Participants
Participants were college students recruited from an introductory psychology class at University of Texas at Austin. Research protocols were approved by the university IRB (Protocol No. 2018-07-0035). The datasets analyzed in this paper were collected in Fall 2020 (exploratory sample: n = 920, observations = 73,284) and Spring 2021 (confirmatory sample: n = 764, observations = 55,903). Prior to data analysis, we followed an initial data procedure to ensure that only high-quality experience sampling occasions were retained in the final sample:
-
1.
ESM surveys completed too quickly. We computed a threshold based on the number of questions completed in each ESM survey (by multiplying this number by 0.5 s). We subsequently filtered any reports that were completed faster than the threshold.
-
2.
Participant-specific ESM surveys completed too close in time to each other (less than 60 min after the previous report).
-
3.
ESM surveys that took too long to complete (more than 60 min).
-
4.
Participants who indicated in the post-survey that they had not been truthful in the ESM surveys.
We then excluded participants who failed to complete more than 65% of the total required experience sampling reports to gain credit for the assignment. As a final step, we removed participants who were older than 24 years of age since our target population of interest was young adults.
Initial data cleaning procedures resulted in the removal of 32 participants corresponding to 5245 observations for the exploratory sample. Subsequently, we removed 18 participants corresponding to 1528 observations who were older than 24 years of age from the exploratory sample. Finally, we removed 49 participants corresponding to 577 observations because they failed to complete more than 65% of the total number of experience sampling questions required to gain credit for the assignment. Hence, the final sample size of the exploratory sample was 821 participants corresponding to 65,934 observations.
Similarly, initial data cleaning procedures led to the removal of 20 participants corresponding to 3343 observations from the confirmatory sample. We additionally removed 10 participants corresponding to 837 observations who were older than 24 years of age from the confirmatory sample. Finally, we removed 53 participants corresponding to 837 observations because they failed to complete more than 65% of the total number of experience sampling questions required to gain credit for the assignment. Hence, the final size of the confirmatory sample was 681 participants corresponding to 51,500 observations.
In the combined sample consisting of both the exploratory and confirmatory sample, most participants identified as female (69.1%) with a mean age of 18.7 years and were enrolled in either their first (58.7%) or second year of college (26.1%). Most participants identified as Anglo/White (32.5%), Asian/Asian American (21.7%), Hispanic/Latino (25.6%) and African Americans (5.1%).
Procedure
Participants completed a demographic survey during the first week of the semester and a range of personality questionnaires during other weeks of the semester. Participants received seven daily ESM surveys for up to four weeks. Participants received full credit if they provided at-least fourteen days of data with at least four surveys on each of those days. Participants downloaded an application onto their smartphones which sent periodic push notifications to participants about pending surveys. The notifications were programmed to arrive at semi-random times within seven 120-min blocks between 8 am and 10 pm, with a minimum time window of 60 min between each consecutive notification. Participants were permitted to complete surveys on their phones or computers, but notifications expired by the end of each block. Hence, the average time window between two consecutive surveys within the same day was 163 min. The surveys were distributed in the following pattern throughout the day: 22% during the morning, 24% during the midday, 27% during midday, 27% during afternoon and 24% during the evening. All participating students were compensated with class credit and personalized feedback reports that summarized their social media use trends and psychological wellbeing patterns over the course of the semester.
Momentary measures
Wellbeing
Wellbeing was measured using seven adjectives: “happy”, “sad”, “valued and accepted by others”, “lonely”, “worried”, “angry” and “stressed”. Participants responded to each scale using a 1–4 Likert scale (ranging from “not at all” to “a great deal”). The question stem asked participants to indicate their feelings “right now”, explicitly capturing momentary wellbeing at the time of the ESM. The adjectives “angry”, “worried”, “happy”, and “sad” were borrowed directly from past work31. Following past research31,57, we computed momentary affect balance by subtracting the “happy” score from the arithmetic mean score of “sad”, “worried” and “angry”. We treated momentary affect balance and stress as indicators of affective wellbeing. Conversely, we treated momentary feelings of being accepted and loneliness as indicators of social wellbeing. We reverse scored stress and loneliness variables such that higher values on different wellbeing outcomes all indicated “positive wellbeing”. Hence, higher values of loneliness, stress, affect balance and feelings of being accepted all indicate positive wellbeing.
Social media use (vs non-use)
During each ESM survey, participants indicated the activities they had engaged in the past hour using a “select all that apply” multiple choice question. The question stem was “During the PAST HOUR, I spent time doing the following activities (check all that apply)”. The response options consisted of 19 different behaviors (summarized in Table S41) of which one was “Using social media”. We created a new categorical dummy variable that indexed all instances of social media use (vs non-use). All instances of social media use were labelled as “1”. All instances of non-social media use were labelled as “0”. All missing values were preserved. We also assessed participants’ engagement in multitasking as the number of activities performed in the past hour, specifically calculated as the number of activities they indicated performing during the past hour.
Duration of use
If participants selected “using social media” as an activity, branch logic displayed a follow-up question asking participants to rate the duration of their social media use in the past hour on a 4-point scale: 1 = 1–15 min, 2 = 16–30 min, 3 = 31–45 min, 4 = 46–60 min.
Context
At each ESM survey, participants indicated, via “select all that apply” multiple choice response, who they were with during the last hour (social context) and what places they had been in during the last hour (physical context). The social context question stem was: “During the PAST HOUR, I spent time with the following people in-person (check all that apply)”. Participants could indicate having spent time with 8 different categories of people (see Table S41). Based on past research, we created a set of 4 dummy variables from the 8 response options: alone, with family ties, with close ties, and with distant ties (see Table S41). Dummy variables were created such that target social context categories were encoded with a 1, and non-target categories were encoded with a 0 (e.g., Alone = 1, With Other People = 0). All missing values were preserved.
Similarly, the physical context question stem was: “During the PAST HOUR, I spent time in the following places (check all that apply)”. People could indicate having spent time in 13 different types of places. Motivated by theoretical frameworks about psychologically salient physical and social context55,58,59, we created a set of 8 dummy variables from these responses: home, social places, natural places, work places, transit, study places, religious places (see Table S41). Dummy variables were created such that target places were encoded with a 1, while non-target places were encoded with a 0 (e.g., Home = 1, Other Places = 0). All missing values were preserved.
Individual differences and dispositional measures
Demographics
Participants’ age and sex was measured in the demographics survey administered prior to the experience sampling component of the study.
Personality traits
Participants’ Big Five Personality Traits were measured before the start of the experience sampling component of the study. We used the BFI-2 instrument, which consists of 60-items answered using a 5-point Likert Scale28,60. The Big Five Traits measure uses the average of 12 items to measure interindividual differences in extraversion, agreeableness, neuroticism, conscientiousness, and openness. Extraversion captures differences in individuals’ tendency to be gregarious, assertive, energetic, and talkative. Agreeableness captures differences in individuals’ tendency to be trustful, altruistic, modest, and warm. Neuroticism capture one’s tendency to be anxious, angry/hostile, depressed, self-conscious, and impulsive. Conscientiousness captures one’s tendency to be competent, orderly, dutiful achievement striving, self-discipline and deliberative. Finally, Openness captures one’s tendency to be imaginative, have an aesthetic proclivity, preference for variety and curiosity.
Dispositional wellbeing
The following wellbeing tendencies were measured before the start of the experience sampling component of the study: depressive symptoms, satisfaction with life, loneliness, and affect balance.
Depressive symptoms were measured using the Center for Epidemiological Studies-Depression scale that asks participants to indicate a variety of depressive symptoms in the preceding week, including loneliness, poor appetite, and restless sleep30. Higher values corresponded with greater depression symptomatology.
Satisfaction with life was measured using the Diener Satisfaction with Life Scale, as the average of responses provided on a 1 (strongly disagree) to 7 (strongly agree) scale to 5 statements that operationalizes a holistic perspective towards their lived and ideal lives32. People’s satisfaction with life scores are calculated by taking an arithmetic mean of the 5 items of the scale. Higher values corresponded with greater satisfaction with life.
Loneliness was measured using the UCLA loneliness scale, that measures participants’ agreement with 9 statements that ask about the frequency with which participants experience moments of social connection or social disconnection61. Participants responded using a 1 (I never feel this way) to 4 (I always feel this way) scale. Upon reverse scoring a subset of the statements, the final score is calculated by computing an arithmetic mean of all response items. Higher values corresponded to greater loneliness.
Dispositional affect balance was measured using a modified form of the SOEP scales (e.g., Angry, Worried, Happy, Sad, Enthusiastic, Relaxed:31). People indicated the extent to which they felt angry, worried, happy, sad, enthusiastic, and relaxed using a 1 (Very rarely)—5 (Very often) scale. People’s dispositional affect balance was computed by subtracting the mean of their negative emotion scores (e.g., angry, worried, sad, relaxed) from their positive emotion scores (e.g., happy, sad, enthusiastic and relaxed). Hence, positive values corresponded to greater positive affect whereas negative values corresponded to greater negative affect.
Modelling strategy
Data analyses for each of the three research questions was done using multilevel models that accommodated the nested nature of the data (repeated measures nested within-persons). Following usual practice within and between-person effects are disentangled through person-mean centering of all time-varying (Level 1) predictor variables and sample-mean centering of all person-level (Level 2) predictor variables62. Much of the past research has disproportionately focused on examining between-person associations between social media and wellbeing. Hence, to build upon past research, we were especially interested in cross-level interactions (e.g., the extent to which between-person differences in psychological dispositions explain within-person relationships between social media and wellbeing) and within-person moderation effects (e.g. comparing people’s feelings of wellbeing after using social media as compared to when they did not use social media).
For main effect analysis, our general analytic approach consisted of specifying separate models wherein: (1) one of our four possible wellbeing outcomes is specified as a dependent variable and (2) one of two possible operationalizations of social media use are included as predictors. As a result, a total of 8 main effects models were computed. Similarly, for moderation analysis, our general analytic approach consisted of specifying separate models wherein: (1) one of four possible wellbeing outcomes (e.g., feelings of being accepted, loneliness, stress and affect balance) is specified as a dependent variable, (2) one of two possible operationalizations of social media use (e.g., use vs non-use; duration of use) are included as predictors and (3) one of nine possible dispositional variables or one of eleven possible context variables are included as moderators. As a result, a total of 168 moderator models were computed. Across both main effects, dispositional moderators and contextual moderators, a total of 172 models were computed.
What is the relationship between social media use and wellbeing in young adults’ daily lives?
We used frequentist linear regression models in lme463 with random intercepts and random slopes allowed to vary across participants to determine the extent to which social media use and wellbeing are related at the within and between-person levels:
$$\begin{aligned} Wellbeing _{ti} & = \beta_{0i} + \beta_{1i} SocialMedia _{ti} + \beta_{2i} Wellbeing _{{\left( {t – 1} \right)i}} + \beta_{3i} DurationSinceLastResponse _{ti} \\ & \quad + \beta_{4i} Wellbeing _{{\left( {t – 1} \right)i}} DurationSinceLastResponse _{ti} + \beta_{5i} NumberofActivities _{ti} \\ & \quad + \beta_{6i} Weekend _{ti} + \beta_{7i} StudyDay _{ti} + \beta_{8i} Sample _{ti} + e_{ti} \\ \end{aligned}$$
where wellbeing at occasion t for person i is modeled as a function of a person-specific intercept coefficient \({\beta }_{0i}\) that indicates the individual’s prototypical level of wellbeing, a set of person-specific coefficients \({\beta }_{1i}\) to \({\beta }_{7i}\) that indicate the within-person associations between the predictor variables and wellbeing, and residual \({e}_{ti}\) that are assumed normally distributed with standard deviation \({\sigma }_{e}.\) The person-specific coefficients are then modeled as a function of between-person differences. Specifically,
$${\beta }_{0i}={\gamma }_{00}+{\gamma }_{01}{\text{ SocialMedia }}_{i}+{\gamma }_{02}{\text{ Age }}_{i}+{\gamma }_{03}{\text{ Sex }}_{i}+{\gamma }_{04}{\text{ NumberOfResponses }}_{i} + {u}_{0i}$$
$${\beta }_{1i}={\gamma }_{10}+{u}_{1i}$$
$${\beta }_{2i}={\gamma }_{20}$$
$${\beta }_{3i}={\gamma }_{30}$$
$${\beta }_{4i}={\gamma }_{40}$$
$${\beta }_{5i}={\gamma }_{50}$$
$${\beta }_{6i}={\gamma }_{60}$$
$${\beta }_{7i}={\gamma }_{70}$$
where the gammas are sample-level parameters that indicate the intercept and effects for the prototypical individuals, and the residuals \({u}_{0i}\) and \({u}_{1i}\) are residual individual differences in intercept and the within-person association between social media use and wellbeing that are assumed multivariate normal with standard deviations \({\sigma }_{u0}\) and \({\sigma }_{u1}\) and correlation \({r}_{{\sigma }_{u0}{\sigma }_{u1}}\). Of specific interest are the \({\gamma }_{10}\) and \({\sigma }_{u1}\) parameters. We define social media sensitivity as referring to the \({\gamma }_{10}\) parameter whereas \({\sigma }_{u1}\) captures the person-level heterogeneity of social media sensitivity.
What is the relationship between dispositional traits and social media sensitivity?
Dispositional moderators were fit using the lme4 package63 in R with restricted maximum likelihood and missing data (< 0.1%) was treated as missing at random. Statistical significance was evaluated at alpha = 0.05. Dispositional moderators were modelling using the following model:
$${Wellbeing}_{ti}={\beta }_{0i}+{\beta }_{1i}S{ocialMedia}_{ti}+{e}_{ti}$$
$${\beta }_{0i}={\gamma }_{00}+{\gamma }_{01}S{ocialMedia}_{i}+ {\gamma }_{02}{Personality}_{i}+ {\gamma }_{03}S{ocialMedia}_{i}{Personality}_{i}+{u}_{0i}$$
$${\beta }_{1i}={\gamma }_{10}+{\gamma }_{11}{Personality}_{i} +{u}_{1i}$$
where \({\gamma }_{03}\) is the between-person interaction and \({\gamma }_{11}\) is the cross-level interaction.
What is the relationship between context of use and social media sensitivity?
Random intercepts and slopes were specified for social media, context, and their resulting interactions, resulting in complex models that did not converge in a frequentist framework. Hence, we used a Bayesian paradigm for model estimation to facilitate model convergence. The move to Bayesian estimation allowed us to examine the extent to which multiple momentary contexts moderate the relationship between momentary social media use and wellbeing. The expanded model is specified by the following equation:
$$\begin{aligned} Wellbeing_{ti} & = \beta_{0i} + \beta_{1} SocialMedia_{ti} + \beta_{2} SocialMedia_{i} \\ & \quad + \beta_{3} AffectiveWellbeing_{{\left( {t – 1} \right)i}} + \beta_{4} DurationSinceLastResponse_{ti} \\ & \quad + \beta_{5} AffectiveWellbeing_{{\left( {t – 1} \right)i}} DurationSinceLastResponse_{ti} \\ & \quad + \beta_{6} NumberofActivities_{ti} + \beta_{7} Weekend_{ti} + \beta_{8} StudyDay_{ti} \\ & \quad + \beta_{9} SocialMedia_{ti} Context_{ti} + \beta_{10} Sample_{ti} + e_{ti} \\ \end{aligned}$$
$${\beta }_{0i}={\gamma }_{00}+{\gamma }_{01}S{ocialMedia}_{i}+ {\gamma }_{02}{Context}_{i}+ {\gamma }_{03}S{ocialMedia}_{i}{Context}_{i}+{\gamma }_{04}{Age}_{i}+{\gamma }_{06}{Sex}_{i}+ {\gamma }_{07}{NumberOfResponses}_{i}+{u}_{0i}$$
$${\beta }_{1i}={\gamma }_{10}+{\gamma }_{11}{Context}_{i} +{u}_{1i}$$
where \({\beta }_{9}\) is the within-person interaction, \({\gamma }_{03}\) is the between-person interaction and \({\gamma }_{11}\) is the cross-level interaction between average social media use and context. Contextual moderators were fit using the brms package64 with 12,000 iterations (half warm-up) since convergence was not possible with lme4. All models were specified with normal priors and achieved convergence with 12,000 iterations. We did not specify hyperpriors to constrain the hyperparameters of the model. We used the Monte Carlo Markov Chain sampler when fitting the Bayesian models. Specifically, we used 4 separate chains in each model and ensured that they each yielded stable estimates by examining the R-hat values (all values did not exceed 1.01)65. We determined which models yielded coefficients that were larger than 0 by examining the corresponding 95% credibility intervals for each estimate.