Everything Is Obvious
Page 8
Nevertheless, in this crowd, as everywhere, individual people have different tendencies toward violence. Perhaps those who are better off or who are less affected financially by the new policy are less inclined to risk jail time to make a point. Others are more persuaded that violence, although regrettable, is a useful political device. Some may have an unrelated gripe against the police or the politicians or society, and this event is giving them an excuse to vent. And perhaps some of them are just crazier than others. Whatever the reason—and the reasons can be as many and as complicated as you can imagine—each individual in the crowd can be thought of as having a “threshold,” a point at which, if enough other people join in the riot, they will too, but below which they will refrain. Some people—the “rabble rousers”—have very low thresholds, while others, like the president of the student society, have very high thresholds. But everyone has a threshold of social influence, above which they will “tip” from calm to violence. This might seem like a strange way to characterize individual behavior. But the benefit of describing people in the crowd in terms of their threshold is that the distribution of thresholds over the whole crowd, from crazy (“I will riot even if no one else does”) to Gandhi (“I will not riot even if everyone else does”) turns out to capture some interesting and surprising lessons about crowd behavior.12
To illustrate what could happen, Granovetter posited a very simple distribution in which each of the hundred people has a unique threshold. Exactly one person that is, has a threshold of zero, while another has a threshold of one other person, another has a threshold of two other people, and so on all the way up to the most conservative person, who will join in only after all ninety-nine others have. What will happen? Well, first Mr. Crazy—the one with the threshold of zero—will start throwing things, apropos of nothing. Then, his sidekick with the threshold of one (who needs only one other person to riot before joining in) joins him. Together, these two troublemakers prompt a third person—the guy with the threshold of two—to join in as well, and that’s enough to get the threshold-three person going, which is enough to … well, you get the idea: Given this particular threshold distribution, the entire crowd ends up joining the riot, one after the other. Chaos reigns.
Imagine, however, that in the next town over, a second crowd of students, of exactly the same size, has gathered for exactly the same reason. As unlikely as it may sound, let’s imagine that this crowd has almost exactly the same distribution of thresholds as the first one. So closely matched are these two crowds, in fact, that they differ with respect to just one person: Whereas in the first crowd each person had a unique threshold, in this one nobody has a threshold of three, and two people have a threshold of four. To an outside observer, this difference is so minute as to be undetectable. We know they’re different because we’re playing God here, but no feasible psychological test or statistical model could tell these two crowds apart. So what happens now to the crowd’s behavior? It starts out the same: Mr. Crazy leads off just as before, and his sidekick and the guy with a threshold of two join in like clockwork. But then it hits a snag, because nobody has a threshold of three. The next most susceptible individuals are the pair who both have thresholds of four; yet we have only three rioters. So the potential riot stops before it even gets started.
Now imagine, finally, what observers in these two neighboring towns would see. In town A, they would witness an all-out riot, complete with smashed shop windows and overturned cars. In town B, they would see a few loutish individuals jostling an otherwise orderly crowd. If these observers were to compare notes later, they would try to figure out what it was about the people or their circumstances that must have been different. Perhaps the students in town A were angrier or more desperate than those in town B. Perhaps the shops were less well protected, or perhaps the police were more aggressive, or perhaps the crowd in town A had a particularly inflammatory speaker. These are the kinds of explanations that common sense would suggest. Obviously something must have been different, or else how can we explain such dramatically divergent outcomes? But in fact we know that apart from the threshold of a single individual, nothing about the people or their circumstances was different at all. This last point is critical because the only way a representative agent model could account for the different outcomes observed in town A and town B would be if there were some critical difference between the average properties of the two populations, and the averages are for all intents and purposes the same.
The problem sounds similar to the one my students encountered when trying to explain the difference between organ-donor rates in Austria and Germany, but it’s actually quite different. In the organ-donor case, remember, the problem was that my students tried to understand the difference in terms of rational incentives, when in reality it was dominated by the default setting. In other words, they had the wrong model of individual behavior. But in the organ-donor case at least, once you understand how important the default bias is, it becomes clear why the donor rates are so wildly different. In Granovetter’s riot model, by contrast, it doesn’t matter what model of individual behavior you have—because in any reasonable sense the two populations are indistinguishable. To understand how the different outcomes emerge, you must take into account the interactions between individuals, which in turn requires that you follow the full sequence of individual decisions, each unfolding on top of the others. This is the micro-macro problem arriving in full force. And the minute you try to skip over it, say by substituting a representative agent for the behavior of the collective, you will have missed the whole essence of what is happening, no matter what you assume about the agent.
CUMULATIVE ADVANTAGE
Granovetter’s “riot model” makes a profound statement about the limits of what can be understood about collective behavior by thinking only about individual behavior. That said, the model is extremely—almost comically—simple, and is likely to be wrong in all sorts of ways. In most real-world choices, for example, we are choosing between potentially many options, not just the two—riot or don’t riot—in Granovetter’s model. Nor does it seem likely that the manner in which we influence one another in the real world is anything as simple as the threshold rule that Granovetter proposed. In many routine situations, when choosing, say, a new artist to listen to, a new book to read, or a new restaurant to visit, it often makes sense to ask other people for advice, or simply to pay attention to the choices they have made, on the grounds that if they like something you’re more likely to like it too. In addition, your friends may influence which music you choose to listen to or which books you choose to read not only because you assume that they have already done some work filtering out the various options but also because you will enjoy talking about them and sharing the same cultural references.13
Social influence of this general kind is likely ubiquitous. But unlike the simple threshold of Granovetter’s thought experiment, the resulting decision rule is neither binary nor deterministic. Rather, when people tend to like something that other people like, differences in popularity are subject to what is called cumulative advantage, meaning that once, say, a song or a book becomes more popular than another, it will tend to become more popular still. Over the years, researchers have studied a number of different types of cumulative advantage models, but they all have the flavor that even tiny random fluctuations tend to get bigger over time, generating potentially enormous differences in the long run, a phenomenon that is similar to the famous “butterfly effect” from chaos theory, which says that a butterfly fluttering its wings in China can lead to a hurricane months later and oceans away.14
As with Granovetter’s model, cumulative advantage models have disruptive implications for the kinds of explanations that we give of success and failure in cultural markets. Commonsense explanations, remember, focus on the thing itself—the song, the book, or the company—and account for its success solely in terms of its intrinsic attributes. If we were to imagine history being somehow “rerun” many time
s, therefore, explanations in which intrinsic attributes were the only things that mattered would predict that the same outcome would pertain every time. By contrast, cumulative advantage would predict that even identical universes, starting out with the same set of people and objects and tastes, would nevertheless generate different cultural or marketplace winners. The Mona Lisa would be popular in this world, but in some other version of history it would be just one of many masterpieces, while another painting that most of us have never heard of would be in its place. Likewise, the success of Harry Potter, Facebook, and The Hangover would turn out to be a product of chance and timing as much as of intrinsic quality.
In real life, however, we have only one world—the one that we are living in—thus it’s impossible to make the sort of “between world” comparisons that the models say we should. It may not surprise you, therefore, that when someone uses the output of a simulation model to argue that Harry Potter may not be as special as everyone thinks it is, Harry Potter fans tend not to be persuaded. Common sense tells us that Harry Potter must be special—even if the half dozen or so children’s book publishers who passed on the original manuscript didn’t know it at the time—because more than 350 million people bought it. And because any model necessarily makes all manner of simplifying assumptions, whenever we have to choose between questioning common sense and questioning a model, our tendency is to do the latter.
For exactly this reason, several years ago my collaborators Matthew Salganik and Peter Dodds and I decided to try a different approach. Instead of using computer models, we would run a controlled, laboratory-style experiment in which real people made more or less the same kinds of choices that they make in the real world—in this case, between a selection of songs. By randomly assigning different people to different experimental conditions, we would effectively create the “many worlds” situation imagined in the computer models. In some conditions, people would be exposed to information about what other people were doing, but it would be up to them to decide whether or not to be influenced by the information and how. In other conditions, meanwhile, participants would be faced with exactly the same set of choices, but without any information about other participants’ decisions; thus they would be forced to behave independently. By comparing the outcomes in the “social influence” conditions with those in the “independent” condition, we would be able to observe the effects of social influence on collective outcomes directly. In particular, by running many such worlds in parallel, we would be able to measure how much of a song’s success depended on its intrinsic attributes, and how much on cumulative advantage.
Unfortunately, running such experiments is easier said than done. In psychology experiments of the kind I discussed in the previous chapter, each “run” of the experiment involves at most a few individuals; thus conducting the entire experiment requires at most a few hundred subjects, typically undergraduate students who participate in exchange for money or course credit. The kind of experiment we had in mind, however, required us to observe how all these individual-level “nudges” added up to create differences at the collective level. In effect, we wanted to study the micro-macro problem in a lab. But to observe effects like these we would need to recruit hundreds of people for each run of the experiment, and we would need to conduct the experiment through many independent runs. Even for a single experiment, therefore, we would need thousands of subjects, and if we wanted to run multiple experiments under different conditions, we’d need tens of thousands.
In 1969, the sociologist Morris Zelditch described exactly this problem in a paper with the provocative title “Can You Really Study an Army in a Laboratory?” At the time, his conclusion was that you couldn’t—at least not literally. Therefore he advocated that sociologists should instead study how small groups worked, and rely on theory to generalize their findings to large groups. Macrosociology, in other words, like macroeconomics, couldn’t ever be an experimental discipline, simply because it would be impossible to run the relevant experiments. Coincidentally, however, the year 1969 also marked the genesis of the Internet, and in the years since, the world had changed in ways that would have been hard for Zelditch to imagine. With the social and economic activity of hundreds of millions of people migrating online, we wondered if it might be time to revisit Zelditch’s question. Perhaps, we thought, one could study an army in the laboratory—only this lab would be a virtual one.15
EXPERIMENTAL SOCIOLOGY
So that’s what we did. With the help of our resident computer programmer, a young Hungarian named Peter Hausel, and some friends at Bolt media, an early social networking site for teenagers, we set up a Web-based experiment designed to emulate a “market” for music. Bolt agreed to advertise our experiment, called Music Lab, on their site, and over the course of several weeks about fourteen thousand of its members clicked through on the banner ads and agreed to participate. Once they got to our site they were asked to listen to, rate, and if they chose to, download songs by unknown bands. Some of the participants saw only the names of the songs while others also saw how many times the songs had been downloaded by previous participants. People in the latter “social influence” category were further split into eight parallel “worlds” such that they could only see the prior downloads of people in their own world. Thus if a new arrival were to be allocated (randomly) to World #1, she might see the song “She Said” by the band Parker Theory in first place. But if she were allocated instead to World #4, Parker Theory might be in tenth place and “Lockdown” by 52 Metro might be first instead.16
We didn’t manipulate any of the rankings—all the worlds started out identically, with zero downloads. But because the different worlds were carefully kept separate, they could subsequently evolve independently of one another. This setup therefore enabled us to test the effects of social influence directly. If people know what they like regardless of what other people think, there ought not to be any difference between the social influence and independent conditions. In all cases, the same songs should win by roughly the same amount. But if people do not make decisions independently, and if cumulative advantage applies, the different worlds within the social influence condition should look very different from one another, and they should all look different from the independent condition.
What we found was that when people had information about what other people downloaded, they were indeed influenced by it in the way that cumulative advantage theory would predict. In all the “social influence” worlds, that is, popular songs were more popular (and unpopular songs were less popular) than in the independent condition. At the same time, however, which particular songs turned out to be the most popular—the “hits”—were different in different worlds. Introducing social influence into human decision making, in other words, increased not just inequality but unpredictability as well. Nor could this unpredictability be eliminated by accumulating more information about the songs any more than studying the surfaces of a pair of dice could help you predict the outcome of a roll. Rather, unpredictability was inherent to the dynamics of the market itself.
Social influence, it should be noted, didn’t eliminate quality altogether: It was still the case that, on average, “good” songs (as measured by their popularity in the independent condition) did better than “bad” ones. It was also true that the very best songs never did terribly, while the very worst songs never actually won. That said, even the best songs could fail to win sometimes, while the worst songs could do pretty well. And for everything in the middle—the majority of songs that were neither the best nor the worst—virtually any outcome was possible. The song “Lockdown” by 52 Metro, for example, ranked twenty-sixth out of forty-eight in quality; yet it was the no. 1 song in one social-influence world, and fortieth in another. The “average” performance of a particular song, in other words, is only meaningful if the variability that it exhibits from world to world is small. But it was precisely this random variability that turned out to be large. For example, by chang
ing the format of the website from a randomly arranged grid of songs to a ranked list we found we could increase the effective strength of the social signal, thereby increasing both the inequality and unpredictability. In this “strong influence” experiment, the random fluctuations played a bigger role in determining a song’s ranking than even the largest differences in quality. Overall, a song in the Top 5 in terms of quality had only a 50 percent chance of finishing in the Top 5 of success.
Many observers interpreted our findings as a commentary on the arbitrariness of teenage music tastes or the vacuousness of contemporary pop music. But in principle the experiment could have been about any choice that people make in a social setting: whom we vote for, what we think about gay marriage, which phone we buy or social networking service we join, what clothes we wear to work, or how we deal with our credit card debt. In many cases designing these experiments is easier said than done, and that’s why we chose to study music. People like to listen to music and they’re used to downloading it from the Web, so by setting up what looked like a site for music downloads we could conduct an experiment that was not only cheap to run (we didn’t have to pay our subjects) but was also reasonably close to a “natural” environment. But in the end all that really mattered was that our subjects were making choices among competing options, and that their choices were being influenced by what they thought other people had chosen. Teenagers also were an expedient choice, because that’s mostly who was hanging around on social networking sites in 2004. But once again, there was nothing special about teenagers—as we showed in a subsequent version of the experiment for which we recruited mostly adult professionals. As you might expect, this population had different preferences than the teenagers, and so the average performance of the songs changed slightly. Nevertheless, they were just as influenced by one another’s behavior as the teenagers were, and so generated the same kind of inequality and unpredictability.17