xtoverid with unbalanced panel/multicollinearity

[William Law]

Hi all. I'm running a fixed effects model that includes interaction terms:

 xtreg EPR LK PAU CO zSM LK_zSM zUB LK_zUB zUD LK_zUD zALMP LK_zALMP zEP LK_zEP i.year, fe cluster(country)

I then estimate the above using RE and use xtoverid to test for random versus fixed effects:

 xi: xtreg EPR LK PAU CO zSM LK_zSM zUB LK_zUB zUD LK_zUD zALMP LK_zALMP zEP LK_zEP i.year, re cluster(country)

  xtoverid

For context, using xtoverid on models that omit interactions strongly suggests a preference for fixed effects. However, using xtoverid after the augmented model gives 'Error - saved RE estimates are degenerate (sigma_u=0) and equivalent to pooled OLS r(198);'. Using xttest0 on the RE model further suggests pooled OLS.

I suppose this might indicate that the interactions remove the need to use either FE or RE. However, my panel is unbalanced, and after research it appears that xttest1 extends xttest0 to unbalanced panels. Using xttest1 strongly suggests the presence of random effects.

In this context, I have two questions:

• Can I trust the result of xtoverid if xttest1 suggests random effects are present? Could it be that the unbalanced nature of my panel makes it dodgy?
• Alternatively, might there be an issue with multicollinearity, if xtoverid suggests FE without interactions, but pooled OLS with them included?

Apologies for the convoluted post, and thank you.

[Carlo Lazzaro]

William:
the first issue relates to the data generating process: are interactions nececessary or not?