The Veterans Administration's discharge abstract system was studied to identify error frequency, source, and effect in five Veterans Administration hospitals. We reviewed 1,829 medical records from 21 services for concordance with the abstract; sampling provided 95% confidence for each service. Of these records, 1,499 (82%) differed from the abstract in at least one item. Of 20,260 items, 4,360 (22%) were incorrect, with three error sources: physician (62%), coding (35%), and keypunch (3%). We projected 2.14 physician and 0.81 coding errors in the average abstract. Eighty-nine percent of projected physician errors were failures to report a procedure or diagnosis. Coding was subjective and errors were synergistic with physician errors. We projected that correction of errors would change 19% of the records for diagnosis-related group purposes and substantially increase future resource allocation. This effect varied considerably by service.