To investigate a number of prospectively collected factors (sociodemographic, medical and behavioural) and their association with incident dementia in a population-based cohort. Nested case-control analysis (at 2 and 6 years) of a population-based cohort study. Individuals aged 65 years and above from five centres in England and Wales: two rural (Cambridgeshire and Gwynedd) and three urban (Nottingham, Newcastle and Oxford). A total of 4,075 individuals from a detailed assessment group, with risk measured at baseline. Incident dementia at 2 and 6 years. Logistic regression was used to calculate crude odds ratios (ORs) for various risk factors and ORs adjusted for age, sex, education and social class. Age (90+ versus 65-69 years OR = 25.6, 95% confidence interval (CI) = 11.6-56.9) and sex (women versus men OR = 1.6, 95% CI = 1.1-2.4) were directly associated with dementia, with a trend by years of education (P(trend) = 0.02) but not social class. Poor self-perceived health (versus good) increased the risk for incident dementia (OR = 3.9, 95% CI = 2.2-6.9). Alcohol and smoking (never, past and current) were neither strongly protective nor predictive. Stroke was strongly related to incident dementia (OR = 2.1, 95% CI = 1.1-4.2), as was Parkinson's disease (OR = 3.5, 95% CI = 1.3-9.3), and exposure to general anaesthesia (GA) was inversely associated with dementia development (OR = 0.6, 95% CI = 0.4-0.9, with a trend with increasing GA exposure; P = 0.003). In this large multicentre and long-term population-based study, some well-known risk factors for dementia, of vascular and Alzheimer's type, are confirmed but not others. The association between self-perceived health-a robust predictor of later health outcomes-and incident dementia, independently of other potential risks, warrants further study.