[glesica]

I agree with your overall point, but I would walk it back just a notch by pointing out that being forced (or at least encouraged) to have a theory first helps to prevent correlation "fishing". If you accidentally stumble upon some really interesting relationship, then that is great and ideally would be shared with the world. But if that sort of thing is allowed to be published without any scrutiny, the incentive becomes to just throw tests at a dataset until it yields something that appears interesting. If you perform enough random tests on a dataset, you are very, very likely to eventually find something "unusual" or "interesting".

[dpark]

Sure. Anyone with the most basic stats knowledge knows that if you dig long enough you'll find random correlations. That's solved by demanding evidence be repeatable rather than requiring the evidence fit into a nice theory.

[glesica]

But asking for a theory can help the situation, that was my point. Of course demanding a theory even in the face of repeated findings isn't helpful, but encouraging a theory, or giving "bonus points" for a theory is, I think, a perfectly reasonable thing to do.