Adolescents’ Emotion System Dynamics: Network-based Analysis of Physiological and Emotional Experience

Abstract

An individual’s emotions system can be conceived of as a synchronized, coordinated, and/or emergent combination of physiology, experience, and behavioral components. Together, the interplay among these components produce emotional experiences through coordinated excitatory positive feedback (i.e., the mutual amplification of emotion concordance) and/or inhibitory negative feedback (i.e., the damping of emotion regulation) processes. Different system configurations produce differential psychophysiological reactivity profiles, and by implication, differential moment-to-moment emotional experience and long-term development. Applying dynamic systems models to second-by-second psychophysiological and experience time-series data collected from 130 adolescents (age 12.0 to 16.7 years) completing a social stress-inducing speech task, we describe the configuration of adolescents’ emotion systems, and examine how differences in the dynamic outputs of those systems (psychophysiological reactivity profile) are related to individual differences in trait anxiety. We found substantial heterogeneity in the coordination patterns of these adolescents. Some individuals’ emotion systems were characterized by negative feedback loops (emotion regulation processes), many by unidirectionally connected or independent components, and a few by positive feedback loops (emotion concordance). The reactivity dynamics of respiratory sinus arrhythmia were related to adolescents’ level of trait anxiety. Results highlight how dynamic systems models may contribute to our understanding of interindividual and developmental differences.

Emotional challenges are prominent in adolescence (Allen & Sheeber, 2008; Dahl, 2001; Hollenstein & Lougheed, 2013). This phase of the life span is characterized by increases in emotional intensity, such as more negativity and less positivity (Hardhavala et al., in press; Larson & Ham, 1993; Rosenblum & Lewis, 2003), as well as the perception or experience of more stressful, emotion-eliciting events (Seidman, Allen, Aber, Mitchell, & Feiman, 1994; Seiffge-Krenke, Aunola, & Nurmi, 2009). Many adolescent-typical behaviors, such as risk-taking, occur in the midst of high emotional intensity, and several adolescent-onset psychopathologies (e.g., anxiety and depression) are related to adolescents’ emotional processing. Most recent frameworks conceptualize these challenges for adolescents as struggles with the maturation of emotion regulation (Ahmed, Bittencourt-Hewitt, & Sebastian, 2015; Allen & Sheeber, 2008; Silk, Steinberg, & Morris, 2003).

Despite fervent recommendations to integrate study of regulatory processes and emotion generating processes (Gross, 2015), to investigate regulatory processes as emotions unfold in real time (Cole, Martin, & Dennis, 2004), and the growing viewpoint that emotion and regulation are not ontologically separable processes (Campos, Frankel, & Camras, 2004; Hollenstein & Lanteigne, 2018; Kappas, 2011; Thompson, Lewis, & Calkins, 2008), investigations into adolescent emotion regulation have focused on either the broad level of strategy use (e.g., suppression and reappraisal use across adolescence; Gullone, Hughes, Neville, & Tonge, 2010; Zimmermann, & Iwanski, 2014) or fine-grained investigation of the neural correlates of emotion regulation (e.g., Hare et al., 2008; Lewis, Lamm, Segalowitz, Seieben, & Zelazo, 2006; McRae et al., 2012). What is missing is the understanding of the moment-by-moment rise and fall of emotions in natural settings that reflects individual differences in emotion regulation (Hollenstein & Lanteigne, 2018).

The dynamic systems approach to development emphasizes the emergence of a structural organization from the dynamics of individuals’ ongoing experiences and transactions with their environment. Individuals are conceived as complex, open systems that self-organize to maintain equilibrium and/or adapt to a changing environment (Magnusson & Cairns, 1996; Molenaar, 2004; van Geert, 1998). Moment-by-moment changes, including the rise and fall of emotions, indicate how the system/individual adapts to perturbations and functions over the long term. Phenotypic endpoints or midpoints (e.g., individual differences in adolescence) both shape and are shaped by the structural organization of the system. In adolescence, the emotional structures that have been built from a decade of emotion dynamics manifest in ways that can persist into adulthood. Specifically, emotional processes in real time that have reciprocal interactions among system components (e.g., appraisals, physiology, expression) can contribute to development and emergence of long-lasting interindividual differences (traits, personality, psychopathology). In the present study, we introduce a novel approach for modeling emotion as a multi-component dynamic system and examining how that system reacts to perturbations. This study thus provides a unique opportunity to examine the linkage between the structural organization of a multi-component emotion system and the emergence of individual differences.

Emotional Concordance and Emotion Regulation

The emotion system has been conceptualized as a synchronized, coordinated, and/or emergent combination of psychophysiological, experience, and behavioral components (for a review, see Hollenstein & Lanteigne, 2014). Mutual amplification of system components contributes to rising of emotion, and here we characterize the degree of this coordination as indicating the extent of emotional concordance. Once activated, however, emotions do not continue to increase in intensity unabated over time but may be met with a countervailing damping force – characterized herein as emotion regulation. Here, elevations of sympathetic arousal (physiological) are met with appraisals (cognitive) that indicate lack of threat, inhibit further escalation of arousal, and prompt return towards baseline levels of arousal. Moment-by-moment, the components of the emotion system may simultaneously coordinate in excitatory positive feedback (i.e., the mutual amplification of concordance) and/or inhibitory negative feedback (i.e., the damping of emotion regulation) processes.

Advances in dynamic systems modeling provide new opportunities to examine individual differences in the real-time manifestation of emotional concordance and emotion regulation processes. In this study, we model and examine individual differences in the mutually excitatory and inhibitory coordination of psychophysiological and experience components of adolescents’ emotion systems. Specifically, we apply dynamic systems models to second-by-second time-series obtained during a social stress task to (a) describe coordinated dynamics of sympathetic nervous system activations (skin conductance level, SCL), parasympathetic nervous system activations (respiratory sinus arrhythmia, RSA), and experience of emotional distress, and (b) examine how variation in those dynamics is related to individual differences in trait anxiety. We consider the social stress task an emotion regulation task for adolescents, where positive feedback loops between components indicate emotion concordance (of anxiety), and negative feedback loops indicate emotion regulation. We chose the dynamic system approach which views emotion as an emergent, self-organizing process (Coan, 2010; Lewis, 2000) because it can quantify the temporal relation between components, especially whether it is positive or negative, so that positive and feedback loops can be identified. This is in contrast to the latent approach, which models emotion as a latent factor that determines component activation, assumes the components are interchangeable, and therefore cannot pinpoint whether these components form positive or negative feedback loops (Coan, 2010). In doing so, we propel understanding of the systems that drive the moment-by-moment rise and fall of emotions, and how differences in the structure of those systems may be related to a wide variety of individual difference factors, including the development of psychopathology (Hollenstein & Lanteigne, 2018).

Consider, for example, the three component systems shown in the top portion of Figure 1. Each of the two panels depicts an emotion system that consists of three change components (depicted as circles): moment-to-moment changes in sympathetic nervous systems activity (ΔSCL), moment-to-moment changes in parasympathetic nervous system activity (ΔRSA), and moment-to-moment changes in experience (Δdistress). Directed (causal) relations among those components are also shown: green arrows indicate positive relations, red arrows indicate negative relations, and lack of an arrow indicates no relation. Exogenous inputs that perturb the system (e.g., ongoing changes in the biopsychosociocultural environment; Ford & Lerner, 1992) are shown by dotted arrows. Although similar in overall structure, these two systems (and the differential equations used to describe them) produce very different emotion dynamics. The system in Figure 1a is characterized by a positive (excitatory) feedback loop, wherein elevations in the individual’s sympathetic nervous system (ΔSCL) prompt elevation in distress (green arrow labeled +0.35), which in turn prompt further elevation in sympathetic arousal (green arrow labeled +0.5) that again elevates distress. Conceptually, the bidirectionality of the positive temporal relations is indicative of mutual excitation – emotional concordance. Over time, inputs into the system escalate towards higher levels. In contrast, the system in Figure 1b is characterized by a negative (inhibitory) feedback loop. Elevations in the individual’s sympathetic nervous system (ΔSCL) prompt elevation in distress (green arrow labeled +0.35). For this individual, however, the change in distress prompts the decrease in sympathetic arousal (red arrow labeled −0.5). The combination of a positive temporal relation and a negative temporal relation is indicative of inhibition – emotion regulation – as negative feedback processes overwhelm the elevations. Over time, inputs into the system decay.

Figure 1.

Conceptual illustration of two three-component emotion system (sympathetic, parasympathetic, and distress). The three components, which are sympathetic, parasympathetic, and distress, are represented by ΔSCL, ΔRSA, and Δdistress, respectively. ΔSCL: first difference of galvanic skin response. ΔRSA: first difference of respiratory sinus arrhythmia. Δdistress: first difference of self-rated distress. Panel A shows a positive feedback (excitatory loop) between change in sympathetic arousal and change in distress, and Panel B shows a negative feedback (inhibitory loop) between change in sympathetic arousal and change in distress. In both panels, the green arrows indicate a positive temporal relation, and the red ones indicate a negative temporal relation; the dashed arrows indicate lag-1 relations and the solid arrows indicate contemporaneous relations. Width of the arrow indicate the absolute value of the temporal relations, which are also labeled on the side of the arrow. The dotted arrows indicate the impulse going into each component. Panel C and D are the time profiles generated by impulse response of Panel A and B, respectively. The reactivity is indicated by the horizontal asymptote (marked by the horizontal dashed line) in each cell of Panel C, e.g., R_(SCL→SCL) = 5.45.

Differences in the structure of this three-component system (e.g., the presence of mutually excitatory or inhibitory loops) imply differences in both individuals’ moment-to-moment experience of emotion, and their long-term development. Differences in experience may be quantified to represent system reactivity – the accumulated dynamic response to perturbation (e.g., a stressful event). From the system descriptions shown in the top portion of Figure 1 (strength and direction of the connections among components), it is possible to examine how the system operates in different conditions. Formally, the system is subjected to an impulse response analysis (Lütkepohl, 2005), wherein an “impulse” (perturbation) given to one component of the system filters through the network of mutually excitatory and/or inhibitory feedback loops to settle at an equilibrium. The equilibrium is a new setpoint due to the allostasis process (Sterling & Eyer, 1988; Schulkin, 2003), the level of which is an index of an individual’s system reactivity. Theoretically (and mathematically), systems characterized by positive feedback loops (mutually excitatory components) and emotional concordance processes will settle at higher equilibria, while systems characterized by negative feedback loops (inhibitory components) and emotion regulation processes will settle at lower equilibria. Stated differently, different system configurations produce differential emotion reactivity, with higher reactivity indicative of greater emotional concordance and lower reactivity indicative of better emotion regulation.

To illustrate, the two collections (matrices) of trajectories shown in the bottom portion of Figure 1 summarize the reactivity of the two systems described above. Each row in Panel C and D represent when the perturbation is sent to one node (SCL, RSA, and distress, respectively), and each column represents how each node responds to the perturbation and settles to equilibrium. For example, the top right cell (SCL→distress) of Panel C indicates the response of distress when the SCL is perturbed, and because these two components formed a positive feedback loop (Panel A), the system excites to and settles at equilibrium (reactivity) of distress = 4.24. These dynamics are quite different from those seen in the top right cell (SCL→distress) of Panel D. Here, in the system characterized by its negative feedback loop (Panel B) perturbation of the SCL filters through a negative (inhibitory) feedback loop to settle at equilibrium (reactivity) of distress = 0.53. Thus, the system configurations produce differential emotion dynamics – and by implication differential moment-to-moment emotional experience.

Emotion System Dynamics and Trait Anxiety

Substantial evidence suggests a strong relation between the level of trait anxiety and emotion system dynamics (Beauchaine, 2001; Carthy, Rienman, Apter, & Gross, 2009; Schneiders et al., 2006; Tan et al., 2012). Among children, for example, a higher level of internalizing symptoms (e.g., anxiety) is associated with greater parasympathetic reactivity to sadness- and fear-inducing film clips (Fortunato, Gatzke-Kopp, & Ram, 2012), as well as to standardized field-laboratory stressors (Boyce et al., 2001). Similarly, among adolescents, anxiety is associated with higher heart rate and less fluctuating heart rate, indicating the deficiency in the parasympathetic system to modulate heart rate (Monk et al., 2001). As noted earlier, phenotypic endpoints such as high trait anxiety or psychopathology may be the result of repetition and the development of long-lasting structure (Lewis, 2000). Emotion systems that organize around the higher equilibria produced by a reactive system characterized by positive feedback loops are likely to lead to a high trait anxiety (or other phenotypic) endpoint (i.e., development of psychopathology; Beauchaine, 2001). In contrast, systems that organize around the lower equilibria produced by negative feedback loops that facilitate regulation are likely to lead to a low trait anxiety endpoint. Referencing Figure 1, an adolescent with the system depicted in Panels A and C is expected to have higher level of trait anxiety than an adolescent with the system depicted in Panels B and D. Viewed cross-sectionally, we expect that differences in the configuration of adolescents’ three-component emotion system will be related to between-person characteristics conceived and operationalized as phenotypes/traits that are measured independently from the dynamic model (e.g., via questionnaire).

The Present Study

In the present study, we examine relations between the configuration of adolescents’ emotion systems (psychophysiology, experience) and trait anxiety. In doing so, we construct dynamic systems models that explicitly integrate emotion generating and emotion regulating processes (Gross, 2015) – positive (mutually excitatory) and negative (inhibitory) feedback loops – to investigate regulatory processes as emotions unfold in real time (Cole, Martin, & Dennis, 2004). Taking the view that emotion and emotion regulation are not ontologically separable processes (Campos, Frankel, & Camras, 2004; Hollenstein & Lanteigne, in press; Kappas, 2011; Thompson, Lewis, & Calkins, 2008), we take a “bottom-up” approach, first using exploratory multivariate time-series methods and second-by-second time-series data collected during a laboratory task to describe the structure of adolescents’ emotion systems, and then examining how differences in the dynamic outputs of those systems are related to individual differences in trait anxiety. Extending prior work based on analysis of between-person (Evers et al., 2014) and cross-variable correlations (Butler, Gross, & Barnard, 2014; Mauss, Levenson, McCarter, Wilhelm, & Gross, 2005), our person-specific network modeling approach and findings forwarded here contribute to understanding of adolescents’ real-time emotional processes and their implications for long-term development. The strength and novelty of this approach includes (a) direct integration of emotion components into a model that includes excitatory or inhibitory feedback loops, (b) estimation of the emotion process as a person-specific network that reflects the unique within-person process, and (c) extraction of the psychophysiological reactivity profile based on the person-specific network and examination of its linkage to long-term development.

Method

Participants

Participants were typically-developing adolescents (N = 183), age 12.0 to 16.9 years (M = 13.60, SD = 1.12) in a mid-sized city in Southern Ontario, Canada, that were recruited via telephone from a database of families that volunteered to be contacted about participating in research studies being done developmental psychologists at a local university. All study protocols were approved by the Queen’s University (IRB# PSYC-082–08, study title: Individual Differences in Psychophysiological Responses While Regulating Emotion in Adolescence).

The present analysis makes use of data from the subsample of adolescents with complete psychophysiological and self-report data. These participants (N = 130, 52% female) were age 12.00 to 16.67 years (M = 13.49, SD = 1.11); identified as White (73%), Unknown (12%), Other (3%), South Asian (<1%), Black (<1%), Chinese (<1%), Latin American (<1%), and multiracial (9%); and were fluent in English (with 95% indicating English as their first language). The analysis sample did not differ from the rest of the sample in terms of age, t(180) = −1.84, p = .07; gender, χ²(1)= .03, p = .87; or ethnicity, χ²(7)= 9.6, p = .21. Thus, the analysis sample is considered representative of full sample.

Laboratory Procedure

Participants came to a university-based laboratory where they were seated in a comfortable chair in an observation room equipped with obscured video cameras that were monitored from an adjacent room. After parental consent and participant assent were obtained, parents left the room for the duration of the study and a female experimenter provided instructions for the participant to fill out questionnaires on a computer (e.g., demographics, socioemotional functioning). Next, the experimenter explained about and applied electrodes to participants’ chest and fingers to measure sympathetic and parasympathetic nervous system activity. Once situated and comfortable with the physiological recording set-up, participants were asked to complete a series of tasks, (0) Self-Reported Experience (baseline, approximately 1 minute): participants reported on their current physical experience and emotional state; (1) Resting Baseline (3 minutes): sitting quietly with no stimulation; (2) Speech Task (3 minutes): participants were instructed to perform a speech on a topic of their choosing, as if to their class at school, and in front of the experimenter; (3) Self-Reported Experience (during stress, approximately 1 minute): participants reported on their physical experience and emotional state during the speech task; (4) Recovery: sitting quietly with no stimulation; (5) Social exclusion induction via Cyberball, self-reported emotions, and recovery; (6) Video-mediated recall of the Speech Task (3 minutes): participants watched the video of their speech performance and used a lever to provide a continuous and dynamic report of their distress level during the Speech Task. After finishing the tasks, participants were debriefed, asked to complete an exit form, and thanked for their participation. Participants received a $20 gift card to a popular bookstore as compensation.

Measures

Parasympathetic nervous system activity.

Respiratory sinus arrhythmia (RSA) is an indicator of parasympathetic nervous system activity. Electrocardiography was recorded from three disposables, pre-jelled, electrodes placed over the distal right collar bone, lower left rib, and lower right rib, connected to battery pack (TEL100M-C; BioPac Systems, 2017) and amplified (MP150; BioPac Systems, 2016) for acquisition using AcqKnowledge 3.9 software (BioPac, 2009). After being checked and corrected by trained research assistants (AcqKnowledge 4.1 software; BioPac, 2016), the cleaned data were passed through a Fast Fourier Transformation (FFT) to obtain power in 0.12 to 0.40 frequency range to obtain a measure of respiratory sinus arrhythmia (RSA), an indicator of parasympathetic nervous system activity (Allen, Chambers, & Towers, 2007). Following previous techniques for RSA time series analysis (e.g., Gates, Gatzke-Kopp, Sandsten, & Blandon, 2015), second-by-second RSA time series were created by estimating RSA in overlapping 16-second windows, the first of which spanned from −8 second before the start of the speech (end of instructions) to +8 seconds into the speech, with the second window spanning from −7 seconds to +9 seconds, and so on up to a final window that spanned from +164 seconds to 180 seconds.

Sympathetic nervous system activity.

Sympathetic nervous system activity was measured as skin conductance level (SCL), specifically as the level of conductance (micromhos) between two SS3 electrodermal response transducers that were attached to the tips of participants’ ring and middle fingers on their non-dominant hand. Continuous recordings at 200 Hz were analyzed using Acqknowledge 4.1 software (BioPac, 2009), visually inspected by trained RAs to identify and remove artifacts, and quantified as the mean level within each second to obtain a SCL time series that was aligned with the RSA time-series.

Experience of distress.

Participants’ experience of distress during the speech task was measured using a video-mediated recall procedure. After completing the stressor tasks, participants watched the video of their speech performance. As the video played, they were asked to indicate the extent to which they felt distressed during the speech using a bipolar rating lever (Gottman & Levenson, 1985), moving the lever up and down to indicate how they felt during the time shown on the video. The lever’s position, between the extremes of calm (= 0) and distressed (= 100) were continuously recorded with a physiological channel (AcqKnowledge 3.9 software) and aggregated and aligned with the other second-by-second time series (using AcqKnowledge 4.1, BioPac, 2009).

Trait Anxiety.

Individual differences in trait anxiety were measured as part of the initially completed computer-based questionnaires using the Beck Anxiety Inventory (BAI; Beck & Steer, 1993), a well-established 21-item questionnaire (Cronbach’s α = .91) wherein individuals indicate on a 4-point scale (0=not at all; 3= severely, I could barely stand it) the extent to which they experience a variety of cognitive and somatic symptoms of anxiety (e.g., unable to relax, fear of worst happening, dizzy or lightheaded, heart pounding/racing).

Data Analysis

There were four steps in the data analysis (see Figure 2 for illustration of Steps 1 to 3). In Step 1, data were pre-processed into a form suitable for time-series analysis and construction of dynamic models. In Step 2, a discrete time state-space modeling approach, unified structural equation modeling (uSEM; Gates et al., 2010; Molenaar & Nesselroade, 2015), was used to construct person-specific networks describing the configuration of the three-component emotion system. In Step 3, these networks were used to compute an impulse response analysis matrix (iRAM) of dynamic reactivity. In Step 4, the Beck Anxiety Inventory (BAI) scores were regressed on the measures of dynamic reactivity. Estimation of person-specific models and computation of iRAM was conducted using the pompom package in R (Yang, Ram, & Molenaar, 2018). An implementation tutorial can be found in the Supplementary Materials.

Figure 2.

Illustration of the analytical steps using one participant’s data. The pre-processed multivariate time-series of physiological and emotional experience data (Panel A), temporal dynamics visualized as 3-component network (Panel B), and the iRAM indicated by the horizontal asymptote in each cell (Panel C). ΔSCL: first difference of skin conductance level. ΔRSA: first difference of respiratory sinus arrhythmia. Δdistress: first difference of self-rated distress. R_(SCL→RSA): reactivity of RSA when SCL is given an impulse.

Step 1: Data Preparation.

The raw time-series data of SCL, RSA, and distress showed evidence of non-stationarity (autocorrelation close to or exceeding 1). Therefore, after scaling so that they had mean of 0 and standard deviation of 1, the time series data was recast as a stationary time series of change score (i.e., first differences),

$$\mathrm{\Delta }y\left(t\right)=y\left(t\right)-y\left(t-1\right)$$ (1)

where y(t) is a vector of the scaled time-series, SCL, RSA, and experience of distress, $y\left(t\right)=\left(\begin{array}{c}SCL\left(t\right)\\ RSA\left(t\right)\\ distress\left(t\right)\end{array}\right)$, and $\mathrm{\Delta }y\left(t\right)=\left(\begin{array}{c}SCL\left(t\right)\\ RSA\left(t\right)\\ distress\left(t\right)\end{array}\right)$

An example of trivariate time series of one individual’s change scores are shown in Figure 2a.

Step 2: Construction of Person-specific Networks.

Each individual’s trivariate time-series of change scores was then modeled as a three-node dynamic network using a unified Structural Equation Model (uSEM, Gates et al., 2010). In brief, the mean-centered multivariate observed time-series Δy(t) is modeled as the output of a mean-centered latent variable time series Δη(t), that is scaled by a factor loading matrix Λ, and a time-series of measurement errors ε(t),

$$\mathrm{\Delta }y\left(t\right)= \Lambda \mathrm{\Delta }\eta \left(t\right)+ \epsilon \left(t\right)$$ (2)

The temporal relations among the set of latent constructs in Δη(t) (the circles in Figure 2b) are then modeled as

$$\mathrm{\Delta }\eta \left(t\right)=\text{A}\mathrm{\Delta }\eta \left(t\right)+{\Phi }_1\mathrm{\Delta }\eta \left(t-1\right)+\zeta \left(t\right)$$ (3)

where Δη(t − 1) is a vector of latent change scores from the prior occasion; A is a matrix of regression parameters that describe the contemporaneous relations among the latent variables (solid arrows in Figure 2b), Φ₁ is a matrix of regression parameters that describe the lag-1 relations (auto- and cross-regressions) among the latent variables (dashed arrows in Figure 2b), and ζ(t) is a multivariate series of on-going dynamic impulses that capture any and all exogenous inputs from the biopsychosociocultural environment (dotted arrows in Figure 2b). Together, the contemporaneous relations in A and auto- and cross-regressive relations in Φ₁ indicate the causal influences among variables through which exogeneous input is processed and diffuses (i.e., dynamic regulation). At the practical level, the uSEM model is estimated using an iterative search process wherein a series of models are constructed and tested for improvements in fit. At each step, Lagrange Multiplier tests (modification indices; Sörbom, 1989) are used to select the path that facilitates maximum improvement in fit. After this element is freed, the model re-estimated, and a new set of modification indices calculated – iteratively adding paths until further addition does not significantly improve model fit. The model expansion was constrained so that only the A and Φ₁ blocks of the model parameter matrix were freed, thus keeping the time-series structure of the model intact. Once the person-specific models were obtained, Φ₁ and A matrices were extracted and drawn as network graphs. Conceptually, the resulting network describes the mutual excitatory and inhibitory processes driving second-to-second changes in an individuals’ experience and psychophysiological activity during the Speech Task. A sample network is shown in Figure 2b, where, for example, positive ΔSCL leads to negative ΔRSA at the next observation (red dashed arrow from ΔSCL to ΔRSA), and the negative ΔRSA in turn leads to negative ΔSCL at the next observation (green dashed arrow from ΔRSA to ΔSCL). These two arrows (which are two temporal relations) form an inhibitory feedback loop between ΔSCL and ΔRSA.

Step 3: Impulse Response Analysis Matrix (iRAM).

Impulse response analysis was then used to quantify the dynamics embedded in the system. The impulse is given to each node respectively, and the emergent response of the system is derived from the configuration of the dynamic network obtained from uSEM. All responses will eventually settle at an equilibrium, the asymptotic level of which are measures of (or a profile of) the reactivity (R) of the multicomponent system.

Specific to the three-component networks examined in this analysis, impulses are given to the three nodes separately, each time tracking changes in all three nodes. Formally, the impulse response for a uSEM is derived from Equation 3 as a one step ahead forecasting process through conversion into a vector auto-regression or moving average model (see Amisano & Giannini, 1996; Gates et al., 2010). The system is set in motion with an initial impulse vector, ζ(0), and latent states are calculated for t = 0 as

$$\mathrm{\Delta }\eta \left(0\right)={\left(\text{I}-\text{A}\right)}^{-1}\zeta \left(0\right)$$ (4)

(to accommodate the contemporaneous relations among latent states) and for each subsequent t as

$$\mathrm{\Delta }\eta \left(t\right)=\left[{\left(\text{I}-\text{A}\right)}^{-1}{\Phi }_1\right]\mathrm{\Delta }\eta \left(t-1\right)={\left[{\left(\text{I}-\text{A}\right)}^{-1}{\Phi }_1\right]}^t{\left(\text{I}-\text{A}\right)}^{-1}\zeta \left(0\right)$$ (5)

For example, an impulse sent into one specific node (i.e., ΔSCL), $\zeta \left(0\right)=\left(\begin{array}{c}1\\ 0\\ 0\end{array}\right)$, $\zeta \left(t\right)=\left(\begin{array}{c}0\\ 0\\ 0\end{array}\right)$ when t > 0. Let [(I − A)⁻¹ Φ₁] = Φ′, then the equilibrium can be computed from Equation 5 as

$$\eta \left(\infty \right)={\sum }_{t=1}^{\infty }\mathrm{\Delta }\eta \left(t\right)=\left[{\sum }_{t=1}^{\infty }\Phi {\prime }^t\right]{\left(\text{I}-\text{A}\right)}^{-1}\zeta \left(0\right)=\left[\Phi \prime {\left(\mathit{\text{I}}-\Phi \prime \right)}^{-1}\right]{\left(\text{I}-\text{A}\right)}^{-1}\zeta \left(0\right)$$ (6)

We denote the expressions of system reactivity as R_{(node 1→node 2)}, meaning reactivity of node 2 when node 1 receives an impulse. After each node is given an impulse, we obtain 3 × 3 matrix of trajectories as shown in Figure 2c. For example, the 3 × 3 iRAM of reactivity scores corresponding to Panel C is $\left[\begin{array}{ccc}{R}_{\left(SCL\to SCL\right)}& R_{\left(SCL\to RSA\right)}& R_{\left(SCL\to distress\right)}\\ R_{\left(RSA\to SCL\right)}& R_{\left(RSA\to RSA\right)}& R_{\left(RSA\to distress\right)}\\ R_{\left(distress\to SCL\right)}& R_{\left(distress\to RSA\right)}& R_{\left(distress\to distress\right)}\end{array}\right]=\left[\begin{array}{ccc}1.98& -7.14& 0\\ 1.45& 0.93& 0\\ 0& 0& 0.31\end{array}\right]$. Importantly from a systems perspective, the impulse given to node 1 flows through all connected nodes so that the final equilibrium of node 2 is based on the structure of the entire network, and all the positive and negative feedback loops embedded in that structure.

Step 4: Associations Between iRAM and Trait Anxiety.

In the final step, we regressed the individual differences measure of interest, trait anxiety (as measured by Beck Anxiety Inventory) on the 9 reactivity scores in the iRAM,

$$BA{I}_i={\beta }_0+{\beta }_1{R}_{\left(\text{SCL}\to \text{SCL}\right)\text{i}}+{\beta }_2{R}_{\left(\text{SCL}\to \text{RSA}\right)\text{i}}+{\beta }_3{R}_{\left(\text{SCL}\to \text{distress}\right)\text{i}}+{\beta }_4{R}_{\left(\text{RSA}\to \text{SCL}\right)\text{i}}+{\beta }_5{R}_{\left(\text{RSA}\to \text{RSA}\right)\text{i}}+{\beta }_6{R}_{\left(\text{RSA}\to \text{distress}\right)\text{i}}+{\beta }_7{R}_{\left(\text{distress}\to \text{SCL}\right)\text{i}}+{\beta }_8{R}_{\left(\text{distress}\to \text{RSA}\right)\text{i}}+{\beta }_9{R}_{\left(\text{distress}\to \text{distress}\right)\text{i}}+{\epsilon }_i$$ (7)

where β₁ to β₉ indicate the extent to which individual differences in each aspect of the system dynamics are independently related to individual differences in trait anxiety.

In follow-up analyses used to check robustness of the relations some additional models were examined. Individuals’ mean level of SCL, RSA, and distress during the task were included as covariates to further isolate the added value of the dynamic reactivity scores for explaining between-person differences in trait anxiety. Baseline and baseline-to-task change scores were regressed on the trait anxiety scores to determine how the new dynamic reactivity approach compared to more traditional change score approach. Models were fit in R (R Development Core Team, 2008), with statistical significance of the final models assessed at p < .05.

Results

Emotional Concordance and Emotion Regulation

Person-specific networks were derived from the trivariate time-series using uSEM. Of the total 130 network models, 119 fit the data well, as indicated by at least three of the following fit criteria: RMSEA ≤ 0.08, SRMR ≤ 0.08, CFI ≥ 0.95, NNFI ≥ 0.95. Models from 11 individuals did not fit well and were set aside. Mostly, these individuals’ time-series had very low variance – suggesting that they were not affected by the social-stress inducing speech task. The lack of perturbation precludes study of how their emotion systems are configured.

There was substantial heterogeneity in the coordination patterns of these adolescents – some individuals’ systems were characterized by negative feedback loops (emotion regulation processes), many by unidirectionally connected or independent components, and a few by positive feedback loops (emotion concordance). Specifically, of the 119 viable networks, 29 networks (24.4%) contained at least one negative feedback loop (regulatory process), 56 networks (47.1%) had only unidirectional temporal relation, 32 networks (26.9%) had no cross-component temporal relation (completely independent components), and 2 networks (1.7%) contained at least one positive feedback loop (concordant process). Specifically looking at connections between psychophysiology and experience components, 12 networks (10.1%) contained at least one cross-system feedback loop, and another 33 networks (27.7%) contained unidirectional cross-system connections.

The structure of each individual’s system (lagged and contemporaneous relations embedded in the networks) was extracted and used in the impulse response analysis to calculate person-specific reactivity scores. As seen in Table 1, system reactivity scores (R_{(node1→node2)}) derived from the iRAM analyses ranged between −26.05 and +20.69. Of note, average R_(SCL→SCL), R_(RSA→RSA), and R_{(distress→distress)} were all above zero, indicating that “carryover” or “emotion inertia” is an important part of individuals’ emotion system dynamics (see e.g., Kuppens, Allen, & Sheeber, 2010).

Table 1.

Descriptives and correlations of emotion system dynamics (reactivity) and trait anxiety measures

Variable	1	2	3	4	5	6	7	8	9	10	11	12	13
1. R(SCL→SCL)	-
2. R(SCL→RSA)	0.17	-
3. R(SCL→distress)	−0.21	−0.16	-
4. R(RSA→SCL)	−0.32	−0.32	0.22	-
5. R(RSA →RSA)	0.17	0.23	−0.02	−0.27	-
6. R(RSA →distress)	−0.11	−0.11	0.19	0.13	−0.05	-
7. R(distress →SCL)	0.02	−0.07	−0.04	0.03	−0.16	−0.15	-
8. R(distress →RSA)	−0.04	−0.13	0.04	0.04	−0.23	−0.2	0.67	-
9. R(distress → distress)	−0.09	−0.01	0.07	0.06	−0.16	0.06	−0.2	−0.1	-
10. BAI	−0.01	−0.04	−0.02	0.04	0.16	−0.11	0.09	0.09	0.15	-
11. Task SCL	−0.02	−0.12	0.04	−0.02	0.15	−0.01	−0.02	0.05	−0.15	−0.11	-
12. Task RSA	0.02	−0.07	−0.03	0.24	−0.1	−0.04	0.06	0.03	−0.02	0.07	0	-
13. Task distress	0.09	0.07	0.04	−0.12	0.09	−0.03	−0.06	−0.11	0.19	0.24	−0.13	−0.1	-
Mean	5.51	−1.47	−0.28	0.22	4.73	0.02	0.05	−0.02	0.87	0.52	6.23	−7.32	44.85
SD	2.66	5.24	1.75	1.28	1.38	0.51	0.60	1.25	0.59	0.42	2.28	0.76	22.13
Min	0.53	−26.05	−11.24	−5.78	0.93	−1.57	−4.56	−9.50	−0.28	0.00	1.17	−10.00	−3.83
Max	14.05	20.69	4.40	3.38	7.90	2.56	2.69	5.36	2.89	1.57	14.81	−2.56	94.52

Note: N = 119. SCL: Skin conductance level. RSA: Respiratory sinus arrhythmia. R_(SCL→RSA): Reactivity of RSA when SCL is given an impulse. BAI: Beck Anxiety Inventory. Task SCL (RSA, distress): mean level of SCL (RSA, distress) during the speech task.

Emotion Dynamics and Trait Anxiety

Results from regression models examining how between-person differences in the emotion system reactivity profile were associated with between-person differences in trait anxiety are shown in Table 2. Two significant relations emerged. Greater reactivity of RSA to perturbation of RSA, R_(RSA→RSA), was associated with higher levels of trait anxiety (β₅ = 0.08, p = 0.01). Similarly, greater reactivity of experience of distress to perturbation of distress, R_{(distress→distress)}, was associated with higher levels of trait anxiety (β₉ = 0.16, p = 0.02).

Table 2.

Results from regression examining relations between individual differences in system dynamics reactivity and trait anxiety

	BAI
Predictor Variable	Estimate (SE)	Estimate (SE)
Intercept, β0	0.001 (0.191)	0.324 (0.457)
R(SCL→SCL), β1	−0.002 (0.016)	−0.006 (0.698)
R(SCL→RSA), β2	−0.006 (0.009)	0.03 (0.427)
R(SCL→distress), β3	−0.01 (0.023)	0.056 (0.567)
R(RSA→SCL), β4	0.027 (0.034)	−0.009 (0.366)
R(RSA →RSA), β5	0.079* (0.03)	0.081** (0.009)
R(RSA →distress), β6	−0.076 (0.079)	0.036 (0.536)
R(distress →SCL), β7	0.068 (0.086)	−0.013 (0.563)
R(distress →RSA), β8	0.025 (0.042)	−0.081 (0.335)
R(distress → distress), β9	0.156* (0.068)	0.113 (0.099)
Task SCL, β10	--	−0.021 (0.214)
Task RSA, β11	--	0.043 (0.398)
Task distress, β12	--	0.004* (0.03)
Multiple R2	0.107	0.170
Adjusted R2	0.032	0.075

Note: N = 119. SE = standard error. BAI = Beck Anxiety Inventory. R_(SCL→RSA): Reactivity of RSA when SCL is given an impulse. Task SCL: mean level of SCL during the speech task.

^*p <0.05,

^**p<0.01.

Follow-up analyses controlling for differences in mean level of SCL, RSA, and distress during the task, indicated robustness of the relation between trait anxiety and system reactivity of RSA, with the R_(RSA→RSA) scores adding value for prediction of differences in trait anxiety (β₅ = 0.08, p = 0.009). Probing further, we checked whether the differences in trait anxiety were related to more traditional measures of task-level changes in RSA change. They were not. Individual differences in trait anxiety were not related to baseline RSA (β = 0.04, p = 0.59) or traditional RSA change scores, calculated as the difference between mean level of RSA during the task and the baseline level of RSA (β = 0.02, p = 0.75). In sum, we found that dynamics of the three-component dynamic system, specifically individual differences in how perturbations to RSA flowed back to RSA (R_(RSA→RSA)) are related to adolescents’ long-term (i.e., trait) emotional experience.

Discussion

The present study used a bottom-up dynamic system modeling approach to describe individual differences in the dynamics driving moment-to-moment changes in adolescents’ physiological activity and cognitive appraisals during a social stress-inducing task. Results indicated substantial heterogeneity in the coordination patterns – with some adolescents’ emotion systems characterized by negative feedback loops (emotion regulation processes), many by unidirectionally connected or independent components, and a few by positive feedback loops (emotion concordance). Quantifying the system dynamics using a new impulse response analysis approach, we found that differences in the dynamics surrounding reactivity of RSA were related to adolescents’ level of trait anxiety, above and beyond mean levels of SCL, RSA, and distress. The modeling approach and findings contribute to both how real-time emotion processes are conceptualized, and what we know about how multicomponent feedback processes contribute to individual differences and long-term development.

Emotional Concordance and Emotion Regulation

Using a bottom-up time-series models approach to describe the person-specific temporal dynamics of a three-component (SCL, RSA, distress) emotion system provided an opportunity to explicitly integrate emotion generating and emotion regulation processes (Gross, 2015) through accommodation of both positive (mutually excitatory) and negative (inhibitory) feedback loops. The person-specific networks showed heterogeneity in configurations of the emotion system, which provided evidence that the emotion generating and regulating processes are person-specific. Different system configurations may manifest or are optimized to support different kinds of dynamics. Positive feedback loops, for example, facilitate coordinated elevation and escalation in multiple components of the system – concordances that promote action readiness in threatening situations. Only 2 individuals (1.7%) in this study had networks characterized by strong positive feedback loops. One interpretation of this result is that the task demands did not pull for the fight or flight responses facilitated by positive feedback loops. Negative feedback loops, in contrast, facilitate coordinated stabilization – regulation processes that facilitate return to equilibrium. One interpretation of the result that 29 individuals’ networks (24.4%) contained at least one negative feedback loop is that the task demands pulled for the regulatory processes that facilitate stability and speech performance. In sum, the prevalence of the different network configurations may be a function of the context in which the emotion dynamics were measured. Positive feedback networks, and concordance, are notably difficult to capture in less than extreme situations (Hollenstein & Lanteigne, 2014).

Evers and colleagues (2014) proposed a dual-process theory stating the autonomic system and the reflective system (e.g., experience and behavior) are independent and provided between-person evidence to support this theory. Of the 119 individual networks obtained here, 45 (37.8%) had links that suggested second-by-second interdependency between the physiology and experience components (i.e., links between the sympathetic-distress or parasympathetic-distress nodes). These network configurations suggest that, for at least a substantial portion of adolescents, the autonomic and reflective components of the emotion system were not distinct. Because the phenomenon of interest is within individual, the present results do not appear to support the dual-process theory.

Alternatively, it may be that the adolescent emotion system is not yet at adult levels and that the separation of the autonomic and reflective systems has not been fully realized. In general, the novelty of the present analyses combined with the degree of physical and emotional changes occurring during adolescence (Hollenstein & Lougheed, 2013) make it challenging to generalize these results beyond this age period. The predominance of heterogeneity in our results may reflect the diversity of developmental progress in the maturation of emotion systems. Our sample ranged in age from 12 to 17 years, spanning the period of the most pronounced and rapid change. Neuroscience-informed models of adolescent development highlight the cascade of emotion-related changes across adolescence. Steinberg’s (2008) dual-process model highlights the rapid maturation of the limbic system in early to mid-adolescence that is not met with sufficient self-regulatory control until later in adolescence. The more recent imbalance model goes beyond these dual systems to articulate the patterns of connectivity and efficiency within and across cortical and subcortical networks (Casey, Galavan, & Somerville, 2017). Thus, without longitudinal and age comparisons, it is uncertain to what degree the dynamic patterns we revealed are developmentally stable.

Emotion System Dynamics and Trait Anxiety

We found reactivity of RSA is associated with trait anxiety. In general, individual differences in RSA have been found across a wide range of psychopathologies and across age (Beauchaine, 2015). RSA reactivity in particular may reflect prefrontal cortex functioning related to self-regulation and thus has been suggested as a marker of real-time emotion processing (Thayer, Hansen, Saus-Rose, & Helge Johnsen, 2009). Thus, the relation between RSA dynamics and trait anxiety may reflect individual differences in adolescents’ capacity for self-regulation as well. Results align with previous literature showing relations between extent of RSA withdrawal, rather than baseline levels of RSA, is related to level of internalizing (including anxiety) symptoms during childhood (Fortunato et al., 2012). Interpreted with respect to the suggestion that RSA reactivity is associated with risk for the development of psychopathology (Beauchaine, 2001), the results of this study suggest that those risks may carry into adolescence. Emergence of trait anxiety may result from an emotion system that canalizes around the higher equilibria produced by a reactive system that lacks the negative feedback loops that facilitate regulation. Generalizing from the second-to-second time-scale out to the hour-to-hour timescale, this interpretation is consistent with findings from experience sampling studies that indicate greater reactivity of negative emotions and less effective use of regulation strategies is associated with higher level of internalizing symptoms (Silk et al., 2003; Schneiders et al., 2006; Tan et al., 2012; Carthy et al., 2009). Similarity of relations across time-scales of inquiry highlights the possibility that person-specific network modeling approaches like the one used here can be used to unpack how the configuration of the system driving the moment-to-moment dynamics of individuals’ emotions contributes to interindividual differences in development as moments accumulate into hour, weeks, and months of emotional experience.

Limitations and Future Directions

The results of this study must be interpreted with respect to some limitations in design and implementation. The participants of this study are somewhat homogeneous in ethnicity, age, and levels of trait anxiety, with only 26.6% of the sample having a moderate to severe level of anxiety (Beck & Steer, 1993). Before generalizing to the larger population, it will be useful to engage with other populations, including individuals with more severe psychopathology, and with both younger and older individuals. Here, measurement of the individual differences variables was done contemporaneously with the measurement of the moment-to-moment dynamics (in the same session). Our contention is that the moment-to-moment dynamics contribute to emergence and development of trait-level or age-specific differences. However, the links between the short-term dynamics and the long-term changes requires repeated measurement of both the dynamics and the individual differences measures. Measurement burst designs, wherein participants complete the same protocol at 6-month or 1-year intervals for many years would be particularly useful (see Ram & Gerstorf, 2009).

Measurement of individuals’ emotion system configuration and dynamics was done using second-by-second time-series data collected during a single stress-induction task. As such, the derived system configurations may be somewhat situation specific. It is not yet clear how much the context matters, but there is strong evidence that physiological reactivity does differ by tasks (e.g., Fortunato et al., 2012). Future studies should examine if and how individuals’ system configurations differ across contexts, and in which contexts the individual differences in network configuration are most informative.

The analytical method used in this paper to describe and quantify system dynamics is limited both in scope – only covering three components of the emotion system (SCL, RSA, distress) – and in time-scale – only examining temporal relations that manifested contemporaneously (in the same second) or at a 1-second lag. The rate of change of each emotion system component may not be the same. SCL, for example, may respond more rapidly to perturbations than self-reported appraisals. Following from studies that have used within-subject correlations across a +/− 10 second window to establish person-specific lags for subsequent analysis (Butler et al., 2014; Dan-Glauser & Gross, 2013; Sze, Gyurak, Yuan, & Levenson, 2010), it will be important to extend the uSEM method used here in ways that support examination of negative and positive feedback loops that may manifest at other time-scales and within higher dimensional systems.

Conclusion

In this paper we conceptualized emotion as an emergent, dynamic process, and used a dynamic system approach to model the emotion generating process and emotion regulatory process simultaneously. We highlighted the connection between the theoretical account of emotion concordance and regulation and the temporal dynamics as two types of feedback loops, and also emphasized on the connection between the temporal dynamics and the reactivity aspect of emotion as an integrated form of the dynamics. In this regard, the present study addresses the gap between the process accounts that permeate theories of emotion and development and the means to test those accounts (Granic, Hollenstein, & Lichtwacrk-Aschoff, 2016; Richters, 1997). These dynamic approaches go beyond coarse average summaries to capture the rich real-time processes that characterize biological, behavioral, cognitive, and emotion dynamics as they transpire in the real world. We look forward to what emerges next.