A network approach to psychopathology: new insights into clinical longitudinal data

Abstract

In the network approach to psychopathology, disorders are conceptualized as networks of mutually interacting symptoms (e.g., depressed mood) and transdiagnostic factors (e.g., rumination). This suggests that it is necessary to study how symptoms dynamically interact over time in a network architecture. In the present paper, we show how such an architecture can be constructed on the basis of time-series data obtained through Experience Sampling Methodology (ESM). The proposed methodology determines the parameters for the interaction between nodes in the network by estimating a multilevel vector autoregression (VAR) model on the data. The methodology allows combining between-subject and within-subject information in a multilevel framework. The resulting network architecture can subsequently be analyzed through network analysis techniques. In the present study, we apply the method to a set of items that assess mood-related factors. We show that the analysis generates a plausible and replicable network architecture, the structure of which is related to variables such as neuroticism; that is, for subjects who score high on neuroticism, worrying plays a more central role in the network. Implications and extensions of the methodology are discussed.

Conflict of interest statement

Competing Interests: The authors have declared that no competing interests exist.

Figures

Figure 1. Estimated population network at baseline.

The six items are: C = cheerful, E = pleasant event, W = worry, F = fearful, S = sad and R = relaxed. Solid green arrows correspond to positive connections and red dashed arrows to negative connections. Only arrows that surpass the significance threshold are shown (i.e., for which the p-value of the t-statistic is smaller than 0.05). Arrows can be either red, indicating a negative relationship (i.e., ), or green, indicating a positive relationship (i.e., ). Furthermore, the strength of the relation from item k to item j (i.e., an extremer value for ) is translated into the thickness of the arrows: the thicker the arrow between two nodes, the stronger the nodes or items are related. Note that item responses can also be predicted from the previous state of the item itself. These arrows are the self-loops in the network.

Figure 2. Inter-individual differences of the arrows of the network from Figure 1.

The thickness of the arrows is based on the size of the standard deviation of the random effects. To construct the figure, we have put a cutoff of 0.1 on the standard deviation and only the standard deviations above the cutoff are shown with a non-transparent arrow. As the threshold for the standard deviation of the random effects 0.1 was chosen because it represents large inter-individual differences. The average coefficient of the self-loops (i.e., autoregression coefficients) is about 0.2 with a random effects standard deviation of 0.1. Therefore, assuming a normal distribution, the range from 0 to 0.4 represents 95% of the individual self-loop coefficients. With a larger cutoff, such as 0.2, also individuals having negative self-loops would be taken into account. However, more than 95% of the population has a positive self-loop strength.

Figure 3. Individual networks (at baseline) of two different persons.

Figure 4. Centrality (betweenness) of each item in the network as a function of level of neuroticism at baseline.

Low, mid, and high neuroticism are shown from left to right. The labels of the items are abbreviated by their first letter (C = cheerful, S = sad, R = relaxed, W = worry, F = fearful and E = event). The black dots are the model-based estimate of betweenness, the darkgrey vertical lines represent 50% confidence intervals and the light grey vertical lines represent 95% confidence intervals (as estimated from the bootstrap method). Together, the median, 50% and 95% confidence intervals give information on how the node centrality for every item in all three networks is distributed.

Figure 5. Correspondence between the basis dataset and the validation dataset.

Left panel: Representation of the correspondence between the population network coefficients (fixed effects) of the basis dataset (x-axis) and the validation dataset (y-axis). Right panel: Representation of the correspondence between the inter-individual differences (random effects) of the basis data (x-axis) and the validation data (y-axis).

Figure 6. Centrality (betweenness) of each item in the network as a function of level of neuroticism in the validation dataset.

Low, mid, and high neuroticism are shown from left to right. The labels of the items are abbreviated by their first letter (C = cheerful, S = sad, R = relaxed, W = worry and F = fearful). The black dots are the model-based estimate of betweenness, the darkgrey vertical lines represent 50% confidence intervals and the light grey vertical lines represent 95% confidence intervals (as estimated from the bootstrap method). Together, the median, 50% and 95% confidence intervals give information on how the node centrality for every item in all three networks is distributed.