Methodological issues in the development of a pharmacogenomic algorithm for warfarin dosing: comparison of two regression approaches.

Abstract

To ascertain whether multiple polynomial regression (MPR) has any advantage over multiple linear regression (MLR) in developing pharmacogenomic algorithms. Two pharmacogenomic algorithms were developed based on MPR and MLR models from a warfarin pharmacogenomic data set (derivation cohort [n = 125] and validation cohort [n = 115]). The MPR model showed better correlation with therapeutic dose (r = 0.62 vs 0.52); better diagnostic utility in distinguishing the warfarin-sensitive and warfarin-resistant patients (area under the receiver operating characteristic curves: 0.89 vs 0.81); and lower rate of underestimation (13.9 vs 20%) compared with the MLR model. Rate of overestimation was higher in the MPR than the MLR (10 vs 6.7%) model. The MPR approach has advantages over the MLR approach in predicting accurate and safe dose.