Methodology · Edition 1 · 2026

How the Industry Collaboration Index is built

The index measures collaboration with industry as behaviour in the public record, not as reported income or reputation. Every input is open, every rule is published here, and every figure on a university page can be traced to the works behind it.

1. Who is covered

The 1,192 universities of the Research Collaboration Index cohort, so the two indices can be read side by side. 3 cohort entries are left out because they are not universities or are attribution artefacts; each is listed with its reason on the transparency page. A university is ranked once it has at least 100 verified industry co-authored works in the window (1,009 are ranked; 183 are listed, with their evidence, as not yet ranked).

Which university an author belongs to. The open record credits a university with every institution filed under it, including national laboratories and institutes it shares with others. The ICI credits an author to a university only when the author's institution is the university itself, or when the author's own affiliation text names the university, or when the institution is a medical school, hospital or campus that carries the university's name. So a paper from a national laboratory counts for the university that manages it only where its authors name that university.

2. The evidence

• Research co-authored with companies, 2021-2025: articles, reviews and book chapters in the open scholarly record (OpenAlex, CC0) with at least one cohort university and at least one organisation typed as a company. 726,414 candidate works were examined; 385,292 have at least one verified university and company link. The 2025 record is still filling in as late publications are indexed, and coverage may not be equally complete for every university. Trends that compare recent years with earlier ones are therefore labelled observed publication activity and need local confirmation.
• Subject totals for every university, by subfield, from the same record, to set expected rates.
• Clinical trials (reported, not scored) starting 2021-2025 in ClinicalTrials.gov that pair an industry sponsor with a university collaborator, or a university sponsor with an industry collaborator: 4,228 distinct trials (4,665 university and trial pairs). They are shown on profiles for context. They do not enter the score in Edition 1, because a single registry and exact name matching cover some countries far better than others.
• Funded partnerships (reported, not scored): EU Horizon 2020 and Horizon Europe projects starting 2021-2025 in which a cohort university and at least one company take part (CORDIS; 6,170 projects; universities matched by official name and website, companies are private for-profit participants other than charities and grant-management services), and, for UK universities, UK Research and Innovation projects with company partners (Gateway to Research). Across 366 ranked European universities the number of EU projects with companies agrees with the number of verified industry co-authored works at a Spearman rank correlation of 0.65, and across 62 ranked UK universities the number of UKRI projects with named company partners agrees at 0.83: related, but not the same thing. They are not scored because comparable registries do not yet exist for most research systems.

Data: the open scholarly record as harvested in September 2026. Works from the latest year are still being added to the open record, and fewer of their links can yet be confirmed: the verified share of candidate works by year is 2021 55%, 2022 56%, 2023 56%, 2024 53%, 2025 49%. Many health-sector joint works are publications of company-sponsored multicentre clinical trials; they are real co-authorship, and the sector indices show where they concentrate.

Not used in Edition 1: patents (co-invention and industry patents citing university research join in Edition 2, September 2027, once a worldwide patent build is validated), funded partnerships in the score, and income figures reported by universities (available only for some countries, so they would reward data availability rather than collaboration). Edition 2 will test a funded-partnership measure as more national registries open.

3. Verifying every company link

Open data decides that an author works at a company by parsing an address. That goes wrong often enough to change rankings: a department of "Biometry" read as Zimmer Biomet, a street called Mizuho-ku read as a bank, an NHS Biomedical Research Centre read as Oxford BioMedica. So every university and company link passes these checks before it counts:

1. Industry only. Publishers, hospitals and clinics, university-owned organisations and known data artefacts are excluded even when tagged as companies. Market-producing state-owned enterprises count as industry, following the OECD Frascati Manual, and their share is shown on every profile.
2. The author's own words. The company's name must appear in the affiliation text the author wrote: as a whole phrase, as all of its distinctive words, or as its registered acronym. Place names and generic words (research, technologies, Oxford) never count on their own, and a name that appears only as part of a street or district address is rejected.
3. Company groups. National subsidiaries and research labs are combined under one group (for example every Pfizer entity), using the open record's own parent links and identical base names.
4. University units. A company name inside the name of a university's own institute, centre, laboratory or sponsored chair (the Maersk Mc-Kinney Moller Institute, a Saudi Aramco Chair) is the university's unit, not a partner. So is a company that appears only inside the same affiliation text as the university.
5. Dual affiliations. Where the only company author also holds a post at the university, the work counts half, and never toward durability or sustained partnerships.
6. Consortium papers. Works with 100 or more authors (the open record lists at most 100) count toward co-creation only: not toward durability, scale (including the network) or breadth, because they do not show a two-way relationship.
7. Hand audit. The organisations with the most links were checked by hand; research institutes, charities, public bodies and university-owned companies that the open record types as companies are excluded, each with a published reason. Companies whose names are also places, buildings or university units (a hospital acronym, a street in Tokyo, a university named after an oil company) must be written in their own company form.
8. Contractors at federal laboratories. A US company is set aside when at least half of its authors (over at least 50 author appearances) also name a US federal laboratory (NASA, NOAA, a national laboratory) among their own affiliations: its staff are working inside a government lab. This one rule covers every such company, including those that operate a laboratory under contract. The 15 companies it removed are listed on the transparency page.
9. Company staff at science parks. An author who works for a company located in a university's science or innovation park counts for the company, not the university.

Result: of 1,966,368 candidate university and company links, 1,029,249 were verified and 937,119 (47.7%) set aside because the affiliation text did not confirm them. Many set-aside links are to real companies whose name the author did not write out. A hand audit on 3 October 2026, before the second round of verification fixes, found 91.8% of 200 verified links correct (95% interval 88.2 to 95.5%); about 85% among the smallest companies. In a sample of set-aside links from that audit, about 13% were in fact the right company, so the index misses some true links as well as keeping some false ones. A fresh, stratified internal audit of this edition, after those fixes, is complete and is awaiting an independent human check of 30 items; its result will be published here as a dated update. Update, 6 October 2026: two further internal audits of this edition, the second of them independent of the first, agree with the published figure; the independent audit found 92.9% of 200 verified links correct (95% interval 88.4 to 95.7%). Both were adjudicated with an AI assistant, not by a person. Both found that errors concentrate in links the open record no longer attributes to the company, so on 6 October 2026 every verified link was rechecked against the current record and 14,537 links it no longer supports (the company or the university no longer appears on the paper) were set aside before the final scores were calculated. An independent human check of 30 audit items is pending; its result will be published here as a dated update. Every count and reason is published.

4. The four pillars

• Co-creation is field-normalised: each university's verified industry works are compared with the number expected from its own output in each of 252 subfields at the cohort's industry rate for that subfield, so a medical or engineering profile is neither rewarded nor penalised for its mix. The ratio is shrunk toward the cohort average by empirical Bayes (Gamma-Poisson), so small numbers cannot produce extreme values.
• Durability is a partnership-weighted share. Its counting unit is the company-work incidence: each verified joint work counts once for every company group on it (half where the only company author also holds a post at the university), and consortium papers are left out. Durability is the share of a university's incidences that fall in sustained partnerships (company groups with five or more single-affiliation joint works, active in at least three of the five years), shrunk toward the cohort average with 20 pseudo-incidences. Example: a paper written with three company groups, two of them sustained partners, adds 3 to the denominator and 2 to the numerator; a paper whose only company author is also listed at the university adds 0.5 to the denominator (and 0.5 to the numerator if that company is a sustained partner), and never counts toward the five works that make a partnership sustained.
• Scale is the mean of three percentiles: reach (company groups with two or more joint works), sustained partnerships (the count), and influence, which is PageRank (damping 0.85) in the two-sided network of universities and company groups, with fractional edge weights (a work shared by several universities and companies is split between them). Consortium papers are left out of the network. Scale explicitly captures network size.
• Breadth is Gini-Simpson diversity (small-sample corrected) across the main research fields of a university's company partners. Partner countries are not scored, so a university in a country with a large domestic industry is not penalised; each profile shows its domestic and state-owned shares instead.

Terms. Field-weighted citation impact (FWCI): citations compared with the world average for works of the same field, type and year (1.00 is average). PageRank: a network measure in which a link to a well-connected partner counts for more. Percentile: the share of ranked universities with a lower value.

5. Scores, bands and uncertainty

Each pillar becomes a percentile among the ranked universities; the score is the weighted average of the percentiles (0 to 100). Bands follow the Research Collaboration Index: Platinum for ranks 1 to 100, Gold 101 to 300, Silver 301 to 600. Every rank carries a sensitivity range: the index is re-run 200 times, resampling the evidence (Poisson) and moving every weight up or down by as much as 25%, and the range shown holds 90% of the resulting ranks, widened where needed so that it always contains the published rank. These are sensitivity ranges under those stated scenarios, not formal tests of equality: overlapping ranges do not prove two positions are the same, and they do not model every source of error (affiliation, company-grouping and sector-classification errors are not resampled). Leading positions are sensitive to evidence and weighting choices. Compare the published sensitivity ranges alongside the point estimates. 351 universities have a sensitivity range that crosses a standing-band boundary; each profile carries a note such as "Published Platinum standing; the sensitivity range crosses the Platinum/Gold boundary."

Validation against an independent open ranking

The CWTS Leiden Ranking Open Edition 2025 (CC0) publishes, for 2020-2023, the share of each university's publications co-authored with organisations the open record types as industry, without checking the affiliation text. Across the 1,147 universities in both, the share of output co-authored with industry ranks the universities in much the same order (Spearman rank correlation 0.82), and the ICI's field-normalised co-creation measure correlates at 0.80. The two are not identical, by design: they cover different years, and the ICI counts only links the authors' own affiliation text confirms. In the ICI's own ledger, a typical university keeps 50% of its candidate industry works after verification; those that lose most are those whose own addresses, buildings or departments were read as companies (the University of Haifa's address on Mount Carmel, for example, was read as a company called Carmel). This is a check on the measure, not an endorsement by CWTS.

6. Sectors

Each of the 4,516 OpenAlex research topics is mapped to at most one of 21 sectors by published rules on the topic's name and field (emerging sectors are tested first, so a battery-materials topic counts as Batteries, not Materials). The full mapping is downloadable. Every topic was then read by hand, and topics whose names matched a rule but whose substance belongs elsewhere were reassigned by name (the list is in the code and the download). Where a topic straddles two sectors, the title of each work settles it: ground-vehicle aerodynamics belongs to mobility, not aerospace; a work in a security topic with no security content belongs to its actual subject; robotics and space works must be about robots or space. A university is ranked in a sector with at least 20 verified industry works in it; the sector score weights volume 30%, reach 25%, repeat partnerships (three or more works over two or more years) 25% and network influence within the sector 20%. Its pool is the number of universities eligible in that sector, shown on every sector page. A sector ranking is published only when at least 10 universities are eligible: Robotics and Autonomous Systems has 4, so its ranking is withheld, and 21 sectors are covered with 20 sector rankings published. A sector's award pool is different: it counts university and company partnerships (five or more joint works over two or more years), not universities, so a sector can have award finalists while its university ranking is withheld. Sectors with fewer than 50 eligible universities are labelled provisional: their positions move more with small changes in evidence. About half of all verified industry works sit in no sector, because their topics belong to no industry (basic science, for example); they count fully in the overall index.

7. The Sector Awards

Candidates are verified university and company-group partnerships with at least five joint works in the sector over at least two years (for the Global Partnership of the Year, ten works over three years), excluding consortium papers. They are scored on volume (40%), years active (30%) and field-weighted citation impact shrunk toward 1.0 (30%). The ten highest-scoring become finalists, with no more than one per university, two per company group and four per university country in a category (the caps shape the shortlist, never the rankings); five reserves are held in case a finalist withdraws or a correction removes one. Each finalist is represented by an academic at the university and a researcher at the company who co-wrote the partnership's most-cited joint paper not already used elsewhere (guidelines, protocols, perspectives and papers with more than 30 authors are not used as evidence), each with at least two joint works in the partnership and neither listed at both organisations; no person represents more than one finalist. The winner in each category is the finalist with the highest partnership score in the frozen Edition 1 record; no panel, nomination or payment plays any part. The finalists and the winners are announced together with the index at 09:00 BST, Thursday 15 October 2026, after the corrections and opt-out window closes at 17:00 BST on Tuesday 13 October; a researcher's opt-out never changes a winner (the person is simply not named), and a partnership leaves a shortlist only through an accepted correction or its university's withdrawal, when the next reserve takes its place. On each award page the winner is listed first and the other finalists follow in alphabetical order of university. The Sector Awards recognise partnerships between universities and independent companies. A partnership is not eligible for an award when (1) the company or its ultimate parent is state-owned or state-controlled; (2) the company, its ultimate parent or the university appears on a UK, EU, UN or US sanctions or restricted-party list, including the US Entity List, the Non-SDN Chinese Military-Industrial Complex Companies list and the Section 1260H list; or (3) the company is primarily a defence or military producer. The screen applies to every country in the same way. It affects award eligibility only: every verified partnership still counts in the index, the sector indices and each university's evidence, because the index measures the open scholarly record as it is. In Edition 1 the screen covered every university and company group in the award candidate pool (2,175 company groups and 1,041 universities, screened on 6 October 2026 against the US Consolidated Screening List (including the Entity List, the SDN list, the Non-SDN Chinese Military-Industrial Complex Companies list and the Military End User list), the Section 1260H list (FR Doc 2026-11571, dated 5 June 2026, published 10 June 2026), the UK Sanctions List, the EU consolidated financial sanctions list and the UN Security Council consolidated list). It made 319 company groups ineligible (285 state-owned or state-controlled, 52 on a list, 19 primarily defence producers; some meet more than one rule) and 17 universities on a list. In each category the next eligible partnerships took their places under the same caps. The robotics sector ranking is withheld because fewer than 10 universities are eligible for it, but the Robotics award still runs on its own screened partnership pool: the award pool and the ranking pool have different eligibility rules.

Changes made before publication (5 and 6 October 2026)

• Network rule. The code now follows the published rule: consortium papers (100 or more authors) are left out of the PageRank network. Against the earlier build this moved no university between bands, changed no finalist and no winner; 175 of 1,016 ranks moved, by at most 5 places.
• Company groups. Seven company groups that the open record split in two (a parent company and its own subsidiaries filed as two groups) are now each one group, identified by the top of the ownership chain. With the network rule, this moved 8 universities across a band boundary (each within 2 places of it), and changed shortlists in three categories; it changed no winner.
• Award eligibility screen. Introduced as described above. It changes award shortlists and winners only; no index, sector or university figure depends on it.
• Verification refresh (6 October 2026). Every verified link was rechecked against the current open record and 14,537 links it no longer supports were set aside. Where a university has been reorganised and the record now files its authors under the successor institution, the link stays with the university (137 links). Edition 1 keeps the cohort as defined, the same entities as the Research Collaboration Index: links the record now attributes to the University of Colorado Anschutz Medical Campus remain with the cohort entity for the University of Colorado Denver (532 links). That split will be revisited for the 2027 edition.

8. What the index is not

It is not a measure of research income, teaching, reputation or overall quality, and a lower position is not a judgement on a university. Industry structure differs between countries, which is why every profile also shows its position in its country and region. Collaboration that leaves no trace in publications (confidential contract research, consultancy, licensing, patents) is not visible to it; Edition 2 (September 2027) adds patent evidence once a worldwide patent build is validated. Market-producing state-owned enterprises count as industry, following the OECD Frascati Manual; their share is shown on every profile and every such company is marked. Company names appear as they do in the open record; listing is not an endorsement, and the index takes no position on any company's regulatory or sanctions status.

Commercial disclosure. Global Research Partnerships sells services to universities, including the Industry Partnership Blueprint, and some universities in the index are clients. No commercial relationship affects any figure: every score is computed by published rules from open data, and the Edition 1 figures are frozen at launch, once the correction window has closed and every accepted correction has been applied.

Independence. The index is free, public and independent. No commercial relationship with any university or company changes any position, and no one can pay to be included, excluded or moved.