This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Mental Health, is properly cited. The complete bibliographic information, a link to the original publication on http://mental.jmir.org/, as well as this copyright and license information must be included.
Depression is the leading cause of diseases globally and is often characterized by a lack of social connection. With the rise of social media, it is seen that Twitter users are seeking Web-based connections for depression.
This study aimed to identify communities where Twitter users tweeted using the hashtag #MyDepressionLooksLike to connect about depression. Once identified, we wanted to understand which community characteristics correlated to Twitter users turning to a Web-based community to connect about depression.
Tweets were collected using NCapture software from May 25 to June 1, 2016 during the Mental Health Month (n=104) in the northeastern United States and Washington DC. After mapping tweets, we used a Poisson multilevel regression model to predict tweets per community (county) offset by the population and adjusted for percent female, percent population aged 15-44 years, percent white, percent below poverty, and percent single-person households. We then compared predicted versus observed counts and calculated tweeting index values (TIVs) to represent undertweeting and overtweeting. Last, we examined trends in community characteristics by TIV using Pearson correlation.
We found significant associations between tweet counts and area-level proportions of females, single-person households, and population aged 15-44 years. TIVs were lower than expected (TIV 1) in eastern, seaboard areas of the study region. There were communities tweeting as expected in the western, inland areas (TIV 2). Counties tweeting more than expected were generally scattered throughout the study region with a small cluster at the base of Maine. When examining community characteristics and overtweeting and undertweeting by county, we observed a clear upward gradient in several types of nonprofits and TIV values. However, we also observed U-shaped relationships for many community factors, suggesting that the same characteristics were correlated with both overtweeting and undertweeting.
Our findings suggest that Web-based communities, rather than replacing physical connection, may complement or serve as proxies for offline social communities, as seen through the consistent correlations between higher levels of tweeting and abundant nonprofits. Future research could expand the spatiotemporal scope to confirm these findings.
Each day, 313 million global users share 500 million messages or tweets on the popular social networking site Twitter [
Each year in the United States, about 7% of adults suffer from depression, but only half seek professional help [
Other research has focused on why users communicate about mental health through Web-based communities like Twitter. For example, the study #WhyWeTweetMH resulted in several themes for tweeting about mental health, including a sense of community, raising awareness, combating stigma, a space for expression, and coping and empowerment [
Possibly, lack of neighborhood amenities, such as parks and museums, may indicate fewer destinations for social interaction between residents. In this way, social bonds may be influenced by a neighborhood’s structural and functional characteristics [
Lack of neighborhood amenities and thus limited social interactions may be factors that influence use of Twitter for seeking connection about depression. To explore community characteristics that may lead to higher numbers of Twitter users seeking Web-based support, this study employed spatial data to understand how built and demographic features of communities relate to use of Twitter for depression support. Our findings may inform community efforts to increase social interaction and provide support for residents’ mental health. This study provides unique information that has not yet been analyzed by examining geographical location and community amenities in the context of tweets about depression.
Our research was determined to be Nonhuman Subjects research by the Michigan State University Institutional Review Board. All researchers involved received ethical training by the same institutional review board.
Tweets were downloaded through Twitter’s public streaming data [
Twitter restricted data capture to a random sample of 10% of total tweets from public content, and we compiled it into a database [
To predict expected tweets by county, we compiled demographic data commonly associated with depression diagnoses. We collected demographic data from the 2010 Census and the American Community Survey 2015 5-year estimates [
To understand the relationship between community characteristics and levels of tweeting, we compiled data for characteristics that could provide support, in-person treatment, opportunities for social interaction, or indicators of active community residents. These characteristics included parks or protected open spaces, places of worship, museums, active voter rates, mental health care providers, nonprofit organizations (organized by National Taxonomy of Exempt Entities code), K12 schools, and owner-occupied housing units. Likewise, we compiled data on characteristics that might hinder community support, including vacant housing units. Community characteristic data came from several sources (see
Nonprofit organizations were sorted into groups by purpose using Stata v11.1 (StataCorp, College Station, TX, USA): health, human services, public and societal benefit, religion, and education. Counts of nonprofit organizations by group, places of worship, mental health care providers, and vacant and owner-occupied housing units were summed by county. A rate per 100,000 population was then calculated. The spatial locations of museums, K12 schools, parks, and county boundaries were mapped using ArcGIS v10.5 (Esri, Redlands, CA, USA) [
Active voter totals were all from the fourth quarter of 2016. We aggregated active voter counts by county and then divided these counts by the active voter population. We then calculated the percent of active voters. For states that do not report active voter totals, we used that state’s voter turnout rate for the 2016 presidential election.
We assumed that, in theory, community levels of support seeking through Twitter should be predicted by demographic characteristics associated with depression and poor mental health [
Our focal interest was on understanding areas that have higher or lower tweeting than expected. So, after fitting the model, we calculated the difference between observed and expected counts and categorized these residuals into tweeting index value (TIV) tertiles, whereby 1=undertweeting, mean −0.37 (SD 0.53); 2=tweeting as expected, mean −0.02 (SD 0.01); and 3=overtweeting, mean 1.36 (SD 1.55). These TIVs were then used to examine community characteristic trends.
Demographic and community characteristic data sources.
Variable | Data sourcea | Year |
Percent aged 15-44 years | US Census Bureau | 2010 |
Percent female | US Census Bureau | 2010 |
Percent white | US Census Bureau | 2010 |
Percent single-person household | US Census Bureau | 2010 |
Percent below the poverty line | US Census Bureau (American Community Survey 5-year estimates) | 2015 |
Owner-occupied housing unit rate | US Census Bureau (American Community Survey 5-year estimates) | 2015 |
Vacant housing unit rate | US Census Bureau (American Community Survey 5-year estimates) | 2015 |
Places of worship rate | Association of Religion Data Archives | 2010 |
Mental health care providers rate | County Health Rankings & Roadmaps | 2015 |
Museum rate | Institute of Museum and Library Services | 2017 |
K12 schools per 100,000 children | United States Geological Survey | 2016 |
Percent active voter | State Board of Electors | 2016 |
Percent area occupied by park | State Geographic Information Systems Data Portals | Various |
Nonprofit organizations rates | Urban Institute National Center for Charitable Statistics Data Archives | 2005 |
School-aged population | US Census Bureau | 2010 |
County boundaries | United States Geological Survey | 2016 |
aTweet counts were collected by the research team and are not included here. Nonprofits were broken down into National Taxonomy of Exempt Entities classes for analysis, but these classes are not shown here. Rates per 100,000 total population unless otherwise noted.
During analysis, we decided to test whether our results might be driven by inclusion of one county—Washington, DC—because this county appeared to be an outlier for several reasons. For example, while its museums and other community amenities are very high, the population tends to be younger, and tweet counts were the highest. For this reason, we conducted a sensitivity analysis whereby the above model was also fitted without DC included. However, our incident rate ratios (IRRs) changed <1% for all independent variables (see
The suite of community characteristics outlined previously was selected due to each factor’s potential role in promoting or hindering support, in-person treatment, opportunities for social interaction, or indicators of active community residents. We calculated means of each characteristic by TIV. We then calculated a ratio of TIV5:1 and a Pearson correlation coefficient (
Mapping captured #MyDepressionLooksLike tweets revealed that most counties in the study region were not using the hashtag during the observation period (
Maryland, New York, and Pennsylvania had the highest tweet counts in our study region. However, their statewide means showed they were similar to other states. Washington, DC was an outlier in descriptive statistics of counties by state (
We found a positive, statistically significant correlation between several independent variables and tweet count (
Seeking to understand the relationship between the built environment and the levels of tweeting to connect about depression, we assembled a suite of community characteristics and related them to TIVs. We observed a
#MyDepressionLooksLike tweets by county for 11 states and the District of Columbia (2016).
Regression results used to calculate tweeting index values (n=245) for counties in the study region.
Independent variables | Incident rate ratio | 95% CI | |
Percent aged 15-44 years | 1.11 | .02 | 1.02-1.21 |
Percent female | 1.70 | <.001 | 1.25-2.29 |
Percent white population | 0.99 | .37 | 0.96-1.02 |
Percent single-person household | 0.90 | .03 | 0.82-0.99 |
Percent below poverty level | 1.05 | .31 | 0.95-1.16 |
Tweeting index values by county for 11 states and the District of Columbia. TIV: Tweeting index values.
Community characteristics by tweeting index values.
Community characteristics | TIVa1b | TIV2b | TIV3c | TIV3:TIV1 | ||
K12 schools per 100,000 children, mean | 199 | 289 | 212 | 1.07 | 0.19 | .004 |
Museums rate, meand | 17 | 33 | 20 | 1.18 | 0.20 | .002 |
Percent area occupied by park, mean | 9 | 9 | 10 | 1.11 | 0.04 | .49 |
Places of worship, meand | 104 | 177 | 93 | 0.89 | 0.16 | .01 |
Vacant housing rate, meand,e | 6 | 17 | 7 | 1.17 | 0.16 | .01 |
Owner-occupied housing rate, meand,e | 26 | 29 | 25 | 0.96 | 0.08 | .21 |
Percent active voter, mean | 76 | 72 | 74 | 0.97 | -0.15 | .02 |
Mental health care providers rate, meand | 209 | 157 | 293 | 1.40 | 0.09 | .18 |
Nonprofits (all) rate, meand | 412 | 527 | 579 | 1.41 | 0.23 | <.001 |
Nonprofits (health) rate, meand | 70 | 91 | 84 | 1.20 | 0.15 | .017 |
Nonprofits (human services) rate, meand | 147 | 184 | 195 | 1.33 | 0.23 | <.001 |
Nonprofits (public or societal benefit) rate, meand | 41 | 50 | 80 | 1.95 | 0.24 | <.001 |
Nonprofits (religious) rate, meand | 12 | 14 | 17 | 1.42 | 0.13 | .04 |
Nonprofits (education) rate, meand | 76 | 85 | 97 | 1.28 | 0.13 | .04 |
aTIV: tweeting index values.
bUndertweeting.
cOvertweeting.
dRate per 100,000 people.
eValues in thousands.
Most prior research on use of Twitter to connect about depression focused on interpersonal variables. This research focusing on community factors uniquely contributes to the literature. In this study, three positive, statistically significant associations were found to predict the count of tweets from people connecting about depression, including percent female, percent population aged 15-44 years, and percent single-person households. These findings correspond with other studies showing higher rates of depression among women and single-person households [
While we found several significant associations, some anticipated predictors of tweet counts were not confirmed. Percent below poverty level and percent white were not significantly associated with tweet counts. Our insignificant findings for area-level poverty contrast with some published data showing an association between poverty and depression [
In examining community characteristics related to overtweeting or undertweeting, rates of K12 schools, museums, places of worship, vacant housing rates, and health nonprofits exhibited a
This paper expanded the literature about depression in relation to Twitter by exploring and providing information about community characteristics that correlate with people turning to a Web-based community to connect about depression. Before this study, literature about depression relating to Twitter mainly consisted of how to interpret depression based on tweets and why people tweet about their mental health issues, such as depression. This included findings underlying the positive impact of Web-based platforms such as Twitter in discussing mental health issues, but also pointed to social media use and adverse consequences, including increased depressive symptoms. While often overlooked, people’s environment strongly impacts the state of their mental health [
Our study has several limitations. First, future research could expand the study area beyond the northeastern United States. As seen from Park et al’s 2013 study in South Korea, people from all over the world use Twitter to seek connections, so understanding their physical communities may also lend insight into the research question. Future studies could also expand the time frame from which tweets were gathered. Increasing the time frame for capturing data would allow this paper to have greater population validity and allow for inferences about how neighborhood changes relate to depression. Because of our limited spatiotemporal scope and use of NCapture, our sample size was smaller than originally anticipated, thus presenting generalizability concerns. Additionally, we used the geotagged location to characterize the community of Twitter users; this partially limited our sample and may have introduced error. Geotags may not be representative of the Twitter user’s community; rather, they could indicate places users were traveling to temporarily. Additionally, we used counties as a proxy for “community” since existing community-level data is often reported at the county level. In the future, we suggest a smaller geographic scale such as a city or town because these units are usually more representative of a person’s community.
Our study is one of the first to explore built and social environmental contributions to the use of Twitter to connect about depression. Communities that overtweet and undertweet were more likely to have lower rates of K12 schools, museums, places of worship, vacant housing rates, and health nonprofits. These communities were also likely to have higher rates of active voters. Especially evident in our study is that communities with higher rates of nonprofits exhibited higher than expected levels of tweeting—suggesting that lack of community investment may influence Web-based connection seeking. Urban planning efforts may usefully promote amenities to bolster social interactions and lessen isolation, thus ultimately offering opportunities for social support for depression.
Tables used for sensitivity analysis.
Descriptive statistics for counties in study region by state.
incident rate ratio
tweeting index value
This study was funded by Michigan State University’s Provost Undergraduate Research Initiative. We would like to thank Josh Vertalka, Katherine Bogen, Heather McCauley, and Amanda Rzotkiewicz for their contributions to the research process.
None declared.