Demographers have spent decades asking a question that science's own gatekeepers have mostly not asked about themselves: when people pair up, how much do they sort on background, and how much of that is preference rather than the mere fact that some kinds of people are more common in the pool? We propose to bring that measurement toolkit to the people who run journals. The plan is to treat each appointment to an editorial board, and each assignment of a reviewer by a handling editor, as a pairing between two scholars. We then ask whether those pairings match on "pedigree" (doctoral institution prestige, current institution prestige, and advisor lineage) more than the composition of the eligible pool and the appointees' own track records would predict.
Getting this measurement right matters because the existing evidence points at editors as the place where bias does the most damage, but it does so only in simulation and with bias assigned at random. An agent-based model in Editorial behaviors in peer review finds that 10% biased editors cut the quality of accepted papers by 11.01%, against 6.81% for 10% biased reviewers. Nobody has measured whether real editorial boards reproduce themselves along pedigree lines. Nobody has measured whether modern automatic reviewer matching, which leans on citation links, adds to that sorting.
Core Claim
Raw homophily statistics, such as the correlation between an appointer's and an appointee's institutional prestige, mix two things. One is genuine sorting. The other is shifts in who is in the pool. If elite-institution PhDs make up a growing share of the eligible population, raw correlations rise even when no one's preferences change. Two ideas from the assortative-mating literature separate these.
The first is relative measurement. Trends in Educational Assortative Marriage in China Over the Past Century ranks each person by percentile within their own cohort's distribution, which is how it can say that sorting nearly tripled across cohorts born 1906–1995 without being fooled by educational expansion. The second is the conditional "matching doors" test. Marriage of matching doors: Marital sorting on parental background in China shows that parental class and household registration predict spousal sorting even after controlling for the spouses' own attainment. Our academic analogue asks whether doctoral pedigree predicts appointer–appointee matching after controlling for the appointee's own publication and citation record.
Our claim is that these two adjustments change substantive conclusions about editorial boards. Specifically, a margin-adjusted, achievement-conditioned sorting index will (a) recover planted sorting parameters that raw coefficients misstate when pool composition drifts, and (b) find residual pedigree sorting in real board renewals. We also claim that the citation factor in current reviewer matchers measurably increases pedigree closeness between authors and assigned reviewers. That second claim is falsifiable on public benchmarks.
Research Program
Panel of editorial boards. The first component is data construction, since the file contains no editorial appointment dataset. IEEE Transactions publish a yearly masthead listing the editor-in-chief, associate editors, and the board's tiered structure; the file indexes examples for Information Theory, Systems, Man, and Cybernetics, and Cybernetics. We will assemble a panel of these mastheads over many years. Each newly appearing member is an "appointee." The incumbent editor-in-chief and the senior editors in the same area are the plausible "appointers." Members are linked to publication records for affiliation, doctoral institution, advisor where recoverable, and achievement measures.
Eligible pool and pedigree scores. For each journal-year, the eligible pool is every scholar who published in that journal or its core citing venues within a fixed window. Institutional prestige is converted to a percentile within that year's pool, which follows the relative-measurement move of the China trends paper.
Sorting index. The measurement itself compares the observed cross-tabulation of appointer and appointee prestige tiers with what random draws from the eligible pool would produce. It is estimated as a log-linear model on that table, in which the marginal terms absorb pool composition and an interaction term $\delta$ captures sorting. This adopts the stated aim of Sorting in the Marriage Market: A New Approach to Measuring Assortative Mating, which is to measure sorting intensity relative to marginals rather than through raw correlation; we do not claim to reproduce that paper's specific estimator. The matching-doors version adds the appointee's own achievement as a covariate. A $\delta$ that survives this control is pedigree sorting that is not explained by track record.
Reviewer assignment. The second component moves from boards to day-to-day reviewer assignment. Chain-of-Factors Paper-Reviewer Matching ranks reviewers in three stages: semantic similarity first, topic refinement second, citation links last. Because scholars cite their own students, advisors, and institutional peers, we hypothesize that the final citation stage raises pedigree closeness between author and reviewer. We will compute the same sorting index on the top-ranked reviewer lists in the KDD 2020 benchmark, which supplies 174 papers, 737 reviewers, and 3,480 annotated pairs with Microsoft Academic Graph metadata and therefore affiliations, and we will ablate each factor. We will then trace the cost of removing that sorting. This is done by adding a cap on pedigree closeness to the assignment step and recording how expertise precision and the core-stability property of Group Fairness in Peer Review respond on its CVPR 2017/2018 and ICLR 2018 assignment data.
Interpretive lens. Wealth Dynamics Over Generations: Analysis and Interventions is used as a lens, not a deliverable. It shows that when a gatekeeper's signal mixes talent with inherited advantage, persistent traps appear in about 30% of simulated parameter settings. Our achievement-conditioned estimate gives an empirical handle on how much weight pedigree carries relative to track record in real appointments. That indicates whether the trap regime is even plausible for editorial governance, though our data alone cannot prove a trap exists. The referral model of The Role of Referrals in Immobility, Inequality, and Inefficiency in Labor Markets motivates the expectation that homophilous appointment lowers match quality. That paper assumes the network is fixed from outside the model. Our board panel instead observes the network that appointments create.
Baselines
Naive reference. The "do nothing" reference is a permutation null: appointees are reshuffled at random among appointers within the same journal-year, holding both marginals fixed. Any index must place the real data outside this null before we interpret it.
Raw homophily statistic. The first real competitor is the raw Pearson correlation, or network assortativity coefficient, between appointer and appointee prestige percentiles. It is the strongest simple competitor because it is what most bibliometric studies would report, and it already uses percentiles. It will be computed on exactly the same pairs and pool definitions. Our index beats it if, on synthetic data with a fixed planted sorting parameter and drifting pool composition, our estimate stays within its confidence interval of the true value while the raw coefficient drifts. On real data, the test is whether the two give different trend conclusions across years.
Achievement-only model. The second competitor is a regression of appointment on the candidate's own citations and publications with no pedigree term. If this model fits as well as the pedigree-augmented one, the matching-doors claim fails.
Existing matchers. For reviewer assignment, the strongest existing method is Chain-of-Factors, which reports P@5 of 55.68 (soft relevance) and 28.24 (hard) on NIPS 2006, against 49.41 for TPMS and 52.94 for SPECTER, all on that benchmark. No reported result was found on Lacuna for pedigree closeness of any matcher's assignments, nor for CoF on KDD 2020 under our protocol. We will therefore rerun CoF, SPECTER-only, and TPMS-style matching on KDD 2020 with identical candidate pools and no extra tuning. For stability, the reference is CoBRA's reported 0% core violations against TPMS's violation factor of 3.134 on CVPR 2017.
Ablations. The ablation that isolates our key ingredient removes the citation stage from CoF and measures the change in the sorting index and in P@5. A pedigree-capped assignment counts as a success if it lowers the sorting index to within the permutation null while losing less than one P@5 point relative to unconstrained CoF on KDD 2020. A loss larger than that is reported as the price of de-sorting, not hidden.
Why not just…?
Why not just rely on conflict-of-interest filters? These filters typically exclude recent co-authors and shared current affiliation. They do not see shared doctoral institution, advisor lineage, or prestige tier, which are exactly the background traits the matching-doors literature finds operating beneath observed attainment. They also bind reviewers, not the handling editor who assigns them, and the editor is where Editorial behaviors in peer review locates the larger damage. The test is direct: we apply a standard filter first and then compute the sorting index on what remains. If residual $\delta$ is indistinguishable from the null, this alternative wins and we will say so.
Why not just report a raw assortativity coefficient? This is the real threat, because if pool composition is stable over the study period the two measures will agree. The planted-parameter simulation settles when they diverge. Real data then shows whether the composition of journal pools moved enough over the panel years to matter. If it did not, the margin adjustment becomes a robustness check rather than a headline, and we will present it that way.
Why not just conclude that sorting reflects merit? If elite-pedigree scholars are appointed because they have stronger records, the achievement-only model will absorb the pedigree effect. Because of that possibility, the conditional test, not raw sorting, is the central estimate. One caveat applies. Achievement is itself partly shaped by earlier gatekeeping, and the wealth-dynamics model makes this concern precise, so controlling for it gives a conservative lower bound on pedigree sorting rather than a clean causal estimate.
Proposed Experiments
Synthetic recovery. The first decisive check is synthetic. We generate appointment tables with known $\delta$ while the elite share of the pool rises over twenty simulated years, then compare recovery error for raw correlation and for our index.
Board panel. The second check is the IEEE masthead panel. We estimate $\delta$ with and without achievement controls, separately for editor-in-chief-to-senior-editor and senior-editor-to-associate-editor appointments, and test whether sorting is stronger at the top tier as the editor-multiplier result would suggest.
Reviewer matching. The third check runs on KDD 2020. We measure the sorting index of top-5 reviewer lists from TPMS-style, SPECTER, and full CoF matching, and from CoF without its citation stage. If removing the citation stage lowers pedigree closeness at small cost in P@5, the claim that current matchers amplify pedigree sorting is supported. If it does not, the claim is refuted.
Stability cost. The fourth check applies the pedigree cap on the CVPR and ICLR assignment data. We report core violation rates and egalitarian welfare beside those of CoBRA and TPMS, to show whether de-sorting conflicts with sub-community stability. That conflict is open because CoBRA already trades away the minimum per-paper score to reach 0% violations.
References
- Editorial behaviors in peer review
- Trends in Educational Assortative Marriage in China Over the Past Century
- Marriage of matching doors: Marital sorting on parental background in China
- IEEE Transactions on Information Theory publication information
- IEEE Transactions on Systems, Man, and Cybernetics publication information
- IEEE Transactions on Cybernetics publication information
- Sorting in the Marriage Market: A New Approach to Measuring Assortative Mating
- Chain-of-Factors Paper-Reviewer Matching
- Group Fairness in Peer Review
- Wealth Dynamics Over Generations: Analysis and Interventions
- The Role of Referrals in Immobility, Inequality, and Inefficiency in Labor Markets