How to use survey weights in the calculation of the standardized mean difference after cem?

[Leo Nard]

Dear all,

My question is about the use of survey weights in the calculation of the standardized mean difference after using the coarsened exact matching method.

To calculate the difference in grades between students (with gender as covariable), I proceed as follows:

cem sex, tr(treated)

Then I get cem_weights that I multiply with the weight of my survey:

gen weights=cem_weights*survey_weights

Then I run the regression to estimate the causal effect from cem output as follows:

regress grad treated [pw=weights]

To verify that my individuals did not differ too much between the treatment and control groups I wanted to calculate the standardized mean difference for each covariate from the stddiff function by proceeding like this

stddiff sex, by(cem_weights)

It works. However, I can't find instructions for incorporating the weights from my survey. Do you have an idea?

Thanks a lot,
Léo Nard

[Leo Nard]

Any idea? Or maybe my question doesn't make sense?

Thanks

[David Radwin]

I have a few suggestions.

1. First, a request on posting etiquette. As noted in FAQ 12.1, "If you are using community-contributed (also known as user-written) commands, explain that and say where they came from: the Stata Journal, SSC, or other archives." cem was written by Matthew Blackwell, Stefano Iacus, Gary King, and Giuseppe Porro and is available from the Stata Journal (findit cem). stddiff was written by Ahmed M. Bayoumi and is available from SSC (ssc describe stddiff).
2. I'm not sure your approach of multiplying the survey weights by the cem weights and using them as sampling weights (pweights) will yield valid results. For an approach to combining matching with regression, see publications cited and discussion in this or this.
3. I'm not sure why it is helpful to use weights in any fashion for calculating standardized mean differences. Weights are used to generalize results from a sample to a population. But most matching methods (including coarsened exact matching) discard observations that are outside common support or cannot be matched. This winnowing of the sample is good for causal inference, but it also means that the sample no longer represents the population even with weights. That might be fine for causal inference (e.g., using the approach linked above), but it's not clear why a diagnostic test like standardized differences should be generalized to the population.

[Leo Nard]

Dear David,

Thank you for your response
1) I apologize, I didn't know this rule and will use it in my next posts.
2) Thank you for the references, I was using this discussion as a basis for the multiplication of the weights.
3) I totally understand what you are saying, it is very clear, thank you!

Best