Dear Stata users,

I have an unbalanced panel data-set (missing about 3.5% of observations, with i=159, t=9, n=1385 out of 1431) with a binary outcome (1;0). I understand that the way to model this outcome is to use xtlogit, and the first step is to find out whether fixed or random effects estimation is appropriate. There is a recommendation on this forum to use the Hausman test, just as one would with a continuous outcome, and there does not seem like there have been any newer suggestions ever since (here). My problem is that it turns out that there is no within variation for many i's, so when I run xtlogit y x, fe where Y is the binary outcome and X is a set of predictors, Stata omits multiple observations leaving only 29 groups and 245 observations.

Code:
. xtlogit Y X i.year, fe
note: multiple positive outcomes within groups encountered.
note: 130 groups (1140 obs) dropped because of all positive or
      all negative outcomes.


Conditional fixed-effects logistic regression   Number of obs      =       245
Group variable: id                              Number of groups   =        29
Given that, I would like to ask your help with the following questions:
  1. Would it still be reasonable to rely on "hausman fe re" results to pick the right estimator (i.e. xtlogit fe or re)? This seems to be the case in the advice mentioned earlier.
  2. If so, what is the right course of action if the the fixed effects is preferred but the vast majority of observations shows no within variation and is omitted?
  3. If the random effects is supported by Hausman test, is it a sound approach given the absence of within variation in many cases?
  4. Given these data and assuming xtlogit fe or re are not appropriate in this case, what would be a reasonable alternative estimator?
Thank you in advance for any help.