csdid with repeated cross-sectional data

[Gene Korku]

Hello FernandoRios ,

I would like to implement csdid using repeated cross-sectional data. In my setup my data is at the household level. Different households are sampled each year. I am looking to examine the impact of a state-level policy on my outcome of interest. Policies are implemented in a staggered nature and some states have no policy (control group)

In this case what is the appropriate specification when using csdid,

csdid outcome policy, ivar(state) time(year) gvar(policy)

OR

csdid outcome policy, time(year) gvar(policy)

Also, would cluster(state) be appropriate?

[FernandoRios]

Second option and cluster state sounds appropriate

[Marlene Fiesole]

Hi Fernando,

I was wondering, are State/Treatment Level Fixed effects automatically included in the second option:
csdid outcome policy, time(year) gvar(policy) ?

You mention fixed effects at the treatment level also on your github when discussing CSDID with a repeated cross section, but I believe there is not a final answer.

In case they are not accounted for could we think to just demean the data at the treatment level and then run the command (the one I mention in this post) ?

Thanks in advance!

[Jeff Wooldridge]

Marlene: csdid, in the panel data case, is like a differencing estimator, and so the fixed effects are eliminated -- just like when using the fixed effects estimator in levels. It works differently with a repeated cross section but the idea is the same. Still, it is easy to use jwdid (or hdidregress) to explicitly do a fixed effects version of the estimation. If you're not interested in including controls, then csdid has no advantages because it is, in effect, just a bunch of 2 x 2 DiDs using the period before the intervention as the reference period. jwdid can be more efficient because it essentially averages all pre-treatment periods to form a reference period.