Universities beat federal agencies at turning research dollars into patents, and mid-size STEM/medical schools out-produce elite research universities per dollar

Summary

For universities, this corpus can compute a real patents-per-federal-dollar rate by joining NSF HERD institutional R&D expenditure to USPTO patent grants through the shared institution identity in ref.canonical_org_entity. Over 2018-2022, mid-size, STEM/medical-heavy schools such as Texas A&M-Corpus Christi, the University of Toledo, and Baylor University post patents-per-federal-dollar rates 2-5x higher than the elite research powerhouses (MIT, Caltech) that receive far more total federal R&D money and file far more total patents. Scale (raw R&D dollars, raw patent counts) and efficiency (patents per dollar) point to different "winners."

For federal agencies, this corpus and the public record together cannot support a clean per-agency invention-per-dollar figure: patent records here do not carry the government-interest funding statement that would attribute a specific patent to its funding agency, and GAO's own FY2020-2024 review of iEdison invention-disclosure data (the metric agencies actually use) reports disclosures only in aggregate across the four dominant funders (DOD, HHS/NIH, DOE, NSF = 97% of all disclosures) rather than broken out per agency per year. What is computable here is agency R&D obligations: HHS ($74.1B) and DOD ($72.4B, summing its sub-agencies) obligated roughly 2x NSF's ($6.9B) and 5x DOE's ($13.7B) FY2022 total, so on obligations alone HHS and DOD are the largest inputs — but a defensible inventions-per-dollar ranking of agencies is not something this data supports, and is flagged as a genuine gap rather than estimated.

Methodology and known limitations

• University side: research.nsf_herd_by_institution filtered to funding_source='Federal', rd_field='All', summed 2018-2022 (a window chosen to avoid patents.patent_grants' interior data gap for 2023 and its anomalous 2024 row count, both confirmed via data_coverage), joined through ref.canonical_org_entity (nsf_herd_inst_id to patents_assignee_id) to a distinct patent count per assignee from patents.patent_grants. Only 222 of thousands of NSF HERD-surveyed institutions carry a resolved patent-assignee match in this warehouse, so this ranks the subset with a confirmed identity link, not every federally funded university.
• Known mapping caveat: canonical_org_entity.patents_assignee_id is documented as lossy many-to-one — a name can map to the wrong assignee id (the schema's own comment cites a measured Schlumberger case). One result, National Defense University showing 290 patents against $41.6M in federal R&D, is almost certainly such a mismatch (a small military graduate school with no engineering patent program) and has been excluded from the ranking below rather than reported as fact. The remaining top entries are plausible but not individually re-verified against USPTO assignee text, so treat the university ranking as directionally reliable, not exact to the patent.
• Agency side: research.nsf_federal_rd_obligations covers only FY2022 in this warehouse (confirmed via data_coverage — 721 rows, one fiscal year). It reports obligations by funding agency times performing sector, not by resulting invention — there is no join key in this corpus from a federal obligation to the specific patent(s) it funded (that link lives in a patent's government-interest statement, which PatentsView does not carry). Government-wide invention counts by agency come instead from GAO-26-107971 (iEdison data, FY2020-2024), which reports only that DOD+DOE+HHS/NIH+NSF combined account for 97% of the government's 56,260 disclosed inventions — not a per-agency, per-year count usable for a dollar-weighted rate.