On 2004-08-19, at 16.53, Stanislav Kolenikov wrote:
If you have a
dataset of say 10000 individuals, and you also want to take some
sample design clustering into account... you are doomed to wait for a
few weeks for the model to converge.
There are however techniques to speed up -gllamm- that can be extremely
efficient with large datsets, if the number of response patterns are
limited. Lets say you have 10.000 observations but only 5*3*10 possible
response patterns; if you use -contract , freq(_freq)- you end up with
only 150 observations with varying frequency weights. This dataset
should be estimated with -gllamm , weight(_freq)- in about the same
amount of time as a dataset with only 150 individuals.
unless you have the latest Cray at your disposal, the model should be
kept to a moderate size. [ .... ] (It took a few days with
-oprobit- link and the panel structure with just one random effect on
my computer.)
That is so true. It just happened that two days ago I got a paper
accepted for publication that summarise 323 hours (!) of computing time
(Mac G4 800Mhz). A total of 10 models of varying complexity and
dependent variables were estimated, 3 binary (logit) and 7 ordinal
(ologit). The data was longitudinal with 60 repeated measures of 17
subjects - 1020 obs. With only two random effects I expect the models
to converge in about 1/10 of the time. With 4 random effects I expect
one model to converge in 200-300 hours and with 5+ random effects I
probably need a Cray (or maybe a G5).
Michael
*
* 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/
