Modeling Of Bayes Propensity Score And Application In Signal Detection Of Adverse Drug Reactions

Background:Noadays, the amount in spontaneous reporting system(SRS) of drug adversereaction(ADR) was incresingly growing. And early ADR signal detection had greatinfluence on health safety of patients and thus become to be the focus of public healthresearch. Hence, it will be very important to apply appropriate data mining and statisticalmethods to implement post-marketing surveillance(PMS) in SRS data. However,ignoringthe confounding factors or misuse of statistical methods to balance them may affect theaccuracy of signal. Our group has been applied traditional propensity score anlysis(PSA) tocontrol coufouding in survillance system for ADR signal detection. But there has beenarosed several problems when use of traditional propensity score anlysis, such as:(a)ignoring the linear relationship of continuous covariate and logit (y);(b) the influence ofuncertain propensity score(PS) to treatment effect;(c) unable to make use of priorinformation;(d) unaccurate in rare events of outcome or numerus covariates.Aim:To establish several Bayesian propensity score analysis (BPSA) such as Bayesianpropensity score stratification (BPSS), Bayesian propensity score matching(BPSM)、Bayesian propensity score weighting(BPSW) and Bayesian propensity score regressionadjusting(BPSR). A simulation study was conducted under the different circumstance ofsample size, treatment effect and prior information precision to explore the accuracy andprecision of treatment effect estimation. Finally, the optimal Bayesian propensity scoremodel will be chosen and applied in the ADR signal detection to avoid the false signals.Methods:In this study, the Monte Carlo method was used to simulate data sets.We firstestablished the model of treatment and covariates, and then set up the model of outcome,treatment and covariates. Different types of covariates were simulated based on theirassociation with treatment and outcome, including covariates associated with bothtreatment and outcome, covariates associated with treatment and covariables associatedwith outcome. Continuous and binary categorical covariates was also simulated accordingto the type of data distribution. We had set five treatment effect(t1.5,1.2,0.8,0.5,0.2), four prior information precision (B0,1,10,100) and3sample sizes(N=50、100、250). First, a Bayesian Logistic model was established to estimate propensity score (PS), and then incorporated in the treatment effect model in differentways, such as matching, stratification, weighting and regression adjusting. Finally, sevenBayesian propensity score model was established. The evaluation indicator of treatmenteffect includes point estimate, standard deviation, bias (absolute bias and relative bias),mean square error, and the coverage of confidence interval. The software of SAS and Rwere used to programming the model.In the end, we chose the Bayesian propensity score matching (BPSM) with the priorinformation precision of100and applied it in the SRS data of Shanghai food and drugadministration in2009. The accuracy of signal detection in BPSA, PSA and general datamining methods was judged according to comparison with Chinese drug adverse reactiondictionary, literature and drug specification.Results:BPSM and BPSS plus Cochran-Mantel-Haenszel (CMH) with a higher priorinformation precision (B100) both made improvement in accuracy and stability oftreatment effect under the circumstance of small sample size (n=50), especially BPSM hadgreat advantage in point estimation and the length of confidence interval. When the samplesize set to100, BPSM and BPSS plus CMH produced similar results. To increase thesample size to250, their difference was further decreased and the point estimation wascloser to true effect.The results was deeply affected by the way of PS in treatment effect model no materthe model of BPSA or PSA. Matching and stratification could be used as the first-choicefor their better performance in treatment effect estimation. The stability of weighting andregression adjusting was poor，which was closely affected by the association betweentreatment and outcome.A case study was conducted in the ADR signal detection.‘Azithromycin and localnumbness’ was detected as a suspected signal by signal detection method of ROR, PRR, ICand MHRA. But we had found that it was a false signal by comparing with Chinese drugadverse reaction dictionary and reviewing literature and drug specification. Univariateanalysis was conducted in the covariates of age, weight, sex, concomitant drug andtriggering time of ADR and showed that all of them were unbalanced. After balanced thecovariates by PSA and BPSA,‘Azithromycin and local numbness’ turned to a negativesignal. The statistics of the four signal detection method using BPSA model was muchlower than the threshold, but it was closer to threshold by using the PSA model. All the covariates included in the model was balanced using BPSA model whereas weight andtriggering time of ADR was still unbalanced by PSA model.Conclusions:BPSM and BPSS plus CMH had great advantage in the accuracy and stability oftreatment effect estimation under the circumstance of higher prior information precision(B100) and small sample size (N=50). In case study, it is demonstrated that BPSMwith a higher prior information () could make full use of information in SRS tobalance the baseline covariates, controlling coufonding bias and reduce the false signals insignal detection.