BACKGROUND: Gender differences in health care utilization in older Americans may be confounded by nursing home residence. Medicare data contain several files that can be used to create a measure of nursing home residence, but prior work has not addressed which best account for potential confounding. Simpson's paradox occurs when aggregated data support a different conclusion from what the disaggregated data show. We describe such a paradox that appeared when we sharpened our definition of "nursing home residence" while examining gender differences in Medicare utilization at the end of life. METHODS: To understand gender-specific health care utilization at the end of life, we conducted a retrospective analysis of a national random sample of Medicare beneficiaries aged 66 or older who died in 2001 with Parts A and B data for 18 months before death. We sought to associate each of total hospital days and costs during the final 6 months of life with numbers of primary care physician visits in the 12 preceding months. In addition to demographics, comorbidities, and geography, "nursing home residence" was a potential confounder, which we imputed in two ways: 1) from skilled nursing facility bills in the Part A Medicare Provider Analysis and Review (MedPAR) file; and 2) from Berenson-Eggers-Type-of-Service codes indicating widely spaced doctor visits in nursing homes obtained from Medicare's carrier file. CONCLUSION: Gender differences in Medicare utilization are strongly confounded by nursing home resident status, which can be imputed well from Medicare's carrier file, but not MedPAR. Inc. All rights reserved.