If there’s one thing that social scientists, economists and policy-makers would give their eye teeth for, it’s finding out how large-scale social patterns arise. Behavioural economist Joshua Epstein thinks he’s getting closer to understanding such dynamics by building rich models that represent how people act, and how their interactions translate into wide-scale change. As he tells Liz Else, he’s applying this approach to myriad problems, from how to minimise deaths in a flu pandemic, to uncovering the fate of lost cultures – and even to “growing” the Ten Commandments from scratch.
How does your work differ from what other modellers have been trying to do for years?
We create artificial societies of individual “agents” who behave like simple but plausible humans. We can simulate large numbers of people who differ from each other in various ways – their rules of behaviour, information, genetics and ethnic identities – and who interact locally. We can look at how those local interactions generate macroscopic patterns in society. That’s why I call this “generative social science”. We are building agent-based societies that are much richer than was possible 20 years ago. Mathematical economics and game theory focus on static equilibrium, and typically posit highly rational individuals. We are interested in bounded rationality – more like Homo sapiens than Homo economicus – and in dynamics. Of course, our method raises as many questions as it answers.
Advertisement
What sort of questions?
For example, how do you know that the set of micro rules you’ve set for your agents is the only one that would generate the macroscopic social pattern you are interested in? There might be alternative agent rules that could generate the same pattern. That would be an embarrassment of riches, but you’d have to do more scientific work to figure out which was the most tenable explanation.
How well does the method work in the real world?
It can help us understand everything from the way economic classes emerge in societies, to how epidemics spread, to the dynamics of civil conflict. Take epidemics. We looked at how to contain a smallpox epidemic. We based our data on 49 outbreaks of smallpox in Europe between 1950 and 1971. We had data on the size distribution of the outbreaks, and on the distribution across transmission settings: homes, schools, workplaces and hospitals. We built an artificial society based on the real data, including these same social units. We then calibrated the model by adjusting the number of person-to-person contacts per day in these various settings so that, when the biological agent was released, the results matched the real-life epidemic statistics.
“We’ve looked at how to contain a smallpox epidemic”
Have you looked at pandemic flu too?
Yes. We’re building agent-based infectious disease models for the National Institutes of 91ɫƬ. This is all part of an NIH programme called the Models of Infectious Disease Agent Study, or MIDAS. I think it’s fair to say that our kind of modelling has really taken over in large-scale infectious disease policy. Recently I’ve been looking at the value of restricting international travel during flu pandemics.
Isn’t that a bit draconian?
We’re finding that travel restrictions can make an important difference in the rate at which flu pandemics propagate around the globe. We have an article in the 2 May issue of PLoS One that discusses models in which airline travel into and out of a country is stopped the minute a threshold number of cases is observed there. The effects are complex. One surprising result is that in some cases travel restrictions can actually make the epidemic worse.
That sounds counter-intuitive.
That’s what we thought. We finally discovered that the restrictions were interacting with the global seasonality of flu. There is a high season and a low season in terms of flu prevalence. If it breaks out in Hong Kong in the low season and you restrict travel into the US, you will delay the arrival of the infection until the US high season and the resulting US epidemic will be worse. The benefit of this kind of global-scale modelling is that it includes both behaviour and planetary dynamics, such as seasons. There has been little work on modelling the role of behaviour and psychology – how people adapt to crisis situations, how they panic, how they refuse to take vaccine – all phenomena that could loom large in a real pandemic. I’ve been pressing, with success, for a focus on the behavioural dimensions of epidemiology.
Have you looked at civil violence?
Yes. In 2002 I published a paper on an agent-based model of civil violence in Proceedings of the National Academy of Sciences (vol 99, p 7243) which includes revolutions against central authority (such as the Russian, French and Iranian revolutions), and violence between distinct groups (such as Tutsi and Hutu in Rwanda). These models help illuminate social configurations – what Lenin called revolutionary situations – that are likely to eventuate in political violence. We’d like to “grow” the French revolution.
Have you done this sort of “computational history” before?
One of my favourite studies is our model of an ancient people called the Kayenta Anasazi. They lived in Long House Valley in north-eastern Arizona from around AD 900, and suddenly abandoned these lands around 1300. Why they vanished is a central enigma for archaeologists of the American Southwest. We created a computational model using both real data from AD 800 to 1300 about climate, crop production and drought severity, and artificial agents that form families and farm the land.
What did you discover?
Placing artificial Anasazi where the true ones were in AD 900, we were able to grow the observed spatial settlement patterns and population dynamics with high fidelity. We found that the Long House Valley area could have supported the Anasazi environmentally, suggesting that they abandoned it for social or cultural reasons. The artificial Anasazi also established settlements in places that have not been excavated yet, and it would be fun to take a look, to see if the true Anasazi also did. The trouble is that these are not public lands and it’s a delicate matter to go digging in there. The point about our work is not that it will solve the mystery of the Anasazi, but that agent-based modelling gives us a new kind of empirical research tool that may provide explanations no one has thought of before.
What’s on your to-do list?
Be a great dad, learn Ravel’s Gaspard de la Nuit. Find out if we can “grow” ethical systems like the Ten Commandments.
Profile
Joshua Epstein entered Amherst College as a musician, went on to study mathematics and, at MIT, mathematical social science. He is now a senior fellow in economics at the Brookings Institution in Washington DC, director of its Center on Social and Economic Dynamics and a member of the Santa Fe Institute external faculty. His latest book is Generative Social Science: Studies in agent-based computational modeling (Princeton University Press, 2007).
