Dude, where’s my psychohistory?

In a feature story for Nature at the beginning of the month, Laura Spinney writes about the latest generation of academics hoping to quantify historical patterns and make universal claims about “cliodynamics,” as U Conn’s Peter Turchin hopes to call his work. We’ve seen this movie before as quantitative social science history in the 1960s, cliometrics in the 1970s, and 1990s attempts by social scientists to use (bastardized) chaos models to describe everything from commodity and stock prices to … okay, generally commodity and stock prices. For the most part, those of us who are trained in history and in quantitative social sciences understand the limits of quantified models of human behavior. Maybe those who are more ambitious along these lines were inspired by Isaac Asimov’s Foundation series, with Asimov’s fictional discipline of psychohistory; Paul Krugman has said as much in terms of his attraction to economics. ((In that blog entry, Krugman writes at the end, “I wanted to be a psychohistorian when I grew up, and economics was as close as I could get.”)) But I think Krugman would agree with me that we don’t have psychohistory, despite the value of quantitative work in history.

Why not?

Exogeneous shocks shock. Some of the best empirical quantification of human behavior is descriptive more than predictive. My favorite example is from demography, where Ansley Coale and Sam Preston capped off more than a century of work in mathematical demography in 1982 with a complete model of population dynamics. I’m biased because I took several courses with Sam in grad school, but it’s beautiful and essential reading in the field. Demographers have contributed lots of other great models of population change, so the Coale-Preston work is just an example. And… it’s descriptive, not predictive. Once HIV spread around the world, demographers’ models are helpful in describing the consequences of changes in mortality for specific populations, but they could not predict HIV. Epidemics, volcano eruptions, and lots of other events are going to be outside any effective model of society, or an effective model for the data available.

I’m a prisoner of my model. In her article, Spinney focuses on one research project to explore cycles of violence in human societies. The wonderful thing about a periodic pattern is that there is a wealth of analysis techniques one can use, from polynomials to linear combinations of sine waves. The problems with trying to make claims about periodicity in human behavior? That “wealth of analysis techniques” is one of them, if you are trying to model a system, because it is just too easy to fit a flexible model to any data that looks vaguely periodic. Social scientists all too often stop looking for models at the first one that suffices. That’s great if you’re developing something that is descriptive and awful if you are trying to generate a best model of behavior. As described by Spinney, Turchin’s work looks like it may fall in this category.

Corollary: the Peter Principle of clever techniques. Social scientists and others have created some incredibly clever analytical techniques, and then others get to abuse them. The econometric instrumental-variable approach to estimating behavior is a classic example. James Heckman, a deserved Nobel Prize winner in economics, has observed in print that the instrumental-variable approach is sometimes more valuable for identifying the existence of effects than for specifying their size. Another example is the propensity-score matching design for identifying treatment effects in a non-randomized study. As my statistician uncle David Salsburg pointed out to me, using a highly parametrized propensity-score model assumes too much robustness for the matching model; even the entire universe of polynomial functions has a measure of roughly zero in comparison with all potential matching models. ((My uncle’s advice: compare treated/untreated cases in quintile bands on the chosen model; then the fragility of the model is primarily around the quintile boundaries. Addendum: This is one of the several ways to use propensity-score analysis without worshipping it. Another is to use it to eliminate outliers who are far from being on the edge of the behavior in question.))

Information loss. One of the fundamental assumptions of physical models is that a system does not lose information–one faces measurement error, but at least theoretically a system does not lose information. In contrast, social behavior loses information all the time, in at least two senses. ((Because humans are part of the physical world, of course we are part of a universe that does not lose information, but humans in themselves are not the universe or a closed system.)) First, human interactions are sensitive to some small differences but not all small differences. I spoke with a fellow chair by phone the other day about some planning for the fall. There is plenty of her behavior that I paid attention to, mostly the semantic meaning of her words. There is plenty of her behavior that didn’t affect me, and some of it I wouldn’t know about (was she multitasking during the phone call? did she roll her eyes at some of the things I said?). After we hung up, I took some actions, and here is the relevant piece: there is no way you could retrace all of her behavior during the phone call from the actions I took afterwards. That is fairly typical about people: you lose information about some part of social behavior from later behavior. So at some level, physicists’ models of the universe are inappropriate for human interactions.

There is an alternative: viewing concepts from physical sciences as potentially useful metaphors rather than exact analogues. Even here, sociologist Howard Beckman’s warning about metaphors is warranted, but there is always some use in learning about other fields as long as you do not take those models too literally. Sean Carroll’s notion of spontaneous social symmetry breaking is one example: he argues that it could be useful in explaining the difference between theoretical color-blindness in a world that never experienced prejudice and in a world where we have a long history of discrimination. In fact social-scientists do use that model of human behavior, except we generally call it by the awkward term path-dependence or cumulative advantage (aka the Matthew Effect), and most social scientists probably have not been exposed to physicists’ deeper notions of symmetry breaking.

But the bottom line is that modeling human behavior can be inspired by ideas in the physical sciences but will not be able to squeeze out the same type of exactitude Spinney described as an ideal.