Instrumental variable analysis using multiple databases: An example of antidepressant use and risk of hip/femur fracture

Abstract

Background: Instrumental variable (IV) analysis can reduce bias due to unmeasured confounding, yet it has not been widely used in pharmacoepidemiologic studies. Objectives: To assess the performance of several IVs across multiple databases in a study of antidepressant use and risk of hip/femur fracture (HF). Methods: Information on adult patients with at least one prescription of a selective serotonin reuptake inhibitor (SSRI) or tricyclic antidepressant (TCA) during 2001-2009 was extracted from three European databases: THIN (UK, n = 570139), BIFAP (Spain, n = 250865), and Mondriaan (Netherlands, n = 22474). Conventional Cox model and two-stage IV analysis were applied to estimate the risk of HF associated to initiation of SSRI vs. TCA. IVs were created using the proportion of SSRI prescriptions per practice (PSP) or using the number, one (PPP1), five (PPP5) or ten (PPP10), of previous prescriptions by a physician. Quantitative methods (e.g. correlation (r), standardized difference (SDif)) were used to assess the validity of IVs. 95% confidence intervals (CI) in IV analysis were estimated using bootstrapping. Results: Conventional analysis showed that SSRI use was associated with an increased risk of HF in BIFAP and THIN, hazard ratio (HR) 1.35 [95%CI 1.18-1.56] and 1.35 [1.26-1.44], respectively and similarly in Mondriaan (though not significant), HR 1.36 [0.86-2.15]. The IVs PSP (THIN and BIFAP) and PPP10 (THIN and Mondriaan) were strongly associated (r>0.15) with SSRI prescribing and independent of confounders (SDif