Fixed Effects with a new Grouping

[Harry Toube]

Hi everyone!

I am running some panel data containing 30 provinces observed across 5 years over two variables. I have run a fixed effects model taking year- and province-fixed effects to estimate the relationship between Var1 and Var2.

Now, I am looking to run this fixed effects model using Region-fixed effects instead of province-fixed effects where a Region is a larger macroarea containing multiple provinces. I have already tagged each province by the region to which it belongs, but I am not sure how to run this updated fixed effects regression without losing within-region variation. I have tried averaging Var1 and Var2 for each Region and running:

xtset Region Year
xtreg Var1 Var 2 i.Year, fe r

But this somehow feels wrong?

Any and all help is very much appreciated!

[Carlo Lazzaro]

Harry:
you may feel uncomfortable with your second code as provences are nested within Regions.
Therefore, the main issue here seems to whether or not a macroarea panel data regression makes sense (personally, I do not feel that way).
In addition (and, in my opinion, much more substantive), you cannot get any informative result with two predictors only (and 150 observations are not that much, too).

[Harry Toube]

Hi Carlo,

Thank you for the feedback.

I am inclined to agree about the observation size, but I am really working in communication with someone more experienced on this topic than myself (though presently unavailable, hence the question not being directed at him) and for reasons internal to the project, it makes sense to run with the small sample size.

With regards to the provinces being nested in Regions, I am a bit of a 'noob' with STATA and do not quite understand why this would be a problem/how it would be possible to run the specification with Region-fixed effects in spite of it?

Thank you for the help!

Best,
Harry

[Carlo Lazzaro]

Harry:
1) thanks for clarifications;
2) switching from provinces to regions as your -panelid- creates no technical problem. My concerns were about its meaning (but it is not my research).
Therefore, the code you shared in your first post should work.
3) The concern about the (too) small number of predictors (whatever your -panelid- is) still apllies, though.