On the Control of Psychological Networks

Abstract

The combination of network theory and network psychometric methods has opened up a variety of new ways to conceptualize and study psychological disorders. The idea of psychological disorders as dynamic systems has sparked interest in developing interventions based on results of network analytic tools. However, simply estimating a network model is not sufficient for determining which symptoms might be most effective to intervene upon, nor is it sufficient for determining the potential efficacy of any given intervention. In this paper, we attempt to remedy this gap by introducing fundamental concepts of control theory to both psychometricians and applied psychologists. We introduce two controllability statistics to the psychometric literature, average and modal controllability, to facilitate selecting the best set of intervention targets. Following this introduction, we show how intervention scientists can probe the effects of both theoretical and empirical interventions on networks derived from real data and demonstrate how simulations can account for intervention cost and the desire to reduce specific symptoms. Every step is based on rich clinical EMA data from a sample of subjects undergoing treatment for complicated grief, with a focus on the outcome suicidal ideation. All methods are implemented in an open-source R package netcontrol, and complete code for replicating the analyses in this manuscript are available online.

Keywords: complicated grief; control theory; intervention science; network psychometrics; personalized medicine; psychopathology.

Figure 1.. Symptom Dynamics of Complicated Grief.

Blue edges represent positive relations, red represents negative relations. Autoregressive effects were rescaled to improve edge contrast, and edges with absolute values below .05 were omitted from the visualization. Note that this network represents the lagged relation between the changes in symptom values (i.e., the use of an integrated VAR) rather than the lagged relations between the symptom values themselves. .

Figure 2.

Networks, Intervention and Variable Values for Systems 1 and 2. Networks represent dynamics operating on change (as in Figure 1). Both systems were subjected to the same series of disturbances, which was generated as white noise.

Figure 3.

Intervention strength and symptom values when (A) intervening on only suicidal ideation with strength of 1 and (B), when intervening on suicidal ideation with a strength of 1, and all other symptoms with a strength of .1. Note that the same sequence of random disturbances were applied to each simulation. Results show that the broader intervention in Panel B leads to more reduction in suicidal ideation and thoughts: loved ones than the single target intervention in Panel A. Additionally, the broad nature of the intervention in Panel B only reduced the spread of symptoms other than suicidal ideation and thoughts: loved ones slightly more than the single target intervention in Panel A.

Figure 4.

Intervention Strength and Symptom Values for Interventions A, B and C. Intervention A primarily targets suicidal ideation, Intervention B primarily targets positive activities (reverse coded), and Intervention C primarily targets. thoughts: future. The simulataneous application of these interventions led to a reduction to 0 for both suicidal ideation and thoughts: future, and a ~ 3.75 reduction in positive activities (reverse coded).

Figure 5.. Symptom Values and Intervention Strength for Interventions A, B, C with Unequal Symptom Importance and Intervention Costs.

Note: Interventions A and C have been specified to be incompatible (i.e. applying them simultaneously costs more than applying both separately). The simultaneous application of these interventions led to a similar reduction to 0 for both suicidal ideation and thoughts: future as in Fig. 4, but only a ~ 1.5 point reduction in positive activities (reverse coded) (compared to a 3.75 reduction seen in Fig. 4).