Statistical design of personalized medicine interventions: The Clarification of Optimal Anticoagulation through Genetics (COAG) trial

Background: There is currently much interest in pharmacogenetics: determining variation in genes that regulate drug effects, with a particular emphasis on improving drug safety and efficacy. The ability to determine such variation motivates the application of personalized drug therapies that utilize a patient's genetic makeup to determine a safe and effective drug at the correct dose. To ascertain whether a genotype-guided drug therapy improves patient care, a personalized medicine intervention may be evaluated within the framework of a randomized controlled trial. The statistical design of this type of personalized medicine intervention requires special considerations: the distribution of relevant allelic variants in the study population; and whether the pharmacogenetic intervention is equally effective across subpopulations defined by allelic variants. Methods: The statistical design of the Clarification of Optimal Anticoagulation through Genetics (COAG) trial serves as an illustrative example of a personalized medicine intervention that uses each subject's genotype information. The COAG trial is a multicenter, double blind, randomized clinical trial that will compare two approaches to initiation of warfarin therapy: genotype-guided dosing, the initiation of warfarin therapy based on algorithms using clinical information and genotypes for polymorphisms in CYP2C9 and VKORC1; and clinical-guided dosing, the initiation of warfarin therapy based on algorithms using only clinical information. Results: We determine an absolute minimum detectable difference of 5.49% based on an assumed 60% population prevalence of zero or multiple genetic variants in either CYP2C9 or VKORC1 and an assumed 15% relative effectiveness of genotype-guided warfarin initiation for those with zero or multiple genetic variants. Thus we calculate a sample size of 1238 to achieve a power level of 80% for the primary outcome. We show that reasonable departures from these assumptions may decrease statistical power to 65%. Conclusions: In a personalized medicine intervention, the minimum detectable difference used in sample size calculations is not a known quantity, but rather an unknown quantity that depends on the genetic makeup of the subjects enrolled. Given the possible sensitivity of sample size and power calculations to these key assumptions, we recommend that they be monitored during the conduct of a personalized medicine intervention.