Insignificant time dummies (both indivdually and jointly) when using Random Effect Logit

[Ashley Batten]

Hello all,

I am using a random effect logit model and have been advised it is best to include year dummies using the 'i.year' command. However, all of these time dummies that have not been removed due to collinearity and such are not statistically significant. I spoke with someone knowledgable on econometrics as whole and his advice was that omitting the time dummies was completely my call, however I am unsure as to what dropping them as opposed to keeping them in would mean for my model. Would anyone be able to advise as to how I should be interpreting this? Should I be dropping the time dummies and running the re model without them? I do lose 1200 observations when compared to not using 'i.year' but am not sure if this could be a justification for omitting the dummies.


I cannot find any literature on this subject, so if anyone could point me to any it would be appreciated.

Thanks!

[Phil Bromiley]

Very often, folks run tests that the coefficients are jointly zero and then drop them if you cannot reject the hypothesis that all the time dummies are zero.
Alternatively, run it both ways. If you get the same results, then you can report both as a robustness check. I would make sure I understood why you're losing piles of observations when you include year - surely you don't have missing on year.