Panel data with two way (id and day) fixed effects

[Christoph Winter]

Hi,

I am using Stata 12/SE.

I have a panel data set with daily prices (about 700 days) for 15000 firms. I want to estimate a two-way fixed effects model with id and time fixed effects.

Code:

xtset id date
xtreg var1 var2 i.date, fe

However, Stata calculates for days without any result. I think there are just too many dummy variables.

Is there a way to improve the process?

Best,
Christoph

[Marcos Almeida]

Hi, Christoph,

I gather you don't need to add "date" after xtreg, since you did xtset this variable.

You might well try again the commands without "date" and see if you succeed.

Below, a link to the Stata manual on the subject:

http://www.stata.com/manuals13/xtxtreg.pdf

Hopefully that helps.

Best,

Marcos

[Christoph Winter]

Thanks for your commend. However, I am not sure whether I got your point.
So, what I want to control for are id fixed effects as well as time fixed effects (i.e. cost shocks for all firms at the same day).
Why do you think I do not have to include i.date?

[Andrew Musau]

Try running OLS with both firm dummies and time dummies. This is equivalent to two-way fixed effects

reg x1...xn i.id i.date

Also verify that your date variable has integer values e.g. 1- 700

[Carlo Lazzaro]

Christoph:
as an aside to other skillful insights, I would advise you to add -cluster- option to standard errors in pooled OLS.

[Sergio Correia]

Dear Christoph,

"stata calculates for days" is usually because your computer memory (RAM) is too small to create all the required dummies. Instead of -xtreg fe- you could try -reghdfe- (install it from ssc with "ssc install reghdfe"). It works like -areg- but you can put more than one variable in absorb().

Best,
Sergio

[Christoph Winter]

Thanks, guys!
I tried areg var1 i.date, absorb(id) and reghdfe var1, absorb(FE_DATE=i.date FE_FIRM=i.id) and got the same residuals.
It took a few minutes but it both worked.
However, is there a reason why this works just fine and xtreg doesn't or is it likely that I made a mistake?
Best, Christoph

[Andrew Musau]

Christoph: As Sergio says, it is likely that your computer memory is too small to estimate all dummies when running -xtreg-

To check this, try using a sub-sample of the data, running -xtreg- with say 500 firms and 50 days. This should work.

-areg- is probably efficient in memory usage.

Incidentally, if Stata had a direct command to run 2way fixed effects (without having to add time dummies), this would not be an issue.

[Christoph Winter]

Ok, I see, it is not that easy
I also tried _regress var1 i.date, absorb(id). It also works but I get different residuals. Is there an explanation for that?