From BRUCE@uqpharmacy.pharmacy.uq.edu.au Fri Aug 25 00:52:16 1995
Subject: Residuals in NONMEM Hi NM-users,
Can anybody clear up the following for us?
We are assessing predictive performance of a "best" NONMEM model using POSTHOC Bayesian estimation in an independent group of subjects. Are the weighted residuals (WRES) produced by POSTHOC equivalent to the "Standardised Residuals" (SRES) used, e.g. by Sam Vozeh's group for this purpose (e.g. see J Pharmacokinet Biopharm 1990;18:161-173, and Clin Pharmacol Ther 1991;50:55-65)?
If they are then the mean of the WRES apparently can be t-tested against 0, and the standard deviation of the SRES can be t-tested against 1. Is our thinking on this correct? Thanks in anticipation of any advice.
Pharmacy Dept, Univ Qld,
From lewis Fri Aug 25 10:36:00 1995
Subject: validation Bruce raises a perrennial issue. I am not sure I agree with using residuals in the way that he contemplates and Vozeh suggests; I am sure that "validation" does not mean the same thing to everyone (or maybe any two people!).
In my (current and perhaps temporary) view validation is meaningless; predictive performance is meaningful. I like to examine how good the predictions of the model are in a validation (or other more clever sample - re-use) set. Do they capture the effect of the covariates present there as they did in the index set? Are they good enough for clinical purposes? etc?
Stu has a nice piece in the Bourne archive the last newsletter, no. 11, gives explicit directions how to access this) that discusses validation more broadly. I advise anybody thinking about the topic to read it.