Thursday, 7 June 2012

The Cognitive Microfoundations Project: a behavioural economics world tour

There has been much talk about microfoundations on the economics blogs in the last few months [Noahpinion, Mark Thoma, Simon Wren-Lewis twice, Andrew Gelman twice, Karl Smith, Paul Krugman twice, Robert Waldmann, Rajiv Sethi from 2009]. The idea of microfoundations is that a model of the overall economy should be consistent with how individual people act. The aggregate behaviour of variables like GDP, government deficits and unemployment should be derived by adding up the choices of individuals, not by treating the whole population as if it were a single entity.

(A microfounded model might start off like this: "Imagine N agents, each of which has income yn, consumes cn and saves sn. Then yn = cn + sn. For each agent, sn varies with the interest rate r according to the following relation..." while a non-microfounded model is more likely to start: "Total spending in the economy is C and saving is S. C+S must sum to Y, total income. S varies with the interest rate r...")

But does the microfoundations approach really work? It seems a good idea in principle. It works well in some other fields like physics and chemistry (though less so in biology). Building things from the ground up protects us against falling into certain mathematical traps. Some concepts (like the idea of people trading different goods with each other) don't really even exist at the aggregate level, so are hard to talk about without microfoundations. The idea that we can understand things in this level of detail is an appealing one.

Unfortunately, the idea of microfoundations has come to be closely associated with rational agent theory. Most microfounded economic models are implementations of DSGE (dynamic stochastic general equilibrium), which assume a population of rational utility-maximising agents who are given certain preferences and resources and respond logically to those. Readers of this blog, or of any behavioural economics book, will be unsurprised to hear that real people do not maximise utility in the way DSGE models insist - as demonstrated in numerous psychology experiments. Economists usually respond to this objection in one of two ways, neither of them quite satisfactory.

Response one: to claim that rational utility maximisation is close enough to the truth to describe the economy reasonably well. Sure, there are exceptions: people might not always discount future earnings in a consistent way, and sometimes we buy things because they’re on sale and not because our utility from the product exceeds the price paid - but those are minor errors, they mostly cancel each other out, we learn to be more rational over time, and the limits imposed by our income force us to act fairly rationally. So, DSGE models, maybe with a couple of small tweaks, are still the best way to describe the economy and work out how to manage it. We can still make inferences about how tax rates will change the choices of individual workers, or how interest rates will affect investment and savings decisions, and draw conclusions from that about how the whole economy will evolve.

Response two: to agree that individual rational agent models are too far from the truth to be useful but then to give up. For many, the failures of economic forecasting in the leadup to the 2008 crisis prove this. There are better ways to describe individual decisions - behavioural economics gives us some hints - but these are mathematically too hard to build models with. Therefore we shouldn’t bother with microfoundations - instead, we should reason from aggregates, such as the total amount of money, production, employment and debt in the economy. It is possible to work out, for example, that if companies try to save more money (as we can see they currently are), individuals try to pay off their debts (as they are), and governments try to cut their deficits (as they say they are) something must give. The model may not tell you which one will fail, but it can tell you that something must. These models can’t describe all economic phenomena because the aggregates don’t always tell you enough, but maybe they are all we have.

The first response is wishful thinking. The second is fatalism.

What if there is another way? Maybe, by choosing the right models from cognitive psychology and behavioural economics, and aggregating them in the right way, we can develop an accurate representation of large-scale systems after all. Then perhaps we can get the benefits of a microfounded model - which lets us understand many different economic phenomena, and gives us confidence via experiments that its conclusions are sound - but with greater accuracy, predictive power and robustness than today’s DSGE models.

Such models, microfounded not on rational utility theory but on real cognitive processes, might focus on specific domains such as consumer product markets or labour markets. They might let us explore the effects of specific economic policies such as tax or interest rate decisions. Eventually, they might develop into a unified theory that can be used to investigate any aspect of the economy - the cognitively sound equivalent of Arrow-Debreu general equilibrium theory.

Can this be done? It’s too early to say for sure, but it’s one of the most important questions for the economics discipline to ask itself.

So this year I’m going on tour. I will travel to wherever I can meet researchers in different economic domains and work out with them how psychology can be incorporated into their models. Although it might be possible to work out cognitive microfoundations from first principles, I suspect it will be more practical to start asking what kind of foundations will illuminate each different economic domain.

My initial objective is to work with people in each of the following disciplines:
  • Consumer behaviour
  • Competition and market organisation
  • Labour economics
  • Trade and international economics
  • Fiscal policy
  • Development economics
  • Monetary theory
  • Industrial organisation
  • Personal finance
  • Financial markets and asset pricing
  • Environmental economics
  • Health economics

I have a few collaborations lined up already, but there’s no restriction to just one in each field. So if you work in one of those areas - or would like to propose another - get in touch and I can add your location to my itinerary.

So far I’ve been to Madrid, Barcelona, Marseille, Paris and Honolulu. From today, my immediate plans are:
  • Until 13th June: San Francisco and Berkeley.
  • 13th-19th June: Atlanta.
  • 19th-30th June: the northeastern US - DC, NYC and all points between.
  • July: the UK and South Africa.
If you’re near any of those locations why don’t we meet up? If we discover anything useful there’s a co-author credit in it for you.

1 comment:

isomorphismes said...

Surprised you didn't mention multi-agent models. I remember in 2004 seeing X Gabaix's course on naive agents at the OCW project mention just a few ideas for short-calculating agents.

I also worked with a trader who wanted to use multi-agent modelling to make money in the markets. I steered him away from that b/c I don't think it's tractable at his timescale (intraday) w/o riddikulus paralellisation.

I can dig up a jillion papers about modelling with multiple agents (although I'm sure you could pretty easily as well). But my big question would be: if you remember Krugman's "Dishpan" article, he mentions thi downside of simulations. You get a lot of results but they're not analytic -- just tables of distributions and it's unclear whate reflects your assumptions and what reflects your inference. (Though I guess with a lot more computation and thought you could work that out.) But still you'll never get something as tight as an "if-then" mathematical result.

So, on the one hand, we have an obvious way of having different kinds of people interact; on the other hand, what would we get any better than the kind of Alife dissertation people love to trash: 150 pages of results with no wisdom.

Thus the attraction of satisficing models and bounded-rationality models, a topic I always meant to look into but quit the field too soon to accomplish. Last paper I perused was Becker, something about satisficing game theoretic solutions using Brouwer's theorem. At that point I guess there's a need for much more abstract math (to handle the slippery optima) as well as some psychologists in the room.