You're right, many of these studies use weak methods to infer causality that don't fully rule out that possibility. Usually the problem is that they fail to acc

You're right, many of these studies use weak methods to infer causality that don't fully rule out that possibility. Usually the problem is that they fail to account for poor diagnosis accuracy - someone with (for instance) active schizophrenia might not be diagnosed for quite some time after it presents, and any window where the person is undiagnosed but symptomatic is a period where self medication of symptoms could appear later in the data as having "caused" the symptoms when they are finally diagnosed. I don't think this is typically deliberate, but it's sloppy, at the very least confidence levels should be adjusted to account for assumed error rates. Some studies fall apart completely (in the sense that you cannot reject the hypothesis that increased marijuana use is driven by current symptom strength, not the other way around) under reasonable assumptions about how correlated symptom strength and diagnosis probability are. I don't think this is just a theoretical problem, either. At least anecdotally, most of the people that I've known with serious mental illnesses like schizophrenia were undiagnosed for many, many years before finally figuring it out, despite clearly having a lot that was "wrong". There are massive stigmas that make sufferers hesitant to disclose the true severity of their symptoms, even when asked, until they become unbearable.