TWFE - Fixed effects at two levels

[Daisy Dang]

Dear Statalisters,

I am estimating the effect of a goverment support program on individual health, using difference-in-difference with continuous treatment. The amount of support varies across provinces. Data covers 2 years: 2019 & 2021 (2019 as the base year).
I include in my equation both individual FE and province FE.
From this post: https://www.statalist.org/forums/for...nt-level-to-id , I understand that if I use individual FE, any time-invariant province-level effects are automatically built in.
In my case, the treatment is at province level. Including province-level FE will capture all time-invariant province characteristics correlated with the amount of support.

The estimates with and without province FE are totally opposite.
I am not sure whether including province FE is better in my case.
I would appreciate to receive any advice on this.

Many thanks.

[Jeff Wooldridge]

The only reason you can include both individual and province FE is that some people switch provinces; otherwise, the province FEs will drop out. Is that the case?

As per the FAQ, you'll get higher quality responses if you show your data structure through -dataex- and then show us what you typed and what Stata produced. Otherwise, we're a bit in the dark.

[Daisy Dang]

Thanks Jeff for your reply.
Yes, about 30% of my sample switches provinces. Why this is the only reason to include both individual and province FE?
Is there any disadvantage of adding province FE (besides individual FE) in my case, especially on the variation of the treatment?

[Carlo Lazzaro]

Daisy:
as we know, the -fe- estimator wipes out time-invariant predictors: hence, no coefficient for within-panel constants will be calculated.
On the top of that, the -fe- estimator likes time-vaying predictors with enough within-panel variation..
Eventually, as Jeff wisely reminded you to, please stick with the FAQ and:
a) show what you typed and what Stata gave you back;
b) share an example/excerpt of your dataset via -dataex-.
Thanks.