Statistical approaches to pharmacodynamic modeling: motivations, methods, and misperceptions.
We have attempted to outline the fundamental statistical aspects of pharmacodynamic modeling. Unexpected yet substantial variability in effect in a group of similarly treated patients is the key motivation for pharmacodynamic investigations. Pharmacokinetic and/or pharmacodynamic factors may influence this variability. Residual variability in effect that persists after accounting for drug exposure indicates that further statistical modeling with pharmacodynamic factors is warranted. Factors that significantly predict interpatient variability in effect may then be employed to individualize the drug dose. In this paper we have emphasized the need to understand the properties of the effect measure and explanatory variables in terms of scale, distribution, and statistical relationship. The assumptions that underlie many types of statistical models have been discussed. The role of residual analysis has been stressed as a useful method to verify assumptions. We have described transformations and alternative regression methods that are employed when these assumptions are found to be in violation. Sequential selection procedures for the construction of multivariate models have been presented. The importance of assessing model performance has been underscored, most notably in terms of bias and precision. In summary, pharmacodynamic analyses are now commonly performed and reported in the oncologic literature. The content and format of these analyses has been variable. The goals of such analyses are to identify and describe pharmacodynamic relationships and, in many cases, to propose a statistical model. However, the appropriateness and performance of the proposed model are often difficult to judge. Table 1 displays suggestions (in a checklist format) for structuring the presentation of pharmacodynamic analyses, which reflect the topics reviewed in this paper.