Dear all,
this is more of a general econometrics question. On several occasions
(looking at different questions using different econometric
specifications) I have encountered a strange(?) problem when using
-clogit- (-xtlogit, fe-) to estimate a fixed effects logit with time
dummies: The model converges easily and quickly without time dummies,
but the log likehood becomes non-concave when adding the time dummies.
When using a linear probability model with time dummies (estimated by
-xtreg,fe-) all estimates appear normal in the sense that point
estimates are reasonable, there are no dropped time dummies, missing
standard errors or something similar. Estimates are also qualitatively
similar to the conditional logit estimates without time dummies.
Sample sizes differ somewhat between the various occasions where I
encountered this problem but are usually around 100,000
person-year-observations from roughly 10,000 individuals (unbalanced
panel, excluding individuals that were dropped due to constant
outcomes, individuals are observed for at most 22 periods). Further
control variables include some dummies and some continous variables
(exact number depending on the specific model, but usually not more
than 20, some of the dummies are "1" for very few cases, e.g. 80). All
of these appear to have sufficient within-individual variation.
On of the occasions I experimented with a cubic time trend instead of
the time dummies. Here, the linear and the quadratic term were dropped
due to multicollinearity. While this seems to be root of my problem, I
am still not sure why this multicollinearity does not cause problems
in the linear probability models. Additionally, I am not sure how to
find out what causes this multicollinearity.
I would greatly appreciate any comments. In particular, am I missing
something (theoretically) obvious here (like some famous result that
the conditional logit model is not identified with time dummies...)?
Or is this problem data specific and I am just unlucky to encounter
this on different occasions in a row? Additionally, if anyone has a
good idea on how to find the problematic variables, I would love to
hear it.
Best regards,
Nils
*
* 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/