My irritation with statistics is that you are basically testing the accuracy of your model. If you thoroughly understand the subject of your model, the model will likely reflect reality. If you don't thoroughly understand the subject of your model, your statistical analysis will discourage you from understanding it further.
That's true. It's been postulated by some statisticians that a perfectly controlled experiment quickly becomes tautological. You have to live with ambiguity and pay of that is choosing the model you think is most appropriate. But the thing you can't do if you want to learn anything worthwhile is to choose your study question post hoc, as they've very clearly done here. You see that with drug studies all the time: "Well it didn't work in the way we said it would, but if you only look at white women who are between 35 and 47 with a history of familial diabetes and who spend more than 6 hours a day watching Real Housewives reruns it's a home run! (n = 4 out of 1800 studied)"