Pros, Cons and Necessity of Robust Standard Errors in Fixed Effect Models

[Nina Scherl]

Hi,
I am having an unbalanced panel data set with about 300 firms over 35 years.
I am running fixed effect models to check for differences in my DVs between certain time-periods for which I included dummies.

So far I have used the following general command:

xtreg DV TimeDummyBefore TimeDummyAfter i.Year, fe cluster(ID)

which does not show the hypothesized results.

If I run the model without using robust standard errors, some results turn significant.
Therefore, I was wondering whether I actually have to use robust standard errors.
So what are pros and cons of robust standard errors in fixed effect models and how can I find out whether I need to use them?

Thanks,
Nina

[Nina Scherl]

I forgot to mention that I am analyzing subgroups of these 300 firms. So once I look at about 50 firms and once at the remaining 250. I am not sure whether this is relevant or not.

[Carlo Lazzaro]

Nina:
you're dealing with a panel with a large N and a quite large T dimension: hence, it is questionable that -xtreg- is the right choice here; I woud take a look at -xtgls-, as autocorrelation may bite across 35 years (I would skim through the iterature in your research field and see what other researchers did in the past when presented with the same research topic).
As far as the -robust- issue is concverned, under -xtreg- it accomodates for both autocorrelation (which can be a real issue over 35 years) and/or heteroskedadsticity: hence, it seems wise to use -robust- (or -cluster()-) option in your case (please note that, unlike under -regression-), under -xtreg-, both options do the very same job), but this should have nothing to do with statistical significance.
I fail to follow your need of doing different regressions on subsamples: you can simply include a categorical variable among the predictors that split your firms in two subsamples and run an unique regression.

[Nina Scherl]

Hi Carlo,
thank you very much for your helpful comments. I really appreciate them!

(1) Concerning the subsamples issue, I have followed your advice.
In my dataset I have a variable called Characteristic which takes values of 0 or 1 depending on the characteristic of the firm and is for the analysis at hand constant over time (so always 0 or always 1). I am including interaction terms now and run the following command:

xtreg DV TimeDummyBefore TimeDummybefore*Characteristic TimeDummyAfter TimeDummyAfter*Characteristic Characteristic i.Year, fe cluster(ID)

As I am running a fixed effect model, I thought the variable Characteristic would get omitted as it is constant over time.
The hint Stata gives me however, why the variable Characteristic is omitted,is the following: "note: Characteristic omitted because of collinearity".
Is this of any concern? What do I have to do about it?
And am I allowed to include such interaction effects when not all of the variables themselves are included?

(2) Concerning the robustness-issue, I understand the advantage and will use it for my model.

(3) Concerning the xtgls-command, I have one very similar paper at hand which also uses a FE model to estimate the effects.
I investigate a special period for firms and I want to check how the performance of the firm differes before (during 5 years before), during and after (during 5 years after) this special period. For some firms this period is quite long: let's say 20 years, for other firms the period only lasts for about 1-2 years. So for some firms I include 20+5+5=30 years of data for others only 2+5+5=12 years of data. And for some even less than that if I do not have data for the full 5 years before or after the special period.
By knowing this would you still advise me to look into the xtgls command?

Thanking you in anticipation.
Best regards, Nina
(Stata 14.0)

[Carlo Lazzaro]

Nina:
thanks a lot for providing further details.
I will try to reply point-to-point to your questions:
(1) I'm fine with your approach, but your code can be probably improved.
You seem to have created categorical variables and interactions by hand, whereas you can rely on the wonders of -fvvarlist- notation.
I would re-write your code as follows (assuming that your data are in -long- format; otherwise, see -reshape-):

xtreg DV i.TimeDummy##i.Characteristic  i.Year, fe cluster(ID)

where -TimeDummy- includes both -before- and -after-.
Omission due to collinearity is something that you should not worry about (and there's nothing you can do about that, but changing your model specification. However, if your current model specification gives a fair and true view of the data generating process, go ahead and forget collinearity).

(2) Ok.

(3) If the paper you mentioned was published (so that you can include it in your reference list and throw reviewers a bone if requested), go -xtreg-. Obviously, I assume that you have already checked (via the user-written programme -xtoverid-, that you can install via -search xtoverid-; conversely, -hausman- is not the way to , as it allows default standard errors only) that -fe- outperforms -re- specification in your case.