Background When comparing transfused versus nontransfused patients, erythrocyte transfusion is consistently associated with increased mortality. Nonetheless, unmeasured confounding may unduly influence this comparison. This unmeasured risk may have less influence on comparisons of patients undergoing surgery at hospitals with differing transfusion rates. Methods Administrative databases were used to conduct a population-based cohort study of patients who underwent elective hip- or knee-replacement surgery from 1999 to 2008 in Ontario, Canada. The authors used Cox proportional-hazards models to determine the adjusted association of hospital-specific erythrocyte transfusion rates (i.e., comparing hospitals with differing transfusion rates) with postoperative mortality. For comparison, they also determined the adjusted association of patient receipt of transfusion (i.e., comparing transfused vs. nontransfused patients) with mortality. Results Of 162,190 patients, 23% (n=37,015) were transfused. Hospital-specific transfusion rates at the 66 included hospitals ranged from 10.3 to 57.9%. Compared with nontransfused patients, transfused patients experienced increased adjusted 30-day (hazard ratio 2.32; 95% CI, 1.91-2.83) and 1-yr mortality (hazard ratio 1.75; 95% CI, 1.60-1.91). However, when hospitals were categorized into quartiles based on hospital-specific transfusion rates, mortality rates were similar (highest transfusion quartile vs. lowest transfusion quartile: 30-day mortality, hazard ratio 1.11, 95% CI 0.82-1.50; 1-yr mortality, hazard ratio 1.02, 95% CI 0.82-1.26). Conclusions The association of transfusion with postoperative mortality differed significantly when comparing transfused versus nontransfused patients, as opposed to comparing hospitals with differing transfusion rates. This discrepancy raises questions about the true relationship between transfusion and mortality.