According to Fredrickson (1998), positive emotions can be defined from two perspectives: (1) their subjective quality as containing some kind of positive experience and (2) their functions by enhancing well-being and initiating the described upward spiral. Accumulating research on positive emotions since then supports their functional role in various domains, such as in the clinical field (e.g., Lambert et al., 2012; Lamers et al., 2012; Sin and Lyubomirsky, 2009), in pain management (Hanssen et al., 2017; Thong et al., 2017), social connections (Brown and Fredrickson, 2021), physical health, and health behavior (Fredrickson et al., 2021; Pressman et al., 2019). More specifically, singular positive emotions, such as gratitude, happiness, or hope (Chang E. C. et al., 2022; Chang O. D. et al., 2022; Cohn et al., 2009; Xiang and Yuan, 2021), have been explored for their unique effects, supported by the notion that they do not necessarily all account for the same latent construct (Roth and Laireiter, 2021).

A diverse body of research on the BBT emerged throughout the past decades, mainly in terms of applying the theory in diverse fields of research and less about testing the model per se, which was done mainly by Fredrickson and her group (e.g., Fredrickson, 2004, 2013; Garland et al., 2010). Indeed, the model offers a well-applicable framework on how positive emotions, personal and social resources, cognition, and outcomes could interact either in each moment or across time. The following summary of recent research on the theory is not exhaustive but rather gives a selected overview of current findings and approaches.

In their study, Xiang and Yuan (2021) sought to examine the predictive effect of gratitude on satisfaction with life, with mindfulness and both benign and malicious envy as potential mediating factors. The theoretical framework posits that gratitude may have a broadening effect and potentially influence envy-related processes. Similarly, Johnson et al. (2021) investigated the relationship between mindfulness, positive emotions, resilience, and perceived stress. Testing multiple models against each other, non-judging and non-reacting were identified to be directly and indirectly related to resilience and perceived stress and emotions. Using structural equation modeling (SEM), Denovan and Macaskill (2017) investigated the relationship between resilience, leisure coping, and positive affect and their impact on stress and flourishing. The study showed that coping and resilience directly influenced flourishing but not perceived stress. Many similar and additional studies were conducted during the recent years, investigating the pathways between typical variables from positive psychology to test parts of the BBT in various contexts (e.g., Chang E. C. et al., 2022; Chang O. D. et al., 2022; Chen et al., 2021; Enete et al., 2022; Johnson et al., 2021).

Although many studies have been conducted in this field, three major limitations remain: (1) Definition and operationalization: Due to the ambiguity of the model’s definition, the broadening term is not clearly conceptualized, which hampers its operationalization. (2) Design and methods: Often, cross-sectional data are tested improperly with statistical models that assume a causal structure of the data, like structural equation modeling, regression analyses, or mediation models. (3) Partiality: Many studies have only examined specific facets of the theory, such as the broadening or building effects, neglecting to investigate their temporal sequence, functions, and reciprocal relationships with positive emotions, satisfaction with life, and mental health.

The present studies seek to overcome some of these limitations and deliver first insights into the model’s general mechanisms without assuming a causal structure to the data and without having collected longitudinal data. Using network modeling, the connections between the variables can be tested while controlling for the influence of the remaining ones in the network. While this cannot substitute for a thoughtful longitudinal assessment, the presented studies can provide preliminary guidance for a future re-conceptualization of the theory.

Recent theorizing and methodological advances allow researchers to adopt a new approach for analyzing multivariate and interdependent data through network modeling. This approach treats systems of variables as interconnected networks (Fried and Robinaugh, 2020), which can be investigated by a class of statistical applications, so-called network models (Boorsboom et al., 2022). The major advances and primary applications of these models are within the domain of psychopathology, where relationships between symptoms can now be modeled in more detail (Boschloo et al., 2015; Fried et al., 2015; Isvoranu et al., 2017). Unlike network models in other scientific branches, psychological networks are defined by variables and their relations with each other, rather than specific entities (Deserno et al., 2022). Models of this kind are called “Pairwise Markov Random Fields” (PMRF; Epskamp et al., 2022), where variables are referred to as nodes, which are connected through edges conveying information about the strength and direction of the relationship. If a relationship between two variables is depicted in a network model, these two variables are conditionally dependent. While dependence can be termed in many ways, the definition of independence is straight forward. Two variables (e.g., X1 and X2) are considered independent if knowledge of one does not provide information about the other. If this independence holds after controlling for a third variable (e.g., Y), or more, the variables are considered conditionally independent (X1 ⫫ X2 | Y; Waldorp et al., 2022). The relevance of a particular node in the network is quantified by measures of centrality, such as strength, betweenness, or closeness (Opsahl et al., 2010; Robinaugh et al., 2016). Most research is aiming beyond merely drawing a network graph but rather to conclude on an optimal model, given the data. This model ranges somewhere between a fully independent, with no connections between any nodes, and a saturated model, including every possible edge (Blanken et al., 2022). For continuous datasets, Gaussian Graphical models (GGM) and partial correlations are commonly used to compute the core model. The partial correlations refer to the edge between two nodes after conditioning on all other nodes in the network. To identify the optimal model, which is parsimonious without sacrificing performance, different algorithms can be applied, such as thresholding, pruning, model search, and regularization (Isvoranu and Epskamp, 2021). Regularization, an application penalizing complexity in statistical models, is borrowed from machine learning and applied in this article due to its generalizable network structures (Blanken et al., 2022; Hastie et al., 2001). Regularization can use different approaches; for multivariate estimation in GGMs, the graphical Least Absolute Shrinkage and Selection Operator (LASSO) is recommended (Friedman et al., 2008). LASSO creates sparse networks through setting edge weights to zero (Tibshirani, 1996), contrary to other methods, which shrink parameters. This procedure involves the tuning parameter λ which is hence directed through the Extended Bayesian Information Criterion (EBIC, Chen and Chen, 2008), which is specifically suitable for PMRFs (van Borkulo et al., 2014). The rigor of the EBIC is controlled through the manually set hyperparameter γ (set to 0.5 in this article). After estimating the model, its stability can be assessed through non-parametric bootstrapping (Fried et al., 2022; Hastie et al., 2015). Therefore, the computed bootstrapped confidence intervals cannot be used for significance testing in regularized models, which control for error variance already. The stability of centrality estimates can be computed through case-drop bootstrapping (Epskamp et al., 2018). Lastly, different groups can be compared for network equivalence, using network comparison tests (van Borkulo et al., 2023). These involve a re-estimation of networks by group, a permutation procedure which constructs the respective test statistic and finally the test of significance against the distribution of the statistic. In the present paper, tests for network invariance and global strength are applied. In case of specific hypotheses, targeted edges can be compared between groups as well (Fried et al., 2022).

In two studies (N_total = 614), we explored the relation between commonly used components of BBT, including positive emotions, current thought-action-potentials (broadening), resources (building), and respective outcomes (e.g., satisfaction with life). To our knowledge, this approach is new to the field and is expected to deliver the most complete picture of dynamics within the BBT. While the BBT is a model with temporal components, we are yet confident that a cross-sectional network approach should be sufficient to take the equivalent of a snapshot of the between variable relationships. While we purposely avoid causal claims, we argue that due to the recursive nature of the hypothesized model, we can get useful and informative insights, even by cross-sectional data, as in our case, which can inform both theory and future studies.

To test the relation between all components of the BBT, six core measures were identified from the literature and selected as indicators of positive emotions. To buffer against measurement dependencies, two measures for broadening and building were selected. After data collection and cleaning, a regularized network model, using multivariate estimation was computed between all variables. The analysis aimed to test the hypothesized directional links between positive emotions, broadening, building, and positive outcomes, with a particular focus on the expected strong association between positive emotions’ broadening function and positive outcomes. Consistent with the theoretical framework, positive emotions were expected to have weaker direct connections with building variables, while building measures were anticipated to have a positive association with both broadening indicators and positive outcomes.

Descriptive statistics for all measures are presented in Table 1. In most cases instruments showed satisfying reliability. Only for cognitive coping (CRI), Cronbach’s alpha was below 0.80. The data can be found in the section of supplementary material as well as further fit indices.

Set out to explore the cross-sectional interplay of relevant variables, used in BBT-research, study 1 delivered a series of interesting results. The original theory postulates a causal path from positive emotions to broadening, from there to building, preliminary terminating in positive outcomes and in reverse resulting in again enhanced positive emotions. Although the temporal sequence was not the subject of our specified design, we were still able to observe many of the components described in the theory. From the estimated network, the variables for positive emotions, resources and life satisfaction showed strong relations and appear to be connected in a robust manner. On the other hand, a serial or stepwise connection from positive emotions via broadening and building to satisfaction with life was not observed, as positive emotions and satisfaction with life mainly related directly to each other. Contrary to the theory, the broadening variables were either excluded (TST) or only loosely related to the rest of the variables. Cognitive coping was related to resilience but was not connected to positive emotions. It seems noteworthy that this pattern would have been unlikely to be found in a non-regularized manner, which optimizes the variables toward partial correlations and elimination of non-meaningful edges.

Study 2 was planned and conducted to replicate the findings and to test alternative measures for the broadening variable to explore the measurement dependency of the findings of study 1.

Using network modeling and network analysis, we investigated the interplay of variables, commonly involved in the broaden-and-build theory of positive emotions. We found substantial support for a strong interplay between positive emotions, building measures, and positive life outcomes. The role of broadening measures used to capture the process of changed thought-action repertoires, however, remained questionable. In both studies, they did not seem to act as hypothesized by connecting the model parts in a meaningful way. Due to the loose connection of the broadening measures to the rest of the model, the overall model cannot be supported. An exploratory series of network comparisons from our two independent studies did not reveal meaningful differences in model composition except from employment status. It appears noteworthy at this point that we examined a model with distinct temporal components in a cross-sectional manner and therefore recommend interpreting the findings only in a preliminary manner.

The current study raises questions about the role of the broadening component within the framework of BBT. While the other components of the model seem to be well interrelated, broadening measures were found to be rather isolated from the remaining system. This may be due to the lack of a temporal order in our study. However, to our knowledge, the theory does not make any clear statement on the timing of the variables within the model. Given the hypothesis of a continuous upward spiral of the model, it appears plausible to expect that the interrelationship between the components should be observable in a cross-sectional approach. Therefore, as one central consequence, we assume that the model is well-functioning without the broadening component as well. Until the role and measures of broadening are not captured better in a conceptual and empirical sense, we can conclude that positive emotions, resources, and life satisfaction are well interrelated and meaningful for each other, also without the variable of broadening the mind as a necessary intermediate stage or phase. In our view, this conclusion remains valid even if we keep in mind that we applied a cross-sectional design in both studies.

As demonstrated in the current studies, network modeling holds potential for the field of positive psychology. The approach allows to reach beyond often limited insight into multivariate patterns without overly strong assumptions. As seen in the case of the BBT, directionality, and causality might be not straightforward. While precisely planned studies on specific effects clearly contribute to the understanding of fine-grained mechanics, mere hypothesis testing of specific emotions or specific traits cannot be the sole advance toward understanding complex systems of human functioning. To gain clearer perspectives on broader landscapes of measures, the network approach can offer a strong toolbox to observe complete models or complex relationships and facilitate the generation of future hypotheses.

Although, as argued earlier, from our perspective longitudinal data are not strictly necessary to study the BBT, the cross-sectional nature of our two studies limits the interpretability of the results we obtained. Since the model also makes causal assumptions about the temporal course of each variable, it seems reasonable and necessary to also use prospective longitudinal data to test these assumptions. Beyond that, the selection of the variables, while drawn from a specific literature, remains arbitrary. While optimism and resilience are well-established resources in positive psychology (Färber and Rosendahl, 2020; Heekerens and Eid, 2021; Shanahan et al., 2021), future studies are needed to test the generalizability of the observed patterns. Last, we observed the weakest reliabilities in measures of broadening, which might put the general robustness of these measures into question.

Two central implications emerge from our results: The first relates to the conceptual level of the theory and the second to its empirical testing. At the conceptual level, necessary further development and sharpening of the BBT, especially with respect to its broadening component, is needed. At the moment, this component seems to be neither conceptually nor theoretically adequately conceived. One can assume that the phenomenon of broadening the mind may be too short-sighted and include other, perhaps more important theoretical components, such as cognitive or affective ones. These could be thought of and formulated as counterparts to the well-known worrying or rumination elements of vicious cycles (Moorey, 2010). At the same time, these could also play an important role in broadening and opening of the mind in the sense of the BBT.

Empirically, continuing research presented here in a careful causal manner, the sustained application of network models appears justified. This enables researchers to observe a complex system of variables in various ways and reduce the complexity to the most meaningful substrate. To gain further insight into the BBT and Positive Psychology from a network perspective, two directions appear specifically promising: (1) a longitudinal extension to monitor the development of the variables interplay across time; (2) large-scale samples on more variables could offer a better understanding and insight for future theorizing. Overall, however, research on BBT should also change its focus in the future. Applications should not be the primary instance of research, but first and foremost the theory which must be tested with greater rigor and developed further.

The BBT elicited mixed evidence in our two studies. While many of the model’s components showed meaningful relations and were robust across demographic groups, broadening consistently deviated from the model’s predictions. Based on our findings, we discussed two major implications for future research, to further develop and refine BBT theoretically, as well as conducting longitudinal studies with bigger samples and considering a broader range of measures. All in all, the primary focus of research should lie on the advancement of the theory and less on further searches for new applications.