0

I am running a panel data analysis with the following problem:

  • the variance of the individual effects in the random effects model is negative

I've read that one possibility is to set the variance to zero, which transforms the random effects model in a fixed effects model: link. Does that mean that the random effects model becomes a fixed effects model if I assume the variance to be zero and not negative?

mpiktas
  • 35,099
Grunez32
  • 21

1 Answers1

1

Recall that usual estimates of individual effect variance are calculated under assumption that the regression disturbances are not serially correlated. Hence if you get negative estimates, you should check whether serial correlation is present.

If serial correlation is not present then you should test whether individual effect is present at all. If you cannot reject hypothesis that it is zero, this means, that there is no individual unobserved effect and you can use simple pooled regression.

If the serial correlation is present this means that the usual random effects estimator is not suitable. You should then either include additional variables into your model, or use a GLS approach. The standard fixed effects estimator would be consistent in such case, but you should then use robust standard errors, since serial correlation would still be present.

mpiktas
  • 35,099
  • Thank you for your answer. Additionally, I have to say that if I test the fixed effects model agains OLS the results indicate to use the fixed effects model. However, I cannot compare fixed effects with random effects (Hausman test) to determine which model to use (as the random effects model cannot be generated due to the negative variance). Hence, it is possible to say that that the fixed effects model is the correct model? – Grunez32 Dec 15 '15 at 10:33
  • I've updated my answer. Hausman test has a serial correlation robust version, but for that you still need the estimate of individual effect variance. Since it is negative, this means that there is something wrong with your model and switching to fixed effects would not help. – mpiktas Dec 15 '15 at 11:07
  • I have serial correlation in my model. Therefore, it is able to use the fixed effects model under the consideration of robust standard errors. – Grunez32 Dec 15 '15 at 11:11