Dynamic network models reveal personalized patterns of well-being in young adult daily lives Authors: Saida Heshmati, Jaymes Paolo Rombaoa, Sean H. Merritt, Alexander P. Christensen Published: 2025 ABSTRACT Despite increased focus on well-being in the psychological sciences in recent years, cross-sectional and group-level methods fall short of capturing the dynamics of individual people’s well-being in daily life. Such an investigation is especially important for young adults who are in a developmental phase in which they are more likely to experience fluctuations in their daily lives. We used ecological momentary assessment over one week in two samples of first-year college students (Sample 1, N = 103, assessments = 2,535; Sample 2, N = 76, assessments = 1,796) and a dynamic network approach (Dynamic Exploratory Graph Analysis). This approach estimates, within each person, how well-being elements fluctuate across days and co-vary with one another, recovering both their differentiation (exclusivity) and their interdependence (connectivity). In this approach, elements of well-being—measured via the PERMA framework—are modeled as a dynamic network to address the following research questions: (1) How do elements of well-being in young adults’ daily life form a dynamic network? (2) Does the structure of well-being hold across all young adults, or is there heterogeneity? (3) Is there synchrony among well-being elements, and do they drive a person’s well-being network? Findings suggest that whereas group-level dynamic well-being networks align with theoretical models, each person’s individual experiences vary significantly, with each person demonstrating unique well-being network structures and synchronous elements. These findings underscore the importance of considering individual variability in well-being, highlighting the necessity for personalized approaches to understand and promote well-being among young adults.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-24897-6. Introduction In recent decades, psychological research has increasingly recognized that well-being is not a singular or static trait but a complex, multifaceted experience shaped by an individual’s goals, social relationships, and contexts1–4. This shift has moved the field toward understanding well-being as a dynamic and multidimensional personalized process that unfolds within individuals over time5–8. Yet, despite this theoretical progress, much of the empirical research continues to rely on static, group-level approaches that may obscure important within-person variation. Understanding how well-being is structured and experienced in everyday life—within individuals, across time, and in response to the changing demands of development—is fundamental to advance theory and improve intervention. Well-being as a multidimensional and dynamic construct The psychological study of well-being has long acknowledged its multidimensional nature from various perspectives and conceptualizations9,10. Diener’s foundational model of subjective well-being11 distinguished between affective well-being—the presence of pleasant emotions and the absence of unpleasant ones—and cognitive well-being, defined as overall life satisfaction and satisfaction within specific domains (e.g., work, family, health). These components differ not only in content but also in their temporal stability: affective well-being is more reactive to short-term changes, whereas cognitive well-being tends to be more stable, though still susceptible to fluctuation over time12,13. Complementing these nuances, Ryff’s model of psychological well-being14 offers a more eudaimonic perspective, positing six core dimensions: autonomy, environmental mastery, personal growth, positive relations with others, purpose in life, and self-acceptance. These dimensions emphasize optimal psychological functioning and the realization of human potential, rather than just the presence of pleasure or satisfaction. Ryff’s model also highlights developmental change, suggesting that certain dimensions (e.g., personal growth or purpose in life) may fluctuate in salience and expression depending on life stage and daily life events or emotions15,16. These dynamics broaden the definition of well-being to include both the pursuit of meaning and the ongoing adaptation to life’s demands. Together, these foundational theories laid the groundwork for understanding well-being as a complex, evolving construct, shaped by both enduring traits and momentary experiences, and spanning emotional, evaluative, and psychological dimensions. Beyond structure, daily-life studies show that well-being domains fluctuate meaningfully within individuals. Positive and negative affect show pronounced within-person variation as measured through various state affect measurement tools17–19, and their within-person factor structure may differ from the between-person structure20. Beyond affect, evaluative well-being (life satisfaction) also varies on short time scales, with stable individual differences in how people update their cognitive evaluations of life satisfaction13. Eudaimonic components fluctuate as well: daily purpose shows reliable within-person variability16, and engaging in daily eudaimonic activities covaries with next-day well-being21. Experiences of love in everyday life—central to the Relationships facet—also display robust within-person dynamics and are linked to psychological well-being and personality22. Taken together, this evidence suggests that short-term variability and person-specific co-variations among the elements are characteristic of well-being processes, motivating within-person, multidimensional approaches to complement traditional between-person models. Building on these frameworks, the PERMA model, a multidimensional framework of well-being introduced by Seligman23, explains well-being through five core components —Positive emotions, Engagement, Relationships, Meaning, and Accomplishment—that span both hedonic and eudaimonic traditions. It’s argued that whereas these elements are not an exhaustive list of factors that make up a person’s experience of well-being in life (Seligman, 2018), they (1) “can be defined and measured independently of other elements while being interconnected (connectivity),” and (2) “people pursue each element for its own sake and not just to serve another element (exclusivity)”24 (p. 2). This emphasis on both exclusivity and connectivity aligns with systems-based approaches that argue for well-being as a network of interacting elements rather than a unidimensional score25,26. Since its introduction, PERMA has been the focus of a growing body of research examining its structure and utility26–28. Some scholars have questioned whether PERMA offers added value beyond existing models such as Subjective Well-Being (SWB); for instance, Goodman and colleagues found that PERMA and SWB were highly correlated, raising concerns about redundancy29. In response, Seligman24 clarified that PERMA was not intended to measure a different kind of well-being, but rather to identify constituent elements that contribute to it. He argued that PERMA’s utility lies in its ability to disentangle the components of well-being, making it especially useful in applied and intervention contexts. Seligman also proposed six criteria for evaluating the model’s elements, including that they be independently definable and measurable. Although critiques have raised concerns about potential halo effects and non-orthogonality among the elements, these issues highlight the need for modeling approaches that can detect underlying causal associations or shared variance across components. Recent network psychometric evidence supports this view. Using exploratory graph analysis (EGA30,31 in a representative U.S. adult sample, Merritt et al.32 found that the five PERMA elements (plus health and negative-emotion items) formed separable yet interconnected dimensions that were largely distinct from life-satisfaction and positive/negative-affect measures of Subjective Well-Being; only the emotion items overlapped. Importantly, even when life-satisfaction and positive/negative-affect items were added to the network, the five PERMA communities remained intact and the centrality hierarchy—for instance, Positive Emotion and Relationships as the most influential hubs—did not change32. The network model showed excellent fit indices and illustrated Seligman’s “exclusivity” and “connectivity” criteria more clearly than traditional factor analysis, leading the authors to describe PERMA as a “web of well-being.” Their results indicate that PERMA captures nuanced, non-redundant patterns of interrelation among antecedents of well-being even after controlling for shared variance with SWB, and they underscore the value of network approaches for isolating halo effects and pinpointing leverage points for intervention. EMA-adapted PERMA measures likewise capture robust intra-individual variability across all five elements and link those dynamics to global well-being33. Together, these findings provide the empirical rationale for our questions: testing whether PERMA elements remain differentiated in daily life (exclusivity), how strongly they covary within persons (connectivity), and to what extent those within-person relations are shared versus person-specific during the college transition. Well-being in young adults and the college transition Longitudinal studies show that major life transitions—whether normative (marriage, parenthood, retirement) or non-normative (unemployment, disability)—are typically followed by short-term changes in well-being and gradual hedonic adaptation, though the magnitude and duration of these effects vary by domain and person34,35. The move from high-school to university is a similarly salient transition: it combines changes in residence, social networks, and role demands, and thus is expected to produce detectable, if heterogenous, shifts in multiple well-being domains. This multidimensional approach to well-being is particularly informative in the context of emerging and young adulthood, especially those transitioning to college. Although it is hypothesized that well-being is dynamic across the lifespan, early college years are marked by ongoing development in self-regulation, identity formation, and social adaptation36,37. These processes introduce daily fluctuations in emotions, purpose, social connectedness, and achievement—core components of both hedonic and eudaimonic well-being. Rather than implying that instability uniquely necessitates dynamic modeling, we argue that this life stage offers an ideal context to observe well-being as it unfolds in real time. First-year college students are more likely to face irregularities and fluctuations in their day-to-day life, participating in risk-taking behaviors, frequent social gatherings, and often irregular work schedules which can lead to greater within-person variability in well-being38,39. At the same time, the first college year offers abundant opportunities for mastery and intellectual growth—challenging coursework, research and internship experiences, and new leadership roles—which can enhance Engagement and Accomplishment. Interactions with expanded peer networks, greater autonomy, and exploration of personal values frequently bolster Positive Emotion, Relationships, and Meaning40–42. We do not assume that well-being is more dynamic in this life stage compared to others, but we suggest that developmental and contextual features of this period may make those dynamics more observable and differentiated. Whereas many developmental changes unfold over months or years, fluctuations in well-being-related experiences among college students—such as stress, mood, motivation, sense of purpose, and connection—can occur over much shorter timescales. These fluctuations are often driven by shifting academic demands (e.g., exams, deadlines), transitions in social roles and relationships (e.g., changing friend groups, forming or ending romantic partnerships), and variable engagement in leisure and work activities. These contextual changes, in turn, can lead to observable variation in health-related behaviors and psychological states. For instance, studies have found that college students’ sleep quality and quantity vary significantly across days and weeks, often in relation to academic pressures and social engagement43. Other research has found that daily stressors and positive events have an influence on college students’ daily feelings of purpose15.Eating and physical activity behaviors also shift across the semester depending on schedule demands and living arrangements44. Daily leisure activities—including both serious and casual leisure—are highly variable and intertwined with students’ social and emotional regulation strategies38. These fluctuations are embedded in a broader period of life characterized by dense transitions in social roles and life structure45, suggesting that well-being processes during this phase are temporally dynamic. Importantly, these contextual fluctuations may differentially impact specific domains of well-being. For instance, a socially active weekend may enhance positive emotion and perceived connection without necessarily improving a sense of purpose or accomplishment, whereas an academic failure may affect meaning and mastery but leave relationships intact. In fact, network research conducted in this age group has found that certain domains of well-being (e.g., Positive emotions and Relationships) were more strongly connected to other domains of the well-being network in young adults’ daily life, whereas others (e.g., Meaning) were more weakly connected or idiosyncratic46. Such divergence highlights that well-being components do not always rise and fall together—and may even counteract one another at times. As such, the different domains offer heterogeneity in the potential representation of well-being in each person. Thus, this developmental window provides a theoretically and empirically supported context in which to observe meaningful short-term variation (e.g., momentary, daily, weekly) in well-being and its contributing factors. A multidimensional approach is therefore essential not only for distinguishing which well-being domains are impacted by specific experiences, but also for examining how these domains co-vary—or decouple—within individuals across time. A dynamic modeling approach allows us to capture these intraindividual fluctuations and the evolving interplay among well-being components. Further, emerging adulthood is a period of heightened individual differentiation: people construct meaning and satisfaction in increasingly idiosyncratic ways. Idiographic, dynamic models are thus particularly well-suited to study how each individual organizes their well-being system over time. Although future research should examine whether similar dynamics occur in other life stages, our current focus is on a population where both developmental theory and contextual variability suggest strong relevance for a dynamic, person-specific approach. A dynamic systems approach to well-being In parallel, the science of well-being is shifting toward a dynamical systems perspective, moving beyond theoretical frameworks that view well-being as a fixed trait46. Dynamical systems are usually defined as systems or a set of interconnected components (e.g., observable variables) that are dynamic in that the components change across time. This shift challenges the concept of an “ideal,” and unchanging state of well-being and mental health as a stable endpoint6,47. A dynamical systems approach emphasizes the fluid and evolving nature of well-being, highlighting how it unfolds through ongoing interactions that cultivates stable traits over time. Further, it enables an examination of how well-being changes over short and long time-scales, the variability among individuals, and their impact on a person’s overall journey toward well-being46. There is an emerging consensus on the necessity of idiographic, person-centered approaches in psychological science7,48. The prevailing use of nomothetic approaches in psychological research, which seek to establish general laws and commonalities across individuals, are insufficient to explain individual-level experiences. These methods presuppose that psychological traits and dynamics, such as those associated with well-being, are consistent across people (i.e., ergodicity49,50 with a growing body of research highlighting the limitations of this assumption51. For instance, although group-level analyses might identify common dimensions of personality across a population, individual-level analysis reveals a much broader spectrum of personal characteristics that vary not only in number but also in kind52,53. This distinction is crucial in the study of well-being in young adults, where heterogeneity in personal significance and the subjective nature of well-being processes play pivotal roles31. Idiographic approaches allow researchers to delve into the unique configurations of well-being elements for each individual, acknowledging that the interrelations among these elements can differ dramatically from person to person. Dynamic network models are one method that can investigate how the components and the whole of well-being relate across time. In a network, well-being is represented as a system of interconnected ‘nodes’ (circles), which represent the items of the PERMA survey, and ‘edges’ (lines), which represent conditional associations (i.e., partial correlations) between the nodes. In a dynamic network, the edges represent how two components change together across time given all other components. Nodes that consistently change together can form densely connected “clusters” known as ‘communities,’ which are consistent with factors of a factor model (e.g., Meaning construct in PERMA; when the data are generated from a latent variable model54. This representation is afforded by Dynamic Exploratory Graph Analysis55 (DynEGA). DynEGA estimates the first-order derivatives of each variable’s time series for each person using the Generalized Local Linear Approximation (GLLA) method54 to capture their rate of change (i.e., how quickly a variable increases and decreases) across a time window of consecutive measurement points (e.g., five). Afterwards, networks and communities are estimated from these derivatives for each person and the sample. Specifically, sample-level networks represent the structure of well-being domains across people (i.e., the aggregate), where the derivatives of each person are “stacked” on top of each other to represent a single data matrix. Conversely, individual-level networks represent the idiographic networks for each person and only their derivatives are used for network estimation. Similarities and differences between the individual-level and sample-level as well as pairwise individual-level structures can be examined at various degrees. At the broadest structure, networks can be compared to examine how similar individuals are to each other and the sample. At the intermediate structure, communities can be compared to determine whether dimensions are similar (even if the networks appear different). At the lowest structure, centrality measures can be compared to determine how similar nodes are connected within the networks. Taken together, these comparisons provide evidence for whether the dynamics of well-being are homogeneous across the sample. The present research In the present investigation, we aimed to explore well-being in first-year college students as a dynamic network, characterized by temporal momentary changes and individual variability. We conduct a two-study investigation using ecological momentary assessment (EMA56 to examine well-being in first-year college students through the lens of the PERMA framework. This period of young adulthood, especially the transition to college, is marked by significant fluctuations in daily routines and social activities, making it a critical phase for examining well-being. Our central aim is to test—during this salient life transition—whether the five PERMA elements display (a) exclusivity, clustering into their expected communities within individuals, and (b) connectivity, forming a coherent, interacting network that shifts from day to day. The start of university offers an ideal context for this test: new academic challenges can boost Engagement and Accomplishment, expanding peer networks can strengthen Positive Emotion and Relationships, and evolving values may reshape Meaning—yet the same demands can also strain these domains. This approach aligns with the shift in psychological sciences toward more nuanced and process-oriented understandings of mental health and well-being52,57, recognizing that static and nomothetic models do not adequately capture the complexity of human experiences58. Accordingly, our investigation is organized around three concept-driven questions that translate PERMA theory into the lived context of the first university semester. First, we ask whether students’ moment-to-moment reports of well-being elements still resolve into the five distinct PERMA dimensions when they are grappling with new academic, social, and autonomy demands. If the classical structure emerges at both the sample and individual levels, it would suggest that Seligman’s building blocks remain psychologically separable even in a turbulent developmental window; on the other hand, if, for example, Engagement and Accomplishment fuse into a single academic-mastery cluster, that would indicate a transition-specific blending of experience. Having established the dimensional map, we next examine the architecture of the resulting networks. Specifically, we ask which PERMA elements act as the most central hubs—those with the greatest node-strength—in students’ day-to-day well-being networks. Theory and prior college-transition findings26 lead us to predict that Positive Emotion and Relationships will emerge as high-influence hubs, mirroring the social-affective salience of the first semester; Engagement is expected to occupy a mid-level position, whereas Meaning and Accomplishment may function as weaker or more variable hubs. By ranking node strengths in every individual network and comparing those ranks with the pooled, sample-level pattern, we can see whether first-year students share a common set of organizing “anchors” or whether the center of gravity in the well-being web differs from one student to the next. ​​Finally, we probe the temporal synchrony of the elements: which pairs rise and fall together across days, and are those couplings shared across individuals? We expect a widespread synchrony between Positive Emotion and Relationships, reflecting day-to-day shifts in social life being strongly tied to positive emotional experiences, whereas links such as Meaning and Accomplishment may be synchronous for some students but not others, signaling personalized routes to thriving—or strain—during the transition. Together, these questions allow us to evaluate PERMA’s exclusivity and connectivity propositions in a way that is conceptually anchored in the developmental realities of emerging adulthood. Method Participants and procedures The first sample comprised 103 first-year college students enrolled at undergraduate colleges in Southern California aged 18 to 20 years old (M = 18.41, SD = 0.57). Participants in this sample were from a larger multi-week intervention study across multiple semester cohorts (e.g., Fall 2021, Spring 2021), but the data presented here is based on the pre-intervention period. First-year students were recruited via flyers posted around the undergraduate college campuses and social media to participate in a longitudinal study of well-being in daily life. Then, recruits completed online or in-person interviews with a research assistant. Informed consent was obtained from all subjects before starting the study. Briefly, the intervention study had participants complete one week of EMAs, followed by four weeks of two types of mindfulness-based training conditions or an active control condition, and a final week of EMAs post-intervention. Participants completed EMAs at semi-random time points within four blocks (9:00–11:00 a.m., 12:00–2:00 p.m., 3:00–5:00 p.m., 6:00–9:00 p.m.) with no two surveys administered within one hour of each other. We selected four daily prompts to balance the need for capturing meaningful within-day fluctuations—guided by foundational work on EMA and Experience Sampling Methods59,60—and to cover key periods of academic activity, social interaction, and evening reflection, when well-being components are known to vary15,38,43. The EMAs were signaled four times each day for 1 week. Out of a total of 2,884 possible EMAs (103 people \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\times\:$$\end{document} 7 days \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\times\:$$\end{document} 4 signals), this sample had 2,535 available EMAs to analyze (sample compliance rate = 88%), with participants completing an average of 24 (SD = 3) out of the 28 signaled surveys (individual compliance rate = 86%). The compliance rate for this sample was acceptable (above 80%), according to EMA guidelines by Stone and Shiffman57. The second sample comprised 76 first-year college students (age: M = 18.41 years, SD = 0.61) during the weeks following the first state-mandated university lockdowns in California due to the COVID-19 pandemic in Spring 2020. First-year students were recruited in a manner similar to Study 1 and underwent pre-study interviews with a research assistant. Recruits were informed about the study (one-week daily life study, no intervention) and provided informed consent before starting this study. These students had measures identical to Study 1 but were not involved in a larger intervention study with multiple weeks of data collection. We used the one week of EMA data available from Study 2 to replicate Study 1. Out of a total possible 2,128 EMAs (76 people \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\times\:$$\end{document} 7 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\times\:$$\end{document} 4), the second sample had 1,796 EMAs to analyze (sample compliance rate = 84%), with participants completing an average of 23 (SD = 3) out of 28 signaled surveys (individual compliance rate = 82%). Lastly, components of the data in Study 2 have been used elsewhere61 but have not been analyzed using the current methods. Participants completed a battery of trait and EMA measures. Participants were paid proportionally to their survey response rates, with a maximum payment of $30 for trait measures and one week of EMAs. Demographic information for both samples is provided in Table 1. Data for both samples were part of a research study approved by the Institutional Review Board at Claremont Graduate University (CGU IRB #3661). All research was performed in accordance with the Declaration of Helsinki. Table 1Demographic characteristics of two samples of first-year students.DemographicsSample 1(N = 103)Sample 2(N = 76)Gender % Female8571 % Male1124 % Non-binary34 % Prefer not to say11Race/ethnicity % Asian or Pacific Islander4436 % Black, African American813 % Hispanic, Latino1414 % Native American, Eskimo, Aleut11 % Other213 % White, Caucasian4646The total number of ecological momentary assessments was 2,535 for Sample 1 and 1,796 for Sample 2. Participants could “check all that apply” for Race/Ethnicity. Table 1Demographic characteristics of two samples of first-year students.DemographicsSample 1(N = 103)Sample 2(N = 76)Gender % Female8571 % Male1124 % Non-binary34 % Prefer not to say11Race/ethnicity % Asian or Pacific Islander4436 % Black, African American813 % Hispanic, Latino1414 % Native American, Eskimo, Aleut11 % Other213 % White, Caucasian4646The total number of ecological momentary assessments was 2,535 for Sample 1 and 1,796 for Sample 2. Participants could “check all that apply” for Race/Ethnicity. Demographic characteristics of two samples of first-year students. The total number of ecological momentary assessments was 2,535 for Sample 1 and 1,796 for Sample 2. Participants could “check all that apply” for Race/Ethnicity. Measure mPERMA Momentary well-being was assessed using Momentary PERMA (mPERMA33, the adapted version of the PERMA-Profiler27 that is validated for intensive longitudinal designs and measuring elements of well-being in daily life. Like the PERMA-Profiler, mPERMA measures various components of well-being23,24. To emphasize that participants were answering questions about their momentary well-being, participants were prompted with the phrase “in this current moment,” and phrases such as “in general” were removed. There were three items based on each block of PERMA for a total of 15 items: Positive emotions (e.g., “I am feeling joyful”), Engagement (e.g., “When I noticed this survey, I was absorbed in what I was doing”), Relationships (e.g., “I feel helped and supported by others”), Meaning (e.g., “I have a sense of direction in my life”), and Accomplishment (e.g., “I am making progress towards accomplishing my goals’’). Each item was presented in a randomized order on participants’ smartphones to reduce biased responses due to order effects. For each item, participants used a slider scale that ranged from 0 (not at all) to 100 (completely). Data analysis plan Network and dimension Estimation The networks and dimensions at the sample- and individual-level were estimated following the DynEGA procedure31. First, to estimate networks for each sample and individual within their respective samples, the first-order derivatives for each variable within each person were computed using the GLLA55. This resulted in a n by 15 matrix for each person where the rows are equal to the n first-order derivatives and the columns represent each PERMA item. Next, the individual-level networks were estimated on each person’s first-order derivative n by 15 matrix. The sample-level networks were estimated by “stacking” these derivative matrices such that the rows represented each participant’s first-order derivatives consecutively and the columns represented each PERMA item. In all cases, the graphical least absolute shrinkage and selection operator (GLASSO62 with extended Bayesian information criterion (EBIC63 for model selection was used to estimate each network. The resulting network in each case represents the PERMA items as nodes and the partial correlations of the first-order derivatives as edges. On each network, the lower-order Louvain algorithm64 was applied. This algorithm begins with each node as its own community (dimension) and proceeds by merging nodes one-by-one into an existing community that maximizes its gain in modularity65. Modularity is a statistic that quantifies the extent to which nodes are more densely connected within their community relative to their connections to other communities. This algorithm automatically determines the number and content of the communities. The result is a so-called “community solution” where each node belongs to one community (dimension). For more technical details of this section, see SI2. Technical Data Analysis Plan in the Supplemental Information: https://osf.io/95qvr?view_only=f64237b42caf48f39288b8bb83072116. Network, dimension, and node similarity To evaluate the similarity of sample-to-individual level and pairwise individual-level networks, we used an information-theoretic metric called Jensen-Shannon Distance66 (JSD). JSD quantifies how different two networks are based on their topology (i.e., the structure and organization of the network connectivity). The more similar the two networks are, the smaller the distance67, ranging from 0 (identical) to 1 (completely different). This metric was converted to a similarity by computing 1 – JSD (Jensen-Shannon Similarity; JSS). To our knowledge, there are no benchmarks for what is considered “high” similarity for JSS. To provide rough estimates of what might be expected as “high,” a small-scale simulation study was conducted using each sample-level network. This simulation had four conditions: (1) keeping the edge weights the same but randomizing the structure, (2) keeping the structure the same but randomizing the edge weights, (3) keeping the structure and edge weights the same, but incrementally increasing the proportion of edge weights that are swapped with one another, and (4) keeping the structure and edge weights the same, but incrementally increasing the proportion of edges that are dropped. These conditions allowed us to determine the extent to which the structure and edge weights affect JSS. Based on our results, we provide tentative benchmarks of 0.90 and above as “high” similarity, 0.80 and above as “moderate” similarity, and 0.70 and above as “low” similarity. Values below 0.70 were consistently associated with differences related to random structures and edge weights suggesting substantial differences. Technical details for this small-scale simulation can be found in SI 3. Jensen-Shannon Similarity Benchmark Simulation of the Supplemental Information. Dimension similarity was computed using the Adjusted Rand Index (ARI68. The ARI quantifies the similarity between two community solutions or how the PERMA elements are sorted into different communities based on the Louvain algorithm by counting the matching pairs and accounting for the expected number of matching pairs by chance. Values can range from − 1 (completely different) to 1 (exactly the same) with zero representing similarity expected by random chance. ARI was computed between the sample- and individual-level dimensions to determine the similarity of the individual-level networks to the sample-level dimensions. A permutation test was performed to determine whether the ARI value was significantly different from zero69. Centrality similarity used node strength (absolute sum of a node’s connections to other nodes), which was compared between sample-to-individual level and pairwise individual-level networks. Node strength is commonly used in network psychometrics and is often referenced as “centralness” in terms of rank order70. Consistent with this usage, comparisons used Spearman’s correlation to determine the similarity of centrality. For more technical details of this section, see SI2. Technical Data Analysis Plan in the Supplemental Information: https://osf.io/95qvr?view_only=f64237b42caf48f39288b8bb83072116. Pairwise node synchrony Definitions of synchronicity vary. We define synchrony as “significant overlap” in the rate at which two (or more) variables change together across time71. To quantify significant overlap, Unique Variable Analysis72 (UVA) was used. UVA uses a graph theoretic measure called weighted topological overlap73 to quantify the extent to which the connections of two nodes are similar (i.e., the frequency and strength with which two nodes are also connected to other nodes as well as each other). Values ≥ 0.25 are considered statistically redundant as to be capturing “synchronous” information. Synchronous variable pairs were counted across people representing the frequency of individuals with those synchronous pairs. For more technical details of this section, see SI2. Technical Data Analysis Plan in the Supplemental Information: https://osf.io/95qvr?view_only=f64237b42caf48f39288b8bb83072116. Analytic considerations Number of observations per person The dynamic network modeling approach used in this study is robust across simulations with sample sizes of 50 with fifty observations per person54,74. In this study, the sample size is larger than the minimum size of previous simulation investigations suggesting that the sample-level results are well powered. To our knowledge, the number of observations per person has not been investigated using DynEGA with fewer than 50, which leaves open the question of how robust the individual-level results are in the current study. In comparison with recently published empirical studies using other dynamic network modeling methods, the average number of observations (24 and 23, respectively) in this study are consistent with previous research5,75. Missing data Although missing data was minimal, even a few missing observations in key places could obstruct achieving reliable results. The main issue of missing data in the present approach is in the estimation of the derivatives. In Boker and colleagues76, bias from missing values and time interval misspecification were evaluated in the time delay embedding (i.e., consecutive windows of time used to estimate the derivatives) and found that “doing nothing” (i.e., ignoring the missing data) almost always performed as well as the best correction method of row-wise loading correction using full-information maximum likelihood. From their results, the primary concern of missingness, where the bias of missing data is not completely mitigated, occurs at the beginning or ending time series. In the first sample, there were two people who had missing data with one person who had one response missing for one item on the very last measurement occasion and the other person who had missing data for 2 items on one measurement occasion at the end. There was one person who had missing data for 4 items on one measurement occasion at the end. In the second sample, there were no missing data. In sum, there was low missingness overall as well in cases where this method would be most affected. Rating scale A continuous slider rating scale has come under scrutiny in recent years due to its challenging interpretation for participants, which can have unstable reference points and lead to inflated variability77,78. In a recent simulation study, Zhang and colleagues79 found the choice between Likert and slider scales to depend on the study’s aims. For dimensionality and psychometric properties, similar conclusions can be reached for Likert scales with five or more categories and slider scales. One study using the current analytic approach found that continuous slider scales provided results that were more consistent with theoretical structures than Likert scales80. Nevertheless, more work is needed to evaluate the consequences of response scales in intensive longitudinal data and modeling. Openness and transparency This study was not preregistered. The materials (including survey items, data, and code) used in this study are available at Open Science Framework (OSF). All analyses were conducted using R (version 4.4.4). DynEGA was applied using the EGAnet package (version 2.0.7) and associated results were visualized using the ggpubr (version 0.6.0), GGally (version 2.2.1), and ggplot2 (version 3.5.1) packages in R. Data used for this study, analysis scripts, and results are available on OSF (https://osf.io/qvws2/?view_only=f64237b42caf48f39288b8bb83072116). Descriptive statistics for the intraclass correlation coefficients and the between- versus within-person zero-order correlations are available in Tables S1–S4 within Section SI1. Variability of PERMA Items in Between- and Within-Person Dynamics, in the Supplemental Information: https://osf.io/95qvr?view_only=f64237b42caf48f39288b8bb83072116. Results Dynamic PERMA dimensions For Sample 1 and 2, the five theoretical PERMA dimensions were identified in the sample-level networks (Fig. 1). At the individual-level, the median number of dimensions was five for both samples (Table 2). On the one hand, the individual-level dimensions appear to be consistent with the sample-level dimensions given the majority (more than 50%) of individuals had 5 dimensions in both samples. On the other hand, the item content of each individual’s dimensions were unique such that no one person had the exact same structure as the theoretical PERMA dimensions. Whether these results indicate homogeneity or heterogeneity is explored next. Fig. 1Sample-level dimensions for Sample 1 and Sample 2. Nodes (circles) represent mPERMA variables, blue edges (lines) represent positive regularized partial correlations with the thickness indicating magnitude, and the color of the nodes denotes the community (dimension) they belong to: Positive (P; red), Engagement (E; blue), Relationships (R; green), Meaning (M; orange), and Achievement (A; yellow). Sample-level dimensions for Sample 1 and Sample 2. Nodes (circles) represent mPERMA variables, blue edges (lines) represent positive regularized partial correlations with the thickness indicating magnitude, and the color of the nodes denotes the community (dimension) they belong to: Positive (P; red), Engagement (E; blue), Relationships (R; green), Meaning (M; orange), and Achievement (A; yellow). Table 2Frequency and proportions of dimensions in the individual-level networks.Number of dimensionsSample 1n (%)Sample 2n (%)One0 (0.0%)4 (5.3%)Two0 (0.0%)0 (0.0%)Three1 (1.0%)1 (1.3%)Four20 (19.4%)15 (19.7%)Five59 (57.3%)41 (53.9%)Six20 (19.4%)14 (18.4%)Seven3 (2.9%)1 (1.3%)Displays the number of network dimensions identified at the individual level for both samples. Table 2Frequency and proportions of dimensions in the individual-level networks.Number of dimensionsSample 1n (%)Sample 2n (%)One0 (0.0%)4 (5.3%)Two0 (0.0%)0 (0.0%)Three1 (1.0%)1 (1.3%)Four20 (19.4%)15 (19.7%)Five59 (57.3%)41 (53.9%)Six20 (19.4%)14 (18.4%)Seven3 (2.9%)1 (1.3%)Displays the number of network dimensions identified at the individual level for both samples. Frequency and proportions of dimensions in the individual-level networks. Displays the number of network dimensions identified at the individual level for both samples. Similarities and differences Network similarity As a point of reference, the two sample-level networks were highly similar with a JSS of 0.914 (Fig. 1). The sample-to-individual level comparison (left in Fig. 2) demonstrated lower similarity, on average, and substantial variability (Sample 1: M = 0.751, SD = 0.089, range = 0.338–0.965; Sample 2: M = 0.762, SD = 0.085, range = 0.541–0.968). On average, the individual networks had low similarity with the sample-level networks. Although the similarities to their respective sample-level were heterogeneous across individuals (ranging from substantially different to highly similar), the heterogeneity was consistent across the two samples, t(177) = -0.857, p = 0.392, d = 0.130. The pairwise comparison of the individual-level networks revealed similar heterogeneity in both samples (Sample 1: M = 0.693, SD = 0.107, range = 0.181–0.987; Sample 2: M = 0.701, SD = 0.091, 0.424–0.997), and was hardly different, t(6697) = -3.872, p < 0.001, d = 0.086. On average, the individual-level networks were more different from one another than their respective sample-level networks. These results suggest that the individual-level networks were heterogeneous when compared to each other and their respective sample-level networks but that this heterogeneity was consistent across samples. Fig. 2JSS distributions of the sample-level and individual-level networks. Displayed are the Jensen-Shannon Similarity distributions of the sample-level compared to individual-level networks (left); and the pairwise JSS across the individual-level networks (right), where the color of the boxes represents the strength of the similarity from 0.50 (orange) to the sample-to-individual midpoint around 0.75 (white) to 1.00 (blue). Sample 1 is the top row; Sample 2 is the bottom row. JSS distributions of the sample-level and individual-level networks. Displayed are the Jensen-Shannon Similarity distributions of the sample-level compared to individual-level networks (left); and the pairwise JSS across the individual-level networks (right), where the color of the boxes represents the strength of the similarity from 0.50 (orange) to the sample-to-individual midpoint around 0.75 (white) to 1.00 (blue). Sample 1 is the top row; Sample 2 is the bottom row. Dimension similarity Dimension similarity focused exclusively on sample-to-individual level comparisons of community solutions. The individuals whose communities had the highest, median, and lowest values of ARI with their respective sample-level communities are displayed in Fig. 3. In Sample 1, the highest (ARI = 0.811) and median (ARI = 0.236) similarities were significantly different (p < 0.001 and p = 0.010, respectively) than would be expected from random (ARI = 0). The lowest (ARI = -0.174) similarity did not significantly differ (p = 0.901) from random. Similarly, in Sample 2, the highest (ARI = 0.811) and median (ARI = 0.174) similarities were significantly different (p < 0.001 and p = 0.029, respectively) than would be expected from random community solutions whereas the lowest (ARI = -0.174) similarity did not significantly differ (p = 0.902). Across individuals in Sample 1 and 2, thirty-seven (35.9% of the sample) and thirty (39.5% of the sample; respectively) had similarities with their sample-level communities that were not significantly different from random community solutions. Taken together, these results suggest that although the majority (> 60%) of the individuals had community structures that were more similar than random, there was a substantial proportion of individuals who were not. These results suggest that the sample-level dimensions are not wholly representative of all individuals in their respective samples. Fig. 3Sample-level community comparisons with individual-level communities. Sample 1 (top) and Sample 2 (bottom) community comparisons with the sample-level communities (leftmost) compared against the highest (second from the left), median (second from the right), and lowest (rightmost) similarities of the individual communities. Nodes (circles) represent mPERMA variables, blue and red edges (lines) represent positive and negative regularized partial correlations (respectively) with the thickness indicating magnitude, and the color of the nodes denotes the community (dimension) to which each variable belongs. Sample-level community comparisons with individual-level communities. Sample 1 (top) and Sample 2 (bottom) community comparisons with the sample-level communities (leftmost) compared against the highest (second from the left), median (second from the right), and lowest (rightmost) similarities of the individual communities. Nodes (circles) represent mPERMA variables, blue and red edges (lines) represent positive and negative regularized partial correlations (respectively) with the thickness indicating magnitude, and the color of the nodes denotes the community (dimension) to which each variable belongs. Centrality similarity At the node-level, despite the strong similarity between node strength for the sample-level networks (r = 0.793, p < 0.001), there was substantial heterogeneity for the rank-order associations between the sample-level and individual-level node strengths for Sample 1 (r’s between − 0.425 and 0.636) and Sample 2 (r’s between − 0.586 and 0.571; left in Fig. 4). When comparing each sample’s distributions (Sample 1: M = 0.096, SD = 0.260; Sample 2: M = 0.094, SD = 0.252), there was no significant difference, t(177) = -0.048, p = 0.962, d = 0.007. On average, the sample-level rank-order of node strength had a small relation to the individual-level rank-order of node strength. Fig. 4Between- and within-person node strength comparison in two samples. Displayed are the Spearman’s rho distributions of the sample-level and individual-level centralities (left); and the pairwise Spearman’s rho across the individual-level networks (right), where darker boxes represent greater magnitudes and the colors blue and orange represent positive and negative associations, respectively. Sample 1 is the top row; Sample 2 is the bottom row. Between- and within-person node strength comparison in two samples. Displayed are the Spearman’s rho distributions of the sample-level and individual-level centralities (left); and the pairwise Spearman’s rho across the individual-level networks (right), where darker boxes represent greater magnitudes and the colors blue and orange represent positive and negative associations, respectively. Sample 1 is the top row; Sample 2 is the bottom row. For the pairwise individual-level centrality similarities, there was similar heterogeneity in Sample 1 (M = 0.029, SD = 0.267, range = -0.796–0.764) and Sample 2 (M = 0.013, SD = 0.264, range = -0.782–0.786) such that the difference was negligible, t(8101) = -2.627, p < 0.001, d = 0.061. On average, there was no relation of the rank-order of node strength between individuals suggesting that the importance of any one PERMA element was substantially different from person-to-person. Taken together, these results suggest that despite consistency in centrality at the sample-level (three most central for Sample 1: P2, P1, E2, and Sample 2: P2, P3, P1) there was consistent and considerable variability at the individual-level in both samples. We treat this as a descriptive pattern rather than a definitive ranking, as centrality estimates varied substantially across individuals in both samples. Synchronous components Pairwise synchrony patterns were largely consistent across both samples at both the sample- and individual-level (Fig. 5). Four out of five possible variable pairs that were identified as synchronous were found in both samples and were similarly some of the most frequently identified variable pairs at the individual-level (E1–E3, A1–A3, P1–P2, A1–A2). Generally, the most frequently identified synchronous variable pairs at the individual level for both samples were variables that belonged to the same theoretical P, E, or A element dimension at the sample-level (e.g., 28% for P2–P3 in Sample 1; 20% for P2–P3 in Sample 2). Specifically, 11–45% of participants in both samples showed at least one within-element synchronous pairing (green bars). This result is not surprising as these variables were crafted to represent the same underlying construct. That said, the most synchronous variable pair (E1 – E3) in both samples ranged from 38 to 45%, which was less than 50% of their respective samples. Unexpectedly, this suggests that even for within-element synchronous pairings—which should theoretically represent the same well-being element— no variable pair was synchronous for the majority of people. In other words, more individuals deviated from the sample-level patterns of synchrony pairings (62% in Sample 1, 55% in Sample 2) rather than matched sample-level patterns. Surprisingly, there were several variable pairs that were not in the same dimension (cross-element synchrony pairings; yellow bars) at the sample-level that were synchronous for a few individuals. These cross-element synchrony pairings (yellow bars)were less frequent for participants in both samples, with ranges of less than 15% (e.g., P1–E2 was 14% in Sample 1, R1–M1 was 9% in Sample 2). There were several synchronous cross-element pairings across many people but the majority of pairs were identified in 10% or less individuals for both samples; this suggests heterogeneous processes across individuals. In sum, these patterns of sample-level, cross-element variable pairings reflect the relative unique processes of each person’s well-being and support the different network and dimension structures reported earlier. Fig. 5Frequency distributions for individuals with synchronous variable pairs. Frequencies of the number of individuals with variable pairs reaching synchrony (wTO > 0.25) in Sample 1 (top) and Sample 2 (bottom). The bar colors represent whether the variable pairs were in the same (green; within-element) or different (yellow; cross-element) PERMA dimension in the respective sample-level network. Values at the top of the bars represent the sample-level wTO values and the gap represents the cut-off between variable pairs that are synchronous at the sample-level. Frequency distributions for individuals with synchronous variable pairs. Frequencies of the number of individuals with variable pairs reaching synchrony (wTO > 0.25) in Sample 1 (top) and Sample 2 (bottom). The bar colors represent whether the variable pairs were in the same (green; within-element) or different (yellow; cross-element) PERMA dimension in the respective sample-level network. Values at the top of the bars represent the sample-level wTO values and the gap represents the cut-off between variable pairs that are synchronous at the sample-level. Summary Broadly, there is consistency at the sample-level in network structure, dimensionality, and centrality across both samples. At the individual-level, this consistency turns to heterogeneity for the between sample- and individual-level structures as well as pairwise similarities between individuals across network structure, dimensionality, and centrality. The distributions of the similarities of the network structure and centrality, however, did not significantly differ. These individual-level results support the conclusion that there is consistency of heterogeneity. Discussion We investigated how elements of well-being, as experienced in daily life by first-year young adult college students, form dynamic networks, examining both sample-level trends (i.e., across all people) and individual-level patterns (i.e., each person’s idiographic well-being structure). Using a multidimensional perspective of well-being, theorized to be made up of various dimensions that form a web of well-being24,32, we implemented a dynamic psychometric network approach to examine how well-being changes as a system over time. From this analysis, several key conclusions emerged. First, the structure of well-being was analyzed over time at both sample and individual levels and revealed how well-being elements theorized in the PERMA model form dynamic networks. At the sample-level, our analysis identified the theoretical PERMA dimensions, which were replicated in a second sample. The global connectivity patterns (within and between PERMA dimensions) were highly similar (indicated by the JSS results). We also evaluated centrality at the sample-level—both samples had positive emotion items among the most strongly connected items to other items in the network (one Engagement item had similar elevated connectedness). This is consistent with previous findings using single time points26. Also, this pattern is descriptive rather than a definitive ranking, because centrality estimates can vary substantially across samples and individuals81,82. Accordingly, we use the sample pattern as a reference point for interpreting person-specific deviations, not as evidence that any element is universally most influential. We also examined whether these sample-level findings held true at the individual-level. The individual-level results highlighted the fluid nature of well-being experiences in first-year college students where the strength and nature of interactions between elements varied considerably from the sample-level. In fact, we found that although around 50% of individuals’ well-being networks had five communities, no single individual had exactly the same sample-level dimensions (which were consistent with the theoretical dimensions). Further, at least 40% of the individuals in both samples did not have the same number of communities as the sample-level. Departures from the theoretical PERMA dimensions reflect alternative constellations (elements of different theoretical dimensions forming alternative dimensions), consolidations (two or more elements functioning as a single cluster for a given person), or divergences (an element differentiating into multiple clusters). In practical terms, this finding means the same experience (an exam, a new friendship) can propagate through different parts of a student’s well-being system, so targets inferred from the sample-level model won’t fit everyone. Methodologically, this result clarifies that the five-dimension structure is a useful average, not a person-level template—hence the value of pairing sample summaries with idiographic analyses. Substantively, person-specific organizations point to different leverage points for support (e.g., social, cognitive, or mastery-related pathways) depending on how an individual’s network is configured. Second, our synchrony analysis—assessing elements rising and falling together from day to day—offered a complementary view of short-term coupling. A substantial minority of the sample showed within-element linkages, most notably within Engagement (≈ 38–45% for E1–E3), with smaller but visible rates for Positive emotion (e.g., ~ 20–28% for P2–P3). For example, the frequent E1 and E3 (indicators of Engagement) linkage suggests that, for a notable subset of students, different facets of engagement (e.g., feeling absorbed and staying involved in an activity) tend to co-activate across days. Practically, this pattern is consistent with the idea that creating conditions for immersion—uninterrupted work blocks, clear goals, or personally meaningful tasks—may move the whole “engagement cluster” together, and for some students could indirectly support adjacent elements (e.g., progress or positive affect). At the same time, because many students did not show this coupling, engagement may be more context-specific for others, implying that support should be tailored (e.g., focusing on task fit or reducing competing demands rather than simply increasing time on task). These synchrony patterns are descriptive; we did not test moderators that might account for who shows which pairings, so we avoid attributing them to individual differences and flag explaining this heterogeneity as a priority for future work. Notably, no within-element pairing characterized a majority of students. Cross-element pairings were rarer for any specific pair (each ≤ 15%; e.g., Positive Emotion with Engagement ~ 14%, Relationships with Meaning ~ 9%). Overall, more students deviated from the sample-level synchrony pattern than matched it (about 62% in Sample 1; 55% in Sample 2). Practically, this pattern suggests reasonable starting points for monitoring and support lie within elements (e.g., building daily engagement and progress), but the most effective levers will differ across students; cross-element routes (such as meaning-building through relationships) seem to apply to smaller subgroups. Hence, group summaries are useful heuristics, but selecting targets is best guided by each student’s observed synchrony profile. Taken together, these patterns support PERMA’s exclusivity (distinct communities) and connectivity (time-varying covariation), while underscoring substantial person-specific organization. Such differences suggest that each person’s well-being is shaped by a unique combination of factors, which may not be fully represented by aggregate, sample-level models. For example, a first-year college student may be experiencing periods of academic success, showing strong connections between Accomplishment and Positive emotions in their well-being network, while another student may find meaning through personal relationships, showing stronger connections between Meaning and Relationship dimensions. This divergence in the composition and dynamics of individual well-being networks highlight the inherently person-specific nature of well-being in young adults and that well-being is not a static construct but a continuously evolving interplay of various factors. Each individual’s network of well-being elements can change in response to external pressures and internal traits, highlighting the need for dynamic approaches in both research and practical applications. Implications and future directions For young adults, especially those transitioning to college, our findings—replicated in two studies—have significant implications. This developmental period accompanied by a drastic change in environment and lifestyle for those transitioning to college, is marked by substantial changes in daily routines, social interactions, and responsibilities, which can lead to fluctuations in well-being. First-year college students face unique challenges, including adapting to a new academic environment, establishing new social connections, and often living away from home for the first time. These changes can result in heightened stress, homesickness, and a sense of uncertainty or they can result in greater confidence, stronger relationships, and new purpose. Additionally, the demands of academic performance, time management, and navigating new social dynamics can create a complex and often overwhelming experience. Recognizing that each student’s well-being network is unique allows for the development of targeted strategies that consider individual differences in experiences and needs. For example, some students might benefit from interventions that emphasize positive emotions and engagement in enjoyable activities, whereas others might need support in building strong relationships or finding purpose in their academic and personal lives. By tailoring interventions to the specific dynamics of each student’s well-being network, we can promote better mental health and overall well-being, helping first-year college students navigate this important transition with greater ease and success. We note however, that although we observed heterogeneity in structure and synchrony, the present study did not test whether specific person-level characteristics or context account for these differences. A clear next step is to examine whether stable traits (e.g., personality, baseline well-being), contextual factors (e.g., residence, workload), and transition experiences systematically predict person-specific network features—such as the number/content of communities, node-strength ranks, and counts of synchronous pairs. Practically, this means treating these idiographic features as outcomes and modeling them as a function of traits and context (e.g., multilevel models with random slopes, or mixture/profile approaches) to determine whether the observed heterogeneity reflects systematic individual differences rather than unsystematic dispersion. This variability also underscores the importance of developing a deeper understanding of the mechanisms that drive these differences. There is a need for dynamic models that can more accurately reflect the fluid and evolving nature of well-being, which would ideally incorporate methods that capture these fluctuations over time and adapt to individual differences. More recently, theoretical perspectives and methodological approaches centered on capturing individual differences in well-being dynamics (e.g., GoHiAR83 ; well-being as a dynamic process in mobile health interventions46,84 are on the rise, and investigations examining well-being as a process are emerging. Such approaches would provide a deeper understanding of how well-being elements interact and change, offering a richer, more nuanced perspective than is possible with traditional static models. Our findings also highlight the need for integrating intra- and inter-individual approaches in investigations of psychological constructs48,85. Although inter-individual research helps identify general trends and common patterns across large groups of people, intra-individual research focuses on the unique experiences of individuals as they live their daily lives48. Combining these approaches could lead to a richer understanding of fluid psychological constructs such as well-being in fluctuating developmental stages (e.g., young adulthood), one that acknowledges and incorporates the variability and personalization evident in our findings. Our findings prompt rethinking well-being intervention designs and implementations for young adults, especially those transitioning to college, moving beyond one-size-fits-all approaches that overlook the nuances that characterize a person’s well-being. Moreover, the dynamic nature of these networks—where the importance and influence of certain elements can shift over time—suggests that well-being interventions should be adaptable and responsive to changes in an individual’s life context. The Process-Based Therapy approach51,57 is one example of such an approach in the clinical sciences (e.g., Li et al.83 and Westhoff et al.86). Limitations Our operationalization of well-being was confined to the five PERMA domains. Because network models are conditional on the variables included, partial correlations (and subsequently, networks) can change when additional constructs are introduced87,88. Although prior work indicates that PERMA communities remain separable—and hub rankings stable—when life-satisfaction and affect items are added32, the present conclusions should be interpreted as applying to a PERMA-centered representation of well-being. Broadening the construct space (e.g., authenticity, self-acceptance, personal growth, autonomy/competence, physical health) and running sensitivity analyses with alternative item sets would strengthen generalizability. Additionally, our mPERMA “Meaning” items primarily index purpose (direction/goal pursuit) rather than the broader tripartite conception of meaning (purpose, coherence, significance). This narrower operationalization may conflate meaning with purpose and underrepresent coherence/significance dynamics; accordingly, inferences about “Meaning” here should be read as momentary “purpose”. Future EMA work should disaggregate these facets to test the robustness of network structure and edges. The samples comprised first-year college students. We do not claim that well-being is more idiographic in this life stage than in others; rather, students’ first semesters offer a unique window in which dynamics may be especially observable. Even so, generalizability beyond college settings is limited. Replication with non-college emerging adults, other age groups, diverse cultural contexts, and different life transitions—and tests of measurement equivalence for momentary PERMA—are needed. Each participant contributed approximately 28 signals over one week. This density is adequate for the present dynamic network estimation, but it constrains inference about slower processes, rarer events, or the longer-term stability of centrality rankings. Designs with longer or multi-burst sampling, greater within-day density, and mixed signal- and event-contingent prompts would permit finer time-scale comparisons and more precise person-specific estimates. Conclusion The current investigation of well-being as a dynamic system challenges the status quo of its static study in the psychological sciences. Our results show that dynamic representations of well-being are not monoliths but rather mosaics of interacting elements. Despite using the same elements (PERMA items), each person had a unique organization that was different from their respective sample. Such a finding highlights the heterogeneity in the processes of people’s well-being and marks a critical step toward redefining how well-being is conceptualized and studied in young adults. By embracing the complexity and individuality revealed in our analysis, both researchers and practitioners can better serve the diverse needs of individuals in this stage of life. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1