Page 60 - GTM-4-2

Global Translational Medicine Sleep and emotion rhythmicity in tweets

of “Tweets” with positive emotionality declining and ranging from –1 (negative) to 0 (neutral), to +1 (positive).
negative emotion increasing with progression of the day. However, this approach conflates positive and negative
Despite these early advancements, further investigation emotion into a single dimension, assuming their
into how social media use can be used to study sleep-wake conditional dependence. This assumption departs from
behaviors has been remarkably scant, in part due to a lack the substantial body of literature, indicating that positive
of established methods to detect and estimate sleep using and negative emotion are independent dimensions that
social media activity. are differentially affected by sleep loss and circadian
3
trajectories. Therefore, in accordance with the previous
In an elegant effort, Roenneberg used tweets to estimate 1
2
sleep opportunity by analyzing the circadian rhythmicity study which explored circadian aspects of emotional tweet
content, we also derived measures of positive and negative
of a single social media account (~6.5 h in the case of affect for each tweet using a natural language processing
Donald Trump). However, links between these digital rest- (NLP) model (Linguistic Enquiry and Word Count [LIWC]
activity rhythms (dRARs) and clinically or socially relevant analysis ), using the “positive emotion” and “negative
4
indices, such as the emotional content of tweets, were not emotion” output variables which give an overall positivity
examined, potentially due to the complexities involved and negativity score for each tweet based on a validated
in integrating these dRARs with the highly complex linguistic dictionary of 6400 words. We subsequently
yet rich nature of tweet contents. Extending this work, calculated mean positive and negative affect across hourly
we developed a methodological process to examine the and monthly bins to assess circadian and seasonal effects.
relationship between a user’s sleep-wake patterns and the Models were conducted with Zero-Inflated Poisson (ZIP)
daytime emotional content of their tweets. As these tools regression controlling for time and month of each tweet to
are ultimately intended for use at the level of the individual limit confounding. In separate sensitivity analyses, we also
user, here we present proof-of-concept data from a single tested the relationship of time and month with positive
user (Kanye West “@Ye”) to identify periods of sleep and negative emotion using ZIP regression to test main
disturbance, based on a user’s habitual sleep window and and interaction effects of hour, month, and affect (positive
evaluated their impact on next-day emotional state. vs. negative), both to serve as a validation check for

1.1. An individual-level proof of concept: their pre-specified inclusion as covariates, and to test the
Introducing dRARs analysis sensitivity of the model to detect known effects observed at
the population level in a single user. Finally, we visualized
As demonstrated by Roenneberg, case studies of high- data using a double plot, adapted from a common method
2
volume Twitter (“X.com”) users may represent such a used in circadian actigraphy research, along with semantic
means through which meaningful longitudinal data can networks for the emotional constructs identified by the
be validated. In this analysis, 1868 tweets were extracted LIWC analysis (Figure 1).
from the user profile of Kanye West (available at www. 1
twitter.com/@Ye) across 258 days using the Twitter. Our model confirmed findings at the population level
com application programming interface and a custom that tweets fluctuate significantly across the 24-h cycle
programming script written in Python. To approximate with positive emotion declining with the progression of
nocturnal from diurnal tweets, data were mean-averaged the day (P < 0.001), whilst negative emotion increased
across the period, and a sliding window function applied and peaked towards the evening (P < 0.001) (Figure 1).
Next, we tested the effect of nocturnal tweeting on next-
to determine the least active 5 h (L5 index) of the 24-h
period, a common method used to determine the nadir of day emotional tweet content whilst controlling for time-
circadian quiescence and define the “nocturnal” window. of-day of the daytime tweets, month, and the tweet-word
We subsequently identified days in which @Ye’s tweets were count, to mitigate their potential confounding impact
(Table S1). Our models demonstrated that increased
preceded by Twitter activity during the previous nocturnal nocturnal tweeting was significantly associated with
window and derived the number of tweets during that negative emotional content of tweets the following day
window, as well as the total number of nocturnal tweets (Bonferroni-corrected P ≤ 0.001) and was also linked to
over the previous 7 days (Supplement File for detailed significantly reduced positive emotion the following day
methods). (Bonferroni-corrected P ≤ 0.001). When examining tweets
The most frequently adopted approach to study Twitter across the previous week, increased nocturnal tweeting
and other social media content is sentiment analysis. was also significantly associated with decreased positive
Sentiment analysis offers a low-dimensional representation emotional content (Bonferroni-corrected P = 0.041), but
of emotion by integrating several complex word valences not with negative emotional content (Bonferroni-corrected
into one measure of relative positive/negative emotion, P = 0.432).

Volume 4 Issue 2 (2025) 52 doi: 10.36922/gtm.5176