Correct, the study design in the original article cannot answer causal questions (although of course it can generate hypotheses). For example, if an omniscience

Correct, the study design in the original article cannot answer causal questions (although of course it can generate hypotheses). For example, if an omniscience informed us "actually, people lose sleep in middle age as a consequence of the early stages of dementia", that would make sense as well. The best causal analysis is interventional, but many times (such as in this case) it would be both extremely challenging and likely hard to do in any ethical fashion (I can't imagine how it could be done, but perhaps someone could). Short of that, you're left with instrumental variables. Genetics offers a very nice instrumental variable, since the causal arrow largely goes one way if you setup your study design carefully. That is generally performed as Mendelian randomization, which has several assumptions but, if met, has nice properties. https://en.wikipedia.org/wiki/Mendelian_randomization https://en.wikipedia.org/wiki/Mendelian_randomization