[R] Question re complex survey design and cure models

[Thomas Lumley]

On Wed, Jul 13, 2011 at 6:13 AM, Sze Liu <szeliu at hsph.harvard.edu> wrote:
> Hello all,
> I am using AddHealth data to fit a cure, aka split population model using nltm.  I am not sure how to account for the complex survey
> design - does anyone have any suggestions?  Any help would be greatly appreciated!


This will be not be easy, because the nltm functions don't accept any
sort of weights.    Two possible approaches:

1/ modify nltm to accept weights for the likelihood contributions.
The point estimates from nltm should then be correct, though the
standard errors will be wrong.   You can now use as.svrepdesign() in
the survey package to create replicate weights (jackknife or
bootstrap, essentially), and then use withReplicates() to apply the
replicate weights to nltm and compute standard errors.

2/ use svymle() in the survey package to do the estimation.  This
needs the loglikelihood and score functions for a single observation,
which you may be able to extract from the nltm code.


   -thomas
-- 
Thomas Lumley
Professor of Biostatistics
University of Auckland