RE: likelihood model for missing SEX data

From: Jakob Ribbing jakob.ribbing@farmbio.uu.se Subject: RE: [NMusers] likelihood model for missing SEX data Date: Wed, 5 Jul 2006 11:11:39 +0200 Nick, Anthe, I could be wrong, but would not the suggested approach (estimating the probability of SEX=0 at the same time as estimating the effect of SEX on CL) cause an upward bias in the magnitude of the estimated-SEX effect? The individual probabilities [1] would have to be distorted so that the subjects with missing SEX do not change the estimated SEX effect (when the imputation is based only on the DV). More importantly, if there are other covariates available that could predict SEX (e.g. WT), these should be used. The latter could change (improve) the estimated SEX effect. Unless the effect of SEX is substantial, the imputation could be based only on the other covariates since DV does not add much information. Imputing best-guess SEX based only on other covariates is done without the random distortion and before estimating any covariate model in NONMEM. If best-guess SEX is not predicted well by the other covariates this approach will lead to an estimated sex-effect which is biased low (just as when imputing with the typical value), but I guess the expected SEX could be used if SEX is treated as a continuous covariate and using a linear covariate model? Best regards, Jakob 1. Carlsson, Kristin C., Radojka M. Savic, Andrew Hooker, Mats O. Karlsson. Mixture models in NONMEM - how to find the individual probability of belonging to a specific mixture, and why this can be useful information. PAGE 15 (2006) Abstr 956 [www.page-meeting.org/?abstract=956]