Applying Weights from a IPWRA to Sub-Group Analyses

[Mitch Lingo]

Hello,

I am using

teffects ipwra

command to look at whether a financial incentive does or does not increase retention. Both my outcome (retention) and treatment (financial incentive) are dichotomous, requiring a logit for each equation in the IPWRA (leaving help from this link moot).

In the analysis, I provide an overall look at retention at 1 year, 5 years, and 10 years. Due to varying differences between employees, I would like to look at sub-classifications. For example differences by gender may occur from women to pursue staying at home as they have children with a partner. My original intent was to just use the

if gender==0

and

if gender==1

options to perform each analysis on its own. Then it was pointed out to me that the analysis would use different weights between the overall, male, and female analyses.

So I am wondering if there is a way to use weights created from the overall analysis and apply those weights to analyzing the various subpopulations in an IPWRA? If not, that's cool. Just trying to gauge the possibility.

TIA

[Clyde Schechter]

Instead of doing separate regressions by sex, why not add treatment X sex interaction to the model?

[Mitch Lingo]

I gave that a thought and a try earlier. From my understanding of the command structure, it only allows for dichotomous treatments. So the if command was going to be my workaround. The model is something like this as of right now.

teffects ipwra (retention i.gender y2 y3 y4, logit) ///
            (receivedaward i.gender y2 y3 y4), vce(cluster unit)

Also, even if the interaction were to work in the teffects command, the command would apply a different set of weights than the weights in the overall analysis.

[Clyde Schechter]

OK, I see the problem. I have made very little use of -teffects- myself, and not in a context like this. I don't know if Stata will accept this, but have you tried:

teffects ipwra (retention i.gender i.gender#i.receivedaward y2 y3 y4, logit) ///
    (receivedaward i.gender y2 y3 y4), vce(cluster unit)

If the mention of receivedaward in the first equation causes Stata to reject that, try it with homebrew interaction.

[Mitch Lingo]

Yeah, that will not work either as the treatment variable cannot be in the first equation.

I am pretty sure the homebrew (assuming you are getting at categorical variable of Female-No Award, Female Award, Male-No Award, Male-Award) interaction will not work either. Reason being that whatever ATE the IPWRA comes up with will only have one ATE for "receivedaward" instead of one for each Gender-Award dyad as specified by the second equation.

Thanks for the ideas.