Simon Wren-Lewis responded to my post on microfoundations in macro, wondering what was wrong with *mainly *micro-founded macro [is this what his blog name really refers to?] if small ad-hoc interventions helped the model fit the data better.

The presumption in this question must be that there isn’t a modification to the microfoundations that would also help the model describe the data better (than the solely microfounded model you started with). I’m not sure about Simon’s specific example, but this is a legitimate enough question.

If the objective is to describe the data better, perhaps also to forecast the data better, then what is wrong with this is that you can do better still, and estimate a VAR – a system of equations where everything is allowed to be a function of everything else lagged.

In fact, if you want to take this microfoundations plus ad-hoc modification model to the data, and construct what Sims (he cites Haavelmo in his Nobel lecture for this project) calls a ‘probability model’, then what is wrong with Simon’s proposed model is that it incorporates theoretically unsubstantiated – in his language ‘incredible’ – identification restrictions.

To re-emphasise a point in my previous post, you *can* use this model for policy experiments. In other words, a Sims like model is an attractive option even if you aren’t solely trying to explain in sample, or forecast the data out of sample. Provided, that is, that you convince yourself of the following: that the experiment does *not* constitute so much of a departure from the past that the statistical laws of motion you estimate will move around so much in response to the policy as to invalidate the inference about what you should do that you made from the original model. This approach resembles in part the one Simon toys with, weighing the costs of Lucas-Critique problems against the possible benefits of describing the data better. Only here we are talking about an entirely statistically founded, rather than microfounded model. A topical example of this approach in action is the literature using VARs to measure the fiscal multiplier, including work by Romer and Romer, Cloyne, Blanchard and Perotti, Caldera and others.

Another way to put Sims’ point would be this [partly inspired by a Nick Rowe post]. Once you make that modification, you don’t know what you have any more. You might hope that you have a system of equations that describes what consumers and firms *do*, and one that does it in a more data-congruent way than before. But in fact your hopes might be dashed. Really all you have is a system of equations linking numbers that the statistics agency collects. And no good reason to have paired down that system of equations by eliminating variables that could appear in them.

Commenting on Simon’s post, Noah Smith writes ‘YES YES A THOUSAND TIMES YES’ at the suggestion that you could add a little statistical realism to the microfounded model. Noah is extremely sceptical of microfoundations. So much so that he requests a post to explain why they might have any merit at all. So, he should be saying: NO NO GET RID OF ALL THE MOTHER&&&&&&G MICROFOUNDATIONS WHILE YOU ARE AT IT.

As I said in my previous post, ad-hoc modifications seem attractive if they are a guess at what a microfounded model would look like, and you are a policymaker who can’t wait, and you find a way to assess the Lucas-Critique errors you might be making. Or you are a generous researcher who wants to try to help someone cleverer or more persistent to confirm your guess. Or you want to convince someone of the same ilk that it’s worth trying to confirm the guess, because you demonstrate that (conditional on the guess proving correct) there is some great prize at stake, some significant revision of our diagnosis of past events, or policy prescription for responding to some future event.

Yates:

“… ad-hoc modifications seem attractive if they are a guess at what a microfounded model would look like, …”

No.

The modifications are useful if they are a good guess of how the actual economy really behaves. You are making a further assumption that a microfounded model is the One True Model. This additional assumption is entirely unjustified. This is the same error of circular reasoning you made in your last post in appealing to the Lucas Critique in defending microfoundations.

Let me elaborate because yesterday you were unable to see the error.

A simple statement of the Lucas critique would be that if agents’ behaviour changes in response to policy changes, and if our models do not incorporate such behaviour, then they will not be useful for policy evaluation in actual economies. So far, so good, but note that this is an if-then statement. Whether it is true of the real world, and if so, to what extent, is an empirical question.

An implication is that a fully specified model that correctly specifies the behaviour of the actual economy will need to correctly specify how agents react to policy changes and will thus be immune to the Lucas critique. The circularity comes now when one simply assumes that a microfounded model (based on tastes, technology, endowments, and rational expectations) is such a model, and then argues that such microfounded models are immune to the policy critique.