Maryan Morel, Benjamin Bouyer, Agathe Guilloux, Moussa Laanani, Fanny Leroy, Dinh Phong Nguyen, Youcef Sebiat, Emmanuel Bacry and Stephane Gaiffas
Abstract. Background and Purpose Existing screening works provide point estimates for drug-outcome pairs risk assessment. We propose a flexible approach based on dynamic risk estimates to support alert generation while providing additional information on risk qualification (delay, shape) and LOD-specific biases. We illustrate this approach by studying the longitudinal effect of anxiolytic, hypnotic, antidepressant, and neuroleptic molecules on fractures using SNDS, a French large healthcare claims database. Methods. We follow French new users who were 65 y.o. or older in 2014 for up to four years. We use ConvSCCS, a flexible longitudinal model based on self-control case series. This model alleviates several observational claims data issues and does not require precise assumptions on risk timings. The presence of eventual indication biases is assessed by esti- mating dynamic pre-exposure relative risks. Results. Pre-exposure risk estimates suggest the presence of confounding by indication in anxiolytics, hypnotics and neuroleptics estimates, while it is not the case for antidepressants. Tricyclic antidepressants exhibit lower relative risk than other antidepressants. Zolpidem relative risk is consistently higher than Zopiclone across all sensitivity analyses. Conclusion. This approach complements existing screening methods as well as clinical or observational risk quantification studies by providing granular and dynamic risk estimates for many molecules using a single model. It could be used to map molecules and adverse events, pointing out the presence of eventual biases or associations for further investigation.