From: David L. Van Brunt, Ph.D. <dlvanbrunt_at_gmail.com>

Date: Thu 23 Jun 2005 - 00:18:52 EST

David L. Van Brunt, Ph.D.

mailto:dlvanbrunt@gmail.com

R-help@stat.math.ethz.ch mailing list

https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html Received on Thu Jun 23 00:24:15 2005

Date: Thu 23 Jun 2005 - 00:18:52 EST

With such a wide range of backgrounds here, I thought I'd toss this out here to get ideas.

I've lucked into some clinical trial data where schizophrenic patients were
randomly assigned to start on one of three drugs, then were followed

"naturalistically" over a year (or more, depending on when they enrolled).

They were assessed with a standard battery of instruments at baseline, and
at regular intervals. Apart from the initial random assignment, treatment
decisions were left up to patients and their doctors. Thus, one of the main
outcomes of the trial is the "all-cause time to discontinuation" which is
believed to be the patient-centric tipping point where the risks and costs
of using the drug outweight the benefit, so the patient either switches or
just stops. The outcome variable, therefore, is a censored time variable. I
have the number of days, and a flag to identify if that number is actual
discontinuation or just the end of observation.

Now, the a priori analysis is done and gone (Cox regression of which drug had longest time to d/c), and a more clinical question has come up: using these data, can we build a model to predict, for each patient, the drug which is likely to be best for that individual. Sounds like a bayesian opportunity to me, a separate model for each drug based on an analysis of the patients randomized to each drug, then apply each model to a holdout sample, and see if patients matched to therapy did better than patients who were not.

If I were just predicting a single continuous measure, or a dichotomous
outcome, I'd have no problem. The question is, given that the outcome is a
censored time variable, which approaches would get me closest to the

"posterior probability of success" or "expected number of days, give or take

Y" in the clinical sense, not the frequentist sense ("if you repeated this
study K times,...")?

Just interested in hearing some thoughts on this.

David L. Van Brunt, Ph.D.

mailto:dlvanbrunt@gmail.com

[[alternative HTML version deleted]]

R-help@stat.math.ethz.ch mailing list

https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html Received on Thu Jun 23 00:24:15 2005

*
This archive was generated by hypermail 2.1.8
: Fri 03 Mar 2006 - 03:32:56 EST
*