Purpose
Different strategies applicable to control for confounding by indication in observational studies were compared in a large population‐based study regarding the effect of bisphosphonates (BPs) for secondary prevention of fractures.
Methods
The cohort was drawn from healthcare utilization databases of 13 Italian territorial units. Patients aged 55 years or more who were hospitalized for fracture during 2003–2005 entered into the cohort. A nested case‐control design was used to compare BPs use in cohort members who did (cases) and who did not experience (controls) a new fracture until 2007 (outcome). Three designs were employed: conventional‐matching (D1), propensity score‐matching (D2), and user‐only (D3) designs. They differed for (i) cohort composition, restricted to patients who received BPs straight after cohort entry (D3); (ii) using propensity score for case‐control matching (D2); and (iii) compared groups of BPs users versus no users (D1 and D2) and long‐term versus short‐term users (D3).
Results
Bisphosphonate users had odds ratios (95% confidence interval) of 1.20 (1.01 to 1.44) and 0.95 (0.74 to 1.24) by applying D1 and D2 designs, respectively. Statistical evidence that long‐term BPs use protects the outcome onset with respect to short‐term use was observed for user‐only design (D3) being the corresponding odds ratio (95% confidence interval) 0.64 (0.44 to 0.93).
Conclusions
User‐only design yielded closer results to those seen in RCTs. This approach is one possible strategy to account for confounding by indication. Copyright © 2014 John Wiley & Sons, Ltd.