The CAUSALTRT procedure uses various estimation methods to adjust for the effects of confounding variables by fitting models for the treatment assignment T or the outcome Y, or both. For the estimation methods, you specify the outcome variable Y in the MODEL statement and the treatment assignment T in the PSMODEL statement.
Depending on the estimation method that you specify in the METHOD= option in the PROC CAUSALTRT statement, you must also provide specific additional modeling information in the MODEL and PSMODEL statements, as shown in Table 1.
Table 1: Estimation Method Requirements
| METHOD= | Additional Specification |
|---|---|
| IPW, IPWR, or IPWS | A model for the treatment assignment in the PSMODEL statement. |
| REGADJ | A model for the outcome in the MODEL statement. If you do not specify an outcome model, PROC CAUSALTRT fits an intercept-only model separately for each treatment condition. |
| AIPW or IPWREG | Both an outcome model in the MODEL statement and a model for the treatment assignment in the PSMODEL statement. |
The model for the treatment assignment is also called the propensity score model.
If you do not specify the METHOD= option in the PROC CAUSALTRT statement, the CAUSALTRT procedure determines a default estimation method based on which models are specified in the MODEL and PSMODEL statements, according to following criteria:
When only a treatment model is specified (in the PSMODEL statement), the default estimation method is IPWR. For example:
proc causaltrt;
model y;
psmodel trt = x1 x2;
run;
When only an outcome model is specified (in the MODEL statement), the default estimation method is REGADJ. For example:
proc causaltrt;
model y = x1 x2;
psmodel trt;
run;
When both the treatment and outcome models are specified, the default estimation method is AIPW. For example:
proc causaltrt;
model y = x1 x2;
psmodel trt = x1 x2;
run;
When neither an outcome nor a treatment model is specified, the default estimation method is REGADJ with an intercept-only outcome model. For example:
proc causaltrt;
model y;
psmodel trt;
run;
If you specify the ATT option in the PROC CAUSALTRT statement, then you are estimating the average treatment effect for the treated (ATT) instead of the default average treatment effect (ATE). In this case, estimation methods are limited to either IPWR or regression adjustment. Therefore, if models for both the treatment assignment and outcome variable are specified without the METHOD=IPWR or METHOD=REGADJ option, PROC CAUSALTRT issues an error.
The CAUSALTRT procedure fits only the models necessary for the estimation method used. For example, if a model for the treatment assignment is specified but the REGADJ estimation method is requested, then PROC CAUSALTRT does not fit the treatment model and therefore does not produce output or displays that involve the propensity scores or weights.
For more information about the estimation methods that the CAUSALTRT procedure implements, see the sections Estimating the Average Treatment Effect (ATE) and Estimating the Average Treatment Effect for the Treated (ATT).