"Health" is a loosely defined word and there are many goals and metrics. We can definitely study some metrics of "health" and study them with high confidence if

"Health" is a loosely defined word and there are many goals and metrics. We can definitely study some metrics of "health" and study them with high confidence if the goal is well defined and it can be monitored with fine granularity. For example, if many people exercise and then the blood glucose drops for the next day, we can see its good effect, and probably it's easier to figure out what hormonal mechanism or neural pathway affect them. If the goal is higher-level health issue, like longevity, it is trickier because of, as you said, many confounding issues. I suspect there would be cyclic causality between exercise and longevity/overall health, and it would be quite difficult to separate the cause and the effect. IMO, maybe the best way is, making educated guess based on the studies with clearer metrics.