[FernandoRios]

Hi Daniel
I think the problem you are experiencing is that you are overfitting the model.
For instance, consider the last table you provide, and consider the cohort 1982. When estimating the ATTGT at T=1982 you are using 59 observations for the post treatment period, and 61 for the pretreatment period.
They are what I call effective sample size. So given that small sample, I would be very careful in using controls, because you will fall into perfect prediction or dropping observations when you try IPW or DRIPW.

reg is more forgiving because it just uses OLS, which automatically drops/omits perfectly colinear covariates. But the results will be very sensitive.
So, I would start by reconsidering what variables you want to control before using csdid.
Best wishes
Fernando

[Daniel Perez Parra]

Thanks for your reply FernandoRios ,

Do you mean the overlapping assumption? where for every value of my controls, there is at least a small probability that the observation is not treated. For instance, given that regions define whether the cohort gets treated or not (the rollout is by region), that wouldn't respect the overlapping assumption.

[FernandoRios]

Precisely!
Region would be a terrible control.
Your logit would not be able to estimate anything, and your OLS will create nonsensical iterpolations

[Daniel Perez Parra]

Precisely!
Region would be a terrible control.
Your logit would not be able to estimate anything, and your OLS will create nonsensical iterpolations

Alright, now I understand. However, using religion and ethnicity as controls still reproduces the same problem, and I really don't understand why since there are observations of almost all ethnicities and religions across the treated cohorts. Is there any way(command if available) of testing the overlapping assumption? Btw I took another survey and I could duplicate the number of observations to check whether this helped to overcome the overlapping barrier, but it didn't. Sadly, not including ethnicity and religion as my controls would make me suspect of no parallel trends assumption in my DD model.

[CODE]
tab gvar_gh ethnicityds

| ethnicityds
gvar_gh | amhara oromo tigrie affar somalie guragie sidama other | Total
-----------+----------------------------------------------------------------------------------------+----------
0 | 250 168 50 1,693 1,814 16 10 83 | 4,084
1982 | 1,822 1,746 244 9 282 708 3 297 | 5,111
1985 | 3,191 2,935 12 2 9 49 3 341 | 6,542
1986 | 116 73 6 0 0 239 619 2,291 | 3,344
1989 | 78 10 2,679 3 1 0 1 28 | 2,800
1990 | 1,204 1,632 96 2 9 184 4 2,573 | 5,704
-----------+----------------------------------------------------------------------------------------+----------
Total | 6,661 6,564 3,087 1,709 2,115 1,196 640 5,613 | 27,585


gvar_gh | orthodox protestan moslem other | Total
-----------+--------------------------------------------+----------
0 | 195 34 3,839 16 | 4,084
1982 | 2,924 359 1,796 32 | 5,111
1985 | 3,410 825 2,211 96 | 6,542
1986 | 543 2,178 415 208 | 3,344
1989 | 2,660 3 134 3 | 2,800
1990 | 1,565 1,603 2,343 193 | 5,704
-----------+--------------------------------------------+----------
Total | 11,297 5,002 10,738 548 | 27,585


[\CODE]

[FernandoRios]

I wonder if there is something else going on with your data and model
Can you email me and send me your data (only what is required for the model) along with the exact command you are using?
I want to check what is going on
F

[Daniel Perez Parra]

Done! Thanks a lot, it's very appreciated. Let me know if you need something more from my side.

Daniel.