Exponential Random Graph Modeling for Complex Brain Networks

Exponential random graph models (ERGMs), also known as p* models, have been utilized extensively in the social science literature to study complex networks and how their global structure depends on underlying structural components. However, the literature on their use in biological networks (especially brain networks) has remained sparse. Descriptive models based on a specific feature of the graph (clustering coefficient, degree distribution, etc.) have dominated connectivity research in neuroscience. Corresponding generative models have been developed to reproduce one of these features. However, the complexity inherent in whole-brain network data necessitates the development and use of tools that allow the systematic exploration of several features simultaneously and how they interact to form the global network architecture. ERGMs provide a statistically principled approach to the assessment of how a set of interacting local brain network features gives rise to the global structure. We illustrate the utility of ERGMs for modeling, analyzing, and simulating complex whole-brain networks with network data from normal subjects. We also provide a foundation for the selection of important local features through the implementation and assessment of three selection approaches: a traditional p-value based backward selection approach, an information criterion approach (AIC), and a graphical goodness of fit (GOF) approach. The graphical GOF approach serves as the best method given the scientific interest in being able to capture and reproduce the structure of fitted brain networks.

💡 Research Summary

The paper introduces Exponential Random Graph Models (ERGMs), also known as p* models, as a principled statistical framework for modeling whole‑brain connectivity networks, a domain where traditional neuroscience approaches have largely relied on single‑feature descriptive statistics (e.g., clustering coefficient, degree distribution) or generative models that reproduce only one such feature. Recognizing that the brain’s functional and structural networks are shaped by the simultaneous interaction of many local properties, the authors argue that ERGMs—widely used in social network analysis—offer a way to quantify how a set of interacting local graph statistics collectively give rise to the observed global architecture.

Data and preprocessing
The authors use resting‑state functional MRI and diffusion MRI data from a cohort of healthy adults (approximately 20–30 subjects). After standard preprocessing (motion correction, spatial normalization, parcellation into cortical and subcortical regions), binary adjacency matrices are constructed for each participant by thresholding connection weights to achieve comparable network densities across subjects.

Candidate statistics
A suite of local graph statistics is defined as potential model terms: simple edge count, two‑star (degree‑related) terms, triangle counts, geometrically weighted edgewise shared partners (GWESP), and geometrically weighted dyadwise shared partners (GWDSP). These capture, respectively, overall sparsity, degree heterogeneity, clustering, and higher‑order triadic closure mechanisms that are hypothesized to be important in brain networks.

Model selection strategies
Three distinct strategies for selecting which statistics to retain in the final ERGM are compared:

1. p‑value‑based backward elimination – Starting from a full model, terms with non‑significant coefficients (p > 0.05) are iteratively removed. This approach mirrors traditional hypothesis testing but can be unstable in high‑dimensional network settings and may discard terms that are jointly important.

2. Information‑criterion (AIC) selection – Models are compared using the Akaike Information Criterion, balancing goodness‑of‑fit against model complexity. While AIC penalizes over‑parameterization, it does not directly assess whether the simulated networks reproduce key topological features of the empirical data.

3. Graphical Goodness‑of‑Fit (GOF) – For each candidate model, a large number of networks are simulated. The simulated networks are then compared to the observed networks across a battery of diagnostics (degree distribution, clustering coefficient distribution, path length, modularity, etc.). Visual and quantitative discrepancies guide the retention or removal of terms.

Results
The p‑value approach yields a parsimonious model but often eliminates GWESP or GWDSP terms, leading to simulated networks that under‑represent the high clustering and triadic closure typical of brain graphs. The AIC‑selected model improves fit modestly but still shows systematic deviations in the tail of the degree distribution and in modularity measures. The GOF‑guided model, which retains edge, two‑star, and especially GWESP terms, produces simulated networks whose degree, clustering, path length, and community structure closely match those of the empirical brain networks. The positive GWESP coefficient indicates that shared partners (i.e., the presence of common neighbors) strongly promote edge formation, consistent with the brain’s tendency toward densely interconnected local modules.

Interpretation and implications
The study demonstrates that ERGMs can capture the interplay of multiple local mechanisms—sparsity, degree heterogeneity, and triadic closure—in shaping whole‑brain network topology. By providing a statistical test for each term, ERGMs allow researchers to move beyond descriptive summaries and to formally evaluate competing mechanistic hypotheses about brain connectivity. The authors argue that the graphical GOF approach is most aligned with neuroscientific goals because it directly assesses whether the fitted model reproduces the structural signatures of interest.

Future directions
The authors propose extending this framework to clinical populations, where differences in ERGM parameters (e.g., reduced GWESP in Alzheimer’s disease) could serve as biomarkers of network pathology. They also suggest integrating weighted ERGMs to preserve connection strength information and combining ERGMs with longitudinal data to study network evolution over development or disease progression.

Conclusion
Overall, the paper provides a thorough methodological roadmap for applying ERGMs to brain networks, validates the approach on healthy adult data, and establishes graphical GOF as the preferred model‑selection criterion for preserving the complex architecture of neural connectivity. This work opens the door for more nuanced, hypothesis‑driven investigations of how local wiring rules give rise to the emergent functional and structural organization of the human brain.