Statistical Modelling for Adaptive Design of Clinical Trials: Optimality and Efficiency

与传统的临床试验方法相比，反应自适应设计（RAD）具有伦理优势，但它也在数据中引入了依赖性。大多数关于RAD的统计学文献集中在平均响应模型的统计方法上，而没有考虑群体异质性。忽略这种异质性可能会导致结论的偏倚。此外，暴露于不同环境因素的患者对治疗的反应可能非常不同。所有这些都是随机临床试验推荐的有效治疗方法并不总是在日常卫生保健中起作用的主要原因。本研究将开发新的RAD来倡导RAD的伦理优势，提出统计方法来解释RAD对各种人群异质性的依赖性，并为拟议的设计建立有效的推理方法。本研究将探讨自适应随机化对统计估计的影响，并利用其影响来提高治疗效果的学习效率。本研究考虑的群体异质性包括患者和机构特征及其与治疗的相互作用以及患者对一系列治疗的反应，本研究将建立具有协变量和相互作用，特别是定性相互作用的RAD的统计方法，以及用于动态治疗区域临床试验的RAD。研究结果将为适应性随机化条件下相关数据的估计以及异质性人群临床试验的最优设计提供统计学文献。使用RAD对定性交互进行自适应学习，可以提高检测亚群的效率。为聚类反应开发的统计方法将为多中心试验增加有效的伦理设计。将自适应随机化方法应用于动态治疗区域（DTR）临床试验中，可以实现对最优DTR的自适应学习，从而提高伦理效益和统计效率，解决了临床试验设计和分析中面临的方法学挑战。所开发的方法将人口异质性在RAD学习倡导的道德利益，提高统计能力。这项研究将增加临床试验的创新设计和有效的统计方法的统计文献。由此产生的工具箱可以推广到其他临床环境中，从业者希望有效地确定亚群的最佳治疗方法。

Response adaptive design (RAD) has ethical advantages over traditional methods for clinical trials, but it also introduces dependency into data. Most statistical literature on RAD focused on statistical methodologies for mean response models without considering population heterogeneity. Ignoring such heterogeneity could introduce bias into conclusions. In addition, patients exposed to different environmental factors may respond to treatments very differently. All these are the main reasons why the efficacious treatments recommended by randomized clinical trials do not always work in day-to-day health care.This proposed research will develop new RAD to advocate the ethical advantages of RAD, propose statistical methods to account for the dependency introduced by RAD for various population heterogeneities, and establish efficient inferential methods for the proposed designs. This research will investigate the influence of adaptive randomization on statistical estimation, and employ that influence to improve the efficiency in learning on treatment effect. The population heterogeneities considered in this research include patients' and institutional characteristics and their interactions with treatments as well as patients' responses to a series of treatments.This research will establish statistical methods for RAD with covariates and interaction, especially qualitative interaction, as well as for RAD for dynamic treatment region clinical trials. The results will add to the statistical literature on estimation for dependent data under adaptive randomization as well as optimal designs of clinical trials for heterogeneous populations. The use of RAD to adaptively learn on qualitative interaction may improve the efficiency of detecting subpopulations. Statistical methods developed for clustered responses will add efficient ethical designs for multi-center trials. The integration of adaptive randomization into clinical trials of dynamic treatment region (DTR) will enable adaptive learning on the optimal DTR thus improving the ethical benefit and statistical efficiency.This proposed research addresses the methodological challenges in design and analysis of clinical trials when facing population heterogeneities. The developed methodologies incorporate population heterogeneities in RAD learning to advocate the ethical benefits and to improve statistical power. This research will add innovative designs of clinical trials and efficient statistical methods to statistical literature. The resulting toolbox can be generalized to other clinical settings where practitioners hope to efficiently identify optimal treatments for subpopulations.