The CAUSALTRT procedure estimates the average causal effect of a binary treatment T on a continuous or discrete outcome Y, where the observed effect might have been confounded, such as in observational studies.
Two types of modeling are used in PROC CAUSALTRT. One is outcome modeling by a generalized linear model. The other is propensity score modeling of the treatment variable by a logistic regression model (Guo and Fraser 2015; Imbens and Rubin 2015). For example, the following statements specify both outcome and treatment modeling for estimating the causal treatment effect of Drug on Diabetes (both are assumed to be binary here):
proc causaltrt data=drugs;
class Gender;
psmodel Drug = Age Gender BMI;
model Diabetes = Age Gender BMI;
run;
In the PSMODEL statement, you specify that the treatment variable is Drug and the corresponding propensity score model contains Age, Gender, and BMI as predictors of the treatment variable. The propensity scores in this model refer to the probabilities of drug use in individuals. In the MODEL statement, you specify that the outcome is Diabetes, which is modeled by Age, Gender, BMI, and Drug. Notice that the treatment effect of Drug on Diabetes is made implicit here. With the modeling of both the outcome and propensity scores, PROC CAUSALTRT uses a doubly robust estimation method (Lunceford and Davidian 2004) to estimate the potential outcome means, the average causal treatment effect, and their standard errors.
In fact, depending on whether or not you specify the effects on the right-hand side of the equal sign in the PSMODEL and MODEL statements, PROC CAUSALTRT estimates causal treatment effects by four different types of methods:
specification of effects in both the PSMODEL and MODEL statements: doubly robust methods such as the augmented inverse probability weighting method and the inverse probability weighted regression adjustment method
specification of effects only in the PSMODEL statement: three types of inverse probability weighting methods—basic, with ratio adjustment, and with ratio and scale adjustments
specification of effects only in the MODEL statement: regression adjustment method
specification of effects in neither the PSMODEL nor MODEL statement: unadjusted effect by computing the observed outcome mean difference between the treatment and control groups
The CAUSALTRT procedure can estimate two types of causal effects: the average treatment effect (ATE) and the average treatment effect for the treated (ATT). However, not all estimation methods can estimate both types of causal effects.
Other main features of PROC CAUSALTRT include the following:
a class of generalized linear models for modeling continuous and binary outcomes
asymptotic and bootstrap methods for computing standard errors and confidence intervals
diagnostic plots for the propensity score model, including various plots of propensity scores or weights
output data sets for propensity scores, inverse probability weights, and the predicted potential outcomes
For more information about PROC CAUSALTRT, see Chapter 39, The CAUSALTRT Procedure, and Lamm and Yung (2017).