I think the match variables can be used since they have different purposes.
When you match on gender with PSMATCH it's to ensure there's no different distributions of gender that could affect the outcome. Simpsons Paradox often occurs when you have mismatched sizes of groups, ie one group is larger than others.
When you use it in the analysis, it's to determine if the different genders has an effect on the outcome between the different treatments.
@Chris9 wrote:
Should variables used in PROC PSMATCH (or PROC LOGISTIC prior to PSMATCH) be re-used in the outcome analysis?
For example, if we match on gender, should we use gender in the outcome analysis? Here are the imagined steps:
1. Calculate the propensity score with gender as a variable (e.g., treated = gender) in PROC LOGISTIC or PSMATCH
2. Match on gender with PSMATCH
3. Run the outcome analysis with PROC UNIVARIATE using gender in the model
See this for example:
https://documentation.sas.com/?docsetId=statug&docsetTarget=statug_psmatch_examples05.htm&docsetVers...
Why should gender be in the final analysis? Just to check that it's balanced and has no effect? Couldn't it be just left out if the standardized differences are low (e.g., less than 0.10)?
