Dear Statalisters
I am not quite sure as to direction here; any advice would be most welcome.
I have a multi-record per patient survival data set with 28 day (from acute
diagnosis) mortality as the outcome. A Cox model with (significant)
time-varying covariates gives a "good" fit , by conventional means (residual
analysis etc) .
A log normal AFT model (parameterized in the time-ratio sense) seems to do a
"good" job as well (again, by conventional diagnostics). The shape of the
baseline Cox model hazard (using stkerhaz, recently posted) certainly has a
log-normal profile.
Apart from "intrinsic" scientific sense of what drives the model, are there
any "objective" criteria to select the better model; the following have
occurred to me
visual inspection of residual plots has a certain subjective element
plotting survival curves from the 2 models against K-M estimates to give an
"observed-predicted" comparison
AIC / BIC comparisons are OK for comparing parametric models, but seems not
to be appropriate for the Cox, in a cross-comparison sense, where there is a
partial-likelihood. This begs the question of the status of the likelihood
in the Cox model, as per page 126-128 in the green-book (An introduction to
survival analysis using Stata); is it possible to recover the "full"
likelihood or is it OK to use the partial likelihood?
This problem arose in the context of a certain bias in the medical
literature towards the Cox model.
Thanks in advance for any contribution on this matter
john moran
John L Moran
Intensive Care Unit
The Queen Elizabeth Hospital
28 Woodville Road Woodville
SA 5011
Australia
Tel 61 08 8222 6463
Fax 61 08 8222 6045
E-mail: [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/
