Take a typical consumption function where consumption depends on current income and other things. Income is endogenous. In CC models using 2SLS, first stage regressors might include variables like government spending and tax rates, possibly lagged one quarter. Also, lagged endogenous variables might be used like lagged investment. If the error term in the consumption equation is serially correlated, it is easy to get rid of the serial correlation by estimating the serial correlation coefficients along with the structural coefficients in the equation. So assume that the remaining error term is iid. This error term is correlated with current income, but not with the first stage regressors, so consistent estimates can be obtained. This would not work and the equation would not be identified if all the first stage regressors were also explanatory variables in the equation, which is the identification criticism. However, it seems unlikely that all these variables are in the equation. Given that income is in the equation, why would government spending or tax rates or lagged investment also be in? In the CC framework, there are many zero restrictions for each structural equation, and so identification is rarely a problem. Theory rules out many variables per equation.
Pointer from Mark Thoma.
I am afraid that Ray Fair leaves out the main reason that I dismiss macroeconometric models, namely the “specification search” problem. As you can gather from the quoted paragraph, there are many ways to specify a macroeconometric model. Fair and other practitioners of his methods will try dozens of specifications before settling on an equation. As Edward Leamer pointed out in his book on specification searches, this process destroys the scientific/statistical basis of the model.
I have much more to say on this issue, both in my Science of Hubris paper and in my Memoirs of a Would-be Macroeconomist. In the latter, I recount the course that I took with Fair when he was a visiting professor at MIT.
Other remarks:
1. On DSGE, I think that the main vice is the “representative agent” consolidation. It completely goes against the specialization and trade way of thinking. Fighting the whole “representative agent” modeling approach is a major point of the Book of Arnold, or at least it is supposed to be. (I may have been too terse in the macro section of my first draft.)
2. VAR models are just a stupid waste of time. As I said in a previous post, we do not have the luxury of saying that we construct models that correspond with reality. What models do is allow us to describe what a possible world would look like, given the assumptions that are built into it. VAR models do not build in assumptions in any interesting way. That is claimed to be a feature, but in fact it is a huge bug.
I think that the project of building a model of the entire economy is unworkable, because the economy as whole consists of patterns of specialization and trade that are too complex to be captures in a model. But if you forced me to choose between VAR, DSGE, and the old-fashioned stuff Fair does, I would actually use that. At least his model can be used to make interesting statements about the relationship of assumptions to predicted outcomes. But that is all it is good for, and for my money you are just as well off making up something on the back of an envelope.
Isn’t another critique that when people decide to change their risk preference, or liquidity preference, or propensity to hold cash or to consume, or time structure of production, or animal spirits or whatever you prefer to call it, none of the other variables matter?
You say: “I think that the project of building a model of the entire economy is unworkable, because the economy as whole consists of patterns of specialization and trade that are too complex to be captures [sic] in a model.”
This makes me wonder, based on my background in physics and computers, just how big a model would actually be necessary. In physics you can glean quite a bit of information about a plasma that has 10^19 particles per cm^3 using a ratio of 10^6 or 10^7 real particles to one computer particle. We do it this way sometimes because the math (typically Gaussian stuff) breaks down for complicated plasma. What ratio would you need of “computer entrepreneur” or “computer customer” or “computer taxpayer” to the real thing before you started getting statistically useful results?
Admittedly in physics you have some nice clean rules you can apply to every particle, like F=ma and such, but even then, to capture real physics you do some much more complicated modeling on a subset of the total particles on each time tick to learn more. I’m sure a “big enough” model could do similar things with people, and no doubt the theory of how the statistics would work is already being studied.
The other thing, though, is the data. In climate science you are trying to incorporate as many streams of data as you can. Back-testing your model against historical data and then extrapolating is the whole deal. You could claim that a good model of the economy would need too many streams. One per person? What sampling rate of streams for each n people leads to good conclusions? Related: I am being actively recruited by a large hedge fund right now that claims to use millions of data sources in their macro models. How many economic data streams does Amazon have? Facebook?
At what point is a big enough computer, with a big enough model, and enough inputs, statistically useful even if you see the world’s trade as information in a Hayekian way? For my perspective that computer does not seem so big nor so far-fetched.
Computer power is not the solution. The problem is too much of what James Manzi calls causal density (his book Uncontrolled explains this, as do the two essays I mentioned in the post)
I guess the question this post raises would be: what in your opinion does good macroeconomic research look like? Do you see a role for simplifications such as the representative agent or full information rational expectations in the context of a model that focuses on something else? Krugman’s ZLB paper I would argue is an example of work that uses these simplifications in a very insightful way, and not because the assumptions are true. http://www.brookings.edu/~/media/projects/bpea/1998%202/1998b_bpea_krugman_dominquez_rogoff.pdf
I’m sympathetic to the PSST framework as being a useful broad framework, but I guess I also see value in models that throw away large parts of that complexity in order to focus on one interesting piece at a time. A complex reality might be best understood with a patchwork of simple models of different aspects of it.
Models that throw away large parts of complexity are fine, as long as you are modest about them. Solow was always very modest about his growth model, but later generations of economists came to worship such models. Krugman pounds his fist on the table and insists that his model is the one right way to interpret reality. Such hubris is unjustified.
Here’s a fantastic article with a very nice tool demonstrating the overspecification problem in macroeconomics: http://fivethirtyeight.com/features/science-isnt-broken/
Want to show that Republicans (or Democrats) have a positive (or negative) effect on the economy? Tweak your variables and your model until you get an answer that is significant in the direction you want! All answers are possible.