Heterogeneous Peer Effects with Endogenous Network Formation

Duong Trinh, Santiago Montoya-Blandón · 2026

Read Heterogeneous Peer Effects with Endogenous Network Formation by Duong Trinh, Santiago Montoya-Blandón (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-24850.txt.

Introduction

Duong Trinh and Santiago Montoya-Blandón build a framework for peer effects when people choose their own networks. Outcomes spill over between peers, but responses differ from person to person. Links also form endogenously: unobserved traits drive both who connects and what outcomes follow. Ignoring that selection corrupts spillover estimates. Their model, called SCHSAR, jointly handles link formation and outcome determination. A finite mixture structure lets peer effects vary across latent types.

Estimation is fully Bayesian, using data augmentation to sidestep failures of standard methods. A simulation study validates the approach. The application studies an innovation network among United States firms. Peer effects on corporate research and development spending come out positive and significant, yet clearly heterogeneous. The correction for self chosen links is what makes the estimates credible. Policy shocks to research spending would therefore land unevenly across firms, which targeted policy design can exploit.

Catalog record

CREATORDuong Trinh, Santiago Montoya-Blandón
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE138 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.24850

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Heterogeneous Peer Effects with Endogenous Network Formation Duong Trinh, Santiago Montoya-Blandón Abstract This paper introduces a new econometric framework for modeling social interactions with heterogeneous peer responses, addressing endogenous link formation. Our Selection-corrected Heterogeneous Spatial Autoregressive (SCHSAR) approach jointly models link formation and outcome determination. We incorporate a finite mixture structure to capture heterogeneity in peer effects and account for unobserved individual-specific factors driving both network formation and outcome equations, addressing network endogeneity for credible estimation of heterogeneous spillover effects. We propose a fully Bayesian data augmentation approach for estimation and inference, overcoming challenges posed to standard likelihood-based methods. A simulation study validates our approach. Our empirical application to an innovation network among U.S. firms reveals significant positive, yet heterogeneous, peer effects on corporate R&D investments, after accounting for endogenous network formation. The findings highlight varying firm behaviors in response to exogenous R&D policy shocks and and quantify firm-level direct and spillover effects, offering valuable insights for evidence-based and targeted policy design. Keywords: Bayesian inference, Social-spatial interaction, Finite mixture models, Heterogeneous peer effects, Network formation, R&D investment. JEL classifications: C11, C31, C51, D85, O32. 1 Introduction Modeling social interactions and spillovers is a significant topic in economics, as understanding how individuals’ outcomes are shaped by those of their peers within networks is essential for both theoretical insights and policy design, with applications ranging from education ( [ ref-patacchini2017heterogeneous ] ; [ ref-tincani2018heterogeneous ] ) and labor markets ( [ ref-ioannides2004job ] ) to industrial organization ( [ ref-de2020consumption ] ; [ ref-fainmesser2016pricing ] ) and beyond. More often than not, networks exhibit substantial heterogeneity in spillover patterns across agents. While weighted or directed networks can capture variations in exposure or interaction intensity, they cannot account for the differences in how agents translate peer behavior into their own decisions, which results in heterogeneous marginal responses to peers and thus heterogeneous peer effects. Capturing such heterogeneity and distinguishing between those who amplify versus those who absorb spillovers is thus crucial for avoiding unintended policy consequences ( [ ref-hong2013heterogeneous ] ; [ ref-leung2022causal ] ; [ ref-bargagli2025heterogeneous ] ; [ ref-auerbach2026local ] ). The challenge is compounded by the fact that individuals may select into relationships based on factors that also affect outcomes, creating endogeneity problems. For instance, in education, students form friendship networks based on both observed and unobserved traits, and peer influences on academic outcomes vary across individuals ( [ ref-calvo2009peer ] ). Similarly, in innovation, firms strategically choose R&D collaboration partners, and the impact of these alliances on innovation activities is not uniform across firms ( [ ref-powell1996interorganizational ] ). Labor market outcomes are shaped by professional networks that workers selectively form, with effects on job search and wages differing by individual characteristics ( [ ref-bayer2008place ] ), while in agricultural technology adoption, farmers’ advice networks influence their adoption decisions heterogeneously ( [ ref-conley2010learning ] ). Critically, failing to account for endogenous formation of links or substantial heterogeneity can produce severely incorrect inferences about the peer effects. As detailed network data becomes increasingly available, the demand for credible econometric frameworks has intensified, yet existing methods remain inadequate for accommodating the complex realities of economic and social interactions. 1 1 1 Recent surveys (see, e.g., [ ref-bramoulle2020peer ] ; [ ref-kline2020econometric ] ) have acknowledged empirical challenges in estimating peer effects in networks. This paper addresses these longstanding methodological challenges by proposing a unified framework, the Selection-corrected Heterogeneous Spatial Autoregressive (hereafter referred to as SCHSAR) model, that simultaneously tackles both heterogeneity and endogeneity concerns when modeling network interactions. At its core, the SCHSAR framework jointly models network formation and individual outcomes through two stages, corresponding to two equations. The network formation equation explicitly expresses dyadic links as the realization of decision-making processes influenced by observed characteristics and unobservable individual-specific factors. The outcome equation features a mixture structure in which individuals belong to finite latent types, each governed by distinct set of parameters when they interact with peers in a network. This data-driven grouping captures the rich heterogeneity of peer effects, allowing both the magnitude and direction of the effects to vary across unobserved clusters. Notably, network endogeneity may arise due to unobservable factors affecting both stages, creating the need to correct potential selection bias. The SCHSAR model addresses this issue by accounting for these factors that link the two equations. Overall, SCHSAR offers a flexible and robust approach to draw credible conclusions about peer effects in settings where both individual heterogeneity and network endogeneity are present. We develop a Bayesian Markov chain Monte Carlo (MCMC) algorithm to estimate the proposed SCHSAR model, offering a straightforward implementation with several practical advantages. First, the algorithm leverages Bayesian data augmentation to achieve computational efficiency while maintaining statistical rigor. By sampling latent variables alongside model parameters from the joint posterior distribution, we overcome the intractable high-dimensional integration problem typically faced in maximum likelihood estimation. This eliminates the numerical optimization challenges while seamlessly incorporating model constraints through prior specifications. The unified Bayesian approach facilitates one-step inference across all unknown quantities, providing principled uncertainty quantification. Second, our implementation is able to deliver rich individual-level insights through posterior prediction. The algorithm generates complete posterior predictive distributions for the individual latent type indicators, enabling a data-driven probabilistic classification of individuals into latent clusters of heterogeneous interaction effects. Moreover, this facilitates an examination of how latent heterogeneity correlates with observable characteristics, enriching the interpretability of the model. Last but not least, the Bayesian framework naturally streamlines inference on policy-relevant quantities that are challenging to analyze from a frequentist perspective. For example, spillin and spillout effects, which are highly nonlinear functions of the model parameters, can be computed directly from posterior samples as estimation byproducts. Comprehensive simulation evidence validates the computational tractability and convergence of our Bayesian implementation. Our Monte Carlo experiments also demonstrate favorable frequentist properties, showing that the SCHSAR model delivers valid inference and improved performance relative to more naive approaches that either ignore endogenous network formation or assume homogeneous interaction effects. Having developed valid tools for estimation, we apply the proposed SCHSAR framework to examine the formation of an innovation collaboration network among U.S. firms and heterogeneous interaction effects on firm performance, with a focus on their R&D efforts. Using a sample of 1,\!150 firms, the empirical analysis reveals significantly positive, yet heterogeneous, peer effects on corporate R&D investments, even after controlling for selection bias due to endogenous network formation. Allowing for heterogeneity uncovers two latent firm types: approximately 34 percent constitute a peer-driven type with a larger estimated network effect of around 0.215 and weaker own-price elasticity near -2.2 , while the remaining 66 percent of firms belong to a self-driven type with a smaller network effect of around 0.127 and stronger own-price elasticity near -9.5 . These elasticities capture how sensitive a firm’s R&D investment is to changes in the R&D tax price. The peer-driven type is more responsive to changes in peers’ R&D investments, while the self-driven type is more sensitive to its own cost structure. Additionally, we disentangle notable transmitters and absorbers in response to exogenous R&D policy shocks – firms that excel at absorbing peer innovation are rarely those best at transmitting it – highlighting distinct roles within the network. Furthermore, this framework

Read the complete text