Jun 22, 2018 7:00 AM

Twitter Users Are Analytical in the Morning, Angsty at Night

What an analysis of 800 million tweets and 7 billion words published to Twitter between 2010 and 2014, across the 54 largest cities in the UK, reveals (and doesn't) about the British state of mind.

Wake up. Grab phone. Unlock. Open Twitter. Absorb tweets. Scroll. Absorb tweets. Scroll. Absorb tweets. What do they say?

They say:

X content

This content can also be viewed on the site it originates from.

Well yes, that. But what else?

Look closer. Disregard the topics; pay attention to the words. Soak up not just a few tweets, but a few million. Take them in not merely when you wake up, but every hour, every day, for years. What do you see now?

If you're Nello Christianini, you see patterns. Specifically, you see diurnal trends—daily rhythms—in the way humans use words. In a study published in the latest issue of PLOS ONE, Christianini and his colleagues analyzed 800 million tweets and some 7 billion words published to Twitter between 2010 and 2014, across the 54 largest cities in the UK, to study what they could reveal about the ways the British population thinks and feels on a 24-hour cycle.

"I don't want to know what specifically individual people are discussing," says Christianini, a computer scientist at the University of Bristol. "I want to know if I can measure trends in psychological states from a massive, textual time series of tweets." To know, that is, whether and how a society's expressions vary throughout the day.

The idea's not terribly far-fetched. Our circadian rhythms—the biological timekeepers of things like blood pressure, hormone levels, and metabolism—are coupled to our mental states. (Disruptions to circadian rhythms are strongly associated with psychological conditions ranging from major depression to seasonal affective disorder.) If you can deduce people's thoughts and emotions from what they say or do, perhaps you can detect their internal states by studying how their expressions change.

Social networks—public ones like Twitter, especially—make that kind of monitoring possible at a grand scale. They let you collect a lot of samples at high frequency without relying on self reports (humans are notoriously unreliable narrators of their lived experiences). Sure, the performative demands of the Twitterverse means that tweets probably aren't a pure reflection of self, either. But they are, to the sheer delight of computational social scientists everywhere, publicly available and in bountiful supply.

In fact, scientists have already dived in. In 2010, researchers at Northeastern and Harvard Universities analyzed 300 million tweets from across the US to see how Americans' moods fluctuate on a daily and weekly basis. A year later, Cornell sociologists Scott Golder and Michael Macy analyzed tweets from more than 2 million people in 84 countries.

Golder and Macy also tracked the language of individual Twitter users, so they could distinguish between an individual’s mood changes and broader trends in the population. That's an important distinction: Do happy people just tend to tweet in the morning, or are people, broadly speaking, happier in the morning? You can only know with this type of analysis, which appeared in Science as the first global-scale picture of how moods fluctuate across cultures on a daily, weekly, and seasonal basis.

Christianini's team's work takes the Twitter-mining approach in new directions by monitoring not just changes in mood, but in thinking styles. "Mood is only one little part," Christianini says. "This is about cognitive processes, it's about concerns and interests."

To do it, they compared their harvested tweets against the Linguistic Inquiry and Word Count, a text-analysis app that correlates large lists of words and word stems with specific aspects of human psychology. Macy and Golder's study looked at two of the most well-studied lists, which are associated with positive and negative affects. But Christianini's team cross referenced their Twitter data with all 73 of the variables defined within the LIWC, including processes (like certainty and tentativeness), emotions (like anxiety and anger), social concerns (like family and friends), and time-orientation (like past- or future-focused).

Their findings reflect not just variations in mood, but styles of thought. Analytical thinking—which correlates with frequent use of nouns, articles, and prepositions—seems to peak early in the day, along with an increased concern with things like power and achievement. Late at night, however, existential thinking dominates. By 3:00 am, positive emotions are at their lowest, and topics like death and religion have peaked. At the population level, anyway. Well, the British population level.

That's the thing about these findings: They can only tell us so much about the circadian bases for our daily mental states. "I applaud the fact they went so far beyond the two dimensions that [Golder] and I looked at—I think that's terrific, and it needed to be done," says Macy, coauthor of the cross-cultural tweet-analysis from Science. "But in many ways their methods are a bit of a throwback."

How so? The geographic limitation, for one, raises questions about whether the variations are cultural or biological in origin. Some of the psychometric variables in the LIWC are less well-characterized than others, which limits the conclusions you can draw from their overlap with tweets. And, perhaps most significantly, because the study didn't monitor individuals, it can’t tell whether the 3 am spike is because people's thoughts generally turn toward topics like death and religion after dark, or if existentialists just prefer to tweet late at night.

Future studies would benefit from an international scope, and the tracking of individual users' tweets, though Christianini says that last bit could be tricky—especially in the post–Cambridge Analytica age, when psychometric research has become uncomfortably associated with privacy violations. "I am very proud: I don't follow individuals," Christianini says. “I look at anonymized, aggregated collected data. So I cannot know if it’s the same person changing or not.”

But someone else could. Golder and Macy did. Twitter, after all, is a public medium. If you're worried about subject privacy, you could always anonymize your user data by, say, assigning each tweeter a number. Security experts will tell you de-identification techniques are never a perfectly secure bet—but then even ethical investigations always carry some measure of risk.

You can imagine how understanding people's thoughts and emotions could be used for good: to inform clinical interventions for mood management, say, or help improve productivity. But you can also imagine employers leveraging psychometric techniques to screen job applicants for certain personality types. "The truth is, and I say this as a concerned scientist, if something is possible, technically legal, and profitable, someone will do it," Christianini says.

Macy agrees. "The fact that you or I can't think of a way that someone will misuse this data doesn't mean someone else won't."