3 ms·

Thanks for this. As always with big population studies, there's a huge risk of getting your result by failing to control for something. Or, conversely, missing

Thanks for this. As always with big population studies, there's a huge risk of getting your result by failing to control for something. Or, conversely, missing a real relationship by controlling for something else it predicts. Or for most studies, both. Probably you missed some key confounder, and probably you also controlled away some of the real effects. So "correlation found" ends up meaning "the missed confounders beat out the improper controls". (Does being a 'night owl' cause depression, which shortens lifespans? It certainly sounds plausible, so controlling for psychological disorders raises an eyebrow. And so on, for a dozen other controls.)

bunderbunder 9y ago
In turn, thank you. And yeah, one of the other things that I find annoying about these studies that follow this paradigm of, "Control for everything you can think of, and then make a big deal of any residual effect, as if it's somehow implying an independent causative relationship," is that they don't seem to really distinguish between probable mediators and probable confounders. Which makes sense, I suppose. If you're just on an easter egg hunt for p-values, then I guess you don't really need to care about how an analysis decision might affect the practical utility or theoretical value of your findings.