Do case-only designs yield consistent results between them and across different databases (DB)? Hip fractures associated with Benzodiazepines (BZD) as a case study

Background: The case crossover (CXO) and selfcontrolled case series designs (SCCS) are increasingly used in pharmacoepidemiology. In both designs relative risk estimates are obtained within persons rather than between persons thus implicitly controlling for fixed confounding variables. Objectives: To examine the consistency of relative risk estimates of hip/femur fractures (HF) associated with the use of BZD across case-only designs and across two different DB, when same protocol and analytical methods are applied. Methods: CXO and SCCS studies were carried out in BIFAP (Spain) and CPRD (UK). For the CXO, exposure to BZD was divided into non-use, current (up to 30 days after the end of last supply), and recent (1-60 days after). A case moment with four control moments (each 90 days apart) were defined from index date (HF); odds ratios (OR; 95%CI) of current use vs. non-use were estimated using conditional logistic regression with adjustment for co-medications (AOR). For the SCCS, exposure to BZD was divided similarly, but current use was subdivided into:1-30; 31-60; 61-182; 183-365; and >365 days. A conditional Poisson regression was used to estimate incidence rate ratios (IRR; 95%CI) of current use as compared to non-use, adjusted for age. To investigate possible event-exposure dependence we also evaluated the relative risk excluding a pre-exposure time of 30 days. Results: In the CXO current use of BZD was associated with an increased risk of HF in both DB, BIFAP [crude OR= 1.70 (1.50-1.92); AOR= 1.47 (1.29-1.67)] and CPRD [crude OR= 1.75 (1.60-1.92); AOR= 1.55 (1.41-1.67)]. In the SCCS IRRs for the first current period was 0.79 (0.68-0.92) in BIFAP and 1.42 (1.27-1.59) in CPRD. However, when we removed the 30 day pre-exposure period from non-use, the IRR for first current period was 1.40 (1.21-1.62) in BIFAP and 1.59 (1.42-1.78) in CPRD. Conclusions: CXO designs yielded consistent results across DB, while SCCS did not. However, once we accounted for the event-exposure dependence, estimates derived from SCCS were more consistent across DBs and with CXO results