Homophily and Long-Run Integration in Social Networks

We model network formation when heterogeneous nodes enter sequentially and form connections through both random meetings and network-based search, but with type-dependent biases. We show that there is “long-run integration,” whereby the composition of types in sufficiently old nodes’ neighborhoods approaches the global type distribution, provided that the network-based search is unbiased. However, younger nodes’ connections still reflect the biased meetings process. We derive the type-based degree distributions and group-level homophily patterns when there are two types and location-based biases. Finally, we illustrate aspects of the model with an empirical application to data on citations in physics journals.

💡 Research Summary

The paper develops a stochastic model of network formation in which heterogeneous agents (nodes) arrive sequentially and create directed links through two distinct mechanisms: a random‑meeting stage and a network‑based search stage. Each node belongs to a type θ drawn from a finite set Θ with ex‑ante probabilities p(θ). When a new node is born it forms n outgoing links. A fraction mr of these links (the “parents”) are chosen by a type‑dependent random meeting process: the probability that a node of type θ connects to a node of type θ′ is p(θ,θ′). If p(θ,θ′)=p(θ′) the meeting is unbiased; otherwise the process is biased, capturing homophily when p(θ,θ)>p(θ). The remaining fraction ms=n−mr links are formed by “search”: the new node selects uniformly at random among the neighbors of the previously chosen parents. The search stage may be unbiased (uniform over neighbors) or biased; most of the analysis assumes it is unbiased.

Three notions of integration are introduced. Weak integration holds whenever the probability of receiving a new link grows with a node’s degree, which is guaranteed as long as some links are formed by search and the degree‑preferential attachment effect is present. Long‑run integration is a stronger property: as a node ages, the composition of types among its neighbors converges to the global type distribution. The authors prove that this occurs only when the search stage is unbiased; the random‑meeting stage may remain arbitrarily biased. Partial integration describes a monotonic convergence of the neighbor‑type composition toward the global distribution without full convergence; this arises under specific conditions on the bias structure.

A special case with exactly two types and location‑based meeting bias is examined. Here, each type tends to reside in a particular geographic or social “location,” and random meetings are proportional to the local population shares. The model yields explicit formulas linking a node’s age (or degree) to its local homophily, showing that older nodes become less homophilous because search increasingly dominates link formation. The authors also derive the degree distributions for each type, extending the classic preferential‑attachment results to heterogeneous agents. They find that while the minority group experiences fewer intra‑group links (size effect), homophilic bias pulls in the opposite direction, creating a tension that nevertheless leads to identical overall in‑link distributions for both groups, independent of their relative sizes.

The theoretical results are illustrated with an empirical application to citation data from American Physical Society journals (1985‑2003). Papers are treated as nodes, their fields as types, and citations as directed links that can be formed either directly (random meeting) or via the reference lists of already‑cited papers (search). Empirically, the share of citations a paper receives from its own field declines with age, while cross‑field citations increase, consistent with the model’s partial integration prediction. Moreover, the pattern suggests that the search component (following reference lists) is less biased than the initial random discovery of papers, supporting the theoretical claim that unbiased search drives integration.

Overall, the paper makes several contributions: (1) it introduces the first stochastic growth model that incorporates both type‑dependent random meetings and a potentially unbiased network‑based search, allowing a rigorous analysis of how homophily evolves over time; (2) it clarifies the conditions under which long‑run integration emerges, highlighting the crucial role of unbiased search; (3) it provides analytical expressions for degree distributions and group‑level homophily in a heterogeneous setting; and (4) it validates the theory with real citation data, showing that network mechanisms can indeed mitigate initial homophilic biases. The findings have policy relevance: encouraging unbiased search mechanisms (e.g., broader literature reviews, recommendation systems that expose users to diverse sources) can reduce segregation and improve information diffusion across groups.