Accuracy of basic cancer patient data: results from an extensive recoding survey.

Abstract

The accuracy of data coded from the medical records of 985 patients from 22 major U.S. cancer centers was checked by recoding during 1978-81. The 29 items covered demographics, diagnosis, and therapy. Original codes were compared to recodes, and disagreements were classified as major or minor. The highest rate of major disagreements, 23%, was for stage of disease, followed by 10% for histology and 7% for site. Major disagreement rates for most other items were under 7%. Only 3% of a large sample of major disagreements involved justifiable differences in interpretation; the others were due to errors in the use of records. Major disagreement rates varied by a factor of 10 across sites, 4 across centers, and 2 across stage of disease. For several items the code "unknown" was overused and led to disagreements. A new procedure is presented for analysis of disagreement rates. The results from this procedure can guide training effort to improve coding accuracy.