Bayesian hierarchical methods in the detection of potentially teratogenic first-trimester medications.

ABSTRACT

PURPOSE: Bayesian hierarchical models (BHMs) have been used to identify adverse drug reactions, allowing information sharing amongst adverse reactions and drugs expected to have similar properties. This study evaluated the use of BHMs in the routine signal detection analyses of potential first-trimester teratogens, where these models have not previously been applied. METHODS: Data on 15 058 malformed foetuses exposed to first trimester medications (1995-2011) from 13 European congenital anomaly (CA) registries were analysed. The proportion of each CA in women taking a specific medication was compared with the proportion of that CA in all other women in the dataset (55 CAs × 523 medications). BHMs were grouped by either medications or CAs or by both simultaneously, and the results compared with analysing each medication-CA combination separately and adjusting for multiplicity using a double false discovery rate (FDR) procedure. The proportions of "high-risk" medications (medications which have been shown to carry a moderate to high risk of foetal malformations) identified as potential signals were compared, as well as the total number of potential signals requiring follow up (the effective workload). RESULTS: BHMs identified more high-risk medications than the double FDR method, but the effective workload was larger. A BHM grouping both medications and CAs, for example, identified 23% of high-risk medications compared with 14% by the double FDR; however, there was an increase from 16 to 71 potential signals requiring follow up. CONCLUSION: For comparable effective workloads, BHMs did not outperform the double FDR, which is comparatively straightforward to implement and is therefore recommended for continued use in teratogenic signal detection analyses.