A question about longtitudinal data analysis: what is the difference between xt..., pa vs. xt..., fe ?

[Kenny Don]

Hello , I am a master student and start to learn the panel data, it took me a week to read all the stuff of panel data and Stata command. Now I have a question when I learn GEE analysis.
the xtgee Y X, family( ) link( ) corr( ) is equal to xt... Y X, pa corr( ) , for example xtgee Y X, f(b) l(logit) c(exc) will give exactly the same results as xtlogit Y X, pa c(exc)
So I assume GEE model can be explained as "population average" , but what is the different between xtlogit Y X, pa (which is equal to GEE) and xtlogit Y X, fe
I know the difference between fixed effects and random effects, but if I am not intrestest at individual level variances, how to choose the model ?
Will you use GEE which is same as xtlogit Y X, pa or will you use xtlogit Y X, fe ? This two commands will give different results, so I am confused, Thanks !

[Carlo Lazzaro]

Kenny:
welcome to this forum.
You tried to compare two different models: hence, no wonder that they giveback different coefficients.
See -xtlogit- entry, page 299, Stata .pdf manual, release #16.

[Kenny Don]

Thank you so much! Carlo, I read the Stata.pdf manual, and have some further questions. Since the fixed effects model is equal to conditional regression model, we treat the cluster as "matched" data , the problem is you will need at least one positive outcome in each cluster (for binary outcomes regression) , otherwise Stata will drop out all the clusters which had constant Y ( all Ys within cluster are 0 or 1) , if the clusters in the dataset has many equal Ys clusters, the xtlogit Y X, fe will give much different result compared with xtlogit Y X, pa, because population average use all the clusters information in the dataset whereas the fixed effects model only used the clusters which had Y = 1 and Y = 0
So I want to ask:
1. Is the fixed effect model in Stata only considered for matched analysis? Assume individual within clusters are very similar
2. If I have unbalance dataset, which has only small amount of the clusters , GEE is more appropriate, is this right ? What does the population average mean, does it mean you only use the the average X and average Y build the regression model?
Thanks a lot