Random-effects Model Based Poisson Distribution Bayesian Meta-analysis

Objective Meta-analysis refers to the statistical analysis method which collects and integrates a series of related studies.In the clinical practice,it is common to see the cases that there were no direct comparison evidences or the need to select the best intervention measure for patients from multiple interventions,which becomes indirect comparison meta-analysis or multiple treatment comparisons(MTCs,also called network meta-analysis,NMA).Traditional meta-analysis could only compare two intervention groups,while NMA can compare synthetically multiple intervention measures.Currently,NMA can be implemented through frequency and Bayesian statistics.Whether traditional or Bayesian meta-analysis,the synthesis results obtained mainly are curative effect analysis.However,as to sparse data like adverse events,the research of its theoretical method is still inadequate.To fill in this gap,this paper aims to explore the suitable meta-analysis method for sparse data.Method This research takes four interventions(Ursodeoxycholic acid(UDCA),S-adenosylmethionine(SAMe),UDCA plus SAMe,placebo)for intrahepatic cholestasis of pregnancy(ICP)as the example.The content of this thesis mainly contains two parts.In the first section,we introduce meta-analysis of two treatment groups(comparison UDCA and SAMe for pruritus improvement)using command function “bayesmh” in Stata version 14.0 software,and then describe the two sets of program code in detail;in the next section,we show how to perform Bayesian NMA for sparse data using WinBUGS.We perform meta-analysis for each pregnancy outcome of ICP pregnancies(pruritus improvement,caesarean section,premature delivery,fetal distress and neonatal intensive care unit(NICU))employing non-informative prior distribution with Poisson likelihood function.After getting posterior probability distribution for each model parameters,we evaluate the models through deviance information criterion(DIC)and convergence diagnostics.Results Based on the inclusion and exclusion criteria,altogether thirteen eligible articles were included in the final meta-analysis.In the part of performing meta-analysis of two interventions using Stata software,similar results were obtained from the two sets of program code.The diagnostic plot of the model effect indicates that the overall model fit is relatively satisfactory.In the part of Bayesian NMA,there were no statistics significant differences between the interventions in five pregnancy outcomes.The fitting effects for models of pruritus improvement,caesarean section,premature delivery are perfect.While due to the proportions of zero are excessive,DIC and convergence diagnostics indicates that the models fit are not satisfactory enough in fetal distress and NICU.Conclusion Stata version 14.0 can do well in Bayesian meta-analysis for two comparison groups,while it can't perform well for multiple comparisons.The researcher successfully constructed a random-effects model based on Poisson distribution Bayesian meta-analysis which can deal with sparse data meta-analysis well.But when the proportion of zero is excessive,the suitable models still need to be further explored.