Hi Maarten,
thanks for your input, but I'm not sure that this is the problem. The
year effects are identical for everyone, while my variables of
interest (assume a dummy for the sake of simplicity) switch in
different years for different people. In other words, the model is
y_it = 1{a_i + g_t + tau*d_it + u_it > 0}, where a_i is the individual
fixed effect, g_t is the common time effect and d_it is the variable
of interest. What you have in mind sound to me like a model of the
sort
y_it = 1{a_i* g_t + tau*d_it + u_it > 0}, that is a model with
individual specific time effects which would indeed exhaust all
available information.
What could be the case though is that I end with too few individuals
for whom the dependent variable and the variable of interest change
jointy in each year. In fact, I just tried to use two-year intervals
annd now everything runs fine and looks reasonable (at least at a
first glance).
Thanks again and best wishes,
Nils
On Tue, Jan 5, 2010 at 2:35 PM, Maarten buis <[email protected]> wrote:
> --- On Tue, 5/1/10, Nils Braakmann wrote:
>> I have a large panel of roughly 50,000 observations over 20
>> years. I model the effects of a time-varying variable on individual
>> employment probabilities using -clogit- with the individuals as
>> groups. The trouble starts when I add year dummies, which leads to
>> severe convergence problems (nonconcave likelihood for several
>> hundred iterations, etc.).
>
> So you have data on individuals observed for multiple years.
> With a fixed effects model you have taken out all the information
> you might obtain from comparing individuals. This is a good thing
> in the sense that by only comparing individuals with themselves
> you are more likely to compare like with like, but it is also a
> bad thing as you are throwing away information. If your data is
> collected annually, and you added year dummies, then it seems to
> me that the fixed effects in combination with the year dummies
> will have exhausted all the information that is present in your
> data, and that the effect of any additional variables are thus not identified, and Stata will show that by not converging.
>
> Hope this helps,
> Maarten
>
> --------------------------
> Maarten L. Buis
> Institut fuer Soziologie
> Universitaet Tuebingen
> Wilhelmstrasse 36
> 72074 Tuebingen
> Germany
>
> http://www.maartenbuis.nl
> --------------------------
>
>
>
>
> *
> * 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/
>
*
* 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/
