Statistical Inference with Data-Dependent Treatment Allocation Rules
研究了在随机化优胜者规则这种自适应设计下，如何为二分类响应治疗差异构造精确和近似的置信区间，发现精确无条件方法在小到中等规模试验中表现更好，且分析时不能忽略设计本身。
Abstract In comparing two treatments with dichotomous responses, the randomized play-the-winner rule (Wei and Durham 1978) tends to assign more study subjects to the better treatment. For ethical reasons, this property is desirable for studies on human subjects. The randomized play-the-winner rule, which is a modification of Zelen's play-the-winner rule (Zelen 1969), is not deterministic and is less vulnerable to experimental bias than other adaptive designs. Recently, this design has been used in a trial to evaluate extracorporeal membrane oxygenation (ECMO) for treating newborns with respiratory failures at the University of Michigan. In this article, exact conditional, exact unconditional, and approximate confidence intervals for the treatment difference are studied from a frequentist point of view with the randomized play-the-winner rule. For small and moderate-sized trials, the exact unconditional procedures perform much better than the conditional ones because of the adaptive nature of the designs. Furthermore, we find that the design used for the trial should not be ignored in the analysis. The large-sample unconditional confidence intervals based on likelihood ratio statistics are not very sensitive to the design and perform well for moderate-sized trials. On the other hand, the intervals derived from the maximum likelihood estimates behave poorly under the adaptive design. All of the procedures are illustrated with the Michigan ECMO data.