We assessed the interobserver agreement on the clinical diagnosis of dementia syndrome and dementia subtypes as part of a cross-national project on the prevalence of dementia. Fourteen clinicians from the participating countries (Canada, Chile, Malta, Nigeria, Spain, and the United States) independently assessed the diagnosis of 51 patients whose clinical information was in standard records written in English. We used the DSM-III-R and ICD-10 criteria for dementia syndrome, the NINCDS-ADRDA criteria for Alzheimer's disease (AD), and the ICD-10 criteria for other dementing diseases, and measured interobserver agreement. We found comparable levels of agreement on the diagnosis of dementia using the DSM-III-R (kappa = 0.67) as well as the ICD-10 criteria (kappa = 0.69). Cognitive impairment without dementia was a major source of disagreement (kappa = 0.10). The kappa values were 0.58 for probable AD, 0.12 for possible AD, and rose to 0.72 when the two categories were merged. The interrater reproducibility of the diagnosis of vascular dementia was 0.66 in terms of kappa index; the diagnoses of other dementing disorders as a whole reached a kappa value of 0.40. This study suggests that clinicians from different cultures and medical traditions can use the DSM-III-R and the ICD-10 criteria for dementia effectively and thus reliably identify dementia cases in cross-national research. The interrater agreement on the diagnosis of dementia might be improved if clear-cut guidelines in the definition of cognitive impairment are provided. To improve the reliability of AD diagnosis in epidemiologic studies, we suggest that the NINCDS-ADRDA "probable" and "possible" categories be merged.