A 61-million-person experiment on Facebook shows how ads and friends affect our voting behaviour

On November 2nd, 2010, more than 61 million adults visited Facebook’s website, and every single one of them unwittingly took part in a massive experiment. It was a randomised controlled trial, of the sort used to conclusively test the worth of new medicines. But rather than drugs or vaccines, this trial looked at the effectiveness of political messages, and the influence of our friends, in swaying our actions. And unlike most medical trials, this one had a sample size in the millions.

It was the day of the US congressional elections. The vast majority of the users aged 18 and over (98 percent of them) saw a “social message” at the top of their News Feed, encouraging them to vote. It gave them a link to local polling places, and clickable button that said “I voted”. They could see how many people had clicked the button on a counter, and which of their friends had done so through a set of randomly selected profile pictures.

But the remaining 2 percent saw something different, thanks to a team of scientists, led by James Fowler from the University of California, San Diego. Half of them saw the same box, wording, button and counter, but without the pictures of their friends—this was the “informational message” group. The other half saw nothing—they were the “no message” group.

By comparing the three groups, Fowler’s team showed that the messages mobilised people to express their desire to vote by clicking the button, and the social ones even spurred some to vote. These effects rippled through the network, affecting not just friends, but friends of friends. By linking the accounts to actual voting records, Fowler estimated that tens of thousands of votes eventually cast during the election were generated by this single Facebook message.

Mobilising voters

The study’s origins were seeded at 2010’s South by Southwest conference. Fowler was talking about his new book on social networks. Cameron Marlow, an “in-house sociologist” at Facebook was talking about the company’s interest in working with external researchers. People said they should talk to each other. “Facebook has always had an interest in politics, and for many elections, they have used the “megaphone” at the top of the News Feed to send a message to encourage people to vote,” says Fowler. “It was a natural fit to take what they normally did and add an alternate version, plus a control group.”

Facebook added just one hitch. They wanted to encourage as many people to vote as possible, which is why the “social message” group makes up 98 percent of the sample. Still, with 600,000 people in each of the other two categories, Fowler’s team have amassed a sample size that most researchers would salivate over.

The team found that 20 percent of the “social” users who saw their friends’ faces clicked the “I Voted” button, compared to just 18 percent of the “informational” users who saw the faceless message – a difference of 2 percentage points. Their odds of clicking for polling information also went up by 0.26 percentage points.

Talk is cheap, though. Fowler’s team also matched some the users’ online actions against actual voting records. They found that the social users increased the proportion of people who actually voted by 0.39 percentage points compared to either of the other groups. The just-the-facts-ma’am approach didn’t do anything. Facts only mattered when paired with social pressure.*

The differences may seem small, but try scaling them up. If the proportion of voters rises by 0.39 percentage points and you’re dealing with 61 million people, you’ve got 238,000 extra votes. And as Fowler reminds us, the 2000 US presidential election between George Bush and Al Gore came down to just 537 votes in Bush’s favour, in the state of Florida.

But David Nickerson, a political scientist from the University of Notre Dame, says, “It is hard to interpret [this] finding without getting completely into the weeds.” Due to privacy concerns, the team couldn’t rifle through individual records, and relied on an algorithm to do the job instead. On top of that, they could only do this for a third of all the users. Even so, Nickerson says that they have produced the best estimate we have about how online advertising could lead to behaviour in the fleshy world. Damon Centola, who studies collective behaviour at the M.I.T. Sloan School, agrees.  “This study does a very nice job of demonstrating the significant implications of an experimentally controlled on-line action for its off-line counterpart,” he says.

“It makes sense that weak messages can alter behavior by very small amounts,” adds Andrew Gelman, a statistician and political scientist at Columbia University. “It says that Facebook is another advertising channel.” But what a channel! It’s not up there with door-to-door canvassing, but as Nickerson explains, “Every 1000 ads placed create 3 votes. If each ad impression cost a campaign 2 cents, it would be lead to $6.67 per vote, which is very cost-effective.”

The influence of (close) friends

The online messages didn’t just affect the people who saw them. Facebook is, after all, a social network. But anyone who uses Facebook knows that the people we accept as “friends” aren’t all equal. Fowler’s team measured the strengths of these ties by counting the frequency of Facebook interactions, and confirmed that they match the frequency of flesh-space ones. They then arbitrarily defined “close friends” as the 20 percent of people with whom a user interacts with most often.

For each close friend who got the social message, a user’s odds of clicking “I Voted” went up by 0.099 percentage points; their odds of finding out about polling sites went up by 0.012 percentage points; and their odds of voting went up by 0.224 percentage points. The other “ordinary friends” increased the odds of clicking the button, but nothing else – they influence the smallest of behaviours, but not the more substantial ones.

Gelman is sceptical about these figures. “They are measuring tinier effects on the order of one-tenth to one-hundredth of a percentage point,” he says. “That sounds like zero to me.  Or, to put it another way, these effects are not zero, but they are on the same order of magnitude as other things in the air, such as recent news and political advertisements.”

Fowler fully admits that the effects are small, but he notes that they were the result of a single message. And, as before, they sum up to big population-wide changes. The team calculated that friends generated an extra 886,000 button-clicks, close friends created an extra 559,000 still, and close friends of close friends created an extra million.

When it came to actual votes, ordinary friends did nothing, but close friends generated an extra 282,000. Combined with the 60,000 votes produced by people seeing the message directly, that accounts for about 0.14 percent of all the votes cast during the 2010 election. And all from a single Facebook message!

“We may not personally feel moved by a single message like the one shown on Facebook, but the combination of millions of messages and billions of friends means that it does make a difference,” says Fowler. “The network multiplies whatever it is seeded with, and today it is easier to seed than ever.” The network is key, he stresses. “The indirect effect of social contagion on real-world voting behaviour was four times larger than the direct effect.  If we had only measured the direct effect, we would have missed the whole story.”

Gelman is wary. “Once you start playing around with small effects and looking at statistical significance, you can find all sorts of things in a dataset that might not replicate,” he says. But he adds that such research is worth doing, given how much time we spend online. And that may be the lasting legacy of this study.

Psychology experiments are often under-powered—that is, they don’t include enough people to produce statistically solid results. The internet, and social networks like Facebook, could change that by allowing scientists to carry out research on an unprecedented scale. It’s cheap and the results have “external validity”, meaning that they’re relevant to what people actually do in life, rather than in a stark controlled laboratory.

“It’s a brand new world!” says Fowler. He thinks that such experiments could help psychologists to do detailed studies on very specific groups of people. “[That] is the first step in understanding not just average human behaviour, but the behaviour of specific types of individuals in specific types of environments,” he says. “There are many human psychologies, not just one.”

Reference: Bond, Fariss, Jones, Kramer, Marlow, Settle & Fowler. 2012. 61-million-person experiment in social influence and political mobilization. Nature http://dx.doi.org/10.1038/nature11421

* Of course, this could have been because the people were responding to their friends’ social cues, or simply because we pay more attention to faces. The study doesn’t really tell us which explanation works, and while both speak to the inflence of social factors, they involve very different routes.

More on psychology and politics:

• The power of nouns – tiny word change increases voter turnout
• Sports results can affect election results
• Subliminal flag shifts political views and voting choices
• Voters use child-like judgments when judging political candidates
• Political attitudes linked to startle reflexes
• Undecided voters aren’t really undecided – the hidden side of decision-making