Medical records have often been found to be less reliable than interviews to patients when data on the initial signs and symptoms of cancer, and the out-of-hospital diagnostic process are sought; in spite of this, a large body of research on "diagnostic delay" in cancer is based on clinical records. As part of a study on delay in neoplasms of the digestive tract we analyzed the agreement on the type and date of the initial symptom between hospital records and a structured personal interview. Records were abstracted for a random sample (N = 60) of 183 patients interviewed. Concordance on the date of the first symptom was deemed to exist if the difference was +/- 30 days. The Kappa index (kappa) and the overall proportion of agreement (with its corresponding 95% confidence interval) were used. Medical records and structured personal interviews were concordant on the type of the first neoplastic symptom in only 61% of cases (kappa = 0.50): 67% in esophagus cancer (kappa = 0.49), 60% in stomach cancer (kappa = 0.52), and 61% in colorectal cancer (kappa = 0.50). Records underestimated the occurrence of anorexia as first symptom and overestimated weight loss and dysphagia. Only 56% of cases were date-concordant, the agreement being lower in colorectal cancer (46%) than in esophageal (67%) and stomach cancer (75%). Records indicated the first symptom to have occurred at a later date than interviews in 33% of cases; overall, a study based on hospital records would have underestimated the symptom to diagnosis interval by 2.2 months per patient. Only 40% of cases were totally (symptom and date) concordant. Marked discrepancies may exist between the information contained in medical records and what patients report during a structured interview. The quality of medical records data on the duration and nature of cancer symptoms should be assessed before its use in etiologic and evaluative research.