Enhanced Clinical Trial Design of a Proof-of-Concept Study via Bayesian simulation analyses
Patanjali Ravva (1), Jonathan French (1), Sarah Dubrava (2) and Hélène M Faessel (3)
(1) Global Pharmacometrics, Pfizer Inc, USA; (2) Statistics-Primary Care, Pfizer Inc, USA; (3) Clinical Pharmacology-Primary Care, Pfizer Inc, USA
Background: A proof-of-concept (POC) study was proposed to assess the efficacy and tolerability of 3 doses of a MR tablet versus placebo (P). Each MR dose (2 higher, 1 equivalent to IR) was also to be compared to the commercial IR tablet included as a reference. Based on the observed response rates from historical IR clinical trials the study was powered on an odds ratio (OR) of 3.77 (15% P, 40% IR) for standard pairwise comparisons. An estimated equal allocation of 72 subjects per arm would allow comparing each MR dose to P with 90% power and 5% type I error. However, the power to detect a difference between MR and IR was 19%. Thus a clinical trial simulation study was undertaken to compare how well different designs of approximately equal cost could meet the POC objectives.
Methods: A Bayesian logistic regression model was used with priors on the IR and log OR responses. A mild dose-response relationship was assumed over the MR dose range. A beta prior distribution was used for IR; the shape parameters were derived from historical P-controlled clinical trials and informativeness of the priors. A normal prior distribution was used for the log OR and MR response rates. Simulation was used to assess the impact of reallocating subjects from the IR and P groups to MR treatments by borrowing information from historical data. Prior information was varied across scenarios. 10,000 trials were simulated for each design scenario and analyzed using WinBUGS 1.4.1.
logit[P(yi = 1| TRTi)] = α + β1*(TRTi = IR) + β2*(TRTi = MR1) + β3*(TRTi = MR2) + β4*(TRTi = MR3)
yi = 1 if the ith patient is a responder, otherwise yi = 0; TRTi = treatment group
Results: Use of prior information resulted in an increased power for the pairwise comparison of MR to IR while maintaining the power for the comparison to P. The power of a sample size of 40 subjects for IR and P with 75 subjects for each MR dose was 63% versus 19% for equal allocation. The weight of the prior can be viewed as 35 additional subjects worth of information in the IR and P treatment arms. Trial performance metrics for the selected design was also explored using model-based analysis and the understanding of the IR exposure-response relationship.
Conclusions: An efficient POC design was achieved through simulation analysis. Formally using historical information in a Bayesian analysis allowed additional patients to be allocated to MR groups and increased the power to establish POC.