Using Social Media For Large Behavioral Studies Is Fast and Cheap, But Fraught With Biases and Distortion
Carnegie Mellon, McGill Researchers Say Higher Methodological Standards Are Needed
By Byron Spice / 412-268-9068
PITTSBURGH—The rise of social media has seemed like a bonanza for behavioral scientists, who have eagerly tapped the social nets to quickly and cheaply gather huge amounts of data about what people are thinking and doing. But computer scientists at Carnegie Mellon University and McGill University warn that those massive datasets may be misleading.
In a commentary published in the Nov. 28 issue of the journal Science, Carnegie Mellon's Juergen Pfeffer and McGill's Derek Ruths contend that scientists need to find ways of correcting for the biases inherent in the information gathered from Twitter and other social media, or to at least acknowledge the shortcomings of that data.
And it's not an insignificant problem; Pfeffer, an assistant research professor in CMU's Institute for Software Research, and Ruths, an assistant professor of computer science at McGill, note that thousands of research papers each year are now based on data gleaned from social media, a source of data that barely existed even five years ago.
"Not everything that can be labeled as 'Big Data' is automatically great," Pfeffer said. He noted that many researchers think — or hope — that if they gather a large enough dataset they can overcome any biases or distortion that might lurk there. "But the old adage of behavioral research still applies: Know Your Data," he maintained.
Still, social media is a source of data that is hard to resist. "People want to say something about what's happening in the world and social media is a quick way to tap into that," Pfeffer said. Following the Boston Marathon bombing in 2013, for instance, Pfeffer collected 25 million related tweets in just two weeks. "You get the behavior of millions of people — for free."
The type of questions that researchers can now tackle can be compelling. Want to know how people perceive e-cigarettes? How people communicate their anxieties about diabetes? Whether the Arab Spring have been predicted? Social media is a ready source for information about those questions and more.
But despite researchers' attempts to generalize their study results to a broad population, social media sites often have substantial population biases; generating the random samples that give surveys their power to accurately reflect attitudes and behavior is problematic. Instagram, for instance, has special appeal to adults between the ages of 18 and 29, African-Americans, Latinos, women and urban dwellers, while Pinterest is dominated by women between the ages of 25 and 34 with average household incomes of $100,000. Yet Ruths and Pfeffer said researchers seldom acknowledge, much less correct, these built-in sampling biases.
Other questions about data sampling may never be resolved because social media sites use proprietary algorithms to create or filter their data streams and those algorithms are subject to change without warning. Most researchers are left in the dark, though others with special relationships to the sites may get a look at the site's inner workings. The rise of these "embedded researchers," Ruths and Pfeffer said, in turn is creating a divided social media research community.
As anyone who has used social media can attest, not all "people" on these sites are even people. Some are professional writers or public relations representatives, who post on behalf of celebrities or corporations, others are simply phantom accounts. Some "followers" can be bought. The social media sites try to hunt down and eliminate such bogus accounts — half of all Twitter accounts created in 2013 have already been deleted — but a lone researcher may have difficulty detecting those accounts within a dataset.
"Most people doing real social science are aware of these issues," said Pfeffer who noted that some solutions may come from applying existing techniques already developed in such fields as epidemiology, statistics and machine learning. In other cases, scientists will need to develop new techniques for managing analytic bias.