present is conceptually identical with other tree-based methods, but differs in its implementation as it is based on a Q statistic.…”

Estimating high-dimensional heterogeneous treatment effects in the context of climate and health: a case study exploring the role of effect modification in the relationship between precipitation shocks and diarrheal illness

“…However, it is also important to note that these techniques do not guarantee that identified covariates are true effect modifiers. Covariates may have large influence in fit-the-fit CART models or large variable importance due to high correlation with true effect modifiers (Jawadekar et al 2023). Variable importance should not be interpreted as the proportional influence of heterogeneity or the likelihood of a covariate being a true effect modifier.…”

Comparison of methods that combine multiple randomized trials to estimate heterogeneous treatment effects

“…5 We ultimately selected these three approaches because they represent two distinct classes of methods for estimating heterogeneous treatment effects (see Brantner et al 14 ) and seem to be used in practice, especially the causal forest. 17,18 Specifically, the first two approaches are multi-step procedures that involve first estimating the conditional outcome mean under treatment or control and then combining the two into one CATE function, while the causal forest involves tree-based partitioning of the covariate space by treatment effect.…”

“…The single‐study methods that exist can be grouped into multiple categories, as delineated by Brantner et al 14 For ease of comparison, three approaches are included that are user‐friendly and have been shown to be effective in previous literature: the S‐learner, X‐learner, 4 and causal forest 5 . We ultimately selected these three approaches because they represent two distinct classes of methods for estimating heterogeneous treatment effects (see Brantner et al 14 ) and seem to be used in practice, especially the causal forest 17,18 . Specifically, the first two approaches are multi‐step procedures that involve first estimating the conditional outcome mean under treatment or control and then combining the two into one CATE function, while the causal forest involves tree‐based partitioning of the covariate space by treatment effect.…”