[FernandoRios]

No, based on what you describe, your results are just as they are.
Infact, if you are using no controls, then you could validate your results by obtaining all the estimates by hand.
Something like
tab year gvar, sum(confidence_govt) noobs nofreq nost
And then construct the DIFF in DIFF based on that data. This is what CSDID would be doing for you. So it would be a good exercise.

[manon ritto]

thank you very much!
And do you know why when I add controls such as a democracy index that I interact with time it is not working anymore?

This is how I create the control for democracy for example:

forval t = 1981/2023 {
gen demo_idx`t' = demo_idx if year_of_survey == `t'
replace demo_idx`t' = 0 if year_of_survey != `t'
}

The regression;
reghdfe confidence_govt treatment Gender age demo1981-cont1996 demo1998-cont2022, absorb(year_of_survey Country) vce(cluster Country)


The csdid code:
csdid confidence_govt gend1981-cont1996 gend1998-cont2022, cluster(Country) time(year_of_survey) gvar(start_year) method(dripw) /// ///
agg(event) saverif(rif_example) wboot replace rseed(1)


(I took out one year for colinearity issues, I do not interact the controls gender and age with time)

I have a lot of observations (something around 400 000). As earliest treatments are around 2000, should I reduce the scope?

Thank you again.

[FernandoRios]

You can’t interact variables with time because they are already interacted implicitly the way csdid works

[manon ritto]

I understand now, thank you!

Regarding the absence of ATT before treatment, do you know the reason for it not appearing? And is there a way to fix this issue?

One last question: my data is quite unbalanced, with varying numbers of observations per year and irregular intervals between each available year. Could this imbalance potentially impact the analysis, and if so, is there a way to control for it?