[paulvonhippel]

Back to the dots. In our data, there are 20 treatment cohorts, so I would expect 20 dots. Instead I see over 100 dots. Should I not expect one dot per cohort?

[FernandoRios]

No it’s one dot per cohort per period. Basically every single attgt estimate is a dot or combination of cohort periods

[FernandoRios]

Regarding fixed effects, you can’t because then you wouldn’t have overlapping condition
you could do it under linear model in which case I recommend jwdid

[paulvonhippel]

Adding a dummy for each school didn't work out so well. The following code produced dots without the school dummies but Xs with them. I'm not sure I understand why.

csdid obese ib(5).ethnicity_code ib(freq).school_id, time(spring) gvar(cohort) agg(simple) cluster(school_id) dripw notyet
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx

[FernandoRios]

Hi Paul,
Few of points.
1. The way you are calling CSDID will use "dripw" (default) method for DID. If you want "DRIPW" you will need method(dripw)
2. All Propensity based methods (dripw drimp ipw stdipw) require strong overlapping in covariates (Specially drimp). Now, adding school_id as dummies will have two problems.
a) Incidental parameter problem that occurs on the logit (IPT) model estimation.
b) If some schools are not treated, and others are fully treated (every student is), then this will cause perfect prediction, which will create the errors you see (Xs)
3. Perhaps a better approach would be using method(reg), because it will work by creating extrapolations, being more relax about the overlapping condition. However, this leniency does not apply when the covariates are categorical.
4. I think one option to add (by force) Fixed effects could be using something similar to Mundlak approach. That is (in your case) add average of all ethnicity dummies to the model, by School.

Hope this helps
Fernando

[paulvonhippel]

Thanks FernandoRios! My responses:

1. I think you're telling me to replace the option dripw with the option method(dripw). Is that right? You say method(dripw) is the default method, so by specifying dripw I'm just asking explicitly for what csdid would do anyway. Am I understanding correctly?
2. (b) I think what you're saying is that if treatment is applied at to entire schools, not individual children, then school dummies will be collinear with the treatment cohorts. Am I understanding correctly?
3. Thanks for the suggestion. I'll stick with dripw for now.
4. I think your suggestion of adding school means would work, but I also think that I can center the X and Y variables around their school-level means, which is what the xtdata, fe command does. That's another way to put in school fixed effects.

Best,
Paul

[paulvonhippel]

Another question: what value should the gvar() variable have for cohorts that never receive treatment? The documentation says this:

e.g., 0 Never treated, 10 treat at t=10, 20 treated at time 20, 25 treated at time 25.

I suppose it's arbitrary to use the value of 0 for "never treated." Any value would do, as long as it's distinctive, right? The software can figure out from the treatment dummies that the cohort was never treated.

[FernandoRios]

Thanks FernandoRios! My responses:

1. I think you're telling me to replace the option dripw with the option method(dripw). Is that right? You say method(dripw) is the default method, so by specifying dripw I'm just asking explicitly for what csdid would do anyway. Am I understanding correctly?
2. (b) I think what you're saying is that if treatment is applied at to entire schools, not individual children, then school dummies will be collinear with the treatment cohorts. Am I understanding correctly?
3. Thanks for the suggestion. I'll stick with dripw for now.
4. I think your suggestion of adding school means would work, but I also think that I can center the X and Y variables around their school-level means, which is what the xtdata, fe command does. That's another way to put in school fixed effects.

Best,
Paul

1) Yes, that was my suggestion. Method() explicitly indicates the method to be used for estiation.
2) Correct, School Dummies will be colinear, which would lead to perfect prediction.
3) Not sure if centering would work. Specially if you are still using them in combination with the logit (dripw)

Lastly, It has to be zero. That is how I coded CSDID. Zero is never treated, any other positive value is period of treatment. In R never treated is coded as "." if I recall correctly.
F

[paulvonhippel]

When you say you're not sure if centering would "work," what do you mean? The software wouldn't run, or you wouldn't trust the resulting estimates?

[paulvonhippel]

By the way, is it possible to use [iweight=] and method(dripw) at the same time. Both weight the estimates, no?

[FernandoRios]

Not iweight. All weights are used as pweights
and correct. I’m just not clear if that would work on the propensity score calculations.
The work by prof Wooldridge suggests that mundlack works with nonlinear models, but I’m not aware about the logic with respect to models with demeaned data.
worth to try

[paulvonhippel]

Regarding the Mundlak approach: It wouldn't be enough to only put in the school's mean for the covariates (e.g., ethnicity). Wouldn't you also have to put in the school means for the treatment dummy?

[FernandoRios]

Not sure on that. It would certainly be a problem for overlapping again.