Foreign-trained faculty and the collaborative organization of high-impact U.S. science

1. Introduction

• Prior research leaves open how foreign-trained faculty’s contributions are organized within U.S. institutions and whether embeddedness, networks, or research portfolios relate to their elevated impact.The paper treats organizational placement and network position as part of how internationally accumulated capabilities become embedded in domestic scientific production.Source: §1 · p. 2
• The study links faculty rosters to OpenAlex publications and combines workforce benchmarking, within-stratum comparisons, and analyses of prior collaboration networks and cognitive portfolios.Training origin is based on terminal-degree institution location; the panel covers 2011–2020 and more than 300 major U.S. universities.Source: §1 · p. 2
• Foreign-trained faculty make up approximately 11–12% of the workforce but account for 13–14% of publications and 14–16% of elite output.The reported elite-output representation is closely related to organizational embeddedness.Source: §1 · p. 2
• Within the same institution, scientific domain, rank, and year, the elite-output difference narrows substantially, while productivity and collaboration differences persist.Foreign-trained faculty are strongly represented in research-intensive institutions.Source: §1 · pp. 2–3
• Foreign-trained faculty have larger and broader prior collaboration networks, and collaborator reach is more strongly associated with subsequent elite output among them.After accounting for prior publication volume, their topic breadth is slightly narrower and more cognitively concentrated, despite greater raw topic breadth.Source: §1 · p. 3
• The paper frames international training as scientific capacity embedded in institutions and relationships, rather than simply additional human capital.Its comparison of social and cognitive organization identifies collaboration-network structure, not unusually broad portfolios, as the more consistent distinction associated with foreign training.Source: §1 · p. 3

2.1. International training, mobility, and scientific performance

• International mobility is associated with higher publication rates, citation impact, and international collaboration, although relationships vary across fields, destinations, and career stages (Franzoni et al., 2012, 2014; Gu et al., 2024; Han et al., 2024; Jonkers & Cruz-Castro, 2013; Scellato et al., 2015).Studies also connect mobility premiums to combinations of geographic, institutional, academic, and disciplinary factors.Source: §2.1 · p. 3
• Because U.S. universities differ in resources and opportunities and hiring reflects stratification, aggregate performance differences can reflect both training trajectories and subsequent organizational placement.Comparing faculty within similar organizational settings connects mobility research with institutional stratification and cumulative advantage (Zhang et al., 2022).Source: §2.1 · pp. 3–4

2.2. Mobility, collaboration networks, and knowledge recombination

• Mobility may broaden access to people and knowledge through collaboration networks, while cognitive breadth may independently support knowledge recombination and impact; the two mechanisms should not be conflated.The literature leaves open whether international training is expressed mainly through broader individual portfolios or relational structures connecting scientists to distributed knowledge.Source: §2.2 · p. 4

3.1. Data sources and record linkage

• The linked panel combines annual AARC tenure-system rosters with OpenAlex records for 2011–2020; author matching uses name, institutional, and disciplinary information with confidence thresholds.Approximately 81% of covered faculty are linked at medium or high confidence, and residual linkage error remains possible, particularly for common names and sparse publication records.Source: §3.1 · pp. 4–5

3.2. Training origin and outcome measures

• Foreign-trained faculty have terminal degrees from institutions outside the United States; elite output counts field- and year-normalized top-1% cited papers.Training origin does not denote nationality, citizenship, birthplace, immigration status, ethnicity, or a scientist’s full educational and career trajectory.Source: §3.2 · p. 5

3.3.1. System-level representation

• Annual system-level comparisons measure workforce, publication, top-1% paper, and corresponding-authorship shares, fractionally allocating mixed-origin papers between training-origin groups.The elite representation ratio compares the foreign-trained share of top-1% papers with their share of observed faculty.Source: §3.3.1 · p. 6

3.3.2. Decomposition of the aggregate elite-output difference

• Nested logistic models add organizational covariate blocks sequentially and estimate standardized foreign-trained–U.S.-trained probability contrasts rather than comparing logit coefficients directly.Counterfactual predictions retain each observation’s covariates and are averaged over the observed covariate distribution.Source: §3.3.2 · p. 6

3.3.3. Within-context organizational comparisons

• Within-context comparisons retain strata containing both training groups and average their contrasts across institution, scientific domain, rank, and year.The 38,223 eligible strata are used to compare publication output, elite output, FWCI, mixed-origin team participation, degree centrality, and corresponding authorship.Source: §3.3.3 · pp. 6–7

3.4.1. Construction of prior collaboration networks

• Prior collaboration networks use coauthors from the preceding three years and distinguish collaborator reach, breadth, and tie depth.Reach counts distinct collaborators; breadth accounts for concentration across ties; tie depth is the mean number of shared publications per collaborator.Source: §3.4.1 · p. 7

3.4.2. Training-origin differences in prior network position

• Network-position models compare standardized reach, breadth, and tie depth by training origin, adjusting for year, scientific domain, and institutional research intensity.The analysis includes faculty-years with a nonempty prior collaboration network; reach and breadth are log-transformed before standardization.Source: §3.4.2 · pp. 7–8

3.4.3. Prior network reach and subsequent elite output

• The subsequent elite-output analysis models current-year top-1% publications against prior network reach and tests whether the reach association differs by training origin.It includes faculty-years with a prior network and at least one current-year publication; standardized predictions vary training origin and reach while retaining observed organizational covariates.Source: §3.4.3 · pp. 8–9

3.5. Cognitive-portfolio measures and models

• Cognitive portfolios use publications from the three years before each outcome year, separating prior research structure from subsequent elite output.Each prior publication is assigned its OpenAlex primary topic.Source: §3.5 · p. 9

3.5.1. Topic breadth

• Raw topic breadth is measured as exponential Shannon diversity, but is adjusted for prior publication volume and year because portfolio size strongly predicts observed breadth.Prior portfolio size explains R-squared = 0.49 of variation in raw breadth; the adjusted measure represents breadth greater or smaller than expected for portfolio size and year.Source: §3.5.1 · p. 9

3.5.2. Cognitive span

• Cognitive span measures the topic-share-weighted mean hierarchical distance between distinct topics, distinguishing intellectual dispersion from effective topic breadth.Topic distances range from different topics in the same subfield to different domains.Source: §3.5.2 · p. 10

3.5.3. Training-origin differences in cognitive portfolios

• Training-origin differences in adjusted breadth and cognitive span are estimated in standardized units, controlling for year, scientific domain, and institutional research intensity.Standard errors are clustered by faculty member.Source: §3.5.3 · p. 10

3.5.4. Joint cognitive configuration and elite output

• A grouped-binomial model tests whether the association between breadth and subsequent elite output depends on cognitive span, adjusting for training origin and organizational context.Training origin is an adjustment variable rather than part of the focal cognitive interaction.Source: §3.5.4 · p. 10

3.6.1. Robustness to network and outcome definitions

• Robustness analyses compare one-, three-, and five-year network lookbacks and replace the top-1% outcome with a top-10% indicator.Lookback windows are tested on their full eligible samples and a common sample of 980,186 faculty-years.Source: §3.6.1 · p. 10

3.6.2. Heterogeneity across scientific and organizational contexts

• Heterogeneity analyses test network interactions across career stages, institutional research intensity, scientific domains, and 12 foreign training systems represented by at least 500 faculty.Pooled Wald tests assess whether interactions differ across domains and country-specific training systems.Source: §3.6.2 · p. 11

4.1. Contribution relative to workforce presence

• Foreign-trained faculty were about 10–11% of tenure-system faculty but produced 13–14% of publications and 14–16% of top-1% papers; their elite representation ratio rose from 1.34 in 2011 to 1.48 in 2020.The representation ratio remained above 1 throughout the decade.Source: §4.1 · p. 11, Figure 1 · p. 12

4.2. Organizational position and within-context differences

• Their publication shares exceeded workforce shares in most domains, and representation was strongest in high-research-intensity institutions, with Mathematics and Computing and Humanities notable exceptions for top-1% representation.The pattern was visible at both public and private nonprofit universities.Source: §4.2 · p. 12, Figure 2 · p. 13
• Within 38,223 institution–domain–rank–year strata, differences in elite output, mean field-weighted citation impact, and corresponding-author share were greatly reduced.The nested decomposition attributed 13.2% of the elite-output difference to measured organizational factors, leaving 86.8% outside those coarse measured factors; exact institution matching explained more than coarse categories.Source: §4.2 · pp. 12–13, Figure 2 · p. 13
• Foreign-trained faculty retained nearly four-fifths of their publication difference within strata and had a higher mixed-training-origin team share than U.S.-trained faculty.The publication difference was 1.13 papers per faculty-year raw and 0.89 within strata; team-share differences were 0.30 and 0.27, respectively.Source: §4.2 · p. 13, Figure 2 · p. 13

4.3. Prior collaboration networks and elite output

• Over the prior three years, foreign-trained faculty averaged 89.1 collaborators versus 64.8, with greater collaborator breadth and slightly deeper ties; adjusted differences in reach, breadth, and tie depth remained positive.Adjusted differences were 0.130 SD in log reach, 0.104 SD in log breadth, and 0.037 SD in tie depth.Source: §4.3 · p. 14, Figure 3 · p. 15
• Prior network reach was more strongly associated with subsequent top-1% output for foreign-trained faculty, and the adjusted origin gap shifted from −0.34 percentage points at −1 SD reach to +0.23 at +1 SD.The foreign-trained × reach interaction was OR 1.091 (95% CI, 1.054–1.130). The model is associational; unobserved funding or team roles may influence both network size and later high-impact work.Source: §4.3 · p. 14, Figure 3 · p. 15

4.4. Cognitive portfolios provide a contrasting pattern

• Although raw topic counts were higher for foreign-trained faculty, their breadth was slightly narrower after adjustment for prior portfolio size and year, and their cognitive span was also modestly lower.Adjusted breadth was −0.0487 SD and cognitive span was −0.0426 SD; both were associated with lower top-1% probability, while the positive breadth × span interaction indicated that breadth’s negative association weakened as span increased.Source: §4.4 · p. 15, Figure 4 · p. 16
• The results point to a stronger distinguishing signal in collaboration networks than in cognitive breadth once observable topic volume is equalized.The paper reports that network reach is more strongly associated with elite output among foreign-trained faculty, whereas their topic portfolios are not broader after equalizing the number of observable topics.Source: §4.4 · p. 15, Figure 4 · p. 16

5.1. Alternative temporal definitions of collaboration

• An 8–10% larger multiplicative increase in top-1% output odds is associated with a one-SD increase in log network reach for foreign-trained faculty across network windows.The estimates are 1.081, 1.091, and 1.099 in the one-, three-, and five-year full eligible samples, respectively; corresponding common-sample estimates are 1.081, 1.091, and 1.098.Source: §5.1 · p. 17, Figure 5 · p. 17

5.2. Scientific-domain heterogeneity

• An 8-degree-of-freedom Wald test found no strong evidence that the reach interaction systematically differs across the nine broad scientific domains studied (χ²(8) = 11.98, p = 0.152).Adjusted interaction estimates range from 0.967 in Education to 1.197 in Mathematics and Computing; only Mathematics and Computing and Medicine and Health have estimates statistically distinguishable from 1.Source: §5.2 · pp. 17–18

5.3. Heterogeneity across foreign training systems

• χ² = 27.16, p = 0.004 rejects equality of the 12 country-specific reach interactions, which range from 0.845 for Switzerland to 1.284 for Israel.Estimates exceed 1 and are statistically significant for the United Kingdom, Canada, India, and Israel; the patterns may reflect selection, field composition, destinations, cohort histories, language, or which graduates obtain U.S. faculty appointments, so they are not causal rankings.Source: §5.3 · p. 18, Figure 7 · p. 19

5.4. Additional robustness checks

• The positive network interaction also appears across career stages and institution types and remains positive for top-10% output, though attenuated from the top-1% estimate.The top-10% interaction is 1.052 (95% CI, 1.030–1.074), compared with 1.091 for top-1% output.Source: §5.4 · p. 19

6.1. Mobility advantages are organizationally situated

• Within institution, domain, rank, and year, elite-output differences narrow sharply, while publication-volume and mixed-training-origin collaboration differences remain.The observed advantage therefore includes both placement within the research system and differences within comparable organizational settings.Source: §6.1 · p. 20

6.2. International mobility carries relational capital, not simply additional ties

• Foreign-trained faculty have greater prior collaborator reach and breadth, and reach is more strongly associated with their subsequent elite output; this pattern is stable across windows and broadly similar across domains.Country-specific network–output associations vary substantially, so international training is a heterogeneous source of relational capital.Source: §6.2 · p. 20

6.3. Mobility and knowledge recombination

• Foreign-trained faculty appear cognitively broader in raw comparisons, but the difference disappears after accounting for larger prior publication portfolios.The separation of social reach from cognitive breadth suggests that access to heterogeneous knowledge may be distributed across collaborators rather than reflected in broader individual portfolios.Source: §6.3 · p. 20

6.4. International training as a feature of scientific capacity

• The paper connects mobility, career-context, and network research by identifying organizational structure, collaboration networks, and research portfolio as separable features of scientific capacity.Its focus is the destination research system, where international training is associated with these features.Source: §6.4 · p. 21

6.5. Limitations and future research

• The analyses identify organizational and relational patterns, not selection mechanisms or causal effects; terminal-degree location captures only one element of mobility history.Coauthorship records publication ties but not mentorship or tacit knowledge transfer; the authors call for research reconstructing full career sequences.Source: §6.5 · p. 21

• Yan, Ni, and Zheng report conceptualization and methodology; Gu reports methodology and validation, while individual roles also include software, analysis, data curation, visualization, supervision, and writing.All four authors contributed to writing the original draft or reviewing and editing.Source: CRediT authorship contribution statement · p. 21

• The author–work spine contains 12,109,441 linked faculty–publication observations for 209,959 U.S.-trained and 26,435 foreign-trained faculty.Source: A1. Analytical samples and linkage… · p. 24, Table A1 · p. 24
• High-confidence links require S ≥ 0.80 and Δ ≥ 0.05; medium-confidence links require 0.50 ≤ S < 0.80 and Δ ≥ 0.01.Source: A1. Analytical samples and linkage… · p. 24, Table A1 · p. 24
• Approximately 81% of records have medium- or high-confidence linkage coverage; the table reports denominators for main and auxiliary analyses.Definitions and model equations are provided in the main text rather than repeated in the table.Source: A1. Analytical samples and linkage… · p. 24, Table A1 · p. 24

Checks overview

• The appendix retains career-stage, institutional-environment, and broader-threshold checks, with Table A2 placing their focal estimates on a common scale.Source: A2. Additional robustness checks · p. 24, Table A2 · p. 25

Career-stage heterogeneity

• The career-stage check plots prior-three-year reach by training origin and interaction odds ratios; its rank-linked sample retains 99.62% of main-model faculty-years.Rank models use the main-sample reach scale.Source: Figure A1 · p. 24, A2. Additional robustness checks · p. 26

Institutional-environment heterogeneity

• The institutional-environment check compares reach and interaction odds ratios across highest-research and other institutions; the latter estimate is less precise because its stratum is substantially smaller.Source: Figure A2 · p. 25

Broader high-impact threshold

• The foreign-trained × reach interaction is stronger for top-1% than top-10% cited output, but remains positive at both thresholds.Both estimates use the same 1,119,314 faculty-year sample, exposure scale, fixed effects, and clustered inference.Source: Figure A3 · p. 25

Estimates and interpretation

• Tables A3 and A4 provide the complete numerical estimates behind Figures 6 and 7; heterogeneity interpretation relies on joint tests rather than isolated p-value comparisons.Source: A3. Complete field and training-system… · p. 26, Table A3 · p. 27, Table A4 · p. 28

Scientific domains

• The omnibus test of reach × origin × domain terms was not significant (Wald χ²(8) = 11.98, p = 0.152).Domain-specific models use the main-model reach scale and adjust for year and institutional research intensity.Source: A3. Complete field and training-system… · p. 27, Table A3 · p. 27

Foreign terminal-degree systems

• Country-specific reach interactions are unequal (Wald χ²(11) = 27.16, p = 0.004), and all 12 are jointly nonzero (Wald χ²(12) = 40.55, p = 5.83 × 10⁻⁵).Each coefficient compares a focal foreign system with the common U.S.-trained reference; estimates are associational, not national rankings.Source: A3. Complete field and training-system… · p. 28, Table A4 · p. 28