-csdid- with State-year fixed effects

[Parul Gupta]

Hello all

I have a repeated cross-section of districts. Each district has several households (fresh sample selected every year). The treatment is given at district level. I am using -csdid- (ssc install drdid, ssc install csdid) to estimate the treatment effects.

My baseline regression is

        csdid `outcome' `covariates' [weight=hh_mult], time(year) gvar(gvar) cluster(distID) long

I now want to include state-year fixed-effects as a robustness check.

I tried the following:

csdid `outcome' `covariates' i.stateID#i.year [weight=hh_mult], time(year) gvar(gvar) cluster(distID) long

but everything got omitted.

Is there a way to implement this? I read the related threads but couldn't find a convincing answer.

[FernandoRios]

That wont work
1) Implicitly, every variable in the controls is already interacted with the time dummies
2) Adding StateID will only works if you have all Potential states (Not treated and treated ) within each State

[Jeff Wooldridge]

It might be useful to add that pure regression-based methods, such as jwdid, don't have a problem with this scenario. If there are empty cells then the ATTs won't be estimated, and of course they don't contributed to aggregated effects. I believe csdid with method(reg) also would be okay, but it can be less efficient than jwdid. How many districts per state, out of curiosity?

If you want to mimic the typical event study estimation, you'd use the following:

 
  jwdid `outcome' `covariates' i.state [weight=hh_mult], tvar(year) gvar(gvar) cluster(distID) never

As Fernando said, this will interact the state dummies with all year dummies and cohort x year dummies. You will see missing estimates given the problems you report with csdid.

[Parul Gupta]

Thank you Jeff and Fernando, for your responses.

@Jeff: There are large and small states, so number of districts per state ranges from 5 to 30.

@Fernando: I don't understand your second point. Districts are nested within states; in each state, there are some treated districts and some untreated.

I created state-year dummies and used them directly in csdid, and it gave me estimates, however, the sign and significance did a complete 180, and parallel trends tests also fail now. I am not sure how to interpret this.

[Jeff Wooldridge]

Parul: You initially said "everything got omitted" and now you say it gave estimates but they appear to not make sense to you. These are very different. If you show your full output we might be able to provide more help.

The potential problem with csdid is that even if you have treated districts within each state you might not have treated districts for every treated cohort. So if no districts in state A are treated in, say, 2020, but some are treated in 2021, this will cause problems with the propensity score estimation because some treatment probabilities are zero once you condition on state in 2020. Theoretically, this actually shouldn't be a problem because the estimates are of the ATETs, and this should allow zero treatment probabilities. But csdid, like other teffects commands, does not allow perfect predictions in the logit estimation even if it is only perfectly predicting being a control unit.