The CAUSALTRT Procedure

The CAUSALTRT procedure estimates the average causal effect of a binary treatment, T, on a continuous or discrete outcome, Y. Although the causal effect that is defined and estimated in PROC CAUSALTRT is called a treatment effect, it is not confined to effects that result from controllable treatments (such as effects in an experiment). Depending on the application, the binary treatment variable T can represent an intervention (such as smoking cessation versus control), an exposure to a condition (such as attending a private versus public school), or an existing characteristic of subjects (such as high versus low socioeconomic status). The CAUSALTRT procedure can estimate two types of causal effects: the average treatment effect (ATE) and the average treatment effect for the treated (ATT). For more information about the causal effects that the CAUSALTRT procedure can estimate, see the section Causal Effects: Definitions, Assumptions, and Identification.

The CAUSALTRT procedure implements causal inference methods that are designed primarily for use with data from nonrandomized trials or observational studies. In an observational study, you observe the treatment T and the outcome Y without assigning subjects randomly to the treatment conditions. Instead, subjects "select" themselves into the treatment conditions according to their pretreatment characteristics. If these pretreatment characteristics are also associated with the outcome Y, they induce a specious relationship between T and Y and hence cloud the causal interpretation of T on Y. Therefore, estimating the causal effect of T in observational studies usually requires adjustments that remove or counter the specious effects that are induced by the confounding variables.

To adjust for the effects of the confounding variables, you can model either the treatment assignment T or the outcome Y, or both. Modeling the treatment leads to inverse probability weighting methods, and modeling the outcome leads to regression adjustment methods. Combined modeling of the treatment and outcome leads to doubly robust methods that can provide unbiased estimates for the treatment effect even if one of the models is misspecified. For more information about which model specifications are required for different estimation methods and how default estimation methods are determined, see the section Outline of Estimation Method Requirements. For an introduction to causal inference from nonrandomized data, see Imbens and Rubin (2015); Morgan and Winship (2015); Berzuini, Dawid, and Bernardinelli (2012).