Large models, small models and Brexit

Non-economists with no interest in modelling techniques can skip to paragraph starting 'How is this all related to Brexit'.

I promised to look at some of the other papers in the OxREP volume “Rebuilding macroeconomic theory” besides my own, but as usual other things - including Brexit - got in the way. In this post I want to talk about the paper by Haldane and Turrell, which is about Agent Based Models, or ABMs. Right at the end of this post, however, I will come back to Brexit.

As a result of the microfoundations hegemony, any paper talking about a different modelling strategy often feels it must start by describing some drawbacks of that hegemony, and this paper is no exception. I might talk about that some other time, but instead I want to recommend what I think is one of the most realistic discussions of what ABM can or cannot do I have read.

As you might guess from the name, ABMs model the economy as a collection of a large number of different agents, each of which behaves in a specified way. The authors generalise the idea of a choice between internal and external consistency that I talk about in my paper to also include a degree of heterogeneity.

As you can see, ABMs are all about allowing as much heterogeneity as you wish. This is not to say that other methods cannot do heterogeneity (they can), but ABMs major in this dimension, and in practice often keep the behaviour of agents relatively simple compared to a DSGE. (A slight quibble: I would argue that as DSGEs are internally consistent by definition, the orange square representing them should be a slimmer and perhaps taller rectangle.) ABMs (within the bounds of tractability) owe no allegiance to any school of thought: the paper has a nice table of the many different types of consumption function used in a range of ABM studies.

As the macroeconomy is indeed made up of many different types of agents who may be doing different things, and whose interaction may produce unexpected results, it seems like ABMs can only be a good thing. But this additional freedom brings a large cost. Because, and unlike some hard sciences, there is a large amount of uncertainty about how people actually behave, we cannot treat any model as a black box, the output from which has to be accepted without question. No civil servant or central bank economist can go to politicians or governors and simply say it is what the model said.

Exactly the same problem can arise with SEMs, simply because of their complexity or disaggregation. It could also arise from a complex DSGE. The first question any economist asks when seeing an output from any large and complex model is does the result make sense given the smaller theoretical models they carry around in their head. It is why I proposed for SEMs the process I called theoretical deconstruction, where model properties were either reduced to familiar results from simpler models, or show the limitations of those simpler models. Again, as the paper notes, a similar process needs to, and in some cases has, happened with results from ABMs.

How is this all related to Brexit? The results showing how different degrees of Brexit would do the economy damage to different extents that I talked about in my last post were produced by trade theory’s equivalent of ABMs, called computable general equilibrium (CGE) models. These allow for considerable heterogeneity (across sectors and countries) in modelling trade. As Chris Giles recounts in this excellent piece, the model is more complex than anything the Treasury had before Brexit, and was built specifically to help with Brexit.

As Chris writes

“It must have come as a bit of a shock to government economists that the moment some results of this new model were leaked this week, ministers rushed to deny the usefulness of the tools they commissioned. Such models are “always wrong”, declared Steve Baker, a junior Brexit minister, on Tuesday.”

As I note in a postscript to my last post, he went further on Thursday to suggest that civil servants had deliberately cooked the model to sabotage Brexit.

How do we know that this didn’t happen, apart from the implausibility that so many civil servants could concoct such a conspiracy. Precisely because in this case the results from a highly disaggregated model broadly agrees with most other studies, and also common sense: the more difficult you make trade, the less there will be and the more costly that will be for UK output. Chris ends with some words that should be sent to every journalist in the country.

“Ministers now have a choice. They can opt for an honest Brexit in which they argue in public that people should pay an economic price for their policies. Or they can opt for a dishonest Brexit, pretending they have a secret plan for economic nirvana and trashing their own internal economic evidence. Ministers’ initial reaction in disowning the analysis suggests deception is the government’s central Brexit strategy. People talk about a crisis in economics. After this episode, it is the crisis in politics that should really concern us.”

Misunderstanding macroeconomic models

The first half of this post is meant for non-economists, but it ends with a couple of points on OLG modelling

I recently wrote a post on the Eggertsson and Mehrotra paper on secular stagnation, because I thought the paper was interesting. A much more critical post from Unlearning Economics (UE) has just appeared in Pieria. UE says it “helps to illustrate the troubles faced by contemporary macroeconomics”. One of UE’s complaints seems to reflect a misunderstanding, often shared by non-economists, about what much academic macromodelling is designed to do.

UE objects to the fact that the model assumes that the amount the young can borrow (the degree of leverage) is exogenous, which means that there is no attempt to explain where this constraint on the borrowing of the young comes from. UE also complains that the model contains no banks, and no investment in physical capital. In other words, the model is much too simple. It is a natural enough idea: to explain what might be currently going on, you need a more complex model that includes everything that could be important.

There is certainly a place for this kind of more elaborate model. Christiano, Eichenbaum and Trabandt in this paper want to argue that a model based on New Keynesian theory can track what has happened over the last ten years. Their model has 40 equations. If I was trying to do a similar exercise, I would want to augment the standard New Keynesian framework with at least the following: nominal wage stickiness as well as price stickiness, a financial sector that endogenised both the cost and rationing of credit, a model of consumption which allowed for credit constraints and precautionary saving, a housing market, a model of the labour market that combined matching with rationing (as here), and something that allowed recessions to have long lasting (hysteretic) impacts on labour supply and technical progress. However large models like this will involve many macroeconomic ‘mechanisms’, and it will generally be unclear which mechanisms are important at driving particular results or explaining particular facts. We do not want to treat the elaborate model as a black box, but instead we want to understand its properties.

To understand complex models, we need much simpler models. (I once - in this paper - called the process of relating complex models to simpler models ‘theoretical deconstruction’.) In fact it is often sensible to start with the simpler model. For example, a particular issue with secular stagnation is to show how the natural real interest rate can be negative for decades rather than years (i.e. beyond the Keynesian short term)? What mechanism can do this? As I explained in my post, neither a standard representative agent model nor a standard two period overlapping generations model (OLG model, where the two generations are those earning and those retired) will give you that result. What Eggertsson and Mehrotra show is that a very simple three period OLG model (which adds a young generation that borrows) where borrowing by the young is constrained (they would like to borrow more but cannot) can provide just that mechanism.

That is a key point of the paper. The paper is not designed to explain where borrowing constraints come from: there is now a big literature on that. Thankfully the authors do not feel compelled to microfound these constraints. Instead the paper simply offers and explores a mechanism whereby an increase in these borrowing constraints could move the natural interest rate into negative territory, and for it to stay there. Having established that result, it is for subsequent work (which the authors intend to do) to see if that mechanism survives complicating the model, by for example adding investment.

Suppose the endeavour is successful, and a more complex but realistic model is able to provide an account of secular stagnation that includes other important mechanisms and which is based on a realistic set of parameter values. That would be a success, but those not familiar with all the work would ask: why does this model allow real interest rates to be negative when the standard models we know do not. The reply would be that the three period OLG structure was critical, and to see why have a look at the original, simple model.

Now you might say the authors should wait until they have built the more realistic model before creating what could turn out to be a research path that might fail to achieve its goal. That would be quite wrong, because the more debate there is within the academic community when ideas are at their early stages the better. I want to give an example of this, but here I will go into territory that will probably only interest macroeconomists.

It might be the case, for example, that the authors intuition that their results will survive introducing other assets like physical capital can be shown to be wrong very quickly. Indeed, Nick Rowe has already made such a claim, arguing that the presence of land as an asset ensures a positive real interest rate. If Nick was right this could be enough to kill the research programme, without any more time being wasted. Whether he is right is another matter: this paper by Rhee may be relevant in that respect.

Here I just want to add a final thought. Within an OLG framework, it may not be necessary to establish the existence of a steady state with negative real interest rates. The typical period in an OLG model lasts two or more decades. So if the dynamics of such a model involved some overshooting, it might be possible to generate prolonged periods (in years) of negative interest rates even if the steady state real interest rate was positive. To be honest I’m not sure what might give rise to overshooting of this kind, but that may just reflect my inadequate imagination.