Incidence rates of non-Hodgkin's lymphomas (NHLs) have nearly doubled in recent decades. Understanding the reasons behind these trends will require detailed surveillance and epidemiological study of NHL subtypes in large populations, using cancer registry or other multicenter data. However, little is known regarding the reliability of NHL diagnosis and subtype classification in such data, despite implications for the accuracy of incidence statistics and studies. Expert pathological re-review was completed for 1526 NHL patients who were reported to the Greater Bay Area Cancer Registry and who participated in a large population-based case-control study. Agreement of registry diagnosis with expert diagnosis and with International Classification of Diseases for Oncology-2 (Working Formulation) subtype classifications was measured with positive predictive values and kappa statistics. Agreement of registry and expert diagnoses was high (98%). Thirty patients were found on review not to have NHL; most of these had leukemia. For subtypes, agreement of registry and expert classification was more moderate (59%). Agreement varied substantially by subtype from 5% to 100% and was 77% for the most common subtype, diffuse large cell lymphoma. Seventy-seven percent of 128 registry-unclassified lymphomas were assigned a subtype on re-review. Our analyses suggest excellent diagnostic reliability but poorer subtype reliability of NHL in cancer registry data information that is critical to the interpretation of lymphoma time trends. Thus, overall NHL incidence and survival statistics from the early 1990s are probably accurate, but subtype-specific statistics could be substantially biased, especially because of high (15-20%) proportions of unclassified lymphomas.