OBJECTIVE: To estimate the percentage reduction in incidence of dementia that would be obtained if specific risk factors were eliminated. DESIGN: Prospective seven year cohort study. SETTING: General population, Montpellier, France. PARTICIPANTS: 1433 people aged over 65 with a mean baseline age of 72.5 (SD 5.1) years. MAIN OUTCOME MEASURES: Diagnosis of mild cognitive impairment or dementia established by a standardised neurological examination. RESULTS: Cox models were constructed to derive hazard ratios and determine confounding and interaction effects for potentially modifiable risk factors for dementia. Mean percentage population attributable fractions were calculated with 95% confidence intervals derived from bootstrapping for seven year incidence of mild cognitive impairment or dementia. The final model retained crystallised intelligence (population attributable fraction 18.11%, 95% confidence interval 10.91% to 25.42%), depression (10.31%, 3.66% to 17.17%), fruit and vegetable consumption (6.46%, 0.15% to 13.06%), diabetes (4.88%, 1.87% to 7.98%), and apolipoprotein E epsilon4 allele (7.11%, 2.44% to 11.98%). CONCLUSIONS: Increasing crystallised intelligence and fruit and vegetable consumption and eliminating depression and diabetes are likely to have the biggest impact on reducing the incidence of dementia, outweighing even the effect of removing the principal known genetic risk factor. Although causal relations cannot be concluded with certainty, the study suggests priorities that may inform public health programmes.