RE: An approach for imputing missing independent variable (covariate)

From: "Piotrovskij, Vladimir [JanBe]" <VPIOTROV@janbe.jnj.com> Subject: RE: An approach for imputing missing independent variable (covariate) Date: Fri, 22 Sep 2000 15:11:58 +0200 There is no doubts multiple imputation is superior. However, in real life population pharmacokineticists doing data analysis do not use it. They have a dilemma: either to exclude records with missing covariates (usually just a few) or to perform a reasonable imputation. Most often means or medians of available values are used that is clearly not the best solution as it introduces a bias into the final model parameter estimates. My approach was aimed to reduce that bias by using information contained in DV. However, the risk of a bias still remains because the final estimates are affected by imputed values which are indeed considered as true covariates. What I would suggest as a palliative: 1. Develop a model using available covariates only (some information is lost, but no bias, and SEs are OK) 2. Generate estimates for missing covariates by inverting regression equations as I siggested, but DO NOT perform interations. 3. Do posthoc step only to obtain conditional estimates for individuals with missing covariates. Vladimir