Impact of violations of the assumptions of the self-controlled case series design in pharmacoepidemiological studies: An example of antidepressants use and the risk of hip fracture

Abstract

Background: The self-controlled case-series (SCCS) design has been applied to control for time-fixed (un) measured confounding in pharmacoepidemiological studies. Although previous studies acknowledged that violations of the key SCCS assumptions lead to biased exposure effects, little is known about the impact of the violations in empirical studies. Objectives: The aim of this study was to evaluate the impact of various levels of violation of assumptions of the SCCS design and different definitions of observation/risk periods in a study of antidepressants use and the risk of hip/femur fracture (HF). Methods: Information on adults with an HF who used antidepressants at any time during the observation period 2001-2009 was extracted from the UK THIN (6632 cases) and the Dutch Mondriaan (136 cases) databases. The incidence rate ratio (IRR) using this design was defined as the rate of events during exposed periods and during all other observed periods. The IRR of HF was estimated using conditional Poisson regression. Results: The IRRs appeared extremely biased when all subjects were censored at their first/last HF or when the analysis was restricted to subjects experiencing hip fracture after initiating antidepressant use. For example, in THIN, IRRs for >365 days of exposure were 1.26 [1.13-1.42] when complete follow-up was considered and 40.1 [32.2-49.9] when censoring was at the first event. However, modest censoring at the first or last event (up to 20%) had a minor impact on the IRRs. Additionally, results were consistent when including subjects who were exposed at the start of follow-up and for different risk period definitions. Conclusions: The SCCS design is sensitive to violations of the assumptions and yields apparently biased estimates when a significant number of subjects are censored at the event or when the analysis is restricted subjects who experienced hip fracture after initiating antidepressants. The performance of this design may differ across studies and across databases. Therefore, in each SCCS study, correct specification of the SCCS design should be carefully assessed and reported.