Social Modelling

Yesterday Shankar Vedantam, the great science reporter for the Washington Post, had an article on the applications of computer modeling and analysis to social and economic problems. The article began with a timely piece on an analysis by Yaneer Bar-Yam of the New England Complex Systems Institute that suggests that specific changes in the regulatory framework for the stock market made by the Bush administration resulted in more instability in the financial systems, which in turn allowed the bursting of the housing bubble and the sub-prime mortgage crisis to trigger much wider economic problems.

Vedantum properly cautions against putting too much credence in the results of such models. However, they can be very useful. Often systems with complex feedback loops can behave in ways that are quite unintuitive, and a model which explores the effect of changes in a simplified version can help understand that behavior of the larger system.

What really caught my attention was his discussion of some of the results of V. S. Subrahmanian at the University of Maryland.

At the University of Maryland, for example, computer scientist V.S. Subrahmanian and political scientist Jonathan Wilkenfeld have built a computational model to predict how different situations amplify the likelihood of violence in the Middle East. One conclusion of their model is that the militant group Hezbollah is more likely to lob rockets into Israel when elections are being held in Lebanon -- some proportion of the attacks are meant to impress a domestic audience. The conclusion is not necessarily counterintuitive: A skilled political watcher could have told you the same thing. But if pundits intuitively know how 100 different issues might influence outcomes, computational models can tell you the relative importance of each variable.

Another model by the Maryland group shows that infant mortality levels predict the likelihood of political instability in a country better than any other single measure. Again, this is not a shocker. Anyone can guess that countries with poor public health are on less secure footing. What the model does is tell us to pay preferential attention to infant mortality over, say, hunger or poverty or religious strife.

Both of these results may help develop intuition. I would caution, however, that "correlation is not causality". It may well be that some combination of conditions -- ignorance, poverty, the growth of urban slums -- result in both more political instability and higher reported infant mortality. It can also be that a model assuming one set of causal relationships results in behavior of the output variables of the model system that closely parallel the observed behavior of the real system even if the causal relationships in the real system are different. That may be an important problem if one seeks to change the behavior of society based innovations that worked in the model, but depend on unrealistic causal assumptions.