Social networks outsmart cognitive biases

In 2010, the New York City-based restaurant Serendipity 3 revealed its $69 hot dog, winning the Guinness World Record for the world’s most expensive hot dog. Served on a toasted pretzel roll with truffle butter and covered in foie gras, the award-winning hot dog made the restaurant’s $18 cheeseburger seem like a steal. And that’s the point, says Professor Damon Centola of the Annenberg School for Communication at the University of Pennsylvania.

“If you put an absurdly expensive item on the menu, it anchors people’s expectations,” Centola says. “A $69 hot dog makes an $18 cheeseburger seem reasonable by comparison.” And it works: once Serendipity 3 introduced its $69 hot dog, cheeseburger sales skyrocketed.

Numerous studies have examined how anchoring bias affects individuals — psychologist Daniel Kahneman won the 2002 Nobel Prize in Economics for showing how pervasive and inescapable anchoring bias is — but it’s never been studied in networks.

In a new study published in the Journal of Social Computing, Centola, the Elihu Katz Professor of Communication, Sociology, and Engineering, and Annenberg doctoral candidate Calvin Isch tested whether groups that were initially subjected to anchoring bias could recover their ability to make rational decisions as a result of exchanging their opinions within peer networks.

In an experiment involving 1,600 people, Centola and Isch showed participants a photo of 246 pennies in a pile and asked them to guess the number of coins. Participants either worked alone (as solitary individuals) or as members of connected networks. To establish an anchor, participants were first assigned either a low (118) or high (353) anchor value and asked to indicate whether there were more or fewer pennies than this anchor. After answering this question, participants were then asked to estimate the number of coins in the pile.

The anchors worked just as predicted. Without any anchor, participants' median guess was 185 pennies. With the low anchor, the median guess dropped to 141, becoming less accurate (as a result of the harmful anchor), but with the high anchor, the average guess jumped to 200, moving closer to the actual number (a result of the helpful anchor).

After providing their initial guesses, participants in the network groups exchanged opinions with their peers. They were then given the opportunity to reflect and revise their guesses based on their peers’ guesses. Participants in control groups working alone were also given the option to reflect and revise their guesses, but were not provided with any information about other guesses.

Participants in the control groups did not improve their guesses across rounds of revision, the researchers found, but the exchange of opinions within peer networks significantly improved accuracy for participants in both the low- and high-anchor groups. Participants who had been exposed to a harmful anchor moved away from it, back toward a more accurate guess, while those exposed to a helpful anchor moved farther in that direction. 

Overall, participants working in networks reduced their errors by 22%, while those working alone showed no reduction in their errors. “Social influence in egalitarian networks allowed errors to cancel out, undoing the harmful anchoring bias, and making individuals much more rational in their estimates,” Centola explained.

Centola and Isch found a surprising reason for this: it stems from participants’ confidence in one another. Participants who were confident in their own answers were not necessarily more accurate, but those with greater “confidence in others” tended to be more accurate and more engaged in the social learning process.

“Ironically,” Centola says, “people who were the most accurate reported less confidence in their own answers, but greater confidence in others. It was network members’ confidence in one another that enabled them to collectively unbias themselves and find more accurate answers than participants in the control groups.”

The findings indicate that anchoring bias is not always negative — it’s context-dependent, and so are the effects of networks.

"It’s a useful lesson for online networks. Sharing opinions in an egalitarian way enables even explicitly biased populations to overcome biases that are impossible for individuals to escape on their own,” Centola says.

“Experimental Evidence That Social Learning in Structured Information-Sharing Networks Corrects Anchoring Bias” was published in the Journal of Social Computing.