A Simulation Study of Bioequivalence Test in a Two-stage Cross-over Adaptive Design

• Main Authors: Huang, Yi-Chun, 黃意珺
• Other Authors: Hsueh, Huey-Ming
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/w9z83j

碩士 === 國立政治大學 === 統計學系 === 107 === The two-stage adaptive design is developed to increase the efficiency of clinical trials during drug development. Not only that we may obtain a statistical conclusion in the interim analysis and have an early stop of the trial at the first stage, but also that we can have a more adequate sample size re-estimation for the second stage of the experiment. On the other hand, to minimize the data variation, we consider the crossover design of two periods and two sequences in every stage. This research mainly focuses on testing the bioequivalence of a test generic drug to a reference brand-name drug. We consider the two one-sided t-tests (TOST) (Phillips, 1990) to compare the means of the maximum plasma concentration of a patients taking the two different drugs. Instead of integrating a bivariate non-central t-distribution for power calculation (Maurer et al., 2018), we use an R-package to simulate sample data from normal population distributions to calculate the power of the TOST in a cross-over design. The sample size can be determined subsequently. Through a simulation study, we find that the type one error rate of the TOST in the adaptive cross-over design is acceptable mostly. However, the power of the test noticeably exceeds the required level, the sample size estimation tends toward conservatism. Hence, there is a room of improvement for this adaptive design. Besides the data of a continuous end-point, we also investigate the applicability of the TOST in an adaptive cross-over design with the data of a binary end-point. Similarly, we simulate numerous samples of data from Bernoulli population distributions to evaluate the power of the TOST. According to an intensive simulation study, we find that the application of such design with respect to binary data may produce an inflated type I error rate and an insufficient power. Furthermore, the first-stage-stopping-rate is not good as the data of continuous end-point.