Hillel Alpert <[email protected]> asks:
> The first question is on fitting an xtmixed model.
> While fitting models on unbalanced data using xtmixed, I encounter the
> following comment messages:
> "Hessian becomes unstable or assymetric"
> "Not concave"
> How can these be addressed; for example which types of variables or
> parameter would be best to drop?
Have a look at the -emonly- and -emiter()- options to -xtmixed-, which select
optimization via EM (only) and set the number of EM iterations. Most of the
time, if you set the EM iterations large enough you can get a feel for which
are the offending variance components (i.e., those that really want to be zero
and should be taken out of the model).
EM maximization does not use Hessians, hence it doesn't suffer from problems
of lack of concavity. The flip side is that it can be very slow to converge,
not guaranteed to converge at all, and you don't get standard errors. As a
result of this, by default EM is relegated to starting-value duty in
-xtmixed-.
> Would selecting out cases based on rare values of a variable (for example
> the first year in which there were few observations) be useful?
Perhaps, but try the above first.
> The second question is on testing individual effects in the model:
> How can we test whether a time-trend observed for the panel overall is seen
> across individual subjects on the panel and how can we differentiate for
> which ones the time-effect is seen and for wihich it is not?
> I have seen an example using the post-estimation "test" command to compare
> the model with time as random effect in the nesting of individuals with the
> model without this effect. The example shows time included as a fixed effect
> as well, although I wonder why it is included in both places.
> xtmixed hamd week endog || id: week
The overall test of trend-within-panel can be done by using -lrtest- to
compare the above with
. xtmixed hamd week endog || id:
which does not have a trend at the panel level. The reason the former syntax
has week in both places is that in the fixed-effects specification (the
first), you have an overall, average slope due to week. In the random-effects
specification, -week- indicates that you want panel-specific random deviations
from that overall average slope. If you didn't have -week- in the
fixed-effects part, you would be assuming that the average slope over panels
is zero -- probably not what you want.
As for testing for trend effects panel by panel post-estimation, you can
currently get panel-specific slopes, otherwise known as BLUPs of random
effects -- see option -reffects- to -predict- after -xtmixed-. However,
inference would require the standard errors of these estimated effects,
something currently not supported in -xtmixed-. Needless to say, we are
working on that (among other things).
--Bobby
[email protected]
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
