OBJECTIVE— To define the test characteristics of four methods of screening for diabetic retinopathy. RESEARCH DESIGN AND METHODS— Four screening methods (an exam by an ophthalmologist through dilated pupils using direct and indirect ophthalmoscopy, an exam by a physician's assistant through dilated pupils using direct ophthalmoscopy, a single 45° retinal photograph without pharmacological dilation, and a set of three dilated 45° retinal photographs) were compared with a reference standard of stereoscopic 30° retinal photographs of seven standard fields read by a central reading center. Sensitivity, specificity, and positive and negative likelihood ratios were calculated after dichotomizing the retinopathy levels into none and mild nonproliferative versus moderate to severe nonproliferative and proliferative. Two sites were used. All patients with diabetes in a VA hospital outpatient clinic between June 1988 and May 1989 were asked to participate. Patients with diabetes identified from a laboratory list of elevated serum glucose values were recruited from a DOD medical center. RESULTS— The subjects (352) had complete exams excluding the exam by the physician's assistant that was added later. The sensitivities, specificities, and positive and negative likelihood ratios are as follows: ophthalmologist 0.33, 0.99, 72, 0.67; photographs without pharmacological dilation 0.61, 0.85, 4.1, 0.46; dilated photographs 0.81, 0.97, 24, 0.19; and physician's assistant 0.14, 0.99, 12, 0.87. CONCLUSIONS— Fundus photographs taken by the 45° camera through pharmacologically dilated pupils and read by trained readers perform as well as ophthalmologists for detecting diabetic retinopathy. Physician extenders can effectively perform the photography with minimal training but would require more training to perform adequate eye exams. In this older population, many patients did not obtain adequate nonpharmacological dilation for use of the 45° camera.