Individual FE in Staggered Diff-in-diff: Poisson

[Sarita Vaswani]

Hi everyone,

I’m studying the effect of a policy on crimes against women using district-level panel data. The policy was implemented at the state level, but the timing varied across states. Since my outcome variable is a count, I’m inclined to use a Poisson model. I’m following the framework outlined in "Simple Approaches to Nonlinear Difference-in-Differences with Panel Data" by Prof. Wooldridge. I’ve written my code and also experimented with the jwdid package. However, as many of you know, this approach does not allow for individual (district) fixed effects.

In my context, I believe district fixed effects are important because the only available time-varying district-level control is population. Fixed effects would help absorb a lot of unobserved heterogeneity across districts that group FE likely cannot. So my question is: Is there a package or approach that allows me to estimate a nonlinear DiD model (e.g., Poisson) with individual fixed effects?

[FernandoRios]

hi Sarita
Not sure why you say it doesnt allow it
if you use jwdid, method(ppmlhdfe) you can use ivar(district) to control for those fixed effects.
Perhaps there is something different in your mind?
F

[Sarita Vaswani]

Hi Fernando,

Thank you for getting back. This is what I was running:

jwdid dowrydeaths totalpop , ivar(district_num) tvar(year) never gvar(treat_timing) method(poisson)

It only takes into account the group fixed effects. And since my read of the paper - Simple Approaches to Nonlinear Difference-in-Differences with Panel Data - was that the approach requires group FE, I was wondering if there is a reason that we don't include individual FE. I will try to use the code you mentioned.

[FernandoRios]

Right poisson doesn’t accept individual fixed effects
ppmlhdfe does

[Jeff Wooldridge]

Without control variables and a balanced panel the estimated coefficients are the same. So saying you “need” district FEs is incorrect. With covariates they are different, but often close. If you insist on district FEs you can use xtpoisson with i.year and fe or, as Fernando says, ppmlhdfe. But you won’t get reliable estimates on the level that allows a comparison with the linear model.