Re: st: logistic fixed-effect models with Stata and SAS

[Simon Falck <sfalckstata@gmail.com>]

On a prior question related to spatial econometrics and spatial panels, 
I made the suggestion to take a look at the work of Paul Elhorst, who 
developed several useful applications that may be useful to you.

http://www.stata.com/statalist/archive/2013-01/msg01385.html

You may want to look into the possibilities of using Matlab, since it is 
more flexible than Stata. Several of Elhorst applications are best 
implemented using Matlab.

Simon


On 2013-08-01 16:28, Yigit Aydede wrote:
> Hi,
>
> My research is to understand the moving decisions from one region to another on a dataset that has around 800K people in 285 regions that covers the entire country.  The number of people in each cluster (region) is ranging from 3000 to 120000.  The dependent variable is 1 if the person has moved, 0 otherwise.  I have a bunch of variables that control individual characteristics. The success ratio (for 1s) is around 3 percent.  So complementary log-log has a better fit.
>
> But, I have 2 major problems: (1) correcting the spatial correlation within each region, (2) removing unobserved regional fixed effects. So Xtlogit,fe or clog (conditional logit) with clustered SEs are the best options.
>
> Stata is not able to handle both of these for a big dataset.  There is a long discussion about this on Statalist in 2012 so that at the end they recommend SAS for this situation.  I also contacted Stata Tech Support, they have confirmed that Stata cannot handle xtlogit, fe or clogit on my dataset.
>
> I turned to SAS (logistic regression with Strata), but it's not able handle it either. It gives "insufficient memory" as an error message.
>
> I'll start looking for different methods such as creating a random and smaller subsample.  I've started reading the literature on this but I need a kind of head start.  I appreciate if you would suggest possible options on my research.
>
> I've run cloglog on it with regional dummies and clustered SEs, but the literature is very clear on this: regional dummies to control fixed effects in nonlinear models give biased results.  But, I am not clear if this "incidental parameters problem" in logistic regression (with dummies) asymptotically disappears (or gets smaller) if T (in my case people) goes to infinity, as I have very large T.  Any hint on this is also much appreciated.
>
> Thanks  a lot for your time and help
> --Yigit
>    
>      
> Yigit Aydede
> Department of Economics
> Saint Mary's University
> Halifax, NS, B3H 3C3
> Canada
>
> T: (902) 420-5673
> F; (902) 420-5129
>
>
>
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/faqs/resources/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/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/