. American ; Psychological Association © 2019 American Psychological Association ISSN: 0022-3514 Journal of Personality and Social Psychology: Personality Processes and Individual Differences 2020, Vol. 119, No. 6, 1478-1496 http://dx.doi.org/10.1037/pspp0000272 Is Weil-Being Associated With the Quantity and Quality of Social Interactions? Jessie Sun University of California, Davis Kelci Harris University of Victoria Simine Vazire University of California, Davis Social relationships are often touted as critical for well-being. However, the vast majority of studies on social relationships have relied on self-report measures of both social interactions and well-being, which makes it difficult to disentangle true associations from shared method variance. To address this gap, we assessed the quantity and quality of social interactions using both self-report and observer-based measures in everyday life. Participants (N = 256; 3,206 observations) wore the Electronically Activated Recorder (EAR), an unobtrusive audio recorder, and completed experience sampling method self-reports of their momentary social interactions, happiness, and feelings of social connectedness, 4 times each day for 1 week. Observers rated the quantity and quality of participants' social interactions based on the EAR recordings from the same time points. Quantity of social interactions was robustly associated with greater well-being in the moment and on average, whether they were measured with self-reports or observer reports. Conversational (conversational depth and self-disclosure) and relational (knowing and liking one's interaction partners) aspects of social interaction quality were also generally associated with greater well-being, but the effects were larger and more consistent for self-reported (vs. observer-reported) quality variables, within-person (vs. between-person) associations, and for predicting social connectedness (vs. happiness). Finally, although most associations were similar for introverts and extraverts, our exploratory results suggest that introverts may experience greater boosts in social connectedness, relative to extraverts, when engaging in deeper conversations. This study provides compelling multimethod evidence supporting the link between more frequent and deeper social interactions and well-being. Keywords: social interactions, well-being, extraversion, experience sampling, naturalistic observation Supplemental materials: http://dx.doi.org/10.1037/pspp0000272.supp Social relationships are often touted as critical for well-being (Axgyle, 2001; Myers, 2000). Indeed, a clear conclusion from previous research on social interactions and well-being is that people feel happier in moments when they are interacting with others, and that happier people tend to spend more time interacting with others. Across a range of methods, including not only retrospective and momentary self-reports (Kushlev, Heintzelman, Oi-shi, & Diener, 2018; Lucas, Le, & Dyrenforfh, 2008; Rohrer, Richter, Brummer, Wagner, & Schmukle, 2018; Srivastava, An- gelo, & Vallereux, 2008; Watson, Clark, Mclntyre, & Hamaker, 1992), but also mechanical clickers for counting social interactions as they occur (Sandstrom & Dunn, 2014b), and observer ratings based on unobtrusive audio recordings of everyday behavior (Mehl, Vazire, Holleran, & Clark, 2010; Milek et al., 2018), studies consistently show that the amount—or quantity—of social interactions one has is associated with greater well-being. Less is known, however, about how much the quality of social interactions—including what happens during a social interaction This article was published Online First October 24, 2019. ® Jessie Sun, Department of Psychology, University of California, Davis; Kelci Harris, Department of Psychology, University of Victoria; ® Simine Vazire, Department of Psychology, University of California, Davis. Simine Vazire acquired funding. Simine Vazire and Kelci Harris designed and supervised data collection for the larger study. All authors conceptualized the current research goals. Jessie Sun supervised Electronically Activated Recorder coding, analyzed the data (with input from Kelci Harris and Simine Vazire), and drafted the manuscript. All authors provided critical revisions to the manuscript and approved the final manuscript for submission. Data collection for this article was supported by a grant from the National Science Foundation to Simine Vazire (BCS-1125553). A por- tion of these findings were presented at the Association for Psychological Science Annual Convention in New York City, New York, May 21-May 24, 2015. We are grateful to Luke Smillie and Rich Lucas for comments on a draft of this article, and to the many research assistants who ran participants and coded the EAR recordings. The quantitative data, R scripts, and Mplus input and output files required to reproduce the analyses reported in this article are available at https://osf.io/23vpz. Codebooks for all measures in the larger study are available at https:// osf.io/akbfj. Correspondence concerning this article should be addressed to Jessie Sun, Department of Psychology, University of California, Davis, 200 East Quad, Davis, CA 95616. E-mail: jesun@ucdavis.edu 1478 SOCIAL INTERACTIONS AND WELL-BEING 1479 and who it is shared with—matters for well-being. This literature is less conclusive because of its reliance on self-report measures of both the quality of social interactions and well-being. To address this gap, we use a multimefhod approach that harnesses naturalistic observations of social interactions to clarify whether and which qualities of social interactions are related to within-person fluctuations and between-person differences in well-being. Finally, given the empirical and theoretical importance of trait extraversion to both social behavior and well-being, we examine whether the associations between the quantity and quality of social interactions and well-being are similar for introverts and extraverts. Compared to experimental paradigms, our naturalistic method emphasizes ecological validity but prevents us from drawing causal conclusions. Still, grounded in the perspective that "it is necessary to know the thing we are trying to explain" (Asch, 1952/1987, p. 65; see also Rozin, 2001), we believe that establishing the robustness and magnitude of effects in the real world provides an important foundation for future experimental tests that can shed light on causal explanations. Effect sizes based on observations of real-world phenomena can also constrain theories about causal links between social interactions and well-being. Therefore, we focus on thoroughly describing the associations between several aspects of naturally-occurring variations in social interactions and well-being, rather than addressing issues of causality. Is the Quality of Social Experience Related to Weil-Being? Not all social interactions are equal—rather, social interactions are flavored by attributes such as conversational features (e.g., conversational depth and self-disclosure), relational features (e.g., relationship type, closeness, acquaintanceship, liking), and the purpose of the social interaction (e.g., entertainment, work, chores), that make each social interaction unique. Theory and intuition both suggest that the quality of one's social interactions should matter for well-being, over and above the sheer amount of time spent in social interactions. For example, Baumeister and Leary (1995) argued that people have a universal basic need to form and maintain strong, stable interpersonal relationships. Crucially, according to Baumeister and Leary, mere social contact is not enough to fulfill this need to belong; instead, interactions should be not only frequent but also pleasant (or at least free from conflict), and people need to perceive a bond that involves "stability, affective concern, and continuation into the foreseeable future" (p. 500). At first glance, it seems that a plethora of studies show that the quality of social experience is related to well-being. For example, people who report that their relationships are more satisfying and supportive tend to report greater subjective well-being (for a review, see Lyubomirsky, King, & Diener, 2005). However, such associations are based on self-reports of both relationship quality and subjective well-being, which raises concerns about potential artifacts (for methodological critiques, see Lucas & Dyrenforfh, 2006; Lucas, Dyrenforfh, & Diener, 2008). For example, people's positive perceptions of their lives could lead to halo effects in which they think that all domains of their lives (including their social relationships) are going well, regardless of whether things are objectively going well in those domains. Studies that use self-reports to measure both the quality of one's social interactions and well-being cannot easily disentangle true associations from associations due to this type of halo effect or other sources of shared method variance (e.g., response styles). Such studies likely produce inflated estimates of the association between the quality of social interactions and well-being. To date, only a handful of studies have examined whether non-self-report measures or manipulations of social interaction quality are associated with well-being, and the results have been mixed. The strongest evidence for a link between the quality of naturalistic social interactions and well-being comes from studies that use the Electronically Activated Recorder (EAR; Mehl, 2017), a device that unobtrusively records audio snippets of people's everyday behaviors. Human coders subsequently code these recordings for audible behaviors, including the quantity and quality of social interactions. Using this method, Mehl and colleagues (2010) found that happier people tend to have more substantive conversations—an effect that was later replicated (Milek et al., 2018). Although the size of the association between life satisfaction and the percentage of substantive conversations was fairly small (r = .15; Milek et al., 2018), this finding is compelling because there is no method overlap between behavioral observations of conversational depth and self-reported life satisfaction. Another important quality of a social interaction is one's relationship with the person with whom it is shared. Specifically, interactions with close others—although not without their own unique challenges—afford the opportunity for more responsive, accepting, and authentic interactions (at least as subjectively experienced), compared to interactions with distant others (Venaglia & Lemay, 2017). Yet, some studies suggest that even interactions with strangers and weak ties can be quite pleasant. For example, bus and train commuters who were instructed to interact with a stranger reported more positive experiences than those who were instructed to remain in solitude (Epley & Schroeder, 2014). Even having a brief but genuine social interaction with a Starbucks barista appears to have hedonic benefits, compared to completing the transaction as efficiently as possible (Sandstrom & Dunn, 2014a). These studies do not imply that closeness is irrelevant to well-being—only that minimal social interactions with those on the peripheries of our social networks can be surprisingly rewarding. Fewer studies have examined whether interactions with close others are more rewarding than interactions with distant others, and have found mixed results. In the laboratory, participants who were randomly assigned to interact with a stranger felt just as happy as those assigned to interact with their romantic partner (Dunn, Biesanz, Human, & Finn, 2007). Similarly, a study of naturally occurring social interactions did not find systematically larger well-being benefits of interacting with strong ties than weak ties (Sandstrom & Dunn, 2014b, Studies 2a-2b). A recent experience sampling study, however, showed that people felt happier when they interacted with close others, compared to distant others—whether closeness was indexed by relationship type or subjective closeness (Venaglia & Lemay, 2017). Similarly, Mueller and colleagues (2019) found that people tended to feel happiest after interactions with friends, followed by interactions with family members, others, and colleagues. Thus, overall, it is unclear whether people benefit more from interacting with close (vs. distant) others. 1480 SUN, HARRIS, AND VAZIRE In sum, very few studies have used non-self-report methods to examine the association between the quality of social experience and well-being, and existing studies have produced mixed results. The main goal of our study is to provide a strong test of the associations between the quality of social interactions and well-being, by using self- and observer-reports of multiple conversational and relational aspects of social interaction quality. Trait Extraversion as a Potential Moderator A second longstanding question is whether and how the associations between social interactions and well-being differ for those who are more or less extraverted ("extraverts" and "introverts", for short-hand). Extraversion describes the tendency to be talkative, assertive, outgoing, and sociable. Considerable evidence supports the theory that extraversion reflects reward sensitivity—the extent to which people are motivated to pursue rewards, and enjoy those rewards once they are attained (for reviews, see DeYoung, 2015; Smillie, 2013). Because many human rewards are social, extraverts, compared to introverts, should derive greater enjoyment from social interactions (the social reactivity hypothesis; Srivas-tava et al, 2008). Yet, the social reactivity hypothesis has received only mixed support. Early studies found that extraverts and introverts experienced similarly large boosts in momentary positive affect when they spent more time socializing (Lucas, Le, & Dyrenforth, 2008; Srivastava et al., 2008). Similarly, commuters who were instructed to interact with a stranger reported more positive experiences than those instructed to remain in solitude, whether they were more or less extraverted (Epley & Schroeder, 2014). One recent study did find that extraverts had a stronger positive association between social time and average momentary mood than did introverts—but this did not generalize to global positive affect (Kushlev et al., 2018, Study 3b). Similarly, even though extraverts experienced a larger increase in positive affect after a "cocktail party" interaction than did introverts, a large proportion of introverts who had expected to feel worse after socializing actually felt better (Duffy, Helzer, Hoyle, Helzer, & Chartrand, 2018). It is also unclear whether the quality of social interactions matters more for introverts. In her popular book Quiet, Susan Cain (2012) speculated that introverts "prefer to devote their social energies to close friends, colleagues, and family" and "have a horror of small talk, but enjoy deep discussions" (p. 11). Theoretically, however, extraversion involves an affiliative component characterized by the enjoyment of close interpersonal bonds, as well as a more general sensitivity to rewards (Depue & Morrone-Strupinsky, 2005; DeYoung, 2015; Smillie, 2013). These theoretical perspectives suggest that, if anything, extraverts should show a stronger association between deep social interactions and well-being, compared to introverts. Here, the empirical evidence is once again inconclusive. A meta-analysis of four studies suggests that extraversion does not moderate the association between unobtrusively captured depth of conversation and self-reported life satisfaction (Milek et al., 2018). Nor are we aware of evidence that introverts benefit relatively more from interacting with close (vs. distant) others than do extraverts. One study of over 50,000 social interactions found no moderating role of extraversion on the within-person associations between the type of interaction partner and momentary happiness (Mueller et al, 2019). If anything, Sandstrom and Dunn (2014b; Study 2a) found that each additional interaction with a "weak tie" (but not a "strong tie") predicted a greater increase in belonging for introverts, compared to extraverts (but note that this interaction effect did not generalize to subjective well-being). Thus, our study also aims to address the open question of whether the associations between the quantity and quality of social interactions and well-being differ for introverts and extraverts. The Present Study There are still many open questions about how social interaction is related to well-being, and whether trait extraversion moderates any of the associations between aspects of social interactions and well-being. Our key goal is to examine whether self- and observer-based measures of the quantity and quality of social interactions converge on similar conclusions, in order to disambiguate true associations from methodological artifacts. To do so, we use an intensive multimethod approach to capture repeated self- and observer ratings of social interactions and self-reported well-being, and to examine effects at the within- and between-person levels. We measure naturally occurring fluctuations in social interactions and well-being in participants' everyday lives to provide an ecologically valid test of these associations. To assess the quantity of social interactions, we use self- and observer-ratings of the presence or absence of social interactions. Assessing the quality of social interactions—especially using non-self-report methods, and in a way that clearly distinguishes between the quality of the social interaction and how the participant feels about the interaction—is much less straightforward. Here, we measure four variables that capture differences in the quality of social interactions. We use self- and observer-ratings of two conversational features (conversational depth and self-disclosure) that reflect deeper, more intimate interactions. We also use self-reports of two relational features (how much participants knew and liked their interaction partners). We opted not to analyze the observer ratings of these two relational features because we decided (without looking at the results) that these are inherently self-defined variables, and that observer ratings are very unlikely to contain valid variance not captured by self-reports. Of course, these four variables (conversational depth, self-disclosure, knowing, and liking) do not capture all of the ways that social interactions can differ. However, we believe that they are strong candidates for variables that (a) capture variability in the quality of everyday social interactions, (b) could potentially be related to well-being, and (c) can be validly measured repeatedly in everyday life, and, in the case of conversational depth and self-disclosure, with both self- and observer reports. We examine how the quantity and quality of social interactions are associated with two distinct dimensions of well-being: feelings of happiness and of social connectedness (which both feature in several taxonomies of well-being; e.g., Butler & Kern, 2016; Huppert & So, 2013; Ryff, 1989). This allows for a more finegrained understanding of the distinct correlates of different dimensions of well-being (e.g., Baumeister, Vohs, Aaker, & Garbinsky, 2013; Dwyer, Dunn, & Hershfield, 2017; Sun, Kaufman, & Smillie, 2018), while facilitating more general conclusions about the link between social interactions and well-being broadly construed (rather than only the affective component of well-being). Whereas SOCIAL INTERACTIONS AND WELL-BEING 1481 happiness is a more general indicator of well-being that is influenced by many everyday experiences besides social interactions (e.g., Kahneman, Krueger, Schkade, Schwarz, & Stone, 2004), feelings of social connectedness are conceptually more closely linked to the quantity and quality of one's social interactions. Importantly, however, the feeling of social connectedness (as we operationalize it in this study) is also distinct from the quantity and quality of social interactions. Our social connectedness measure captures subjective feelings of connectedness versus loneliness, rather than the presence or absence of a social interaction, or features of the social interaction (e.g., the depth of the interaction, or how well the participant knew the interaction partner). In addition, people can feel more or less socially connected even when they are not interacting with others, whereas the quality of social interaction variables—by definition—only apply when a social interaction occurs. We generally expected to see positive associations between all aspects of social interactions and well-being, but expected the effect sizes to be smaller for observer-based (vs. self-reported) measures of social interactions. Apart from these general expectations, we had no specific predictions about how effect sizes might vary across the analyses. We also had no predictions about the moderating role of extraversion (given the mixed findings reviewed above). Method Ethics and Open Practices Statement We used data from the first wave of the longitudinal Personality and Interpersonal Roles Study (PAIRS). Data collection and coding procedures were approved by Institutional Review Boards at Washington University in St. Louis (IRB ID: 201206090; Study Title: Personality and Intimate Relationships Study) and University of California, Davis (IRB ID: 669518-15; Study Title: Personality and Interpersonal Roles Study). Other published articles have used the PAIRS dataset (for a full list of citations, see https://osf.io/3uag4), including the experience sampling method (ESM) happiness and positive emotion variables (Sun, Schwartz, Son, Kern, & Vazire, 2019; Weidman et al., 2019; Wilson, Thompson, & Vazire, 2017), and the ESM quantity of social interaction, conversational depth, and self-disclosure variables (Wilson, Harris, & Vazire, 2015) used in this dataset. Of these, the most closely related paper (Wilson et al., 2015) examined between-person correlations among friendship satisfaction, the average quantity and quality of self-reported social interactions with friends, and trait extraversion, but did not include any EAR data or analyses examining wifhin-person associations. This is the first article that we know of that examines associations between social interactions and well-being using both self- and observer-reports in everyday life, using any dataset. Note also that this dataset is not the same as the EAR datasets used in the research reviewed above (Mehl et al., 2010; Milek et al, 2018). Below, we describe the measures and procedures relevant to the current article. Several parts of the description of procedures and analytic specifications have been closely adapted from a previous article that used different variables from the same dataset (Sun & Vazire, 2019). Codebooks for all measures in the larger study are available at https://osf.io/akbfj. Although ethical considerations prevent us from making the audio files publicly available, the quantitative data, R scripts, and Mplus input and output files required to reproduce the analyses reported in this paper are available at https://osf.io/23vpz. We did not preregister any of these analyses, as the data were collected years ago and we were familiar with the dataset and had run some analyses before starting this project. Thus, all results are exploratory and any interesting patterns should be interpreted with caution. Participants and Procedure Overview The study involved 434 students at Washington University in St. Louis, who were recruited in 2012 and 2013 via flyers and classroom announcements across the campus. Participants completed a measure of trait extraversion as part of a battery of questionnaires during an initial laboratory-based assessment ($20 compensation). For the next two weeks, they completed ESM measures of social interactions and well-being four times per day (for the opportunity to win $100; odds of winning were 1 in 10 if all ESM reports were completed). In addition, most participants (N = 311) wore the EAR for the first week ($20 compensation), providing audio recordings of their everyday lives that were later coded for social interaction variables. We ended data collection when we reached the end of a semester and had recruited at least 400 participants. After exclusions (described below), the final subset of 256 participants (178 women, 77 men, one gender not reported) used in the current analyses ranged in age from 18 to 29 years (M = 19.17, SD = 1.78) and identified as Caucasian (n = 144), Asian (n = 61), Black (n = 25), American Indian or Alaska Native (n = 1), Other (n = 18), or did not disclose their ethnicity (n = 7). See the Appendix for a demographic comparison of participants who were included versus excluded from the current analyses. ESM and EAR Procedures ESM. The ESM portion of the study began after participants completed the laboratory-based assessment. Four times per day (at 12 p.m., 3 p.m., 6 p.m., and 9 p.m.) for 15 days, participants received a text message notification and were emailed a link to a survey that contained ESM measures of their social interactions and well-being in the hour that preceded the notification (11 a.m.-12 p.m., 2 p.m.-3 p.m., 5 p.m.-6 p.m., and 8 p.m.-9 p.m.). ESM data exclusions. In line with exclusion criteria applied in previous papers that used the PAIRS ESM data (e.g., Finnigan & Vazire, 2018; Wilson et al, 2017), we excluded ESM reports (a) if they were completed more than 3 hr after the notification was sent, (b) if participants completed fewer than 75% of the items, (c) if participants used the same response option for at least 70% of the items, or (d) if participants indicated that they were asleep during the entire target hour. We also excluded practice ESM surveys that were completed during the participant's initial laboratory session. After these exclusions, 10,949 reports from 406 participants remained (across the 2-week ESM period, including the week in which participants were not wearing the EAR). EAR data collection. During the first week (6-8 days) of the ESM protocol, 311 participants wore the EAR, implemented through the iEAR app using an iPod Touch device. The EAR component of the study was optional, was only offered during 1482 SUN, HARRIS, AND VAZIRE nonsummer months of the study, and was not an option when all of the researchers' iPod Touches were in use by other participants. The EAR was programmed to record 30 s audio snippets of participants' ambient sounds, every 9.5 min from 7 a.m. to 2 a.m. Participants were encouraged to wear the EAR as much as possible and to wear it clipped to a waistband or the outside of their pockets (i.e., not inside a bag or pocket). Although participants had no way to tell when the device was recording, they were told that they could decide to not wear the EAR at any time for any reason. After 3-4 days, participants returned to the lab to upload their data (due to device memory limitations), and then continued wearing the device before returning it after another 3-4 days. EAR data exclusions. Upon returning the device at the end of the week, participants received a compact disk with their recordings, so that they could listen to and erase any files they did not want the researchers to hear. Only a few participants (n = 15) chose to erase a total of 99 files. After deleting these files, along with files from six participants who withdrew, and one participant who only had silent recordings (suggesting that the microphone malfunctioned), 152,592 usable recordings from 304 participants remained. EAR codings. Research assistants from Washington University in St. Louis (n = 8) and University of California, Davis (n = 137) listened to the audio files recorded during the same hours as the ESM reports (11 a.m.-12 p.m., 2 p.m.-3 p.m., 5 p.m.-6 p.m., and 8 p.m.-9 p.m.), and coded participants' social interactions (and other variables), across two coding tasks (described below). Some research assistants were involved in both tasks, but were only assigned participants that they did not code in the other task. Thus, across the two tasks, the codings for each participant were provided by different sets of research assistants. Hour-level codings. For the hour-level codings, for each of their assigned participants, coders listened to the six or seven 30 s files (3-3.5 min total) from each ESM-matched hour, coded whether or not the participant interacted with others, rated participants' conversational depth and levels of self-disclosure during that hour (if the participant was interacting with others), then moved onto the next ESM-matched hour for that participant. Because research assistants joined and left the lab at different times, each participant was coded by a different set of coders. Initially, we aimed to have each participant coded by three coders. However, as the interrater reliabilities based on three coders were low, we decided to add three more coders, so that each participant was coded by at least six coders. Between the two sets of codings, we made minor changes to the coding protocol (see the online supplemental materials), in hopes of increasing intercoder reliability. File-level codings. For the file-level codings, coders again listened to only the audio files that were part of each ESM-matched hour. However, unlike for the hour-level codings, they listened to and coded each file separately, coding whether or not the participant was interacting with others during each file (rather than providing a single binary judgment for the entire hour). For this coding task, all but four participants were coded by three or more coders. EAR coding exclusions. Coders only rated participants' social interactions in hours and files that contained sufficient acoustic information. For the hour-level task, we only kept hours that at least three coders rated as being informative (i.e., no technical difficulties, and participants appeared to be awake and wearing the EAR; see the online supplemental materials for details). Based on these criteria, 807 out of 5,222 hr (15.45%) were uninformative (and excluded from further analyses). For the file-level task, we only kept files that at least two coders rated as being informative. Based on these criteria, 4,208 out of 31,417 files (13.4%) were uninformative (and excluded from further analyses). Measures See Table 1 for within- and between-person reliability coefficients for all composites. Social interactions. Quantity of interactions. Self-reported. In the ESM surveys, participants completed the item, "From [11 a.m.-noon/2 p.m.-3 p.m./5 p.m.-6 p.m./8 p.m.-9 p.m.], were you interacting with other people?" (response options: no, one person, two people, three to five people, more than five Table 1 Descriptive Statistics and Inter-Correlations Among All Observed Variables Descriptive statistics Between-person correlations Variable M SDWP SDBP 1 - ICC(l) wwp "bp 1 2 3 4 5 6 7 8 9 10 1. Binary interactions (ESM) .78 .40 .13 .90 2. Binary interactions (EAR) .70 .44 .13 .93 .94 .98 .40 3. Continuous interactions (EAR) .33 .32 .09 .92 .93 .90 .35 .67 4. Conversational depth (ESM) 3.11 0.96 0.45 .82 .04 -.05 .01 5. Conversational depth (EAR) 2.71 0.42 0.15 .88 .70 .24 .01 -.08 .12 -.05 6. Self-disclosure (ESM) 2.53 1.05 0.48 .83 .20 -.03 .07 .43 .10 7. Self-disclosure (EAR) 2.36 0.54 0.19 .89 .78 .41 .10 -.06 .23 .10 .54 .32 8. Knowing (ESM) 3.80 1.06 0.38 .89 .14 .16 .15 .28 -.15 .12 .01 9. Liking (ESM) 4.21 0.76 0.36 .82 .27 .20 .26 .38 -.12 .21 .07 .62 10. Happiness (ESM) 3.46 0.81 0.49 .73 .82 .98 .28 .21 .20 .27 -.09 .26 .01 .22 .35 11. Social connectedness (ESM) 3.57 0.80 0.44 .77 .52 .48 .49 .33 .41 .33 -.05 .33 .08 .39 .54 .66 12. Trait extraversion 9.29 2.84 .90 .27 .21 .19 .16 -.06 .17 .10 .16 .23 .35 Note. Means were computed from the aggregate observed mean for each person. SDWP = within-person SD; SDBP = between-person SD; wwp = within-person omega reliability; wBP = between-person omega reliability; ESM = experience sampling method; EAR = Electronically Activated Recorder. ICC(l), the intraclass correlation, represents the proportion of total variance (obp + trWp) that is due to between-persons variability (obp; i.e., mean-level differences on a variable across the week), so 1 — ICC(l) denotes the % of total variance due to within-person variability (owp; i.e., fluctuations around a person's mean level). These between-person correlations are based on the aggregate observed mean for each person, which is why they are different from the latent self-observer agreement correlations reported in-text. Correlations > 1.151 are significant at p < .05, not corrected for multiple comparisons. SOCIAL INTERACTIONS AND WELL-BEING 1483 people). We recoded these responses into two categories that denote whether a social interaction occurred (coded as 1) or not (coded as 0) during the target hour. Because participants were not provided with an explicit definition of what counted as "interacting with other people," these self-reported social interactions could have included computer-mediated social interactions. Observer-based. We had two observer-based measures of the quantity of social interactions. The first measure was a binary measure based on whether the participant had interacted at all in the target hour, analogous to the self-report measure described above. After listening to the six or seven 30 s files for the hour, EAR coders completed the same item as in the ESM survey, "Was the participant interacting with other people?" (response options: no, one person, two people, three to five people, more than five people), with respect to the entire hour. We recoded each coder's responses into two categories that denote the absence (coded as 0) or presence (coded as 1) of a social interaction during the target hour. Then, we aggregated the responses across coders by recoding the hour-level score to 0 {no interaction) if the majority of coders said that the participant did not interact during that hour, and to 1 {interaction occurred) if at least half of the coders said that the participant interacted during that hour. The second measure provided a separate, continuous measure of social interaction during the same hours, using data from the file-level codings (i.e., codings of each of the six or seven 30 s sound files in a given hour). Coders completed the item, "During this file, was the participant interacting with other people?" (0 = no, 1 = yes) for each individual file (rather than the entire hour). We aggregated the scores across coders by recoding the file-level score to 0 {no interaction) if the majority of coders said that the participant did not interact in that file, and to 1 {interaction occurred) if at least half of the coders said that the participant interacted in that file. Then, we aggregated the file-level scores to a continuous hour-level score by taking the mean of all of the informative files in that hour (up to seven files). This continuous variable could range from 0 {no social interactions in any of the sound files in that hour) to 1 {social interaction occurred in all six or seven files in that hour). Quality of interactions. Self-reported. If participants indicated that they had interacted with at least one person in the target hour, they completed four additional 1-item measures about the quality of their interactions. Participants rated the depth of their own conversations ("How superficial (i.e., shallow) to substantive (i.e., deep) were the conversations?"; on a scale ranging from 1 [very superficial] to 5 [very substantive]), and how much they self-disclosed ("How much did you self-disclose?"; on a scale ranging from 1 [not at all] to 5 [a lot]) during the target hour. Participants also reported on two relational features—how much they knew and liked the people they interacted with ("How well do you [know/like] them?"; on a scale ranging from 1 [not at all] to 5 [very well/very much]). Observer-based. If EAR coders indicated that the participant had interacted with at least one person in the target hour, EAR coders rated the depth of the participants' conversations ("How superficial (i.e., shallow) to substantive (i.e., deep) did the conversations sound?"; on a scale ranging from 1 [very superficial] to 5 [very substantive]), and how much the participant self-disclosed ("How much do you think the participant self-disclosed?"; on a scale ranging from 1 [not at all] to 5 [a lot]) during the target hour. Coders also had the option to select "No way to tell" (rather than a number on the 1-5 scale). EAR coders also completed measures of how much participants knew and liked the people they were interacting with, but we chose not to analyze these measures as we thought that it would be difficult for EAR coders to tell how much participants knew and liked their interaction partners. Given the subjective nature of these variables, we decided that, unlike conversational depth and self-disclosure, which can be observed by others, the observer-based measures of knowing and liking one's interaction partners) would be unlikely to contain any valid variance not captured by self-reports. Momentary well-being. As part of the ESM survey, participants completed self-report measures of their momentary feelings of happiness and social connectedness during the target hour (e.g., "from 11 a.m. to 12 p.m."). Happiness. To measure feelings of happiness, we averaged two items: "How happy were you?" (on a scale ranging from 1 [not at all] to 5 [very]) and "How much positive emotion did you experience?" (on a scale ranging from 1 [none at all] to 5 [a lot]). All participants had data on the happiness item, but data on the positive emotion item was missing for 51 of the 256 participants, as this item was added after data collection had begun. Social connectedness. To measure feelings of social connectedness, we averaged together two items: "did you feel 'close, connected' to others?" and, reverse-scored, "how lonely were you?" (on a scale ranging from 1 [not at all] to 5 [very]). Trait extraversion. Participants completed the Big Five Inventory (BFI-44; John & Srivastava, 1999), which includes an eight-item measure of trait extraversion. Responses were made on a 15-point scale (ranging from 1 [disagree strongly] to 15 [agree strongly]). We computed z-scores for trait extraversion prior to using them in the moderation analyses. These z-scores were computed separately for participants who were included in the quantity of social interaction analyses and participants who were included in the quality of social interaction analyses (see final sample details below). Data Included in Final Analyses Quantity of interactions. To hold the time points constant across all quantity analyses, we first excluded observations that were missing either ESM or EAR data, resulting in 3,292 observations that had both ESM and EAR data. Then, we excluded 33 participants who had fewer than five observations, resulting in 3,206 observations across 256 participants for these analyses. Quality of interactions. Participants and observers only reported on the quality of social interactions when the participant had been interacting. Participants and observers agreed about whether or not the participant had interacted with someone in the past hour 70.12% of the time (agreement is weakened by the fact that EAR coders only heard 3 to 3.5 min of the hour). To hold the time points constant across all quality analyses, we only included the 1,836 time points for which participants and observers agreed that the participant had interacted with someone. Then, we excluded 64 participants who had fewer than five social interactions, resulting in 1,641 observations across 192 participants for these analyses. Data Analysis The data had a multilevel structure, with observations (Level 1) nested within participants (Level 2). To model the within- and 1484 SUN, HARRIS, AND VAZIRE between-person associations that social interaction variables had with well-being, we used Muthen and Asparahouv's (2009) general multilevel structural equation modeling (MSEM) framework, implemented using Mplus Version 8.3 (Muthen & Muthen, 1998-2017) and facilitated by the R package MplusAutomation (Hallquist & Wiley, 2018). MSEM uses latent variable decomposition, which allows for Level 1 and Level 2 effects to be simultaneously estimated. We ran separate models for each of the nine predictors (self-reported and observer-rated binary interactions, observer-rated continuous interactions, self-reported and observer-rated conversational depth and self-disclosure, and self-reported knowing and liking) and the two well-being outcomes (feelings of happiness and social connectedness), with models either including or excluding trait extraversion as a moderator (described below). Measurement models. Latent variables. We modeled EAR-coded conversational depth and self-disclosure as latent variables, to account for intercoder unreliability in these predictors (see Figure 1). For these latent variables, we used coders as indicators. Some hours were coded by Between model Coder 1 Within model Coder. Coder 6 Figure 1. Measurement model for the Electronically Activated Recorder (EAR) observer-based conversational depth and self-disclosure variables. The six residuals at each level were constrained to equality (for each respective level). more than six coders, but to reduce model complexity, for the latent variables, we only included data from up to six coders (see the online supplemental materials for details). Thus, every variable had six indicators (with each indicator representing the observed score from a given coder, for a given participant). For a given participant (e.g., Participant 1), all ratings from coder 1 were from the same coder (e.g., Research Assistant 1). However, for a different participant (e.g., Participant 2), Coder 1 could have been a different research assistant (e.g., Research Assistant 2). To model the interchangeability of coders, we fixed all loadings for the six indicators to 1, constrained the six residual variances to be equal, and allowed the variance of the latent observer-rated variable to be freely estimated. Observed variables. All other variables were modeled as observed variables in the structural models described below. These included self-reported binary social interactions, conversational depth, and self-disclosure; observer-rated binary and continuous social interactions; self-reported happiness and social connectedness; and self-reported trait extraversion. Reliability estimates. We conducted multilevel confirmatory factor analyses (MCFA; Geldhof, Preacher, & Zyphur, 2014; Shrout & Lane, 2012) to obtain level-specific omega (oo) reliability estimates for the EAR-coded social interaction variables and the ESM happiness and social connectedness variables. Because the ESM happiness and social connectedness variables each only had two indicators, we constrained the factor loadings for the two items to be equal at each level for these MCFA models. To estimate the reliability of the trait extraversion measure, we computed coefficient a) using the MBESS package (Version 4.4.3; Kelley, 2018). These reliability estimates are reported in Table 1. Structural models. We illustrate the MSEMs in Figure 2. In all models, y denotes the outcome variable (happiness, social connectedness, or the individual "close, connected" and "lonely" items used in the supplemental analyses; see online supplemental materials), x denotes the social interaction predictor variable, and z denotes the moderator variable, trait extraversion. The subscripts i and j denote observations at time i for person j. For simplicity, Figure 2 does not depict the measurement model used for the EAR conversational depth and self-disclosure variables (shown in Figure 1). Models for main effects. In the first set of models (see Figure 2, Models A-C), we regressed each well-being outcome onto each social interaction predictor at both the within- and between-person levels, with random intercepts and random slopes. This allows each participant to have a different mean level of well-being, and a different association between each social interaction variable and well-being. For the quantity of social interaction analyses (see Figure 2, Model A), we estimated both the within- Owl) and between-person (PB1) effects in the same model. For these analyses, the person-level estimates of well-being were based on latent variable decomposition. However, for the quality of social interaction analyses, we estimated the within- and between-person effects in two separate models (see Figure 2, Models B-C). Because the quality variables only applied when a social interaction occurred, using person-level estimates of well-being based on this subset of time points would only enable inferences about the association between the average quality of interactions and average well-being during social interactions (rather than overall well-being across all time points, including hours in which the participant was not interact- SOCIAL INTERACTIONS AND WELL-BEING 1485 Figure 2. Path diagrams representing the multilevel structural equation models used in the study. Bold coefficients denote the key parameter(s) of interest in each model. Squares represent observed variables, circles represent latent variables, and filled-in circles represent random slopes (labeled as Sy) or interactions (labeled as XjXzj). Double-headed arrows represent variances and covariances. ing). Thus, to draw inferences about the associations between the average quality of social interactions and overall well-being, we computed person-level happiness and social connectedness aggregate scores for each person using all 3,206 time points, and estimated the between-person effects predicting these observed scores. This ensures that the between-person effect ((3B1, Model C) represents the association that the quality of social interactions has with well-being across all time points, not only the time points in which participants were interacting with others (represented by the PB1 effect in Model B, which we are not interested in). Models for interaction effects. In the second set of models (see Figure 2, Models D-F), we added trait extraversion as a moderator of the association between aspects of social interactions and well-being. As for the main effects, we used one model to estimate the cross-level and between-person interaction effects for the quantity of social interaction predictors (Model D), but used two separate models to estimate the cross-level (Model E) and between-person (Model F) interaction effects for the quality of social interaction predictors. For reasons described above, the between-person well-being variable was latent for the quantity of social interaction model (Model D) and observed (aggregated across all time points) for the between-person quality of social interactions model (Model F). Trait extraversion was modeled as an observed, z-scored between-person variable (z-). To estimate the cross-level interaction effect (PB3), we regressed the random slope (Sy) onto trait extraversion. To estimate the between-person interaction effect (PB4), we constrained the mean of the latent between-person social interaction variable to zero (which centers the predictor). Then, we regressed the between-person well-being variable onto the interaction between trait extraversion and the latent between-person social interaction variable X z-). In all moderation models, we modeled the main effect of trait extraversion on average well-being 0B2), as well as the covariance between trait extraversion and the social interaction predictor. Estimation and inference criteria. We used the Bayes estimator in Mplus Version 8.3 (Muthen & Asparouhov, 2012), with the default set of diffuse (i.e., noninformative) priors. We use the 95% credibility interval (CI) around the standardized effects ((3) as inference criteria for the range of plausible population values of the effect sizes. We standardized the quality of social interaction effects against the standard deviations of both the predictor and the well-being outcome. The binary and continuous social interaction variables are on a readily interpretable metric, so we only standardized the wifhin-person effects against the standard deviations of the well-being outcome variables 1486 SUN, HARRIS, AND VAZIRE (but standardized the between-person effects against the standard deviations of both the predictor and the well-being outcome). Thus, the wifhin-person (3 for the binary social interaction variable represents the standard deviation increase in momentary happiness or social connectedness when participants were interacting versus when they were not, and the wifhin-person (3 for the continuous variable represents the standard deviation increase in momentary happiness or social connectedness when participants were interacting in 100% of the files in the target hour versus when they were interacting in none of the files in that hour. All standardized effects were computed based on the standard deviations of the respective levels (i.e., wifhin-person or between-person). Results Preliminary Analyses Descriptive statistics and intercorrelations among all variables are shown in Table 1. Omega reliability estimates showed that 70% to 94% of the wifhin-person fluctuations in each of the social interaction variables, as assessed by three to six EAR coders per participant, were due to meaningful fluctuations (rather than random noise). The two happiness items reliably assessed true fluctuations in momentary happiness (a)WP = .82). Although the composite of the two social connectedness items had lower reliability (d)WP = .52), we nevertheless chose to combine them for conceptual reasons (see Supplemental Table SI in the online supplemental materials for item-level results). EAR coders also reliably assessed between-person differences in quantity of social interactions (a)BP ^ .90), but showed much lower reliability when assessing between-person differences in conversational depth (d)BP = .24) and self-disclosure (a)BP = .41). MSEM corrects for attenuation of point estimates due to measurement error, but greater measurement error results in less precise estimates. Next, we assessed the extent of agreement between ESM self-reports and EAR observer reports of social interactions. Latent wifhin-person correlations based on MSEMs showed that participants and observers agreed moderately on when participants were interacting or not, (r = .39, 95% CI [.35, .42]), and on moment-to-moment fluctuations in conversational depth (r = .31, 95% CI [.25, .37]) and self-disclosure (r = .31, 95% CI [.25, .36]). One reason that agreement was not higher may be that EAR coders only listened to 3 to 3.5 min of each hour that participants reported on. Latent between-person correlations based on MSEMs also showed that participants and observers agreed moderately on which participants interacted more often (r = .52, 95% CI [.25, .70]), and self-disclosed more on average (r = .51, 95% CI [.12, .80]). However, there was no self-observer agreement on which participants had deeper conversations on average (r = —.33, 95% CI [ — .64, —.08]; note that this association was between latent variables and that the observed association was much smaller [-.05], see Table 1). Quantity of Social Interactions Within-person effects. Do people feel happier and more socially connected when interacting with others? We found that this was the case for both self-reported and observer-rated social interactions (see Table 2 and Figures 3-4). Indeed, the entirely positive within-person slopes in Figures 3-4 show that every u cd d c I % 1) oj Table 2 Predicting Happiness and Social Connectedness From Social Interactions Happiness Social connectedness Self-reported (ESM) Observed (EAR) Self-reported (ESM) Observed (EAR) Predictor interactions interactions interactions interactions ß R2 ß R2 R2 R2 Quantity of interactions Binary interactions Within 0.59 [0.49, 0.68] .07 0.45 [0.37, 0.53] .05 1.08 [0.98, 1.16] .20 0.69 [0.61, 0.77] .10 Between 0.32 [0.12, 0.49] .10 0.21 [0.01, 0.42] .05 0.52 [0.37, 0.65] .27 0.38 [0.18, 0.55] .14 Continuous interactions Within 0.79 [0.68, 0.90] .08 1.08 [0.97, 1.17] .13 Between 0.16 [-0.02, 0.33] .03 0.46 [0.24, 0.63] .21 Quality of interactions Conversational depth Within 0.12 [0.08, 0.17] .07 0.04 [-0.03, 0.10] .04 0.19 [0.14, 0.22] .11 0.15 [0.08, 0.21] .10 Between 0.37 [0.16, 0.55] .14 -0.24 [-0.57, 0.06] .06 0.44 [0.29, 0.61] .20 -0.08 [-0.32, 0.18] .01 Self-disclosure Within 0.21 [0.16, 0.27] .08 0.14 [0.08, 0.20] .04 0.27 [0.22, 0.33] .12 0.19 [0.13, 0.25] .06 Between 0.38 [0.19,0.56] .14 0.05 [-0.20, 0.34] .01 0.45 [0.24, 0.59] .20 0.16 [-0.13, 0.47] .03 Knowing Within 0.27 [0.21,0.31] .08 0.31 [0.26, 0.36] .15 Between 0.29 [0.08,0.51] .09 0.56 [0.41,0.78] .31 Liking Within 0.36 [0.32, 0.41] .14 0.42 [0.38, 0.46] .19 Between 0.47 [0.30, 0.59] .22 0.68 [0.54, 0.82] .46 Note. ESM = experience sampling method; EAR = Electronically Activated Recorder. The within-person quantity of interactions (3s are only standardized with respect to the well-being outcome. All other coefficients are standardized with respect to both the predictor and outcome. R2 = proportion of variance explained at each level. 95% credibility intervals are shown in brackets. SOCIAL INTERACTIONS AND WELL-BEING 1487 "5 u cd d c 1) oj 50 40 >. U g30 |20 it 10 ESM EAR cfl 4 W a) a (0 0.00 0.25 0.50 0.75 1.00 1.25 1.50 Slopes for Binary Interactions -60 -40 -20 0 20 40 60 80 % Files with Interactions (EAR) -3-2-10 1 2 Conversational Depth (ESM) -3 -2-10 1 2 Conversational Depth (EAR) -2-10 1 2 Self-Disclosure (ESM) ■2-10 1 2 Self-Disclosure (EAR) -3-2-10 1 2 Knowing (ESM) -1 o 1 Liking (ESM) Figure 3. Within-person associations between aspects of social interactions and momentary happiness. The top-left panel shows the histograms for the unstandardized within-person associations between momentary happiness (outcome) and whether or not participants interacted with someone in the past hour (predictor) as rated by the self (experience sampling method [ESM]) or observers (Electronically Activated Recorder [EAR]; estimated in two separate models). The solid vertical lines show the mean slopes. For the remaining panels, each thin line represents the within-person association between each social interaction variable and momentary happiness for each person, and the thick line shows the average within-person association. The x-axis shows raw deviations from each person's mean social interaction state, whereas the y-axis shows the uncentered 1-5 score on momentary happiness. See the online article for the color version of this figure. single participant tended to feel happier and more socially connected when they interacted in the past hour, compared to when they did not. Specifically, when participants self-reported interacting (vs. not) in the past hour, their momentary happiness was on average 0.59 SD higher, and their momentary social connectedness was on average 1.08 SD higher. The effects were slightly smaller, but still substantial, for observer-rated interactions: When participants were observed interacting (vs. not) in the past hour, their momentary happiness was on average 0.45 SD higher, and their momentary social connectedness was on average 0.69 SD higher. Does the amount of social interaction within an hour also matter? The more fine-grained continuous measure based on observers' codings showed that participants generally reported greater momentary happiness and feelings of social connectedness when they were observed interacting during a greater proportion of 30 s recordings in a given hour (see Table 2 and Figures 3-4). The effect sizes showed that participants were on average 0.79 SD higher in happiness and 1.08 SD higher in social connectedness when they were observed to be interacting in all recordings in a given hour (vs. none of them). Between-person effects. Do people who interact more with others on average also feel happier and more socially connected on average? Participants who had a greater proportion of hours in which they self-reported or were observed interacting with others tended to report greater average feelings of happiness and social connectedness (see Table 2). For the continuous measure of social interactions, participants who were observed interacting in a greater percentage of files on average reported greater average 1488 SUN, HARRIS, AND VAZIRE 70 5" 50 S 40 3 CT30 ■t 20 10 ESM EAR S V) lit o n -5 C 2 O o 1 0.00 Conversational Depth (ESM) -2-10 1 2 Self-Disclosure (ESM) 5 o) (a £4 73