Directed Exponential Random Graph Model
Also known as: Directed ERGM, p-star model (directed), directed p* model, directed Markov graph model
The Directed Exponential Random Graph Model (Directed ERGM) is a family of statistical models for directed networks that estimates the probability of observing a given directed graph as a function of structural configurations — such as reciprocity, transitive triads, and in-degree centralization — and node or dyad covariates, enabling principled inference about the social processes that generate directed ties.
Key highlights
- Simultaneously models structural network tendencies and covariate effects, unlike regression on node attributes alone.
- Explicitly separates in-degree and out-degree effects, reciprocity, transitive closure, and cyclic closure that undirected models conflate.
- Provides interpretable log-odds coefficients for each structural configuration and covariate with standard errors.
- Grounded in a full probability model, enabling formal hypothesis testing and model comparison via AIC and BIC.
- Flexible specification: the analyst can include or exclude configurations based on theoretical expectations about tie-formation processes.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use a directed ERGM when the research question concerns the generative processes behind directed ties — who sends ties to whom, and why — and when tie direction carries substantive meaning, such as in citation, advice-seeking, following, or dominance networks. It suits networks of roughly 20 to 500 nodes where the analyst wants to control for structural tendencies such as reciprocity or transitive closure while estimating covariate effects. Avoid it when the network exceeds a few hundred nodes without specialist approximations, when ties are undirected (use standard ERGM instead), when the primary goal is community detection or prediction rather than generative inference, or when MCMC fails to converge, signalling a degenerate model specification.
Strengths & limitations
- Simultaneously models structural network tendencies and covariate effects, unlike regression on node attributes alone.
- Explicitly separates in-degree and out-degree effects, reciprocity, transitive closure, and cyclic closure that undirected models conflate.
- Provides interpretable log-odds coefficients for each structural configuration and covariate with standard errors.
- Grounded in a full probability model, enabling formal hypothesis testing and model comparison via AIC and BIC.
- Flexible specification: the analyst can include or exclude configurations based on theoretical expectations about tie-formation processes.
- Computationally intensive MCMCMLE estimation scales poorly beyond a few hundred nodes without approximations or sub-network sampling.
- Model degeneracy — the fitted model placing nearly all probability on near-empty or near-complete graphs — is a common failure requiring iterative reconfiguration.
- The intractable normalising constant prevents standard likelihood-ratio tests; model comparison relies on indirect criteria.
- Results depend heavily on configuration choice; different specifications can yield conflicting conclusions about the same network.
- Longitudinal directed processes require separate temporal ERGM (TERGM) extensions, which add further complexity.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
What distinguishes a directed ERGM from a standard undirected ERGM?
In a directed ERGM every potential tie between nodes i and j is represented as two separate arcs — i to j and j to i — each with its own probability. This allows the model to separately estimate reciprocity, in-degree spread, out-degree spread, and distinct directed triangle subtypes that collapse or disappear when ties are treated as undirected.
How do I handle model degeneracy?
Degeneracy arises when the model assigns near-zero probability to all networks except empty or complete ones. It typically results from including Markov-family triangle counts without their lower-order counterparts, or without using geometrically weighted alternatives such as GWESP. Start with a sparse model, add configurations one at a time, and always inspect goodness-of-fit simulations before reporting.
Can I apply directed ERGM to a network with thousands of nodes?
Standard MCMCMLE becomes impractical above roughly 500 to 1000 nodes. For larger networks, consider snowball-sampled sub-network estimation, contrastive divergence approximations, or alternative generative models such as the stochastic block model, which scales more easily to large graphs.
What software supports directed ERGM estimation?
The statnet suite in R, principally the ergm package, supports directed networks natively via directed network objects. PNet provides a Windows GUI widely used in social science. Bayesian estimation is available through the Bergm package in R.
How should I interpret a positive reciprocity coefficient?
A positive reciprocity coefficient means — after controlling for all other included configurations — a pair of nodes is more likely to form a mutual directed tie than chance alone predicts. In practical terms, knowing that A sends a tie to B increases the estimated probability that B also sends a tie to A, conditional on network density and other modelled effects.
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.
- 2.Frank, O. & Strauss, D. (1986). Markov graphs. Journal of the American Statistical Association, 81(395), 832-842.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Directed Exponential Random Graph Model. ScholarGate. https://scholargate.app/network-analysis/directed-exponential-random-graph-model