Background: The SF-36 Health Survey questionnaire has been proposed as a generic measure of health outcome. However, poor rates of return and high levels of missing data have been found in elderly subjects and, even with face-to-face interview, reliability and validity may still be disappointing, particularly in cognitively impaired patients. These patients may be the very patients whose quality of life is most affected by their illness and their exclusion will lead to biased evaluation of health status. A possible alternative to total exclusion is the use of a proxy to answer on the patient’s behalf, but few studies of older people have systematically studied patient-proxy agreement. Objective: To compare the agreement between patients, lay and professional proxies when assessing the health status of patients with the SF-36. Methods: The SF-36 was administered by interview to 164 cognitively normal, elderly patients (Mini-mental State Examination 24 or more) referred for physical rehabilitation. The SF-36 was also completed by a patient-designated lay proxy (by post) and a professional proxy. Agreement between proxies and patients was measured by intraclass correlation coefficients (ICCs), and a bias index. Results: Professional proxies were better able to predict the patients’ responses than were the lay proxies. Criterion levels of agreement (ICC 0.4 or over) were attained for four of the eight dimensions of the SF-36 by professional proxies, but for only one dimension by lay proxies. In professional proxies, the magnitude of the bias was absent or slight (<0.2) for six of the eight dimensions of the SF-36 with a small (0.2–0.49) negative bias for the other two. Lay proxies showed a negative bias (i.e. they reported poorer function than did the patients themselves) for seven of the eight dimensions of the SF-36 (small in two and moderate (0.5–0.79) in five). Conclusions: For group comparisons using the SF-36, professional proxies might be considered when patients cannot answer reliably for themselves. However, in the present study, lay proxy performance on a postal questionnaire showed a strong tendency to negative bias. Further research is required to define the limitations and potentials of proxy completion of health status questionnaires.