No result for F-test in fixed effects regression (clustered)

[Karolin Knapp]

Hi everyone,

Stata does not display a F value for the following command. Unfortunately, I cannot display the output as I am using confidential data, but I can share the information below.

Does anyone know what I could do about this ? The corresponding OLS and random effect regression does yield an F-test for overall significance and it is ok. Do I have to worry about this then?

xtreg invbv i.exbo_d i.exbo_d#c.epa_pos_scale i.exbo_d#c.epa_neg_scale i.exbo_d#c.opa_pos_scale i.exbo_d#c.opa_neg_scale i.brsect_d fyear_c empt_c, fe cluster(idnum)

Fixed-effects (within) regression Number of obs = 8,028
Group variable: idnum Number of groups = 2,816

F(20,2815) = .
Prob > F = .

Many thanks in advance for any helpful comment!

[Clyde Schechter]

The commonest cause of this problem is having more predictors than there are clusters, but that it not the case here: you have only 20 predictors, and 2,816 clusters. The other situation that can give rise to this is if some of the clusters are singletons.

The question, after you identify your singleton cluster(s), is what, if anything to do about it. The singleton clusters do not in any way invalidate the results for the individual predictor variables, and they are all there. It is only a problem for the omnibus model F-statistic. But usually the omnibus model F-statistic is of no real interest anyway. So first decide whether you really need it: usually it is just ignored anyhow. If you really do need it, the solution is to combine any singleton cluster with some other cluster--the trick is to do it in a meaningful way. If your clusters were countries, you wouldn't want to combine, say Afghanistan with Brazil--the two countries are very dissimilar in almost all respects. So you would want to combine like with like to the extent possible.