Objective: A previous study reported a U-shaped association between fasting insulin and dementia in a 5-year follow-up of a male cohort, which was now re-investigated in a representative population of women followed over 34 years, and taking into account the incidence of diabetes. Methods: Fasting values for insulin and glucose were obtained from serum samples in 1212 non-diabetic women aged 38-60 at the 1968 baseline. Risk of dementia was assessed by Cox proportional hazard regression adjusting for insulin, glucose, and other covariates, and, in a second model, after censoring for incident cases of diabetes. Incident diabetes was considered as a third endpoint, for comparison with dementia. Results: Over 34 years, we observed 142 incident cases of dementia. The low tertile of insulin displayed excess risk for dementia, hazard ratio (HR) = 2.34, 95% confidence interval = (1.52, 3.58), compared to the medium tertile, but the high tertile of insulin did not, HR = 1.28 (0.81, 2.03). These associations were also seen for dementia without diabetes comorbidity. In contrast, high but not low insulin predicted incident diabetes (115 cases), HR = 1.70 (1.08, 2.68) and HR = 0.76 (0.43, 1.37), respectively. Conclusions: Our results revealed a non-linear association between fasting serum insulin and dementia in a female population, with high risk at low insulin values that was not attributable to preclinical dementia or impaired insulin secretion. This condition suggests a new pathway to dementia, which differs from the metabolic pathway involving diabetes.