Study Of Causl Effect EstimationMethods Based On Propensity Score And Application In Assessing Post-marketing Adverse Drug Reaction

Background and purposeDrug safety is a basic livelihood project, and the issue of drug safety research has alwaysbeen the focus in public health research. Although before a drug launching to market it hasexperienced a large number of animal experiments and clinical trials, the post-marketingdrug safey surveillance is very important, due to the results of animal experiments are notsufficient to predict the safety of human drug application and in clinical trials theirconditions of drug applications have some differences with the actual clinical practice. Forexample, study duration is short, the sample size is small, and there are strict inclusioncriterias for subjects. In all pharmacovigilance tools, spontaneous reporting systemprovides maximum information at the least cost and it is the primary means of monitoringadverse drug reactions at present. Due to we usually have little information aboutbackground events incidence in the entire population, the number of patients exposure totargeted drugs or the under reporting rates, it is hard to obtain a reliable expected numberof drug-event combination. The most popular analyzing tool of these data is data mining.While all these data mining tools are trying to detect the association between drug andevent other than get the causal effect of a drug to event. Besides, these tools didn’t takeconfounders such as age, indication, weight, concomitant medication etc. into consider.Due to the data features in spontaneous reporting system, the popular methods to controlconfounders such as stratifying and logistic regression are not applicable.Our poupose of this study is to introduce a new method to assessing the adverse drugreaction in spontaneous reporting system, taking into account both the impact ofconfounding factors on drug and adverse events, and analyzing the relationship betweenthe drugs and adverse events in a causal view. And we are aim to further confirm thesuspicious signal after discovery the signal using usual data mining tools and before finalconfirm through expert evaluation, drug epidemiology surveys or case studies and thus toprovide an evidence for drug risk management, evaluation and decision-making.MethodsWe introduced the Rubin’s causal model framework into spontaneous reporting systemdata analysia and built a Rubin’s model based on the characteristics of the spontaneousreporting system data. And we clarified the definition of study population and potential outcome. Meantime we comprehensively reviewd the theory of several approaches toestimating average causal effects using propensity score, and in particularly introduced theproperties of these methods when the outcome variable is dichotomous.Several parameters were set in line with the characteristics of the spontaneous reportingsystem data to compare the properties of several estimating average causal effects methodsusing propensity score via Monte Carlo simulation method. Several senarios were set toinvestigate the accuracy and efficiency of estimates through calculating the bias, standarddeviation and MSE of estimated causal effects. These senarios include constructing rightand wrong propensity score models, setting different relationship strengths between thecovariates and treat covariates and different relationship strengths between the covariatesand outcome variables. Besides, we simulated a study to compare the performance ofcausal effects between the two methods using Bayesian propensity score and traditionalpropensity score.We reorganized the data of U.S. FDA spontaneous reporting system (FAERS) reported inthe year of2011and2012, and detected the suspicious drug-evetn combination usingconventional data mining methods, combined with the number of reports. We chose thecases who may take the targeted drugs as our study observations according to the definitionof our study population. Then we estimated the causal effects using several methods basedon propensity score to examine whether these methods can be applied to real data.ResultsWe have established a causal model of adverse reactions evaluation in spontaneousreporting system. We take all possible cases that would take targeted drugs as studypopulation, define possibility of advents ouccurence while taking or not taking targeteddrugs as potential outcomes, define the variables those could be related to the treatmentassignment or outcome as covariates, and define the average difference of potentialoutcomes as treatment causal effects.Simulation results indicate that the popular version of stratification via estimatedpropensity scores based on within-stratum sample mean differences and a fixed number ofstrata can lead to biased inference due to residual confounding, and the effect of this biasbecomes more serious with increasing sample size. A modified stratification based insteadon within-stratum regression estimates of treatment effect can eliminate this bias andachieve dramatic improvements in efficiency, and this approach is not sensitive to mismodelling propensity score model while adding more variables. Besides this approachhas a relative high estimate efficiency compared to other methods. But it is hard to estimatea variance when the relationship between outcome variable and covariates is nonlinear. Theweighting methods to estimate the causal effect are usually able to get unbiased estimates.The simulation results also showed that the most widely used weighing methods usingfixed sample size has a relative low estimating effiency for it doesn’t use full sampleinformation. Besides, the double robustness method has a robust property in that itcontinues to lead to unbiased estimation of the average causal effect even if the regressionmodels involved do not coincide with the true relationship, according the analyst broadprotection against misspecification not available with these other approaches. Under oursimulation settings, we didn’t find big differences of casual effects between the twomethods using Bayesian propensity score and traditional propensity score. But the twostratification methods didn’t performance well with large bias and low efficiency when thesample size was small.The results of analysis of FAERS data reported in2011and2012showed that the use ofbisphosphonates can have a higher incidence of fractures than not using bisphosphonates.And the causal effects are IPW1：0.1083(0.0028,0.2138)； IPW2：0.1086(0.0049,0.2123)； DR：0.1065(0.0028,0.2102)； S：0.0711(-0.0544,0.1966)； SR：0.1123(0.0068,0.2178).ConclusionsAfter introducing the concept of causal model into spontaneous reporting system dataanalysis, we can obtain more intuitive interpretation of results. Methods estimating causaleffects based on propensity score are applicable in assessing adverse drug reaction inspontaneous reporting system. These methods can overcome the limits of ignoringconfounders using conventional data minig methods and can obtain more credible results.It will serve as an additional method to assess adverse drug reaction after data mining andbefore the final confirm. Results of the Example showed that bisphosphonates on fractureincidence may have a causal relationship, suggesting that we need to conduct further studyof this combination, such as Meta-analysis, large-scale drug epidemiology, case studies andso on.