Influence of clinical trial design to detect drug effect in systems with within subject variability
Chenhui Deng, Elodie L. Plan and Mats O. Karlsson
Pharmacometrics Research Group, Department of Pharmaceutical Biosciences, Uppsala University, Uppsala, Sweden
Objectives: In many clinical development programs, identification of drug effect is attempted through hypothesis test based on data from a cross-over (XO) or, more commonly, a parallel (PA) group trial. One question relating to the absolute and relative merits of the two trial types concerns the capability of detecting drug effects in systems with within-subject variability (WSV). In a previous study, systems with WSV were analyzed using dynamic parameter IOV (dIOV) or stochastic differential equations (SDE) . The aim of this study is to investigate type I error calibration and power to detect drug effects in PA and XO trials with WSV.
Methods: Stochastic simulations and estimations (SSE) were applied with a model including an exponential decay placebo effect and a linear dose effect. A dIOV was introduced on the maximum placebo effect for simulations. Models with or without drug effect, and with or without dIOV, SDE for system noise (SDEs) and SDE for parameter variability (SDEp), were fitted to the simulated datasets. Randomization tests were performed to investigate the type I error and power was determined following type I error control. A novel mixture model approach was proposed to investigate the drug effect in XO studies.
Results: Uncalibrated type I errors were high for XO, but not PA, trials. The power to detect existing drug effects was much higher for XO than PA trials even after type I error calibration. Among models with or without WSV misspecification, similar results were obtained for PA trials, whereas they varied for XO trials, with increased power when WSV was more adequately described (true model>SDEs>SDEp>no WSV). The randomization test-derived test statistics for decrease in OFV were stable and close to the χ²-distribution value in PA trials, but dissimilar and large among explored structural models with different drug effects in XO trials. When applying the proposed mixture model approach in XO trials, cutoffs, similar for different drug effects among different models, were considerably lower than with the above method and in agreement with χ²-distribution.
Conclusions: XO trials are more powerful than PA trials; however assessment of drug effects in XO trials is made more complex in the presence of WSV. We have presented alternatives to handle this in terms of (i) extended models for WSV, (ii) randomization test for XO trials, and (iii) a mixture model approach for appropriately contrasting models with or without drug effects.
 PAGE 23 (2014) Abstr 3194 [www.page-meeting.org/?abstract=3194] Acknowledgement: This work was supported by the DDMoRe project and FP7-HEALTH-2013-602552.