The authors hypothesized that cognitive impairment is associated with falls in older adults, but that transfer bias may obscure this association in cross-sectional community studies. The bias would arise if demented patients who fall are relatively unavailable to community surveys due to death or institutionalization. To test this hypothesis, a "dose-response" relationship between falls and cognitive impairment was tested for using data from a longitudinal cohort study of 157 patients with Alzheimer-type dementia. In a cross-sectional analysis of baseline data, when 96% of the cohort were community residents, the association between falls and cognitive impairment was insignificant (odds ratio for a 10-point change in Mini-mental State score = 1.2, 95% confidence interval 0.76-1.9). Yet cognitive impairment at baseline predicted falls during three-year follow-up (OR = 1.8, 95% CI 1.1-3.0). Both severity of dementia and falls were risk factors for death or institutionalization. Patients at highest risk for leaving the community during follow-up were those who became non-ambulatory. The authors conclude, based on longitudinal data, that there is a dose-response effect between cognitive impairment and falls. Transfer bias probably obscured the association at baseline. These data emphasize the importance of prospective studies of falls.