Group-year FEs do not absorb the group-level treatment

[Phuongmh Le]

Hi Statalist,

I'm running a DiD with a continuous treatment at the group (industry) level. The treatment does not vary over time. The post-treatment period has one year (2020).

The interaction term between the treatment and year is expected not to be significant when the year-industry fixed effects are included, as it is absorbed by the year-industry fixed effects. However, in my case I find it significant. Could you please suggest some explanations?

Thank you very much.

Phuong

[Carlo Lazzaro]

Phuong:
mamy of your interactions are omitted due to collinearity.
My guess is that you should consider a more parsimonious model.
As an aside, please use CODE delimiters to share what you coded and what Stata gave you back (as per FAQ). Thanks.

[Phuongmh Le]

Dear Carlo,

Thank you very much for your reply

You are right about the multi-collinearity. The coefficients on c.treatment#i.year have very large VIF. Including time-industry dummies is clearly not a good choice.
Maybe I should include more-aggregated industries to control for the time-varying industrial heterogeneity.

However, there is something I still don't understand. Multi-collinearity should increase the standard errors and lead to insignificant estimates. While in my case, including time-industry dummies make the coefficients on c.treatment#i.year significant.

As for the CODE delimiter, since the stata output is too long, I can't put it within the delimiter.

Best,
Phuong

[Carlo Lazzaro]

Phuong:
as the limits of the 95% CO of your paraneters do overlap, my guess is that there's a serious misspecification in your model.
Again, I'd consider a more parsimonious model, possibly on the grounds of the literature in your research field.
As an aside, whenever I face problems with sharing what I did via CODE delimiters, there's often something to deal with in my regression (or whatever else I'm working on).

[Phuongmh Le]

Dear Carlo,

Thank you again for your suggestion. I will have to consider other models then.

All the best,
Phuong