Blanchard joins calls for Structural Econometric Models to be brought in from the cold

Mainly for economists

Ever since I started blogging I have written posts on macroeconomic methodology. One objective was to try and convince fellow macroeconomists that Structural Econometric Models (SEMs), with their ad hoc blend of theory and data fitting, were not some old fashioned dinosaur, but a perfectly viable way to do macroeconomics and macroeconomic policy. I wrote this with the experience of having built and published papers with both SEMs and DSGE models.

Olivier Blanchard’s third post on DSGE models does exactly the same thing. The only slight confusion is that he calls them ‘policy models’, but when he writes

“Models in this class should fit the main characteristics of the data, including dynamics, and allow for policy analysis and counterfactuals.”

he can only mean SEMs. [1] I prefer SEMs to policy models because SEMs describe what is in the tin: structural because they utilise lots of theory, but econometric because they try and match the data.

In a tweet, Noah Smith says he is puzzled. “What else is the point of DSGEs??” besides advising policy he asks? This post tries to help him and others see how the two classes of model can work together.

The way I would estimate a SEM today (but not necessarily the only valid way) would be to start with an elaborate DSGE model. But rather than estimate this model using Bayesian methods, I would use it as a theoretical template with which to start econometric work, either on an equation by equation basis or as a set of sub-systems. Where lag structures or cross equation restrictions were clearly rejected by the data, I would change the model to more closely match the data. If some variables had strong power in explaining others but were not in the DSGE specification, but I could think of reasons for a causal relationship (i.e. why the DSGE specification was inadequate), I would include them in the model. That would become the SEM. [2]

If that sounds terribly ad hoc to you, that is right. SEMs are an eclectic mix of theory and data. But SEMs will still be useful to academics and policymakers who want to work with a model that is reasonably close to the data. What those I call DSGE purists have to admit is that because DSGE models do not match the data in many respects, they are misspecified and therefore any policy advice from them is invalid. The fact that you can be sure they satisfy the Lucas critique is not sufficient compensation for this misspecification.

By setting the relationship between a DSGE and a SEM in the way I have, it makes it clear why both types of model will continue to be used, and how SEMs can take their theoretical lead from DSGE models. SEMs are also useful for DSGE model development because their departures from DSGEs provide a whole list of potential puzzles for DSGE theorists to investigate. Maybe one day DSGE will get so good at matching the data that we no longer need SEMs, but we are a long way from that.

Will what Blanchard and I call for happen? It already does to a large extent at the Fed: as Blanchard says what is effectively their main model is a SEM. The Bank of England uses a DSGE model, and the MPC would get more useful advice from its staff if this was replaced by a SEM. The real problem is with academics, and in particular (as Blanchard again identified in an earlier post) journal editors. Of course most academics will go on using DSGE, and I have no problem with that. But the few who do instead decide to use a SEM should not be automatically shut out from the pages of the top journals. They would be at present, and I’m not confident - even with Blanchard’s intervention - that this is going to change anytime soon.

[1] What Ray Fair, longtime builder and user of his own SEM, calls Cowles Commission models.

[2] Something like this could have happened when the Bank of England built BEQM, a model I was consultant on. Instead the Bank chose a core/periphery structure which was interesting, but ultimately too complex even for the economists at the Bank.

More on Stock-Flow Consistent models

This is a follow-up to this post, but which is prompted by this Bank of England paper, which builds a stock-flow consistent model for the UK. If you are not familiar with the term ‘stock-flow consistent’ (SFC) then read on, because in a sense this post is all about why I think the way the authors and others define this class of models is misleading.

SFC models are popular with Post-Keynesians, and the definition you find on Wikipedia is “a family of macroeconomic models based on a rigorous accounting framework, which guarantees a correct and comprehensive integration of all the flows and the stocks of an economy.” Now I suspect any mainstream macroeconomists would immediately respond that any DSGE model is also stock-flow consistent in this sense. This point is made in a post by Noah Smith, and it is completely valid, although otherwise I think his account of the weaknesses of SFC models is wide of the mark.

If you think this is a trivial debate about titles, take this description of the pros and cons of SFC compared to DSGE models taken from the paper:

Take the cons (merits of DSGE compared to SFC) first. Number one is almost definitional: DSGE models have to be microfounded, but SFC models start with aggregate relationships. But that is not a defining feature of SFC models, because there is a long tradition of macro modelling that is not microfounded but starts with aggregates, a tradition that begins well before DSGEs with the simultaneous creation of national accounts data, econometrics and Keynesian economics. This tradition goes by many names: ‘Structural Econometric Models’ (SEMs), ‘Cowles Commission’ (favoured by Ray Fair) or most recently ‘policy models’ (see Blanchard). I’ll just call them aggregate models here.

A key question, therefore, is what marks SFC models out from other aggregate models? The authors obviously think there is something, because of their second ‘con’. The third and fourth ‘cons’ are common to many large SEMs. (I once wrote a paper on how to mitigate the first of these problems.) The fifth ‘con’ just follows from the first.

At first sight the sixth ‘con’ does the same, but I would argue that if there is anything that characterises SFCs among aggregate models it is this. Aggregate models would generally involve an extensive discussion of the theoretical origins of the relationships they used, but if this paper is anything to go by that is less true for SFCs. If you think this last point is unfair, look at the discussion of the consumption function (before equation 4).

This failure to acknowledge the existence of other aggregate models is even more apparent among the ‘pros’. The first and second can be true for any model, including a DSGE model, but the third is critical. It is true, but again it is also true for many aggregate and some DSGE models. As I argue in my previous post, the key point about the archetypal DSGE model is that it does not need to track household wealth, because there is no attempt by consumers (given the theory) to achieve some target value of wealth.

The fourth is true for any model, including DSGE models. The fifth is true for any aggregate model as long as expections variables are explicitly identified. The sixth is also almost bound to be true of any aggregate model, because starting with aggregates and being eclectic (and potentially internally inconsistent) with theory allows you to more closely match the data than DSGEs.

To summarise, if you were to ask how this model compares to other aggregate (non-microfounded) models, the answer would probably be that it takes theory less seriously and it has a rather elaborate financial side.

The New Classical counter revolution had many good and bad consequences, but one of the undesirable consequences was, it seems, to define the equivalent of a year zero in macroeconomics, where nothing that was not in the New Classical tradition created before (or even after) this revolution is deemed to exist. The same should not be true for heterodox economists. If you are going to effectively return to a pre-DSGE tradition, please do not pretend that tradition did not exist.

There is a well known UK professor of econometrics who was very fond of admonishing authors who failed to cite work that they were either extending or just copying. The intention here is not just to do the same. One of the big dangers with any kind of elaborate aggregate model is that you can get bizarre model properties from not thinking enough about the theory, or imposing enough because of the theory. Knowing some of the authors I doubt that has happened in this case. But it would be a mistake for others to believe that the properties of their model show the importance of accounting rather than the theory they have used.