Pyy-Martikainen Marjo <[email protected]> asks:
> I am trying to estimate a Cox model with shared frailty but the estimation
> seems to take a lot of time: the job started yesterday afternoon and is
> still going on. I am using a data set with appr. 6000 spells of
> unemployment belonging to appr. 1500 individuals. In the model I am trying
> to estimate, spells by the same individual share the same frailty. I have
> only one explanatory variable (a dummy for gender) and I use Efron's method
> for handling ties. By restricting the data set to 700 spells the estimation
> takes about an hour. But I would like to use all my spells...
> Has anybody any experiences about the speed of computation when estimating
> this model -is it always this slow and are there any ways to speed up the
> estimation?
Cox models with shared frailty are fitted via penalized likelihood, where
the frailty variance is estimated via repeated Cox model fits where the
group-level frailties are directly estimated.
Thus, with 1500 groups, this amounts to running repeated Cox regressions
where you are estimating 1500 frailties plus the number of other covariates
in your model.
Stata's current implementation calculates the partial likelihood exactly,
which in your case involves the inversion of 1500 x 1500 matrices, hence
the slowness. There exist in the literature faster sparse matrix
approximations to these likelihoods, and their implementation in Stata is
something we are looking into.
--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/
