st: Re Lilian tesman- Predict mortality

[Kay Walker <kay.walker@internode.on.net>]

Have you created your model for finding predictors of death from 
variables collected from ONLY the ones who have died, or from others? I 
would make the model from the ones who died and look at the order of the 
strengths of the predictors from best downwards, adding them in or out 
Step wise. I would also look at different types of regression eg. 
various polynomials as there are conditions where a "symptom" can become 
worse, or actually seem to lessen before death- so you get a 
curved/wavelike sequence of measurements on that variable- the human 
manifestations of variables often don't work the same way that numbers 
do. At this stage you wouldn't be doing a logistic regression as you 
only have one end point- death. The timing or order of variables may be 
important in real life as well, which might force you to have early/late 
versions of variables, eg. temperature can be high at night, normal 
during the day; heart rate can be fast during the acute phase , then 
slow down in those becoming well, but rise again in those who are going 
to die.
After developing the best fit- or a selection of possibles, enter the 
measurements from ones who haven't died into the logistic model to see 
if there is a discernible pattern anything like the deceased ones.
Depending on your data you might be better off doing a discriminant 
analysis on the dead vs. live and getting predictors from the variables 
which separate the groups best.
I've only done this sort of modeling on diseases that are rare- like 
Duchennes Muscular Dystrophy- NOT on large population groups- so I might 
have given you a pile of garbage!.
*
*   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/