Understanding and Measuring Psychological Stress Using Social Media

Sharath Chandra Guntuku,1 Anneke Buffone,1 Kokil Jaidka2

Johannes C. Eichstaedt,1 Lyle H. Ungar1
1
University of Pennsylvania, 2 Nanyang Technological University
{sharathg@sas, annekeb@sas, jeich@sas, ungar@cis}.upenn.edu,[email protected]
arXiv:1811.07430v2 [cs.CL] 4 Apr 2019

Abstract
A body of literature has demonstrated that users’ mental health
conditions, such as depression and anxiety, can be predicted
from their social media language. There is still a gap in the sci-
entific understanding of how psychological stress is expressed
on social media. Stress is one of the primary underlying causes
and correlates of chronic physical illnesses and mental health
conditions. In this paper, we explore the language of psycho-
logical stress with a dataset of 601 social media users, who
answered the Perceived Stress Scale questionnaire and also
consented to share their Facebook and Twitter data. Firstly, Figure 1: Overview of the approach taken to build language
we find that stressed users post about exhaustion, losing con- based stress prediction model.
trol, increased self-focus and physical pain as compared to
posts about breakfast, family-time, and travel by users who
are not stressed. Secondly, we find that Facebook language is
more predictive of stress than Twitter language. Thirdly, we very severe, alleviating psychological stress and promoting
demonstrate how the language based models thus developed a healthy lifestyle is more efficient compared to treating a
can be adapted and be scaled to measure county-level trends. more acute and chronic condition (Wilkinson 2005).
Since county-level language is easily available on Twitter us- People are increasingly using social media platforms in
ing the Streaming API, we explore multiple domain adaptation
algorithms to adapt user-level Facebook models to Twitter lan-
order to inform others about their mental states, solicit so-
guage. We find that domain-adapted and scaled social media- cial support, as well as keep records of their daily activities,
based measurements of stress outperform sociodemographic preferences, and interests. Notwithstanding the challenge of
variables (age, gender, race, education, and income), against working with a non-random, non-representative sample of
ground-truth survey-based stress measurements, both at the social media users, studies have identified the markers of self-
user- and the county-level in the U.S. Twitter language that disclosure which concern depression (Guntuku et al. 2017c),
scores higher in stress is also predictive of poorer health, less schizophrenia (Ernala et al. 2017), ADHD (Guntuku et al.
access to facilities and lower socioeconomic status in counties. 2017b), alcohol consumption (Liu, Weitzman, and Chunara
We conclude with a discussion of the implications of using 2017), and personality (Guntuku et al. 2017a). With respect to
social media as a new tool for monitoring stress levels of both stress, the linguistic features of event-related stress have been
individuals and counties.
predicted from social media posts about experiences such as
travel and work (Lin et al. 2014); however, these findings
Introduction cannot be applied to improve the psychological understand-
ing of stress, because people suffering from chronic stress
Stress is defined as perceived distress caused by an interaction
do so irrespective of stressful events. For instance, preparing
between a person and their environment (Cohen, Miller, and
for an exam is a stressful event, while chronically feeling
Rabin 2001). While people can handle stress better or worse
overwhelmed with responsibilities is trait-related stress. An-
depending on their general coping skills, experiencing stress
other research gap is that the previous work has focused on
too frequently is known to affect well-being and physical
known stressors collected using search keywords (Thelwall
and mental health negatively. Stress is seen to be a single
2017). However the labels thus acquired likely have person-
pervasive trait influencing health, through a broad range of
ality confounds (Preotiuc-Pietro et al. 2015), emphasizing
negative affective states and somatic pathways (McEwen
the need for using stronger ground truth. Instead, we antici-
and Stellar 1993). Considering that the symptoms associated
pate that insights into psychological stress could help in (a)
with depression and other severe mental health conditions are
designing social-media-based interventions to enable a low-
Copyright c 2019, Association for the Advancement of Artificial stress lifestyle, and (b) developing a better understanding of
Intelligence (www.aaai.org). All rights reserved. regional variations in stress.
Table 1: Items on the Cohen’s stress scale: Each question is from a large number of people for this research study. Adapt-
assessed on a Likert Scale. (-) indicates reverse coded items. ing models trained at user-level to predict stress in counties
In the last month, how often have you: has multiple applications in monitoring health and well-being
- been upset because of something that happened unexpectedly? in counties, especially where survey data is hard to collect.
- felt that you were unable to control the important things in your life?
- felt nervous and ”stressed”? Thus motivated, in this paper we address the following re-
- felt confident about your ability to handle your personal problems? (-) search questions:
- felt that things were going your way? (-)
- found that you could not cope with all the things that you had to do? RQ1 How does social media language of users who are
- been able to control irritations in your life? (-) stressed differ from those who are not?
- felt that you were on top of things?(-)
- been angered because of things that were outside of your control? RQ2 How do Facebook and Twitter language differ in pre-
- felt difficulties were piling up so high that you could not overcome them? dicting user-level stress? Since county-level language
is easily available on Twitter, can off-the-shelf domain
adaptation algorithms be used to improve prediction
The ubiquitous nature of smart devices and Internet ac- on Twitter language?
cess in almost all parts of the world means that social media RQ3 How do domain-adapted social media-based measure-
is a potentially powerful tool to measure the psychological ments of stress at the county level correlate with health
states and behaviors of people at both micro- (individual) and behaviors and socioeconomic characteristics at the
macro- (county) levels. However, except for a few studies, county level?
little has been done to explore how to scale language models
to study regions, and no work yet has attempted to do this for
stress. Although some studies have analyzed the geographic RQ1: Differential Language Analysis of
variation in social media language corresponding to chronic Stressed Users
illnesses (Culotta 2014), depression (Bagroy, Kumaraguru, Methods
and De Choudhury 2017), well-being (Schwartz et al. 2013a),
heart-disease (Eichstaedt et al. 2015), and happiness (Quer- User-level social media data: We deployed a survey on
cia et al. 2012), they are often challenged by (a) a limited Qualtrics2 (a platform similar to Amazon Mechanical Turk),
understanding of how to scale user-level models to measure comprising several demographic questions (age, gender, race,
counties, and (b) the lack of ground truth about a large pop- education, and income) and the Cohen’s 10-item Stress scale
ulation. Recent work has shown the need to use weighting (Cohen, Kessler, and Gordon 1997) (Table 1), and invited
and scaling techniques to transform user-level language esti- users to share access to their Facebook status updates and/or
mates from Twitter to county-level estimates (Rieman et al. Twitter usernames. Users received an incentive for their par-
2017), in order to avoid the ecological fallacies reported in ticipation, and we obtained their informed consent to access
the studies mentioned above. However, there are challenges their Facebook and Twitter posts. All users were based in
associated with transferring predictive models from one so- the US. This study received approval from the IRB of our
cial media platform to another, because of differences in institution.
self-disclosure on Facebook vs. Twitter (Jaidka et al. 2018). Out of all users who took the survey, 601 users completed
This study contributes with (a) effective off-the-shelf meth- the survey and had active accounts with more than 900 words
ods for cross-domain adaptation of user-level stress models on both Facebook and Twitter. We collected their Facebook
to measure county-level stress, and (b) validation against posts by using the Facebook Graph API and downloaded
ground-truth survey data built on over two million responses. their Twitter posts using the Twitter API. Of these 601 users,
To summarize, the research gaps in previous work are the 265 self-identified as female. The mean age of the sample
lack of language models to predict psychological stress, a was 38. The stress scores range from 6 to 39 (mean 30). Each
limited understanding of scaling and transforming Facebook item in the scale is scored on 0-4, with an absolute maximum
language models to work on Twitter, and the lack of valida- summing to 40.
tion against the region-level ground truth. We show in Figure Features: We process all our social media posts using the
1, how our study uses transfer learning to adapt user-level HappierFunTokenizer available with the DLATK package
models trained on their Facebook language, to predict county- (Schwartz et al. 2017) which is emoticon- and social media-
level stress from the county-level Twitter language. This is aware. We then represent the language of each user and
necessary because it is easier to train predictive models on county as a set of features. In the dictionary-based method,
the language of a small population of social media users, we transform social media language into numerical features
but it is expensive to survey entire counties for training pur- representing percentage proportions of lexical categories in
poses. Furthermore, county-level social media language is an existing dictionary. In the data-driven method, we trans-
only available for Twitter, where approximately 20% of all form language into numerical features which represent the
public posts are geo-tagged with their location information, proportions of word clusters which are statistically similar
and they can be easily mined by using Twitter’s Streaming according to their frequency distributions.
API1 . On the contrary, Facebook needs user authentication LIWC: We use Linguistic Inquiry and Word Count
for accessing their posts which is resource intensive to collect (LIWC) (Pennebaker, Booth, and Francis 2007), a dictionary
1 2
https://developer.twitter.com/en www.qualtrics.com/Survey-Software
Figure 2: Words and phrases associated with a) high-stress users (red) and b) low-stress users (blue). The size of the word
indicates the correlation strength and the color indicates frequency (darker is more frequent). Correlations are controlled for age
and gender, and are significant at p < .01, two-tailed t-test Bonferroni p-correction.

comprising 73 different psycholinguistic categories (e.g., top- ful stimulus and determining own coping abilities (Staal
ical categories, emotions, parts-of-speech) to represent the 2004), this result is very intuitive. It should be noted that
language of each user and county as the normalized frequency increased self-focus in stressful situations is likely adaptive,
distributions. but a prolonged self-focus in one’s thoughts, especially in
Topics: In the data-driven method, we represent the lan- the context of negative affect, is known to psychologists
guage of each user and county as normalized frequency dis- as rumination and linked to negative effects for health and
tributions for a set of topics derived using Latent Dirichlet well-being (Moberly and Watkins 2008). Hate could signify
Allocation. These topics are an open-source resource avail- aggressive or angry affect, which also tends to be maladap-
able through the DLATK package (Schwartz et al. 2017) and tive and could signify the experience of frustration while
were trained on a corpus of over 20 million Facebook statuses perceiving resources fall short of perceived demands.
(Schwartz et al. 2013b). Words and phrases such as “me”, ”I had”, “feel like”, ”I
TensiStrength: To detect direct and indirect expressions don’t” and ”I hate” are significantly correlated with high
of stress or relaxation, a stress lexicon (Thelwall 2017) is stress. Furthermore, the language of high stress appears to
used. We obtain stress scores at the sentence level, from each be marked by expressions of perceived lack of control and
post on Facebook and Twitter and aggregate them to users by expressions of a need state or lack of resources (“struggling
calculating mean stress scores. with”, “tired of”, “I need”), as well as negative-angry affect
User engagement We extracted features such as the num- (“hate”, “I hate”). Further, high stress language seem to be
ber of posts made between 12am-6am, mean message length, comorbid with mental health conditions (“depression”, “de-
and number of URLs & hashtags. These features were shown pressed”, “anxiety”, “bipolar”). It is interesting that language
to be predictive of stress by (Lin et al. 2014). reflects the adverse effects that stress can have on health (Wat-
son and Pennebaker 1989). The language of low stress has
Results prominent positive affect (“excited to”), discussions of meals
We identify the linguistic characteristics indicative of high (“breakfast”), as well as feelings of social inclusion (“joined
and low psychological stress based on the social media lan- the”). The language of low stress comprises the discussion
guage of individual users on Facebook. We conducted a sim- of meals, specifically “breakfast”, which may indicate relax-
ilar analysis on Twitter and present them in the appendix. ation and enjoyment inasmuch as meals are often taken in the
Since we explore several features simultaneously, we con- company of others, social inclusion, as well as with positive
sider coefficients significant if they are less than a Bonferroni- affect (“excited to”).
corrected two-tailed p of 0.01 (i.e., when examining 1000 LIWC: LIWC categories from Facebook language that sig-
features, in the case of words and phrases, a passing p-value nificantly correlate with stress are shown in Table 2. The top
is less than 1x10−5 ; when examining 2000 topics p-value is correlated categories in Facebook comprise first person sin-
less than 5x10−6 , and when examining 73 LIWC categories gular pronouns (“I”), indicating increased self-focus, which
p-value is less than 1.3x10−4 ). corroborate previous findings (Pennebaker and Lay 2002)
Ngrams: We extract 1-,2-, and 3-grams from all posts to that self-references by individuals increase in emotionally
analyze significant associations between words & phrases vulnerable situations. Users on Facebook are more likely
and stress. Figure 2 provides a visualization of the Pearson to use adverbs such as “very” and “really” to emphasize
correlations of words and phrases with stress on Facebook. their point, and more likely to explicitly use words denoting
In general, the language of stress has a prominent self-focus. negative emotions, such as “hurt” and “anger”. Mentions of
Given that psychologists have found stress to be a state result- negative emotion confirms our expectations, because stress is
ing from an individual assessing the demands of the stress- an aversive state (Cohen, Kessler, and Gordon 1997). Filler
Table 2: Pearson correlations (top 5) between stress score and Table 3: Pearson correlations between stress score and Topics
LIWC features extracted on Facebook, when controlling for age and extracted from Facebook, when controlling for age and gender.
gender. All features are significant at p < .01, two-tailed t-test and Topic labels are manually created. All features are significant at
Bonferroni corrected. p < .01, two tailed t-test and Bonferroni corrected. The top 5 topics
LIWC are shown for both positive and negative correlations.
Positively Topics
most frequent words r
correlated Positively most frequent words used by high-stress
r
1st Person Singular I .22 correlated people
Adverbs very, really .17 i’m, tired, hungry, bored, exhausted,
Negation no, not, never .16 Exhaustion freaking, sleepy, sooooooooo, stressed, .21
Negative Emotion hurt, sadness, anger .17 grumpy
Fillers I mean, you know .15 i’m, sick, tired, feeling, hearing, tire, fed,
Feeling hurt .21
Negatively bullshit, assuming, hurting, numb
most frequent words r don’t, feel, anymore, beg, begging, hon-
correlated
1st Person Plural we -.18 Physical Pain estly, knees, clue, hollow, creeping, sym- .20
Affiliation friend, social -.17 pathy
Positive Emotion love, nice, sweet -.12 feel, sick, crap, feeling, ugh, hate, feels,
Feeling sick sucks, crappy, bleh, worse, miserable, .19
sickness, icky, =(
i’m, it’s, i’ve, don’t, lost, mind, quarter,
word associations may indicate a lack of self-esteem or feel- Losing control .19
wouldn’t, anymore, control, reason
ing down on oneself, and hedging as a result (Abouserie Negatively most frequent words used by low-stress
1994). Negation further suggests a ’lack of’ things, men- r
correlated people
tioned and experienced by those who are also prone to high great, lunch, nice, dinner, family, en-
stress. On the other hand, the first person plural pronouns joyed, church, wonderful, afternoon, sun-
Family and eating .13
(“We”), and the LIWC “Affiliation” categories are negatively day, kids, evening, shopping, meeting,
correlated with stress, which implies that a high-stress indi- hubby
vidual often depicts themselves as isolated, with a certain kings, delhi, leon, mumbai, rocks,
disassociation from their social circles. Negative correlations Travel queens, reached, rains, phew, bak, travel- .12
of stress with positive emotion reflects that those individuals ling, royal, metro, indians, mahal
reporting lower stress are significantly more likely to express
positive emotion.
Topics: Topics which significantly correlate with stress (a) LIWC features and (b) Topic features (c) TensiStrength
are provided in Table 3. Exhaustion is typical of prolonged scores, and (d) engagement features (such as time of posts,
stress (McManus, Winder, and Gordon 2002). Feeling hurt, number of posts, number of posts between 12am-6am) for
physical pain, and feeling sick are known correlates of stress users in four folds, and testing on the users in the held out
(Gil et al. 2004). Lack of control also signifies the concept fold.
of resources falling short of demands, and possibly reducing In the five-fold cross-validation setting, we perform linear
their ability to cope with stressors (Gray, Waytz, and Young regression with several regularization methods such as ridge,
2012). elastic-net, LASSO and L2 penalized SVMs and find that
Our findings also reiterate previous research on language elastic-net showed marginally superior performance over the
use in mental health (De Choudhury et al. 2013), as they others. Accordingly, we report results only using elastic-net.
describe symptoms of physical pain and sickness, besides The performance was measured by calculating Pearson’s r
expressing a lack of control and negative emotions. On the over the aggregated predictions from the five folds.
other hand, the mention of family meals on Facebook has We first evaluate the above features to predict stress within
a negative correlation with stress, suggesting a good social domain - i.e., we train and test on the same platform. Engage-
support network and taking time to spend with loved ones as ment features are part of the feature set used in (Lin et al.
well as travel, are ways people relax. 2014) for predicting user-level stress. We also examine how
the models perform when compared to sociodemographic
RQ2: Predicting User-Level Psychological variables, namely age, gender, race, income, and education.
Stress using Facebook and Twitter Then, we evaluate how models trained on Facebook per-
form at predicting stress from Twitter language in a cross-
Methods domain setting. Previous studies showed that predictive per-
We utilize the Facebook and Twitter data from the same users, formance changes in cross-domain applications (Jaidka et
described in the previous section, to evaluate the performance al. 2018; Zhong et al. 2017). Therefore, we then attempt to
of supervised models trained on Facebook and Twitter lan- improve the cross-domain prediction performance by apply-
guage at predicting users’ stress for a held-out set. We stratify ing domain adaptation. The motivation for cross-domain and
our set of users into five folds with a uniform distribution domain adaptation experiments is to build an accurate pre-
of age and gender traits in each fold. We conduct a cross- dictive model on Twitter language to then scale it to county
validated weighted linear regression for stress, training on level Twitter language.
Table 4: Within Domain: Stress Prediction Performance (Pearson’s as LIWC is stable across both platforms. We also combined
r) based on 5-fold cross-validation using different features and mod- both Facebook and Twitter corpora, and the model trained
els trained and tested on the same domain. Social media language on both platforms together gives a marginal improvement in
adds to and outperforms sociodemographic variables at predicting prediction performance.
stress. Given the marked distinction between expressions of stress
D: Sociodemographic variables in Facebook and Twitter language, we investigate several ap-
Feature Pearson r proaches to improve the predictive performance of models on
Age, Gender, Race Twitter. This problem can be viewed as a domain adaptation
.25
Income & Education task, where we are adapting from a source domain: users’
SM: Social media language language on one platform, to a target domain: the same users’
Feature Facebook Twitter language on another platform.
User Engagement (Lin et al. 2014) .11 .05
TensiStrength (Thelwall 2017) .17 .11 Table 5: Cross Domain: Stress Prediction Performance (Pear-
LIWC 2015 .29 .22 son’s r) based on 5-fold cross-validation. FB: Facebook; Tw:
Twitter.
Topics .31 .18 Trained on FB Trained on FB+Tw
Language + Sociodemographic Feature
Tested on Tw Tested on Tw
D + SM .33 .26 LIWC 2015 .23 .24
Topics .15 .17
Results for Within Domain Predictions
Table 4 shows the performance of predicting stress using so- Results with Domain adaptation
ciodemographic variables and social media language. Within
the domain, Facebook does better than Twitter by a slight Most prior work on domain adaptation has focused on the
margin. On Facebook, both LIWC and Topics outperform case where some labels are available on both the source and
sociodemographic variables (age, gender, race, income, and target domains and is usually done by combining (often in
education). Topics outperform LIWC on Facebook (r=.305), some weighted fashion) training data sets or, less commonly,
and LIWC outperforms Topics on Twitter (r=.218). These cor- trained models from the source and target domains. A simple
relations, which are considered a high correlation in measur- and effective supervised method was proposed by Daumé
ing internal traits (Meyer, Finn, and others 2001), show that (Daumé III, Kumar, and Saha 2010) which applied a super-
linguistic features perform reliably well at predicting stress. vised heuristic mapping from labeled data the source and
User engagement features performed rather poorly when com- target domains, to a higher dimensional feature space, which
pared to linguistic features indicating that trait-prediction is are used to train standard classifiers or linear regression mod-
a different task when compared to state-prediction. A similar els. This approach has demonstrated the best performance in
observation has been made while using user engagement fea- several comparative evaluations conducted for image-, text-
tures for predicting other traits (Preotiuc-Pietro et al. 2016). and sentiment-classification (Pan and Yang 2010).
The correlation between stress scores collected from sur- We test two standard domain adaptation techniques for im-
vey and user-aggregated TensiStrength scores was .17 for proving the cross-platform performance of predictive models:
Facebook and .11 for Twitter. TensiStrength was developed one supervised approach; Easy Adapt (Daumé III, Kumar,
by annotating keyword-selected Twitter posts, which makes and Saha 2010), which uses labeled data from both the source
them inapplicable for measuring psychological stress. Conse- and the target, and one unsupervised approach: Transfer Com-
quently, we dropped engagement features and TensiStrength ponent Analysis – TCA (Pan et al. 2011), which requires no
from subsequent analysis. We also limited the posting period labels on the target domain. While there are several candi-
to the previous one month (consistent with the time period date algorithms within both supervised and unsupervised
in the Cohen’s survey questionnaire) and observed that the approaches, we chose Easy Adapt and TCA for their simplic-
performance shows a non-significant drop by 0.02 in r. ity in application.
The standard notation used throughout this section is that
Results for Cross-Domain Predictions XS refers to labeled observations in the source domain and
XT refers to the test set in the target domain.
Since our goal is to predict stress in counties using Twitter, we
examined how models trained on Facebook perform on Twit-
ter (shown in Table 5). Performance drops (by 5% compared EasyAdapt : We define our problem according to the im-
to within domain performance) when Facebook models are plementation described by Daumé (Daumé III, Kumar, and
used to predict stress from the Twitter language. Specifically, Saha 2010). Let X1 denote the original feature space for
topics see a larger drop (by 50%) possibly due to not being the user in the source domain, X1 = RF . We construct an
able to generalize across platforms. We have seen that expres- augmented feature space X1f = R3F , by creating an platform-
sions of stress in Facebook language vs. Twitter language specific, and user-specific version of each feature in X. For
are significantly different in terms of the vocabulary used. this, we define Φs : X −→ X e to transform feature vec-
Consequently, a standardized theory-driven dictionary such tors corresponding to the platform-specific and user-specific
feature spaces. The mappings are defined by the following is used in order to ensure that any word-to-outcome corre-
equation: lations observed were stable, resulting in a dataset of 2710
counties. On an average, each county had 8,892,568 words.
Φs (x) = x, x, 0 , Φt (x) = x, 0, x (1) County Health outcomes: We use two sets of county
s
Here, Φ (x) is the feature vector for the source domain, health outcomes in our work: 1. Gallup-Sharecare Well-Being
Φt (x) is the corresponding vector for the target domain. data and 2. County health statistics for the United States.
0 = 0, 0, ..., 0 ∈ RF is the zero vector. From the Gallup-Sharecare Well-Being Index, we use the
We thus augment the original feature space with the labeled stress outcome to validate the stress predictions made by our
data points (for the same set of individuals) from the target language model. Gallup data is collected as a part of 1,000
domain, excluding the held out sample for testing. Next, we telephone interviews conducted every day across the US
train a regression model between this augmented feature (Huppert and So 2010). The questions on the interview range
space and stress scores of the users. We similarly transform from topics such as health behaviors, work environment to
the feature space of the held out sample before prediction. social and county factors, and financial security. We specifi-
cally used the stress fields from the Gallup data aggregated
to counties.
Transfer Component Analysis (TCA) TCA exploits the From US County Health Rankings and Roadmaps portal3 ,
Maximum Mean Discrepancy Embedding (MMDE) for com- we obtained county socioeconomic characteristics and health
paring the distributions between the source and target domain, behaviors which provides access to county-level health fac-
based on the Reproducing Kernel Hilbert Space (RKHS). tors from a wide range of sources, including Behavioral Risk
The empirical estimate of the distance between DS and DT , Factor Surveillance System, American counties Survey, and
Dist(DS , DT ), is the National Center for Health Statistics.
n1 n2
1 X 1 X
|| si − ui ||H (2) Results
n1 i=1 n2 i=1
After analyzing the performance of different models at pre-
where u and s are individual observations in DS and DU , H dicting stress in the cross-domain setting, we use the domain-
is a universal RKHS and φ : X −→H. adapted model trained on LIWC features to predict stress
After applying domain adaptation (results in Table 6), per- using Twitter language from US counties (shown in Figure 3).
formance increases by 16% when compared to using Face- We validate our language-predicted stress with county-level
book models alone to predict stress from Twitter language stress reported by Gallup, and further with county-level esti-
without any domain adaptation. TCA is seen to outperform mates of health behaviors and socioeconomic characteristics.
EasyAdapt.

Table 6: Domain Adaptation: Stress Prediction Performance

(Pearson’s r) based on 5-fold cross-validation, trained on
Facebook and tested on Twitter.
No Domain Transfer Component
Feature EasyAdapt
Adaptation Analysis
LIWC 2015 .23 .23 .27
Topics .15 .16 .18

RQ3 Evaluating County Level

Language-Predicted Stress
Methods Figure 3: Map showing stress predicted from Twitter lan-
Data The language of US counties is obtained from Twit- guage from US counties. Counties which had less than 100k
ter. It comprises of the normalized frequency distributions of tweets are shown white.
words, averaged by the number of individuals who spoke
those words, extracted from a set of geo-located tweets First, we use the Gallup-Sharecare Well-Being data, with
(Schwartz et al. 2013a). The authors constructed this dataset county-aggregated stress labels (termed as ‘Gallup stress’
by collecting a 10% sample of Twitter posts between the hereon), as a means of validating our language-predicted
years 2009-15 from the TrendMiner project (Preotiuc-Pietro stress. We compare Gallup stress and language-predicted
et al. 2012). Tweets were geo-mapped to US counties using stress (referred to as ‘Twitter stress’ hereon). We aim to train
latitude/longitude coordinates and the self-reported location the best performing user-level model, as this directly impacts
field when available, following the approach described in the county-level prediction accuracy. We experiment with
(Schwartz et al. 2013a). In this manner, 20% of the tweets models trained on individual platforms (best r = 0.26), and
could be successfully mapped, resulting in over 1.7 billion
3
geo-located tweets. A threshold of 100,000 words per county www.countyhealthrankings.org/
find a significant difference in predicting Gallup stress after Table 7: Correlation between county level health measures
domain adaptation (r = 0.34). These correlations are con- and stress. All correlations are significant at p < 0.01.
sidered a high correlation in psychology, especially when Language-based stress predictions predict health behav-
measuring outcomes like stress (Meyer, Finn, and others iors and clinical care facilities above and beyond socio-
2001). Part of this improvement is due to the amount of demographic controls
information that Facebook contains over Twitter (Table 4). County-level Outcomes Language Correlation
However, Facebook alone is insufficient as the Twitter lan- With With SES
guage is substantially different from Facebook (Tables 5 and No Controls controls
6). Health Behaviors
Predictive performance of sociodemographic variables MV Mortality Rate .40 .11
(age, gender, income and education) and Twitter language are % Physically Inactive .44 .31
r=.24 and r=.34 respectivelt. Twitter language outperforms % Smokers .29 .30
sociodemographic variables at predicting stress in counties. % Poor mental health days .14 .13
Teen Birth Rate .23 -
Having validated the language-estimated stress in counties,
Drug Poisoning Mortality Rate .13 -
we examine the correlation of stress with socioeconomic char-
Clinical Care
acteristics and health behaviors in US counties. The results Preventable Hosp. Rate .41 .36
are shown in Table 7. Mental health providers rate -.39 -.19
Primary Care Physician Rate -.43 -.33
Health Behaviors Counties with higher stress have higher % Could not see doctor due to
.17 -
mortality, a higher percentage of physically inactive peo- cost
ple, a higher percentage of poor mental health days, and a Socioeconomic Environment
higher percentage of smokers, teen pregnancy, and higher Median Household Income (ln) -.27 -
Education (% College
rates of drug poisoning. These are all factors which have -.34 -
graduates)
known correlation and causality structures as reported by % With Access to Exercise
(Stults-Kolehmainen and Sinha 2014; Steptoe et al. 1996; -.45 -.264
Facilities
Watson and Pennebaker 1989). Individuals living in high-
stress counties appear to suffer from poor health, affected
both by aspects in their individual lives as well as by the ity with mental health conditions (Watson and Pennebaker
nature of their counties. Using social media language as a 1989), pronounced self-focus (Pennebaker and Lay 2002)
real-time proxy to assess a county’s stress, and consequently, and feelings of hurt, physical pain, and being sick, which
health, could help design interventions and policies to allevi- also known correlates of stress (Gil et al. 2004). Our find-
ate stressors. Language has the specific advantage of getting ings also reiterate previous research on language use in
at the causes/stressors that counties are experiencing, and this mental health (De Choudhury et al. 2013). An important
could greatly complement and also inform traditional survey insight from our work was that the predictive utility of dif-
mechanisms (such as Gallup). ferent platforms varies, as has been seen in prior research
looking at differences in self-disclosure (Jaidka et al. 2018;
Socioeconomic Characteristics Counties which are rich Zhong et al. 2017). This motivates the need for transfer learn-
and educated express less stress (Karlamangla, Singer, and ing when the stress model trained on Facebook language
Seeman 2006). This is observed by the low correlation be- needs to be applied to Twitter language. This observation has
tween language-predicted stress and Median Household In- implications for other researchers interested in training and
come (r=-.271) and Education level, % people who attended applying machine learning models across domains.
some college, (r=.345) in counties. Furthermore, counties On applying the adapted stress model to the language of
where people have access to exercise facilities also have low counties, we obtain face-valid results about the relationships
stress (Salmon 2001). Our language-based findings of stress between stress and ground-truth statistics about county health.
and income are corroborated by an analysis of the Gallup data Our findings clarify the association of stress with socioeco-
reported by (De Neve and Ward 2017). High-stress counties nomic factors (Baum, Garofalo, and Yali 1999), suggesting
also have higher avoidable hospitalization rates and have less that it is the socially and economically deprived counties
access to providers for mental and physical health. These which are likely to face higher stress. Urban counties, with
associations suggest that individuals in high-stress counties pressures of work and a fast-paced routine, are less likely to
also face other economic and infrastructural challenges. experience debilitating, psychological stress (Lederbogen et
al. 2011). This is an important way to differentiate the stress-
Discussion ful events in an urban lifestyle with the trait-based stress
Our work provides some of the first insights into the lin- experienced in deprived or rural counties.
guistic manifestation of psychological stress on social me-
dia, while also showing the feasibility of applying models Limitations and Future Work
trained at the user-level to predict stress in counties with While there are several advantages and merits of using social
validity. Language associated with stress is indicative of ex- media to measure stress, one of the apparent shortcomings of
haustion (McManus, Winder, and Gordon 2002), comorbid- linguistic analysis is that they do not offer causal insights. Our
methods anticipate that the language of counties is indicative (Sharma and Gedeon 2012; Guntuku, Roy, and Weisi 2015;
of stress. It is possible that people might stop using social Samani et al. 2018), particularly transfering learning across
media while they are stressed, which could impact the perfor- modalities (Guntuku et al. 2019).
mance of our user-level model (Corrigan et al. 2001). How-
ever, we did not find any correlation (positive or negative) Privacy and Ethical Considerations
between the raw number of posts and psychological stress. Prediction performances obtained in this study indicate that
To address such scenarios, monitoring multiple data sources psychological stress can be inferred with some accuracy from
such as phone-based sensors can be useful in promptly reach- public (Twitter) or semi-public (Facebook) social media data.
ing out to the afflicted individuals (Singh and Agarwal 2016). While the purpose of delivering social support and mental
Further, it would be interesting to disentangle the language health services motivates this effort, these algorithms also
indicative of short-term stressors versus long-term chronic raise some important privacy and ethical questions.
stress, potentially by collecting data around recent stressful From the perspective of privacy concerns, organizations
events (divorce, moving house, losing job etc.). with vested interests (e.g. insurance companies) may be moti-
Some studies have shown preliminary findings that con- vated to infer this information automatically. As being chron-
nect social media usage to poorer mental health, indicating ically stressed can lead to stigma at multiple levels, data pro-
that undesirable usage of social media might cause depres- tection and ownership frameworks are necessary to make sure
sion among its users (O’Keeffe, Clarke-Pearson, and others the data is not used against the users interest (McKee 2013).
2011). While we can see that social media can uncover sev- Few users realize the amount of health-related information
eral behavioral and psychological insights into the stressful that can be inferred from their digital traces; consequently,
lifestyles of people, further research is warranted to precisely transparency about the indicators derived and the purposes
determine the causality pathways. behind such inference should be part of ethical and policy
While one way of representing counties is by their geo- discourse.
graphic distribution (as in the case of US counties), another
way could be by their residents’ occupation (‘work’ coun- Conclusion
ties), and students (‘college’ counties). Monitoring stress
levels in such counties would especially be useful in devising Using a sample of users who have accounts on both Twit-
programs or initiatives to encourage a low-stress lifestyle ter and Facebook, we uncovered insight into the language
amongst wherever individuals are particularly vulnerable to of stress in individuals. Our results also showed that stan-
stress. While social media (and Twitter in particular) are not dardized linguistic dictionaries, such as LIWC, outperform
representative of the real-life users, several insights have engagement attributes such as user posting behavior, in pre-
been obtained using public Twitter data in public health dicting stress. Our results complement psychological survey
(Yang et al. 2018). Further, the underlying predictive model data by deepening our understanding of the environment
can be applied to several online forums. For example, it that may contribute to both individual and county-level well-
could be applied to colleges across the U.S. to understand being. Language estimates can be used to monitor stress in
the specific chronic stressors in different college campuses different social settings, to devise initiatives encouraging a
(extending the recent work (Saha and De Choudhury 2017; low-stress lifestyle. Our model will help measure the im-
Bagroy, Kumaraguru, and De Choudhury 2017)). It is espe- pact of such interventions in real time. Further, language
cially promising that such measurements can be effective has the specific advantage of getting at the causes/stressors
while being unobtrusive. Monitoring stress using social me- that counties are experiencing thereby enabling personalized
dia also offers the advantage of discovering stress associa- interventions.
tions that are otherwise not easy to get by traditional surveys The findings shed light on the nuanced relationship of so-
(Holdeman 2009). Moreover, such a tool could monitor stress cioeconomic status with psychological stress and suggest
in real time, potentially detecting unforeseen events before insights for other psychological factors, such as locus of con-
they happen (Reason 1995). trol, the degree to which people feel like being manipulated
by their environment vs. perceive a sufficient level of self-
On the technical front, while our application of domain
control and in the counties where they live and work. We
adaptation bridged the gap between two social media plat-
believe that the potential for this technology to be used in
forms – Facebook and Twitter, we note that further work
both micro and macro scales is tremendous. Our computa-
is warranted to address the challenges associated with spa-
tional techniques, while useful to monitor stress in counties,
tially correlated terms introduced when aggregating tweets
can also be used to give real-time feedback to individuals
to counties. The study by (Rieman et al. 2017) has explored
in those counties. Such feedback about individuals’ social
such problems and their implications for county-level pre-
media usage along specific dimensions, can help in design-
dictions from language. Also, several works that focus on
ing simple but effective technology-assisted interventions in
socio-geographical studies have used advanced autoregres-
cultivating mindfulness and stress control (Miller, Fletcher,
sive modeling techniques which could help in developing
and Kabat-Zinn 1995).
more robust models, even for counties with missing data
based on their spatial neighborhood association with other
counties that have data points. Further, it would be interesting References
to examine stress from multiple modalities (text, images and [Abouserie 1994] Abouserie, R. 1994. Sources and levels
sensor data), which can potentially complementary insights of stress in relation to locus of control and self esteem in
 university students. Educational psychology 14(3):323–330. unifying moral judgment. Psychological Inquiry 23(2):206–
[Bagroy, Kumaraguru, and De Choudhury 2017] Bagroy, S.; 215.
Kumaraguru, P.; and De Choudhury, M. 2017. A social media [Guntuku et al. 2017a] Guntuku, S. C.; Lin, W.; Carpenter,
based index of mental well-being in college campuses. In J.; Ng, W. K.; Ungar, L. H.; and Preoţiuc-Pietro, D. 2017a.
Proceedings of the 2017 CHI Conference on Human Factors Studying personality through the content of posted and liked
in Computing Systems, 1634–1646. ACM. images on twitter. In Proceedings of the 2017 ACM on web
[Baum, Garofalo, and Yali 1999] Baum, A.; Garofalo, J.; and science conference, 223–227. ACM.
Yali, A. 1999. Socioeconomic status and chronic stress: does [Guntuku et al. 2017b] Guntuku, S. C.; Ramsay, J. R.; Mer-
stress account for ses effects on health? Annals of the New chant, R. M.; and Ungar, L. H. 2017b. Language of adhd
York Academy of Sciences 896(1):131–144. in adults on social media. Journal of attention disorders
[Cohen, Kessler, and Gordon 1997] Cohen, S.; Kessler, 1087054717738083.
R. C.; and Gordon, L. U. 1997. Measuring stress: A guide [Guntuku et al. 2017c] Guntuku, S. C.; Yaden, D. B.; Kern,
for health and social scientists. Oxford University Press on M. L.; Ungar, L. H.; and Eichstaedt, J. C. 2017c. Detecting
Demand. depression and mental illness on social media: an integrative
[Cohen, Miller, and Rabin 2001] Cohen, S.; Miller, G. E.; review. Current Opinion in Behavioral Sciences 18:43–49.
and Rabin, B. S. 2001. Psychological stress and antibody [Guntuku et al. 2019] Guntuku, S. C.; Preoţiuc-Pietro, D.;
response to immunization: a critical review of the human Eichstaedt, J. C.; and Ungar, L. 2019. What twitter pro-
literature. Psychosomatic medicine. file and posted images reveal about depression and anxiety.
[Corrigan et al. 2001] Corrigan, P. W.; Green, A.; Lundin, R.; In Thirteenth International AAAI Conference on Web and
Kubiak, M. A.; and Penn, D. L. 2001. Familiarity with and Social Media.
social distance from people who have serious mental illness. [Guntuku, Roy, and Weisi 2015] Guntuku, S. C.; Roy, S.; and
Psychiatric services 52(7):953–958. Weisi, L. 2015. Evaluating visual and textual features for
[Culotta 2014] Culotta, A. 2014. Estimating county health predicting user likes. In 2015 IEEE International Conference
statistics with twitter. In CHI. ACM. on Multimedia and Expo (ICME), 1–6. IEEE.
[Daumé III, Kumar, and Saha 2010] Daumé III, H.; Kumar, [Holdeman 2009] Holdeman, T. C. 2009. Invisible wounds
A.; and Saha, A. 2010. Frustratingly easy semi-supervised of war: Psychological and cognitive injuries, their conse-
domain adaptation. In Proceedings of the 2010 Workshop on quences, and services to assist recovery. Psychiatric Services
Domain Adaptation for Natural Language Processing, 53–59. 60(2):273–273.
Association for Computational Linguistics. [Huppert and So 2010] Huppert, F. A., and So, T. T. 2010.
[De Choudhury et al. 2013] De Choudhury, M.; Gamon, M.; Flourishing across europe: Application of a new conceptual
Counts, S.; and Horvitz, E. 2013. Predicting depression via framework for defining well-being. Gallup.
social media. In ICWSM, 2. [Jaidka et al. 2018] Jaidka, K.; Guntuku, S. C.; Buffone, A.;
[De Neve and Ward 2017] De Neve, J., and Ward, G. 2017. Schwartz, H. A.; and Ungar, L. H. 2018. Facebook vs.
Does work make you happy? evidence from the world happi- twitter: Cross-platform differences in self-disclosure and trait
ness report. Harvard Business Review. prediction. In ICWSM.
[Eichstaedt et al. 2015] Eichstaedt, J. C.; Schwartz, H. A.; [Karlamangla, Singer, and Seeman 2006] Karlamangla,
Kern, M. L.; Park, G.; Labarthe, D. R.; Merchant, R. M.; A. S.; Singer, B. H.; and Seeman, T. E. 2006. Reduction
Jha, S.; Agrawal, M.; Dziurzynski, L. A.; Sap, M.; Weeg, in allostatic load in older adults is associated with lower
C.; Larson, E. E.; Ungar, L. H.; and Seligman, M. E. 2015. all-cause mortality risk: Macarthur studies of successful
Psychological Language on Twitter Predicts County-level aging. Psychosomatic medicine 68(3):500–507.
Heart Disease Mortality. Psychological Science 26:159–169. [Lederbogen et al. 2011] Lederbogen, F.; Kirsch, P.; Haddad,
[Ernala et al. 2017] Ernala, S. K.; RIZVI, A. F.; BIRNBAUM, L.; Streit, F.; Tost, H.; Schuch, P.; Wüst, S.; Pruessner, J. C.;
M. L.; KANE, J. M.; and DE CHOUDHURY, M. 2017. Rietschel, M.; Deuschle, M.; et al. 2011. City living and
Linguistic markers indicating therapeutic outcomes of social urban upbringing affect neural social stress processing in
media disclosures of schizophrenia. Proc ACM Hum-Comput humans. Nature 474(7352):498.
Interact 1(1):43. [Lin et al. 2014] Lin, H.; Jia, J.; Guo, Q.; Xue, Y.; Li, Q.;
[Felson, Scheff, and Retzinger 1993] Felson, R. J.; Scheff, Huang, J.; Cai, L.; and Feng, L. 2014. User-level psycho-
T. J.; and Retzinger, S. M. 1993. Review essay: Shame, logical stress detection from social media using deep neural
anger, and aggression. network. In ACM MM.
[Gil et al. 2004] Gil, K. M.; Carson, J. W.; Porter, L. S.; Sci- [Linn et al. 1985] Linn, L. S.; Yager, J.; Cope, D.; and Leake,
pio, C.; Bediako, S. M.; and Orringer, E. 2004. Daily mood B. 1985. Health status, job satisfaction, job stress, and
and stress predict pain, health care use, and work activity life satisfaction among academic and clinical faculty. Jama
in african american adults with sickle-cell disease. Health 254(19):2775–2782.
Psychology 23(3):267. [Liu, Weitzman, and Chunara 2017] Liu, J.; Weitzman, E. R.;
[Gray, Waytz, and Young 2012] Gray, K.; Waytz, A.; and and Chunara, R. 2017. Assessing behavioral stages from so-
Young, L. 2012. The moral dyad: A fundamental template cial media data. In Proceedings of the 20th ACM Conference
 Companion on Computer Supported Cooperative work & [Quercia et al. 2012] Quercia, D.; Ellis, J.; Capra, L.; and
social computing (CSCW), volume 2017, 1320. NIH Public Crowcroft, J. 2012. Tracking gross community happiness
Access. from tweets. In Proceedings of the 15th ACM Conference
[McEwen and Stellar 1993] McEwen, B. S., and Stellar, E. Companion on Computer Supported Cooperative work &
1993. Stress and the individual: mechanisms leading to dis- social computing (CSCW). ACM.
ease. Archives of internal medicine 153(18):2093–2101. [Reason 1995] Reason, J. 1995. Understanding adverse
[McKee 2013] McKee, R. 2013. Ethical issues in using social events: human factors. BMJ Quality & Safety 4(2):80–89.
media for health and health care research. Health Policy [Rieman et al. 2017] Rieman, D.; Jaidka, K.; Schwartz, H. A.;
110(2):298–301. and Ungar, L. 2017. Domain adaptation from user-level face-
[McManus, Winder, and Gordon 2002] McManus, I.; book models to county-level twitter predictions. In Proceed-
Winder, B.; and Gordon, D. 2002. The causal links between ings of the Eighth International Joint Conference on Natural
stress and burnout in a longitudinal study of uk doctors. The Language Processing (Volume 1: Long Papers), volume 1,
Lancet 359(9323):2089–2090. 764–773.
[Meyer, Finn, and others 2001] Meyer, G. J.; Finn, S. E.; et al. [Rouhizadeh et al. 2016] Rouhizadeh, M.; Ungar, L.; Buf-
2001. Psychological Testing and Psychological Assessment: fone, A.; and Schwartz, H. A. 2016. Using syntactic and
A Review of Evidence and Issues. American Psychologist. semantic context to explore psychodemographic differences
[Miller, Fletcher, and Kabat-Zinn 1995] Miller, J. J.; in self-reference. In Proceedings of the 2016 Conference on
Fletcher, K.; and Kabat-Zinn, J. 1995. Three-year follow-up Empirical Methods in Natural Language Processing, 2054–
and clinical implications of a mindfulness meditation-based 2059.
stress reduction intervention in the treatment of anxiety [Saha and De Choudhury 2017] Saha, K., and De Choudhury,
disorders. General hospital psychiatry 17(3):192–200. M. 2017. Modeling stress with social media around incidents
[Moberly and Watkins 2008] Moberly, N. J., and Watkins, of gun violence on college campuses. Proceedings of the
E. R. 2008. Ruminative self-focus, negative life events, and ACM on Human-Computer Interaction (CSCW).
negative affect. Behaviour research and therapy 46(9):1034– [Salmon 2001] Salmon, P. 2001. Effects of physical exercise
1039. on anxiety, depression, and sensitivity to stress: a unifying
[O’Keeffe, Clarke-Pearson, and others 2011] O’Keeffe, theory. Clinical psychology review 21(1):33–61.
G. S.; Clarke-Pearson, K.; et al. 2011. The impact of social [Samani et al. 2018] Samani, Z. R.; Guntuku, S. C.; Moghad-
media on children, adolescents, and families. Pediatrics dam, M. E.; Preoţiuc-Pietro, D.; and Ungar, L. H. 2018.
127(4):800–804. Cross-platform and cross-interaction study of user person-
[Pan and Yang 2010] Pan, S. J., and Yang, Q. 2010. A survey ality based on images on twitter and flickr. PloS one
on transfer learning. IEEE Transactions on knowledge and 13(7):e0198660.
data engineering 22(10):1345–1359. [Schwartz et al. 2013a] Schwartz, H. A.; Eichstaedt, J. C.;
[Pan et al. 2011] Pan, S. J.; Tsang, I. W.; Kwok, J. T.; and Kern, M. L.; Dziurzynski, L.; Lucas, R. E.; Agrawal, M.;
Yang, Q. 2011. Domain adaptation via transfer component Park, G. J.; Lakshmikanth, S. K.; Jha, S.; Seligman, M. E. P.;
analysis. IEEE TNN. and Ungar, L. H. 2013a. Characterizing Geographic Variation
[Pennebaker and Lay 2002] Pennebaker, J. W., and Lay, T. C. in Well-being using Tweets. ICWSM.
2002. Language use and personality during crises: Analyses [Schwartz et al. 2013b] Schwartz, H. A.; Eichstaedt, J. C.;
of mayor rudolph giuliani’s press conferences. Journal of Kern, M. L.; Dziurzynski, L.; Ramones, S. M.; Agrawal,
Research in Personality. M.; Shah, A.; Kosinski, M.; Stillwell, D.; and Seligman, M.
[Pennebaker, Booth, and Francis 2007] Pennebaker, J. W.; 2013b. Personality, Gender, and Age in the Language of
Booth, R. J.; and Francis, M. E. 2007. Linguistic inquiry and Social Media: The Open-vocabulary Approach. PloS ONE.
word count: Liwc [computer software]. Austin, TX: liwc. net. [Schwartz et al. 2017] Schwartz, H. A.; Giorgi, S.; Sap, M.;
[Preotiuc-Pietro et al. 2012] Preotiuc-Pietro, D.; Samangooei, Crutchley, P.; Ungar, L.; and Eichstaedt, J. 2017. Dlatk:
S.; Cohn, T.; Gibbins, N.; and Niranjan, M. 2012. Trend- Differential language analysis toolkit. In Proceedings of the
miner: An architecture for real time analysis of social media 2017 Conference on Empirical Methods in Natural Language
text. In Sixth International AAAI Conference on Weblogs and Processing: System Demonstrations, 55–60.
Social Media. [Sharma and Gedeon 2012] Sharma, N., and Gedeon, T.
[Preotiuc-Pietro et al. 2015] Preotiuc-Pietro, D.; Eichstaedt, 2012. Objective measures, sensors and computational
J.; Park, G.; Sap, M.; Smith, L.; Tobolsky, V.; Schwartz, techniques for stress recognition and classification: A sur-
H. A.; and Ungar, L. 2015. The role of personality, age and vey. Computer methods and programs in biomedicine
gender in tweeting about mental illnesses. In NAACL HLT, 108(3):1287–1301.
volume 2015, 21. [Singh and Agarwal 2016] Singh, V. K., and Agarwal, R. R.
[Preotiuc-Pietro et al. 2016] Preotiuc-Pietro, D.; Carpenter, 2016. Cooperative phoneotypes: exploring phone-based be-
J.; Giorgi, S.; and Ungar, L. 2016. Studying the dark triad havioral markers of cooperation. In Proceedings of the 2016
of personality through twitter behavior. In CIKM, 761–770. ACM International Joint Conference on Pervasive and Ubiq-
ACM. uitous Computing, 646–657. ACM.
[Staal 2004] Staal, M. A. 2004. Stress, cognition, and human On Facebook, words and phrases such as “me”, “i had”,
performance: A literature review and conceptual framework. “feel like”, “i don’t” and “i hate” are significantly correlated
[Steptoe et al. 1996] Steptoe, A.; Wardle, J.; Pollard, T. M.; with high stress, while words and phrases such as “lunch”,“a
Canaan, L.; and Davies, G. J. 1996. Stress, social support great year” and “to see” are significantly correlated with
and health-related behavior: a study of smoking, alcohol low stress. On the other hand, on Twitter, words and phrases
consumption and physical exercise. Journal of psychosomatic which indicate low stress are “excited to”, “! check”, “fresh”,
research 41(2):171–180. “win” and “experience”.
[Stults-Kolehmainen and Sinha 2014] Stults-Kolehmainen, The lack of words in the bottom right quadrant suggests
M. A., and Sinha, R. 2014. The effects of stress on physical that that language of high stress on Twitter is a subset of the
activity and exercise. Sports medicine 44(1):81–121. language of high stress on Facebook. These findings motivate
our proposal to adapt stress models trained on Facebook
[Thelwall 2017] Thelwall, M. 2017. Tensistrength: Stress language, a more inclusive corpus, to Twitter language, for
and relaxation magnitude detection for social media texts. the purpose of county-level stress prediction.
IPM.
[Watson and Pennebaker 1989] Watson, D., and Pennebaker,
LIWC: From Table 8, we observe that Twitter has some
J. W. 1989. Health complaints, stress, and distress: exploring
similarities with Facebook - such as the use of adverbs, and
the central role of negative affectivity. Psychological review
the negative correlation with the use of “Positive Emotion”
96(2):234.
words. Unexpectedly, on Twitter, the category most strongly
[Wilkinson 2005] Wilkinson, R. G. 2005. The impact of correlated with stress is “Comparisons” and not first person
inequality: How to make sick societies healthier. The New singular pronouns. We anticipate that this reflects the ten-
Press. dency of Twitter users to drop self-referencing pronouns, as
[Yang et al. 2018] Yang, Q.; Tufts, C.; Ungar, L.; Guntuku, has been identified in previous work (Rouhizadeh et al. 2016).
S.; and Merchant, R. 2018. To retweet or not to retweet: Un- Comparisons and differentiation could indicate lack of sat-
derstanding what features of cardiovascular tweets influence isfaction with status quo, indicating possibly pessimism or
their retransmission. Journal of health communication 1–10. expressing perceived resources to fall short of demands, or
[Zhong et al. 2017] Zhong, C.; Chan, H.-w.; Karamshu, D.; simply lower perceived satisfaction with life of those prone
Lee, D.; and Sastry, N. 2017. Wearing many (social) hats: to stress (Linn et al. 1985). Insecurities of those higher in
How different are your different social network personae? stress may be what is indicated by correlations with nonflu-
ICWSM. encies (Felson, Scheff, and Retzinger 1993). A “focus on the
past” tends to be maladaptive and may indicate that those
Appendix high in stress might find themselves stuck in what happened
Here, we describe the experiments detailing the language dif- and in ruminative, self-deprecating thoughts. This is rein-
ferences between Facebook and Twitter and their associations forced by the negatively correlated category “Power”, which
with stress. suggests that the individuals under stress express a lack of
control and superiority over their environment (Abouserie
1994).
Usage Statistics: In terms of pure usage statistics, we find
that the our 601 participants had an average of 1238 Facebook
and 1649 Twitter posts. The mean number of words they Topics: Twitter topics which are positively correlated with
wrote per post was 17.1 for Twitter and 23.8 for Facebook. stress (Table 9) comprise of words and emojis depicting nega-
This difference is expected given the 140-character limit tive emotions, possible sarcasm and feelings of awkwardness.
imposed on Twitter posts. However, the median number of We did not find any topics which are negatively correlated
words per post – 17 for Twitter and 18 for Facebook – is much with stress, on Twitter when controlled for age and gender.
closer, suggesting that our participants were more varied in
their posting behavior, when it came to Facebook. TensiStrength: The Pearson correlation between the mean
stress values for Facebook and Twitter, was 0.28 and signif-
N Grams: Next, we consider platform differences in ex- icant at p < 0.01. This difference is expected because the
pressing stress, at the word- and phrase-level. Figure 4 pro- TensiStrength lexicon was developed on Twitter posts, and as
vides a 2D visualization of the Spearman correlations of indi- such it might not generalize well to Facebook language. The
vidual words with stress on the X-axis, and with social media finding does suggest that stress is expressed very differently
platform on the Y-axis. The words along the X-axis are highly on Twitter versus Facebook, and motivates our decision to
correlated with stress, but less preferential towards individual adapt Facebook stress models to Twitter.
social media platforms. The words at the Y-axis poles are
highly correlated with platform usage but less predictive of
stress. For the purpose of this plot, we have discarded words
with non-significant correlations as well as ρ <= 0.05. Thus,
the words along the diagonals are the ones which best distin-
guish Twitter usage from Facebook usage for the purpose of
expressing stress.
Figure 4: Correlations of words/phrases with Social Media platform (Facebook on the positive y-axis and Twitter on the negative
y-axis) and Stress Scale. Correlations are significant, Bonferroni corrected at p < 0.01

LIWC - Twitter Topics – Twitter

Positively Positively most frequent words used by high-stress
most frequent words r r
correlated correlated people
Comparisons greater, best, after .164 hate, stupid, −− , > . <, gah, ugh, frig-
Hatred gin, effin, sigh, grr, >:,wtf, annoying, urgh, .165
Adverbs very, really .144 dammit
Nonfluencies er, hm, umm .136 funny, it’s, thinks, people, isn’t, finds, hilari-
Differentiation hasn’t, but, else .134 Sarcasm .159
ous, ironic, entertaining, nicer, unfortunate
Focused on Past ago, did, talked .134 Feeling feels, weird, kinda, weird, bit, strange, odd,
.150
Negatively awkward hmm, suddenly, sort, awkward, dunno
most frequent words r
correlated Table 9: Pearson correlations between the stress score and
Power superior -.159 Topics extracted from Twitter posts, when controlling for age
Positive Emotion love, nice, sweet -.107 and gender. Topic labels are manually created. All features
are significant at p < .01, two tailed t-test. No negative
Table 8: Pearson correlations (maximum 5) between the stress correlations were found for Twitter posts
score and LIWC features extracted on Twitter, when control-
ling for age and gender. All features are significant at p < .01,
two tailed t-test.