New trends in network modeling of psychopathology

The network theory, as outlined in D. Borsboom's seminal pa­per in this journal1, postulates that mental disorders may be viewed as emergent phenomena from complex systems, in which symptoms and external factors directly influence each other in a network of interacting components. From this theoretical frame of mind arose network psychometrics2, 3, the field of study concerned with estimating multivariate statistical models that allow for network representations from data. While these techniques have been predominantly applied to cross-sectional datasets, in which patients are measured only once in efforts to gain insight into nomothetic relationships, they are also promising in forming patient-specific network models designed to uncover idiographic relationships that may facilitate case conceptualization and further guide treatment. The suite of techniques developed in network psychometrics allows for unique ways to explore relationships in symptom data and beyond, as well as to check the robustness and replicability of these results. Research inspired by the network theory of mental disorders is now vast and taking many directions. Here we introduce some of the most promising new trends in this field. A meta-analytical framework. As a result of the rapid expansion of the network analytic framework in psychopathology research, and in response to critical arguments on the generalizability of network models obtained from single samples, the novel meta-analytic methodology (i.e., meta-analytic Gaussian network aggregation, MAGNA) represents a central advancement in the field. MAGNA allows researchers to perform meta-analyses of network models, making it possible to aggregate results across multiple studies, and providing a statistical and objective framework to summarize research findings4. This methodology has recently been applied to the field of post-traumatic stress disorder, where correlational structures of 52 different samples with a total sample size of N=29,561 have been analyzed under a single pooled network model. The upcoming years are likely to see a rise in meta-analysis of various mental disorders, as well as across a multitude of population subgroups, ultimately leading to stable and generalizable network structures that can further be used in theory formation, as well as in innovative prevention and intervention strategies. From exploratory to confirmatory network analysis. Network models to date are explorative in nature, allowing researchers to search best-fitting models that could explain links present in empirical data. As a result, most research in the field of network psychometrics is hypothesis-generating and exploratory. As we become more acquainted with the operating principles of such models, moving from exploratory to confirmatory analysis is a natural next step. We see more often that setting up hypotheses about the presence and strength of links in a network structure is intuitive and common, and many new research avenues and developments are currently directed toward confirmatory studies. To this end, it is increasingly more common for network analyses to be at least in part pre-registered, registering not only the dataset to be analyzed and the analytical plan, but also expected findings in the network architecture. In addition, dedicated software now exists to perform confirmatory network analysis5, which allows for assessing the fit of a pre-defined network architecture in the same manner as a confirmatory factor analysis allows for assessing the fit of a predefined factor model. Finally, recent lines of research investigate the inclusion of prior theoretical knowledge in network estimation. Longitudinal analysis. The vast majority of network analyses are performed on datasets in which every subject is measured only once, often termed cross-sectional data. While cross-sectional analyses based on large sample sizes can certainly be informative, it has been recognized that the interpretation of such results may be troublesome6. Most notably, cross-sectional data cannot distinguish between-person effects (e.g., a person that is on average more anxious also experiences on average more depressed mood) from within-person effects (e.g., whenever a person feels more anxious than his/her average, he/she often also experience higher levels of depressed mood). Longitudinal analysis, in which network models are estimated from either intensive time-series data or large sample panel data, has grown popular in modeling such within-person fluctuations over time. While psychiatric symptoms are often defined over a long period of time, and therefore fluctuate less over time by definition, the problems they represent certainly can fluctuate over time, as can the mood states associated with mental health problems. To this end, longitudinal network analysis presents powerful avenues for future research. Formal mathematical modeling. While the field of network psychometrics arose from the conceptual thinking in the network theory of mental disorders, it is not necessarily tied to network theory, which conceptualizes mental health as a system of interacting components. Where network psychometrics takes a theory-agnostic bottom-up approach of estimating relationships between symptoms from data, an increasingly popular alternative is to use theory-laden top-down approaches, instead forming networks by using theory to inform mathematical equations that aim to explain how data are generated. While such formal modelling is popular in many fields of science (e.g., such models have been used to inform policy regarding the COVID-19 pandemic), they are yet rarely used in psychopathology. This approach shows great promise, however, both in the formation of nomothetic theories aiming to explain common phenomena7, as well as in idiographic case conceptualizations of patients in clinical practice8. Embracing the complexity of mental health. In attempting to obtain identifiable multivariate models that are estimable from practically obtainable datasets, network psychometrics may make concessions that make the models used deviate from the complex systems thinking that inspired network theory1. For example, psychometric network models estimated form cross-sectional data mostly include only pairwise interactions, do not feature phase transitions between multiple stable states, and include only linear effects, while complex systems thinking often involves the presence of feedback loops, phase transitions between multiple stable states, interactions at different time-scales, and non-linear effects. It may be that the understanding of mental health requires all these concepts and more, being an interplay not only of symptoms but also of numerous other factors, ranging from biological to sociological factors, and from fast effects that span seconds to slow effects that span a lifetime9. The development of methods that capture and explain this complexity may be one of the great challenges that mental health research is to face in the coming decades.