Interpret xtoverid and Mundlak or: How to choose fixed or random effects.

[Luca Haseney]

Hello Stata community,

I try to decide if I should choose a random or fixed model for my analysis.
To do this I utilize:
1. the Mundlak approach following:
https://blog.stata.com/2015/10/29/fi...dlak-approach/

With two time invariant variables. Alternatively I just included one.
bysort FIRM: egen mean_x3 = mean(x3) ... bysort FIRM: egen mean_x6 = mean(x6) quietly xtreg y x1 x2 x3 mean_x2 mean_x3, vce(robust) estimates store mundlak
test mean_x3 ... mean_x6 Result:

chi2( 10) = 28.44
Prob > chi2 = 0.0015


2. Alternatively I utilize xtoverid
xtreg y1 x1 ... x6, re
xtoverid , robust cluster(ID)
Sargan-Hansen statistic 24.298 Chi-sq(5) P-value = 0.0002

My question:

Is it fair to go for a fixed model, instead of a random effects model or did I miss something important/misused the tests?
I think the results suggest with enough confidence that I should utilize fixed effects.

[Joro Kolev]

The null hypothesis for both tests is that the random effects assumptions are not violated.

With both tests you decisively reject this null. So no statistical considerations the fixed effects model is appropriate for your data.

[Luca Haseney]

Dear Mr. Kolev,

Am I right that this then means that random effects are not appropriate?
But I can not infer from this that fixed effect is the way to go?

[Joro Kolev]

Dear Mr. Kolev,

Am I right that this then means that random effects are not appropriate?
But I can not infer from this that fixed effect is the way to go?

Loosely speaking you are rejecting the random effects model assumptions for the alternative set of fixed effects model assumptions.

So in the random vs fixed effects you should choose the fixed effects.

The fixed effects model might be inappropriate for some other reasons which are not tested here.

[Luca Haseney]

Dear Mr. Kolev, thank you very much, I got it!

Is there a suitable way to test or tell if the fixed model is appropriate?
From what I know it is just tested if random or fixed effects should be applied.

[Carlo Lazzaro]

Luca:
as the community-contributed module -xtoverid- does not complain about the lack of a panel-wise effect and the -re- null is soundly rejected, I'd feel confident about going -fe-.

[Luca Haseney]

Dear Mr. Lazzaro,

thank you very much! I will go that way.