It is relatively straightforward to control for things like "exercise from other activities" and underlying comorbidities in an attempt to compare "like with like" even for non-randomized studies.

One should also keep in mind that while correlation =/= causation, showing a strong correlation is a step along the path toward showing causation.