Four national samples, totaling 1,692 cases selected from ESRD patients incident in 1986 and 1987, were analyzed to document the reliability of the USRDS database. For these cases, the reported data in the USRDS database were compared with on-site source documents (medical records, hospital discharge summary records, transplant summaries, and billing records). The results show an overall agreement rate of 90.6 percent for over 50 variables. Agreement rates on the most important fields range from a low of 87.0 percent for point-in-time variables (eg, modality of treatment at the end of the year), to a high of 94.9 percent for dates (eg, date of birth, death, first ESRD service). The study concludes that (1) appropriate definitions and protocols are needed in the assessment of the primary disease causing ESRD, and (2) formal definition of when the patient becomes an ESRD patient needs review. This study points to the strengths and limitations of the USRDS database and indicates areas that need to be addressed to improve current and future data collection efforts.