Doing OK? Check social media
Researchers are looking at online behavior, word choice to gauge public mental health
Which was the saddest day of them all?
This is the question you may be asking yourself, surveying the wreckage of 2020 thus far.
There are so many contenders to consider: Was it Thursday, March 12, the day after Tom Hanks announced he was sick and the NBA announced it was canceled? Was it Monday, June 1, the day protesters were tear gassed so that President Donald Trump could walk to his Biblewielding photo op?
Actually, it was neither, according to the Computational Story Lab of the University of Vermont. Instead, the lab offers this answer: Sunday, May 31. That day was not only the saddest day of 2020 so far, it was also the saddest day recorded by the lab in the last 13 years. Or at least, the saddest day on Twitter.
The researchers call it the Hedonometer. It is the invention of Chris Danforth and his partner Peter Dodds, both trained mathematicians and computer scientists and the co-directors of the lab. The Hedonometer has been up and running for more than a decade now, measuring word choices across millions of tweets, every day, the world over, to come up with a moving measure of well-being.
In fact, the last time The New York Times checked in with the Hedonometer team, in 2015, the main finding to emerge was our tendency toward relentless positivity on social media. “One of the happiest years on Twitter, at least for English,” Danforth said recently with a note of rue. That result now seems an artifact from an ancient era. “Since then it has been a long decline.”
What has remained constant is this: “Happiness is hard to know. It’s hard to measure,” he said. “We don’t have a lot of great data about how people are doing.”
The Computational
Story Lab is part of a small but growing field of researchers who try to parse our national mental health through the prism of our online life. After all, never before have we had such an incredible stockpile of real-time data — what’s known as our “digital traces”— to choose from.
And never has that stockpile of information towered as high as it does now, in the summer of 2020: In the first months of the pandemic, Twitter reported a 34% increase in daily average user growth. Without our normal social life as antidote and anchor, our social media now feels more like real life than ever before.
Since 2008, the Hedonometer has gathered a random 10% of all public tweets, every day, across a dozen languages. The tool then looks for words that have been ranked for their happy or sad connotation, counts them, and calculates a kind of national happiness average based on which words are dominating the discourse.
On May 31, the most commonly used words on English language Twitter included “terrorist,” “violence” and “racist.” This was about a week after George Floyd was killed, near the start of the protests that would last all summer.
Since the beginning of the pandemic, the Hedonometer’s sadness readings have set multiple records. This year, “there was a full month — and we never see this — there was a full month of days that the Hedonometer was reading sadder than the Boston Marathon day,” Danforth said. “Our collective attention is very ephemeral. So it was really remarkable then that the instrument, for the first time, showed this sustained, depressed mood, and then it got even worse, when the protests started.”
James Pennebaker, an intellectual founder of online language analysis and a social psychologist at the University of Texas at Austin, became interested in what our choice of words reveals about ourselves — our moods, our characters — exactly at the moment when the internet was first supplying such an enormous stockpile of text to draw from and consider.
“These digital traces are markers that we’re not aware of, but they leave marks that tell us the degree to which you are avoiding things, the degree to which you are connected to people,” said Pennebaker, the author of “The Secret Life of Pronouns,” among other books. “They are telling us how you are paying attention to the world.”
Munmun De Choudhury, a professor in the School of Interactive Computing at Georgia Tech, is also examining digital data for insights into well-being.
De Choudhury’s work over the years has focused not only on population studies, like the Hedonometer, but also on the individual.
In 2013, she and colleagues found that by looking at new mothers on social media, they were able to help predict which ones might develop postpartum depression, based on their posts before the birth of their babies. One of the most telling signs? The use of first-person singular pronouns, like “I” and “me.”
“If I’m constantly talking about ‘me,’ it means that my attention has inward focus,” De Choudhury said. “In the context of other markers, it can be a correlate of mental illness.”
This finding first emerged in the work of Pennebaker, but De Choudhury said that particular study was “eyeopening” for her. “We were pleasantly surprised that there is so much signal in someone’s social media feed that can help us make these predictions,” De
Choudhury said.
You may be wondering if Twitter is really a representative place to check the state of the general population’s mental health. After all, many of its users tend to refer to it by such nicknames as “hellsite” and “sewer.”
Some studies have shown that frequent social media use is correlated with depression and anxiety. Can we really discern our national happiness based on this particular digital environment and the fraction of the population — 1 in 5 in 2019 — that regularly use Twitter?
Angela Xiao Wu thinks we cannot. Wu, an assistant professor of media, culture and communication at NYU, argues that in the rush to embrace data, many researchers ignore the distorting effects of the platforms themselves.
We know that Twitter’s algorithms are designed to keep us hooked on our timelines, emotionally invested in the content we are presented with, coaxed toward remaining in a certain mental state. “If social scientists then take your resulting state, after all these interventions that these platforms have worked on you, and derive from that a national mood? There’s a huge part of platform incitement that’s embedded in the data, but is not being identified,” she said.
Indeed, Johannes Eichstaedt, a computational social scientist at Stanford, and a founder of the World Well Being Project, concedes that the methods like the ones his own lab uses are far from perfect. “I would say it’s about a C+,” he said. “It’s not that accurate, but it’s better than nothing.”