2019 年 40 巻 1 号 p. 15-34
Missing data is a common problem in longitudinal clinical trials, and mixed-effects models for repeated measures (MMRM) have been widely applied to circumvent the resulting bias effectively. However, many standard inference methods of MMRM lead to the inflation of type I error rates for the tests of regression coefficient parameters when the longitudinal dataset is small and incomplete. Permutation inference methods have been established as accurate inference methods under small sample settings. In this article, we propose two effective permutation-based inference methods for the analyses using MMRM. One is the permutation of the treatment assignment variable and the other is the permutation of weighted residuals estimated by the reduced model under null hypothesis. We conducted numerical evaluations via simulation studies under realistic situations to evaluate performances of the proposed methods. The two methods generally provided valid inference results and performed relatively well compared with the current standard methods, even for small and incomplete datasets. Applications to a postnatal depression clinical trial are also presented.