He writes,

So we choose a microfoundation because it gives us the aggregate answer suggested by the data, and not because of evidence that this microfoundation is appropriate. We then insist that everything in that model has to be consistent with this microfoundation, and that our model has been built from only thinking about what individual agents do. Have we just replaced ad hoc with post hoc?

Read the whole thing for context. Pointer from Mark Thoma. My comments:

1. My joke is that the Phillips Curve went from being an empirical finding in search of theory to a theory in search of empirical support.

2. Empirically, the only thing of which we can really be sure is that the best predictor of inflation is the lagged dependent variable (this is hardly the only macroeconomic variable to which this applies). After that, the magnitude (and even the sign) of the relationship between wage growth and employment varies quite a bit depending on how you do your specification searching.

3. Once again, let my plug my macro memoir.