Savage: http://arxiv.org/abs/1609.01774
Aronow v. King

There are several good quotes in there. Two of my favorites:
“Put simply: KR’s proposed heuristic – comparing the robust standard error to the classical standard error – does worse than provide no information about misspecification in this setting: it provides misinformation.”
“To summarize, KR’s recommendations demand that we specify a full parametric model that reduces the phenomenological description of the observable data to a finitedimensional function. KR then advocate that we readjust the model based on the data until we find a model that fits. That is, when the data does not match this infinitely restrictive model, we should change the assumptions until it does.”