By Stefan McCabe
On February 3, the NULab co-sponsored a talk by Michael Macy titled “Can We At Least Agree on Science?” The talk, a synthesis of previously published work and a preview of a forthcoming paper, was an excellent example of how computational methods and large-scale datasets can shine new light on existing controversies in sociology and political science.
Macy’s project is motivated by a long-standing puzzle: why do political (and cultural) positions that seem to have nothing in common cluster into a few common political identities? The strangeness of this phenomenon was eloquently stated by Steven Pinker in The Blank Slate:
The right-left axis aligns an astonishing collection of beliefs that at first glance seem to have nothing in common. If you learn that someone is in favor of a strong military, for example, it is a good bet that the person is also in favor of judicial restraint rather than judicial activism. If someone believes in the importance of religion, chances are she will be tough on crime and in favor of lower taxes. Proponents of a laissez-faire economic policy tend to value patriotism and the family, and they are more like to be old than young, pragmatic than idealistic, censorious than permissive, and in a business rather than a university or government agency. The opposition positions cluster just as reliably: if someone is sympathetic to rehabilitating offenders, or to affirmative action, or to generous welfare programs, or to tolerance of homosexuality, chances are good that he will also be a pacifist, an environmentalist, an activist, an egalitarian, a secularist, and a professor or student.
This really is a strange phenomenon. How far can it be stretched? Macy wants to see how far it extends: does the same political polarization affect the ostensibly neutral sciences?
The team’s method for testing this is quite clever. Amazon provides recommendations for each book (“customers who bought this item also bought…”), so we can infer the presence of a bipartite network connecting books and customers. Because it’s Amazon, there’s a lot of data to work with (1.3 million books and 26 million co-purchase links). While the physical and life sciences are not as politicized, there is evidence of polarization. In some subdisciplines (e.g., climatology) the effect is stark, in other subdisciplines the effect is weak, but it’s almost always present. Liberals and conservatives do read different books about science. There’s also a broader trend: conservatives prefer applied science, liberals prefer basic science. (This effect was ingeniously operationalized using the ratio of patent to article citations.) Even though Amazon’s recommendation system is an algorithmic black box, the results held up when replicated using Barnes & Noble’s website.
Professor Macy also presented some of the research that went into his paper, “Why Do Liberals Drink Lattes?” Using a similar bipartite network of Twitter co-followers, Macy and his colleagues assigned an ideology score to different brands. They find evidence for many common stereotypes: liberals really do love Priuses (but not as much as they love Tesla); conservatives ride Harleys. Many of these displays of cultural-political polarization are presented at a companion website, Lifestyle Politics.
It was a very fun talk; Professor Macy is a witty and engaging speaker and the methods used are quite interesting. Nevertheless, it’s hard to see this data and not feel some degree of pessimism about the current political situation. It turns out that the answer to the question posed in his title—“Can We At Least Agree On Science?”—is essentially “no.” Where the audience for science is increasingly polarized, I fear that controversial research may become increasingly risky for scholars to pursue.
If Macy’s talk made me pessimistic as a citizen, it nevertheless made me optimistic as a social scientist. His work shows how network methods can scale to very large datasets (and illustrates some of the methodological challenges posed by Big Data and network measures). Although the work was primarily empirical, he tied the empirical data to speculation on causal mechanisms through a very clever agent-based model, showing how simulation can provide a bridge between empiricism and theory.