I think stats is the culprit. Or lack of it. If there's one area of study that ties together observations and models, it's stats. But stats suffers from the same problem that programming does: people think it's a tool that they just need to know a little bit of, so they don't defer to the proper authorities in the area.

Of course this is just anecdotes, but I do have a friend who did stats as a phd and this is the attitude that seems to be common: stats (and code) are minor tools that won't affect the outcome of our experiment.