csdid: Interpreting ATT with covariates and testing gender heterogeneity

[Adriano Ruggiero]

Dear all,

I am estimating a staggered-adoption DiD with csdid (Callaway & Sant’Anna) and my baseline is

csdid outcome, ivar(id) time(year) gvar(first_treat) method(drimp) wboot rseed(999) notyet long2

I would appreciate guidance on three related issues: (i) when I include a conditional variable or other controls, how should I interpret the reported effects ATT? Furthermore, is it possible to use only time-invariant variables?
I want to assess gender heterogeneity without dropping observations; I know one can run a restricted sample, e.g.,

csdid outcome if gender==1, ivar(id) time(year) gvar(first_treat) method(drimp) wboot rseed(999) notyet long2

,

but this removes many observations and may change the composition of comparison units (“not-yet treated” controls). Within csdid, is there a recommended approach to obtain female and male effects jointly while preserving the original control structure and directly testing whether they differ (e.g., an officially supported postestimation route to aggregate ATT(g,t) by a grouping variable—something like an over(gender)—or a unified estimation that produces a formal test of the female–male difference)?
Consider that gender does not change for the same ID over time.

Furthermore, where should I place the conditioning variable: is the best/only supported way to include gender as a control in xvar() (so it enters the DR/propensity components), or is there any supported syntax to place it “next to the dependent variable” in csdid?


Any pointers to the correct syntax, postestimation commands, and references for interpreting ATTs with covariates and estimating/contrasting gender-specific effects without resorting to restricted samples would be greatly appreciated.

Many thanks,

Adriano Ruggiero

[Jeff Wooldridge]

You can use jwdid and include gender as a control, and this will report the moderating effects of gender.

[Adriano Ruggiero]

Thank you, Prof. Jeff Wooldridge

that’s very helpful. I’ll try jwdid with gender included as a control, as you suggest.

For consistency with the rest of my project, most of which is based on the Callaway & Sant’Anna framework, is there a comparable way to assess moderation by gender within that approach?

[FernandoRios]

Not directly
you already considered doing this by group. And that is the only approach.

[Adriano Ruggiero]

Thank you, Professor Jeff Wooldridge

My understanding is that including gender as a control adjusts for composition, but to actually report moderation (ATT by gender) in jwdid we use post-estimation with over(gender). For example, in my setting:

jwdid outcome gender, ivar(id) tvar(year) gvar(first_treat)

estat simple, over(gender)

This way I keep sex as a time-varying control and then obtain ATTs separately for men and women.

Is it correct in this way or I have to use gvar()/gtvar()

[Aruppillai Thayaparan]

i need a help from this forum

i have stata 17 and need to run conditional average treatment effects and heterogeneity effects and also multinomial endogenous switching regression models with commands and interpretation

[email protected].

[Aruppillai Thayaparan]

how i use svyselmlog command to run multinomial endogenous switching regression mopde in stata 17?