Study On Method And Application Of Bayesian Meta Analytical Predictive Prior Borrowing Historical Information In Clinical Trials With Binary Outcomes

In clinical trials,especially for small-scale trials,borrowing of historical information is common to help the design and analysis of a new trial,so as to save sample size effectively,reduce the duration and the cost of clinical trials,improve efficiency of the current trial,while taking into account ethical issues.Among the numerous prior approaches under the Bayesian framework,which is the best to make use of the historical control data should compare their operating characteristics.Based on Bayesian hierarchical model,this paper selects several methods for incorporating the historical data,including meta analytical predictive prior,robust meta analytical predictive prior,modified power prior of full Bayesian method and empirical power prior.To modified power priors,if the power parameters are independent,it’s written as MMPP Ind,or not,then DMPP.Under different degrees of data heterogeneity,we investigate the operating characteristics of each approach on the binary endpoints.In order to better comparison,the method that only using the current data for Bayesian analysis is also chosen.In the third chapter,we embark on a case study to illustrate the construction of prior distribution and posterior distribution of the chosen approaches and compare their operating characteristics,which includes five studies to access the efficacy of extreme therapy in the relapsed or metastatic head and neck squamous cell carcinoma.In the fourth chapter,three simulation scenarios are set to compare the power and type I error rate of the methods.The simulation shows the following results.Firstly,MMPP Ind and DMPP lead to type I error rate got inflated,when the proportion of heterogeneous historical trials is 40% or 50%and the treatment effect of heterogeneous historical research is 0.4 or 0.7.Secondly,EPP method can lead to an increase in type I error rate compared with only using the current data for Bayesian analysis,and the type I error rate of EPP is the highest among all methods except the situation that the type I error got inflated.Thirdly,in most cases,MMPP Ind and DMPP with a smaller correlation coefficient between power parameters which are barely noticeable difference has the higher power than the other approaches.While the few cases there is heterogeneity between historical data and current data,EPP has the highest test efficiency.Fourthly,in general,borrowing historical control data can improve the test power,while the power may be lower when the historical data and the current data are incompatible.Fifthly,corresponding to 0.6,0.2,or 0.9,the power and type I error of each approach under the situations of the smaller treatment effect of heterogeneous historical study 0.4,0.1,or0.7 are higher.Finally,MAP usually has higher power than RMAP,but the power of RMAP will be higher and it shows better robustness under the prior-data conflict.In these approaches making use of the historical control data,if importance is attached to improve the power,MMPP Ind,DMPP with a smaller correlation coefficient between power parameters and EPP are preferable to meta analytical predictive priors.However,if attached to control the type I error,MAP or RMAP is the preference.