Bootstrap validation of pharmacodynamic models defined via stepwise linear regression.

Abstract

When a pharmacodynamic model is to be considered as the basis for individualized drug dosing, validation of the model is clearly warranted. Rigorous validation is problematic when the training data set to be modeled has too few data points and no independent test data set exists. A simulation method known as the bootstrap lends itself particularly well to this dilemma. Bootstrap sampling allows simulation of needed test data sets that mimic the initial data set. Model validation is then undertaken by repeating the model formulation procedure on the bootstrap samples. For illustration, a pharmacodynamic model for leukopenia was constructed by stepwise linear regression from data of 41 patients with cancer treated with the drug amonafide. Stepwise regression analyses were then repeated for 100 bootstrap samples, which verified the initial selection of covariates for the model. Next the regression parameters and residual error standard deviation of the model were repeatedly estimated for 200 additional bootstrap samples. The bootstrap results confirmed the initial formulation of the pharmacodynamic model from the training data set.