Introduction to Causal Analysis Procedures

The CAUSALMED procedure estimates causal mediation effects in situations where the observed effects might have been confounded, such as in observational studies.

In causal mediation analysis, a treatment variable T is posited to causally affect an outcome variable Y through a mediation process that is reflected by the values of a mediation variable M. The total causal effect of T on Y is decomposed into two parts:

To estimate causal mediation effects, PROC CAUSALMED uses the regression adjustment approach proposed by Valeri and VanderWeele (2013) and VanderWeele (2014). For a comprehensive review, see VanderWeele (2015). In this approach, you specify parametric models for the outcome variable and the mediator variable. In addition, confounding covariate effects are dealt with during model estimation.

For example, the following statements specify a causal mediation analysis in which you investigate the causal effects of the treatment variable Smoking (smoking behavior of mother) on the outcome variable Death (infant death):

proc causalmed data=birthwgt;
   class LowBirthWgt Smoking Death AgeGroup Race Drinking;
   mediator LowBirthWgt = Smoking;
   model Death = LowBirthWgt | Smoking;
   covar AgeGroup Race Drinking;
run;

You specify the outcome regression model in the MODEL statement and the mediation model in the MEDIATOR statement. Together, these models specify that the causal effect of Smoking on Death is mediated by the mediator variable LowBirthWgt (low birth weight) and that the interaction effect, as well as the main effects of Smoking and LowBirthWgt, is included in the outcome model. In addition, you specify in the COVAR statement that the covariates AgeGroup, Race, and Drinking might have confounded the observed effects. These confounding effects would then be adjusted for when PROC CAUSALMED fits the outcome and mediator models and estimates various causal mediation effects.

The main results from PROC CAUSALMED include the total effect of T on Y, the natural indirect effect of T on Y that is mediated by M, the natural direct effect of T on Y that is not mediated by M, and the percentage of total effect that is due to mediation. Standard errors for these estimates are computed by analytic formulas or bootstrapping.

PROC CAUSALMED supports binary and continuous variables for the mediator and the following combinations of data type and model for the outcome response:

Other main features of PROC CAUSALMED include the following:

For more information about PROC CAUSALMED, see Chapter 38, The CAUSALMED Procedure, and Yung, Lamm, and Zhang (2018)