Formalizing psychological interventions through network control theory

Julia Elina Stocker, Georgia Koppe, Hanna Reich, Saeideh Heshmati, Sarah Kittel‐Schneider, Stefan G. Hofmann, Tim Hahn, Han L. J. van der Maas, Lourens Waldorp, Hamidreza Jamalabadi

TL;DR

This paper demonstrates that Network Control Theory (NCT)—a framework typically used in engineering and neuroscience—can accurately model and predict the effects of psychological interventions. By treating human attitudes as a dynamic system, the researchers showed they could mathematically estimate the "energy" required for change and predict post-intervention outcomes with high accuracy ($r = 0.84$).

What Problem Does This Paper Address?

Psychological research has increasingly used "networks" to represent complex behaviors, such as how symptoms of depression or specific attitudes interact. However, these models are usually static snapshots; they describe a state but fail to explain how that state changes when we intervene (e.g., through therapy or persuasion).

The authors identify a major gap: we lack a formal, mechanistic way to quantify "how much" a psychological variable will change following an external perturbation. They propose that NCT can bridge this gap by providing the mathematical tools to characterize the "controllability" of a psychological network and predict the cascade of effects an intervention will trigger.

Methods

The researchers applied a Linear Time-Invariant (LTI) model to psychological data, represented by the equation: $$x(k + 1) = Ax(k) + Bu(k)$$ Where:

  • $x$ is the psychological state (e.g., attitudes).
  • $A$ is the interaction matrix (the network structure).
  • $B$ is the input matrix (how the intervention hits the nodes).
  • $u$ is the intervention itself.

They compared two ways of building these networks:

  1. Self-constructed: Based on participants' own subjective drawings of how they perceived their attitudes to be causally linked.
  2. Data-driven: Using Dynamic Mode Decomposition with Control (DMDc) to infer the underlying dynamics directly from the observed data.

They then calculated key control-theoretic metrics:

  • Average Controllability (AC): The ability of a node to drive the system into many different, easy-to-reach states.
  • Modal Controllability (MC): The ability to drive the system into "difficult-to-reach" states.
  • Control Energy: The theoretical "effort" or power required to move the system from one state to another.

Datasets

The study used a publicly available dataset (Hoekstra et al., 2018) featuring 30 participants.

  • Task: Participants answered an 11-item questionnaire about attitudes toward meat consumption.
  • Intervention: Participants were presented with 11 hypothetical scenarios (one per item) designed to shift their opinions.
  • Data Points: Pre-intervention ratings, post-intervention ratings, and participant-drawn causal maps of their own attitudes.

Key Insights and Ideas

  • Intuition vs. Reality: One of the most striking "aha" moments was that self-constructed networks were poor predictors of change. Participants' subjective understanding of their own mental "wiring" did not align with how their attitudes actually shifted. The data-driven (DMDc) model was far superior.
  • Energy as a Resistance Metric: The study found that "Control Energy" acts as a mathematical proxy for psychological resistance. If the theory predicted a high energy requirement to reach a certain state, the intervention was significantly less effective (lower sensitivity).
  • Controllability Matters: Nodes with high "Average Controllability" were more sensitive to interventions. This suggests that in the future, therapists or designers could identify "influential" symptoms or thoughts that, if targeted, would cause the largest positive ripple effect across the entire mental network.

Key Results

  • High Prediction Accuracy: The data-driven generative model predicted post-intervention responses with a correlation of $r = 0.84 \pm 0.15$. In contrast, the self-constructed model showed no predictive power ($r = -0.04$).
  • Sensitivity and Specificity: Interventions were generally effective (positive sensitivity), but there was high variability. NCT metrics successfully explained this variability: Average Controllability positively correlated with sensitivity ($r \approx 0.20$), while Modal Controllability correlated negatively ($r \approx -0.19$).
  • Energy Correlation: Theoretical control energy was negatively associated with intervention success ($r = -0.22$ for the generative model), confirming that "harder to move" states in the model were indeed harder to change in the participants.

Limitations and Future Work

  • Sample Size: This was a proof-of-concept study with only 30 participants. Larger, more diverse samples are needed to validate the findings.
  • Linearity Assumption: The authors used a linear model for simplicity, but human psychology is famously non-linear. Future work could explore non-linear control theory.
  • Temporal Dynamics: The interventions were "one-shot" scenarios. Real-world interventions (like months of psychotherapy) involve repeated "inputs" and effects that decay or strengthen over time, which this model doesn't yet account for.
  • Clinical Application: While this study looked at attitudes toward meat, the real prize is applying this to clinical disorders like depression or PTSD, where identifying the "optimal" node to target could revolutionize treatment planning.

Summary updated

Abstract

Despite the growing deployment of network representation to comprehend psychological phenomena, the question of whether and how networks can effectively describe the effects of psychological interventions remains elusive. Network control theory, the engineering study of networked interventions, has recently emerged as a viable methodology to characterize and guide interventions. However, there is a scarcity of empirical studies testing the extent to which it can be useful within a psychological context. In this paper, we investigate a representative psychological intervention experiment, use network control theory to model the intervention and predict its effect. Using this data, we showed that: (1) the observed psychological effect, in terms of sensitivity and specificity, relates to the regional network control theoretic metrics (average and modal controllability), (2) the size of change following intervention negatively correlates with a whole-network topology that quantifies the "ease" of change as described by control theory (control energy), and (3) responses after intervention can be predicted based on formal results from control theory. These insights assert that network control theory has significant potential as a tool for investigating psychological interventions. Drawing on this specific example and the overarching framework of network control theory, we further elaborate on the conceptualization of psychological interventions, methodological considerations, and future directions in this burgeoning field.
BibTeX
@inproceedings{stocker2023formalizing,
  author = {Julia Elina Stocker and Georgia Koppe and Hanna Reich and Saeideh Heshmati and Sarah Kittel‐Schneider and Stefan G. Hofmann and Tim Hahn and Han L. J. van der Maas and Lourens Waldorp and Hamidreza Jamalabadi},
  title = {{Formalizing psychological interventions through network control theory}},
  booktitle = {Scientific},
  year = {2023},
  doi = {10.1038/s41598-023-40648-x},
}
Figure