Linking hospital discharge and death records - accuracy and sources of bias

Background and Objective: The aim of this study was to develop and apply an automated linkage algorithm to 10 years of California hospitalization discharge abstracts and death records (1990 to 1999), evaluate linkage accuracy, and identify sources of bias. Methods: Among the 1,858,458 acute hospital discharge records with unique social security numbers (SSNs) from 1 representative year of discharge data (1997), which had at least 2 years of follow-up, 66,410 of 69,757 deaths occurring in the hospital (95%) and 66,998 of 1,788,701 of individuals discharged alive (3.7%) linked to death records. Linkage sensitivity and specificity were estimated as 0.9524 9 and 0.9998 and positive and negative predictive values as 0.994 and 0.998 (corresponding to 400 incorrect death linkages among out-of-hospital death record linkages and 3,300 unidentified record pairs among unlinked live discharges). Results: Based upon gold standard linkage rates, discharge records for those of age 1 year and older without SSNs may have 2,520 additional uncounted posthospitalization deaths at 1 year after admission. Gold standard comparison for those with SSNs showed women, the elderly, and Hispanics and non-Hispanic Blacks had more unlinked hospital death records, although absolute differences were small. The concentration of unidentified linkages among discharge records of traditionally vulnerable populations may result in understating mortality rates and other estimates (i.e., events with competing hazard of death) for these populations if SSN is differentially related to a patient's disease severity and comorbidities. Conclusion: Because identification of cases of out-of-hospital deaths has improved over the past decade, observed improvements in patient survival over this time are likely to be conservative. (C) 2004 Elsevier Inc. All rights reserved.