Except that you are not picked from the whole population at large, but from a subset of it. Namely the subset that (weakly) believes they have dementia. This changes the incidence rate (or likelihood function) probably by at least an order of magnitude.