Research And Application Of Bayesian Instrumental Variable In Active Surveillance Of Adverse Drug Reaction

Background:Drug safety is closely related to public health,social harmony and stability,and human long-term development,and it is a key issue in the field of public health.With the development of medical technology,more drugs have been approved.At the meantime,there are increasingly adverse drug reactions(ADRs)followed,which pose a threat to public health.Therefore,it is essential to improve drug safety surveillance and evaluation measures to detect,confirm and deal with ADRs as soon as possible.And to some extent,it will help to lower drug-induced risks and reduce health hazards.ADR surveillance is an important way to ensure drug safety.Active surveillance is a variety of activities,behaviors and research based on pharmacoepidemiology methods to identify medical product–adverse outcome associations that may be safety signals.Compared with passive monitoring relying on ADR spontaneous reporting system(SRS),active surveillance is supposed to formulate detailed monitoring plans in advance,including the collection scheme of ADRs or adverse events.And through the implementation of the plan,to achieve the purpose of collecting ADR information in an all-round way.Therefore,active surveillance can overcome the limitations of passive monitoring,such as under-reporting,lack of information and population of drug users.The results of active surveillance have more reliable and far-reaching impact on clinical drug use,drug regulation and social economy.However,active surveillance is mostly observational study.It is inevitable that there are confounding factors,especially potential or unmeasured confounding factors,such as family status of patients,characteristics of doctors,distance of medical institutions and time factors.These factors are difficult to measure directly in the process of surveillance,but it will bias the relationship between drugs and safety concerns if just ignoring them.Instrumental variable(IV)analysis is commonly used to control bias caused by potential or unmeasured confounding factors,which has been applied in the field of pharmacoepidemiology recently.However,the sample size,the binary treatment and outcome of interest and the low incidence of interest variables in active surveillance have an impact on the suitability of IV analysis.Meanwhile,considering the issue of weak IV,it is urgently needed to explore the method to overcome above difficulties and improve the evaluation efficiency.Objectives:The aim of this study is to establish an IV model combined with Bayesian theory.Comparing the Bayesian instrumental variable analysis model with classical IV models under various data conditions,we want to explore the optimal parameter estimation model to improve the applicability of IV analysis in active surveillance and reduce the estimated bias induced by unmeasured confounding factors.After testifying the performance of the established model,we aim to apply it to the real-world drug safety study to provide effective warning information for drug risk management.Methods:1.Simulation studyFirstly,we reviewed the pharmacoepidemiology literature on IV analysis theory and application and simulated a binary IV according to the physicians’ prescription preference.According to the real data of drug use and ADRs,which are often concerned in drug safety studies,binary treatment and outcome variables are simulated.Secondly,we set different simulation parameters,including different sample size,different strength of IV,different strength of unmeasured confounding factors,and different incidence combination of treatment and outcome variables,24 data scenarios were generated,each simulation scenario was simulated 1000 times to ensure the accuracy of the results.Furthermore,two-stage probit residual inclusion model,probit+logistic residual inclusion model,Limited Information Maximum Likelihood(LIML)model and its Fuller correction(Fuller-LIML)model were constructed as the traditional binary IV analysis method.Meanwhile,the Bayesian statistical idea was integrated into the IV analysis to establish the Bayesian bivariate probit IV analysis model.The Hamilton Monte Carlo(HMC)method was used to get the posterior distribution of the parameters under the specified prior distributions and statistical inference was made based on the posterior distribution.Finally,we comprehensively compared the Bayesian IV model and traditional IV model in different data situation,and the feasibility and applicability of Bayesian IV model in active surveillance were discussed.2.ApplicationsThe models we built were applied to a prospective,open and multi-center post-marketing clinical study of a Chinese herbal medicine used in the treatment of mild or moderate cerebral infarction.The effect of concomitant therapy on adverse drug events(AEs)and serious adverse drug events(SAEs)were evaluated so as to further confirme the drug related risk factors and ensure the patients safety.Results:1.Simulation studyIn the simulation study,the traditional IV models and Bayesian IV model were used to estimate the treatment effect.Bias,standard deviation,mean square error and confidence interval width were used to compare the accuracy and accuracy of the five models in different data situations.When the sample size is small,the Bayesian IV model obtained smaller absolute bias,narrower confidence interval width and showed the best stability.When the sample size is large,the confidence interval widths of the five models were obviously narrowed and the accuracy was improved.When the strength of unmeasured confounder was weak,the absolute bias of the five methods was similar.But Bayesian IV model obtained the narrowest confidence interval width,Fuller-LIML estimation got the smallest mean square error and showed higher accuracy.With the strength of unmeasured confounder increasing,the five models performed steadily.The absolute bias of Fuller-LIML and Bayesian IV model were the smallest and the accuracy were the highest under different strength of IV,and the confidence interval width of Bayesian IV model was the narrowest.When other parameters are consistent,with the increase of the strength of IV,the width of confidence intervals of the five models tends to narrow gradually.Under the combination of low incidence of treatment and low incidence of outcome,Bayesian IV model obtained minimum absolute bias,narrowest confidence interval width,minimum standard error and mean square error.Fuller-LIML nevertheless showed best among the five models under the combination of high incidence of treatment and low incidence of outcome.2.ApplicationsComparing the estimation results of IV analysis models with Logistic regression analysis for applications,it was found that the Bayesian IV model could improve the inference efficacy of concomitant therapy on AEs and SAEs by controlling unmeasured confounders.It was verified that concomitant therapy was the risk factor of AEs and SAEs.Conclusions:To reduce the bias caused by unmeasured confounders in active surveillance of ADR,we should consider the real-world data situations to select the optimal IV analysis model.When the sample size increases and the strength of IV enhances,the confidence interval widths of the five models are narrower and the accuracy of estimation has been improved.When the sample size is large,correction effects of the Fuller-LIML on LIML performs well.Fuller-LIML can effectively reduce the absolute bias,standard error and mean square error,obviously narrow the width of confidence interval,and improve the accuracy and precision of estimation.For the scenario of small samples,weak IV,strong unmeasured confounders,the combination of low incidence of treatment and low incidence of outcome,Bayesian IV analysis model is the best choice.ADR information collected in real-world study is various,so the treatment and outcome of safety evaluation are not limited to binary variables,and it is also challenging to find effective IVs from the data.Due to the long observation period of active surveillance and the restriction of human and financial resources,there may be missing data.Above problems which may be encountered in the practice of IV,it will be better to consider them comprehensively in the design stage of active surveillance,and further improve them with the corresponding methodology in the analytical stage.