Life events as predictors of wellbeing outcomes

Abstract

Life events are noteworthy events that punctuate our lives and shape our experiences; yet their impact on wellbeing remains complex. This paper examines how the occurrence and characteristics of life events affect well-being outcomes, leveraging a year-long observational longitudinal dataset involving 493 adult participants (ages 21–68, mean = 35.85, stdev. = 10.04, 58.42% male and 41.58% female), who completed daily self-reports of affect, stress, anxiety, and sleep. We employed mixed-effects regression models to examine the associations between individual differences (e.g., age, gender, personality traits, baseline wellbeing traits) and life event attributes (e.g., anticipation, valence, intimacy, temporal status, and Facebook disclosure of life events) with wellbeing outcomes. Health-related life events were found to be linked to worsened negative affect and stress, as well as poorer sleep quality. Anticipated events—those expected to occur in the near future—were associated with greater positive affect (f ² = 0.28), lower negative affect (f ² = −0.27), and poorer sleep quality (f ² = −0.12). Similarly, valence—positivity or negativity of an event—was positively associated with positive affect (f ² = 0.24) and negatively associated with negative affect (f ² = −0.21), stress (f ² = −.08), and anxiety (f ² = −0.07). In contrast, event intimacy—defined as how personally sensitive or private an event is—was negatively associated with positive affect (f ² = −0.15). Again, events with a continuous temporal status (i.e., those spanning multiple days) were linked to both higher positive (f ² = 0.31) and negative affect (f ² = 0.27), as well as increased stress (0.20) and anxiety (f ² = 0.13). Interestingly, sharing life events on Facebook is associated with higher positive affect (f ² = 0.22) and lower negative affect (f ² = −0.10), stress (f ² = −0.31), and anxiety (f ² = −0.15), but linked to poorer sleep (f ² = −0.16). These findings provide an empirical understanding of the interplay between life events and psychological responses.

Introduction

Our lives are characterized by dynamic, often unpredictable, trajectories punctuated by “life events”—significant shifts in personal or social circumstances—that have long been recognized as critical inflection points in our lives^1,2,3. These events, ranging from joyous milestones such as marriage, childbirth, and career achievements to challenging adversities such as illness, death of a family member, and financial hardships, have the potential to profoundly reshape our psychophysiological responses and behavioral patterns^4,5,6,7. Therefore, life events, whether positive or negative, individual or collective, can potentially disrupt established routines, trigger emotional fluctuations, and ultimately influence overall wellbeing^8,9.

Theoretical frameworks from psychology help explain why life events—both positive and negative—may be associated with fluctuations in our wellbeing. According to theories of subjective wellbeing^7,9, individuals’ daily affective experiences are influenced not only by stable traits but also by situational factors, such as meaningful life events. Lazarus and Folkman’s stress and coping theory⁶ further emphasizes that individuals’ appraisals of and responses to life events can produce emotional and physiological consequences, particularly in terms of stress and anxiety. Similarly, research on affective dynamics¹⁰ highlights the importance of capturing fluctuations in mood and emotion in response to external events.

These theoretical perspectives motivate the need to empirically examine how specific attributes of life events relate to wellbeing experiences. Yet, despite the pervasive influence of life events, the pathways through which life events shape wellbeing outcomes remain complex and not yet fully understood. Although prior work has established broad associations between life events and wellbeing—such as long-term effects on depression, life satisfaction, and even suicidal ideation^11,12—more granular insights into how specific attributes of these events relate to different dimensions of wellbeing are still lacking. For example, prior research showed that specific types of life events—particularly health-related events—are consistently associated with fluctuations in affective wellbeing¹³. Our study contributes to this space by offering an empirical analysis of the associations between a diverse range of life event attributes—such as anticipation, valence, intimacy, temporal status, and online disclosure—and immediate wellbeing outcomes.

Understanding these short-term associations is valuable not only for theoretical refinement but also for informing more personalized and timely strategies to support wellbeing during life transitions^14,15. By identifying which attributes of events are most strongly linked to wellbeing outcomes, researchers and practitioners can better recognize risk periods and tailor interventions accordingly. Accordingly, this paper aims to explore how life events disrupt mental wellbeing by analyzing the immediate wellbeing outcomes around life event occurrences. Specifically, our work is guided by the research question: How does experiencing a life event influence an individual’s immediate wellbeing?

We conducted a study using data from the Tesserae project¹⁶—a large-scale and longitudinal multimodal sensing research initiative. This dataset includes year-long data from 493 participants who responded to daily self-reported wellbeing surveys on positive and negative affect, stress, anxiety, and sleep. Participants also completed an exit survey based on the PERI life event scale¹⁷ where they self-reported their life events during the past year. Also, 330 of these participants provided their Facebook timeline data, on which prior work manually coded and labeled life event disclosures and corresponding attributes driven by the PERI life event scale^18,19. Importantly, both the self-reported life events from the PERI survey and the Facebook life event disclosures included specific timestamps of life event occurrences, which allowed us to temporally align each event with daily self-reported wellbeing data and examine short-term fluctuations surrounding the time of the event.

In this study, we conceptualize our outcome variables within the framework of subjective wellbeing, which includes both hedonic components (positive and negative affect) and broader indicators of day-to-day psychological functioning such as perceived stress, anxiety, and sleep quality^6,8,9,10. These measures reflect short-term wellbeing states rather than clinical indicators of mental illness, aligning with prior research that uses momentary self-reports to capture dynamic aspects of psychological wellbeing^20,21,22. To examine the relationship between life event occurrences and wellbeing outcomes, we developed regression models incorporating both individual characteristics (e.g., demographics, psychological traits) and life event attributes (e.g., type, anticipation, valence, intimacy, and significance). Our findings indicate that models integrating both psychological traits and life event attributes outperformed those relying solely on either of these factors, highlighting the significant role that life events play in shaping wellbeing outcomes. This work contributes to a growing body of research examining life experiences and wellbeing by shifting the focus to event-level granularity and short-term dynamics. Our findings underscore the importance of accounting for not just the occurrence of life events, but also their specific attributes, when assessing psychological wellbeing.

Results

Table 1 reports the performance of the models—M₁ (using only individual differences), M₂ (using only life event attributes), and Table 2 shows a comparison across different types of algorithms—we find that the Support Vector Regressor (SVR) best fits our data. However, to prioritize interpretability—which is the primary goal of our study—we present our ensuing analyses using the linear regression model (with L2 regularization/Ridge) in this paper. Table 3 reports the ANOVA of comparing different model pairs. Overall, we find that based on R², SMAPE, and r, M₃ models perform significantly better than M₁ and M₂—i.e., an individual’s wellbeing outcomes are strongly associated with both individual differences and life event occurrences. To further drill into the findings, we note that M₃ models show high R² for all wellbeing metrics—0.78 for positive affect, 0.77 for negative affect, 0.59 for stress, 0.63 for anxiety, and 0.64 for sleep quality (all with statistical significance). In terms of SMAPE, M₃ models show a mean SMAPE=20.55, which is better than the SMAPE of M₁ models (25.66) and M₂ models (30.31). Similarly, M₃ models show a mean Pearson’s r = 0.56, which is better than r of M₁s (0.28) and M₂s (0.23). Further, ANOVA reveals statistical significance in comparing M₃ models with M₁ and M₂ models, suggesting that the addition of independent variables indeed added significant information to the models. These observations suggest that the occurrences of life events bear significant explanatory power toward fluctuations in an individual’s wellbeing. This is consistent with prior findings on the impact of life events on wellbeing outcomes¹.

Table 1 Performance of regression models predicting immediate wellbeing outcomes (^*p < 0.05, ^**p < 0.01, ^***p < 0.001)

Table 2 Performance of different M₃ regression models predicting immediate wellbeing outcomes (^*p < 0.05, ^**p < 0.01, ^***p < 0.001)

Table 3 Comparing the regression models M₁, M₂, and M₃ with ANOVA tests (^*p < 0.05, ^**p < 0.01, ^***p < 0.001)

Factors associated with wellbeing outcomes

Table 4 summarizes the coefficients and the standardized effect sizes (d) of the variables in the mixed-effects M₃ regression models, where the statistical significance (p-values) are derived from the unregularized linear regression models. For all wellbeing measures, we note that several independent variables show statistical significance. We go over these associations below:

Positive Affect. Here, we observe that life events—varying across types—show negative coefficients, indicating that their occurrences are associated with lowering positive affect. Similarly, positive coefficients for anticipation (0.37) and valence (0.24) suggest that unanticipated or negative life events are likely linked to reduced positive affect, whereas anticipated and positive life events correspond to increased positive affect. Furthermore, life events with higher scope (more individualized) and higher significance are associated with greater positive affect. This observation aligns with prior research on the emotional impact of life events, which emphasizes the role of anticipation, valence, and personal relevance in shaping emotional responses²³.

Negative Affect. For negative affect, health-related life events share a positive coefficient, i.e., the occurrences of health-related events are likely to be associated with higher negative affect. This aligns with prior research showing that health-related stressors, such as illnesses or injuries, could lead to higher levels of negative emotional states²⁴. We also note negative coefficients for anticipation (−0.19) and valence (−0.31)—i.e., life events that are anticipated or positive are likely to be associated with lower negative affect. This observation is consistent with studies indicating that anticipation of positive events or life changes typically reduces negative emotions²⁵. The positive coefficient of scope (0.37) and significance (0.17) indicate that more individualized and more significant events were likely to be associated with high negative affect. The role of personal relevance and perceived significance of life events in intensifying negative emotional responses is supported by prior research²⁶.

Stress. We note similar signs, like in the case of negative affect. Here, health- and work- related events are likely to be associated with high stress. This finding aligns with research indicating that health problems and work-related stressors are major contributors to increased stress^25,27. Additionally, anticipated, positive, or significant life events are likely to be associated with lower stress, consistent with findings that anticipation of positive events or those perceived as beneficial can reduce stress levels ^26,28. Interestingly, we note a negative coefficient for scope (−0.17), i.e., less individualized events are likely to be associated with higher stress.

Anxiety. Interestingly, health-related events show a small negative coefficient, whereas other types of life events do not show statistical significance in association with anxiety. The negative coefficient of valence (−0.07) indicates that positive life events are likely to be associated with lower anxiety, whereas negative life events are likely to be associated with higher anxiety. The positive coefficient of scope (0.15) indicates that more individualized events are associated with higher anxiety. This suggests that personally relevant or individualized events might feel more overwhelming and induce greater anxiety, aligning with stress appraisal theories which emphasize the personal significance of stressors^6,26.

Sleep Quality. For sleep quality as a dependent variable, positive coefficients indicate association with better sleep, whereas negative coefficients are associated with poorer sleep. Therefore, health and work-related events were associated with poorer sleep, aligning with prior research²⁹. The negative coefficients of anticipation (−0.05) and valence (−0.13) indicate that even anticipated and positive events are likely to be associated with poorer sleep. Again, continuous events are likely to be associated with poorer sleep.

Table 4 Linear mixed effect models

Robustness of analysis

We examined the robustness of our analyses with respect to researcher decisions in modeling our data. For instance, we adopted a seven-day time-window (three days before and after the event) approach to calculate immediate wellbeing outcomes. This choice was based on the inherent sparsity of social media data, which rarely provides instantaneous, high-fidelity insights in real-time. Additionally, while life events are recorded as discrete occurrences, they often unfold over a period—such as the birth of a child or a vacation. By averaging wellbeing outcomes over a small time window, we aimed to capture a more stable representation of wellbeing, not sensitive to immediate fluctuations (if any) or a single wellbeing report or immediate. However, to ensure that our analyses and observations are not significantly skewed by our choice of time-window, we also examined two other time windows—three-days and five-days. We repeated the same modeling again, and conducted ANOVA tests between the same type of model but different time-windows. Table 5 summarizes the three models, each using a different time window for averaging wellbeing measures. The models demonstrate comparable performance in terms of R² and Pearson’s r, with the seven-day averaging model performing slightly better on these metrics. Additionally, ANOVA tests reveal no statistically significant differences between the models. We conclude that the models are robust in terms of the choice of time window.

Table 5 Comparing regression models for different rolling time windows of computing immediate wellbeing (three days, five days, and seven days) (^*p < 0.05, ^**p < 0.01, ^***p < 0.001)

Discussion

Our findings revealed multifaceted insights into the relationship between life event occurrences and wellbeing outcomes. Regression analyses revealed that both individual differences and life event attributes are significantly associated with an individual’s wellbeing outcomes of affect, stress, anxiety, and sleep. Given the exploratory nature of our analyses and the absence of strong prior theory linking specific life event characteristics to immediate wellbeing outcomes, our findings are best interpreted as hypothesis-generating. Future work can build on these results to design confirmatory studies that test targeted hypotheses about how event-level attributes interact with individual traits to influence wellbeing.

Our results can also be understood through the lens of stress appraisal theory⁶, which posits that an individual’s interpretation of a stressor—shaped by their personality and context—plays a central role in determining their psychological response. By capturing the valence, intimacy, anticipation, and disclosure of life events, our study provides a more detailed account of how appraisals and coping demands could shape short-term wellbeing outcomes. Our findings further complement and extend prior theoretical and empirical frameworks in psychology and social science. In particular, we position our work in relation to Potter et al.¹³, who emphasize the contextual variability of emotion regulation, and Fried¹⁴, who highlighted the complexity and comorbidity in mental health outcomes. Our event-centered approach also resonates with Kalisch et al.¹⁵’s positive appraisal style theory of resilience, by showing how individuals’ appraisals and contexts are associated with their immediate wellbeing following life events. Therefore, this work lays the groundwork for a contextually grounded framework of wellbeing, where both individual differences and situational appraisals jointly shape psychological responses to life events. Such a model challenges binary notions of stress versus resilience and instead advocates for a continuous spectrum influenced by life event attributes. Future theoretical models could benefit from incorporating these real-time interactions between person and context to better predict psychological adaptation. We list some of our key observations below:

Health-related life events are linked to poorer wellbeing outcomes. We found that health-related life events are strongly associated with worsened negative affect, higher stress, and poorer sleep. These findings support prior literature showing that health events can often act as stressors, triggering prolonged emotional and physiological disturbances². These findings are in line with stress appraisal frameworks and add granularity by specifying which dimensions of wellbeing are most impacted. Health-related events often carry uncertainty, perceived lack of control, and even existential threat—all of which may impair emotional regulation and increase cognitive load². Such repeated or intense stressors can produce compounding psychological tolls. Our results support and extend these frameworks by linking these events to both affective and sleep-related outcomes in the days immediately surrounding event disclosure. The uncertainty and loss of control inherent in health-related disruptions may also deplete coping resources, leading to heightened vulnerability³⁰.

Anticipation of life events is linked with positive affect but sleep disruption. Interestingly, we found that anticipation of events—whether positive or negative—was associated with higher positive affect and lower negative affect, yet also with worse sleep quality. This dual effect could highlight a tension between psychological and physiological responses. Although anticipating and planning for an event can create a sense of agency and emotional uplift, it can also produce cognitive arousal that interferes with sleep onset and maintenance^31,32. This dichotomy is theoretically grounded in dual-process models of affective forecasting and arousal³³, where the emotional benefits of anticipation can co-exist with physiological arousal that undermines restfulness. Anticipation may invoke hope, motivation, or preparedness, thereby elevating positive emotions, but it also engages cognitive rehearsal and vigilance, leading to sleep latency or fragmentation³². These results support theories that even positively valenced anticipatory states can activate the nervous system in ways that disrupt rest³¹.

Event valence strongly predicts affect, stress, and anxiety. As expected, we found that positive events were linked to increased positive affect and reduced negative affect, stress, and anxiety. This finding aligns with a large body of literature showing that positive life experiences buffer stress and improve mood regulation³⁴. In contrast, negative life events can be associated with higher stress and anxiety, as per the emotion theory’s focus on the adaptive value of affective appraisals³⁵. Our findings confirm these theoretical models and add real-world ecological evidence of their relevance in everyday settings.

Intimate events are associated with reduced positive affect. We found that highly intimate events—those deeply personal or sensitive in nature, such as a disease diagnosis—were associated with decreased positive affect. Unlike non-intimate events, which may be external or shared (e.g., experiencing a natural calamity in the community), intimate events likely demand internal coping and provoke existential or identity-related reflection. The lower positive affect may reflect emotional exhaustion or reduced capacity for joy when the event triggers core vulnerabilities, such as illness or loss of autonomy. These effects may also be compounded by the private nature of such experiences, which limits immediate social buffering³⁶. It is also possible that the intimacy of these experiences may diminish opportunities for social support or shared meaning-making, both of which are known to buffer emotional distress³⁷.

Longer-duration events could potentially increase emotional intensity and stress. Events that spanned over a period of time or had ongoing consequences were associated with both positive and negative affect, as well as higher stress and anxiety. This plausibly associates with the “perseverative cognition hypothesis,” which posits that prolonged cognitive engagement with stressors—through worry or rumination—sustains physiological activation and emotional turbulence³⁸. In contrast to discrete, one-time events, ongoing/continuous life disruptions may lack closure, creating continuous changes in emotional response. From a resilience standpoint¹⁵, extended events may erode adaptive capacities over time, increasing risk for sustained distress. Our study contributes empirical support for these theories.

Sharing life events on Facebook improves affect but worsens sleep. We found that life event disclosures on Facebook were associated with higher positive affect, but lower negative affect, stress, and anxiety—indicating an improved emotional wellbeing response. This is consistent with a body of research on the therapeutic effects of online self-disclosures^39,40,41,42. However, we also found that these disclosures are associated with worsened sleep quality. This mechanism may involve nighttime device use and emotional engagement online, which can delay bedtime or impair sleep hygiene⁴³.

Further, our study showcased the need for a more granular approach to identify and understand life event occurrences in wellbeing research and intervention design. Rather than counting the frequency of events alone, it is crucial to consider the type and context of each event—such as its emotional valence, relational intimacy, duration, and disclosure. Reflecting on our original motivation, this study advanced the understanding of how life events are associated with short-term wellbeing by uncovering nuanced associations between specific event attributes and immediate psychological outcomes. While prior work has established broad links between life events and long-term mental health outcomes such as depression or life satisfaction^11,12, our findings offer a more granular and immediate perspective. In doing so, we extend prior research on event-related affective dynamics¹³ and support calls for person- and context-specific models of mental health^14,15. These insights not only enrich theoretical frameworks of stress and emotion regulation but also carry practical implications for designing early-warning tools and adaptive interventions that can respond to the psychological contours of everyday life transitions. This perspective suggests new avenues for personalized wellbeing support. For instance, digital mental health tools can use event-specific data to tailor interventions—offering calming techniques for anticipated or intimate events, or encouraging social connection following negative experiences. Additionally, mental health practitioners can assess how individuals engage with their life events, both offline and online, to better understand the dynamics of mental health. We discuss practical and design recommendations from our study, while noting that these are not to be interpreted as definitive solutions but as conceptually plausible directions for tailoring mental health tools and support systems around people’s lived experiences. Our work provides a foundation for future confirmatory research and iterative co-design efforts with end users and clinical experts. We list some of these opportunities for future digital health and intervention design below:

Event-Aware Wellbeing Apps. Digital mental health platforms–especially those for stress management, mood tracking, or sleep–can be designed to dynamically adapt based on the nature of users’ recent or anticipated life events. For example, when a user logs or anticipates an intimate event (e.g., a medical test result), the app could prompt self-compassion exercises, encourage private journaling, or recommend grounding techniques tailored to identity-relevant stress. For anticipated events, the app might offer preparation modules, helping users manage arousal and avoid over-activation that could disrupt sleep quality. During long-duration life events, the app can offer periodic emotional check-ins, acknowledging the event’s ongoing emotional footprint and reducing the risk of emotional burnout.

Privacy-Aware Life Event Sharing. Given the finding that sharing life events on Facebook is linked to improved affect but worse sleep, designers of digital health systems should create dedicated spaces for controlled event sharing¹⁸. These platforms can encourage sharing during daytime hours to reduce late-night screen time and its impact on sleep, as well as offer a privacy-preserving design, where users receive emotional support without the pressure to overshare or engage constantly. With the advances in generative AI and large language models (LLMs), such systems can also integrate an AI-powered immediate response mechanism for support⁴⁴.

Integration with Passive Sensing and Wearables. Our research also encourages combining self-reported event logs with passive sensing data (e.g., sleep tracking, heart rate variability, phone usage) can help create early-warning systems that detect when a user may be struggling post-event, even if they do not overtly disclose distress. For example, a drop in sleep quality and a rise in late-night phone use after an intimate event may signal the need for a gentle check-in or suggested self-care activity.

Time-Sensitive Intervention Design. The temporally grounded nature of our findings suggests that digital health systems should consider not only what types of life events occur, but also when interventions are deployed in response to them. For instance, anticipatory stress may benefit from pre-emptive interventions that foster calming strategies before the event, while prolonged disruptions may require periodic emotional check-ins and sustained support. This approach aligns with Just-In-Time Adaptive Interventions (JITAIs)⁴⁵, which emphasize delivering support precisely when users need it based on contextual or temporal cues. Integrating life event characteristics into JITAI frameworks could enhance personalization and effectiveness. Additionally, prior research in digital phenotyping and affective computing^45,46,47 underscores the importance of time-sensitive interventions that adapt to users’ fluctuating emotional states. Our findings build on this foundation and suggest new ways to use event-level attributes for temporally informed wellbeing support.

Finally, although our study offers valuable insights, it has limitations, many of which also point to interesting future directions. While the study is longitudinal in data collection, our analyses primarily model wellbeing levels surrounding life event occurrences, using rolling-window averages. Therefore, our work should be understood as observational and descriptive rather than causal or mechanistic in nature. Further, this study is based on non-clinical and self-reported measures of wellbeing, where the use of self-reported measures may be influenced by retrospective recall biases or mood at the time of response. Therefore, this study lays important groundwork for future research that examines both short-term and long-term wellbeing impacts in clinical populations and with clinically validated measures. A further limitation concerns the selective nature of life event disclosures. Because our analyses are entirely based on voluntarily shared data—via self-reported surveys and social media posts—life events that participants chose not to disclose remain unobserved. This may introduce reporting bias and limit our ability to draw inferences about unshared or unarticulated experiences. In addition, while our study considered both anticipated and unanticipated events as well as the valence (positive or negative) of events as independent variables, our analyses did not explicitly differentiate between positive anticipation (e.g., looking forward to desirable events) and negative anticipation (e.g., dreading stressful events). This distinction could have unique implications for wellbeing outcomes and represents an interesting avenue for future work. Additionally, our sample may not generalize across cultures, age groups, or socioeconomic backgrounds. Furthermore, our analysis identifies associational patterns, but cannot confirm causality. Future research can integrate passive data sources, such as sleep trackers and phone usage logs, to supplement self-report and minimize bias. Experience sampling methods can capture event effects in real time and help establish causal pathways. It would also be valuable to explore moderators of these effects, such as personality traits, emotion regulation skills, or offline social support. Our analyses did not incorporate individuals’ prior exposure to similar event types (e.g., repeated health-related events), which could shape both short-term and long-term wellbeing responses. Accounting for cumulative event histories represents an important direction for future research. Finally, investigating how reframing or expressive writing around intimate or long-duration events impacts outcomes can inform therapeutic strategies as well as digital health interventions and platform design.

Methods

We sourced our data from the Tesserae project^16,48,49—a large-scale longitudinal project of studying wellbeing through multiple data modalities that recruited 754 participants. Participants were information workers employed in a variety of job positions and roles (e.g., engineers, consultants, and managers) at various organizations across the U.S. They were enrolled between January 2018 and July 2018, and were asked to remain in the study for either up to a year or until April 2019.

The Tesserae project was approved by the Institutional Review Board (IRB) at the researchers’ institutions. The participants were provided with informed-consent documents, and consent was sought separately for each data modality. They could seek clarification and opt out of any data collection. The data was de-identified and stored in secure databases and servers with limited access privileges.

Individual Differences

The participants provided self-reported data on demographics and psychological traits during enrollment in the Tesserae project. They responded to survey questions on demographics (e.g., age, gender, education, etc.), and psychological constructs: (1) Cognitive Ability assessed by the Shipley scales of Abstraction (fluid intelligence) and vocabulary (crystallized intelligence)⁵⁰, (2) Personality Traits, the big-five personality traits as assessed by the Big Five Inventory (BFI-2) scale⁵¹, and (3) Trait-level Wellbeing, the general positive and negative affect levels as assessed by the Positive And Negative Affect (PANAS-X) scale⁵², the anxiety level as measured via State-Trait Anxiety Inventory (STAI)⁵³, and the quality of sleep as measured via the Pittsburg Sleep Quality Index (PSQI)^54,55. Table 6 summarizes the distribution of self-reported data in our dataset, showing a well-balanced sample across demographics and individual differences. Our participant pool includes 288 male and 205 female participants, with ages ranging between 21 and 68 (mean = 36.57, stdev = 9.88).

Table 6 Descriptive statistics of self-reported demographics and individual differences of 493 participants considered in the study

Self-reported daily wellbeing measures

Participants also responded to daily and periodic surveys on wellbeing measures during their participation. These surveys consisted of short surveys on immediate (state-level) wellbeing, including (1) Positive and negative affect measured by PANAS-Short⁵⁶, (2) Stress as measured by a single-item omnibus question, “how do you rate your current level of stress?” on a scale of 1–5⁵⁷, (3) Anxiety as measured by a daily single item instrument⁵⁸ on a scale of 1–5, and (4) Sleep as measured by a single item MITRE scale measuring the total number of hours (0–24) of sleep in the previous day.

Psychometrics of self-reported surveys

We include the psychometrics of the self-reported surveys in our study. For trait-level attributes, (1) the BFI-2⁵¹ scale of personality traits had reliability estimates as follows: Extraversion (Cronbach’s α = 0.88), Agreeableness (α = 0.83), Conscientiousness (α = 0.86), Neuroticism (α = 0.90), and Openness (α = 0.84); (2) the Shipley-2’s abstraction subscale demonstrated internal consistency α ranging from 0.77 to 0.91 and test-retest reliability of 0.87, while the vocabulary subscale exhibited strong reliability with α≥=0.80⁵⁰; (3) the PANAS-X subscales for positive and negative affect exhibited strong reliability, with α values ranging from 0.84 to 0.90⁵²; (4) the STAI scale showed high internal consistency, with reported α values ranging between 0.86 and 0.95^53,59; and (5) the PSQI scale demonstrated reliability with an α of 0.83⁵⁵.

For state-level wellbeing outcomes, (1) the PANAS-Short scale demonstrated acceptable internal consistency, with α values ranging from 0.78 to 0.87⁵⁶; (2) State anxiety was assessed using a validated single-item measure from Davey et al.⁵⁸, which showed high concurrent validity with STAI (Spearman’s ρ=0.75); and (3) Stress was measured with a single-item self-report scale, and as such, internal consistency metrics are not applicable.

These reliability statistics support the robustness of the psychological constructs assessed and provide confidence in the interpretability of the study’s findings.

Social media data

The Tesserae project asked consented participants to authorize their social media data, particularly Facebook, unless they opted out or did not have an account^16,48,49. The enrollment briefing and consent process explicitly explained that they were expected to continue with their typical social media use. Participants granted access to social media data through an Open Authentication (OAuth) based data collection infrastructure developed in prior work⁴⁸. The OAuth protocol is an open standard, privacy-preserving, and convenient approach for data collection, enabling users to log in and grant third-party access to their data without sharing any personal credentials.

Among the 572 participants who provided access to Facebook data, 242 participants did not make any update between January 2018 and April 2019—the same period when the participants’ self-reported life event occurrences were also collected. The remaining 330 participants made 14,202 posts during this study period—the data used by Saha et al.¹⁸.

Life events data

This paper uses life events data from two sources—1) Self-reported life events survey: participants in the Tesserae project optionally responded to a life events survey at the end of the study. This survey was based on the Psychiatric Epidemiology Research Interview (PERI) life events scale¹⁷. Out of the initial total of 754 participants, 423 participants responded to these surveys with 1547 entries of life events during the study participation period (mean=3.86 events per individual). Figure 1 provides the distribution of these life events by types and by users. Detailed distributions of life events across participants, including the number and types of events disclosed, have been reported in prior work^18,19.

Fig. 1: Distribution of Life Events in the Dataset.

a Life Event Types. b Number of Life Events per User.

This paper uses the union of participants who provided Facebook data (n=330) and self-reported life events data (n=423)—which amounts to 493 participants. Importantly, we note that life event disclosures were entirely participant-driven, and as such, we only analyzed events that were voluntarily shared. Prior work¹⁸ revealed that participants could disclose life events in one modality (e.g., Facebook) but not another (e.g., survey), or vice versa, complicating inferences about unshared experiences. The disclosure of life events on either or both of Facebook and self-reported PERI-based life events survey is likely to be associated with demographic and individual differences as noted¹⁸. Saha et al.¹⁸ conducted an in-depth analysis of how individual differences are associated with disclosure convergence (reporting on both) and divergence (reporting on only one) of the two modalities of self-reported PERI survey and Facebook. In terms of convergence, the likelihood of reporting on both modalities decreased with age, and male participants were less likely to report on both modalities. High conscientiousness was associated with a lower likelihood of disclosing on both modalities, whereas high agreeableness was associated with a higher likelihood of reporting on both. Again, in terms of divergence, male participants are less likely to self-disclose on social media. Also, younger participants and those higher in extraversion were more likely to share life events on Facebook.

Saha et al.¹⁸ qualitatively coded the life event disclosures on 14,202 Facebook posts during the participation period and built a codebook of social media disclosures of life events. The coding was done analogous to the PERI life events scale, and each life event occurrence was labeled with the life event attributes. In particular, the authors formally defined a Facebook post to contain a life event, “if the post describes an event that is directly or indirectly associated with the individual or their close ones, such that it potentially leaves a psychological, physiological, or behavioral impact, or be significant enough to be remembered after a period.”¹⁸. Each of the life event attributes was annotated or labeled independently of one another. We list and describe the life event attributes of the life event occurrences (on the PERI survey and Facebook) below:

Event Type. The PERI scale categorizes life events into six broad types—School, Health, Personal, Financial, Work, and Local. In the PERI survey-based life events data, participants self-reported the type of events, whereas the Facebook data was manually coded based on reading through the posts using the same PERI survey categories¹⁸.

Anticipation. Anticipation is a binary characteristic of life events, indicating whether a life event could be reasonably expected by an individual (1=anticipated) or occurred unexpectedly (0=unanticipated). Anticipated events can be those that one can hope or worry about in the next six months⁶⁰, such as moving, childbirth, or starting a new job. Unanticipated events include sudden or unforeseen occurrences, such as accidents, unexpected layoffs, or medical emergencies. Life events in both the survey and Facebook were labeled with anticipation labels¹⁸.

Valence. Valence is the positivity or negativity of a life event. In the PERI survey, participants self-reported valence of life events on a 7-point scale of “Extremely Negative” to “Extremely Positive.” The VADER⁶¹ tool was used to identify the major sentiment of a life event disclosure across negative, neutral, and positive, and assigned this label as the valence for life event entries¹⁸. For comparability across datasets, these labels were re-scaled on a scale of -1 (negative) to +1 (positive).

Intimacy. Intimacy of a life event constitutes how intimate or how comfortably an individual can open up about a life event to personal, close, trusted others, and public circles of relationships³⁹. Life events were annotated with an intimacy label on a Likert scale of Low, Medium, and High.

Temporal Status. Life events can be grouped based on temporal status—either as continuous (spanning multiple days) or discrete (occurring at a single point in time) events. In the PERI life events survey, participants self-reported the temporal status of each event, and the social media disclosures were manually labeled¹⁸.

Scope. The scope of a life event consists of how much the event directly relates to an individual themselves, or their close ones, or more generic circles⁶². Scope of directness was labeled on a 3-point Likert scale where 1) Low scope events included generic events like bad weather, 2) Medium scope events associate with someone close and leave an indirect effect on the individual (e.g., spouse’s pregnancy, child going to school), and 3) High scope events are unique and direct on an individual (e.g., disease diagnosis)¹⁸.

Significance. Each life event is associated with a degree of significance in an individual’s life⁶³. In the PERI survey, participants self-reported significance (7-point scale of Lowest to Highest Significance). For Facebook disclosures, each event was labeled with a significance rating based on the PERI scale¹⁷. The significance ratings were separately standardized on a min-max scale of 0-1 to make the ratings comparable across the two datasets. We used this scaled score in our ensuing analyses.

Statistical power of participant pool

Although we do not claim absolute generalizability or representativeness of our participant pool, our dataset comprises a diverse sample of 493 participants from across the U.S. (Table 1). To assess the sufficiency of our sample size for significant interpretations, we conducted a power analysis following guidelines by Dattalo⁶⁴. Assuming a small-to-moderate effect size (f ²=0.15), a significance level of α=0.05, a desired power of 0.80, and a maximum of 25 predictors (including individual differences and life event attributes), the estimated required sample size is 172 participants. Our analytic sample of 493 participants exceeds this threshold, providing adequate statistical power to detect moderate effects. While detecting very small effects (f ² < 0.05) would require larger samples, our approach—leveraging a diverse set of individual differences and life event attributes, along with rolling-window-based wellbeing outcomes (described next)—enhances statistical sensitivity and supports the robustness of our findings.

Analytical Approach

We aimed to investigate the relationship between life event occurrences and individuals’ immediate wellbeing outcomes—for the 493 participants’ data considered in our study provided life event occurrences data. For wellbeing outcomes, we used daily self-reported wellbeing measures of positive and negative affect⁵⁶, stress⁶⁵, anxiety⁵⁸, and sleep quality⁶⁵ collected throughout the study period. To overcome data inconsistencies and sparsities, we employed a rolling-window-based moving average approach to enhance the reliability of wellbeing metrics. For each wellbeing measure, we computed a seven-day moving average, incorporating data from the three days preceding, the day of, and the three days following a given date within the study period. This approach allowed us to model the level of wellbeing outcomes surrounding life events, rather than moment-to-moment fluctuations or baseline-relative changes.

To ensure robustness in our measurement, while our primary analyses used a seven-day rolling window, we also sought to ensure that our findings were not driven by this specific choice of seven-day aggregation. Therefore, to assess robustness, we repeated the ensuing analyses with alternative time-windows—three-day and five-day. A consistency in directionality of results across different time windows would indicate that our findings are robust to the choice of temporal aggregation.

We studied to what extent these measures could be explained by individual differences and life event attributes. Therefore, we built several linear regression models (with and without L1/L2 regularization), with individual differences and life event attributes as independent variables and wellbeing outcomes as dependent variables. In addition, the Facebook disclosure of life events was included as an independent variable, and we also included a binary indicator variable denoting whether each participant shared their Facebook timeline data. In particular, we built three kinds of regression models—(1) M₁ models consisted of using only individual differences as independent variables, (2) M₂ models consisted of only life event attributes as independent variables, and (3) M₃ models consisted of both individual differences and life event attributes as independent variables. While it may seem confusing to see positive and negative affect as both independent and dependent variables, we clarify that the regression models used trait-level affect as independent variables. These traits, measured at Tesserae’s entry-point using the PANAS-X scale⁵², represent baseline affect levels. In contrast, the dependent variables were state-level affect, self-reported periodically using the PANAS-Short scale⁵⁶. These state-level measures reflect short-term wellbeing, which our models aimed to predict.

To further clarify about our modeling, although the data used in our analysis was longitudinal—comprising daily self-reported wellbeing and timestamped life events over a year—the regression models were built in a cross-sectional structure. Each row in the model corresponds to a single life event occurrence. For each event, we extracted event-level attributes (e.g., anticipation, valence, intimacy), individual differences (e.g., age, gender, personality traits), and computed wellbeing metrics using a seven-day rolling window centered on the event date. This setup enabled us to examine how individual differences and life event attributes are associated with wellbeing levels around the time of each life event. A mixed-effects linear regression framework offered a more rigorous approach to modeling the nested structure of the data, with life events nested within individuals.

To evaluate the fit of the models, we obtained the goodness-of-fit (R²). We also conducted a k-fold (k = 5) cross-validation and pooled all the observed and predicted values to obtain the symmetric mean absolute percentage error (SMAPE) and Pearson’s correlation (r) to evaluate the performance of the models. Here, SMAPE calculates the percentage of relative errors and is bounded between 0 and 100⁶⁶, and lower values indicate a better predictive performance. In contrast, higher Pearson’s r indicates better predictive performance. Further, we conduct Analysis of Variance (ANOVA) tests to compare each of the pairs of models. ANOVA helps to determine the statistical significance of differences between two models by comparing the change in their sum of squared errors⁶⁷.