Communications of the ACM

tal of 10 tweets or fewer to assemble a sufficient base for understanding how they experienced WFC. The final sample, as of February 2010, included 3,327,715 tweets from 93,700 users, with 35 tweets per user on average over six months in 2009 and 2010.

To account for time-zone differences among Twitter users, we converted the timestamps on all tweets to users’ local time. Our identification of topics in tweets employed Linguistic Inquiry and Word Count, a tool widely used for text analysis, 19 providing typical word dictionaries that measure psychological (such as social) processes, and personal concerns. Personal concerns consist of sub-categories (such as work and home), and social processes consist of sub-categories (such as family). We classified tweets with words in the work category as job-related and tweets with words in the family and home sub-categories as family-related.

We used two sets of variables: WFC and satisfaction. We measured all variables first at the individual level by date and then aggregated them over all active users for each date. To account for individual differences, we measured the baselines of all individuals and standardized all measures by removing baseline differences across individuals.

Following previous conceptualizations, 4, 12 we measured WFC along two dimensions: time and strain. Time-based WFC (TC) is a consequence of competition for an individual’s time from work and family responsibilities. One classical example is “The time I spend with my family (work) often causes me not to spend time in activities at work (family).” 4 Strain-based WFC (SC) arises when role stressors at work (or with family) induce strain in the individual, hampering fulfillment of role expectations in the family (or work) domain. Two classical examples are “I am often so emotionally drained when I get home from work that it prevents me from contributing to my family” and “Because I am often stressed from family responsibilities, I have a hard time concentrating on my work.” 4

TC. We used the proportion of
tweets on work- and family-related top-
ics posted by a user on a given day as
a proxy for the time the user reported
of WFC, particularly time-based and
strain-based WFC, and the relation-
ships between WFC and job and family
satisfaction through the lens of tweets
for the first time. Twitter provides abun-
dant data where user opinions on cer-
tain topics or events can be mined and
is expected to present a precise picture
of dynamics and influence of WFC and
related user experience and percep-
tion. Tweets have been used to exam-
ine the changing patterns of diurnal
and seasonal mood with work, sleep,
and length of day, 11 but they have yet to
be explored to help understand the dy-
namics of WFC. Our research thus takes
a significant step toward expanding re-
search methods for examining WFC.

Social-Media Analytics Method

We used a dataset of Twitter users and
their tweets collected through a com-
bination of random sampling and
social sub-graph extraction that was
representative of the actual population
of the U.S. 8 To study WFC, we filtered
those users who did not have a job
based on whether none of their tweets
involved work-related topics. We also
filtered those users who posted a to-
and work situations fluctuate depend-
ing on one’s circumstances, and a
person may reply differently at differ-
ent times. 20 However, the dynamics of
WFC were not addressed in previous
studies because they generally adopt-
ed one-time, cross-sectional measure
with few exceptions. For instance,
one study published in 2013 collected
survey data at two different points in
time—2004 and 2006—to investigate
the relationship between WFC and pay
satisfaction. 1 Although a diary method
was used to examine WFC and work-
family facilitation, 3 finding consider-
able variation in the same individuals,
the traditional diary method is dif-
ficult to scale up in terms of number
of participants. Further, despite that
previous studies established negative
correlations between WFC and job
and family satisfaction, the strength
of relationships varied greatly from
one study to another, ranging from
nearly negligible to strong. 14, 17 This
variability raises the need to explain
the inconsistent findings.