Working Paper — Please contact whalen@hku.

hk for updated draft

A brief “Law & Technology” contextualization

This paper is structured somewhat differently than most law papers. It is intended for a
generalist audience and therefore must comply with what are usually shorter word count
limits. I’ve included below both the current working draft of the paper, as well as the
Supplementary Information appendices, where we provide some more detail on empirical
methods, robustness checks, etc. If the formal empirical methods seem too foreign at first, try
to get an intuitive or “gut” understanding of what we’re doing rather than focusing on all the
details. I’m happy to discuss the details more in person, and look forward to discussing the
paper with you all. As you read and consider it, it might be helpful to consider two distinct
ways the paper engages with law & technology:
1. From an intellectual property law/innovation policy perspective, this paper raises
questions about the role the law should have (if any) in fostering diversity among
those who produce technology.
2. From a research methods perspective, this paper demonstrates how technology can be
used to reveal biases or inequities in systems that apportion legal rights.

Ryan Whalen
October 18, 2022
Working Paper — Please contact whalen@hku.hk for updated draft

The Glass Ceiling in Boundary Pushing Innovation

ABSTRACT:

We find systematic differences in the patenting experiences between male and female
inventors. Women inventors’ boundary-pushing inventions are rejected more than men’s
whereas women’s conventional inventions show no differences in patenting rates than men’s.
Patent grant rates for men increase when they make novel connections between technological
domains, while women face an increased chance of rejection. Using a diverse sample of about
7 million international patent application records between 2001–2018, we found that female
inventors have consistently lower success rates than men when they seek patent protection for
boundary spanning inventions. Further, while women inventors are penalized for making
boundary spanning inventions, men benefit from doing so and enjoy increasing success rates.
Finally, inventors who use their first initial rather than their first name on patent applications
have success rates for novel inventions that resemble men inventors rather than women
inventors experience. When inventions are mixtures of conventional ideas, women and men
inventors show no differences in patenting success. These differences exist after controlling
for technology, team, and negotiating strategies. When performing innovative work, men are
more likely to be rewarded as boundary spanners, while women are more likely to be rejected
as boundary pushers.

BODY:

In 1816 in Geneva, 19-year-old Mary Shelley, Percy Shelley, Lord Byron, and a friend
challenged each other to come up with a ghost story. Mary Shelley conjured up a radically
novel story. Frankenstein was published anonymously and received high accolades for
innovatively spanning diverse ideas from biology, electricity, and social theory. In the second
edition of the book published a few years later, Mary Shelley was credited with authorship.
Inexplicably, critics panned the book and a few years after that the sale of the book’s copyright
fetched 31 pounds, a quarter of its first year’s earnings.
What was true about the innovation glass ceiling in the 19th century literary world,
remains true in a new frontier of creativity—the world of technological innovation. There is a
a widespread and consistent gender gap in science and innovation work.1–4 This leads not only
to gender inequalities in terms of work opportunities,5 salaries,6 and educational choices,7 but
also to underutilization of resources as society fails to draw from the full diversity of
perspectives available to help solve its most pressing problems.
This gender gap extends to patenting success.2,4,28,29 Figure 1 examines the success rate of
patent applications. It pools all completed patent applications (N=4,754,198) by the inferred
gender of the first listed inventor(s). Fig 1A demonstrates that the patent grant rate for female
inventors is significantly lower than for male inventors (Figure 1) or inventors with unknown
genders. Fig 1b indicates that the gender difference is consistent across years and with little
change in the mean difference despite the growing number of women in science and
technology-based fields.
Figs 1c and 1d reveal that this gender gap extends beyond the success and failure rate of
patent applications. Even when they are granted patent protection, inventions by women lose
more claims during the patent examination process (Fig 1c). Because they provide the legally
enforceable aspect of a patent’s property rights, claims are important measures of patent value30
and so losing claims in examination means the loss of more intellectual property value for
women inventors. We also see that, when they are granted it, women inventors wait longer than
Working Paper — Please contact whalen@hku.hk for updated draft

men to receive patent protection (Fig 1d). Delays in receiving patent protection can slow
business developments, impede product launches, and make it more difficult for start-up firms
to secure investment support (for more information on the gender differences in granted patents
see supplementary text S11).

Figure 1. Women inventors have poorer outcomes than men across a variety of success measures. They are less
likely to be granted a patent than men or inventors with ambiguously gendered names (a). This success rate
difference is consistent across recent years (b). When they are granted a patent, they lose more claims during the
examination process (c), and they experience more delay to grant (d). Analysis includes all granted or
abandoned applications from 2001–18 for majority gender teams (n=4,754,198).

There are many possible mechanisms underlying the patenting gender gap. We explore
these to determine whether the success gap is explained by differences between technical fields,
team sizes, firm sizes, and experience. Finding none of these explain the patenting gender gap
we then turn to the degree to which claimed inventions fit within our pre-conceived notions.
We do this by focusing on the degree to which an invention makes unexpected combinations
of technical information—essentially pushing on the pre-existing scientific and technical
Working Paper — Please contact whalen@hku.hk for updated draft

boundaries. We then examine whether or not men and women inventors are treated differently
when they push these boundaries.

Boundary pushing innovation holds a special place in the innovation process because it
tends to break new ground in conventional domains or establish new research fields.8,9 This
work departs from conventional thinking10 by combining pre-existing knowledge in new
ways11,12 leading to a higher potential for ground-breaking innovation.8,9 Nevertheless, despite
the greater potential of boundary pushing innovations for creating value, their value is harder
to objectively evaluate given the lack of comparative inventions,13 introducing into the
assessment process non-invention related criteria such as the inventor’s gender. For example,
women inventors are less likely to be granted patents than male inventors, except in the domain
women’s health products where they are more likely to be a granted a patent than men
inventors.14 Nevertheless, it is unknown whether boundary pushing innovation and inventor
gender is related.
Gender could interact with boundary spanning scientific and technological discoveries in
two theoretical ways. It is possible that there is a decreased gender gap for women who make
boundary spanning discoveries as the rewards for their novel contributions to knowledge help
overcome the gender barriers they face. Alternately, boundary spanning contributions could be
so new and unusual that their value is unclear and traits such as gender become more salient,
increasing the gender gap.
We engage with this question by leveraging the rich record of innovation success and
failure provided by USPTO patent application data. Using a measure that allows us to estimate
how atypical,11 nonobvious,15 or boundary spanning16 the technical areas combined in a patent
application are, we show that women inventors are more likely than men to have their
applications denied when they make boundary pushing inventions. We also analyze patent
application data from the United Kingdom and Canada and find evidence that this phenomenon
is not unique to the United States. This suggests that there is indeed an innovation glass ceiling.
Identifying this discrepancy reveals a previously undocumented facet of the research gender
gap while also providing further evidence of how biases faced by women in STEM fields
manifest in practice.

Data & Methods

Our empirical materials include the patent classification records for all 6,079,571
applications filed between 2001 and 2018 as well as the application details and examination
records for the 393,391 applications that list more than 2 classification subgroups and are thus
most likely to push scientific and technical boundaries.
Each application is assigned by the patent office into a variety of technical areas it
pertains to, according to the “cooperative patent classification” (CPC) system. We leverage the
tendency to categorize inventions into multiple categories to produce a CPC co-classification
network that measures how typical or atypical a combination of categories is 15,17. The co-
classification network provides insight into which technical areas are commonly combined and
which are more rarely listed together on the same patent application. This gives us insight into
whether an invention makes an incremental improvement to established technologies or
produces something dissimilar to what came before. For instance, consider an invention that
improves upon a thermostat’s temperature sensing function. It might be categorized into CPC
subclasses “F24F” (air-conditioning related inventions), and “G05D” (systems for regulating
non-electric variables like temperature). Such an invention would be common as these
categories have been combined many times in the past, suggesting it makes an incremental
improvement to thermostat technology. A newer type of thermostat invention that allows a
Working Paper — Please contact whalen@hku.hk for updated draft

smart phone to remotely set the temperature might add CPC subclass “H04L” (transmission of
digital information). The fact that H04L has rarely been combined with more traditional
thermostat categories reveals that the remote access thermostat invention is a boundary pushing
invention that represents a new combination of past categories making it less of an incremental
improvement to the status quo and more of a divergence from the technology that came before.
To explore the relationships between gender and success in patenting boundary pushing
inventions, we exclude simple inventions with only one or two CPC subgroup categorizations,
and focus on the patent applications that are most likely to push boundaries. We then use the
co-classification network to compute a metric for each patent application that represents the
degree to which it features common combinations of categories or spans technical boundaries
by combining areas rarely seen together (supplementary text S2). These scores are normalized
within category as z-scores representing how typical or atypical the CPC categories combined
on an application are.17 A score of 0 suggests an average application, negative scores represent
typical inventions with commonly combined CPC categories, and positive scores represent
atypical, boundary-pushing inventions that feature infrequently combined knowledge areas
(supplementary text S3). The face validity of our measure is bolstered in several ways. First,
when using a related method for computing the novelty of a patent,18 we find a high correlation
between that method and our method. Second, previous literature indicates that boundary
spanning research tends to have greater citation impact.11,19,20 Consistent with this prior
research, we find that inventions that made unusual combinations in our data do have higher
future citation impact (supplementary text S8).
We use an established and validated gender inference method to estimate gender
probabilities and assign each inventor to one of three categories: male, female, or unknown21
(supplementary text S1). This automated approach has several advantages. It allows us to
impute gender to a much larger number of inventors than would be possible manually. It also
corresponds to methods used both in the scientific literature22,23 and by the USPTO itself to
describe and analyze the patenting gender gap.24,25 Our estimated gender ratios largely agree
with those published by the USPTO in its studies of the patenting gender gap. For instance, as
of 2016 the USPTO estimated that 20.7% of US-based granted patents had at least one woman
inventor.25 By comparison, our method estimates of this share of patents was 20.3%, within
0.37 percentage points of the reported gender statistics. Name to gender inference algorithms
are capable of high precision,26,27 but they are not without limitations—both because of
ambiguousness in relation to some names and because gender is a non-binary and subjective
state. To address the possible impact of misclassification, we perform analyses with inventor
gender measured in three distinct ways—as a proportion of the team, as the inferred gender of
the first listed inventor, and as the inferred gender of the majority of inventors. We also provide
analyses on a subset of inventions that only have sole inventors and therefore only a single
inventor inferred gender (SI Reference).
We determine the status of each application as either granted (indicating the application
had claims that were accepted and patented), abandoned (indicating the application was not
successful), or pending (indicating the application is still under examination). We then focus
on the granted or abandoned patents and explore the relationships between gender, boundary
spanning innovation, and success at the patent office. Using logistic regression, we measure
the relationship between inventor gender, the atypicality of an invention, and the probability of
being granted patent protection. Our analyses include statistical controls for the size of the
inventing team, the prior successful patenting experience of each inventor, the gender of the
patent examiner assigned by the USPTO to process the application, the patent examiner’s
experience, and the size of the entity applying for patent protection. We also use the primary
cooperative patent subgroup classification as a technology area control variable. As robustness
checks on our findings about this relationship, we also perform analyses using different
Working Paper — Please contact whalen@hku.hk for updated draft

measurements of inventor gender—including inferred gender of the first listed inventor and the
team’s majority gender, and also analyze patent application data from the Canadian and United
Kingdom patent offices.

Findings & Discussion

We first examine alternate mechanisms to explain the patenting gender gap. Technology
area variation in patent success rates could explain these differences if men cluster in
technology areas that have higher granting rates.31 32 However, when we control for success
variation in 125 different technical areas, the female inventor penalty remains. We use a within
CPC subgroup logit regression, showing that even when we control for the different grant rates
in different technical areas the proportion of female members on the inventing team remains
negatively related to grant probability. This suggests that the different rates of success between
male and female patent applicants is not explained by differences between research fields
(supplementary text S4). Relatedly, the organization size of the entity applying for patent
protection also does not explain the observed gender gap.
Inventor and examiner experience could provide other possible explanations for
observed gender differences in patenting success. If inventor experience correlates with
success, this could explain the different success rates enjoyed by male and female inventors.
Research suggests that women are more likely to have shorter careers and drop out of STEM
fields.33 Similarly, research suggests that patent grant rates vary by examiner experience which
could affect our analysis.32,34 However, when we control for both inventor and examiner
experience the gender effect remains (supplementary text S5). In summary, when we hold
constant a wide variety of controls including inventor experience, examiner experience,
examiner gender, team size, and entity size and using a within CPC subgroup specification,
the odds ratio for an all-woman team being granted patent protection is 0.76 times lower than
that of an all-male team (Table 1). This demonstrates that the patenting gender gap is not solely
a product of these various alternate possible explanations.
Finally, the patent examination process involves examiners and applicants exchanging
information and viewpoints to determine patentability including discussions about the claimed
invention, responses to rejections, and possible appeals or requests for further examination.
Research suggests gender variation in negotiation strategies, with women tending to be less
assertive than men in working the review process and pushing back on negative examiner
comments.35 If men are more likely to “not take no for an answer” while “women don’t ask,”36
it could explain gender variation in patenting success. To determine whether this is the case,
we track the negotiation back-and-forth between examiners and applicants for each application.
By focusing on those applicant reactions to the patent office that demonstrate their willingness
to facedown rejection and assert that their invention is indeed patentable, we can compare the
strategies that men and women take when facing rejection. There are three types of examination
transaction of particular interest—responses to the examiner’s “non-final actions” that tend to
be initial rejections by the patent office, responses to a “final rejection” by the examiner, and a
“request for continued examination” that can follow a final rejection. Comparing these
reactions across all applications reveals that there is no significant difference between the rates
at which men and women respond to non-final actions, respond to final rejections, or request
continued examination after a final rejection. Was shown in Fig. 3, we do not find any
differences in assertiveness or negotiation strategies that would explain the patenting gender
gap (supplementary text S6).
Working Paper — Please contact whalen@hku.hk for updated draft

Figure 3. Women and Men Inventors Negotiate equally as Forcefully with the Patent
office for their Inventions. Interactions between examiners and inventors by gender and
interaction (transaction type). We see no significant differences in the way male or
female inventors respond to patent examiners, while we also see that women applicants
are more likely to be granted an extension than men. 4,754,198 granted or abandoned
patent applications are analyzed from the USPTO patent application data.

We next turn to the nature of the invention in question as a possible explanation. Prior
work suggests that a person’s ascriptive characteristics can be interpreted as a signal of quality,
particularly when the object being evaluated is new, different, or falls between conventional
categories.37,38 For example, start-up firms that innovatively combine established categories of
businesses into new businesses are reviewed more negatively than expected by stock
analysts.39,40 This work suggests that boundary-pulling inventions that build on familiar
combinations of technology as opposed to boundary-pushing inventions that combine past
ideas or technology in novel ways may be treated differently depending on the inventor’s
Working Paper — Please contact whalen@hku.hk for updated draft

gender. We find that women inventors are significantly more likely than men inventors to face
rejection when they produce boundary pushing inventions that stray from the status quo, while
inventions by males increase their probability of being granted as they make less typical
combinations of technical areas.
We measure boundary-pulling vs boundary-pushing innovations by measuring the
degree to which the patent is based on familiar or unusual combinations of past technological
domains.11,41,42 Upon submission, the patent office assigns every application a set of technical
classifications that reflect the areas of technology the invention draws from or is relevant to.
Using the method described above, we use the observed combinations of technical
classifications from all previous applications to measure how typical or atypical each
application’s combination of classifications is at the time of submission. To allow comparisons
across technical areas, these are normalized as Z-scores, with low scores for very typical
combinations and high scores for atypical combinations. This provides a quantitative measure
to capture the degree to which an invention pushes technical boundaries or not. Inventions
featuring common combinations of technical areas are more likely to represent typical
inventions that make incremental improvements to known technologies, and will accordingly
have a low atypicality score. Applications featuring rarely combined technical areas are more
likely to take novel approaches to known problems or introduce entirely new product
categories, and will have a high associated atypicality score.

Figure 4. Foreground: Probability of patent application success by invention atypicality and inventor gender. Includes
statistical control for team size, year, examiner gender, examiner experience, applicant entity size, and CPC class.
Background: Distribution of invention atypicality scores by gender.

Figure 4 presents the results of our regression analysis. On the x-axis is the degree to which
the patent is conventional versus boundary-pushing. Values below zero indicate patents that
Working Paper — Please contact whalen@hku.hk for updated draft

are increasingly conventional and increasingly atypical for values above zero. The y-axis
shows the estimated probability of patent acceptance (i.e. that a patent is granted). Plotted lines
represent the changing probability of acceptance for a inventors across five categories: 0%,
25%, 50%, 75%, or 100% women on the team.
The results show a strong interaction effect between gender, boundary-pushing
inventions, and patenting success. First, the probability of having a patent granted uniformly
increases for men inventors the more boundary-pushing the invention. The more men push
boundaries by producing inventions featuring new and unexpected combinations of technical
fields the more likely they are to be rewarded with a patent. Conversely, the probability of
acceptance falls for women inventors the more boundary-pushing the invention.
Second, we find that in the range of boundary conforming inventions the distinctions
between women and men inventors are for the most part statistically indistinguishable (Fig. 4).
In this range of typical inventions, the confidence intervals of the estimated probability of
acceptance overlap except for the inventions of the highest level of boundary-pulling where all
women team have the highest probably of acceptance.
Third, these relationships intensify the more gender homogeneous the team. The
greater the proportion of men on an inventor team, the greater the probability of acceptance
and vice versa for women at all levels of boundary-pushing innovation. For example, a team
of all men vs all women inventors at the highest level of boundary-pushing innovativeness have
a 6% better chance of being patented (.0685 – .0625). But a team of 75% men vs 25% women
inventors at the highest level of boundary-pushing innovativeness have a 3% better chance of
being patented (0.67-0.64). If applications with an atypicality score higher than 0 were granted
at the same rate for women majority teams as for male majority teams, this would result on
average in thousands of more patent grants and a loss of intellectual property valued at
hundreds of millions of dollars (supplementary text S7).
The results are also robust to data sources and hold when we examine UK and Canada,
patent applications or only US domestic origin patent applications. The differences also persist
when we measure gender in a variety of ways including first listed inventor’s gender, team
majority gender, or subset to inventions claimed by exclusively male or female inventors and
teams (supplementary text Add section). When we measure gender as the first listed inventor’s
gender, we see that applications with women listed as the first inventor have the same
atypicality penalty as seen above, but applications where the gender of the first inventor is
uncertain face no such penalty.
Although our available data is less complete for other patent offices, we also find
evidence of similar discrepancies in patent applications at both the UK and Canadian patent
offices, suggesting this is not an isolated occurrence (supplementary text S10).

Discussion

The innovation glass ceiling has widespread implications for science and technology
policy. It suggests the imbalances in STEM education and work opportunities and outcomes
extend beyond those that are established in the literature. In addition to the gender gaps in
employment (5), remuneration (6), promotion (35), and dropout (26), we see that women also
face a gender gap in the challenges of succeeding when they make boundary spanning
technological contributions. The innovation glass ceiling also means that fewer inventions that
make highly atypical combinations of technical fields get granted patent protection. This upsets
the balance of the incentive system established by patent law, while also possibly depriving
society of unrealized groundbreaking inventions that, but for the innovation glass ceiling,
would be patented and brought to market. Furthermore, prior research has shown that women
are more likely to produce inventions aimed at other women, especially in health care fields
Working Paper — Please contact whalen@hku.hk for updated draft

(36). So, an innovation glass ceiling that impedes women from making unusual groundbreaking
inventions could have outsized impact on women and women’s healthcare.
Some evidence suggests that the innovation glass ceiling is possibly a product of
stereotyping, where examiners fall back on beliefs about what correlates with a patentable
application—such as inventor gender and firm size—when assessing atypical inventions.
Alternate causal explanations could focus on the applicant side—for instance firms may be
more likely to abandon prosecution of boundary spanning patent applications by women
inventors. Future work should further explore mechanisms for the innovation glass ceiling.
Understanding precisely how and why female inventors face more barriers to success when
their work is boundary spanning is important to helping address gender gaps in research and
development work.
As with all systematic biases and uneven outcomes across STEM fields, dismantling
the innovation glass ceiling is no simple task. Nonetheless, there are policy changes that could
help reduce its severity. Education is perhaps the easiest intervention. Explaining the
innovation gender gap to patent examiners could help improve their awareness of stereotyping
and help address biases that might affect the patentability assessment when examiners are faced
with difficult decisions about atypical inventions. Another possible intervention would be to
mandate inventor-blind examination at the patent office. This would face logistical challenges
as examiners sometimes interact with inventors rather than their lawyers or patent agents. But
it may help address not only the innovation glass ceiling but also other biases that have
historically led marginalized inventors to use “passing” as white males as a strategy for success
at the patent office (37).
When she was sworn in as the Director of the USPTO in 2022, Kathi Vidal stated that
the patent office needs to do more to foster innovation in groups that have previously been
marginalized by the intellectual property system (38). Previous research on this topic has
largely focused on discrepancies in the rates with which these groups participate in the system
(2, 22, 37), with some work starting to assess the types of inventions produced by inventors in
marginalized groups (36). We expand on these perspectives by looking not to participation
rates or the substance of what is invented, but rather focusing on whether an invention
represents a typical product of existing research areas, or an atypical boundary spanning
combination of fields. We do so because boundary spanning work is often that which expands
the scientific or technical frontier, frames new research problems, and introduces new product
categories (11, 17, 39). If Director Vidal and others wish to ensure that women and other groups
marginalized by the intellectual property system are not excluded from contributing to the
expansion of scientific and technical frontiers, we need more work to help reveal, understand,
and dismantle the innovation glass ceiling.
Working Paper — Please contact whalen@hku.hk for updated draft

References
1. Huang, J., Gates, A. J., Sinatra, R. & Barabási, A.-L. Historical comparison of gender

inequality in scientific careers across countries and disciplines. PNAS 117, 4609–4616

(2020).

2. Jensen, K., Kovacs, B. & Sorenson, O. Gender differences in obtaining and maintaining

patent rights. Nature Biotechnology 36, 307- (2018).

3. Ding, W. W., Murray, F. & Stuart, T. E. Gender Differences in Patenting in the

Academic Life Sciences. Science 313, 665–667 (2006).

4. Frietsch, R., Haller, I., Funken-Vrohlings, M. & Grupp, H. Gender-specific patterns in

patenting and publishing. Research Policy 38, 590–599 (2009).

5. Long, J. S. & Fox, M. F. Scientific Careers: Universalism and Particularism. Annual

Review of Sociology 21, 45–71 (1995).

6. Hoisl, K. & Mariani, M. It’s a Man’s Job: Income and the Gender Gap in Industrial

Research. Management Science 63, 766–790 (2016).

7. Xie, Y., Fang, M. & Shauman, K. STEM Education. Annual Review of Sociology 41,

331–357 (2015).

8. Foster, J. G., Rzhetsky, A. & Evans, J. A. Tradition and Innovation in Scientists’

Research Strategies. American Sociological Review 80, 875–908 (2015).

9. March, J. G. Exploration and Exploitation in Organizational Learning. Organization

Science 2, 71–87 (1991).

10. Kuhn, T. S. The Structure of Scientific Revolutions. (University of Chicago Press, 1970).

11. Uzzi, B., Mukherjee, S., Stringer, M. & Jones, B. Atypical Combinations and Scientific

Impact. Science 342, 468–472 (2013).

12. Fleming, L. Recombinant Uncertainty in Technological Search. Management Science 47,

117–132 (2001).
Working Paper — Please contact whalen@hku.hk for updated draft

13. Whalen, R. Boundary Spanning Innovation and the Patent System: Interdisciplinary

Challenges for a Specialized Examination System. Research Policy 47, 1334–1343

(2018).

14. Koning, R., Samila, S. & Ferguson, J.-P. Who do we invent for? Patents by women focus

more on women’s health, but few women get to invent. Science 372, 1345–1348 (2021).

15. Pedraza-Fariña, L. G. & Whalen, R. A Network Theory of Patentability. The University

of Chicago Law Review 87, (2020).

16. Rosenkopf, L. & Nerkar, A. Beyond Local Search: Boundary-Spanning, Exploration, and

Impact in the Optical Disk Industry. Strategic Management Journal 22, 287–306 (2001).

17. Kim, D., Cerigo, D. B., Jeong, H. & Youn, H. Technological novelty profile and

invention’s future impact. EPJ Data Sci. 5, 1–15 (2016).

18. Whalen, R., Lungeanu, A., DeChurch, L. & Contractor, N. Patent Similarity Data and

Innovation Metrics. Journal of Empirical Legal Studies 17, 615–639 (2020).

19. Mukherjee, S., Uzzi, B., Jones, B. & Stringer, M. A New Method for Identifying

Recombinations of Existing Knowledge Associated with High-Impact Innovation. J Prod

Innov Manag (2015) doi:10.1111/jpim.12294.

20. Kaplan, S. & Vakili, K. The double-edged sword of recombination in breakthrough

innovation. Strategic Management Journal 36, 1435–1457 (2015).

21. Genderize.io | Determine the gender of a name. https://genderize.io/.

22. Ross, M. B. et al. Women are Credited Less in Science than are Men. Nature 1–2 (2022)

doi:10.1038/s41586-022-04966-w.

23. Yang, Y., Tian, T. Y., Woodruff, T. K., Jones, B. F. & Uzzi, B. Gender-diverse teams

produce more novel and higher-impact scientific ideas. Proceedings of the National

Academy of Sciences 119, e2200841119 (2022).

24. Toole, A. A. et al. Progress and Potential: 2020 update. IP Data Highlights 12 (2020).
Working Paper — Please contact whalen@hku.hk for updated draft

25. Toole, A. A. et al. Progress and Potential. IP Data Highlights 20 (2019).

26. Santamaría, L. & Mihaljević, H. Comparison and benchmark of name-to-gender

inference services. PeerJ Comput. Sci. 4, e156 (2018).

27. Karimi, F., Wagner, C., Lemmerich, F., Jadidi, M. & Strohmaier, M. Inferring Gender

from Names on the Web: A Comparative Evaluation of Gender Detection Methods. in

Proceedings of the 25th International Conference Companion on World Wide Web 53–54

(International World Wide Web Conferences Steering Committee, 2016).

doi:10.1145/2872518.2889385.

28. Hunt, J., Garant, J.-P., Herman, H. & Munroe, D. J. Why are women underrepresented

amongst patentees? Research Policy 42, 831–843 (2013).

29. Bell, A., Chetty, R., Jaravel, X., Petkova, N. & Van Reenen, J. Who Becomes an Inventor

in America? The Importance of Exposure to Innovation*. The Quarterly Journal of

Economics 134, 647–713 (2019).

30. Zeebroeck, N. van & Potterie, B. van P. de la. Filing strategies and patent value.

Economics of Innovation and New Technology 20, 539–561 (2011).

31. Cheryan, S., Ziegler, S. A., Montoya, A. K. & Jiang, L. Why are some STEM fields more

gender balanced than others? Psychological Bulletin 143, 1–35 (2017).

32. Sampat, B. N. & Lemley, M. A. Examining patent examination. Stanford Technology

Law Review 2010, 2 (2010).

33. Lerchenmueller, M. J. & Sorenson, O. The gender gap in early career transitions in the

life sciences. Research Policy 47, 1007–1017 (2018).

34. Frakes, M. D. & Wasserman, M. F. Is the Time Allocated to Review Patent Applications

Inducing Examiners to Grant Invalid Patents? Evidence from Microlevel Application

Data. The Review of Economics and Statistics 99, 550–563 (2016).

Working Paper — Please contact whalen@hku.hk for updated draft

35. Amanatullah, E. T. & Morris, M. W. Negotiating gender roles: Gender differences in

assertive negotiating are mediated by women’s fear of backlash and attenuated when

negotiating on behalf of others. Journal of Personality and Social Psychology 98, 256–

267 (2010).

36. Babcock, L. & Laschever, S. Women Don’t Ask. Women Don’t Ask (Princeton University

Press, 2009).

37. Spence, M. Signaling in Retrospect and the Informational Structure of Markets. American

Economic Review 92, 434–459 (2002).

38. Goldin, C. & Rouse, C. Orchestrating Impartiality: The Impact of ‘Blind’ Auditions on

Female Musicians. American Economic Review 90, 715–741 (2000).

39. Zuckerman, E. W. The Categorical Imperative: Securities Analysts and the Illegitimacy

Discount. American Journal of Sociology 104, 1398–1438 (1999).

40. Podolny, J. M. Status Signals: A Sociological Study of Market Competition. (Princeton

University Press, 2005).

41. Youn, H., Strumsky, D., Bettencourt, L. M. A. & Lobo, J. Invention as a combinatorial

process: evidence from US patents. Journal of The Royal Society Interface 12, (2015).

42. Alstott, J., Triulzi, G., Yan, B. & Luo, J. Mapping technology space by normalizing

patent networks. Scientometrics 110, 443–479 (2017).

 Supplementary Materials for

The Innovation Glass Ceiling and the Female

Penalty for Creative Innovation
Tara Sowrirajan, Ryan Whalen, Sourav Medya, Brian Uzzi

This PDF file includes:

Materials and Methods
Figures S1 to S10
Tables S1 to S11
Additional Results

1
S1 Gender Detection

We use a gender classification algorithm called genderize (from genderize.io) to predict the
gender of a person given their first name. Each name is classified a gender with an associated

probability that quantifies the certainty of the assigned gender (1). The three classes are male,
female, and unknown gender.

S1.1 Different Metrics for Gender Composition

We use several different metrics to assess the gender makeup of inventor and attorney teams.
These include majority female (nf emale /ntotal >= 0.5) and majority male (nmale /ntotal >=

0.5), as well as single-gender teams (ngender /ntotal = 1), where n refers to the count of a
particular gender. We also utilize the gender of the first listed inventor in several analyses.

S2 CPC Co-Occurrence Network Computation

Each patent is assigned to a variety of technical areas that it pertains to, according to the “co-

operative patent classification” (CPC) system. We utilize the tendency that inventions are cate-
gorized into multiple categories to create a CPC co-classification network in order to assess the
atypicality of a combination of categories.

This is done by first creating a CPC co-occurrence network called At . At is created cumu-
latively, covering the set of patents in the years at and below t. At (i, j) = n, where n is the
number of times categories i and j co-occur on a patent across all patents in the subset of time

(of year at or below t) we are analyzing.

To normalize for field-level variation in the tendencies of certain categories to be combined
in particular ways by field, we adopt an approach developed in (9) to calculate the standard

scores (z-scores) of CPC code pairs across the space of patents.

2
 oαβ − µαβ
zαβ = (S1)
σαβ
Equation (S1) represents the z-score of a pair of CPC codes α and β. The variable oαβ is the

observed number of co-occurrences of α and β within a patent in the data. Variables µαβ and
σαβ are the expected co-occurrences of the pair of CPC codes and its standard deviation, which
is derived from a null model of the data in which CPC code arrangement is randomized while

preserving CPC code usage and the number of patents in the subset of the data of years at or
below year t.
A positive z-score would be associated with two CPC codes co-occurring more than ex-

pected, and a negative z-score would correspond with two CPC codes that are rarely combined
relative to their expected co-occurrence (9).
In order to compute the expected co-occurrences of pairs of CPC codes, the baseline null

model randomized the arrangement of CPC codes as in (9). Take CPC codes α and β, with nα
and nβ representing their respected occurrences in the set of patents P .
The hypergeometric probability distribution for the number of co-occurrences is given by

Eq. (S2), as derived in (9).

nα |P |−nα



x nβ −x
p(oαβ = x) = |P |

(S2)
nβ

Thus, the expected number of patents combining CPC codes α and β is given in Eq. (S3)
with the variance of µαβ given in Eq. (S4).

nα nβ
µαβ = (S3)
|P |

  
2 nα |P | − nβ
σαβ = µαβ 1 − (S4)
|P | |P | − 1

3
S2.1 Temporal Variance

We compute in a temporally evolving manner At and zαβ where α, β ∈ Φ, where Φ represent

the set of all CPC codes in the set of patents we analyze.

This is done by cumulatively taking the set of all patents P , where P (y) <= t, where y
represents the year of the patent application. We look at applications in the range 2002 to 2018
to construct these z-score metrics, and the data spans years 2001 to 2018. We calculate these

measures based on the subclass level of the CPC code hierarchy.

For instance, Figure S1 depicts the cumulative atypicality adjacency matrix across all CPC
codes in the data through the end of 2017.

Figure S1: 2017 Cumulative Atypicality Network

4
S3 Patent-level Atypicality Construction

From Eq. (S1) we can compute the z-score measure for each pair of CPC codes across all CPC
codes through time. This measure represents how the empirical combinations of areas differ

from their expected combinations, and gives a measure to represent whether a combination of
CPC codes is atypical or conventional. A negative z-score corresponds to an atypical combina-
tion, while a positive z-score corresponds to a conventional combination. A z-score of 0 would

represent a neutral combination, neither atypical nor conventional.

For a given patent application p, say it combines three technological areas, or CPC codes
a, b, and c, with a being the focal, most relevant CPC code. We take all pairs of links with the

focal code a, which in this example would be links a, b and a, c, and take the minimum of their
associated pairwise z-scores as representative of the most atypical link: min(za,b , za,c ). In more
general terms, this can be written as follows:

Latypical = min zxf ,xi (S5)

i∈(1...n)

In Eq. (S5), xf represents the first, focal CPC code of a patent application, and xi where
i ∈ (1 . . . n) represents the n other CPC codes in an application other than the focal CPC code.
Latypical thus represents the minimum pairwise distance between the focal CPC code and all

other area codes in a particular patent application. A positive Latypical corresponds to a high
minimum z-score, representing a conventional combination of CPC codes, while a negative
Latypical corresponds to a low minimum z-score, representing an atypical combination of CPC

codes and thus an atypical patent application.

We can then negate Latypical as in Eq. (S6).

patypical = (−1) · Latypical (S6)

Thus, patypical represents the level of atypicality of patent p, where a positive value for

patypical represents an atypical patent (the more positive, the more atypical), and a negative

5
value for patypical represents a conventional patent (and larger negative values correspond to
more deeply conventional patents). Figure S2 depicts the distribution of patypical across the

space of US patents.

Figure S2: Distribution of atypicality in US patent data of applications with over 2 CPC codes.

S4 CPC Controls and Inter-Field Comparisons

Field-level variation in terms of gender differences in participation rates, thresholds to what is

considered patentable or innovative, and differences in success rates could contribute to gender

differences in success in innovation more broadly. Some of these field-level variations are
presented, for inventor gender related differences in Figure S3 and examiner gender related
differences in Figure S4 in terms of their incidence in the space of patents. These depict field-

level gender differences in different parts of the patenting process from inventors to examiners,

6
specifically, at the broader section level of CPC code for presentation ease, and variation was
found to exist throughout several levels of the CPC hierarchy, for instance, persisting at the class

level. In order to more adequately capture the effects of gender as opposed to effects stemming
from field-level differences, we control for field as measured by focal CPC code at the class
level.

Figure S3: CPC-level variation in participation rate by majority gender composition of inventor
teams.

7
 Figure S4: CPC-level variation in participation rate by gender of patent examiner.

S5 Inventor and Examiner Experience

In order to better isolate the effects of gender, we construct and include several salient features

in our regression models. The experience of the inventor teams and examiners involved could
presumably affect the likelihood of a patent application being accepted (5, 6). For instance,
when binning the data by inventor gender and experience, we can see a gap in success rate

8
between male and female inventors at the same level of experience, as depicted in Figure S5.

Figure S5: Acceptance rate by inventor gender and average inventor experience of the inventor
team.

We measure the average experience of the inventor team to be the mean of the experience of
all the inventor members. Experience is measured by number of patent applications a member
has been on at the application date considered. Examiner experience is measured by the number

of applications an examiner has evaluated to date of the patent application in question.

We validate our measure for inventor experience on a dataset consisting of granted USPTO
patent applications from 1976 to present where inventor names are disambiguated. We construct

a second measure, for granted applications, which consists of the average number of granted
patents previous to the grant date of the focal application for the inventors on the team. We can
only construct this second measure for inventors on granted patent applications due to the nature

9
of the data being diambiguated for only granted patents. As a robustness check, we find that our
measure of inventor experience as measured through average number of previous applications

correlates highly with this second measure from disambiguated data where we measure the
average number of previous granted applications, and this correlation is 0.841, lending validity
to our measure used over the set of all applications.

S6 Argumentativeness Analysis

We utilize the record of patent examination transactions between the applicant and the examiner

by extracting the codes for each transaction from the USPTO Patent Examination Data System
(PEDS). These transactions represent the types of interactions between inventors, attorneys,
and examiners. We want to compare the transaction history between applicants and examiners

to determine if there are differences in the way men/women navigate the patent application
process. To do so, we track the event codes for each application, focusing on those event
codes that represent a substantive exchange between the examiner and the applicant about the

patentability of the invention in question. We find no statistically significant differences in how

different genders navigate patent prosecution.

S7 Patent Value Estimates

Patent value is difficult to estimate precisely, and highly skewed. We estimate the monetary

value of a patent by averaging values from previous estimates. Previous literature estimates of
value averaged across different technical fields is $101,850 (1992 USD) per patent (4). Previous
estimates of value averaged across US entity types is $107,557 (1992 USD) (3). We use the

average across these as a rough estimate of per patent value: $104,703.5 (1992 USD).

10
S8 Impact Measured through Maintenance Fees and Cita-
tion Count

Research shows that patent citations are useful measures of an invention’s value (11). They cor-
respond to the patenting firm’s market value (7), expert assessment of the value of the underlying
invention (2, 8), and objective measures of the effectiveness of the patented technology (10).

There are several ways to assess and estimate the future impact of a patent that has been
granted. One of these measures is through citation count, which we predict the log C + 1 in
regression models, where C represents the number of citations a granted patent has garnered in

8 years. We find that atypicality is positively predictive of log C + 1, while the proportion of
female inventors is negatively predictive. This is depicted in Table S1.

Table S1: Citation Count as Impact measure Regression Results

(USA)
Log(Citation Count) + 1
Proportion Female Inventors -0.1669∗∗∗
(0.0304)
Atypicality 0.0004367∗∗∗
(0.0000508)

Observations 263372
Pseudo R2 0.185
Standard errors in parentheses
∗
p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001
Entity, Year, and Team Size Fixed Effects. This is for granted patents filed 2011 and previously.

Another way to quantify impact is by tracking patent owners’ continued investment in their
patents. Granted patents need to pay fees every four years to maintain their patent, deemed a

’maintenance fee’. We track these event codes for maintenance fees in the data. Presumably,

11
valuable inventions will be paying maintenance fees to maintain their patents. We find that in
predicting an indicator variable of whether any maintenance fee payment was made at any time

(for granted applications), the proportion of female inventors is a negative coefficient while
atypicality is positively predictive, as depicted in Table S2.

Table S2: Maintenance Fee as impact measure: Regression Results

(USA)
Any Maintenance Fee Paid (Indicator)
Proportion Female Inventors -0.13625∗∗∗
(0.03684)
Atypicality 0.0006621∗∗∗
(0.0000581)
Entity 0.091192∗∗∗
(0.0193419)
Observations 113211
Pseudo R2 0.0165
Standard errors in parentheses
∗
p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001
Year, Team Size, and CPC Code Fixed Effects.
This is for patents with grant year at or below 2015 to allow time for accrual of maintenance fees.

Both our measures of impact suggest atypical inventions correspond and predict future value

through both citations garnered and patent maintenance.

S9 Regression Variable Operationalization

We utilize several salient features in logistic regression models to predict patent acceptance.
We additionally use regression models to predict the log C, where C represents the number of
citations a granted patent has garnered in 8 years post grant.

12
 nf emale
The proportion of female inventors on an inventor team is calculated as ntotal
, where n
represents the count of a particular attribute, such as the number of females.

As a robustness check, we also measure inventor team gender composition in two additional
ways. We use both a measure that assigns the team gender based on the gender of the first listed
inventor, as well as an additional indicator variable indicating if the team is majority female
nf emale
(where ntotal
≥ 0.5) or not. These variables are used in regressions as depicted in Tables S4
and S5.
Team size |(Inv)| is calculated according to Equation (S7), where ntotal is the number of

members on an inventor team. Because team size is highly skewed, we transform it into a
categorical variable to capture single-member team, small team, medium team, and large team
dynamics.

(
ntotal if ntotal < 5
|(Inv)| = (S7)
5 if ntotal >= 5
Patent applications can come from big or small entities (corporations), which is coded as a

categorical variable where e(p) = 1 if a patent is from a big entity and 0 otherwise. Inventor
and examiner experience is calculated as described in Section S5. Examiner gender is coded as
a binary variable, where exam(p) = 1 if it is a female examiner and 0 otherwise. We include

fixed effects for year and CPC code at the class level. For the multiple country regression model,
we have an additional categorical variable for country, as depicted in Table S8.

S10 Continuous Operationalization of Gender

One metric for inventor team gender composition we use is the proportion of female inven-

tors on an inventor team, as described previously. The regression results using this metric are
depicted in Table S3. We find that not only is there a negative coefficient for increasing fe-
male inventors on a team, but that the interaction between proportion of female inventors and

13
Table S3: Regressions results for proportion of female authors on inventor team variable oper-
ationalization. There are fixed effects for year, team size, and CPC Code.

(US Patent Applications)

Patent Acceptance
Proportion Female Inventors -0.272∗∗∗
(0.0174)
Atypicality 0.0001534∗∗∗
(0.0000397)
Proportion Female Inventors × Atypicality -0.0005019∗∗∗
(0.000012)
Big Entity 0.590∗∗∗
(0.00806)

Female Examiner -0.102∗∗∗

(0.00793)

Average Inventor Experience -0.00111∗∗∗

(0.0000527)

Average Examiner Experience 0.00118∗∗∗

(0.0000129)

Observations 392622
Pseudo R2 0.0915
The base case for big entity is small entity.
Standard errors in parentheses
∗
p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001

14
atypicality is negative.

S11 Alternate Operationalizations of Gender

As robustness checks, we use alternate measures of inventor team gender composition in ad-
dition to the proportion of female inventors on the team. These include indicator variables of

whether a team is majority female as well as a variable indicating the gender of the first listed
inventor on the application. The regression results for both of these additional gender metrics
are depicted in Tables S4 and S5 and, like the main analysis, are done on the subset of US patent

applications with over 2 CPC codes.

For the first inventor gender regressions results in Table S4, we see patent applications with
women listed as the first inventor face a penalty in terms of predicting patent acceptance as

compared to the male base rate. There is no such penalty for inventors with an undetermined
gender. We observe atypicality with a positive effect as well as the effect of first author gender

x atypicality is negative. Note here that the base case is for male first author inventors for both
the inventor gender main effect as well as the interaction with atypicality.
Similarly, for the majority female inventor team regressions results in Table S5, we see

women majority teams face a penalty in terms of predicting patent acceptance. We observe
atypicality with a positive effect as well as the effect of female majority team x atypicality is
negative. These robustness checks yield results very similar to the gender metric of proportion

female inventors discussed in the main text.

S12 Evidence for Stereotyping in Patent Acceptance

To examine the plausibility of this mechanism of gender stereotyping, we can look to other
possible stereotypes that examiners might rely on when assessing unfamiliar inventions. In
addition to the inventor’s demographic traits, examiners may consider the type of firm applying

15
Table S4: Regressions results for gender of first author of inventor team variable operational-
ization. There are fixed effects for year, team size, and CPC Code.

(US Patent Applications)

Patent Acceptance
First Author Inventor Gender Female -0.144∗∗∗
(0.0124)
First Author Inventor Gender Unknown 0.0304
(0.0350)
Atypicality 0.000120∗∗∗
(0.0000357)
First Author Inventor Gender Female × Atypicality -0.000365∗∗∗
(0.0000785)
First Author Inventor Gender Unknown × Atypicality 0.00000856
(0.0000785)
Big Entity 0.592∗∗∗
(0.00806)

Female Examiner -0.103∗∗∗

(0.00793)

Average Inventor Experience -0.00112∗∗∗

(0.0000527)

Average Examiner Experience 0.00118∗∗∗

(0.0000129)

Observations 392622
Pseudo R2 0.091
The base case for first author gender and its interaction with atypicality is for male first author inventors.
The base case for big entity is small entity.
Standard errors in parentheses
∗
p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001

16
Table S5: Regressions results for majority female inventor team variable operationalization.
There are fixed effects for year, team size, and CPC Code.

(US Patent Applications)

Patent Acceptance
Female Majority Inventor Team -0.179∗∗∗
(0.0128)
Atypicality 0.000122∗∗∗
(0.0000354)
Female Majority Inventor Team × Atypicality -0.000399∗∗∗
(0.0000813)
Big Entity 0.590∗∗∗
(0.00806)

Female Examiner -0.102∗∗∗

(0.00793)

Average Inventor Experience -0.00112∗∗∗

(0.0000528)

Average Examiner Experience 0.00118∗∗∗

(0.0000129)

Observations 392622
Pseudo R2 0.091
The base case for big entity is small entity.
Standard errors in parentheses
∗
p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001

17
for the patent. Legally, firm size should not affect patentability standards, however evidence
suggests that large firms have more success at the patent office. Thus, if examiners fall back

on stereotypes to help them assess the patentability of unfamiliar inventions, we would expect
them to penalize large firms less than others when an application claims a boundary spanning
invention. Indeed, this is what we see when we examine the interaction between firm size

and the degree to which a claimed invention combines rarely combined technical fields, as
depicted in Figure S6. As small entities make increasingly diverse and unusual combinations
of technical fields, their chance of success at the patent office decreases. Large firms have

the opposite dynamic, enjoying increased chance of success as their inventions become more
boundary spanning. While this does not provide conclusive proof, it supports a stereotyping
explanation where examiners rely on extraneous prior correlates of patentability to help them

assess new unfamiliar inventions.

S13 UK and Canada Findings

We find evidence of gender gaps in patent acceptance in the UK and Canada as well. In the
”UK” and ”Canada” columns of Table S8, we see that the proportion of female inventors has
a negative effect on patent acceptance, in line with a negative penalty for female inventors

we witness in the USPTO data. In a regression model predicting patent acceptance with the
US, UK, and Canada patent application data, we find that not only is the proportion of female

inventors negatively predictive and atypicality as a main effect is positive, but the interaction
term between atypicality and female inventor proportion is significant and negative. This can
be seen in the first ”All Countries” column of Table S8. This regression has fixed effects for

country and controls for application year.

For the UK and Canada, since we have access to fewer applications that would not comprise
a full picture of CPC or field-level combinations, we rely on the cumulative CPC code z-score

18
Figure S6: Probability of patent application success by invention atypicality and firm size.
Includes statistical control for team size, year, examiner gender, examiner experience, applicant
entity size, and CPC class.

network from the US patent application data to compute the atypicality of UK and Canadian
applications. This is done using Equations (S5) and (S6) given the CPC codes in the applications
from the UK and Canada while using the cumulative z-score network in the corresponding time

subset of the US application data.

Figures S8 and S7 depict the distributions of atypicality for the UK and Canada respectively.
We find further evidence of differences in acceptance rates between men and women when

evaluating atypical applications in all three countries data. We look at applications with at
least 2 CPC codes in line with considering applications that are more difficult for examiners to
evaluate due to their complexity. The distributions within atypical applications (patypical > 0)

by gender for success are depicted for the three countries studied in Figures S9, S10, and S11.

19
Figure S7: Distribution of atypicality in Canada patent data of applications with over 2 CPC
codes.

We can see that males majority teams have a different distribution of success than female
majority teams for atypical applications across three countries patent systems, and females have

less success. This is further exemplified in Table S6 for both majority gender teams and single
gender teams.

S14 Granted Patent Analysis

Each patent must have at least one independent claim delineating the legal rights to the intellec-
tual property claimed by the applicant. Ultimately, these claims are the enforceable elements of

a granted patent.
We analyze changes to the number of independent claims from the time of application to

20
 Table S6: Success Rates for Atypical Applications Interna-
tionally

USA UK Canada
Male Majority Team 0.658*** 0.661*** 0.852***
Female Majority Team 0.578*** 0.608*** 0.849***
*** p < 0.005
Differences by gender for atypical applications were all significant
when doing a Chi-square test of independence.

Table S7: International Patent Application Logistic Regression Results. There are fixed effects
for year, team size, country for the ”All Countries” results, and fixed effects for year and team
size for the ”UK” and ”Canada” results.

(All Countries)
Patent Acceptance
Proportion Female Inventors -0.309∗∗∗
(0.0155)
Atypicality 0.000302∗∗∗
(0.0000368)
Proportion Female Inventors × Atypicality -0.000521∗∗∗
(0.000111)
Observations 455435
Pseudo R2 0.058
Standard errors in parentheses
∗
p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001

21
Figure S8: Distribution of atypicality in UK patent data of applications with over 2 CPC codes.
Table S8: Female Penalty and Atypicality Reward on International Patent Application Data.
There are fixed effects for year, team size, and CPC Code for the both countries’ results.

(UK) (Canada)
Patent Acceptance
Proportion Female Inventors -0.265∗∗∗ -0.447∗∗∗
(0.0407) (0.0482)
Atypicality 0.0102∗∗∗ 0.00594∗∗∗
(0.000281) (0.000173)
Observations 178424 170686
Pseudo R2 0.227 0.189
Standard errors in parentheses
∗
p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001

22
Figure S9: Distribution of acceptance outcome in UK patent data of atypical applications with
over 2 CPC codes where patypical > 0.

granted versions of the patent text. Let nApp = |Independent Claims at Application Stage|.
Let nGrant = |Independent Claims at Granted Stage|. The features we focus on are denoted
in Equations (S8) and (S9) representing the difference in number of independent claims at

application stage minus grant stage as well as the relative difference in number of claims as
compared to the application stage respectively. This reflects the changes to the number of
independent claims requested by the applicant and those granted by the patent office.

∆Absolute = nApp − nGrant (S8)

nApp − nGrant ∆Absolute

∆Relative = = (S9)
nApp nApp

23
Figure S10: Distribution of acceptance outcome in Canada patent data of atypical applications
with over 2 CPC codes where patypical > 0.

Regressions within granted patents with over 2 CPC codes to predict the quantities regard-

ing the absolute and relative claims at time of filing versus at time are grant, namely metrics
∆Absolute and ∆Relative respectively, are denoted in Tables S9 and S10 respectively. We
see that an increased proportion of female inventors increases ∆Absolute and ∆Relative. This

suggests females face a reduction in the number of claims they are granted from those they re-
quest on their patent applications, suggesting increased claim ‘shrinkage’ over male inventors.
We also find that in regressions on granted patents with over 2 CPC codes, when predicting

the number of granted independent claims ngrant , a higher proportion of female inventors neg-
atively predicts the number of granted independent claims, while atypicality positively predicts
this measure ngrant . This is depicted in Table S11.

24
Table S9: ∆Absolute Regression Results. There are fixed effects for year, team size, and CPC
code.

(USA Granted Patents)

∆Absolute
Proportion Female Inventors 0.3260164∗∗∗
(0.0354938 )
Atypicality 0.0013255∗∗∗
(0.0000707)

Observations 241755
R2 0.0367
Standard errors in parentheses
∗
p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001

Table S10: ∆Relative Regression Results. There are fixed effects for year, team size, and CPC
code.

(USA Granted Patents)

∆Relative
Proportion Female Inventors 0.0649181∗∗∗
(0.0106551)
Atypicality 0.000231∗∗∗
(0.0000212)

Observations 241755
R2 0.0344
Standard errors in parentheses
∗
p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001

25
Figure S11: Distribution of acceptance outcome in US patent data of atypical applications with
over 2 CPC codes where patypical > 0.

Regression results within granted patent applications with over 2 CPC codes to predict

the quantity Tgrant − Tf iling , which represents the time (in days) of grant date from filing
date, are denoted in Table S12. We see that as the proportion of female inventors increases,
so does Tgrant − Tf iling . This suggests that female inventors have to wait longer for patent

protection than their male counterparts. We also observe that increased atypicality increases
Tgrant − Tf iling . This indicates that atypical applications are more likely to take longer to be
granted.

26
Table S11: Regression Results for ngrant . There are fixed effects for year, team size, and CPC
code.

(USA Granted Patents)

ngrant
Proportion Female Inventors -0.1126367∗∗∗
(0.0229627)
Atypicality 0.0003208 ∗∗∗
(0.0000457)

Observations 241755
R2 0.0889
Standard errors in parentheses
∗
p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001

Table S12: Regression Results for time (in days) of grant date from filing date. There are fixed
effects for year, team size, and CPC code.

(USA Granted Patents)

Tgrant − Tf iling
Proportion Female Inventors 32.39971∗∗∗
(4.737681)
Atypicality 0.3355133∗∗∗
(0.0094379)

Observations 241752
R2 0.3891
Standard errors in parentheses
∗
p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001

27
References

1. genderize.io. https://genderize.io/. Accessed: 2022-06-30.

2. Michael B Albert, Daniel Avery, Francis Narin, and Paul McAllister. Direct validation of
citation counts as indicators of industrially important patents. Research policy, 20(3):251–

259, 1991.

3. James Bessen. The value of us patents by owner and patent characteristics. Research Policy,
37(5):932–945, 2008.

4. Timo Fischer and Jan Leidinger. Testing patent value indicators on directly observed patent
value—an empirical analysis of ocean tomo patent auctions. Research policy, 43(3):519–

529, 2014.

5. Michael D Frakes and Melissa F Wasserman. Patent office cohorts. Duke LJ, 65:1601,
2015.

6. Michael D Frakes and Melissa F Wasserman. Is the time allocated to review patent appli-

cations inducing examiners to grant invalid patents? evidence from microlevel application
data. Review of Economics and Statistics, 99(3):550–563, 2017.

7. Bronwyn H Hall, Adam Jaffe, and Manuel Trajtenberg. Market value and patent citations.
RAND Journal of economics, pages 16–38, 2005.

8. Dietmar Harhoff, Francis Narin, Frederic M Scherer, and Katrin Vopel. Citation frequency

and the value of patented inventions. Review of Economics and statistics, 81(3):511–515,
1999.

9. Daniel Kim, Daniel Burkhardt Cerigo, Hawoong Jeong, and Hyejin Youn. Technological
novelty profile and invention’s future impact. EPJ Data Science, 5(1):1–15, 2016.

28
10. Petra Moser, Joerg Ohmstedt, and Paul W Rhode. Patent citations and the size of the
inventive step-evidence from hybrid corn. Technical report, National Bureau of Economic

Research, 2015.

11. Manuel Trajtenberg. A penny for your quotes: patent citations and the value of innovations.
The Rand journal of economics, pages 172–187, 1990.

29