Objectives: To identify the magnitude of confounding bias caused by factors not measured in claims data studies of sedative-hypnotic use and fractures.
Design: Cross-sectional survey data.
Setting: Medicare Current Beneficiary Survey (MCBS).
Participants: Eight thousand seven hundred eighty-five survey participants aged 65 and older.
Measurements: To determine the association between sedative-hypnotic use and potential confounding factors, five factors not measured in Medicare claims data but measured in the MCBS were assessed: body-mass index, smoking, activity of daily living (ADL) score, cognitive impairment, and Rosow-Breslau physical impairment scale. The association between benzodiazepine or zolpidem use and these five factors was estimated. Combined with literature estimates of the associations between confounders and fractures, it was possible to compute the extent of residual confounding bias caused by a failure to adjust for them.
Results: Comparing zolpidem users with nonusers, there was considerable overestimation of an association with hip fractures if ADL scores (21.5% bias) or Rosow-Breslau impairment scales (10.6%) were unobserved in claims data. All five unmeasured confounders together resulted in net confounding of 9.8% (range 0–39%). Comparing benzodiazepine users with nonusers, effect estimates were moderately biased (6.1%). After correction for this bias, the association observed in claims data (relative risk (RR)=1.38; 95% confidence interval (CI)=1.14–1.66) was comparable with that found in a recent study using clinical data (RR=1.20, 95% CI=0.72–2.00). In contrast to the clinical study, the claims data study achieved formal statistical significance because of its much larger size (288 vs 1,222 hip fractures).
Conclusion: Claims data studies tend to overestimate the relationship between benzodiazepine use and hip fractures. After correcting for such bias, a statistically significant association persists. Concluding that there is no relationship at all based on small clinical studies that did not reach statistical significance may be misleading.