You use a propensity score model to estimate the probability that an individual receives a treatment. Then you use the estimated probabilities to create a matched, weighted, or stratified data set to which you apply a subsequent outcome analysis to estimate the causal effect of interest. Thus, a valid propensity model and an appropriate propensity score–based adjustment is critical to a valid estimation of the causal effect.
Checking covariate balance that results from a propensity score–based method (matching, weighting, or stratification) is perhaps the most common practice to justify the appropriateness of the propensity score–based adjustment. Although PROC PSMATCH provides many numerical and graphical tools for checking covariate balance, you can apply only a single logistic regression to model the propensity scores in a given analysis. If the covariate balance is not satisfactory, you might want to try other propensity score models that can lead to better covariate balancing. There are several possibilities for exploring suitable propensity score models:
Add higher-order effects to the existing logistic regression of propensity scores.
Replace the covariates in the existing logistic regression of propensity scores with a new set of covariates.
Model the propensity scores by a technique other than logistic regression, possibly with a new set of covariates and/or a new set of covariate effects.
You can use PROC PSMATCH to explore the first two possibilities, but you must rely on other statistical procedures to explore the third possibility. Fitting a propensity score model in an external procedure might also be necessary when there is complete or quasi-complete separation of the data and the maximum likelihood estimates for the logistic regression model do not exist or are not unique.
When these other procedures produce the estimated propensity scores, you can input the propensity scores into PROC PSMATCH by specifying the PSDATA statement. PROC PSMATCH then examines the covariate balance by using these input propensity scores and completes the remaining tasks of the specified propensity score analysis. For illustrations of using the PSDATA statement, see Example 101.8 of Chapter 101, The PSMATCH Procedure, and Example 1 of Lamm, Thompson, and Yung (2019).