panel data fixed effect

[Olivia Li]

Dear all
I am encountered with a pproblem when runing panel-data DID regression. When I try to absorb year month FE, two different way of absorbing leads to different coefficient estimation (and these two estimations vary from each other a lot)

My question is, what would be the reason for this?

Here is the code that I use:

[CODE]
reghdfe lnwage shock Controls , absorb( ID year month) cluster(ID) resid
reghdfe lnwage shock Controls , absorb( ID year_month) cluster(ID) resid
[END]

Thanks a lot!

[Jared Greathouse]

Uh in your first example you use id year and month fixed effects, in your second one you use id and year-month fixed effects. I don't know why you'd want to do it like this, for your first example you have ID-1 year-1 and 12-1 predictors, and in your second example you have ID-1 and however many year-month predictors there are.

So, that's why

[Carlo Lazzaro]

Olivia:
your first code includes three fixed effects (-panelid-, year and month), whereas the second one includes -panelid- + an interaction between year and month (that I find difficult to justify unless you have relevant seasonality differences across months ).
That said (and without knowing the rests of the outcome that the community-contributed module -reghdfe- gave you back), I would stick with your 1st code.

[Olivia Li]

Uh in your first example you use id year and month fixed effects, in your second one you use id and year-month fixed effects. I don't know why you'd want to do it like this, for your first example you have ID-1 year-1 and 12-1 predictors, and in your second example you have ID-1 and however many year-month predictors there are.

So, that's why

Hi, Jared, the reason I wanna do this in second regression is that I assume each month in each year has a fixed effect. And in the first regression, I assume there is no difference in the same month in different years.
And soory, I am a little confused with the "why" part that you said.

[Jared Greathouse]

Carlo put it more eloquently than I did, but when you include year-month fixed effects, you're essentially including unit specific linear trends.

Example? Say we have 4 units and 10 years. Two way fixed effects means we have 3 unit and 9 time dummies. If we had unit_year fixed effects, we now have one indicator, per state, per year. Or, 26 predictors if my mathematics is right

[Olivia Li]

Olivia:
your first code includes three fixed effects (-panelid-, year and month), whereas the second one includes -panelid- + an interaction between year and month (that I find difficult to justify unless you have relevant seasonality differences across months ).
That said (and without knowing the rests of the outcome that the community-contributed module -reghdfe- gave you back), I would stick with your 1st code.

Hi Carlo, thanks for your kind reply.
It seems that year_month FE (the 2nd one) controls more variations than year month FE (the 1st one), so I am not sure if 1st is the right one. (Although I am with you for the point that the hypothesis of the 1st one complies with economic intuition more than the 2nd one.)

[Jared Greathouse]

You'd only use regression 2 if you think there's legit seasonality. Like year and month specific confounding.

[Carlo Lazzaro]

Olivia:
it's difficult to advise any further without taking a look at your regression outcome table.