Network Econometrics (Peer Effects)
Also known as: Social Interactions Model, Peer Effects Model, Social Network Regression, Ağ Ekonometrisi
Network econometrics estimates how individuals' outcomes are causally shaped by the behaviour and characteristics of their social-network neighbours. Formalised by Bramoullé, Djebbari, and Fortin (2009), the framework embeds a row-normalised adjacency matrix into a linear regression, separating endogenous peer effects (imitation of outcomes), exogenous contextual effects (influence of neighbours' attributes), and correlated effects (shared environment), while using network topology to construct valid instruments.
Key highlights
- Cleanly separates endogenous peer effects, contextual effects, and correlated effects — a classic identification problem in social science.
- Uses observed network structure to construct instruments internally, without requiring external exogenous variation.
- Extends naturally to dynamic, weighted, or directed networks and to nonlinear outcome models.
- Provides a rigorous identification theorem grounded in linear algebra (rank conditions), making assumptions explicit and testable.
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
Apply network econometrics when you have an observed social or economic network, individual-level outcome data, and a research question centred on whether and how peers causally influence each other. The method requires a known, non-trivial network topology with sufficient intransitivity to generate valid instruments. It is inappropriate when the network is fully connected, when group boundaries replace dyadic links (use spatial lag or multi-level models instead), or when the outcome variable violates linearity assumptions without tractable transformation.
Strengths & limitations
- Cleanly separates endogenous peer effects, contextual effects, and correlated effects — a classic identification problem in social science.
- Uses observed network structure to construct instruments internally, without requiring external exogenous variation.
- Extends naturally to dynamic, weighted, or directed networks and to nonlinear outcome models.
- Provides a rigorous identification theorem grounded in linear algebra (rank conditions), making assumptions explicit and testable.
- Requires complete or near-complete observation of the network; missing links introduce measurement error that biases estimates.
- Identification fails in complete graphs and regular lattices where G²X lies in the span of X and GX.
- The linear-in-means specification imposes homogeneous peer effects; heterogeneity across nodes or link types is not captured without extensions.
- Large networks generate high-dimensional instrument sets, rising many-instruments bias; careful selection or LIML is needed.
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
How is network econometrics different from spatial econometrics?
Spatial econometrics typically uses geography-based weight matrices (inverse distance, contiguity) and focuses on spatial spillovers. Network econometrics uses explicitly observed social or economic link structures, allows asymmetric and sparse connectivity, and derives instruments from network topology rather than assuming a geographic decay function. The underlying algebra is similar, but the identification strategy and the interpretation of the weight matrix differ substantially.
What software packages implement this estimator?
In R, the package 'PartialNetwork' (Boucher & Houndetoungan) implements the Bramoullé et al. IV estimator with network-based instruments. Stata users typically code the G matrix manually and apply ivreg2 or xtivreg2. Python implementations are available via custom scripts using NetworkX for adjacency construction and statsmodels or linearmodels for the IV stage.
What sample size and network density are needed for reliable estimation?
No universal threshold exists, but reliable estimation generally requires a network with hundreds of nodes, moderate density (not too sparse to prevent instrument relevance, not complete to preserve intransitivity), and variation in degree across nodes. Simulations in Bramoullé et al. (2009) and subsequent work suggest that star-shaped or locally clustered networks with intransitive triads provide the strongest identification, while overly regular or fully connected networks make identification fragile.
Sources
- 1.Bramoullé, Y., Djebbari, H., & Fortin, B. (2009). Identification of peer effects through social networks. Journal of Econometrics, 150(1), 41–55.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 2). Network Econometrics. ScholarGate. https://scholargate.app/econometrics/network-econometrics