Losing a job can take a toll on mental health. That’s a case that’s been made time and again.
For the first time, researchers are showing that this relationship can be seen in the geo-tagged tweets sent by Americans across the country.
At the Joint Statistical Meeting held in Boston last week, researchers from the U.S. Naval Surface Warfare Center and Johns Hopkins University presented early evidence that counties in the U.S. with higher rates of unemployment also had a higher proportion of Twitter users with depression-signifying language in their tweets.
That isn’t all. In counties with high populations of veterans, more Twitter users showed traces of post-traumatic stress disorder in their public posts. Counties with a lower median household income had a higher proportion of people who had tweets with signs of depression.
The new results suggest it may be possible to do a country-wide assessment of mental health from social media data that is already public, JHU researcher Glen Coppersmith says. His latest work tracked a sampling of tweets emerging from the 100 most densely populated counties in the U.S. for every day of 2013. He and his colleagues compared the results of their analysis with government data on unemployment and median income by region. And lo, correlation.
The new results are an initial demonstration of Coppersmith’s data-sifting tools, but his ultimate goal is to do an early assessment of a region’s mental health quickly and cheaply, and more efficiently than traditional mental health surveys.
The tools could come to use particularly in the aftermath of a natural disaster like Hurricane Katrina or the Boston Marathon Bombing. “If we were FEMA how many counselors should we be sending there?” Coppersmith, who studies computational linguistics, says. “That bit of information may be extremely useful in a traumatic event.”
In the last five years, many groups of researchers have tried to find reliable indicators of people’s mental health in their Twitter and Facebook profiles. At the get-go, they’ve been trying to establish some fundamentals: Does mental illness leave a mark in social feeds?
Coppersmith is one of them. He’s building evidence that algorithms analyzing language in people’s tweets can reveal depression-like language. It’s not a diagnosis, he’s careful to clarify, but an indication that something may be off, sort of like a spiking reading on a thermometer.
Some conditions do leave their stamp, he’s found. Algorithms can spot patterns in language that were indicative of depression, PTSD, seasonal affective disorder and bipolar disorder, he explained at a summer meeting of the Association for Computational Linguistics. Other disorders are more elusive: Alzheimer’s, for example, can’t be reliably traced, at least using current methods.
Coppersmith notes that public data on Twitter was used and that direct messages and private accounts were excluded from the analysis. He is also candid about the limitations to his findings so far. His team only analyzed tweets that are geo-tagged so the sample was restricted to people with smartphones. Also, Twitter users tend to be younger than those on other social networks, so he expects his sample is demographically skewed.