Higher midlife body mass index (BMI) is suggested to increase the risk of dementia, but weight loss during the preclinical dementia phase may mask such effects. We examined this hypothesis in 1,349,857 dementia-free participants from 39 cohort studies. BMI was assessed at baseline. Dementia was ascertained at follow-up using linkage to electronic health records (N = 6894). We assumed BMI is little affected by preclinical dementia when assessed decades before dementia onset and much affected when assessed nearer diagnosis. Hazard ratios per 5-kg/m2 increase in BMI for dementia were 0.71 (95% confidence interval = 0.66–0.77), 0.94 (0.89–0.99), and 1.16 (1.05–1.27) when BMI was assessed 10 years, 10-20 years, and >20 years before dementia diagnosis. The association between BMI and dementia is likely to be attributable to two different processes: a harmful effect of higher BMI, which is observable in long follow-up, and a reverse-causation effect that makes a higher BMI to appear protective when the follow-up is short. •Data from 1.3 million adults from 39 prospective cohort studies were pooled for the analyses.•Higher BMI was associated with increased risk of dementia when the follow-up was long.•When follow-up was short, lower BMI was linked to increased dementia risk probably due to reverse causation.