-gllamm- won't do survival, unless you are willing to treat
-link(log)- as everything you would ever need there. That won't
produce Weibull curve though.
With anything multilevel, you need to think where your random effects
will go. The linear predictor? Constant only, or random slopes as
well? The shape parameter of Weibull (OMG... I am not at all sure this
will work)?
Besides, there won't be a single survival curve anymore. If each
higher order unit contributes an unobservable random effect, you will
have a fuzzy curve: some sort of central tendency curve with random
effect set to zero, and some sort of a confidence band around it based
on the variability of the random effect. You can probably get a little
closer to the best fitting curve using predictions of the random
effects, such as empirical Bayes estimates.
I'd probably think of a hierarchical Bayesian model there. Not that I
am a hugely Bayesian person, but that's one of those cases where
Bayesian machinery works its MCMC magic while frequentists have to
write more and more complicated (multivariate) integration schemes. Of
course in terms of computational time they are likely to be the same
-- the four level chain will take eternity to converge, just as
-gllamm- won't be in much hurry over this kind of a data set. But at
least it terms of setting things up, you just write a bunch of
samplers and have them run on any available cluster :)).
On 1/19/09, Maarten buis <[email protected]> wrote:
> --- John Stephenson <[email protected]> wrote:
> > My understanding is that the random effects in a multilevel model are
> > normally distributed with zero mean so the parameter best estimates
> > will not change whether or not the random effects are included.
>
>
> Unfortunately, these types of arguments don't work well in non-linear
> models. See for example chapter 12 of (Agresti 2002) or chapter 13 of
> (Fitzmaurice, Laird, Ware 2004).
>
>
> > However, if I do need to account for the random effects, can such an
> > analysis be performed in Stata?
>
>
> You can estimate a one level random effects survival analysis model
> using -streg- with the -shared()- option, for more than one level you
> will have to move to -gllamm-, see: www.gllam.org .
>
> Hope this helps,
> Maarten
>
> Agresti, Alan (2002) Categorical Data Analysis, 2nd edition. Hoboken,
> NJ: Wiley.
>
> Fitzmaurice, Garrett M., Nan M. Laird, James H. Ware (2004) Applied
> Longitudinal Analysis. Hoboken, NJ: Wiley.
>
--
Stas Kolenikov, also found at http://stas.kolenikov.name
Small print: I use this email account for mailing lists only.
*
* 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/
