Homophily and Long-Run Integration in Social Networks

1 Introduction

• Weak integration holds when a node’s probability of being found rises with degree, given some search and a possibility of cross-type links; long-run integration requires unbiased search.Partial integration holds under a particular condition on biases, with link biases generally decreasing with age.Source: §1 · p. 3
• With unbiased search, older nodes become less biased because search grows in importance with age and neighbors’ neighbors have a more neutral type composition.Randomly met nodes can still have biased neighborhoods; the mechanism is that these nodes are relatively more connected to their own types.Source: §1 · p. 3
• A two-type location-based model yields an explicit relationship between local homophily and a node’s age or degree, and examines how type shares affect integration.The model ties meeting biases to the population’s type distribution.Source: §1 · p. 3
• Relative homophily is strongest for equal-sized groups and vanishes as group sizes become increasingly unequal; total in-link distributions are identical across groups regardless of relative size.For minority nodes, the size effect makes own-group links rarer while homophilic bias pulls in the opposite direction.Source: §1 · p. 4
• In APS physics-journal citations from 1985 to 2003, a paper’s share of citations from its own field decreases as it ages and becomes more cited.The authors interpret the observed patterns as some evidence of partial integration.Source: §1 · p. 4
• The paper extends stochastic network-formation analysis to heterogeneous agents and studies how homophily changes over time and relates to node degree.Earlier approaches cited in the introduction generally abstract from heterogeneity or make homophily independent of age and degree.Source: §1 · p. 5

2 Homophily in a random meeting process

• Each period adds a node that forms directed links to earlier nodes through random meetings and search, with types shaping the meeting processes.The section focuses on type-based bias in random meetings.Source: §2 · p. 5

2.1 The model

• A fraction of links comes from type-dependent random meetings, while the remainder comes from parents’ neighbors; the search stage may be unbiased or type-biased.The model defines the ratio of random-process links to search-process links as σ.Source: §2.1 · p. 6

2.2 Integration

• Weak integration concerns older nodes being more likely than younger nodes to receive links, whereas partial integration requires their incoming-neighbor type shares to move toward population shares over time.Partial integration requires weakly closer shares over time, and strictly closer shares for some types.Source: §2.2 · p. 7
• Long-run integration is the stronger condition that each type’s share among a node’s incoming links converges to its population share as that node ages.These integration notions concern in-degrees; out-degree dynamics are treated separately elsewhere.Source: §2.2 · pp. 7–8

3.1 Model dynamics

• A mean-field model combines type-biased random meetings with search through the neighborhoods of randomly found parent nodes.The analysis uses a continuous-time approximation that abstracts from starting conditions and short-term fluctuations.Source: §3.1 · pp. 8–9

3.2 Integration

• With some search, nodes satisfy weak integration because the probability of being found increases with degree.The weak integration result also extends to models with biased search.Source: §3.2 · pp. 10–11
• With some search and unbiased search, old nodes' neighborhood type composition converges to population frequencies, even though their share of the population they connect to vanishes.Integration is driven by the changing composition of links, not by nodes eventually connecting to a large fraction of the population.Source: §3.2 · pp. 11–12
• Partial integration requires an additional monotone-convergence condition on A when mr < 1, and the paper does not seek general or necessary conditions for it.The condition imposes restrictions even on strictly positive transition matrices.Source: §3.2 · p. 11
• Unbiased search progressively reduces neighborhood bias as nodes become hubs, but newly formed links remain biased and oldest nodes have the least biased neighborhoods.The reinforcing process works as less-biased neighborhoods become easier for other types to find.Source: §3.2 · p. 12

3.3 On the dynamics of out-degrees

• When ms < 1, out-degree composition converges to a unique steady state, but persistent random links prevent full homogenization.The excess bias in out-links declines over time when the initial distribution is more biased than the steady state.Source: §3.3 · p. 13

4 Location–based biases

• In the two-type model, location choices governed by γ generate type-biased random meetings, while network-based search remains unbiased.When γ = 1/2, locations are independent of type; when γ = 1, meetings are within-group.Source: §4 · pp. 13–14

4.1 Explicit formulas for long–run integration

• Under location-driven biases, the explicit relationship between same-type in-link share and degree is decreasing and convex, and converges to population shares.The relationship is flatter in larger groups, so degree-related differences in integration are smaller there.Source: §4.1 · p. 15

4.2 Cumulative link distributions

• The type-sensitive degree distributions have power-law upper tails, and the ordering of own- versus other-group links depends on whether a group is the majority or minority.For the majority group, F11 FOSD F12; in the minority group, the size and homophily effects can oppose each other, while its high-degree nodes receive relatively more majority-group links.Source: §4.2 · pp. 16–17
• Increasing meeting bias shifts within-group link distributions upward and cross-group link distributions downward; the two groups' overall in-degree distributions are identical.The identical-distribution prediction is independent of homophily biases, relative group sizes, and the share of links formed randomly.Source: §4.2 · p. 17

4.3 Long run homophily and group size

• Group-level homophily increases with the random-meeting share and location bias, while greater reliance on search moves same-type link shares toward population frequencies.As search dominates, homophily tends to disappear; as random meetings dominate, links reflect location-based meeting probabilities.Source: §4.3 · pp. 18–19
• Relative group-level homophily is symmetric in group size, vanishes when one group dominates, and reaches its maximum when groups are equal in size.The location model's natural mixing within locations overcomes location bias at extreme group-size asymmetry.Source: §4.3 · pp. 17–19

5 Integration with biased search

• Search bias is represented by a type-pair matrix Bs, where 1 means no additional bias and values above or below 1 indicate positive or negative bias.The bias captures how a type distorts its relative probabilities of searching parent nodes' out-neighborhoods.Source: §5 · p. 19
• When mr < 1, the biased-search model satisfies weak integration but generically not long-run integration.Long-run integration from the earlier model is the special case with no additional search bias.Source: §5 · p. 20
• With biased search, partial integration holds provided Bs ⊙ A satisfies monotone convergence.Under this condition, neighbor-type distortions decrease over time but do not vanish as a node accumulates links.Source: §5 · p. 20

6 An empirical application to citation data

• The study analyzes 207,912 APS papers and 1,488,866 citations from 1985–2003, treating papers’ first PACS-code digit as their type.The authors truncate the period at 2003 because field composition changed dramatically afterward.Source: §6 · pp. 21–22
• A citation is classified as search when an intermediate paper published between the citing and cited papers cites the cited paper; this rule labels 56.34% of citations as search.The authors note that this rule can overstate or understate search and use it as a strict interpretation because actual discovery paths are unavailable.Source: §6 · pp. 22–23

6.1 Bias in random out-citations

• Random out-citations are more often same-type than population shares would predict across all types, with a mean difference of 0.26 and no correlation between the normalized homophily index and type population share.The difference ranges from about .001 for type 10 to about .61 for type 20.Source: §6.1 · pp. 23–24

6.2 Search bias, long–run integration, and partial integration

• The application’s integration test is limited because intrinsic citation fitness is not modeled and could outweigh age and accumulated in-degree.The analysis instead uses citation order to normalize papers’ time scales, as if they had the same fitness.Source: §6.2 · p. 24
• The empirical test relates whether an in-citation is same-field to its citation order, using separate probit regressions by receiving-paper sub-field.The test uses 1,034,569 citations and treats declining same-field probability with citation order as evidence consistent with partial integration.Source: §6.2 · p. 24
• The share of in-field citations falls by 2.5 percentage points between a paper’s first and 100th in-citation on average, with higher-order citations less likely to be in-field in most sub-fields.Types 00 and 50 show an increase, while type 90 shows no significant change; the aggregate pattern is consistent with partial integration.Source: §6.2 · pp. 24–25

7 Concluding remarks

• Long-run integration occurs if and only if the search channel is type-unbiased, leaving any bias in the initial random-meeting process.Old nodes’ local networks asymptotically resemble the population at large.Source: §7 · pp. 25–26
• The two-type location-bias model yields implications for type-based degree distributions and group-level homophily.The paper notes that old nodes are network hubs, whose characteristics matter for processes such as strategic behavior and diffusion.Source: §7 · p. 26
• The generality of the location-bias results, their application to other formation processes, and multi-type network statistics such as clustering remain open questions.Source: §7 · p. 26

Appendix A Some results on Markov matrices

• M(x) is a Markov matrix because it is a convex combination of probabilities and its columns sum to one.Source: Appendix A · p. 27
• M(x) is a weighted average of the matrices M^μ, and the appendix relates its limit to the matrix whose columns have equal elements.Source: Appendix A · p. 27
• For a positive Markov matrix, the powers M^μ converge elementwise to the corresponding entries of the equal-column limit matrix.Source: Appendix A · p. 27

• In the long run, a type-i node's share of type-j in-links is its type-blind share multiplied by p(j) and a factor constant in type i.Source: §B.1 · pp. 30–31
• The neighborhood composition converges monotonically to population shares when the relevant matrix satisfies the monotone convergence property.Source: §B.1 · p. 31
• Full homogeneity occurs only as m_s approaches 1, although the matrix columns homogenize as m_s increases.Source: §B.1 · p. 31

• F first-order stochastically dominates G exactly when f(x) is at least g(x) for all x ≥ 1 and strictly greater for some x.Source: §B.2 · p. 32
• r1(k), the within-group in-link share, is convex in in-degree for k greater than zero.Source: §B.2 · p. 33
• In the location-based model, the asymptotic limit depends on the initial type distribution p, not on γ.Source: §B.2 · p. 33
• IH1(p) is symmetric around equal group sizes, equals zero at p=0 and p=1, and increases up to p=1/2 before decreasing.Source: §B.2 · p. 35

• For a type-i node, the long-run number of type-j in-links is expressed as a fraction of the type-blind process's overall links.Source: §B.3 · p. 36
• With biased search, the type-i factor remains constant, but the other factor is generally not proportional to p(j).Source: §B.3 · p. 36
• The partial-integration proof applies the monotone convergence property to B_s ⊙ A.Source: §B.3 · p. 36