Science Quality and the Value of Inventions

1 Introduction

• Among patents directly building on science, patent value increases monotonically with the quality of the cited science.The paper reports this relationship across scientific disciplines, technology areas, and time.Source: §1 · p. 2
• Publications with higher scientific quality are much more likely to be cited in patents and are cited at a higher rate.Source: §1 · p. 2
• The results hold across scientific disciplines, technology areas, and time.Source: §1 · p. 2

2 Data

• The dataset covers approximately 43 million Web of Science publications from 1980 onward and more than 4.8 million patent families filed from 1985 to 2012.Patent families must have at least one grant publication at the EPO or USPTO; they protect inventions in developed countries with more than one billion inhabitants.Source: §2 · pp. 2–3
• About 1 million patents link to scientific publications through 7.0 million SNPL references, covering about 2.2 million publications.SNPL denotes scientific articles among the non-patent literature cited by patents.Source: §2 · p. 3
• Scientific quality is measured by publications’ three-year citation counts, while patent value is measured by five-year forward citations from US patents.A patent’s SNPL science quality defaults to the highest-quality publication among its scientific references.Source: §2 · p. 3
• The analysis tests robustness using journal impact factor, alternative aggregations of multiple SNPL references, and other patent-value measures.The default aggregation uses only the highest-quality SNPL reference when a patent has several.Source: §2 · p. 3

• At the top 1% of scientific quality, 40% of publications appear in patent references; the share exceeds 60% at the top 1 permille and reaches 80% beyond the top 1 permyriad.The corresponding average counts in distinct patent families are 7.7 and 21.9 for the top 1 permille and top 1 permyriad, respectively; limiting science citations to the first three years excludes feedback from important patents.Source: §3 · pp. 3–4, Figure 1 · p. 4
• Patents linked to top-quality science receive more than twice as many forward patent citations as patents linked to bottom-quality science, while bottom-quality patents are as valuable as those without science references.The comparison uses five-year US patent forward citations and science quality based on the highest-quality reference when a patent cites multiple publications.Source: Figure 1 · p. 4, §3 · pp. 4–5
• Excluding patents’ self-references yields results close to the main pattern, supporting a link between high-quality science and high-value technology beyond the organizations producing both.The share of self-references is roughly similar across science-quality levels and tends to decrease at higher levels.Source: §3 · pp. 5–6, Figure 2 · p. 6
• The science-quality/value relationship largely persists through indirect patent links: top-10% science-linked patents remain about three times as valuable as bottom-10% patents even far from the science frontier.The positive correlation begins fading only beyond a distance of six steps from the frontier.Source: §3 · pp. 6–7, Figure 2 · p. 6
• Shorter time lags between the patent filing and its highest-quality scientific reference are always associated with higher patent value, and strengthen the positive science-quality/value relationship.The marginal effect of time lag is small at low science quality but patent value rises sharply with shorter lags at high science quality.Source: §3 · pp. 6–7, Figure 2 · p. 6
• The positive relationship is reported as robust across alternative patent-value measures, including EPO citations, monetary-value estimates, and patent-claim length.The monetary measures use limited samples of about 899,000 and 11,000 patents; Table 1 reports descriptive averages for patents without science references and those in the top and bottom 10% of science quality.Source: §3 · pp. 7–8, Table 1 · p. 8

4 Conclusions

• The study challenges the perception that science is an independent upstream activity, detached from technological progress and society.Scientists need to gauge and acknowledge science’s contributions to society and future technical and social advancements.Source: §4 · p. 9
• Scientific quality in publications referenced by patents has a robust, strong relationship with the patents’ impact and commercial value.The findings challenge the view that science quality and technology are largely detached.Source: §4 · p. 9
• The findings are descriptive; industrial users may select promising science through their own assessments, and the observed correlation is unlikely to reflect spurious selection.The exact causes remain for future work; the authors doubt industrial users rely merely on science citation counts or impact measures.Source: §4 · p. 9
• Science quality measured by scientists strongly predicts technological applicability and practical value, but the study does not establish that the alignment is optimal.The findings contradict the notion that citation- and impact-oriented science funding leads to research of little societal value; academic measures may also inform commercial investment decisions.Source: §4 · p. 9

Scientific literature data

• The publication dataset contains 43 million Web of Science research articles published from 1980 to 2016, with bibliographic, affiliation, field, and citation information.3Source: Scientific literature data · p. 11

Patent data

• The patent sample contains more than 4.8 million DOCDB families with an EPO or USPTO grant publication and first filing from 1985 through 2012.4The study includes search- and examination-phase references from EPO, USPTO, and WIPO filings; DOCDB’s enriched XML separates bibliographic fields to improve matching with scientific literature.Source: Patent data · p. 11

Data transformation

• Data preparation uses log(x + 1) for variables with natural zeros, selects the highest available patent-value estimate, and winsorizes specified extreme values.The study restricts scientific references to publications dated no later than a patent family’s first filing year; it also assigns fields by stated tie-breaking rules and drops families without a technology-field classification.Source: Data transformation · pp. 11–12

SNPL references as a measure of knowledge input

• SNPL references are treated broadly as evidence that a cited scientific paper contains information relevant to understanding and developing the technology.The study does not require interpreting each reference as a direct knowledge flow; prior work found SNPL references more closely related to actual knowledge flows than other reference types.Source: SNPL references as a measure of… · p. 13

SNPL matching methodology

• Of roughly 37 million NPL references, 27 million (71.8%) were selected as potential targets, and 13 million (47.1%) met the minimum match-quality score of three.Match quality sums six indicators: year, volume, page(s), first author, journal title, and article title.Source: SNPL matching methodology · p. 14
• A match-quality cutoff of three or higher yielded precision of 0.99 and recall of 0.96 for EPO and 0.92 for USPTO references.Precision is the share of correct matches among algorithm matches; recall is the share of recoverable targets matched successfully.Source: SNPL matching methodology · pp. 14–15
• Because Web of Science coverage is incomplete, matching unmatched references to Scopus added 113,340 SNPL links to 49,254 items, under 2% of the total, which the study excludes.The additional Scopus links covered publication years 1996–2016.Source: SNPL matching methodology · p. 14
• The final dataset contains 952,932 DOCDB patent families with at least one grant publication and one matched scientific-publication reference.The matched publications comprise 2,229,658 distinct scientific articles linked to patent families from 1985–2017.Source: SNPL matching methodology · p. 14

SNPL self-references

• Self-references are SNPL links sharing an inventor-author name or an applicant-publication affiliation, and the study reports analyses both including and excluding these categories.The two categories may overlap.Source: SNPL self-references · p. 15

Related literature on SNPL references

• Prior findings on science and patent value are mixed: Harhoff et al. (2003) report a positive association with NPL-reference counts, while Gittelman and Kogut (2003) find influential science does not necessarily lead to influential patents.Suzuki (2011) finds a strong positive effect on technological value but a weak negative effect on commercial value, with heterogeneity across fields.Source: Related literature on SNPL references · p. 16
• Fleming and Sorenson (2004) describe science as a map that can guide invention toward useful configurations or reduce unproductive search, with returns depending on the difficulty of the inventive problem.Using 16,822 USPTO patents granted in May and June 1990, they report that scientific references increase patent forward-citations with an elasticity of about 10%.Source: Related literature on SNPL references · p. 17

Scientific citations

• Forward citations to publications serve as the main bibliometric measure of scientific quality, reflecting their influence on follow-on research.The measure counts citations within a three-year window from publication. The main specifications assume publication citations are independent of patent citations; robustness analyses exclude citations from patents filed within three years of publication.Source: Scientific citations · pp. 17–18

Journal impact factor

• Journal impact factor (JIF) is a predetermined alternative, but is available only after 1997 and masks quality variation within journals and over time.Because JIF is predetermined when a publication appears, it avoids concerns about truncation and mechanical correlation with patent value.Source: Journal impact factor · p. 18

Patent level aggregation of SNPL references

• Patents with SNPL average 7.4 references, and 64.2% cite multiple publications; patent-level quality is the maximum across references, and self-reference follows the top-quality publication.The maximum is used because forward citations are highly skewed, making the most-cited publication or highest-JIF journal more indicative than the average.Source: Patent level aggregation of SNPL… · p. 18

Patent citations

• Forward citations are the main patent-value proxy, counted as USPTO citations within five years of first filing; EPO citations are analyzed separately for robustness.Patent citations document technological relatedness and novelty or patentability, unlike scientific citations, which recognize relevant prior work. USPTO citations are used in the main specifications; the EPO measure counts examiner-supplied citations only.Source: Patent citations · p. 19

Patent scope

• The alternative scope proxy is the logarithm of the word count in a patent’s first independent claim, with longer claims associated with narrower protection.Patent scope is treated as proportional to value: narrower protection is considered less valuable, and longer claim descriptions tend to specify more features.Source: Patent scope · p. 19

Measures of monetary value

• 1,029,987 patent families have Kogan-based stock-market-return values, including 230,389 with SNPL references.These values proxy private patent-grant value and cover US-member families with at least one publicly listed US-company applicant. PatVal survey assessments cover 11,061 families, including 2,579 with SNPL references; citations and claim length remain indirect measures of economic value, and citation counts may also reflect technological or social value rather than owners’ private value.Source: Measures of monetary value · p. 20

Selection of scientific publications into SNPL references

• Publication selection is modeled as a binary probability or log-transformed count, with scientific quality measured by citations received during the three years after publication.The analyses use field-by-publication-year fixed effects to account for differences in science quality and SNPL frequency across fields and over time.Source: Selection of scientific publications… · pp. 20–21

Science quality and patent value: residualized variables

• Residualization subtracts each patent’s technology-area-by-first-filing-year mean and adds back the overall mean, aligning plotted relationships with regression results.Both science-quality and patent-value variables are residualized using the full patent sample, including patents without SNPL references.Source: Science quality and patent value:… · p. 21

Science quality and patent value: regression models

• Patent-value regressions use five-year USPTO citation counts as the main outcome, with EPO citations, monetary-value indicators, and claim length as alternatives.Outcomes are log-transformed with offset 1; Kogan et al. (2017) measures capture only additional value associated with a US grant, while Giuri et al. (2007) survey wording more precisely targets private patent value.Source: Selection of scientific publications… · p. 20, Science quality and patent value:… · p. 22

Selection of scientific publications in SNPL references

• Science quality consistently has a positive, significant association with articles’ selection into SNPL references, including analyses excluding academic patents and self-references.The analyses vary the reference window and whether they include only the highest-quality reference or first-time applicant–publication pairs.Source: Selection of scientific publications… · pp. 22–23

Main regression results: SNPL science quality and patent value

• Table S-T5 reports elasticity estimates across alternative patent-value measures, adding patent-level controls and applicant fixed effects to a technology-field-and-year baseline.All specifications control for whether a patent has at least one SNPL reference.Source: Main regression results: SNPL science… · p. 23

Alternative measures of SNPL science quality

• Results are similar using journal impact factor or alternative ways of aggregating multiple SNPL publications’ quality at the patent level.The alternatives include the sum, average, and root of the sum of squares alongside the main maximum-based measure; USD-value results are unstable in two early specifications but positive and significant in the most complete specification.Source: Alternative measures of SNPL science… · p. 23, Self-references · p. 23

Self-references

• Positive and significant estimated elasticities persist both among patents with self-references and among those without, so self-references do not drive the overall effects.About 5–10% of patent families include a self-reference, most often an inventor self-reference; applicant self-references occur in 2–4%. The estimates are larger for patents with self-references.Source: Self-references · pp. 23–24

Technology fields, year of patent filing, and applicant countries

• The relationship is particularly strong in Chemistry, also matters in Electrical Engineering, Instruments, and Mechanical Engineering, declines substantially after 2000, and is especially strong in Europe, the US, and Japan.The decline after 2000 is especially pronounced in the US system; results are consistent across the geographic areas considered.Source: Technology fields, year of patent… · p. 24

Interdisciplinarity of SNPL references

• Interdisciplinarity is associated overall with higher patent value except at the top of the SNPL science-quality distribution, while the quality-value correlation remains highly positive for both groups.Patents with interdisciplinary SNPL references are most common at intermediate values of SNPL science quality.Source: Interdisciplinarity of SNPL references · p. 25

Distance of SNPL references

• Patent families with shorter-distance SNPL references have particularly high value and tend to show higher elasticities, while elasticities remain strongly positive and significant at greater distances.This pattern applies to distance and time-distance of references.Source: Distance of SNPL references · p. 25

S-4 Supplementary graphs and tables

• Supplementary figures examine how science quality relates to SNPL citation probability and counts, and how the patent-value relationship varies over time.The time analysis plots interactions by first filing year for US and EP patent forward citations, with and without patent-level controls.Source: S-4 Supplementary graphs and tables · pp. 27, 29
• Maximum science quality is the default measure; sum and square-root-of-sum-of-squares estimates are virtually identical, while average quality yields a more diluted positive correlation with patent value.The comparison uses 5-year US patent forward citations and includes 4,767,844 patents, of which 952,932 have SNPL references.Source: S-4 Supplementary graphs and tables · p. 28
• Additional figures compare patent value by science quality across self-reference categories and by whether SNPL references are interdisciplinary.The captions define self-references as any, inventor, or applicant references and interdisciplinarity by journals associated with at least two WoS field codes.Source: S-4 Supplementary graphs and tables · pp. 29–30

S-5 Regression results

• 0.037 for 5y Cit US and 0.030 for 5y Cit EP are the reported elasticities for maximum 3-year citation quality of SNPL references.The table also reports −0.012 for EP Claim Length, −0.038 for Claim Length, −0.046 for USD Values, and 0.082 for EUR Values.Source: S-5 Regression results · p. 32
• 0.117 for 5y Cit US and 0.087 for 5y Cit EP are the reported elasticities for maximum journal impact factor of SNPL references.The table also reports −0.035 for EP Claim Length, −0.142 for Claim Length, −0.007 for USD Values, and 0.246 for EUR Values.Source: S-5 Regression results · p. 33
• For 3y Cit SNPL ref, elasticities without applicant fixed effects were 0.037 (max), 0.037 (sum), 0.042 (avg), and 0.038 (sq).With applicant fixed effects, the corresponding elasticities were 0.027, 0.026, 0.030, and 0.027.Source: S-5 Regression results · p. 34
• Interdisciplinary coefficients were 0.054, 0.047, 0.040, and 0.037, while science-quality interactions were positive for both single-discipline and interdisciplinary references.The single-discipline interaction coefficients were 0.042, 0.031, 0.035, and 0.027; the interdisciplinary interactions were 0.032, 0.024, 0.023, and 0.017, across 5y Cit US and 5y Cit EP specifications.Source: S-5 Regression results · p. 38
• For 5y Cit US, frontier-distance coefficients fell overall from 0.818 and 0.753 at distance 0 to 0.035 and 0.166 at distance 10.The 3y Cit SNPL ref (max) coefficient was 0.056; estimates are elasticities.Source: S-5 Regression results · p. 39
• For 5y Cit US, short-distance coefficients were 0.141/0.116 versus medium-distance 0.052/0.040; for 5y Cit EP, they were 0.089/0.076 versus 0.039/0.028.Source: S-5 Regression results · p. 40