2001
   Basel, Switzerland

Forward selection and other recipes for disaster

J.G.Wright

Eli Lilly & Co, Lilly Research Centre, Erl Wood Manor, Sunninghill road, Windlesham, Surrey, UK

Many population analyses expend a great deal of time and energy on identifying and quantifying covariate effects. Unfortunately, it is unusual for a covariate to be recorded because there is a clear rationale that it will affect drug disposition. As such the links between many investigated covariates and pharmacokinetic parameters are tenuous. A failure to consider whether there is clinically significant improvement in fit and the impact of multiple testing.have spurred the reporting of spurious or irrelevant covariate relationships. A covariate relationship is only important if it will affect a decision, either in the clinic or in drug development. These problems are amplified in widely applied subset selection procedures, when cascading multiple hypothesis tests are conditioned on each other. Given the limited data on which covariate investigations are conducted, prior knowledge must be used to guide the choice of statistical hypotheses. Indeed, even on very large datsets, as used in economics or weather forecasting, arbitrary stepwise selection procedures are dangerous. This talk will discuss the statistical limitations of covariate selection procedures, and make several suggestions for more rational approaches, including
1. the a priori declaration of physiologically plausible hypotheses
2. streamlining analysis and avoidance of multiple testing whenever possible
3. establishing criteria for pharmacodynamic significance

These issues are illustrated on a pharmacokinetic covariate selection problem that has been validated on separate data.