I am running a fixed effect panel regression with 81 groups x 20 periods, so approx 1620 (unbalanced) observations.
I use the following to create dummies:
*create timedummy
tabulate refper, generate(refdummy)
which to visual inspection looks fine. But when running the xtreg, it removes 6 time dummies due to collinearity. Can anyone explain what is happening here?
From the labelling of your variables, I suspect that the regressors you include are lagged. But the fourth lag of a variable ("_L4") will only be available from the fifth period onwards, so that might explain why the first four time dummies drop out of your regression: you are not actually using any of the data in the first four time periods to estimate the model. Some of the other time dummies might drop out for similar reasons (maybe one of your variables is only available from period 3 onwards, meaning its fourth lag is only available from period 7 onwards). It looks like you might be regressing a differenced variable ("d") on the fourth lag of another differenced variable ("d" and "L4") in which case it makes sense for the first few periods not to be included in estimation. This automatically means you cannot estimate the time dummies for these period.
xtregcommand you used would be consistent with that and is in fact the single piece of Stata information that is most relevant. That aside: there is a statistical question here, but what is it precisely? Do want to know what collinearity means? If not, what kind of explanation do you seek? We don't have your data to look at correlations or scatter plots but you can do that. $\endgroup$