You are not logged in. You can browse but not post. Login or Register by clicking 'Login or Register' at the top-right of this page. For more information on Statalist, see the FAQ.
Hi Amelie. It's always permissible to use a pooled MLE provided you cluster the standard errors for serial correlation. I assume you're using the default in xtmlogit, which is random effects estimation. Running xtmlogit for RE and mlogit both assume any unobserved heterogeneity is uncorrelated with the explanatory variables. If you want a kind of "fixed effects" analysis using mlogit, add the time averages of the time-varying x(i,t) as extra controls. This is the Mundlak device; in the linear case, it would give you the usual FE estimator. You can test the time averages to see if there is correlation between heterogeneity and x(i,t). In all cases, use vce(cluster id).
Hi Jeff, Thank you so much,for the quick reply. I have 20 years pooled together in a panel. I have a solid model specification with all possible observable variables that the theory dictates and I use period effects and vce (cluster id). I understand from your answer it is OK to use "mlogit." I do not need to control for fe or re, So xtmlogit would be over the top. Sorry I missed you at the SEA last week. I am still in NOLA on a scientific retreat. Hope all is well.
Thank you Carlo for your suggestion. I have about 3 million observations (person years) and it is not easy to run an xtmlogit and play by adding and omitting variables. Each run takes a few hours. After Jeff's answer, I feel that I do not need to use xtmlogit and control for re.
SEA was the place to be, and I wasn't there! I couldn't commit to another trip this semester -- but I should've made SEA a priority. Enjoy your remaining time in NOLA.
Comment