Propensity Scores And Application In Signal Detection Of Adverse Drug Reactions

The drug safety is an important problem concerned by people all the time. People werefurther aware of the harm of adverse drug reaction (ADR) by the momentous adverse drugissue in resent years, such as serious cardiovascular events caused by sibutramine.Therefore, it will be increasingly important to implement surveillance for marketed drugs.Spontaneous reporting systems, SRSs, are the main source for early ADR signal detectionin countries individually, and their data was spontaneous reported after marketing.Nowadays, many experts of pharmacoepidemiology and pharmacovigilance haveresearched signal detection technique to extract suspect drug safety signals from so largedatabase. After current in markets, influence factors were difficult to control without strictrandomized controlled trial. So the confounding biases were controlled hardly in SRS.However, the current methods of detecting signals are focusing on the drug-ADRassociation, they neglect confounding factors, such as gender, age, co-medications and soon, which mask or magnify intrinsic associations between drug and ADR. It is inevitablethat many false positive or false negative signals were generated and ability of signaldetecting methods was decreased. Therefore, it will be an farther significant forpharmacovigilance how to make full use of information in SRS in order to controleffectively confounders, so that accuracy of signal detecting are improved and surveillanceof drug safety are strengthened.MethodsPropensity score, PS, was proposed as a method for balancing covariates in80s last century.It was applied in observation studies and non-randomized controlled trial studies.Based on simulation, this study fit PS model by simulating data according to thecharacteristics of SRS, and estimated PS values. The difference of detecting resultsbetween before and after balancing baseline covariates with PS was compared. And theeffect in real data was validated.ResultsIn simulation study, before balancing with PS, both predefined two true positive signalsand two false positive signals were detected by signal detection methods as suspectedsignals. After balancing with PS, the values of true positive signals did not change, whilefalse positive signals were disappeared.Similarly, in real datasets, the ROR of combination ‘Quetiapine-amenorrhea’ as suspected signal was higher than threshold. Expected gender, other covariates were imbalancedbetween treatment and control groups. After balanced covariates by PS, ROR of thiscombination is lower than the threshold, which was suspected as false positive associationby analysis of original data.ConclusionsBecause of limitations of marketing drugs, signal detection methods are likely to generatedfalse positive signals due to lacking analysis covariates. Nowadays, information in SRSwas not applied completely for drug safety research. During this study, it is demonstratedthat PS could make full use of information of SRS to balancing baseline, controllingconfounding bias, reducing false positive signals and providing experts with proofs forevaluating associations of signals. Although there are many limitations, with improvingdata quality and methods, PS will be an effective tool for controlling confounders in SRS.