In the diagnosis of dementia, information obtained from informants or proxies is important; however, little is known about the circumstances under which informants' reports lack accuracy. This study compares informant reports of cognitive status against psychometric tests to identify the degree of, and factors associated with, discrepant reporting. Four areas of patient cognitive ability were examined: memory of recent events, ability to remember a short list of items after a delay, language/naming abilities, and working memory. Primary care practitioners recruited 248 community-dwelling patients aged 75 years or more or aged 50-74 years with suspected memory complaints; 248 friends or relatives acted as informants. More than half of the informants (60%) gave responses consistent with psychometric testing. Informants who underreported patient difficulties tended to report on patients who were diagnosed as having subclinical dementia, were less educated, and had poorer remote memory. Informants who overreported difficulties were more likely to report on those diagnosed with dementia. While informant accounts are critical for the assessment of dementia, we found that in 40% of cases these reports may be inaccurate, particularly when the patient has low education and poor remote memory or when overall cognitive difficulties are mild.