Comparison Of Interim Analysis Methods Based On Information-adjusting In Group Sequential Designs With Small Sample Sizes

ObjectiveDuring the interim analysis of the group sequential trial with small sample sizes,the observed Fisher information may be taken to an extreme value,resulting in abnormal termination,which greatly limits the application of the group sequential design in the field of clinical trials.In this paper,the sequential design of small sample groups is used as the experimental model to study the abnormal termination of interim analysis,and on this basis,we improved the application of the stochastic curtailed procedure and Bayesian interim analysis in the trialMethodsTaking group sequential design as the research object,SAS 9.4 was used to randomly generate and analyze random simulation data at each stage of test.In the first part of this study,we explored the abnormal termination problem that would occur in the traditional group sequential test with small sample sizes and recorded two types of abnormal termination cases,and then proposed Fisher information-adjusting methods and evaluated their effect;In the second part,we conducted comprehensive evaluations of the stochastic curtailed procedure combined with the information-adjusting method by simulation study whose evaluation contents included type I error rate,power and average sample size.Considering the limitations of the stochastic curtailed procedure,we comprehensively compared the statistical characteristics of Bayesian interim analysis under different termination threshold settings in the third part,and summarized the practicability,pros and cons of these methods based on the results of simulation analysisResults(1)Comparison results of effect of the Fisher information-adjusting methods:Under the condition of different sample sizes,the probability-weighted method reduced the occurrence percentage of the two types of abnormal termination to 0%in the simulations.As the sample size decreased,the worst scenario of ratio-weighted method was that percentage of warning cases would increase to 4.8%.In negative simulation study,the percentages of warning case and error case after using inverse gamma prior IG(a=20,b)remained below 0.3%and 0%,while those using non-information prior reached 6.5%and 4.0%(2)Comparison results of statistical characteristics of auxiliary indicators in interim analysisIn regard to the stochastic curtailed procedure combined with the Fisher information-adjusting methods,the conditional power failed to control the type ? error rate,while the Bayesian predictive power with non-information prior could control them in an acceptable range.In the simulation trials with a required sample size of 10 for each group at each stage,when the total stage K was set to 5 and the SCP termination threshold ? was 0.90,the average number of stages and the average sample size were 2.796 and 55.92 respectively.When K was set to 4 and ? was 0.85,they were 2.217 and 44.34 respectively.In the Bayesian interim analysis,when the termination threshold ? was equal to 0.05,Handicap prior could control type I error rate around 0.05.When ? was nominal significance level at each stage,non-information prior could also control type I error rate within the allowable range of simulation error.In terms of power,Handicap prior was around 0.84,which was slightly higher than the no-information prior with nominal significance level as the threshold.In the five-stage,four-stage and three-stage group sequential tests,the average number of stages of the Bayesian interim method using Handicap prior was 3.086,2.554 and 1.982 respectively.ConclusionsThis study proposes the Fisher information-adjusting method for the first time to solve the problem of abnormal termination in the group sequential trials with small sample sizes.Under the condition of small sample sizes,we evaluated the statistical performance of stochastic curtailed procedure combined with information-adjusting method in all aspects,and then the applicability of Bayesian interim analysis with small sample sizes was explored.We recommended that predictive power with non-information prior and Bayesian interim analysis with handicap prior should be used as the auxiliary reference indicators for early decision-making in group sequential clinical trials with small sample sizes.These indicators can shorten the period of group sequential clinical trials and provide reliable conclusions at the interim stage.