Treatment effects with survey data

[John Larsson]

Hello,

I am working with a stratified sample that has sampling ((inverse probability) weights, which I've set up using svyset.
I am interested in testing a categorical outcome based on a categorical variable which conceptually could be thought of as a "treatment."
Does it make sense to use the Stata teffects command in this case, or does it make more sense not to mix the two paradigms? (I.e. Mix the teffects paradigm with the survey strata/cluster/weights specification).
Does it make more sense just to do the conventional logistic regression with the strata/cluster/weights all set up per the svyset command?
Appreciate any feedback,
John L.

[Steve Samuels]

You can estimate the Average Treatment Effect in the Treated ("ATT" or "ATET") by running margins after svy: logistic. See Example 10 in the Manual entry of margins.

You can also get a treatment effects analysis by combining survey and propensity score analyses.. See Dugoff et al. (2014). The models in the simulations of that study are quite simplistic (single confounder, no interactions, no non-linear terms), so I suggest that you also try the margins approach.

DuGoff, E. H., Schuler, M., & Stuart, E. A. (2014). Generalizing observational study results: applying propensity score methods to complex surveys. Health services research, 49(1), 284-303.
Link: http://europepmc.org/articles/pmc3894255

[Steve Samuels]

I should have been more explicit about where the ATT is calculated in Example 10 in the manual entry for margins. It's on page 372 following the paragraph that begins:

To test whether females treated as females is equal to females treated as males, we will need to type...

In the example, consider sex to be the treatment variable and females to be the treated group.

[John Larsson]

Steve, thanks a lot for the info. So would you say that the Stata teffects command is redundant in this particular case? (I.e., in the case where the strata/cluster/probability weights have all been specified.)

[Steve Samuels]

I haven't used the te commands, but I do know that they don't accept a svy prefix. If you read the Dugoff article, you will see the recommendation that the propensity score not be survey-weighted, so perhaps teffects could be used for that part of the analysis.