A Saskatchewan hospital separations database was compared to abstracted hospital records to determine the reliability of the database (i.e. accuracy with which the computer data reflect the charts from which they were coded) and the validity of classifying rheumatoid arthritis status with the database (i.e. the extent to which rheumatoid arthritis mentioned in the database reflected the condition of the patient). A sample of hospitalized subjects fell into three categories: 144 who never had a database diagnosis of any arthritis, 146 who had a database diagnosis of osteoarthritis, and 142 who had a database diagnosis of rheumatoid arthritis. These 432 people experienced 1717 hospitalizations eligible to match a hospital database listing by date, and 1618 matched. Of the remaining 99, 35 were relatively recent and probably had not yet been entered into the database, 39 were possibly entered incorrectly, and 25 could not be matched in any way. Of 150 hospitalizations with a database diagnosis of rheumatoid arthritis, this diagnosis was in the hospital record for 125. Chart documentation of rheumatoid arthritis status was greatest for subjects who, according to the database, were hospitalized by a rheumatologist: of 73 subjects in this category, abstractions showed 69.9% met > or = 5 American Rheumatism Association criteria, 15.1% met < 5 criteria but had a rheumatologist's diagnosis of rheumatoid arthritis, 1.3% met < 5 criteria and had a rheumatoid arthritis diagnosis by a non-rheumatologist, and 13.7% had no mention of rheumatoid arthritis or its symptoms in any medical record abstracted. In summary, reliability of the database was excellent, but validity depended on the source of diagnosis.