Skip to content
Back to search results

Exponential Random Graph Model (ERGM / p*)

Introduced by
Frank & Strauss (1986); extended by Wasserman & Pattison (1996) and Robins et al. (2007)
Year
1986 (foundational); modern ERGM framework 1996–2007
Type
Probabilistic generative network model
Font size

What does it do?

The Exponential Random Graph Model (ERGM), also known as the p* model, is a statistical framework for network analysis that models the probability of an observed network as a function of its local structural features — such as reciprocity, triangles, and degree distribution.

Developed from the foundational work of Frank and Strauss (1986) and extended into the modern framework by Wasserman and Pattison (1996) and Robins et al. (2007), ERGM is the inferential standard for social network analysis, capable of testing whether observed network structures arise by chance or reflect genuine social processes.

The intuition behind the method

When you look at a social network — a friendship graph, a collaboration network, a citation web — you notice patterns: friends of friends tend to become friends themselves, popular actors attract more ties, relationships are often reciprocated. ERGM asks a fundamental question: are these patterns more prevalent than you would expect by chance alone? It does this by treating the entire observed network as a single outcome drawn from a probability distribution, where the probability of that network depends on how many instances of chosen structural configurations (triangles, reciprocal ties, star shapes) it contains. Configurations with positive coefficients are more likely to occur than chance; negative coefficients mean those configurations are actively avoided. The model is estimated not by ordinary regression but by Markov chain Monte Carlo simulation, which generates thousands of plausible networks to compare against the observed one.

History

The probabilistic foundations of ERGM trace back to the p1 model of Holland and Leinhardt (1981), which accounted for reciprocity in directed networks. Frank and Strauss (1986) extended this to a Markov random graph framework capable of capturing higher-order dependence such as triangles, coining what would later be called the p* model. Wasserman and Pattison (1996) formalised the exponential family representation and gave the model its current name. A critical advance came when Snijders (2002) introduced MCMC-MLE estimation, making practical fitting feasible. Robins et al. (2007) contributed the curved exponential family statistics — especially GWESP and GWD — that dramatically reduced degeneracy and opened the method to routine applied use. The statnet suite in R, developed by Handcock, Hunter, Butts, Goodreau, and Morris, became the standard computational implementation.

When to use it

ERGM is appropriate when you have a single observed network (cross-sectional or longitudinal snapshot) and you want to test which local structural processes — reciprocity, clustering, preferential attachment, homophily — generated it. The network should be small to medium in size; MCMC convergence becomes difficult for very large networks (thousands of nodes). The model requires directed or undirected binary ties (weighted variants exist but are more complex). It is not appropriate when the network has fewer than about 20–30 nodes, when you have no clear theoretical hypotheses about structural configurations, or when you need to model continuous time dynamics (use a Temporal ERGM or a Stochastic Actor-Oriented Model instead).

Limitations and considerations

  • MCMC estimation can be slow and may fail to converge for large networks; practical limits are typically a few hundred to a few thousand nodes.
  • Model degeneracy — where the estimated model places near-all probability on empty or complete networks — is a persistent risk requiring careful statistic selection.
  • The model is primarily designed for cross-sectional snapshots; modeling genuine network dynamics requires extensions such as Temporal ERGM (TERGM) or SAOMs.
  • Requires a minimum of roughly 30 nodes for reliable parameter estimation; very small networks yield uninformative estimates.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Provides formal statistical inference — p-values, standard errors, and confidence statements — for network structural hypotheses.
  • Handles the non-independence of network ties that violates the assumptions of standard regression, making it the correct inferential tool for network data.
  • Flexible specification: directed, undirected, bipartite, and weighted variants allow coverage of many network types.
  • Goodness-of-fit diagnostics give a principled way to evaluate whether the model reproduces global network structure from local statistics.

Applications

  • Explaining tie formation in organisational advice networks — testing whether people seek advice from others of similar status or from those they are already friends with.
  • Modelling collaboration patterns among researchers — whether co-authorship forms through transitivity (friends of friends) or is driven by institutional affiliation.
  • Epidemiology and public health — identifying which structural features of sexual contact networks predict disease transmission pathways.
  • International relations — testing whether alliance formation is reciprocal, transitive, or driven by shared attributes such as democracy or trade dependence.
  • Online social platform research — distinguishing whether observed clustering and degree skew in follower networks are structural tendencies or artefacts of platform algorithms.

How it works

How it works

1

The analyst assembles the network data — nodes, edges, and any node or edge attributes — and decides whether the network is directed or undirected. The minimum recommended size is around 30 nodes; smaller networks do not give reliable parameter estimates. Structural features of interest (reciprocity, transitivity, degree heterogeneity, covariate effects) are identified at this stage.

2

A set of network statistics is chosen to represent the hypothesised generative processes. Common terms include edges (baseline density), reciprocity (mutual ties in directed networks), geometrically weighted edgewise shared partners (GWESP, a well-behaved triangulation statistic that avoids degeneracy), geometrically weighted degree (GWD, for degree distribution), and nodal covariates (attributes that may predict tie formation). The choice of statistics directly shapes what the model can and cannot detect.

3

Because the normalising constant of the ERGM likelihood is intractable for all but trivially small networks, parameters are estimated via Markov chain Monte Carlo maximum likelihood (MCMC-MLE). The algorithm iteratively proposes edge additions and deletions, accepting or rejecting each according to the current parameter values, and uses the resulting simulated network distribution to update the parameter estimates until convergence.

4

After estimation, MCMC diagnostics are inspected — trace plots of the sampled statistics should be stationary and mix well, and the t-ratios for all model statistics should be near zero. Goodness-of-fit (GOF) plots compare simulated network distributions of held-out statistics (degree distribution, geodesic distance, edgewise shared partners) against the observed values. A model that passes GOF can reproduce the global network structure from its local statistics. Degeneracy — where the model concentrates all probability mass on empty or complete networks — is a sign of mis-specification and must be resolved by choosing better-behaved statistics.

5

Estimated coefficients are interpreted on the log-odds scale, analogous to logistic regression: a positive coefficient for triangles means each additional triangle increases the log-odds of any given edge being present by that amount, conditional on the rest of the network. Standard errors and z-tests are used for inference. The analyst reports which structural configurations are significantly more or less prevalent than chance and connects those findings to the substantive theory of the network formation process.

Common pitfalls

  • Including a raw triangle count statistic (triangle) without a geometrically weighted version (GWESP) frequently causes degeneracy — always prefer GWESP or similar curved exponential family statistics.
  • Failing to run MCMC diagnostics and goodness-of-fit tests before interpreting coefficients: a converged but mis-specified model will produce biased estimates.
  • Interpreting ERGM coefficients as ordinary regression slopes — they are log-odds of tie presence conditional on the rest of the network, not marginal effects.
  • Applying ERGM to networks with fewer than 20–30 nodes, where parameter estimates are unreliable; use descriptive network statistics in that case.

Frequently asked questions

How does ERGM differ from ordinary logistic regression on dyads?

Logistic regression treats each tie as an independent observation, which violates the fundamental non-independence of network data — the presence of one tie changes the probability of others. ERGM explicitly models this dependence by conditioning on the full network structure and is therefore the statistically correct approach for network data.

What is degeneracy and how do I avoid it?

Degeneracy occurs when the estimated model assigns nearly all probability to either the empty network or the complete network, making the model useless. It is most commonly caused by including the raw triangle count statistic. The solution is to use geometrically weighted versions — GWESP for triadic closure and GWD for degree distribution — which are curved exponential family statistics that are inherently more stable.

How large does my network need to be?

As a practical lower bound, around 30 nodes is the minimum for reliable parameter estimation. Networks of 50–300 nodes are the most common range in applied work. Beyond a few thousand nodes, MCMC convergence becomes very slow and alternative models such as stochastic block models or latent space models may be more practical.

Can ERGM model longitudinal network change?

The standard cross-sectional ERGM models a single network snapshot. For panel data with repeated network observations, the Temporal ERGM (TERGM) or the Stochastic Actor-Oriented Model (SAOM / SIENA) are the appropriate extensions that model tie dynamics over time.

Related methods

Sources

  1. Robins, G., Pattison, P., Kalish, Y., & Lusher, D. (2007). An introduction to exponential random graph (p*) models for social networks. Social Networks, 29(2), 173-191.

    Open DOI reference
    Copied.
  2. Lusher, D., Koskinen, J., & Robins, G. (Eds.) (2012). Exponential Random Graph Models for Social Networks: Theory, Methods, and Applications. Cambridge University Press.

    ISBN: 9780521193566
    Copied.