fixed effects unbalanced panel

[Kate Mamaliga]

Perhaps, this problem has been already discussed. I myself have found that some people have faced with this problem, however I still did not get a clear answer to this problem, and how to proceed in this situation.
I have an unbalanced panel (for most individuals, I have two years of data, and for some individuals only one year). I want to know what stata does to observations (individuals) with only one year of data when I run a fixed effects model with either the xtreg, fe(id) command. It seems like the correct thing to do is to drop observations with only one year of data, but I am getting different results when I explicitly exclude individuals with only one-year of data and when I use Stata's fixed effects commands. In the regression equation i estimate something like that: y_it=var1_it+var_2it+dummy_i+e_it i'm interested in dummy, which is equal to 1 if individual works in the same industry as father's. For those who have it equal to 1 (working in the same industry as father's) dummy is constant in both years. So, I'm doubt if i correctly estimate equation, when in period t i compare the same individuals (who have dummy==1) with a group of some individuals, and in period t+1, I compare these individuals (dummy=1) with a bunch of other individuals, because the data is unbalanced and, subsequently, the comparison group differs between two periods. Of course, the sample composition for the group with dummy==1 can also change from period t to t+1. Can you please help me, what is the best way to proceed in this situation?

[Carlo Lazzaro]

Kate:
the best thing to do is keeping all the osbservations in the dataset (otherwise you 'll end up with a probably biased selected subsample from your original one) and let Stata take care of the unbalanced panel dataset.

[Phil Bromiley]

While I agree with Carlo about keeping data, I think you may have a problem with the single observation panels - your fixed effect essentially drops all panels with one observation. First, I'd first want to understand why you get different parameter estimates when you manually drop these than when Stata essentially does it in the fixed effect.

Second, I wonder if the fixed effect essentially dropping all single observation panels creates a sample selection problem in itself.

[Clyde Schechter]

Phil, Carlo, and Kate,

Alfonso Sánchez-Peñalver and I had a dialog about this issue just yesterday. See http://www.statalist.org/forums/foru...-fixed-effects. The upshot is that if you manually exclude the singleton observations in -xtreg, fe-, the constant term (which is artificial anyway) will change, as will the estimated panel-level effects (which are also artificial), and the variance component estimates. But for the predictor coefficients and standard errors, which are not artificial and are usually what one is most interested in, the results should be the same either way. If Kate is getting different results for those parameters, then something is wrong and she should show the exact code she is running and the exact output from Stata by copy/pasting directly from the Results window or her log file into a code block on this forum. If the coefficients and their standard errors are the same, and the other parameters are the only ones changed, then nothing is wrong and those differences have no substantive consequences anyway.

And I certainly agree that dropping singleton panels risks introducing an appreciable bias into the sample. It won't affect the main results of fixed-effects regression, but it would likely affect other analyses.

[Alfonso Sánchez-Peñalver]

As Clyde Schechter mentions, removing the singletons would change the estimate of the intercept and of the panel-level effects in an -xtreg, fe- estimation. It will not change, however, the sum of both for the panels kept in the estimation, since the sum of both represents the panel-specific constant. The way that Stata calculates the constant and the panel-specific effect is so that the constant is comparable to the constant you get in a -xtreg, re- estimation, which is what led me to realize that removing the singletons may not be appropriate for specification tests such as the one done with -hausman-. Not so much because of the value of the intercept, but also because of the difference in the number of observations used in both estimations.