I was asked to write a bit about microfoundations, in response to a couple of posts by Tony Yates and Simon Wren-Lewis, so just a quick note before I grade exams.
I've never much liked the word "microfoundations," which comes from the title of the Phelps volume. When the Phelps volume came out in 1970, micro and macro looked like they came from people living on different planets, and you had to convince people that it made sense to take ideas from Mars and use them on Venus. We know better now, of course, and economics is one unified whole. People who call themselves macroeconomists make liberal use of game theory, mechanism design, information theory, optimal growth theory, etc., etc. Solid economists worked all that stuff out for us to use, and it would be wasteful to leave it on the shelf.
So, it's not microfoundations, it's just economics - it's the foundation, the plumbing, the electrical work, the walls, the roof, etc. What people mean when they say: "you need some microfoundations," is really "you need more detail in there, so we can understand better what the mechanism is that is at work." We would like our models to be deep - everything explained nicely - but any modeler knows that a model must be simple to be useful. So, if there is too much detail we're in trouble. The modeler may have trouble understanding the damn thing, and he or she may not be able to explain how it works to anyone else.
So, where do you draw the line? How deep a structure do we need to say something useful? I just had a discussion with Noah Smith about Calvo pricing. For me, this is as case where we would like more detail. Typically, in a New Keynesian model, pricing is the friction that makes monetary policy matter. So, for that particular model, it's the process by which prices are set that we're worried about. That's where we want the detailed explanation of what's going on. But Calvo pricing is pretty crude. There are two technologies for changing prices. Either a firm can change a price at zero cost, or it's infinitely costly. And which technology the firm has is determined at random. Well, that leaves a lot of unanswered questions. Why would it be costly for a firm to change its price? Why can't the firm write contingent pricing rules? If it's costly to change prices, surely it must be costly for a firm to change other things, like employment. Why does the firm just serve all the demand the comes in the door when its price is fixed? Basically, there's a lot of unfinished business. We're concerned about how things might change if we put in more detail. Do all the results about monetary policy change or what?
Every piece of research has unfinished business. I have a working paper where quantitative easing (QE) matters, and it matters because short-maturity government debt is better collateral than long-maturity debt. What's it mean for an asset to be better collateral? There is limited commitment, and a borrower can run away with a fraction of an asset that is posted as collateral. What determines that fraction? It's just exogenous. That's a piece of unfinished business. I can think of reasons why that fraction could be endogenous in a deeper model, but I haven't worked that out yet.
This brings us to the Lucas critique. Lucas's ideas were related to earlier ideas that came out of the Cowles Foundation, including the original work on identification and structure. When I took Art Goldberger's econometrics class, he asked us one day to find a definition of structure and bring that into class the next time we met. We found 20 or 30 different definitions, some wildly different. Apparently structure is in the eye of the beholder. To give you an idea how thorny this is, I once had an argument with Bob Lucas about the money demand function, which he is very fond of. He was trying to tell me how remarkable it was that he had found a money demand specification that was stable for 100 years. My reply was something like: "Big deal, a money demand function is not structurally invariant to whatever policies we might want to think about using it for." He said something like: "Unfortunately I think the Lucas critique is used as a bludgeon to do away with ideas one doesn't like." I think he said that later in a public venue. So, even Lucas isn't sure what he thinks of the Lucas critique.
Ultimately, we know our models are going to be wrong. That is, to be useful, a model has to be simple, and simplicity implies it's wrong. Theory is obviously important. The data does not speak directly to us and, in contrast to what Simon seems to think, neither does the data speak to us when we filter it through a VAR - that's just another way to summarize the data. Theory gives us principles on which to organize how we think about the data, so that we're not totally lost.
It's hard to define what a good policy model is. Maybe we need different models for different situations - with the detail in different places. In any event, we hope a good economist knows a good policy model when he or she sees it. However, in some cases it may be difficult to evaluate a policy model's worth unless we actually experiment by putting it into practice.