Cohort controls with CSDID

[Nadia Ali]

Dear all: I have a panel dataset with staggered treatment. There are five groups/cohorts. In the TWFE regression, I usually include cohort fixed effects. Are the cohort fixed effects obsolete when I use the CSDID regression with the "group" aggregation option? How about when I use it with the "simple" aggregation option?

[FernandoRios]

Hi Nadia
CSDID automatically controls for cohort fixed effects (since they are nested within individual fixed effects).
Aggregations relate simply on how to summarize all ATTGTs.
HTH

[Nadia Ali]

Thank you very much for your reply. I should clarify that we randomized within each group, i.e. our randomization was staggered. If the cohort fe is controlled for automatically, should it not be omitted? Instead, I get different estimates depending on whether I include it. Here is my sample code, please note that the variable group is the group variable (1 through 5), each group i has control firms (for which whentreat=0) and treatment firms (for which whentreat= t_i).

csdid yvar i.group , ivar(dossierstring_byte) time(monthyear) gvar(whentreat)

Vs.

csdid yvar, ivar(dossierstring_byte) time(monthyear) gvar(whentreat)

[FernandoRios]

Im guessing that group and whentreated are collinear (or the same). In that case, you are correct. you should not included it as explanatory variable

[Nadia Ali]

Thank you. To clarify: they are not quite collinear: each group has two whentreat values (0 for control units and t_i for treated units).

[Castor Comploj]

Consider a setting where treatment (eligibility) is a deterministic function of age (e.g. at age 21) and I do not want to use RD. I want to include multi-year cohort controls (e.g. a range of 5 birth years), as I do not want to compare people aged 21 (after treatment) with people aged 15, but with people aged 20 (who will later be treated). Does any of this sound feasible in -csdid-? How are individual controls fixing this?

[Castor Comploj]

Answering #6:
It does not make much sense to use diff-in-diff in such a case when there is a clear age-profile of the outcome (ageing would be correlated with the treatment).