Saturday, March 30, 2013

A timid defense of DSGE models

An economic model (see here)
Image source

This post is for economists and/or econ bloggers and/or people with opinions about this issue of mathematical modeling in macroeconomics. So that might be one or two of my usual readers (I won't mention what that is in percentage terms).

Noah Smith has a good post about some problems with DSGE models, particularly when they are solved using linearization methods (I like Smith's blog; I don't like his comment sections as much). He is not the only person with complaints about the workhorse approach to macroeconomic thinking. Stephen Gordon also has a very nice post, again mainly about linearized DSGE. My post is not about Smith's or Gordon's notes; it's a more general comment on this occasionally popular topic.

I don't want to wade too deeply into this debate for the following reasons:

  • Much of the debate is over my head--not technically speaking, but in terms of knowing how to do science well.
  • The blogosphere is a poor place to have this debate.
  • Many of the people involved (but not Smith or Gordon) probably don't know enough about DSGE models to draw useful conclusions (i.e., if you don't know how to build or solve one, you may be unable to make a highly productive contribution to the debate).
  • Political preferences are involved, which makes it hard to tell whether methodology is really the issue.
  • I don't have a strong enough opinion to justify dragging myself and charitable readers through the weeds, aside from thinking that DSGE models are useful tools.
  • I think linearized models are harder to defend than those with global solutions.

But here's what I will say:

  • Not all DSGE models are solved with linearization. Yes, the New Keynesian ones typically are, and those are the ones informing policy the most--and there aren't a lot of great ways around that. But lots of DSGE models can be solved globally without using Taylor-type approximations, in which case approximation error is far smaller (and is basically determined by Curse of Dimensionality and/or numerical precision issues).
  • I don't like it when people act like we face a big choice between using DSGE or doing empirical work. That's a false dilemma. First, why not do both? Second, that distinction does not always exist. This deserves its own post, but for now: DSGE models are just systems of equations, and they can be estimated; and it's not obvious to me that estimating a DSGE system is any more silly than assuming the world resembles the atheoretic linear functional form we use in purely empirical work. I think we all know that there is no such thing as simply "letting the data speak."
  • In this debate, and in debates about formal modeling in general, some critics tend to make the mistake of thinking that a modeler believes that the assumptions and simplifications of the model are true in the real world. They also like to accuse macroeconomists of thinking they are physicists. Those are straw men. Take your straw man and go home.
  • We must get econ pundits to understand that we're all using models, including non-economist bloggers, even if they're not written down as mathematical expressions. Writing a model down in its entirety so that its assumptions are made explicit and its internal workings can be examined by anyone is an act of intellectual humility. It is baffling to me that people who write down their models formally so we can all argue about them are supposedly worse and more arrogant than those who think they can identify a narrative model's assumptions and keep it internally consistent.
  • There are limits to what the "credibility revolution" techniques from applied micro can do for macro. There is little or no clean identification to be had at the aggregate level, and extrapolating empirical results from small regions to the aggregate level can be misleading. These approaches can definitely shed light on macro topics (the excellent Mian and Sufi papers come to mind), but I think it's a mistake to assume that they are sufficient for all macro questions.
  • In typical empirical work and heuristic/narrative theorizing, it's really difficult to avoid partial equilibrium reasoning. A DSGE model, even a very simple one, has more moving parts than my mind alone can keep track of, and it forces agents to obey resource constraints in a way that is really difficult without formality. You don't have to believe literally in "general equilibrium" to appreciate how GE models allow you to think about feedback mechanisms.
For what it's worth, I have personally learned a lot from using DSGE models. Much of the intuition I use to think about macroeconomics originally came from some DSGE model. When I work with them, I frequently have insights that force me to change the way I think. It's possible that those DSGE critics who have not used them could benefit from spending some time with them to see if they learn anything that was hard to think about before. Maybe there are a lot of longtime DSGE users who have never learned anything from them, but I doubt it.

A recent example: Last week, I was messing with an augmented version of this DSGE model (in my version, there is also a corporate sector and some other stuff). Households choose whether to be workers or entrepreneurs. I was messing with the parameter that governs entrepreneurial productivity (specifically, the scale parameter of the Pareto TFP process). I noticed that sometimes when I increase that parameter and solve the model, the share of households that choose to be entrepreneurs fell. My initial expectation was that increasing how much entrepreneurs can produce with a given level of inputs would make entrepreneurship more enticing (relative to earning a wage) and cause more people to do it. But that's partial equilibrium reasoning. I pushed on the model a little and figured out that it depends on the wealth distribution and capital intensity. Roughly speaking, if entrepreneurs consume most of their new income (rather than saving it), and if production is more capital intensive, entrepreneurship will rise: higher TFP raises capital demand, and people don't save enough to mitigate the upward pressure on the interest rate; and wages and interest rates move in opposite directions when production is Cobb-Douglas-esque, so worker income falls while entrepreneurial income rises. On the other hand, if people save most of their new income and production is less capital intensive, the supply of savings can rise enough to mitigate upward pressure on the interest rate (and downward pressure on the wage). It is possible for wages to rise enough to induce some people to abandon entrepreneurship and become workers.

A lot of people would have seen that coming, and I probably would have if I'd thought it through carefully. But there were other things on my mind. The point is that the model forced me to consider the implications of my ideas and to recognize how conditions in one part of the economy affect things that happen elsewhere. This kind of discipline is important when arguing about the potential effects of this or that policy.

I think most people, including me, are used to thinking in partial equilibrium. DSGE models, at the very least, can force us to think a little more broadly.

I don't see why we can't allow for a wide range of methodological approaches in economics, and even in macroeconomics. I suspect that part of the problem is that recent events have raised everyone's level of interest in macroeconomics. People working in other economics fields, along with a lot of noneconomists, want to get in on the action and be part of the conversation. That's a good thing. But some of them can look a bit like Monday Morning Quarterbacks, making big claims about the failure of the field and recommending all sorts of changes to the standard toolkit as if macro can be approached in the same way as their home field. I think it would be a bad idea to discard the DSGE framework, even as I also think that drill-down empirical work is increasingly important (and possible) in macroeconomics (as should be evident by my other posts).

I also suspect that, for a few people, antipathy towards workhorse macro models is driven by politics. This debate often arises in the context of discussions about fiscal multipliers (ha!) or optimal tax policy, and it sometimes appears that people who don't like the political implications of a paper respond to it by rejecting the entire modeling approach. If that's going on, well, we shouldn't let it affect real economics work.

Update: some related thoughts here.


  1. Three cheers for this. Given the arrogant tone from most critics, I wouldn't have minded a more "robust" approach, but this will do nicely. Kudos.

  2. "I don't see why we can't allow for a wide range of methodological approaches in economics, and even in macroeconomics."

    Especially in macroeconomics!

    I was having this exact conversation with another economist in my email this morning. Good post. I think the determination to act like we have a "general model" can hurt both our analysis, and how persuasive we are ... although I suspect this fault is mainly on the side of people who work in the macro area.

  3. Nice post, Ryan. I think/hope we as a discipline are moving away from linearization to more robust approximations. And there are still a lot of very interesting models that cannot be solved and simulated with the current set of DSGE tools.

  4. Matt--Good point, and thanks for sharing the post.

    Kerk--Thanks for teaching me about DSGE models! I got hooked on this stuff when we played around with your immigration model. It felt like Sim City. Maybe that's the real reason I still like DSGE. I definitely feel better about non-linearized models, though it seems like we don't have a choice with some setups.

  5. Hey, I am new reader here, coming to this via your comment on Noah's recent post about maths.

    As a disclaimer, I am one of those people who went through a higher degree in economics and finance and still aren't particularly good at maths. So feel free to ignore what I am going to say since "Many of the people involved probably don't know enough about DSGE models to draw useful conclusions" (I disagree, btw. You don't always need to be an expert to comment on things) but maybe you'll be interested...

    When you speak about your DGSE model and household choice between salaried work/entrepreneurship (a very interesting topic), you end up relying on the fact that savings influence interest rates/investment.

    Sorry but is that entirely not true? No one needs to save money before credit is created. You just need a bank willing to lend and, hey presto, you got some lending/investment/consumption with no savings prior.

    Interest rates are primarily driven by the ST interest rates (ie. the CB) and inflation expectation. Plus risk premiums/competition amongst banks. Not savings.

    How would I translate that into a DGSE?

  6. Hi Frederic, thanks for stopping by. Are you asserting that the quantity of savings has no effect on interest rates? I think most people would be skeptical of that claim.

    I didn't say that one must be an expert to comment on modeling approaches. But dsge is just a tool, and I do think that people who have never used that tool might not be in a great position to claim that it's not useful.

  7. Excellent post, Ryan. Lots of good stuff here.

    I want to push back on your related comment at MR ( and pat you on the back too (so stay with me). You said:

    "Some of the participants (e.g. Yglesias) appear to have limited experience with DSGE, so it’s interesting that they feel confident enough to opine on the subject. Some of them demonstrate a total lack of familiarity with the massive task of trying to model an economy, as when Matthew Klein tweeted “wouldn’t a model that genuinely reflects how world works predict better than other models?” So we have journalists who believe a totally realistic model is attainable. If that is their belief, they have nothing useful to contribute to this particular debate. "

    There are few corners of mainstream macro that are less accessible to econ journalists that DSGE models. General equilibrium blows the mind of most econ PhD students (as you noted partial equilibrium is easier and more applicable/natural to most). Yet anyone with econ 101 can manipulate an IS-LM model (maybe not as magically as Krugman, but still). The predictions of DSGE models can be un-inintuitive, opaque, and emanating from a nest of wack-a-doodle assumptions. Now some of this is fine, all models are wrong. But I look at a lot of real-world data that calls strong intertemporal smoothing motives, high interest rate sensitivity, and low risk aversion (not to mention rationality) into question ... if those fail then most DSGE models are an exercise in computational programming.

    All that said, I think DSGE is an important tool to keep plugging away on. And your points about being concrete and (kinda) transparent in assumptions are important. But DSGE (and all of economic modeling) needs better communication. I am appreciative of practioners who blog on it in an accessible way. I don't have enough knowledge of the DSGE methods to do it well. But your post, like Chris House's and Tyler Cowen's make me happy. If journalists and non-economists, do not understand our methods it the OUR (economists) fault.

    1. Thanks for commenting! Those are great points. I was more pointed in the MR comment than usual. The fault for lack of appreciation of DSGE among non-economists probably lies almost entirely with the economists. I hope I can at least get a few people to recognize that any kind of macro requires some simplification. And, as I said in a comment above, I don't mean to suggest that lack of deep familiarity with DSGE models disqualifies a person from commenting about it. I just want to make sure we all think about how they relate to other forms of analysis.

      Oddly enough, I've thought a lot about your comment on interest rate sensitivity (I think you've mentioned this on Twitter before). Things like that definitely challenge the paradigm.

  8. Not a fan of DSGE, but good point about models. Even James Tobin said in his Nobel lecture that:

    Almost everyone thinks about the economy, tries to understand it, and has opinions how to improve its performance. Anyone who does so uses a model, even if it is vague and informal.

  9. Hi Ryan,

    I'm in a similar boat to Frederic - an advanced degree in economics and finance but not really comfortable with econometrics. I think it very much depends on what sort of mathematical models you are used to. Not being primarily an economist, DSGE models are unfamiliar to me, but I'm at home with more financial models.

    I don't think people who don't use mathematical modelling necessarily fail to think through or write down their implied models. There is more than one language. The application of logic requires mathematical thinking, but it doesn't require mathematical models.

    1. Thanks for commenting Frances! I agree. I do not mean to suggest that non-DSGE approaches are necessarily less thorough, rigorous, or useful. DSGE works for me, but I've read lots of economic analysis (including by you!) that does not rely on them yet conveys insights just as well.

  10. "We must get econ pundits to understand that we're all using models, including non-economist bloggers, even if they're not written down as mathematical expressions."

    This just simply isn't true. It's an assertion, often made by economists, that is based on a very primitive understanding of how people think.

    Personally, I do not use models and I get by just fine.

    1. This looks to be a semantic point. I'm not super interested in that kind of debate; my argument is simply that all economic reasoning requires some assumption and simplification; I think a formal mathematical model is one way to achieve this in a portable way, but there are others. If you're doing analysis without any kind of assumption or simplification, you must have a powerful mind indeed.

    2. Now you're equating "modelling" with "assumptions" and "simplifications". I think you're just muddying the issues massively now.

      Saying that economic reasoning is based on "assumptions" and "simplifications" is so obvious as to be almost a tautology. Saying that all economics is based on models is entirely different.

    3. Ok, so you are making a semantic point. Point taken. I should find a more precise definition of "model".

      If that's the main thing you got out of my post, I did a pretty poor job of conveying what I was thinking.

    4. It was the only part that interested me and it was picked up on a blog I follow. It's also something that I hear repeated all the time and it is never challenged.

      As to DSGE models, I couldn't care less. Anything that uses representative agents is not truly microfounded, so they don't even work on their own terms.

    5. I don't hear this repeated often--but I wonder how much of it really is about semantics. It would be silly to say that people are walking around with formal mathematical models in their heads. But in my view, a model is just a simplification of reality from which the user draws inference, so as you note it's a tautology to say that we're all using models (by my definition). I agree--it seems obvious; but some (not all) of the critiques of DSGE models run along the lines of "it makes this and that unrealistic assumption," which in general is a bad reason (on its own) to reject a model.

    6. I think those critiques are valid too. Models based on unrealistic assumptions will not tell you anything relevant about the real world. We know from experiments that people do not act in line with contemporary microeconomics. And any science that ignores experimental results is not really a science at all.

      But the problems with DSGE in particular and modelling in general are far more profound than this. DSGE, as I said, actually doesn't even work on its own terms. I.e. it purports to be microfounded but actually uses aggregate concepts like the representative agents. Also, I believe that DSGE models fall apart logically in the face of the Sonnenschein-Mantel-Debreu theorem. So, even if you play the game that economists play and wish away experimental results, the DSGE models still don't stand on their own terms.

    7. "Models based on unrealistic assumptions will not tell you anything relevant about the real world."

      Can you point me to some models that are based on realistic assumptions? It seems to me that leaving ANYTHING out of the model amounts to an unrealistic assumption, and since models by definition (well, my definition) leave things out, we're defeating the whole purpose of modeling and economic analysis generally if we say that no unrealistic assumptions are allowed.

      As for the rep agent thing--yes, there is no such thing as a truly microfounded model. I see rep agent as one dimension along which people simplify; other kinds of models simplify along other dimensions (also--not all DSGE models use rep agent). Models with different characteristics allow us to think about different things; yes, they all leave things out, but the quest for a single unifying model that will let us think about anything is futile.

      I would not advocate total reliance on DSGE, or on any class of model for that matter. It seems to me that good analysis requires us to draw on insights from a variety of tools.

    8. 1) But I don't support modelling in economics so why would I need to point you to a "realistic" model. Anyway, leaving out something is substantially different from inserting postulates that are disproved empirically. Real science often leaves out complexities and makes simplifications. It does not, however, include postulates that are disproven by experiment. That would be enough to destroy a scientist's career.

      2) So you agree that DSGE fails even when taken on its own terms? I find it odd that you would then champion it. Also, you didn't mention the SMD theorem. Is a model that is internally inconsistent not also invalid? I mean, if experimental disproof doesn't count for you as an invalidation surely logical inconsistency must, no?