RE: Covariate screening

From: "Sale, Mark" <ms93267@GlaxoWellcome.com> Subject: RE: Covariate screening Date: Thu, 31 May 2001 09:45:58 -0400 All possible covariates? Stuart, Nancy Sambol and Janet Wades paper (J Pharmacol Biopharm, 22(2) 1994 - Interactions between structural, statistical and covariate models in population pk analysis) tells us that we cannot examine covariates in isolation, since interactions can occur. So we would need to look at all combinations of covariates. Lets see, if we have the basic covariates (age, wt, gender, race) on each of 4 basic parameters (V, CL, Ka, Lag), using each of two models (linear and exponential) in addition to no relationship we have: 3^4^4 = 43,046,721 possible models. Clearly we need some sort of screening. We recently had a poster at ASCPT looking at whether a machine learning method (genetic algorithm) can find the "best" among a finite set of candidate models (structural, error and covariate models), comparing it to exhaustive search. Basically it worked pretty well. At some point, we'll compare this "artificial intelligence" method prospectively to "natural intelligence" methods, assuming we can find some around here. Mark