Inference Of Subgroup-level Treatment Effects Via Tree-based Method In Observational Studies

Causal inference is a focused issue that spans many fields,such as clinical medicine and econometrics.It helps practitioners make policy choices or improvements by evaluating the change in outcome caused by treatment,which contains information of great value.As the cost of data acquisition decreases in the age of big data,scholars tend to make causal inferences on observational data,where one of the greatest challenges is to deal with confounding between treatment and outcome variables.Meanwhile,intricate relationships between variables and increasing diversity in population make heterogeneity in causal effects could not be ignored.Scholars have begun applying machine learning methods to identify heterogeneous causal effects in the past decade.The most popular methods are devoted to estimating causal effects at the individual level.However,we argue that individual treatment effects(ITE)are too trivial to utilize in large-scale data analysis scenarios directly.Moreover,ITE can not provide an intuitive explanation of the heterogeneity mechanism for decision-makers.Although there are kinds of literature in recent years that propose to estimate subgroup treatment effects(STE)by building trees or post-processing ITE,problems still exist,like the large bias of estimation or unreasonable subgroup identification strategies.A tree-based algorithm is employed to estimate subgroup-level treatment effects in observational data,called generic causal tree(GCT).GCT adopts the idea of Robinson transformation to construct causal effect estimators based on split and designs a splitting criterion that guides the generation of trees by maximizing the difference of causal effects on the split while controlling their fluctuations.At the same time,we use honest estimation to conduct conditional asymptotic distribution of STEs on a given tree model,ensuring the validity of inference.Through a series of simulation experiments,we demonstrate the advantages of GCT in subgroup identification and effect estimation among existing tree-based methods.Moreover,we verify its effectiveness of inference.