Psychological resilience: To adapt, or to absorb?

One of the earliest definitions of the concept of resilience in the scientific literature is by Holling (Reference Holling1973): “resilience determines the persistence of relationships within a system and is a measure of the ability of these systems to absorb changes of state variables, driving variables, and parameters, and still persist.” Hollis was studying ecological systems, but the phenomenon is observed in a wide range of complex systems and reflects the fact that once a system has settled into a stable state, it will resist internal and external perturbations in order to persist the current state into the future. The term resilience has acquired slightly different meanings in different scientific domains, a recent review identified ecological, organizational, engineering, economic, and psychological resilience, which were in turn defined slightly different within each domain (see Table 1 in Fraccascia et al., Reference Fraccascia, Giannoccaro and Albino2018). Based on the literature of contemporary psychological science, Masten et al. (Reference Masten, Lucke, Nelson and Stallworthy2021) suggest defining resilience as: “the capacity of a dynamic system to adapt successfully through multisystem processes to challenges that threaten the function, survival, or development of the system.”

This definition correctly reflects the way in which the term is used in clinical science, clinical practice, and common language use, that is, the ability to persist a very specific state, one that is considered beneficial to the individual in question, such as a “healthy” physical or mental state, or a developmental trajectory that is “typical.” However, formally, resilience refers to the ability of a complex system to persist any state in response to perturbation (Cui et al., Reference Cui, Lichtwarck-Aschoff, Olthof, Li and Hasselman2023), irrespective of whether the state is a clinical depression, or represents 10 years of sobriety. In fact, from a complex systems perspective, psychotherapeutic interventions are designed to perturb the psychopathological state, to break down its resilience and destabilize it (Olthof et al., Reference Olthof, Hasselman, Strunk, Aas, Schiepek and Lichtwarck-Aschoff2019), allowing the system to reorganize itself into a new stable state that is not pathological (Olthof, Hasselman, Oude Maatman, et al., Reference Olthof, Hasselman, Oude Maatman, Bosman and Lichtwarck-Aschoff2023; see also, Schiepek, Reference Schiepek2009). As mentioned before, the general principles and processes involved in transitions between stable states in complex systems are the same for “healthy” and “unhealthy” states and measures to determine whether a system is about to transition from one stable state into another, are called “resilience loss indicators” (Hasselman, Reference Hasselman2022; Weinans et al., Reference Weinans, Quax, van Nes and Leemput2021). To avoid any confusion, I will use the term psychological resilience to refer to the persistence of states that can be labeled as “healthy,” and use qualifications such as “a stable psychopathological state” for the resilience of states that are maladaptive or “unhealthy.”

In Holling’s (Reference Holling1973) original definition of resilience, the word “absorb” is used in reference to the ability to resist changes that may come from a multitude of sources in the internal or external environment of the system. The definition of psychological resilience uses the term “adaptation” to refer to the same ability, but also adds that adaptation will occur through “multisystem processes”. The properties of adaptation and resilience are two related, but separate characteristics of complex systems. Adaptation refers to the ability of a system to reorganize its internal structure in response to changes in the internal or external environment in order to allow the system to continue to perform its general function. The latter is also included in the definition of psychological resilience as “challenges that threaten the function, survival, or development of the system.” In the context of learning and development, adaptation will generally increase the complexity of the system, or at least expand the degrees of freedom the system has available to generate its behavior. That is, adaption implies new states will emerge, that can of course vary in their degree of stability, their resilience. If a system would be caught in a state that is extremely resilient, it would, by definition, be very difficult to adapt to any changing demands of the environment, due to the “persistence of relationships within a system,” that is, the resistance to change the internal structure is implied by the extreme resilience of the state. This is why the term “absorb” is used to describe the resilience of a state: No adaptation of structure is required; the current state will persist.

From the perspective of a strong complexity commitment, the definition of psychological resilience as provided by Masten et al. (Reference Masten, Lucke, Nelson and Stallworthy2021) should be considered a mixture of the concepts of resilience and adaptation. Psychological resilience describes the observable phenomenon of the remarkable stability of specific “healthy” states that can absorb all kinds of shocks and perturbations, as well as the ability of a system to reorganize its structure in order to be able to continue producing similar resilient “healthy” states under changing demands of the internal and external environment. Given these annotations, there are no objections to commit to a complex system ontology for studying psychological resilience in development based on interaction dominant dynamics.

In their review of the models and methods used to study resilience, the authors point out that: “Still missing in cascade models, however, is the dynamic interplay of complex adaptive systems” (Masten et al., Reference Masten, Lucke, Nelson and Stallworthy2021, p. 531). The authors proceed to discuss a shortlist of multisystem resilience factors that concern processes that span many different systems (brain, body, family, school, community) and spatiotemporal scales (neurobiological, cognitive, emotional, and cultural processes). Although an important emphasis is placed on the complex, nonlinear dynamics involved in understanding resilience of developmental pathways, virtually all conclusions, suggestions, and recommendations are based on a component-dominant causal ontology in which it is assumed possible to trace the individual contributions of component processes to explain psychological resilience in development. That is, the scales and system boundaries that are acknowledged to exist, appear to play no role in data analysis and causal inference, they appear as additional independent components or mediators in the causal chain. This is also reflected in interventions which are described as functioning as a mediator to amplify (or reduce) the effects of resilience or risk factors (i.e., by “installing” a better component into the chain).

From the perspective of the strong ontological commitment to complexity, resilience should be an emergent property, not an identifiable set of risk factors. Empirical studies would focus on examining the contexts in which resilience emerges and when it does not. It is expected that what is a risk factor to one individual may be a protective factor to another, or, that such factors may even switch roles within an individual, phenomena that are in fact briefly discussed by Masten et al. (Reference Masten, Lucke, Nelson and Stallworthy2021) in reference to the work of Rutter (Reference Rutter1987). Dealing with causality in complex adaptive systems remains a difficult topic, in the next section several conceptual tools are introduced that may help researchers get a better grip on complexity, by aligning posited causes, cascades, and contexts along spatial and temporal scales.

Permissive and causative structure

In an interaction-dominant causal ontology it makes sense to distinguish between immediate causal entailment and entailment that is mediative for explaining the behavior of individuals (see Figure 2). The former refers to causes for the current state of affairs whose effects are immediate (e.g., laws of physics, genotype, sociocultural environment, personality), the latter refers to the causes that can be traced as mediators in the realization of the current state of affairs (e.g., age, time of day, quality of sleep last night, current emotional state). The divide between the different types of entailment is an effect horizon, its main purpose is to serve as an explanatory vehicle to indicate there is a structure that is permissive of the behavior in the present, and a structure that is causative (cf. Walker, Reference Walker1983). The permissive structure appears as a set of constants painted on the horizon representing boundary conditions for the causative structure. The causative structure lies within the horizon and is the set of efficient causes of the behavior under scrutiny in the present. Quantum physical phenomena are permissive of phenomena at larger spatial scales; that is, in most circumstances it will not be sensible to declare quantum phenomena as the efficient causes for the macro-scale structure of biological systems, although there is no doubt that the macro state could not manifest itself without the immediate entailment of those quantum phenomena. The same holds for speciation by natural selection: The genome is permissive of the structure and function of organisms belonging to a specific species, for example flight in birds. It is not wrong, but also not very informative, to suggest that the behavior of an individual organism, such as a bird flying away upon sensing danger, was caused by an ancient event in their evolutionary history. This ancient event does however permit the bird to fly away.

Figure 2. The gray boxes represent different effect horizons that can be placed at an arbitrary divide between spatial and temporal scales to separate immediate and mediative effects into permissive and causative structures, respectively. In addition, this illustration shows that upward causation involves structures and processes that have cascading or aggregate effects relative to higher scales. Downward causation involves structures and processes that set parameters and boundary conditions for lower scales.

In cases in which the time scale of interest corresponds roughly to the time scale of daily experiences, the permissive structure will contain variables related to socio-cultural processes and facts from history of nations that may be labelled as “systemic,” these effects are always present and immediate. It is expected that these systemic effects will result in adaptation in the sense of changes to the internal structure of the system. An example is given by Spencer et al. (Reference Spencer, Dupree and Hartmann1997) of the recursive self-appraisal processes in African-American adolescent males who will likely experience the immediate presence of stereotypes and biases as they become aware of their family, community and cultural history. These effects (protective, promoting, or random destabilizing shocks) can interact across time scales as multiplicative cascades which can have catastrophic effects (e.g. producing bi-modal outcomes as described by the cusp catastrophe; Brummitt et al., Reference Brummitt, Barnett and D’Souza2015), that will depend on the particulars of an individual’s unique history of experiences. Indeed, Spencer et al., (Reference Spencer, Dupree and Hartmann1997, p. 818) write: “In the context of such a culture, youth having similar experiences can exhibit either resiliency, or psychopathology.”

Upward and downward causation

The effect horizon represents a divide between spatial and/or temporal scales, which could emerge naturally due to the study design or the variables of interest, however, in principle its placement is completely arbitrary. For example, suppose a researcher is only interested in explaining the behavior of a sample of participants during a laboratory experiment. All characteristics and experiences of the participant before the start of the experiment are considered immediate effects, static parameters. The experimental manipulations are considered the causal structure that mediates a participants’ behavior. As is often the case in sample-based laboratory experiments, some aspects of the permissive structure may be quantified with the purpose of examining the extent of their role in explaining the behavior observed during the experiment, such as their age or educational background, whereas other factors are “controlled for” by random sampling or random assignment of participants to design cells. The causative structure in this case is narrow in the sense that there is not a lot of room to detect any effects of factors that mediate the behavior of interest, which is of course intended by the design. Any effects of variables other than the experimental manipulation must be inferred from between individual differences in values sampled from the immediate causal factors. The causative structure is also shallow, in the sense that it is common to look at just one type of outcome variable at one scale of observation, it’s either response latencies or brain activity or heart rate, or eye movements, but rarely all of those at once.

As a contrast, consider an experience sampling study about mood dynamics, in which participants for 2 months provide daily self-reports about experienced emotions, perceived stress and physical fitness and wear an electronic device which registers heart rate and electrodermal activity. The permissive structure consists of all the things that do not change during the period of observation, such as personality traits, academic achievement, ethnicity. The causative structure covers a much wider range of time scales compared to the laboratory experiment. Factors that are mediative of mood dynamics may concern stress, quality of sleep, but also self-reported mood of the previous week, or month and perhaps an increased heart rate only moments before a mood change. Due to the long observation time, we may expect interactions with the environment like the weather or random events that can be considered external shocks or perturbations, such as the loss of a loved one. The causative structure is also less shallow than the laboratory experiment because of the simultaneous observation of psychological as well as physiological processes, which allows for the determination of coupling strength between the different scales of observation as well as the direction of their coupling.

These examples reveal that there is no privileged scale at which causality resides in complex systems. The phenomena of interest, more specifically, the measurement context chosen to study those phenomena, involves a choice for a specific range of spatiotemporal scales at which mediative causes may be observed. These choices have consequences for causal inference, that is, relative to a scale of interest, the directionality of the causal entailment of other factors is important (see Figure 3). To understand these effects, one can imagine that from the perspective of the scale of interest, slower and faster changing processes have to be “resampled” in order to evaluate their effects, which is commonly referred to as coarse graining (Flack, Reference Flack2017).Footnote ¹ In general, fast processes evolving at shorter time scales can be said to be responsible for bottom-up, aggregate effects (e.g. “mean field” cascading effects), whereas processes at slower scales are more likely to have top-down effects setting boundary conditions for the faster changing processes (Noble, Reference Noble2012; Noble et al., Reference Noble, Tasaki, Noble and Noble2019). By ordering the factors that make up the causative structure relative to a temporal scale of interest, the presence of phenomena such as feedback loops that span multiple scales (referred to as circular, bi-directional, or delayed-feedback causation) can be studied. These and other interesting phenomena (such as scale-invariance) remain invisible, or, are considered logically impossible, if multiscale dynamics are ignored, of there is no commitment to interaction dominant dynamics (Runge, Nowack, et al., Reference Runge, Nowack, Kretschmer, Flaxman and Sejdinovic2019).

Figure 3. The figure displays how, relative to a scale of interest, coarse-graining of slower time scales (upper dashed line) can be understood as changes in parameters that remain constant for longer periods of time, whereas coarse-graining of faster time scales (lower dashed line) can be understood as changes in control parameters that affect the dynamics at the scale of interest. External perturbations (gray arrow) can affect boundary conditions as well as dynamics. Not shown are the downward, boundary setting effects from the scale of interest to the faster time scale and the upward collective dynamics form the scale of interest to the slower time scale.

Applying these concepts to understand the factors involved in psychological resilience in development as discussed by Masten et al. (Reference Masten, Lucke, Nelson and Stallworthy2021) would start with the question of where to place the effect horizon. This will be determined by the phenomenon of interest, obviously, if the purpose is to study life-span resilience, the permissive and causative structures would include different variables compared to a study in which resilience to real-life experienced perturbations (random shocks, systemic stressors) is observed. It is unlikely that the dynamics (and effect magnitudes) at one scale of observation easily translate to other (coarse-grained) scales, something which is for example not accounted for in the cascading model, or any other statistical model for that matter. Cascading effects (effects crossing scale and system boundaries) will often be multiplicative in nature (Kelty-Stephen et al., Reference Kelty-Stephen, Palatinus, Saltzman and Dixon2013), or represent mean-field catastrophic effects (Brummitt et al., Reference Brummitt, Barnett and D’Souza2015), but they will unlikely be of the linear additive kind as assumed in a path cascading models. Such models are also unlikely to be able to capture phenomena such as feedback loops, or, be cast as idiographic models (however, see e.g. Bringmann, Reference Bringmann2021; Ram et al., Reference Ram, Brose, Molenaar and Little2013 for potential solutions)

To summarize, studying causality in complex systems will not yield a neat list of independent efficient causes to explain the dynamics of a phenomenon of interest (cf. Runge, Bathiany, et al., Reference Runge, Bathiany, Bollt, Camps-Valls, Coumou, Deyle, Glymour, Kretschmer, Mahecha, Muñoz-Marí, van Nes, Peters, Quax, Reichstein, Scheffer, Schölkopf, Spirtes, Sugihara, Sun and Zscheischler2019), instead, a system of coupled processes may be uncovered, whose contributions can be evaluated relative to an effect horizon. The effect horizon separates coupled processes into a permissive and a causative structure and the choice to focus on a particular scale of interest determines the kind of causal role (top-down, bottom-up, immediate, mediative) factors can play in explaining the dynamics of the phenomenon of interest. Committing to an interaction dominant causal ontology as described here will have profound consequences for the design of diagnostic tools and intervention methods as well as the evaluation of the current empirical record (see e.g., Hardeman et al., Reference Hardeman, Houghton, Lane, Jones and Naughton2019; Lichtwarck-Aschoff et al., Reference Lichtwarck-Aschoff, Hasselman, Cox, Pepler and Granic2012; Nahum-Shani et al., Reference Nahum-Shani, Smith, Spring, Collins, Witkiewitz, Tewari and Murphy2017; Olthof et al., Reference Olthof, Hasselman, Strunk, Aas, Schiepek and Lichtwarck-Aschoff2019; Schiepek, Aichhorn, et al., Reference Schiepek, Aichhorn, Gruber, Strunk, Bachler and Aas2016). For example, research based on inferring linear associations between independent components based on cross-sectional samples of many individuals, may have uncovered statistically reliable dependencies between risk and protective factors, however, the complex system ontology presented here would predict that these associations represent coarse-grained dynamics that exist only at the chosen scale of interest (i.e., the sample aggregate). There are no guarantees that these associations will generalize from the group level to explain behavior at the level of the individual (see e.g., Fisher et al., Reference Fisher, Medaglia and Jeronimus2018; Wolfers et al., Reference Wolfers, Doan, Kaufmann, Alnaes, Moberget, Agartz, Buitelaar, Ueland, Melle, Franke, Andreassen, Beckmann, Westlye and Marquand2018). Nomothetic sample based studies may have identified boundary conditions that can inform studies that focus on the individual, because if the dynamics of interest can be observed at the level of daily experience of an individual, then that should be the scales of interest for scientific scrutiny.

Causal loop diagrams

A Causal Loop Diagram (CLD) is a visual representation the of dependencies of all factors involved in explaining a phenomenon that can be understood as an emergent, self-organized state of a complex system (Crielaard et al., Reference Crielaard, Uleman, Châtel, Epskamp, Sloot and Quax2022). It is constructed through a process called Group Model Building (Rouwette, Reference Rouwette2016), in which domain experts and stakeholders collaborate to identify the important causal structures. The associations in CLDs can be annotated (indication of direction of effect or nature of feedback loops) based on the empirical record. CLDs can also represent dynamics on multiple scales, which is very similar to the notion of course grained, or emergent dynamics. Although the schematic representation can be helpful, ultimately the goal is to translate the diagram into a so-called System Dynamics Model (SDM) which is a computational version of the CLD, a set of coupled differential equations. The advantage of turning the graphical model into a computational one is that it is possible to simulate “what if” scenarios. For example, an SDM was used study how diabetes may be reduced in three countries in the Caribbean. The SDMs were successfully used to inform health policy decisions (Guariguata et al., Reference Guariguata, Guell, Samuels, Rouwette, Woodcock, Hambleton and Unwin2016). CLDs can become very complex, for example, an attempt to map the multicausality of Alzheimer’s disease (Uleman et al., Reference Uleman, Melis, Quax, van der Zee, Thijssen, Dresler, van de Rest, van der Velpen, Adams, Schmand, de Kok, de Bresser, Richard, Verbeek, Hoekstra, Rouwette and Olde Rikkert2021), resulted in 38 variables and 150 connections between them interacting across 3 different scales.

Figure 4 represents a (fictitious) annotated, multiscale CLD, based on Figure 1 from Spencer et al. (Reference Spencer, Dupree and Hartmann1997). In the original figure, a notion of scale within and between the different domains was already present. Constructing a CLD may further specify the complex interactions within and between the domains. For example, Spencer et al. (Reference Spencer, Dupree and Hartmann1997) report their data was collected in a southeastern metropolitan area of the US. One could use the Group Building Method to build a CLD using experts (scientists, healthcare providers) and stakeholders (policy makers, educators, community members) from the metropolitan area, local experts. The CLD can be annotated and adjusted based on a review of the scientific literature. The CLD would be an idiographic model because it describes the complex dynamics of a particular case, it is unlikely that it would generalize to a northwestern metropolitan area. It would represent “a reflection of the knowledge and assumptions held by a person or group—a shared mental model” (Crielaard et al., Reference Crielaard, Uleman, Châtel, Epskamp, Sloot and Quax2022, p. 7), meaning that it might also change if a completely different group of experts were invited, however, procedures and tests have been developed to evaluate the reliability and validity of the models (Crielaard et al., Reference Crielaard, Uleman, Châtel, Epskamp, Sloot and Quax2022 is an excellent review).

Figure 4. A fictive annotated, multiscale Causal Loop Diagram based on Figure 1 in Spencer et al. (Reference Spencer, Dupree and Hartmann1997). The diagram shown here schematically represents the micro-level associations between important risk contributors in self-appraisal in response to stereotypes and biases. The annotations + and – indicate the direction of a hypothetical effect. Whenever loops emerge, they may be labeled as reinforcing (R, +|+ or –|–), or balancing (B, +|– or –|+). In practice, this model would be constructed based on Group Model Building by domain experts as well as the scientific literature. These diagrams can also be created for the other domains as well as the interactions between the domains (macro-level).

Suppose a multiscale annotated CLD has been made for all domains and underlying variables in Figure 4 (Figure 1 in Spencer et al., Reference Spencer, Dupree and Hartmann1997), what would resilience refer to in the sense of the general stability of a state? An unhealthy macro-state could represent a situation in which most African American adolescents in this southeastern metropolitan area of the US develop some degree of negative learning attitudes. The emergent sate would be called resilient, if it persists for a longer period of time, the development of the negative learning attitudes can likely be regarded as a maladaptive process, permitted by a set of immediate (systemic) factors related to US culture and history, personality, and gender, mediated by the lived experience of negative events that involve stereotypes and biases. We have now used the terms resilience and adaptation to describe the emergence of an “unhealthy” state which reflects their universality and explanatory power in describing complex dynamics. Within the context of psychological resilience, which was reserved for the emergence and persistence of “healthy” states, one would need to define another set of terms that describe the emergence of stable, “unhealthy” states (see Lunansky et al., Reference Lunansky, van Borkulo, Blanken, Cramer and Borsboom2022), which seems redundant if they are governed by the same mechanisms as is assumed in the complex systems perspective.

Now suppose an SDM has been created that can be used to simulate different “what if” scenarios. To simulate what may be required to transition to a resilient state in which learning attitudes are predominantly positive, the focus would be on mediative factors that could be changed in reality, which may inform the design of interventions or policies. Different scenarios can be simulated and this is in fact what has been successfully applied to inform complex policy decisions (cf. Rouwette, Reference Rouwette2016). However, there are no technical objections to change the permissive structure to simulate what happens if for example the immediate effects due to culture and history are reduced or removed. Assuming the result would be positive, this can be a tool to argue for the necessity of more profound cultural changes.

Idiographic System Modeling

Idiographic system modeling can be described as creating a CLD for an individual, to gain insight in the factors that determine what may worsen or ameliorate experienced symptoms of psychopathology (Schiepek, Stöger-Schmidinger, et al., Reference Schiepek, Stöger-Schmidinger, Aichhorn, Schöller and Aas2016). This mixed-method consists of a three-hour co-creative session before treatment starts in which a network is created that represents connections between different factors, very similar to the annotated CLD. The next step is to translate the factors and relationships identified in the network into questions that can be administered on a daily basis (personalized process monitoring), to gain insights into how these factors change over time. During the therapy the data generated by the patient will be used to inform the therapy. This method capitalizes on dynamics over content, the idea is that personalized rather than standardized questionnaires provide an optimal window into the mental, emotional, and physiological states that underly the experienced psychopathological symptoms. Studying the dynamics of these states will reveal information about the stability and interactions of the systems involved. Consequently, personalized questions may be meaningful to the patient, but difficult to interpret for an outsider, for example, in a study of the feasibility of applying personalized process monitoring methods in youth with mild intellectual disability and borderline intellectual functioning, a participant came up with the question “Did your bucket empty by smoking weed?” (Hulsmans et al., Reference Hulsmans, Poelen, Lichtwarck-Aschoff and Otten2023). We do not exactly know what this means, but one can imagine that daily self-reports of these and other personalized questions carry unique information about the mental, emotional, and physical states of this particular participant. The personalized items function as collective variables that tap into underlying dynamics, often the goal is to infer periods of destabilization (loss of resilience) to inform intervention strategies (Fartacek et al., Reference Fartacek, Schiepek, Kunrath, Fartacek and Plöderl2016; Olthof, Hasselman, Strunk, et al., Reference Olthof, Hasselman, Strunk, van Rooij, Aas, Helmich, Schiepek and Lichtwarck-Aschoff2020).

In order to empirically support the claim dynamics over content, it is necessary to show that aspects of the dynamics of personalized questionnaires (dynamical invariants) are associated to outcome measures across different individuals. Olthof et al. (Reference Olthof, Hasselman, Aas, Lamoth, Scholz, Daniels-Wredenhagen, Goldbeck, Weinans, Strunk, Schiepek, Bosman and Lichtwarck-Aschoff2022) studied 404 patients who were treated at an inpatient clinic that uses Idiographic System Modelling, to answer the question whether there is information at the level of the dynamics of these personalized questionnaires that generalizes across patients. The questionnaires consisted of 13 items on average, with an average time series length of 36 days. For each multivariate time series of personalized items, a principal component analysis (PCA) was performed, which was used to create one univariate time series for each patient by projecting only the first principal component. This timeseries represents the primary dimension in which the system changes, it is an abstract representation of the multivariate dynamics and if a transition occurs it is expected to be registered in this component (Lever et al., Reference Lever, van de Leemput, Weinans, Quax, Dakos, van Nes, Bascompte and Scheffer2020). The results revealed characteristic change patterns in the projected time series (no shift, gradual change, one shift, reversed shift, multiple shifts) that were associated to treatment effects. Patients that experienced no shift, or a reverse shift, showed the least improvement. This study shows empirically that the traditional focus on finding general factors of stable pathological or healthy states, may not be representative of the highly idiosyncratic nature of the processes involved in their emergence.

This method can of course also be used to study the emergence of negative learning attitudes in African American adolescents in the southeastern metropolitan area of the US. It would require idiographic system modeling of each participant to gain insight in what kind of factors affect their behavior and self-appraisal. This could already be very informative without personalized process monitoring: Is it the case that all participants will identify the expected immediate causes permissive of developing negative attitudes? Do participants report the same coping mechanisms, how do they express experiencing chronic, or resilient states using their own vocabulary? Personalized process monitoring would also reveal how often participants are exposed to biases and stereotypes during a day and also directly observe how this affects their daily experience of the world. Initially, one might conduct multiple case studies to evaluate the feasibility of daily measurements, as well as the study duration, which should be long enough to be able to register interesting phenomena. The methods described in Olthof et al. (Reference Olthof, Hasselman, Aas, Lamoth, Scholz, Daniels-Wredenhagen, Goldbeck, Weinans, Strunk, Schiepek, Bosman and Lichtwarck-Aschoff2022) can be used to study larger groups with the purpose of identifying associations between the dynamics and outcome variables. The results would be based on the personal experiences of individual participants, which is exactly what PVEST identifies as essential for understanding complex developmental pathways.

Multilayer networks

So far, we have not addressed the multisystemic nature of different factors associated with psychological resilience. To study different systems that are interacting, multilayer networks can be used. Figure 5 displays three different types of multilayer networks, that can be used to represent different types of multiscale interactions. The bottom row displays the “monolayer” version of the network, the question for each situation is whether the interlayer associations displayed as dashed lines in the top row should be treated differently from the association within each layer (solid lines). The interconnected graph (first column) treats variables clustered at a layer as relatively independent entities, for example, layers may represent associations between a group of psychological variables e.g., treating the domains in Figure 4 as relatively independent. The multiplex graph is a multilayer network in which all variables (nodes) are present at all layers (a multilevel graph does allow some nodes to be missing, e.g., the node v1 in Figure 5). Recurrence networks are multiplex graphs in which each node represents a point in time at which a value was observed. Different network layers can represent different groups of variables, physiological, psychological, and endocrine variables measured simultaneously in a multivariate time series (Hasselman & Bosman, Reference Hasselman and Bosman2020; Hasselman, Reference Hasselman2022; Zou et al., Reference Zou, Donner, Marwan, Donges and Kurths2019). To construct a network in which the domains in Figure 4 are layers in the multiplex, a rather intensive study would be required in which all relevant factors are queried at least every day. There are examples of such studies, in which one participant provided about 45 responses per day for a period of 239 consecutive days (Wichers et al., Reference Wichers, Groot and Psychosystems2016), enough data to construct a multiplex recurrence network with 6 different layers (see Hasselman, Reference Hasselman2022). In a hypergraph, edges can connect to more than one node, so a node can belong to different subsets that represent different types of associations or states. Recently, hypergraphs have been applied to model psychological data (Marinazzo et al., Reference Marinazzo, Van Roozendaal, Rosas, Stella, Comolatti, Colenbier, Stramaglia and Rosseel2022). In a multihypergraph, the different edge types are placed on separate layers.

Figure 5. Examples of different options for constructing multilayer networks. See text for details.

Suppose the personalised studies suggested in the previous sections would have been conducted, and multivariate time series data were available representing the idiographic systems of adolescents as they pertain to negative learning attitudes, a multiplex recurrence network could be constructed for each individual. The structure of these networks can be characterized using different network measures that express for example the accessibility of certain states (probability of occurrence), but can also be used to quantify the structural similarities between the different layers in the network (see Hasselman, Reference Hasselman2022). Other layers could represent family members also participating in the study, or more abstract, a network representing a family-level network of relevant variables which interacts with the individual network through course-grained dynamics can be constructed. Network measures, such as the (cross-)clustering coefficient can be used to determine coupling dynamics between different layers (Donges et al., Reference Donges, Schultz, Marwan, Zou and Kurths2011; Feldhoff et al., Reference Feldhoff, Donner, Donges, Marwan and Kurths2012). Such measures can be used in statistical analyses to examine moderation, group differences etc., the main difference to traditional studies being that the data at its highest resolution would represent summaries of the dynamics of the personal experiences of an individual, not estimates inferred from the averaged dynamics of many individuals: “first analyze, then aggregate!”.