One way to address this challenge would be for platforms to
anonymize and release data to researchers. Although some
companies are exploring ways to share data in a privacy-
preserving way (e.g., [19]), data sharing is challenging for
multiple reasons. Companies are limited by privacy policies
and international laws, and sharing disaggregated user data
without the appropriate notice or consent is problematic eth-
ically (in light of privacy concerns) and logistically (e.g., if
a person deletes a post after that dataset is shared with re-
searchers, it is technically challenging to ensure it is deleted
everywhere). Finally, as was shown through the release of
a crawled dataset, it is very difficult—if not impossible—
to fully anonymize networked social media data [79]. Given
the above, it is important that alternative, validated measures
be made available to researchers who do not have access to
server-level data.
Therefore, this paper presents an evaluation of self-report
measures for time spent on Facebook and recommendations
to researchers. As one of the largest social media platforms,
Facebook is the focus of many empirical studies, most of
which employ some measure of site use. We conducted an
analysis comparing server-logged time-spent metrics to self-
reported time-spent survey questions popular in the field. In
doing so, we note that only measuring time spent on plat-
form may offer limited insight into important outcomes such
as well-being, because how people spend their time is often
more important [11, 8]. However, time on platform is an im-
portant variable in numerous studies [18, 31, 35]. Thus, in or-
der to facilitate meta-analyses and support continuity across
past and future scholarship, this study makes the following
contributions: 1) statistical evaluation of self-reported time
spent measures over a large, international sample, 2) assess-
ment of multiple question wordings, and 3) guidance for re-
searchers who wish to use time-spent self reports.
Four problems motivate this work. First, a wide set of so-
cial media usage questions appear in the published literature.
While there have been investigations of the quality of specific
questions [9, 35], no work to date has provided a compre-
hensive analysis of items evaluated against server-level data.
Second, scholars and policymakers care about outcomes of
social media use including well-being [11, 29], social capi-
tal [17, 18, 80, 12], and academic performance [36, 38, 39].
Accurate assessments of social media use in these domains is
critical because of their importance to people’s lives. Third, as
mentioned above, many scholars do not have access to other
sources of data that could contextualize self-report data, such
as server logs or monitoring software. Measurement valid-
ity remains an important consideration for comparative work
within the scientific community. Finally, comparative interna-
tional understanding of social media use is difficult [45] and
rarely conducted, particularly beyond comparisons of or be-
tween Western countries (cf. [23, 50, 65]). International com-
parative work can be particularly fraught due to measurement
error [54, 67, 2]. Because social media is one of the largest
growing sources of information access globally [55], it is im-
portant to assess the accuracy of these questions in different
regions and cultures in order to support this research.
RELATED WORK
Reliability of Self-Reported Media-Use Measures
The measurement of media use has relied for decades on
self-reports [48, 60]. However, these self-reports have been
shown to be unreliable across many domains. Historically,
self-report validity is low for exposure to television and news-
papers [3, 4, 44, 78], general media use [16, 62], and news
access [70, 43, 32, 73, 56]. Self-reported internet and social
media use have also been found to be unreliable. Many stud-
ies across general internet use [5, 61], device use [37], spe-
cific platforms [47], recall of specific types of content [72],
and specific actions taken [66] find low reliability, especially
when compared to logged behavioral data.
For Facebook in particular, a few studies demonstrate the mis-
match between logged data and retrospective, self-reported
use. Studying 45 U.S. college students, Junco [35] found that
there was a “strong positive correlation” (r = 0.59) but “a sig-
nificant discrepancy” between self-reports and laptop-based
monitoring software: participants reported spending 5.6x as
much time on Facebook (145 minutes per day) as they ac-
tually did (26 minutes). That study did not track Facebook
use on mobile phones and participants may have used it more
or less than usual because they knew they were being tracked.
Haenschen [25] surveyed 828 American adults and found that
“individuals underestimate[d] their frequency of status post-
ing and overestimate[d] their frequency of sharing news links
on Facebook.” Burke and Kraut [9, 11] found that self-reports
of time on site among 1,910 English speakers worldwide were
moderately correlated with server logs (r = 0.45). This paper
builds on this prior work by assessing multiple popular ques-
tion wordings at once with a large, international sample and
provides recommendations to researchers on the best ways to
collect self-reports of time spent on Facebook.
Sources of Error in Self-Reported Time Spent
Mental Models of Time Spent. One of the greatest sources of
ambiguity in self-reports of time spent online is that partici-
pants have different mental models for the phenomenon that
researchers care about. For time spent on Facebook, Junco
[35] found that students in an informal focus group reported
thinking about Facebook “all the time,” which may have
caused them to inflate their self-reported time. Attitudes to-
wards social media use—such as “my friends use Facebook a
lot” or “using social media is bad for me”—might also cause
people to report greater or lesser use, respectively [53]. Some
people may include time reading email and push notifications
from Facebook while others might only count time they ac-
tively scrolled through posts or typed comments. Some may
include the time spent on messaging, depending on whether
they are on a device that incorporates it as a separate applica-
tion or not. For people who do not use Facebook every day,
some may estimate their average use across the past week by
including only days in which they opened the app; others may
include zeros for days of non-use. Beyond these differences,
it may be cognitively impossible for participants to recall time
across multiple devices or interfaces.
Wording and Context. Specific words and context also influ-
ence responses to time-spent questions. Common words may