We tend to assume that having a large and expansive network automatically means we will have a collection of diverse perspectives to rely on for information. However, recent research shows that people in networks tend to gravitate toward like-minded people and that most of the people we are likely to meet are already thinking like us. This means that simply trying to meet more and more people won’t work to give us the range of information we need in order to make better decisions and find better opportunities. In fact, this approach can even lead to making disastrous choices.
ON NOVEMBER 8, 2016, Donald Trump shocked America.
Well, Donald Trump shocked half of America.
Sometime around 9:20 P.M., US Eastern time, the predictions of pollsters, data scientists, and media pundits changed dramatically.1 For almost the entire campaign season, including most of Election Day, the consensus was that Trump’s opponent, Hillary Rodham Clinton, was going to win the election and become the country’s first female president. As the results started coming in from more and more counties, particularly those in the Midwestern “Rust Belt” states, Trump’s chances began looking better and better. By midnight, the forecasters who that morning had been assured of Clinton’s victory now gave Trump more than a 90 percent probability of winning.2 At 1:39 A.M., the Associated Press forecasted that Trump would win the state of Pennsylvania. The next moment, Nate Silver, founder of the politics forecasting website Fivethirtyeight.com and the forecaster who correctly predicted the outcome in forty-nine out of fifty states in 2008 and all fifty in 2016, replied to this news: “The AP has essentially called the presidency for Trump.”3
About an hour later, Silver, who that morning had given Trump only a 28.6 percent chance of winning,4 would post on his website: “That’s a wrap. Donald Trump has been elected president of the United States.”5 While Silver was careful to explain that the polling had shown a competitive race and that Clinton had a few weaknesses in the Electoral College, he called the result “the most shocking political development of my lifetime.”6
To be sure, Nate Silver was not the only one who was shocked. The majority of forecasters had favored Clinton to win, and most with a greater assertion of certainty than Silver’s.7 Prediction markets had favored Clinton, as did the majority of public relations and communications professionals surveyed by PRWeek magazine.8 The New York Times claimed weeks before the election that Clinton was “poised to win easily.”9 Just a few weeks before Election Day, Harvard University professors were assuming a Clinton win and discussing how to “build an exit ramp” for Trump supporters, already tackling the challenge of de-escalating the blowback among Trump’s voters.10 Longtime Democratic House leader Nancy Pelosi went even further, claiming as early as June and with great confidence that the Democrats would “win it all,” referring to Democrats’ chances of recapturing the Senate and earning more seats in the House of Representatives. Pelosi called candidate Trump “the gift that keeps giving to us.”11 Even inside of the Clinton campaign, the consensus on Election Day was that victory was a sure thing, and Clinton campaign aides were popping bottles of champagne on the campaign plane early in the day on Tuesday.12 It seemed like everyone was telling everyone else that Clinton’s victory was a sure thing.
In the aftermath of the election, one of the most common emotions among those disappointed by the election results was sheer shock. “The outcome was so certain; how could the reality be anything different? How could we have missed this when everyone we knew was assuring a Clinton victory?” Some even went as far as to wonder how Trump could win when no one they knew had voted for him.13 Despite the shock and surprise, however, signs of a Trump victory were definitely present. They just weren’t noticed.
The Clinton campaign even had warnings about the Rust Belt and about the state of Michigan in particular. But those warnings were so soft and so few that they were largely ignored. A week and a half away from Election Day, the Service Employees International Union started hearing reports of worry on the ground in Michigan.14 The union leaders decided to pull volunteers from campaigning in Iowa and move them to Detroit to shore up efforts there. But even as they were booking hotel rooms, the headquarters of Clinton’s campaign told them to remain in Iowa. The campaign team was certain they would win Michigan by five percentage points.
Not that the entire campaign team was so certain, actually. Jake Sullivan, Clinton’s policy director, was the sole member of the inner circle who expressed concerns that she might lose.15 Sullivan tried often to convince the rest of the team to devote more time and attention to Michigan and the other Midwestern swing states. Sullivan’s concerns weren’t rejected . . . they were just ignored. The inner circle was too busy deciding what traditionally Republican states they wanted to add to their winnings. The overwhelming opinion of the network of Democratic campaigners was that Michigan, and the entire election, was in the bag.
To be sure, most of the Trump campaign’s data and expectations pointed to only an outside chance of winning, but they were cautiously optimistic.16 But the race was certainly far closer than it appeared to a huge portion of the population. On the day of the election, RealClearPolitics, which keeps an aggregate of polls at both the national and statewide levels, was showing Clinton winning by only two Electoral College votes, and many of the states in the Democrat’s corner were well within the margin of error of polling.17 Clearly, many people were seeing only the result that they wanted to see.
Perhaps the most famous Election Day prediction that turned out to be correct was from the liberal activist and filmmaker Michael Moore. Months before Election Day, before either Trump or Clinton had even been officially nominated by their party, Moore said in an interview, “I know that they [the Trump campaign] are planning to focus on Michigan, Ohio, Pennsylvania, and Wisconsin. That’s how he can win the election . . . If he can get those upper kind of Midwestern-type states, then he can pull it off.” A few months later, in July, Moore had strengthened his assertion: he insisted that Trump would win.
His assertion proved true. Though it took time for the election results to be final, in the end Trump won Ohio, Pennsylvania, Wisconsin, and even Michigan. But how did Moore foresee this while so many others did not? For starters, Moore is from Michigan, and he still lives and works there. His early film work focused on the economic plight of the working class in Michigan and other Rust Belt states. While so many other pundits cited demographic and other changes that were making the Rust Belt working class somewhat irrelevant, Moore had firsthand experience of just how strong a force that blue-collar group could be in the general election. But his prediction was either not taken seriously or, if it was, discredited.
In the shocked aftermath of the election, members of the American media began to wonder how they had missed the trend—and why people like Michael Moore saw it. Over time the consensus developed that, indeed, they didn’t understand how to interpret events and data because of their own isolation from opposing views.
There were warning signs. Months before Election Day, New York Times columnist David Brooks (a conservative who still gave Trump little to no chance of winning the nomination) offered an explanation of why even Trump’s nomination was unpredictable. Brooks wrote, “We expected Trump to fizzle because we were not socially intermingled with his supporters and did not listen carefully enough.”18 Brooks was the first journalist, but not the last, to acknowledge not being engaged enough with opposing sides to understand their viewpoints or to notice the signs of Trump’s campaign momentum. Margaret Sullivan, a media columnist for the Washington Post, admitted, “We didn’t take them seriously. Or not seriously enough.”19 Two days after the election, CBS News political correspondent Will Rahn admitted that “we also missed the story, after having spent months mocking the people who had a better sense of what was going on.”20
Meanwhile, Michael Moore had a front-row seat on what was going on in what he considered his hometown. Others were too geographically or ideologically isolated to harbor any doubt. But how could such isolation happen? Is there really a connection between where you live and not only how you see the world but how you think the rest of the world sees itself? The evidence suggests that there is indeed a strong connection. And it affects all of us.
In the early 2000s, the journalist Bill Bishop and the sociology professor Rob Cushing began researching what appears to be an intriguing trend: neighborhoods were becoming increasingly more conservative or liberal.21 This wasn’t just people migrating to liberal or conservative states. Rather, this was happening inside of states, inside of cities themselves, where people appeared to be sorting themselves into neighborhoods based on their ideology.
Bishop and Cushing both live in Austin, Texas, a fairly liberal enclave in a mostly conservative state. They began collecting data on presidential voting records and sorting it by county. Right away the evidence seemed to back up their hunch. Across more than 3,100 counties, a pattern emerged. From 1948 to 1976, Democrats and Republicans were fairly evenly mixed and over that period became even more so. But after the 1976 election, things got very different, very quickly. Migration patterns showed that people began to sort themselves out, and that Democrat and Republican counties began to emerge—they were growing more segregated.
For example, during the polarizing and hotly contested race between Richard Nixon and Hubert Humphrey in 1968, just 37.2 percent of voters lived in a landslide county (places where one candidate won the county by more than 20 percent of the vote). By the year 2000, that number had risen to 45.3 percent.22 Overall, from 1976 to 2004, the gap between the parties at the county level increased in over 2,000 counties, while only about 1,000 counties grew more competitive. This sorting was not just long-distance migration either. It wasn’t people moving to “red” states or “blue” states. Even inside of states, people were migrating to those counties where they felt more comfortable politically.
In California, a traditionally blue state, thirty counties became more solidly Republican, while only eleven counties grew more contested.23 The same thing happened in the other direction. In San Francisco County, 44 percent of voters sided with Gerald Ford in 1976, but by 2004 only 15 percent of voters went Republican. Meanwhile, the number of voters in the county overall hadn’t changed—but for every Republican who left a Democrat moved in.
Bishop and Cushing even examined whether or not gerrymandering—the practice of redrawing congressional districts to gain an advantage—was at play.24 Interestingly, most of the studies they found suggested that partial redistricting actually made incumbent politicians less safe, not more so. There was no observable change in competitive districts immediately following any redistricting for the last three decades they studied. In short, politicians weren’t redrawing boundaries to pick their voters; voters were moving inside of new boundaries to pick their politicians. “Most of America and most Americans were engaged in a thirty-year movement toward more homogenous ways of living,” Bishop wrote, no doubt baffled by the patterns he saw.25
But to a network scientist, this pattern isn’t baffling at all. Instead, it is the exact footprint you would expect from the march of homophily. In networks, opposites don’t attract. Like-minded people do.
Originally coined by Paul Lazarsfeld and Robert Merton in the 1950s, homophily illustrates with data the old adage, “Birds of a feather flock together.”26 In personal relationships, the theory predicts that we are more likely to develop close ties with people who are like us. In social networks, it asserts that networks of individuals will inherently become more segregated and clustered over time. And in study after study, the effect is well documented. Sociologists have seen it wherever they look, from marriages to coworkers to social acquaintances and, yes, even to politics. Homophily on a large scale helps explain the mass migrations that Bishop and Cushing noticed—what they called the “big sort.”
But when it comes to political persuasions, homophily affects not just decisions on where to live but also choices of whom to listen to and what to read. In October 2008, the organizational analyst Valdis Krebs analyzed the network of book purchases on Amazon.com, using data on the books that customers bought as connection points.27 While the network that he analyzed focused on books and the ideas therein, not on actual people, you have to remember that it’s people who buy books, not ideas. His analysis revealed a stark separation of book purchases into not just two but three different clusters. People who bought books about Republican ideas weren’t connected at all to those who bought books about Democratic ideas. Even more surprising, the people buying Democratic books weren’t connected to those buying books about Barack Obama. While Krebs’s analysis foreshadowed Obama’s ability to stitch together a winning coalition of voters, it also foreshadowed the difficulties he would have working with Congress as president.
Sociologists and network scientists have observed the presence of homophily for a long time, but few have found a way to study what triggers it. It’s logical that people form relationships with others whose thinking is similar to their own, because maintaining such a relationship is much easier. But it’s also just as logical that something else is at play. Consider that people tend to choose their friends and even their favored coworkers from a fairly limited pool of the people around them (for example, the people they work with, or the friends of friends they already have). So it could also be that similarities in relationships are a result of locations in a social network. Or it could be both.
It was exactly that question that Duncan Watts (yes, the same Duncan Watts) and the graduate student Gueorgi Kossinets sought to answer. To begin, they needed a network—and a large one—to study. In addition, it had to be a network they could observe over time, since it’s kind of hard to study the origins of something just by looking at a brief snapshot in time. This was a puzzle that previous researchers had struggled to solve. But fortunately for Kossinets and Watts, new technologies offered a solution. Just as we learned from studies of structural holes, Kossinets and Watts believed that email data could be used to draw a rough picture of the connections inside an organization or community. “Reciprocated emails for the most part represent real relationships,” Watts wrote. “It is possible to use email exchanges as a way to observe underlying social networks.”28
Together, they collected data on more than 30,000 students, faculty, and staff at a large university in the United States over the course of an academic year.29 They collected not just the logs of email communication but also data on individual attributes (such as gender, age, department, years in the community) and also records of course registrations (not just courses that students took but also the courses that faculty taught). In total, they collected over 7 million message records—and that was after filtering out all messages sent to more than one person.
By taking all this data together, Kossinets and Watts were able to construct a model of the social network and show how it changed over the course of 270 days (the academic year). In particular, they were focused on new ties that formed over time in the network. Using this model, they found something remarkable.
Homophily did indeed predict who formed a tie with whom. People who were similar to each other but were not acquainted were far more likely than dissimilar people to form a connection over time. However, they found that where someone was positioned in the overall social network of the university also had an effect. Most often, individuals who were close to each other in the structure of the social network were already more similar to each other than distant pairs even before establishing a relationship. Taken together, the effect of where you sit in the network—that is, your proximity to potential connections—appears to influence your choice of who you connect with more than pure similarity.
Over time, what Kossinets and Watts observed was a weak preference for similar individuals made stronger by the way the network structure gradually changed to bring similar people closer together and more likely to form a connection. Their results suggest that homophily is a downward spiral. Small initial preferences matter, but how they change future choices strengthens the appearance of a preference for similarity.
The results also shed light on just how puzzling the problem of homophily turns out to be. Staying tightly clustered inside a network of similar voices, opinions, and characteristics can truly limit your ability to get an accurate picture of the environment around you and to make the best decisions for yourself or your organization. But that same tightly clustered network also makes it more and more difficult to find the new voices, opinions, and characteristics you need to obtain a more accurate picture. The bottom line is that who you know affects how you think, and it also affects which friend of a friend you’re likely to meet, for better or worse.
For the leaders of Gimlet Media, the fear was that homophily’s effects would be for the worse. Gimlet Media is the experiment of Alex Blumberg and his cofounder, Matt Lieber. Alex was a longtime veteran of public radio, having worked as a producer of the hit show This American Life and also as a cofounder of Planet Money, one of the original National Public Radio (NPR) podcasts. Working on Planet Money made Blumberg attuned to the emerging market for podcasts and downloadable audio (usually nonfiction stories or interviews) that could be consumed on demand. But he also saw that listeners would support it. After Planet Money raised over half a million dollars for a T-shirt (and documented the making of that T-shirt in a series of episodes), he had an idea.30 As he studied it more and more, he became convinced that listener commitment was indicative of a big opportunity. NPR is a nonprofit, journalism-focused organization, but he knew that a few for-profit companies were doing really well in podcasting. NPR shows like This American Life and RadioLab were also dominating the podcast market, despite being aired on radio as well.
Would the market respond well to a for-profit podcasting company built with the same kind of attention to quality that NPR shows had become known for? There was only one way to find out.
In August 2014, Blumberg started a podcast—appropriately titled Startup. It was a podcast about starting a podcast company, which he originally called American Podcasting Corporation (a nod to American Broadcasting Company, or ABC), though he and his cofounder quickly changed the name to Gimlet after working with a branding firm on an episode of the show itself. (The name is essentially meaningless, besides being a gin cocktail.) Both the show and the company grew quickly, as Blumberg and Lieber raised $1.5 million in investment from venture capital firms.31 They even raised an additional $200,000 in investment by appealing to listeners of the show who were accredited investors and wanted to be a part of their adventure.32 The investment money fueled more growth projects, and they quickly added two new shows.
That fueled additional interest, in both listeners and investors. After Gimlet received another $6 million in investment in 2015, the company was valued at $30 million.33 That much larger investment meant they needed to spur much larger growth. And growth meant more listeners and more shows; new shows meant more people. But as they looked at their current staff and started planning for their future staff, they noticed something they didn’t like.
In December 2015, the same month they received their $6 million investment, their office was pretty uniform and pretty white. They had twenty-seven employees at the time, and twenty-four of them were white. True to their unique style of transparency, the team at Gimlet decided to make a podcast about it. It was on the show that Blumberg stated why this lack of diversity was so troubling. “We have an obligation to do right by the stories we’ll be telling,” he said. “To tell them not just from one perspective, but with all the nuance and complexity they deserve.”34
To explore the issue, not only for the podcast but also for themselves, Blumberg interviewed several Gimlet employees, especially the employees of color. Despite the potentially uncomfortable experience of “three people of color bravely diving into a sometimes awkward conversation with their white CEO about race and diversity,” Blumberg learned just how easy it is to fall into a homogenous company roster and just how homogenous theirs was.35
It was during an interview with an openly gay colleague that a more interesting question arose. Beyond their surface-level diversity problem, did Gimlet lack ideological diversity? “The vast majority of the staff is sorta like politically liberal, cosmopolitan leaning, you know sort of like Brooklyn-based,” Blumberg told his colleague, but they didn’t appear to be political or religiously diverse. “We don’t have any Evangelicals on staff . . . I don’t think we have anyone who can name one NASCAR driver.”36 At that moment during the interview, the producer, sitting outside the booth listening to the conversation, walked in and asked if he could join the conversation. He was an Evangelical. He went to church every Sunday. He could even name a few NASCAR drivers. And he often felt like the only one; he felt like it was better to keep all of this to himself.
After that interview, Blumberg started to really understand the complexity of their problem. Their goal was clear: “We want to make sure that Gimlet is a place where you feel comfortable sharing your beliefs. Be you Christian, Native American, or transgender, or all of the above,” he said.37 At the same time, Blumberg recognized that the environment and the social network that his company operated inside of were going to make it tough for that to happen.
“We are a largely white organization in a historically pretty white industry,” he continued. “If we just sit around and wait for people of color to apply for the jobs we post, we are going to stay that way.”38 In fact, when they looked at their hiring methods, they found that they were pulling from a homogenous pool, since they mostly relied on refugees from the world of public radio to apply or recruited them directly. Like the university in Kossinets and Watts’s study, they didn’t have a strong preference for similarity, but their place in the larger social network was serving similarity as the only option.
In fact, most of the diverse hires up to that point had come from deliberately acting differently—going out of their way to connect with people from a much different industry or section of the network. They learned that this would be their best chance of solving their dilemma. “A more diverse staff means that there are more professional networks to tap into, and the process of becoming less monochromatic as an organization can take on momentum,” Blumberg said.39
Blumberg’s internal investigation shows just how hard diversity can be. Just as Kossinet and Watts’s study showed that homophily is a downward spiral, you may start out with a preference to connect with similar people, but that shapes your network, making it easier and easier to keep choosing the same type of people. If Gimlet succeeds as a company, whether or not they manage to break free of the temptation to select from a homogenous pool will likely have been a contributing factor.
Opposites don’t attract; similar people do. But when similar people connect, they change the broader social network to make more similar connections more probable. Kossinet and Watts’s study also suggests, however, that it’s possible to reverse course and become an upward spiral. It suggests we can break free from homophily by putting to work the same principles that drive it. Deliberately seeking out new, dissimilar connections moves your place in the network and makes it more likely that your future connections will also be dissimilar. The lesson of homophily is that who you know affects how you think. Because you’re most likely to know people who think like you do, it takes deliberate work to move against the strong current of similarity, but the benefits of escaping that current are legion.
The biggest implication from homophily research is that we are much more likely to make, and to already have, connections with people who are similar to us. While that’s good for making us feel comfortable, it’s bad for making decisions with lots of variables. We need our network to give us alternative perspectives, and to do that we need to know if our network has any alternative perspectives. We need to do an audit of our network.
So here is a quick exercise for doing just that:
Take a look at your most frequent interactions in a given week. Use the call records on your phone or your email outbox to generate a list of twenty to twenty-five people.
Add those names to the first column of a list, with your name at the top.
Then make a few more columns alongside your name and label each column with the category you want to audit—industry, department, function, race, religion, political ideology, etc.
Start listing where each person in the first column fits in those categories. If you don’t know, take a guess and then go find out. (You are probably wrong more often than you suspect.)
Chances are that, for all categories, many of your contacts are going to be pretty similar to you. And that should give you an idea of what you need to work on.
If most of your connections are active on social media, you can fill out this list even faster. Odds are that their profiles include much of this information. Services like LinkedIn, for example, should have individuals’ job and education information, as well as the groups they belong to. Services like Facebook should show you what things your connections “like.” You might be tempted to disconnect with them after seeing some of these things, but don’t . . . if these connections are different from you, you probably need them in your life even more than you think.
For a downloadable template to use when completing this exercise, go to http://davidburkus.com/resources/ and look for networking resources.