Re: st: Two Way Fixed Effect on Unbalanced Panel Data

[David Hoaglin <dchoaglin@gmail.com>]

Hi, Ploy.

As a start, you could separate the effect aij into the "main effects"
for the two countries and the interaction (e.g., replace aij with ci +
cj + dij).

In the second version of your model, you may want to keep the "main
effect" for time (qt in the first version).

You could fit both models by ordinary regression, accompanied by
plotting and diagnosis, to see how the data behave.  (The number of
predictors, however, may be too large for some flavors of Stata.)

That regression approach does not take into account the correlation
structure.  If you regard the panel as consisting of pairs of
countries, a fixed-effects analysis with aij as the pair effect seems
all right.  Those fixed effects for the pairs would account for
structure in the data that does not change over time, and that would
include the country-specific effects ci and cj that I introduced
above.  At the moment, I don't see how to separate aij into ci + cj +
dij within a fixed-effects model, but I have not tried to search for
work on panels of pairs.

I hope this discussion helps.

David Hoaglin

2012/3/17 Ploy T. <ploy.tham@gmail.com>:
> Hi Statalist
>
>
> I have questions regarding two-way fixed effect on large unbalanced
> panel. Here is my model:
>
>
> ln Xijt = β0 + aij + qt + β1 lnGDPit + β2 lnGDPjt + β3 lnPOPit + β4
> lnPOPjt + β5  FTAijt + εijt
>
>
>
> My LHS variable is bilateral export from country i to country j. The
> explanatory variables are GDP and population of both countries, and a
> dummy variable that takes value 1 if both countries are member of an
> FTA. I would like to run the model for
> 1.) one way fixed country effect
> 2.) one way fixed time effect and
> 3.) two way fixed effect.
>
> I have unbalanced panel data for approximately 2,200 country pairs
> with 6 time periods. I'm a new STATA user. So, as far as I know, there
> is no problem on running unbalanced panel data on one-way fixed effect
> (correct me if I'm wromg). But there's no direct method on estimating
> two-way fixed effects on large unbalanced panel data. Is there any way
> or commands that help me get through such problem?
>
> Moreover, if I change my model such that:
>
>
> ln Xijt = β0 + aij +  β1 lnGDPit + β2 lnGDPjt + β3 lnPOPit + β4
> lnPOPjt + β5 FTAijt - lnPit - lnPjt+ εijt
>
>
>
> That is, country-fixed effect and country-and-time fixed effects
> (estimating aij together with lnPit and lnPjt) are estimated together.
> Is there any method to get estimation of such model?

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/