Background: In real world practice, treatments are not randomly allocated but selected based on indications and other considerations. If not taken into account, these selection factors can impact comparative effectiveness estimates in observational studies. Objective: To assess the potential contribution of unmeasured general health status to patient selection in assessing clinical effectiveness of ICDs for primary prevention. Methods: Linking CMS ICD registry, National Cardiovascular Data Registry (NCDR®) ICD Registry™, a heart failure (HF) registry, and Medicare files (2005-2009), we identified HF patients ≥66 years who were eligible for primary prevention ICDs. We compared the risk of nontraumatic hip fracture, skilled nursing facility (SNF) admissions, and 30-day mortality between hospitalized HF patients with and without ICDs. Crude and adjusted hazard ratios were estimated by Cox proportional hazards regression. Results: Compared with 17,853 patients without ICDs, 11,573 patients with ICDs were younger and had lower ejection fractions and more cardiac hospitalizations, but fewer noncardiac hospitalizations and comorbid conditions. Patients with ICDs had greater freedom from unrelated events: hip fracture (HR 0.39; 95% CI, 0.37-0.41, Figure), SNF admission (HR 0.53 95% CI, 0.50-0.55), and 30-day mortality (HR 0.10; 95% CI, 0.09-0.12%). After adjustment for multiple measured patient characteristics, risks in the two groups were less disparate but still substantially lower in ICD patients (HRs ranged from 0.20 to 0.84). Conclusion: The lower risks of these outcomes likely reflect unmeasured differences in comorbidity burden and frailty. These findings highlight potential pitfalls of observational comparative effectiveness research. The findings also support that physicians in clinical practice consider general health status in selecting patients for ICD implantation, avoiding implantation in patients who are less likely to benefit.