Inventor Gender and Patent Undercitation:

Evidence from Causal Text Estimation †

Yael V. Hochberg Ali Kakhbod Peiyao Li Kunal Sachdeva

Rice University & NBER UC Berkeley UC Berkeley Rice University

April 2023

Abstract

Implementing a state-of-the-art machine learning technique for causal identification

from text data, we document that patents lead-authored by female inventors are
under-cited relative to their quality. For the equivalent patent with a lead female
inventor, a patent with a male lead inventor would have received 28% more cita-
tions. Male lead inventors in particular tend to undercite patents with female lead
inventors, while patent examiners of both genders appear to be more even-handed.
Market-based measures of patent value load on the citation counts that would be pre-
dicted for a female lead inventor patent had it been lead-authored by a male, rather
than on actual citation counts. The under-recognition of female-authored patents
likely has implications for the allocation of talent in the economy.

JEL classification: J16, J24, J71, O30, C13

Keywords: Innovation, Gender, Patent, Machine Learning, Big Data, Causal Inference
† We thank Tania Babina, Po-Hsuan Hsu, Amir Kermani, David Levine, Zack Liu, Dmitry Livdan, Ulrike
Malmendier, Asaf Manela, Danqing Mei, S. Lakshmi Naaraayanan, Vira Semenova, Amit Seru, Carolyn
Stein, Dong Yang, Tengfei Zhang, participants at the NBER Productivity Innovation and Entrepreneurship
Meeting, Midwest Finance Association (MFA), York University (Schulich), and the Mid-Atlantic Research
Conference in Finance (MARC) for their helpful comments and suggestions. Yael V. Hochberg is with the
Jesse H. Jones Graduate School of Business at Rice University and NBER, email: hochberg@rice.edu. Ali
Kakhbod is with the Haas School of Business at the University of California, Berkeley, email: akakhbod@
berkeley.edu. Peiyao Li is with the Haas School of Business at the University of California, Berkeley, email:
ojhfklsjhl@berkeley.edu. Kunal Sachdeva is with the Jesse H. Jones Graduate School of Business at Rice
University, email: kunal.sachdeva@rice.edu.

Electronic copy available at: https://ssrn.com/abstract=4269703

 Female inventors appear to face significant obstacles when seeking patents. Women
face significant disparities in the patent approval process (Jensen, Kovács, and Sorenson
(2018)), may face a higher bar for patent grants (Gavrilova and Juranek (2021)), and are
underrepresented among inventors in patent applications more generally (Bell, Chetty,
Jaravel, Petkova, and Van Reenen (2019); Reshef, Aneja, and Subramani (2021)). As a
result, assuming a patent is applied for and granted, we might well expect to observe
a selection effect, whereby the average patent with a female inventor may be of higher
quality than the average male inventor patent, and thus, on average, achieve a higher
forward citation count. Perhaps surprisingly, however, simple analysis of patent cita-
tions suggests that female-authored patents in fact receive fewer forward citations rela-
tive to male-authored patents (Jensen, Kovács, and Sorenson (2018)), suggesting either
that patents granted to female inventors are of lower quality, or, more concernedly, that
the quality of their patents are not fully recognized in the form of forward citations, a
commonly used measure.
Assessing whether a patent is undercited relative to its actual quality is not a trivial
undertaking. Typically, citations serve as the de facto measure of a patent’s quality, even
though the measure is noisy. To determine whether female inventors face systematic
obstacles to citations of their work, versus simply producing lower quality patents, the
econometrician must disentangle actual quality from the citation outcome. In an ideal
setting, the econometrician would either randomize underlying quality across genders
or gender across patents. Natural experiments that mimic this ideal, or suitable instru-
mental variables, however, have been elusive.
In this paper, we utilize novel machine learning techniques that allow for measure-
ment of the causal contribution of gender to citation of patents of similar quality. Our
methodology builds on a burgeoning set of research in the computer science literature
that studies causal identification using textual data (see e.g. Khetan et al. (2022); Shao
et al. (2021)).

Electronic copy available at: https://ssrn.com/abstract=4269703

 The intuition behind these models is straightforward. Our goal is to identify the
expected change in outcome if we apply treatment while holding fixed any mediating
variables affected by the treatment that also might affect the outcome. Here, assuming
that the text of a patent contains sufficient information to adjust for confounding (com-
mon) causes between the treatment and outcome, we can use textual analysis to identify
the causal effect of treatment, as shown in Figure 1.
For our main analysis, we adopt the causal bidirectional encoder representation from
transformers or C-BERT (Veitch, Sridhar, and Blei (2020)). C-BERT estimates causal ef-
fects from observational text data, adjusting for confounding features of the text, such s
the subject or writing quality. It assumes that the text content suffices for causal iden-
tification but is prohibitively complex for standard analysis. C-BERT utilizes causally
sufficient embeddings, (relatively) low-dimensional document representations that pre-
serve sufficient information for causal identification, thus enabling efficient estimation
of causal effects. The causal sufficiency reduces dimensionality yet preserves aspects of
text that predict both the treatment and the outcome while disposing of linguistically
irrelevant information (which is also causally irrelevant). The identification assumption
is that the text contains all information necessary to measure the desired effects (quality
of the patent and forward citations, conditional on gender).
Our main sample covers all utility patents granted by the U.S. Patent Office (USPTO)
from 1976 through 2021. The extended follow-up period allows us to measure the im-
pact of a patent without concerns of forward citations being right-censored. We textual
matching the first names of each inventor to data from the Social Security Administra-
tion and the World Intellectual Patent Organization. We then label each patent by the
inferred gender of the inventors following Desai (2019). For our main analyses, we label
patents as female lead inventor if the first inventor listed on the patent is female, and
male lead inventor if the first inventor listed on the patent is male. Our results are ro-
bust to other labelings of female versus male-authored patents, including restricting to

Electronic copy available at: https://ssrn.com/abstract=4269703

single-authored patents or gender-homogenous teams.
The next step is to train the model. First, we use a pre-trained BERT model provided
by Google (from Python packages Transformers and Tensorflow) to transform the text
of each patent’s abstract1 into a numerical representation. Next, given that the sam-
ple is dominated by male-authored patents, we reduce the sample to an equal set of
male- and female-authored patents by randomly selecting a subsample of male patents
without replacement. Finally, we train two neural networks—one per gender of the
lead patent author—using the BERT numerical representations as inputs and citations
as outputs, representing a mapping from embedding vectors to citation counts. The
first mapping is trained using the subset of data where the patent is female-authored,
while the second mapping is trained using the data where the patent is male-authored.
Unlike the standard OLS approach, the neural network approach captures the complex
and often nonlinear relationships between inputs and outputs, particularly when dealing
with high-dimensional inputs. Having obtained parameters for each gender’s citation-
prediction model, we then take the patent data for each gender and run it through the
prediction model trained on the opposite gender’s inputs and outputs. This produces a
set of counterfactual citation counts for each patent, holding all else equal, and changing
only the gender of the lead author.2
We begin by documenting that even with no adjustments for patent quality or char-
acteristics, there is a statistically significant difference in the number of forward citations
for patents with a female first author versus those with a male first author in our matched
data, consistent with Jensen, Kovács, and Sorenson (2018). This pattern persists when
we control for factors such as the identity of the patent examiner, the identity of the cor-
respondent for the patent submission (usually the law firm involved), the art unit of the
1 we note that our main results are robust to using full patent text and an appropriate textual embed-
ding approach (LONGFORMER) for longer texts instead of BERT.
2 The methodology also incoporates a gender propensity model to ensure a patent’s text is not clearly

identified as male- or female-authored, in order to ensure a quality counterfactual can be computed.

Dropping this restriction only strengthens our results.

Electronic copy available at: https://ssrn.com/abstract=4269703

patent, and the patent issue year. The distributional plots of forward citations confirm
the statistical tests and suggest female-authored patents receive 1 to 3 fewer citations
than male-authored patents. More generally, patents authored by women appear to be
less likely to receive any citations than male-authored patents.
We then use the causal text analysis methodology to mediate for differences in the
quality of patents. We perform two sets of analysis. First, we explore the extensive
margin, including in our sample all patents, including those that receive zero citations
(the modal patent). Second, we explore the intensive margin, restricting our sample to
those patents that receive at least one forward citation. In both cases, we control for a
variety of patent characteristics and fixed effects.
At the extensive margin, applying the C-BERT methodology and comparing actual
citations to those that would be predicted if the patent was authored by an inventor of the
opposite gender, we find that patents with a female lead inventor received approximately
28% fewer citations than an equivalent quality patent in the same art unit, evaluated by
the same examiner, would receive had the lead inventor been male. This difference
equates to approximately 3.8 fewer citations per patent. The impact of this undercitation
is most pronounced for the most impactful patents, with female lead-authored patents
being less likely to reach the top decile of citations. At the intensive margin, we find
similar effects: patents with a female lead inventor received approximately 20% fewer
citations than an equivalent patent would receive had the lead inventor been male, a
difference again of approximately 3.8 fewer citations per patent.
Overall, the results suggest that patents with a female first author would receive
more citations if their first author had been male. The results are robust to a variety of
alternative specifications and do not seem to be attributable to sample selection or model
overfitting. The results are also robust to a variety of different approaches to defining
a “female-authored” patent. For example, we obtain similar results when comparing
patents with a single female author to those with a single male author, or patents with

Electronic copy available at: https://ssrn.com/abstract=4269703

author teams composed only of female authors versus patents authored by teams com-
posed only of male authors. Finally, while our main analysis uses patent abstracts, this
result is also robust to employing the full text of the patent instead of the abstract.3
Our results hold across NBER’s major categories and subcategories of patent technol-
ogy, with some heterogeneity in subcategories.We observe similar patterns when using
the Cooperative Patent Classification (CPC). We observe similar differences in citations in
both emerging and established technology fields. This undercitation of female-authored
patents grows over time, becoming particularly more pronounced since the early 1990s.
We then turn to explore the source of these patterns. Under-citation of female-
authored patents could be driven by the citations inventors themselves include in their
patent applications, or by citations that are added by patent examiners. Controlling for
art unit, issue year, and examiner, patents with female and male first authors both sig-
nificantly undercite patents with a female first authors. Patents with male first authors,
however, do so by approximately double the margin. In contrast, we see little evidence
of either male and female examiners underciting female-led patents. Overall, the results
suggest that the undercitation of female patents is largely due to patent with male first
authors underciting past female-authored patents in their patent applications.
Finally, we explore whether the expected value accorded to the patent by the mar-
ket reaction to its issuance correlates more with actual citations or with the citations
predicted by our methodology. We use the economic value of a patent as measured
by public markets from Kogan, Papanikolaou, Seru, and Stoffman (2017), which is gen-
erally considered to be forward-looking and determined at the time of issuance. We
regress these measures of economic value on our adjusted forward citations, actual for-
ward citations, and a variety of controls. When horseraced against each other, measures
of expected economic value load significantly on our adjusted forward citation measure,
3 Due to the differences in text length between patent full text and patent abstract text, using BERT for
the embedding is computationally infeasible. Instead, we utilize an embedding called LONGFORMER,
which is more suitable for longer texts.

Electronic copy available at: https://ssrn.com/abstract=4269703

but not on actual forward citations. The results suggest that expected economic value,
as measured by market reactions, may be a less biased proxy for patent quality than
standard measures of realized forward citations.
Our results come with several important considerations. First, causal text analysis
relies on the assumption that the patent text used captures all the factors that should
influence the number of forward citations. While it is not possible to test this assumption
directly, it is reasonable to assume that the text of the patent is closely related to its
quality or importance. Second, the predictive ability of any given model in computing
counterfactual outcomes is difficult to assess in these settings. Instead, we plot the loss
function of our model training and find that the fitting error starts to plateau after around
15 epochs of training. Therefore, in order to prevent overfitting, we stop training shortly
afterat 20 epochs. Note that the remaining error is an order of magnitude smaller than
the initial error, so the model has made considerable progress in fitting the data. Further,
we examine the quality of our model fit by synthetically creating counterfactual data to
evaluate the accuracy of our model. As shown in Figure IA4, we find a high accuracy of
the model without bias. Finally, although our evidence suggests women receive fewer
citations for patents of equal quality, we do not argue that this represents discrimination,
as we cannot observe the intent of examiners or inventors. Further research will be
necessary to establish why patents with female lead inventors are undercited.
Our findings have potentially important implications. First, the literature has high-
lighted that innovation is motivated by the expected profits derived from the property
rights granted to patentees, Moser (2005, 2013).4 If women are not equally recognized
for equivalent patents, this may discourage them from entering the innovation economy,
potentially reducing contributions from half of the population. Moreover, this may ex-
acerbate the already substantial wedge between men and women in science, technology,
engineering and mathematics (STEM) fields (Beede, Julian, Langdon, McKittrick, Khan,
4 Inrelated research, the marginal investor values patents Aghion et al. (2013); Hall et al. (2005);
Hirschey and Richardson (2004); Hirshleifer et al. (2013).

Electronic copy available at: https://ssrn.com/abstract=4269703

and Doms, 2011), leading to further inefficient allocation of labor. A second implication
concerns the validity of research that relies on forward citations of patents as a measure
of patent quality. To the extent that female-authored patents are systematically under-
cited relative to their actual quality, the use of forward citations as a measure or control
for quality may be contraindicated. Given the large literature in economics, finance,
and innovation that relies on forward citations to measure patent quality, these findings
suggest that a re-examination of prior findings may be warranted.
Our findings contribute to an emerging literature studying obstacles that inventors
face in the U.S. patent system. Research has studied impediments that women and
minorities face in obtaining patents, with emphasis on the unequal application of laws
(Cook (2014)), unequal opportunities (Cook (2020); Cook and Kongcharoen (2010)), and
discrimination by patent examiners (Desai (2019)). These obstacles all result in depressed
levels of applications and lower success rates for females in obtaining patents, Jensen,
Kovács, and Sorenson (2018). In contrast to this literature, which focuses on causally
identifying differences in patent applications and approvals, our findings focus on a
relatively unexplored question: whether women also face obstacles in citation of their
patents.
Relatedly, our findings contribute to the broad literature studying obstacles that
women face in various research fields. Recent work by Sherman and Tookes (2022)
documents that women face discrimination in financial economics publishing and job
placement. Sarsons, Gërxhani, Reuben, and Schram (2021) and Sarsons (2017) show
women receive less credit attribution for co-authored work in economics, while Card,
DellaVigna, Funk, and Iriberri (2020) and Hengel and Moon (2020) show that, controlling
for quality, female academics in economics receive fewer citations for their work. Koffi
(2021) finds that undercitation in economics is more likely to be of women-authored pa-
pers and that male authors are more likely to cite male-authored papers. Chawla (2016)
and Koffi and Marx (2021) study broader academic fields. Our work suggests parallels

Electronic copy available at: https://ssrn.com/abstract=4269703

in patent citations as well.
In addition, our paper makes an important methodological contribution. A large
literature across many fields has demonstrated that big data has the potential to revolu-
tionize research in general and finance and economics research in particular (Goldstein,
Spatt, and Ye (2021)). In economics, a small but rapidly growing branch of the big
data literature uses natural language processing to quantify text, allowing it to be used
in empirical applied microeconomics research (Gentzkow, Kelly, and Taddy (2019a)). A
partial list of papers in this vein includes the work of Athey and Imbens (2019); Bellstam,
Bhagat, and Cookson (2021); Cong, Liang, and Zhang (2019); Erel, Stern, Tan, and Weis-
bach (2021); Gentzkow, Kelly, and Taddy (2019a); Gentzkow, Shapiro, and Taddy (2019b);
Hanley and Hoberg (2019); Hansen, McMahon, and Prat (2018); Li, Mai, Shen, and Yan
(2021); Loughran and McDonald (2016); Rouen, Sachdeva, and Yoon (2022); Routledge,
Sacchetto, and Smith (2017). Recent advances in computer science have produced new
methods that allow the use of text embedding to mediate and identify causal effects. Our
paper introduces these methods to finance and economics, proposing a new technique,
C-BERT, that uses text as a mediator, allowing us to causally identify the effect of gender
on citations of patents. To the best of our knowledge, we are among the first researchers
to apply deep learning in economics and finance for causal inference using language.

1 Data

Our main analysis uses data on patent content, citations, and attributes. Our main
sample covers all utility patents granted by the U.S. Patent Office (USPTO) from 1976
through 2021. This allows for at least a 20-year follow-up history, extending through the
patent’s expiration.

Electronic copy available at: https://ssrn.com/abstract=4269703

1.1 Patent Content

Our sample of patents comes from the USPTO’s Patent Examination Research Database
(PatEx) dataset. In our main analyses, We study the quality of the patents through the
lens of patent abstracts, as they provide a clear and concise text-only summary of the
core contribution of the patent. Importantly, this is the key text input into the C-BERT
model. Using the abstract of patents presents several key advantages over using the full
body. First, use of abstracts alleviates concerns about differences in the quality of the
figures contained within the patents that could substitute for the quality of the writing.
Second, abstracts are a good proxy for the contents of a patent as well as what inventors
and examiners review. Third, from a practical standpoint, using the full text of the
patents is computationally prohibitive. Even with access to a high-powered computing
cluster, using abstracts in our setting takes several days to complete.
In robustness tests, however, we reproduce our analysis using the full text of the
patents, and substituting the LONGFORMER embedding for BERT, as BERT does not
scale for long texts. All our results remain qualitatively similar (we intend to include
these results in more detail in a future draft).

1.2 Patent Citations

Importantly, there is a key difference between patent citation counts and actual patent
quality. While historically forward citations have been used as a proxy for patent quality,
the key point of our analysis is to determine whether this measure is systematically bi-
ased downwards for female-authored patents. We therefore distinguish between patent
citations (the easily observed outcome for a patent) and quality, which is mediated for
by using the text of the patent. Patent forward citation counts are obtained through use
of data from the USPTO.
Of course, patent forward citations are highly skewed in their distribution, with only

Electronic copy available at: https://ssrn.com/abstract=4269703

a few patents receiving a disproportionately high number of citations. As an alternative
to simple counts of forward citations, we also consider whether a patent receives citations
in the top decile of all patents.

1.3 Gender of Inventors and Examiners

Our main treatment variable is the gender of the lead inventor (first author).5 Inventors,
however, do not disclose their gender when applying for a patent. Because of this, we
must infer gender from third-party sources, (Graham, Marco, and Miller, 2018). To dis-
ambiguate the gender of the inventor, we implement a name disambiguation algorithm
similar to that of Desai (2019). We use the first name of the lead inventor to identify the
gender of the inventor (Tzioumis, 2018).
Starting with the PatentView data, we obtain the first names of each inventor of each
patent. For patents with multiple inventors, we rely on the name of the first inventor due
to that person’s prominence. Next we classify the gender of patent inventors using state-
level data on the frequency of names obtained from the Social Security Administration
(SSA) (Comenetz, 2016). We assign a gender when the percentage of names in the state
belonging to that gender is above 70%.6 If the first name does not match the SSA dataset,
our second step uses a similar process but utilizing a cross-country dataset from the
World Intellectual Property Organization (WIPO) (Martinez, Raffo, Saito, et al., 2016).
We drop patents when there is no distinct gender determination for the lead inventor.
One challenge is that our sample shows that women are underrepresented as inven-
tors on patents (Hunt et al., 2013). As a result, we need to balance our sample across
patents with lead inventors from each gender. To do this, we use all patents with a fe-
male lead inventor and extract a random subsample of patents with male lead inventors
of the same size. We then estimate a propensity model using a one layer logit-linear
5 In further robustness, we consider single-author patents and the gender of the entire team.
6 We take a conservative approach and apply a high confidence interval to reduce Type I errors when
identifying males and females.

10

Electronic copy available at: https://ssrn.com/abstract=4269703

neural network, where the objective function is the binary-cross-entropy between the
predicted treatment indicator and the true treatment indicator. The output of this neural
network is the predicted probability that this patent is written by a female lead author.
Following the C-BERT methodology in Veitch, Sridhar, and Blei (2020), we then drop (i)
all patents in the male subsample whose estimated propensity of being female-authored
is very low (less than 3%) and (ii) the all patents in the female subsample whose esti-
mated propensity of being female-written is very high (greater than 97%). This allows
for any given patent in the remaining sample to have some level of ambiguity, based on
textual analysis, as to whether it was authored by a male of female.7

1.4 Examiner and Inventor Added Citations

Typically, patent applications include a list of related patents and supporting material.
Citations to patents may be added in two ways. First, inventors cite precedent patents in
their applications. Second, examiners will identify additional citations that are missing
from the patent and request that these be included (Farre-Mensa, Liu, and Nickerson
(2022)). Starting in 2001, and more clearly since 2003, the USPTO discloses whether the
citation originated from the examiner or the inventor. For the purposes of the analysis
studying the source of a citation, we create additional citation counts that only record
citations that were explicitly added by examiners and inventors.

1.5 Other Patent Attributes

When an inventor files a patent application with the USPTO, the application is assigned
a USPC class and subclass based on its field of technology. The application is then
assigned to an “art unit” comprised of several examiners who specialize in that particular
technology class and subclass. We use the art unit to which the patent is assigned as our
7 Allour results remain qualitatively similar in nature and stronger in magnitude if we do not exclude
patent whose author gender can be clearly identified from the text content alone.

11

Electronic copy available at: https://ssrn.com/abstract=4269703

proxy for technology type grouping. Our baseline sample contains 898 art units and
11,953 patent examiners.
As an alternative and intuitive patent technology grouping is provided by the NBER
patent category, which is also reported in the USPTO PatentView database.8 The NBER
classification includes six major categories (computer and communications, drugs and
medicine, electrical, mechanical, chemical, and other) and 37 sub-categories. We use
the six categories and 37 sub-categories to examine heterogeneity by patent technology
type, allowing us to present individual subcategory estimates in a digestible manner. In
further robustness tests we also consider the Cooperative Patent Classification (CPC) of
the patents, (see Table IA3).
Patents are typically filed with the assistance of a patent attorney, who may file
many them on behalf of different inventors. The USPTO refers to these law firms (or
legal department of the patent assignee firm) as “customers,” and each such entity is
assigned a customer number. Approximately 60% of observations have a legitimate
customer number. If the observation lacks a valid customer number, we assign it a
common value (“unassigned”). These identifiers are useful because they allow us to
account for possible commonalities in writing style across patent attorneys that may
influence the text of the final submission. Our baseline sample contains 9,516 unique
customers. We label these customer identifiers as “law firm.”
Descriptive statistics for our sample are presented in Table 1.

2 Empirical Strategy

Our analysis presents both methodological and computational challenges. First, we must
represent complex and often subtle differences in the text of the patents in a parsimo-
nious and computationally useful form. Second, we need to relate that text to forward
8 Note, the NBER patent categories are truncated at the end of our sample.

12

Electronic copy available at: https://ssrn.com/abstract=4269703

citations. Finally, we must compute the counterfactual of citations based on the gender
of the inventor.
Below, we outline our empirical strategy to address these challenges. First, we dis-
cuss how we create a high-dimensional representation of text that encapsulates the in-
formation necessary to distinguish patent quality. Second, using this representation, we
provide an overview of the C-BERT methodology and how we train our model. Finally,
we discuss the key identification assumptions implicit in our approach and their validity.

2.1 High-Dimensional Representation of Patent Text

There are a variety of possible approaches to transform text into numerical form. Here
we use a Bidirectional Encoder Representations from Transformers (BERT) approach to
transform the text of each patent into a high-dimensional numerical vector. Developed
by Google (Devlin, Chang, Lee, and Toutanova, 2018), BERT has become the leading ap-
proach in many commercial applications, including Google’s search platform. BERT con-
structs embedding vectors that are numerical representations of the text, which preserve
both the meaning of individual words and the underlying context of each word.9 The
BERT encoder module (Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser, and
Polosukhin (2017)) produces a high-dimensional representation with 768-dimensional
embeddings that each represent the text of a patent’s abstract. We describe the encoder
architecture in detail in Appendix A.

2.2 Causal BERT (C-BERT)

Having created high-dimensional representations of patent text, the second challenge is

to establish the relationship between this data and patent citations. To do so, we use a
novel leading machine learning technique called Causal Bidirectional Encoder Represen-
tations from Transformers (C-BERT) that allows us to causally estimate the contribution
9 See Jha, Liu, and Manela (2022) for an excellent discussion of BERT.

13

Electronic copy available at: https://ssrn.com/abstract=4269703

of language on a binary treatment variable. C-BERT comes from recent advances in com-
puter science, including Khetan, Ramnani, Anand, Sengupta, and Fano (2022); Shao, Li,
Gu, Qian, and Zhou (2021), which has developed methods to use text embedding as a
mediator Veitch et al. (2020). Causal text analysis allows us to use the text of patents as
a mediator to causally identify the role of gender on patent citations. To the best of our
knowledge, ours is among the first papers to apply deep learning to causal inference
with language in the field of economics and finance.
C-BERT is a neural-network-based architecture that estimates counterfactuals of a
binary treatment under the assumption that all of the information (covariate) needed
for causal identification is contained within a given text. As shown in Figure IA2, the
input data for training contains three types of information: the texts of patents, gender
indicators of the inventor(s), and the observed number of citations on the patents. There
are four neural networks that need to be trained: a BERT model for generating text
embeddings, a logit-linear model the maps embeddings to treatment propensities, and 2
two-layer perceptrons that map from embeddings to male and female predicted number
of citations, respectively. The final loss function is a weighted average of the losses of
these four neural networks.
The C-BERT methodology in our context has two key steps. First, it uses a pre-
trained BERT embedding to transform the text of each patent into a high-dimensional
numerical vector. The embedding vectors are numerical representations of the text that
preserve both the meaning of individual words and the underlying context of each word.
Second, C-BERT computes the number of citations an inventor would have received if
that person were assigned the opposite gender. This is accomplished by training two
neural networks, where each model represents a mapping from embedding vectors to
our outcome variable, forward citations, with the first mapping trained using the subset
of patents with a female lead inventor and the second mapping trained using the subset
of patents with a male lead inventor. The two estimated mappings, combined with

14

Electronic copy available at: https://ssrn.com/abstract=4269703

the high predictive performance of neural networks, allow us to approximate the true
mappings.
Armed with our two mappings, we can then estimate the counterfactual of gender
on citation. That is, we can ask the following: how many citations would a patent
whose lead inventor is female have received if the lead inventor had instead been male,
and vice versa? We estimate this by passing the subset of patents with a male lead
inventor through the mapping trained on the embedding vector of female lead inven-
tor patents, and passing the subset of patents with female lead inventors through the
mapping trained on the embedding vector of male-lead-inventor patents. From this, our
mapping estimates the number of citations for the counterfactual number of citations by
gender.
The procedure is depicted in Figure 2. First, we run the trained C-BERT model
where the input data contains the texts of the patents and gender indicators of the
author(s). The texts are first passed through the trained BERT model to generate a
vector embedding for each patent. Then each embedding-gender pair is passed through
a decision step: if the author(s) are male, the embedding is passed to the female citations
network, and, if the author(s) are female, it is passed to the male citation network. The
counterfactual number of citations is then computed by these two networks. In parallel,
regardless of the gender indicators, each embedding is passed through the propensity
network to estimate the treatment propensity of this patent, which is used to identify
patents that are clearly predicted (97%+ probability) to have been written by one gender
or the other irrespective of quality, which are then dropped (as discussed in section 1).
Finally, the output of the model is a set of counterfactual citation-treatment propensity
pairs that each correspond to one patent.
The framework can be expressed more formally in mathematical terms. We denote
the text in the abstract of the ith patent as Wi . We fine-tune the BERT model f to map Wi
to Zi where Zi is the embedding of the abstract. Then we use a logit-linear network g to

15

Electronic copy available at: https://ssrn.com/abstract=4269703

map Zi to a real number, which represents the treatment propensity of this patent. Here
the treatment propensities are the probability that this patent has a female lead inventor.

g( Zi ) = P( Ti = 1| Zi ) = ( g ◦ f ) (Wi )

In addition, we have two citation networks Q1 and Q0 . Q1 maps an embedding

vector to the predicted number of citations if the patent has a female lead inventor, and
Q0 maps an embedding vector to the predicted number of citations if the patent has a
male lead inventor. Mathematically, we define a piecewise mapping Q that represents
the two networks:
Q( Ti , Zi ) = E (Yi ( Ti )| Zi ) = E (Yi ( Ti )| f (Wi ))

where Yi (0) and Yi (1) denote the potential outcomes of the ith patent. In our case,
these potential outcomes are the number of forward citations. Given these mappings
represented by neural networks, we can then estimate the average treatment effect (ATE)
and the average treatment effect on the treated (ATT) using the following equations for
a set of N patents.

N N
ATE = ∑ [E(Yi (1)|Zi ) − E(Yi (0)|Zi )] = ∑ [Qi (1, Zi ) − Qi (0, Zi )]
i =1 i =1

N N
1 1
ATT =
∑iN=1 Ti
∑ Ti [E(Yi (1)|Zi ) − E(Yi (0)|Zi )] = ∑ N Ti
∑ Ti [Qi (1, Zi ) − Qi (0, Zi )]
i =1 i =1 i =1

The resulting output of our C-BERT model is the actual outcome and a counterfactual
outcome. In our application, this is the number of citations and the estimated number
of citations the opposite gender would have received.

16

Electronic copy available at: https://ssrn.com/abstract=4269703

2.3 Assessing C-BERT’s Identification Assumptions

There are three assumptions that the econometrician must consider when applying C-
BERT.

2.3.1 Text Renders the Effect Identifiable

The first necessary condition is that the text of the documents must render the effect
identifiable. Said differently, the effect that the econometrician is measuring must be
measurable directly from the text. Similar to an exclusion restriction within other iden-
tification strategies, this cannot be formally tested. Instead, this condition must be in-
spected and potentially falsified by considering other channels. In the context of this
paper, the quality of the patent should be measurable by the content (text) of the patent
itself. Patent examiners read the text of the proposals to evaluate the novelty of patents
prior to granting a patent. As a result, this necessary condition is likely satisfied in our
context.

2.3.2 Embedding Method Extracts Semantically Meaningful Information

The second necessary condition is that the embedding method extracts semantically
meaningful text information relevant to the prediction of both treatment, T, and out-
come, Y. In our setting, this means that embedding, a lower-dimensional representation
of the text, is sufficient to capture the gender and quality of citations.
To assess the quality of our embedding representations, we consider synthetic tests to
measure the accuracy of our model. To do this, we first compute the synthetic outcomes
of all of the patents across the full dataset. In doing this, we used a random linear trans-
formation that takes a uniformly random 768 vector with values from 0 to 1. Then we
take the dot product of this random vector with each patent’s 768 dimensional embed-
ding. Finally, the resulting values are the synthetic outcome for females, and, for males,
we add a known scalar to the function. In this approach, we know the true treatment

17

Electronic copy available at: https://ssrn.com/abstract=4269703

effect and can evaluate the model. Applying the C-BERT model uncovers the known
true treatment effect with an accuracy of over 90%. This high level of accuracy suggests
that the embedding method clearly extracts semantically meaningful information.

2.3.3 Conditional Outcome and Propensity Score Models are Consistent

Our third and final necessary condition is that the conditional outcome and propensity
score models be consistent. That is, the treatment and control groups should have com-
mon support. To address this, as discussed above, we follow the procedure of Veitch,
Sridhar, and Blei (2020) and drop the patents with either below 3% treatment propensity
or above 97% treatment propensity. In our study, the treatment is the female gender in-
dicator of the lead inventor. Therefore a treatment propensity of at most 3% implies that
this patent, as defined by the embedding of the text, almost certainly has a male lead
inventor. On the other hand, a treatment propensity of at least 97% implies this patent
almost certainly has a female lead inventor. This procedure preserves over 80% of our
data after dropping the propensity score outliers. Importantly, our results remain robust,
suggesting that the conditional outcome and propensity score models are consistent.

3 Do Citations to Patents Differ By Gender of Inventor?

Do forward citation counts for patents differ across the gender of the lead inventor? We
begin by examining the differences in forward citation counts by gender using simple
regression analysis. We then apply the C-BERT model to calculate counterfactuals, and
assess the causal effect of gender on citation counts.

3.1 Comparing Between Genders Without Model Adjustments

We begin by plotting the unconditional differences in citations by gender. The histogram

for citations for male and female lead inventors is plotted in Panel A of Figure 3, and

18

Electronic copy available at: https://ssrn.com/abstract=4269703

visually demonstrates that female receive fewer citations than males.10 On average, male
lead inventors receive significantly more citations than female lead inventors (18 citations
for males, 15 for females, F-stat = 518, Table 1). Testing the difference in distributions, we
find a Kolmogorov-Smirnov statistic of D = 0.066426, with a p-value of = 2.2 × 10−16 ,
further suggesting that male and female forward citation counts come from different
distributions.
We more formally consider the contribution of gender on patent citations by estimat-
ing the following OLS model:

Yi = β 1 I ( FemaleInventori ) + δGrantYear + δArtUnit + δCustomer× Examiner + ε i , (1)

where patent and year are represented by i and t, respectively. Yi is our outcome of inter-
est. Our specification includes fixed effects for customer-examiner pair (δcustomer×examiner ),
art unit (δArtUnit ), and year of grant (δGrantYear ). All standard errors in this paper, unless
otherwise noted, are double-clustered by patent issue year and customer. β 1 is our co-
efficient of interest, where a positive value would indicate that women receive more
citations than males, and vice versa.
The estimates are presented in Table 2. Panel A presents the results for the full
sample of patents (extensive margin), while Panel B presents the estimates for those
patents which receive at least one forward citation (intensive margin).11 Both panels
suggest that female lead investors receive between 0.8 to 4 fewer citations than males,
depending on specifications and controls included.Given that the expected selection ef-
fect from prior literature might predict that we would see higher quality patents–and
thus, higher citation counts–for female authored patents, the patterns from this sim-
ple analysis raise questions. Either the female authored patents being approved are of
10 To address skewness and show more clearly, we the natural logarithm of citations and present a
histogram in Panel A of Figure IA1.
11 Most patents do not receive any forward citations; in general, female lead-authored patents appear

to be less likely to receive any citations than those with a male lead author.

19

Electronic copy available at: https://ssrn.com/abstract=4269703

lower quality, on average, than those of males, or, despite the higher bar for approval of
female-authored patents, the quality of these patents is not being appropriately reflected
in citation counts, with women experiencing undercitation of their inventions relative to
males.

3.2 Adjusting for the Quality of Patents Using C-BERT

To explore this, we turn to our C-BERT model. As a reminder, C-BERT first trains
two mappings, one using only patents from male inventors and a second for female
inventors. Armed with our two mappings, we pass the male patents through the female
mapping, and vice versa. From this, we can estimate the counterfactual number of
citations a patent would have received had its lead author been of the opposite gender,
[
ForwardCitation i . We plot the histogram of predicted citations from C-BERT by gender

in Panel B of Figure 3.12

We compare the actual forward citations to the model-implied citations at the patent
level. Specifically, we calculate the difference in actual versus predicted citations as:

[
Deltai = ForwardCitationi − ForwardCitation i, (2)

where ForwardCitationi is the actual number of citations to a given patent authored

[
by a given gender and ForwardCitation i is the number of citations implied if the lead

inventor had been of the opposite gender. A positive delta implies that a patent has
received more citations than the quality-adjusted number suggested by the opposite-
gender model.
We plot the difference between actual and model implied citations in Figure 4 for
patents with at least one citation. We observe that Deltai appears to be negative on
average for female lead inventors (plotted in red), with a mean and median of -2.69 and
12 To address skewness and to show the distribution more clearly, we the natural logarithm of citations

and present a histogram in Panel B of Figure IA1.

20

Electronic copy available at: https://ssrn.com/abstract=4269703

-0.25, respectively. In contrast, for male lead inventor patents (plotted in blue), we see a
positive delta versus what they would have been predicted to receive with a female lead
author instead, with a mean and median of 1.10 and 0.13, respectively. For both genders,
these difference measures show evidence of skewness. Notably, it is clear that the delta
for any given patent is not strictly positive for males or negative for females.
To study the relative delta, we re-estimate the OLS in Equation 1, replacing the actual
number of forward citations for a patent with Deltai . The estimates are presented in
Table 3. Panel A presents estimates for the full sample of patents, including those with
zero citations (extensive margin), while Panel B presents the estimates for the subsample
of patents which receive at least one citation (intensive margin). Across both panels,
the estimates suggest that patents with female lead inventors are undercited relative to
what would be expected had the patent remained otherwise the same, except that the
lead inventor was instead male. Interpreting our point estimate relative to the sample
mean of the number of forward citations, we find that patents with female lead inventors
receive 3.8 fewer citations than would be predicted if they had had a male first author
instead, controlling for other characteristics and mediating for quality using C-BERT. The
magnitudes are relatively unchanged when including patent-year fixed effects, art-unit
fixed effects, and various customer and examiner fixed effects. The relative stability in
estimates suggests that our analysis does not suffer from a correlated omitted variable,
Oster (2019).
Next, we consider the difference in the propensity for female lead patents to be in
the top-decile of patents by forward citations. To study this we once again estimate
Equation 1 but replace the dependent variable with an indicator that takes the value of
one if the patent is in the top decile of citations based on the mapping for the opposite
gender but is not based on observed data. The estimates are presented In Panel A of
Table 4. The estimates suggest that without any model adjustments, female lead inventor
patents are underrepresented in the top decile of patents by citations, relative to patents

21

Electronic copy available at: https://ssrn.com/abstract=4269703

with a male first author.
We then apply the C-BERT methodology adjustment. In Panel B of Table 4, we re-
estimate the models, where the dependent variable takes the value of one if the patent
based on actual citations was not in the top decile of citations, but under the C-BERT pre-
dicted forward citation, would be (“flipped to top decile”). The estimates suggest that
the probability of a patent flipping into the top decile under the counterfactual is ap-
proximately 1.7 percentage points higher if the lead inventor is female. These estimates
are highly statistically significant and economically meaningful.
These three sets of results represent causal evidence suggesting that female lead in-
ventors are undercited, on average, relative to an equivalent patent granted to their male
counterpart. These differences in forward citations cannot be explained by differences
in art units, time trends, or differences in customers or examiners.

4 Cross Sectional Heterogeneity

Next, we explore whether these patterns of undercitation are uniform across a variety of
dimensions of heterogeneity in patent characteristics.

4.1 Patent Category

First, a reasonable question is whether the underciting of female lead inventor patents
uncovered in our main models holds across all technology categories or whether there is
variation across fields. We next explore this heterogeneity. Specifically, we estimate the
following model.

Yi = β 1 I ( FemaleInventori ) + β 2 I ( FemaleInventori ) × (Category) (3)

+δPatentCategory + δGrantYear + δArtUnit + δCustomer×Examiner + ε i ,

22

Electronic copy available at: https://ssrn.com/abstract=4269703

where the subscript and notation match the prior estimating equations. As in the main
analysis, standard errors are double clustered by year and customer. First, we interact
our female indicators with the six NBER categories to study differences by broad field
categories. Then we explore the 37 subcategories in a similar manner.
The estimates in Table 5 highlight important heterogeneity across patent categories.
Column (1) of Table 5 presents the estimates for the model using the major categories,
where the outcome variable is the actual number of forward citations received by the
patent. Column (2) re-estimates the model using the difference between the actual ci-
tations and the number implied by the C-BERT model for the opposite gender, Delta.
To interpret the overall effects for our variable of interest for each category, we need to
add the coefficients of the indicator for female lead inventor with the interaction term
for each major category. In general, we observe some level of undercitation for all patent
categories, with particularly large disparities for the Drugs and Medical category. Put
differently, if a female lead-inventor patent instead had a male lead inventor, it would
have received significantly more citations, regardless of technology category, echoing
our baseline results.
We can further break down the technology category using the 37 NBER subcate-
gories. Specifically, we estimate the following.

Yi = β 1 I ( FemaleInventori ) + β 2 I ( FemaleInventori ) × (Subcategory) (4)

+δPatentSubcategory + δGrantYear + δCustomer×Examiner + ε i ,

To ease interpretation, Figure 5 presents the sum of the female lead indicator and
the interaction coefficients (Female Lead Inventor × Subcategory) graphically. The raw
estimates are presented in the Internet Appendix in Table IA4.Column 1 of Table IA4
shows the estimates employing the raw citation counts as the dependent variable, while
column (2) uses Delta as a dependent variable. The finer category classification exhibits

23

Electronic copy available at: https://ssrn.com/abstract=4269703

somewhat more heterogeneity than the major classes. Importantly, in all specifications,
we include patent subcategory fixed effects to account for the average level of citations
in a given subcategory. As in the main analysis, standard errors are double clustered by
year and customer. As can be seen clearly in Figure 5, for the vast majority of the tech-
nology subcategories, the estimates suggest that patents with lead female inventors are
cited significantly less than a male lead inventor instead, and these citation undercounts
are often substantial in magnitude.

4.2 Established Versus Emerging Fields

An interesting question is whether the patterns we see across technology fields relate in
some way to whether women are patenting in an established field versus in an emerging
field of technology. It is possible that newer fields may not present as many barriers to
entry or pre-existing biases for female inventors and researchers, given the lack of an
established history of research and researchers, and that we may expect undercitation
patterns to be larger or concentrated in more established fields. On the other hand, the
underlying forces that lead to undercitation for patents with female first authors may
be unrelated to the nature of the field, and relate to gender norms or perceptions more
generally, in which case we would not expect to see a difference.
To explore these issues further, we denote a category as an “emerging field” if the
art unit first appeared within five years of the patent being granted. We then re-run our
models, adding an indicator for an emerging field as well as an interaction between that
indicator and the indicator for a female lead inventor. Our coefficient of interest is the
interaction between the indicator for female inventors and emerging fields.
The estimates are presented in Table 6. While our main result is still apparent, with
patents with female first authors exhibiting an estimated 3.3 to 3.7 fewer citations than
would be predicted if the first author had been male, depending on specification), across
all specifications, we cannot reject the null that there is no additional citation difference

24

Electronic copy available at: https://ssrn.com/abstract=4269703

in emerging versus established fields.

4.3 Time Since Patent Grant

A natural question is whether the undercitation we observe above is present from the
outset or whether it primarily materializes or diminishes later in the life of the patent.
On the one hand, undercitation may be present from the outset but diminish over time
as inventors and examiners become more familiar with the patent and its quality. Alter-
natively, the bias may increase and become more pronounced over time, potentially in-
dicating a self-reinforcing effect that could be harder to overcome. Examining the timing
of the bias in citations can provide valuable insights into the nature of the undercitation
of female inventors and inform potential interventions to address this issue.
To investigate the timing of undercitation, we create separate samples of forward
citations based on the number of years that have passed since a given patent was granted.
Specifically, we divide the post grant period into four sub periods: [0-1) years post
grant, [1-5) years, [5-10) years, and [10-20] years. For each of these subperiods, for
each patent, we collect the forward citations the patent receives during this sub period
post grant. For each subperiods, we re-run our C-BERT methodology to estimate the
delta in citations after mediating for patent quality. This allowed us to examine when
undercitation occurs, relative to the time of patent grant.
The estimates are presented in Table 7. Column (1) presents estimates from forward
citations to patents received in the first year after patent grant, column (2) presents
estimates for forwward citations received in years 2 to 5 after patent grant, column (3)
for citations received in years 6 to 10, and column (4) years 11-20. In each column, the
dependent variable is the Delta estimated from C-BERT using only forward citations
received during that subperiod (by necessity, the number of observations is smaller in
later subperiods as fewer of the patents in our sample will yet have histories of that
length). As can be seen from the estimates in the table, the undercitation for patents

25

Electronic copy available at: https://ssrn.com/abstract=4269703

with a female first author (relative to what would be expected if the first author had
been male) increases over time since patent grant, consistent with undercitation being
self-reinforcing over time. The coefficient on lead female inventor is economically and
statistically insignificant in the first period ([0-1) years), but becomes more pronounced
and strongly statistically significant in the subsequent periods, with estimates of -0.7,
-1.3, and -2.7 for the [1-5) years, [5-10) years, and [10-20] years periods, respectively.
Our estimates in Table 7 are also consistent with our baseline specification. When
we sum the point estimates across the four periods, we obtain results that are similar
in magnitude to those in the baseline specification. Furthermore, when we consider the
magnitude of the estimates relative to the sample means, we see that the bias in citations
increases monotonically over time, with values of 3%, -17%, -20%, and -24.25% for the [0-
1) years, [1-5) years, [5-10) years, and [10-20] years periods, respectively. These findings
suggest that biases in citations may be reinforced by prior biases, leading to a situation
in which overcited patents continue to be overcited and the bias becomes larger over
time.

4.4 Evolution of Biases over the Sample Period

The estimates we present in the prior analyses suggest that across fields, patents with
female first authors are consistently undercited relative to what would be expected for
the same patent had its first author been male. A natural question is whether these
patterns vary over time within the sample, as gender norms and female participation
in the workforce more generally and in science and engineering more specifically have
been changing over time.
Of course, older patents tend to naturally receive more citations. Without adjust-
ments to our initial methodology, our findings may incorrectly suggest a decrease in
bias over time, when in reality, it is simply a reflection that newer patents receive fewer
citations on average. To accurately study the evolution of bias, we restrict our measure

26

Electronic copy available at: https://ssrn.com/abstract=4269703

of forward citations to the first ten years post patent grant. This method avoids the right
censoring problem of forward citations and provides a clearer interpretation, at the cost
of excluding forward citations made after ten years out, where we know more of the
undercitation activity occurs. Because we need to be able to measure ten years of for-
ward citations, we by necessity must exclude patents granted in the last decade of our
sample. We thus create a sample of patents from 1976 to 2011, and measure the number
of forward citations they receive within 10 years. Then, we apply our C-BERT technique
to calculate the bias for each patent.
We estimate models of the following nature:

Yi = β 1 I ( FemaleInventori )
2011
+ ∑ β j I ( FemaleInventori ) × I ( GrantYear = j) (5)
j=1977
+δGrantYear + δArtUnit + δCustomer×Examiner + ε i ,

where β 1 estimates the average undercitation of females across the entire sample, and
the set of coefficients β j estimate the marginal bias in each patent grant year, with 1976
as the year of comparison. Standard errors are clustered in this specification by the grant
year. The omitted group is 1976.
Figure 6 plots the interaction coefficient of for each year from equation Equation 7.
The time-invariant estimate for coefficient on the female lead inventor variable is -0.34;
the interaction coefficients presented in the figure are additive to that number. From
the figure, we observe clearly that the average undercitation of patents with female lead
authors has become more pronounced over time. In comparison to patents from 1976,
those from the late 1970s and early 1980s seem to have been only modestly additionally
undercited. However, starting in the 1990s, the additional undercitation of female-led
patents rises to around 2 citations per patent. Thus, despite a decrease in disparities and
representation of women in the workplace and in science and engineering professions,

27

Electronic copy available at: https://ssrn.com/abstract=4269703

undercitation of patents with female first authors seems to be growing over time.

5 Who Undercites Female Inventors?

So far, we have presented causal evidence that patents with female lead inventors receive
fewer citations than the equivalent patents with male lead inventors. Next we explore
the source of the under-citation: whether it is driven by inventors or examiners, and the
role of their gender.
To set the stage for this analysis, we first discuss how a citation is added to a patent.
When applying for a patent, applicants cite supporting patents whose inventions the cur-
rent patent is building on top of. If, however, the patent examiner deems that there are
additional relevant citations that have not been included by the inventor, the examiner
will also add these to the patent application. As a result, the documented undercita-
tion of patents with female lead inventors may stem from the original inventor-added
citations, additional examiner-added citations, or a combination of both.
To explore the source of the under-citation, we first need to know which citations
in a patent are attributable to the inventor versus the examiner. Starting in 2001, and
more comprehensively starting in 2003, asterisks were added to the USPTO citation
data to identify examiner-added patents in the data. Using this detail, we construct a
new subsample starting from 2003 aggregating forward-citations into four categories: (i)
forward citations added (in a future patent) by male lead inventors, (ii) forward citations
added by female lead inventors, (iii) forward citations added by male-lead examiners,
and (iv) forward citations added by female-lead examiners. Using these groups, we can
then decompose the sources of under-citation of female lead-inventor patents.
We begin our analysis by studying examiner-added citations. For a given patent, we
take all forward citations that occur due to being added to a future patent application
by an examiner. We then break these into forward citations added by female examiners

28

Electronic copy available at: https://ssrn.com/abstract=4269703

and forward citations added by male examiners. Following similar logic to our main
tests, we then apply the C-BERT model, estimating a neural net for male-lead inventor
patents and a neural net for female-lead inventor patents to predict forward citation
counts by examiners of each gender based on the gender of the lead inventor on the
patent of interest. We then run female lead inventor patents through the male neural net
model, and vice-versa, to calculate the C-BERT adjustment to mediate for the quality of
the patent.
Table 8 presents the results of estimation of regression models using the C-BERT
adjustment as the dependent variable. Panel A presents estimates for female examiner
added forward citations, and Panel B presents the estimates for male examiner added ci-
tations. The estimates in Panel A suggest that female examiners contribute minimally to
the under-citation of female lead inventor patents. The coefficient estimates range from
-0.014 citations to -0.029 citations, with only two specification statistically significant at
conventional levels. Panel B similarly shows no real evidence of undercitation of female
lead inventor patents by male examiners. The coefficient estimates range from -0.006
citations to 0.009, with no specification being statistically significant. Taken together, the
estimates suggest that the undercitation bias we observe for female lead inventor patents
is not driven primarily by examiner patents.
Having established this fact, we then turn to forward citations added by future in-
ventors to their patent applications. We conduct a similar analysis to that which we
conduct above with examiners. The estimates from the C-BERT adjustment regressions
are presented in Table 9. Panel A presents the results for female inventor added forward
citations, and Panel B presents the estimates for male inventor added forward citations.
The estimates suggest a clear pattern. First, as can be seen in Panel A, female lead in-
ventors contribute only modestly to the undercitation of female lead inventor patents.
The estimates from the regression models in columns (1) through (4) suggest that fe-
male inventors undercite female lead inventor patents by approximately 0.5 citations. In

29

Electronic copy available at: https://ssrn.com/abstract=4269703

contrast, the contribution of male inventors to the undercitation of female lead inventor
patents is more considerable. The estimates in Panel B suggest undercitation of female
lead inventor patents by male inventors by more than 1.3 citations. This is large, both
economically and statistically, especially in comparison with our main effect.
Taken together, these results suggest that the undercitation of female lead patents
is primarily driven by male lead inventors. Note, however, that we cannot conclude
that this is necessarily discrimination on the part of male examiners and inventors. For
example, these results may stem from men having more and stronger connections to or
familiarity with other male inventors and, as a result, being more familiar with patents
filed by other male lead inventors. These familiarity networks could be boosted by
the presence of a female lead inventor on the current patent. Future research may be
necessary to fully distinguish the reason for the underciting.

6 Robustness Tests

6.1 Definition of “Female Authored Patent”

One potential explanation for our findings is that the way in which we classify patents
by gender may have spurious influence the results. In our baseline method, we used the
name of the first inventor to assign author gender to patents with multiple inventors.
The first name on the patent is likely the most salient, as it is the first name observed
when reading the patent. Of course, it is possible that examiners and inventors may
consider all inventors and not just the first author when attributing gender.
To address the possibility that our definition of author gender spuriously produces
the patterns we observe, we show robustness of the estimates to a number of alterna-
tive approaches to attributing author gender to a patent. First, we limit our sample to
patents with only one author, and re-running our C-BERT model, comparing female
sole-authored patents to male sole-authored patents. This shuts down concerns that the

30

Electronic copy available at: https://ssrn.com/abstract=4269703

gender of additional authors other than the lead author may be driving the baseline
results.13
Using the single-author sample, we still find a statistically significant bias against
female inventors. The estimates shown in Panel A of Table IA5 are consistent with those
in Table 3. As an alternative, we construct a new sample that includes only patents with
all inventors of the same gender, both single authors and teams.14 When we re-run our
C-BERT model using this expanded sample, we find similar patterns of under-citation,
with similar coefficient magnitudes (4 versus 3.8). While do not present estimates from
our additional analyses using these two samples, they are qualitatively similar to those
obtained using a first author gender definition.

6.2 Full Patent Text and LONGFORMER embedding

The results presented up to this point utilize patent abstracts and the BERT embedding.
A natural concern is that the patent abstracts do not have enough content to fully pick up
patent quality for mediation purposes. For robustness, we repeat our analyses replacing
the patent abstract texts with the full patent texts, and utilizing the LONGFORMER
embedding instead of BERT. We use the LONGFORMER embedding instead because
BERT has difficulties handling longer text lengths.
Table IA6 reports the main results using full patent texts, for the earlier subperiod of
our sample, 1986-1994. We observe qualitatively similar, statistically significant, under-
citation results (roughly 2 citations). In future versions of the paper, we intend to extend
this analysis to the full sample period and for all of our estimations.
13 The neural network that measures the propensity for a given patent to have been written by a male
versus a female assures that what we are picking up is quality as indicated by text content as opposed to
writing style.
14 Note, this only grows our sample by roughly 2% because teams of all female inventors that can be

confidently identified in the sample are rare.

31

Electronic copy available at: https://ssrn.com/abstract=4269703

6.3 Overfitting of Model

A standard concern with these types of models is overfitting to the training data. In
our setting, we train two different models by completing multiple passes of our training
dataset through our algorithm, an epoch. While numerous passes of the data help im-
prove the predictive probability of the neural networks, we could have overfit our model
to the data. If so, this would result in relatively poor out-of-sample performance. In the
context of our paper, this would result in incorrect or biased out-of-sample predictions
of the number of citations.
We address this concern by studying the loss function, as presented in Figure IA5, to
ensure a reasonable number of training iterations. Plotting the mean square error (MSE)
per batch against the number of passes of the training dataset, we find two key pieces
of evidence that suggest we have not overfit the model. First, as we increase the number
of epochs, the MSE tends to decrease. Second, we find diminishing improvements to
the error rate as we approach 20 epochs. Taken together, these findings suggest that our
model is unlikely to be overfitted and, as a result, that the model is appropriate and that
reasonable counterfactual citations are predicted from our neural networks.

7 Economic Value of Patent and Citation Bias

We next consider the relationship between the economic importance of a patent, as evalu-
ated by public markets at the time of issue, and forward citations. As we have previously
discussed, undercitation of female-authored patents tends to persist over time and be-
come more pronounced as the years go on. In contrast, the economic value of a patent,
as assessed by public markets, is forward-looking and can be determined at the time of
issuance. An interesting question is whether these forward looking market estimates of
a patent’s economic value relate more closely to actual forward citations, or to the pre-
dicted number of forward citations we obtain out of C-BERT, which adjusts for author

32

Electronic copy available at: https://ssrn.com/abstract=4269703

gender.
To investigate the relationship between economic value and citations, we used the
patent-level measure of economic value proposed in Kogan et al. (2017). This measure is
computed for patent issues for publicly-traded U.S. firms, and utilizes the stock market’s
response to news about patent grants. Specifically, we use the log value of innovation,
deflated to 1982 (million) dollars using the CPI. We then examine whether this measure,
when regressed on actual forward citations and the predicted counter-factual forward ci-
tations fi the first author had been of the opposite gender, loads on one of these measures
in particular versus the other.
Specifically, we estimate:

[
Yi = β 1 I ( FemaleInventori ) + β 2 ForwardCitationi + β 3 ForwardCitation i (6)

+δGrantYear + δArtUnit + δCustomer×Examiner + ε i .

[
The estimates are presented in Table 10. Only the coefficient on ForwardCitation i

loads significantly, suggesting that the market does not appear to undervalue female-
authored patents relative to what it would had that same patent been authored by a
male lead inventor. The estimates provides further support to the notion that actual
measures of forward citations are biased by gender of the inventor.

8 Discussion and Conclusion

We provide causal evidence that patents with female lead inventors are undercited rel-
ative to what they would have received if their patent had a male lead inventor. Our
approach uses new tools in machine learning to disentangle quality from forward cita-
tions, allowing us to show that the most commonly used measure for patent quality in

33

Electronic copy available at: https://ssrn.com/abstract=4269703

fact under-recognizes the quality of female-led patents, relative to equivalent male-led
patents.
Our findings have important economic implications. First, prior literature has high-
lighted that innovative activity is motivated by expected profits derived from the prop-
erty rights granted to the patentee (Moser (2005, 2013)). If female inventors are under-
cited relative to male peers with equivalent patents, and their compensation for their
innovative labor is accordingly harmed, this may discourage women from entering the
innovation economy. Such effects may further exacerbate the gender gap in STEM fields
(Beede et al., 2011), leading to inefficient allocations of labor.
A second important implication of our findings concerns the validity of research that
relies on forward citations as a measure of patent quality. The existence of systematic
gender-related biases in citations may lead to incorrect or misleading conclusions for
research that relies on forward citations as a measure of patent quality. Given the large
literature in economics, finance, and innovation that relies on forward citations as a
proxy for quality, these findings suggest that a re-examination of relevant prior findings
may be warranted.
Our paper also makes an important methodological contribution to the economic
literature by introducing the C-BERT methodology for causal inference. Economics is
steeped in the tradition of borrowing methodological innovations from adjacent fields.
Big data, machine learning, and AI are new approaches that are poised to revolutionize
empirical research in this field, Goldstein, Spatt, and Ye (2021). Causal inference using
text can help researchers in answering key open economic questions. Our paper provides
an initial roadmap for scholars to apply similar approaches in their own spheres.

34

Electronic copy available at: https://ssrn.com/abstract=4269703

References
Aghion, Philippe, John Van Reenen, and Luigi Zingales, 2013, Innovation and institu-
tional ownership, American Economic Review 103, 277–304.
Athey, Susan, and Guido W Imbens, 2019, Machine learning methods that economists
should know about, Annual Review of Economics 11, 685–725.
Beede, David N, Tiffany A Julian, David Langdon, George McKittrick, Beethika Khan,
and Mark E Doms, 2011, Women in stem: A gender gap to innovation, Economics and
Statistics Administration Issue Brief .
Bell, Alex, Raj Chetty, Xavier Jaravel, Neviana Petkova, and John Van Reenen, 2019, Who
becomes an inventor in america? the importance of exposure to innovation, Quarterly
Journal of Economics 134, 647–713.
Bellstam, Gustaf, Sanjai Bhagat, and J Anthony Cookson, 2021, A text-based analysis of
corporate innovation, Management Science 67, 4004–4031.
Card, David, Stefano DellaVigna, Patricia Funk, and Nagore Iriberri, 2020, Are referees
and editors in economics gender neutral?, Quarterly Journal of Economics 135, 269–327.
Chawla, Dalmeet Singh, 2016, Men cite themselves more than women do, Nature 535,
212.
Comenetz, Joshua, 2016, Frequently occurring surnames in the 2010 census, United States
Census Bureau 1–8.
Cong, Lin William, Tengyuan Liang, and Xiao Zhang, 2019, Textual factors: A scalable,
interpretable, and data-driven approach to analyzing unstructured information, In-
terpretable, and Data-driven Approach to Analyzing Unstructured Information (September 1,
2019) .
Cook, Lisa, 2020, Policies to broaden participation in the innovation process, Policy Pro-
posal, The Hamilton Project, Brookings Institution, Washington, DC .
Cook, Lisa D, 2014, Violence and economic activity: evidence from african american
patents, 1870–1940, Journal of Economic Growth 19, 221–257.
Cook, Lisa D, and Chaleampong Kongcharoen, 2010, The idea gap in pink and black,
Technical report, National Bureau of Economic Research.
Desai, Pranav, 2019, Biased regulators: Evidence from patent examiners, Working paper.
Devlin, Jacob, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova, 2018, Bert: Pre-
training of deep bidirectional transformers for language understanding, arXiv preprint
arXiv:1810.04805 .
Erel, Isil, Léa H Stern, Chenhao Tan, and Michael S Weisbach, 2021, Selecting directors
using machine learning, Review of Financial Studies 34, 3226–3264.

35

Electronic copy available at: https://ssrn.com/abstract=4269703

Farre-Mensa, Joan, Zack Liu, and Jordan Nickerson, 2022, Do startup patent acquisitions
affect inventor productivity?, Working paper.

Gavrilova, Evelina, and Steffen Juranek, 2021, Female inventors: The drivers of the gen-
der patenting gap, Working paper.

Gentzkow, Matthew, Bryan Kelly, and Matt Taddy, 2019a, Text as data, Journal of Economic
Literature 57, 535–74.

Gentzkow, Matthew, Jesse M Shapiro, and Matt Taddy, 2019b, Measuring group differ-
ences in high-dimensional choices: method and application to congressional speech,
Econometrica 87, 1307–1340.

Goldstein, Itay, Chester S Spatt, and Mao Ye, 2021, Big data in finance, Review of Financial
Studies 34, 3213–3225.

Graham, Stuart JH, Alan C Marco, and Richard Miller, 2018, The uspto patent examina-
tion research dataset: A window on patent processing, Journal of Economics & Manage-
ment Strategy 27, 554–578.

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

Hanley, Kathleen Weiss, and Gerard Hoberg, 2019, Dynamic interpretation of emerging
risks in the financial sector, Review of Financial Studies 32, 4543–4603.

Hansen, Stephen, Michael McMahon, and Andrea Prat, 2018, Transparency and delib-
eration within the fomc: a computational linguistics approach, Quarterly Journal of
Economics 133, 801–870.

Hengel, Erin, and Euyoung Moon, 2020, Gender and equality at top economics journals,
Working paper.

Hirschey, Mark, and Vernon J Richardson, 2004, Are scientific indicators of patent quality
useful to investors?, Journal of Empirical Finance 11, 91–107.

Hirshleifer, David, Po-Hsuan Hsu, and Dongmei Li, 2013, Innovative efficiency and stock
returns, Journal of Financial Economics 107, 632–654.

Hunt, Jennifer, Jean-Philippe Garant, Hannah Herman, and David J Munroe, 2013, Why
are women underrepresented amongst patentees?, Research Policy 42, 831–843.

Jensen, Kyle, Balázs Kovács, and Olav Sorenson, 2018, Gender differences in obtaining
and maintaining patent rights, Nature Biotechnology 36, 307–309.

Jha, Manish, Hongyi Liu, and Asaf Manela, 2022, Does finance benefit society? a lan-
guage embedding approach, Working paper.

36

Electronic copy available at: https://ssrn.com/abstract=4269703

Khetan, Vivek, Roshni Ramnani, Mayuresh Anand, Subhashis Sengupta, and Andrew E
Fano, 2022, Causal bert: Language models for causality detection between events ex-
pressed in text, in Intelligent Computing, 965–980 (Springer).

Koffi, Marlène, 2021, Innovative ideas and gender inequality, Working paper.

Koffi, Marlène, and Matt Marx, 2021, Cassatts in the attic, Working paper.

Kogan, Leonid, Dimitris Papanikolaou, Amit Seru, and Noah Stoffman, 2017, Techno-
logical innovation, resource allocation, and growth, Quarterly Journal of Economics 132,
665–712.

Li, Kai, Feng Mai, Rui Shen, and Xinyan Yan, 2021, Measuring corporate culture using
machine learning, Review of Financial Studies 34, 3265–3315.

Loughran, Tim, and Bill McDonald, 2016, Textual analysis in accounting and finance: A
survey, Journal of Accounting Research 54, 1187–1230.

Martinez, Gema Lax, Julio Raffo, Kaori Saito, et al., 2016, Identifying the gender of PCT
inventors, volume 33 (WIPO).

Moser, Petra, 2005, How do patent laws influence innovation? evidence from nineteenth-
century world’s fairs, American Economic Review 95, 1214–1236.

Moser, Petra, 2013, Patents and innovation: evidence from economic history, Journal of
Economic Perspectives 27, 23–44.

Oster, Emily, 2019, Unobservable selection and coefficient stability: Theory and evidence,
Journal of Business & Economic Statistics 37, 187–204.

Reshef, Oren, Abhay Aneja, and Gauri Subramani, 2021, Persistence and the gender
innovation gap: evidence from the us patent and trademark office, in Academy of Man-
agement Proceedings, volume 2021, 11626, Academy of Management Briarcliff Manor,
NY 10510.

Rouen, Ethan, Kunal Sachdeva, and Aaron Yoon, 2022, The evolution of esg reports and
the role of voluntary standards, Technical report.

Routledge, Bryan R, Stefano Sacchetto, and Noah A Smith, 2017, Predicting merger
targets and acquirers from text, Working paper.

Sarsons, Heather, 2017, Recognition for group work: Gender differences in academia,
American Economic Review 107, 141–45.

Sarsons, Heather, Klarita Gërxhani, Ernesto Reuben, and Arthur Schram, 2021, Gender
differences in recognition for group work, Journal of Political Economy 129, 101–147.

Shao, Yifan, Haoru Li, Jinghang Gu, Longhua Qian, and Guodong Zhou, 2021, Extraction
of causal relations based on sbel and bert model, Database 2021.

37

Electronic copy available at: https://ssrn.com/abstract=4269703

Sherman, Mila Getmansky, and Heather E Tookes, 2022, Female representation in the
academic finance profession, Journal of Finance 77, 317–365.

Tzioumis, Konstantinos, 2018, Demographic aspects of first names, Scientific Data 5, 1–9.

Vaswani, Ashish, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N
Gomez, Łukasz Kaiser, and Illia Polosukhin, 2017, Attention is all you need, Advances
in Neural Information Processing Systems 30.

Veitch, Victor, Dhanya Sridhar, and David Blei, 2020, Adapting text embeddings for
causal inference, in Conference on Uncertainty in Artificial Intelligence, 919–928, PMLR.

38

Electronic copy available at: https://ssrn.com/abstract=4269703

 (a) Panel A: Model for ATT, Confounding

(b) Panel B: Model for NDE, Mediating

FIGURE 1: DISTRIBUTION OF FORWARD CITATIONS

Citation is the outcome of interest, Gender is the treatment, and Text are the sequence of words. Panel A
depicts the average treatment effect, with the assumption that Text carries sufficient information to adjust
for confounding (common cause) between outcome and treatment. Panel B depicts the natural direct effect
(NDE), where the text is a mediator of the treatment on outcome.

39

Electronic copy available at: https://ssrn.com/abstract=4269703

 Inputs:
● Patent text
● Gender Indicator

BERT

Embedding

patent is written by a female Propensity Network

True False

Treatment
Male Citation Female Citation propensity
Network Network

Counterfactual
number of citations
Output

FIGURE 2: C-BERT ESTIMATION PROCEDURE

The figure illustrates the estimation procedure of C-BERT once the neural networks are trained. The light
blue block at the very top describes the input used for estimation. The green blocks are the four neural
networks trained using the patent data. The blue block describes the decision rule used for counterfactual
estimation. Finally, the red block is the output that combines the outputs of the citation estimation net-
works and the propensity score estimation network.

40

Electronic copy available at: https://ssrn.com/abstract=4269703

 20.0%

15.0%
Density

10.0%

5.0%

0.0%
0 25 50 75 100
Forward Citations
Female Male

(a) Panel A: Observed Forward Citations

20.0%

15.0%
Density

10.0%

5.0%

0.0%
0 25 50 75 100
Expected Forward Citations
Female Male

(b) Panel B: Model Implied Forward

FIGURE 3: DISTRIBUTION OF FORWARD CITATIONS

This figure illustrates the distribution of forward citations. Panel A uses forward citations observed in the
data, while Panel B uses the expected number of forward citations as implied by the model. The horizontal
axis counts the number of citations while the vertical axis measures the percent of the distribution. Red
bars correspond to females, blue bars correspond to males, and purple bars correspond to the overlapping
region. The distribution is truncated at 100 for ease of interpretation. The natural logarithm transformation
of these distributions is presented in Figure IA1. 41

Electronic copy available at: https://ssrn.com/abstract=4269703

 3.0%
Electronic copy available at: https://ssrn.com/abstract=4269703

2.0%
Density

1.0%
42

0.0%

−20 0 20
Bias in Forward Citations
Female Male

FIGURE 4: UNDERCITATION OF FEMALE LEAD PATENTS

This figure illustrates the difference between forward citations and expected forward citations, as defined by Equation 2. The horizontal axis counts
the additional number of citations that a patent should have received after adjusting. Red bars correspond to females, blue bars correspond to
males, and purple bars correspond to the overlapping region. The distribution is truncated between -30 to 30.
 Motors & Engines + Parts
Transportation
Metal Working
Mat. Proc & Handling
Miscellaneous (Mech)
Optics
Electrical Lighting
Power Systems
Nuclear & X−rays
Measuring & Testing
Miscellaneous (Elec)
Electrical Devices
Semiconductor Devices
Sub Categories (NBER)

Drugs
Genetics
Miscellaneous (Drgs&Med)
Surgery & Med Inst.
Information Storage
Computer Peripherials
Computer Hardware & Software
Communications
Electronic business methods and software
Organic Compounds
Miscellaneous (Chemical)
Gas
Resins
Coating
Heating
Furniture,House Fixtures
Pipes & Joints
Apparel & Textile
Miscellaneous (Others)
Agriculture,Husbandry,Food
Receptacles
Earth Working & Wells
Amusement Devices
−20 −15 −10 −5 0 5
Bias In Citation

Others Cmp&Cmm Elec

Category (NBER)
Chemical Drgs&Med Mech

FIGURE 5: DELTA IN CITATIONS BY PATENT SUB CATEGORIES

This figure illustrates the coefficients of Equation 5. For ease of interpretation, each point corresponds to
the linear combination of the baseline result for females and the interaction terms, presented in Table IA4.
Whiskers correspond to a 95% confidence internal. Coefficients are sorted by by patent category and then
by the magnitude of the estimate. Colors corresponds to the patent category as defined by the NBER,
where pink observations correspond the mechanical (Mech), purple corresponds to electrical (Elec), blue
observations correspond to drugs and medical (Drgs&Med), light green observations correspond to com-
puters and communication (Cmp&Comm), dark green observations correspond to chemical (Chemical),
yellow observations correspond to other (Other) categories. The red dotted line is plotted at the zero
intercept, representing a no effect.

43

Electronic copy available at: https://ssrn.com/abstract=4269703

 0
Bias in Forward Citations

−1

−2
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
FIGURE 6: EVOLUTION OF DELTA OVER TIME

This figure illustrates the evolution of citations over time. The horizontal axis corresponds to the year a
patent was granted. The vertical axis corresponds to the delta in forward citations, with negative numbers
corresponding to undercitation. Forward citations are computed within the first ten years the patent was
granted. Each point represents an estimate from a separate estimate, with error bars corresponding to a
95% confidence interval. All estimates include customer, examiner, and examiner art unit fixed effects.

44

Electronic copy available at: https://ssrn.com/abstract=4269703

 0.0
Estimate

−0.5

−1.0
er

er

or
o
in

nt

in

nt
am

am
ve

ve
In

In
Ex

Ex
e

e
al

al
e

e
al

al
m

M
m

M
Fe
Fe

Female Male

FIGURE 7: DELTA IN EXAMINER AND INVENTOR ADDED CITATIONS

This figure illustrates the delta by examiner and inventor added citations and by gender. The red columns
correspond to female-added citations, and the blue columns correspond to male-added citations. Error
bars correspond to the 95% confidence interval. Estimates correspond to column (4) of Table 8 and Table 9.

45

Electronic copy available at: https://ssrn.com/abstract=4269703

 TABLE 1: SUMMARY STATISTICS

This table provides summary statistics on patents and citations. The sample covers patents issued from
1976-01-01 through 2021-12-31. Panel A presents a two-way table of forward citations by gender. Panel
B presents a two-way table of patents in the top decile by gender. Panel C presents a two-way table of
patents by their cooperative patent classification (CPC). ***, **, * denote significance at the 1%, 5%, and
10% level, respectively. Data Source: USPTO.

Gender of Lead Inventor Male Female

N Mean SD N Mean SD Test
Panel A: Difference in Forward Citations

Forward Citation 290786 18.09 51.66 236562 15 45.34 F= 518.967∗∗∗

Panel B: Difference By Top Decile Innovations
Top Decile 290786 236562 X2= 572.671∗∗∗
→ No 260016 89% 216170 91%
→ Yes 30770 11% 20392 9%
Panel C: Difference by Cooperative Patent Classification
CPC Section 290733 236533 χ2 = 4871.596∗∗∗
→ Chemistry 30780 11% 32379 14%
→ Electricity 67174 23% 59396 25%
→ Fixed Constructions 8551 3% 4474 2%
→ Human Necessities 35268 12% 31598 13%
→ Mechanical Engineering 24745 9% 13263 6%
→ Performing Operations 46932 16% 29477 12%
→ Physics 74182 26% 63730 27%
→ Textiles 3101 1% 2216 1%

46

Electronic copy available at: https://ssrn.com/abstract=4269703

 TABLE 2: FORWARD CITATION WITHOUT MODEL ADJUSTMENTS

This table reports estimates of Equation 1 and studies the number of forward citations by the gender of
the lead inventor. The sample includes both patents that received a citation and those that did not receive
a citation. Panel A uses the sample of all patents while Panel B uses patents with a positive number of
forward citations. The sample covers patents issued from 1976-01-01 through 2021-12-31. Standard errors
are clustered at the patent customer and patent issue year level. ***, **, * denote significance at the 1%,
5%, and 10% level, respectively. Data source: USPTO.

Panel A: Extensitve Margin

Forward Citations
(1) (2) (3) (4)
Lead Female Inventor −3.947∗∗∗ −1.443∗∗∗ −1.097∗∗∗ −0.830∗∗∗
(0.266) (0.400) (0.280) (0.270)

Intercept 14.953∗∗∗
(4.497)

Customer FE No No Yes No
Examiner FE No No Yes No
Examiner x Customer FE No No No Yes
Art Unit FE No Yes Yes Yes
Patent Issue Year FE No Yes Yes Yes
Observations 907,996 907,996 907,996 907,996
Adjusted R2 0.002 0.122 0.199 0.375

Panel B: Intensive Margin

Forward Citations
(1) (2) (3) (4)
Lead Female Inventor −3.086∗∗∗ −1.771∗∗∗ −1.306∗∗∗ −0.953∗∗∗
(0.366) (0.345) (0.223) (0.221)

Intercept 18.091∗∗∗
(2.952)

Customer FE No No Yes No
Examiner FE No No Yes No
Examiner x Customer FE No No No Yes
Art Unit FE No Yes Yes Yes
Patent Issue Year FE No Yes Yes Yes
Observations 527,348 527,348 527,348 527,348
Adjusted R2 0.001 0.095 0.182 0.354

47

Electronic copy available at: https://ssrn.com/abstract=4269703

 TABLE 3: MODEL ADJUSTED FORWARD CITATION

This table reports estimates of Equation 2 and studies the number of forward citations by the gender of
the lead inventor. The sample includes both patents that received a citation and those that did not receive
a citation. Panel A uses the sample of all patents while Panel B uses patents with a positive number of
forward citations. The sample covers patents issued from 1976-01-01 through 2021-12-31. Standard errors
are clustered at the patent customer and patent issue year level. ***, **, * denote significance at the 1%,
5%, and 10% level, respectively. Data source: USPTO.

Panel A: Extensive Margin

Delta in Forward Citations
(1) (2) (3) (4)
Lead Female Inventor −3.796∗∗ −4.057∗∗ −3.952∗∗ −3.263
(1.547) (1.537) (1.571) (1.962)

Intercept 0.235∗∗∗
(0.085)

Customer FE No No Yes No
Examiner FE No No Yes No
Examiner x Customer FE No No No Yes
Art Unit FE No Yes Yes Yes
Patent Issue Year FE No Yes Yes Yes
Observations 907,996 907,996 907,996 907,996
Adjusted R2 0.006 0.017 0.051 0.298

Panel B: Intensive Margin

Delta in Forward Citations
(1) (2) (3) (4)
Lead Female Inventor −3.791∗∗∗ −3.853∗∗∗ −3.769∗∗∗ −3.312∗∗∗
(0.779) (0.774) (0.792) (1.136)

Intercept 1.097∗∗∗
(0.191)

Customer FE No No Yes No
Examiner FE No No Yes No
Examiner x Customer FE No No No Yes
Art Unit FE No Yes Yes Yes
Patent Issue Year FE No Yes Yes Yes
Observations 527,348 527,348 527,348 527,348
Adjusted R2 0.006 0.011 0.062 0.328

48

Electronic copy available at: https://ssrn.com/abstract=4269703

 TABLE 4: CITATIONS IN TOP DECILE

This table studies patents that receive forward citations in the top decile. Panel A documents the rela-
tionship between a patent’s lead inventor’s gender and the propensity to receive citations placing them
in the top decile. Panel B documents the relationship between a patent’s lead inventor’s gender and the
model’s prediction a patent would be in the top decile of citations. The sample covers patents issued from
1976-01-01 through 2021-12-31. Standard errors are clustered at the patent customer and patent issue year
level. ***, **, * denote significance at the 1%, 5%, and 10% level, respectively. Data source: USPTO.

Panel A: Citations Without Model Adjustment

Top Decile Patent
(1) (2) (3) (4)
Lead Inventor Female −0.020∗∗∗ −0.011∗∗∗ −0.009∗∗∗ −0.006∗∗∗
(0.003) (0.002) (0.002) (0.002)

Intercept 0.106∗∗∗
(0.022)

Customer FE No No Yes No
Examiner FE No No Yes No
Examiner x Customer FE No No No Yes
Art Unit FE No Yes Yes Yes
Patent Issue Year FE No Yes Yes Yes
Observations 527,348 527,348 527,348 527,348
Adjusted R2 0.001 0.107 0.146 0.185

Panel B: Counterfactual
Flipped to Top Decile
(1) (2) (3) (4)
Lead Inventor Female 0.017∗∗∗ 0.018∗∗∗ 0.018∗∗∗ 0.017∗∗∗
(0.004) (0.004) (0.004) (0.006)

Intercept 0.010∗∗∗
(0.002)

Customer FE No No Yes No
Examiner FE No No Yes No
Examiner x Customer FE No No No Yes
Art Unit FE No Yes Yes Yes
Patent Issue Year FE No Yes Yes Yes
Observations 527,348 527,348 527,348 527,348
Adjusted R2 0.004 0.011 0.016 0.066

49

Electronic copy available at: https://ssrn.com/abstract=4269703

 TABLE 5: CITATION BY NBER CATEOGRY

This table estimates the difference in citations by NBER Cateogry. Column (1) uses the number of forward
citations as its dependent variable, while Column (2) uses the difference in forward citations, as defined
by Equation 2. Estimates include interactions for the patent category based on NBER Categories. All
specifications include NBER Cateogry, Examiner × Customer, and Patent Issue Year fixed effects. The sample
covers patents issued from 1976-01-01 through 2021-12-31. Standard errors are clustered at the patent
customer and patent issue year level. ***, **, * denote significance at the 1%, 5%, and 10% level, respectively.
Data source: USPTO.

Dependent variable:
Forward Citations Delta in Forward Citations
(1) (2)
Female Lead Inventor −0.077 −3.154∗∗∗
(0.395) (0.337)

Chemical × Female Lead Inventor −0.602∗∗ −0.512∗∗∗

(0.263) (0.167)

Computers and Communication × Female Lead Inventor −1.320∗∗∗ −2.837∗∗∗

(0.270) (0.990)

Drugs and Medical × Female Lead Inventor −5.797∗∗∗ −7.260∗∗∗

(0.345) (0.638)

Electrical × Female Lead Inventor −0.767∗∗ 0.025

(0.347) (0.147)

Mechanical × Female Lead Inventor 0.096 0.572∗∗∗

(0.263) (0.157)

NBER Category FE Yes Yes

Examiner x Customer FE Yes Yes
Examiner Art Unit FE Yes Yes
Patent Issue Year FE Yes Yes
Observations 436,458 436,458
Adjusted R2 0.296 0.282

50

Electronic copy available at: https://ssrn.com/abstract=4269703

 TABLE 6: FORWARD CITATIONS, EMERGING FIELDS

This table studies the citations to new fields of innovation. New Field takes the value of one if the art unit
first appeared within five years of the patent being granted. The dependent variable is the difference in
the observed number and the expected number of citations for a patent, as defined by Equation 2. The
sample covers patents issued from 1976-01-01 through 2021-12-31. Standard errors are clustered at the
patent customer and patent issue year level. ***, **, * denote significance at the 1%, 5%, and 10% level,
respectively. Data source: USPTO.

Panel A: Forward Citations

Forward Citations
(1) (2) (3) (4)
Emerging Field −0.099 0.032 −0.071 −0.541
(2.506) (0.551) (0.534) (0.344)

Female Lead Inventor −3.351∗∗∗ −1.703∗∗∗ −1.324∗∗∗ −1.013∗∗∗

(0.524) (0.339) (0.220) (0.160)

Emerging Field × Female Lead Inventor 0.950 −0.246 0.065 0.282

(0.780) (0.430) (0.324) (0.374)

Intercept 18.119∗∗∗
(3.486)

Customer FE No No Yes No
Examiner FE No No Yes No
Examiner x Customer FE No No No Yes
Art Unit FE No Yes Yes Yes
Patent Issue Year FE No Yes Yes Yes
Observations 527,348 527,348 527,348 527,348
Adjusted R2 0.001 0.095 0.182 0.354

Panel B: Delta in Forward Citations

Delta in Forward Citations
(1) (2) (3) (4)
Emerging Field −0.016 0.213 0.308 0.084
(0.185) (0.352) (0.312) (0.202)

Lead Female Inventor −3.612∗∗∗ −3.693∗∗∗ −3.637∗∗∗ −3.305∗∗∗

(0.823) (0.831) (0.847) (1.141)

Emerging Field × Lead Female Inventor −0.643 −0.573 −0.484 −0.036

(0.603) (0.598) (0.568) (0.298)

Intercept 1.102∗∗∗
(0.183)

Customer FE No No Yes No
Examiner FE No No Yes No
Examiner x Customer FE No No No Yes
Art Unit FE No Yes Yes Yes
Patent Issue Year FE No Yes Yes Yes
Observations 527,348 527,348 527,348 527,348
Adjusted R2 0.006 0.011 0.062 0.328

51

Electronic copy available at: https://ssrn.com/abstract=4269703

 TABLE 7: YEARS AFTER PATENT IS GRANTED

This table estimates Equation 1 and studies the difference in forward citations by the number of years after
the patent was granted. Column (1) – (4), study the difference in forward citations 0-1, 2-5, 6-10, and 11-20
years after they are granted, respectively. All specifications use Art Unit, Examiner × Customer, and Patent
Grant Year fixed effects. The sample covers patents issued from 1976-01-01 through 2021-12-31. Standard
errors are clustered at the patent customer and patent issue year level. ***, **, * denote significance at the
1%, 5%, and 10% level, respectively. Data source: USPTO.

Delta in Forward Citations

0-1 Years 2-5 Years 6-10 Years 11-20 Years
(1) (2) (3) (4)
Lead Inventor Female −0.369∗∗∗ −0.626∗∗∗ −1.185∗∗∗ −2.479∗∗∗
(0.105) (0.066) (0.108) (0.252)

Examiner x Customer FE Yes Yes Yes Yes

Art Unit FE Yes Yes Yes Yes
Patent Issue Year FE Yes Yes Yes Yes
Observations 12,238 237,734 258,853 223,519
Adjusted R2 −1.241 −0.247 0.463 0.467

52

Electronic copy available at: https://ssrn.com/abstract=4269703

 TABLE 8: EXAMINER-ADDED CITATIONS

This table studies the source of examiner-added citations for male inventors. The dependent variable is
the difference in forward citations. Panel A uses the difference in forward citations that were added by
female lead examiners as its dependent variable. Panel B uses the difference in forward citations that
were added by male lead examiners as its dependent variable. The sample covers patents issued from
1976-01-01 through 2021-12-31. Note, the source of citations is only available following the start of 2001.
The sample covers patents issued from 1976-01-01 through 2021-12-31. Standard errors are clustered at
the patent customer and patent issue year level. ***, **, * denote significance at the 1%, 5%, and 10% level,
respectively. Data source: USPTO.

Panel A: Citation Added by Female Lead Examiner

Delta in Forward Citations
(1) (2) (3) (4)
Lead Female Inventor −0.029∗∗∗ −0.021∗∗ −0.014 −0.028
(0.011) (0.009) (0.011) (0.018)

Intercept 0.131∗∗∗
(0.038)

Customer FE No No Yes No
Examiner FE No No Yes No
Examiner x Customer FE No No No Yes
Art Unit FE No Yes Yes Yes
Patent Issue Year FE No Yes Yes Yes
Observations 66,757 66,757 66,757 66,757
Adjusted R2 0.0001 0.017 0.062 0.101

Panel B: Citation Added by Male Lead Examiners

Delta in Forward Citations
(1) (2) (3) (4)
Lead Female Inventor −0.006 0.009 0.004 0.003
(0.034) (0.038) (0.037) (0.049)

Intercept 0.060
(0.065)

Customer FE No No Yes No
Examiner FE No No Yes No
Examiner x Customer FE No No No Yes
Art Unit FE No Yes Yes Yes
Patent Issue Year FE No Yes Yes Yes
Observations 180,397 180,397 180,397 180,397
Adjusted R2 −0.00000 0.010 0.046 0.219

53

Electronic copy available at: https://ssrn.com/abstract=4269703

 TABLE 9: INVENTOR-ADDED CITATIONS

This table studies the source of inventor-added citations. The dependent variable is the difference in for-
ward citations. Panel A uses the difference in forward citations that were added by female lead inventors
as its dependent variable. Panel B uses the difference in forward citations that were added by male lead
inventors as its dependent variable. The sample covers patents issued from 1976-01-01 through 2021-12-31.
Note, the source of citations is only available following the start of 2001. Standard errors are clustered at
the patent customer and patent issue year level. ***, **, * denote significance at the 1%, 5%, and 10% level,
respectively. Data source: USPTO.

Panel A: Citation Added by Female Lead Inventors

Delta in Forward Citations
(1) (2) (3) (4)
Lead Female Inventor −0.508∗∗∗ −0.461∗∗∗ −0.473∗∗∗ −0.468∗∗∗
(0.051) (0.053) (0.033) (0.039)

Intercept 0.609∗∗∗
(0.063)

Customer FE No No Yes No
Examiner FE No No Yes No
Examiner x Customer FE No No No Yes
Art Unit FE No Yes Yes Yes
Patent Issue Year FE No Yes Yes Yes
Observations 29,984 29,984 29,984 29,984
Adjusted R2 0.004 0.019 0.075 0.223

Panel B: Citation Added by Male Lead Inventors

Delta in Forward Citations
(1) (2) (3) (4)
Lead Female Inventor −1.363∗∗∗ −1.360∗∗∗ −1.306∗∗∗ −0.990∗∗∗
(0.174) (0.174) (0.172) (0.178)

Intercept 1.083∗∗∗
(0.171)

Customer FE No No Yes No
Examiner FE No No Yes No
Examiner x Customer FE No No No Yes
Art Unit FE No Yes Yes Yes
Patent Issue Year FE No Yes Yes Yes
Observations 282,370 282,370 282,370 282,370
Adjusted R2 0.001 0.010 0.091 0.191

54

Electronic copy available at: https://ssrn.com/abstract=4269703

 TABLE 10: FORWARD CITATIONS AND VALUE OF PATENT

This table studies the relationship between the measures of citations and the market-implied value of
patents. The dependent variable for both panels use the log value of innovation, deflated to 1982 (million)
dollars using the CPI, as calculated in Kogan et al. (2017). The sample covers patents issued from 1976-01-
01 through 2021-12-31. Standard errors are clustered at the patent customer and patent issue year level.
***, **, * denote significance at the 1%, 5%, and 10% level, respectively. Data source: USPTO.

log(dollar )
(1) (2) (3) (4)
Female Lead Inventor 0.083∗∗∗ −0.014 −0.020 −0.050
(0.029) (0.033) (0.029) (0.041)

Forward Citation 0.00003 −0.0001 0.0001 0.0001

(0.0003) (0.0003) (0.0002) (0.0003)

[
Forward Citation 0.004∗∗∗ 0.004∗∗∗ 0.002∗∗∗ 0.003∗∗∗
(0.001) (0.0003) (0.001) (0.001)

Intercept 0.716∗∗∗
(0.138)

Customer FE No No Yes No
Examiner FE No No Yes No
Examiner x Customer FE No No No Yes
Art Unit FE No Yes Yes Yes
Patent Issue Year FE No Yes Yes Yes
Observations 202,865 202,865 202,865 202,865
Adjusted R2 0.008 0.127 0.438 0.321

55

Electronic copy available at: https://ssrn.com/abstract=4269703

 INTERNET APPENDIX
FOR ONLINE PUBLICATION

56

Electronic copy available at: https://ssrn.com/abstract=4269703

Appendix A Explanation of Causal BERT (C-BERT)
C-BERT is a neural network based architecture that estimates counterfactuals of a bi-
nary treatment where all of the covariates needed for causal identification are contained
within a given text. To use C-BERT to identify the effect of gender on the impact of
patents, we first need to train the model. As shown in Figure IA2, the input data for
training contains three types of information: the texts of patents, gender indicators of
the author(s), and the observed number of citations on the patents. There are four neu-
ral networks that need to be trained: a BERT model for generating text embeddings,
a logit-linear model the maps embeddings to treatment propensities, and two 2-layer
perceptrons that map from embeddings to male and female predicted number of cita-
tions, respectively. The final loss function is a weighted average of the losses of these
four neural networks. After the model is trained, we can use it to estimate the coun-
terfactual number of citations of male written patents if they were written by females
and vice versa. As shown in Figure 2, to estimate these counterfactuals, we run the
trained C-BERT model where the input data contains the texts of the patents and gender
indicators of the author(s). The texts are first passed through the trained BERT model
to generate a vector embedding for each patent. Then each embedding-gender pair is
passed through a decision step: if the author(s) are male, the embedding is passed to
the female citations network and if it is written by female(s), the embedding is passed
to the male citation network. The counterfactual number of citations are then computed
by these two networks. In parallel, regardless of the gender indicators, each embed-
ding is passed through the propensity network to estimate the treatment propensity of
this patent. Finally, the output of the model is a set of counterfactual citation-treatment
propensity pairs that each corresponds to one patent.

57

Electronic copy available at: https://ssrn.com/abstract=4269703

Appendix A Encoder Architecture
The encoder architecture works as follows. Let W denote the original input sentence in
words. As shown in Figure IA3, before entering the encoder, W is broken down into
three parts: a token embedding EW T , which represents the content of the sentence; a

segmentation embedding EW S , which labels tokens with the sentence they belong to; and

a positional embedding EW P , which represents the relative distances between each pair of

tokens (a “token” is a word or a part of a word if the word is long). A linear combination
of these three embeddings then goes into the encoder.
The first step of the encoder is a multi-headed attention layer. Its mechanism can
described as follows. Let EW denote the input embedding of the encoder. For a given
token Wi in sentence W, the embedding is denoted EiW . The attention layer calculates
the projection of EiW onto all token embeddings, including itself, using a dot product.
The final output of the single-headed attention layer for each token embedding is a
weighted average of all token embeddings, where the weights are the cosine projection
coefficient of the current token embedding on to each token embedding. A multi-headed
attention layer is analogous to a forest of single-headed attention layers. To construct a k-
headed attention layer using a pk dimensional token embedding, we randomly split the
pk dimensional embedding of each tokens into k groups of p dimensional embeddings.
We then build a single-headed attention layer with one subset of the token embeddings.
Finally, we take a weighted average of all of the output of the k heads.
The output of this multi-headed attention layer is then passed through a normal-
ization layer with residual connection. Residual connection is achieved by passing the
input of the multi-headed attention layer directly to the normalization layer along with
the output of the multi-headed attention layer. This residual connection allows gradients
to directly flow from the input of the multi-headed attention layer to the next layer while
not going through the multi-headed attention layer. After the normalization layer, the
output is passed through a feed forward layer, which converts the output of the normal-
ization layer to the same format as the input of the encoder module. This allows us to
stack multiple encoder modules together, where the previous encoder’s output can be
used as the input for the next encoder. The reason we stack encoders is that the first
encoder learns the contextual relationship between pairs of tokens, the second encoder
learns the relationship between pairs of pairs of tokens, and so forth. For the following
discussions in this paper, we use the word "embedding" to mean the output embedding
of the encoder at the text level.
The pre-trained BERT model uses the encoder architecture to train for two tasks:

58

Electronic copy available at: https://ssrn.com/abstract=4269703

masked language modeling (MLM) and next sentence prediction (NSP). To train the
MLM task, a random subset of tokens in the input sentence is masked with a trivial
embedding vector. Then, after this sentence goes through the encoder, the output goes
through a fully connected linear layer and a softmax layer to predict what the masked
tokens in the original sentence are. This loss is computed using cross-entropy. To train
for next sentence prediction, the encoder takes pairs of sentences as inputs and predicts
whether the second sentence should appear after the first sentence. This loss is computed
using binary-cross-entropy. The final trained BERT model can output embeddings of
sentences or entire texts that represents not only the meaning of the tokens but also the
contextual relationship between tokens and sentences.

59

Electronic copy available at: https://ssrn.com/abstract=4269703

 20.0%

15.0%
Density

10.0%

5.0%

0.0%
0.0 2.5 5.0 7.5
log(Forward Citation)
Female Male

(a) Forward Citations

20.0%

15.0%
Density

10.0%

5.0%

0.0%
0 2 4 6 8
log(Expected Forward Citation)
Female Male

(b) Model Implied Forward

FIGURE IA1: DISTRIBUTION OF FORWARD CITATIONS

This figure illustrates the transformation from forward citations to expected forward citations. Panel
A uses the natural logarithm of forward citations while Panel B uses the natural logarithm of forward
citations expected from our model. The vertical axis in both panels measures the percent of the distribu-
tion. Red bars correspond to females, blue bars correspond to males, and purple bars correspond to the
overlapping region.
60

Electronic copy available at: https://ssrn.com/abstract=4269703

 Inputs:
● Patent texts
● Gender Indicators
● Citations

texts texts and texts and

gender citations
Female Citation
BERT fine-tuning Propensity Network
Network
Male Citation
Network

Loss

Parameter Optimization

FIGURE IA2: C-BERT TRAINING PROCEDURE

The figure illustrates the training procedure of C-BERT once the neural networks are trained. The light
blue block at the top describes the input used for estimation. The green blocks are the four neural networks
that are trained using the patent data. The purple block denotes the loss function of the model which is a
weighted average of the loss of all four networks. Finally, the red block denotes the optimization algorithm
that allows the model to get a step toward fitting the training data.

61

Electronic copy available at: https://ssrn.com/abstract=4269703

 Output

Feed Forward Layer

Normalization Layer

Residual Multi-headed
connection Attention 1 Attention 2 Attention 3
attention

Token Embedding, Segmentation Embedding,

Positional Embedding

FIGURE IA3: ENCODER MODULE

This figure illustrates the structure of the encoder module. The light blue block at the bottom describes
the input. The yellow blocks are the layers within the encoder, and the red block is the output.

62

Electronic copy available at: https://ssrn.com/abstract=4269703

 2.00%

1.50%
Density

1.00%

0.50%

0.00%

−1.0 −0.5 0.0 0.5 1.0

Actual Minus Sythetic Counterfactual

FIGURE IA4: SYNTHETIC DATA TESTS

The figure evaluates the quality of fit of our neural network. The black dashed line is centered at the mean
difference in the data (-0.008). We are unable to reject the null that the true difference in means is equal to
zero (p=0.2693).

63

Electronic copy available at: https://ssrn.com/abstract=4269703

 Loss (MSE per batch)

1
10

15

20
5

Epochs

FIGURE IA5: LOSS FUNCTION

This figure illustrates the loss function of the C-BERT model. The horizontal axis corresponds to the
number of complete passes of the training dataset through the algorithm or epoch. The vertical axis
corresponds to the loss function and is the mean square error per batch.

64

Electronic copy available at: https://ssrn.com/abstract=4269703

 TABLE IA1: DIFFERENCE IN WRITING STYLES

This table reports the difference in the writing style between males and females. The sample covers patents
issued from 1976-01-01 through 2021-12-31 and has at least 120 words. ***, **, * denote significance at the
1%, 5%, and 10% level, respectively. Data source: USPTO, Google Patents.

Gender Male Female

N Mean SD N Mean SD Test

Number of Words 1981500 159.64 43.26 136152 155.22 38.51 F= 1346.691∗∗∗

Sentiment Score 1981500 0.04 0.13 136152 0.05 0.13 F= 35.241∗∗∗
Sentiment 1981500 0.36 0.9 136152 0.37 0.9 F= 6.967∗∗∗
Flesch-Kincaid 197487 23.33 15.01 14278 23.87 15.12 F= 17.13∗∗∗
Flesh 197487 9.78 41.76 14278 7.04 42.2 F= 57.028∗∗∗
Gunning-Fog 197487 27.26 15.91 14278 27.66 16.04 F= 8.487∗∗∗
Coleman-Liau 197487 13.4 2.44 14278 13.63 2.61 F= 119.976∗∗∗
Dale-Chall 197487 11.88 2.42 14278 12.17 2.42 F= 194.12∗∗∗
Ari 197487 26.64 19.66 14278 27.17 19.81 F= 9.695∗∗∗
Linsear-Write 197487 34.08 28.03 14278 34.78 28.54 F= 8.082∗∗∗
Spache 197487 11.57 5.58 14278 11.75 5.62 F= 14.111∗∗∗

65

Electronic copy available at: https://ssrn.com/abstract=4269703

 TABLE IA2: ASSOCIATION BETWEEN GENDER AND FORWARD CITATION

This table uses all patent observations without applying C-BERT. The sample covers patents issued from
1976-01-01 through 2021-12-31. Standard errors are clustered at the patent customer and patent issue year
level. ***, **, * denote significance at the 1%, 5%, and 10% level, respectively. Data source: USPTO.

Forward Citations
(1) (2) (3) (4)
Lead Female Inventor −3.621∗∗∗ −1.023∗∗∗ −0.784∗∗∗ −0.591∗∗
(0.276) (0.310) (0.240) (0.256)

Intercept 15.185∗∗∗
(4.520)

Customer FE No No Yes No
Examiner FE No No Yes No
Examiner x Customer FE No No No Yes
Art Unit FE No Yes Yes Yes
Patent Issue Year FE No Yes Yes Yes
Observations 6,312,796 6,312,796 6,312,796 6,312,796
Adjusted R2 0.0004 0.126 0.206 0.318

66

Electronic copy available at: https://ssrn.com/abstract=4269703

 TABLE IA3: CITATION BY CPC SECTION

This table estimates the difference in citations by CPC Section. Column (1) uses the number of forward
citations as its dependent variable, while Column (2) uses the difference in forward citations, as defined by
Equation 2. Estimates include interactions for the patent category based on CPC Section. All specifications
include CPC Section, Examiner × Customer, and Patent Issue Year fixed effects. The sample covers patents
issued from 1976-01-01 through 2021-12-31. Standard errors are clustered at the patent customer and
patent issue year level. ***, **, * denote significance at the 1%, 5%, and 10% level, respectively. Data
source: USPTO.

Dependent variable:
Forward Citations Delta in Forward Citations
(1) (2)
Lead Female Inventor −1.192∗ −3.537∗∗∗
(0.685) (0.821)

Electricity × Lead Female Inventor 0.054 0.643

(0.568) (0.472)

Fixed Constructions × Lead Female Inventor 0.280 1.676∗∗∗

(0.507) (0.229)

Human Necessities × Lead Female Inventor −1.476∗∗∗ −3.286∗∗∗

(0.350) (1.007)

Mechanical Engineering × Lead Female Inventor 1.356∗ 2.603∗∗∗

(0.768) (0.649)

Performing Operations × Lead Female Inventor 0.796∗ 1.037∗∗∗

(0.400) (0.092)

Physics × Lead Female Inventor 0.435 0.444

(0.453) (0.597)

Textiles × Lead Female Inventor 2.429∗∗ 1.021∗∗

(0.998) (0.487)

CPC Section FE Yes Yes

Examiner x Customer FE Yes Yes
Examiner Art Unit FE Yes Yes
Patent Issue Year FE Yes Yes
Observations 527,266 527,266
Adjusted R2 0.356 0.329

67

Electronic copy available at: https://ssrn.com/abstract=4269703

 TABLE IA4: CITATION BY PATENT NBER SUB CATEGORY

This table estimates the difference in citations by NBER subcategories. Column (1) uses the actual num-
ber of citations as its dependent variable, while Column (2) uses the Delta in citations, as defined by
Equation 2. Estimates include interactions for the patent subcategory based on NBER classifications. All
specifications include Patent Subcategory, Examiner × Customer, and Patent Issue Year fixed effects. Standard
errors are clustered at the patent customer and patent issue year level. ***, **, * denote significance at the
1%, 5%, and 10% level, respectively. Data source: USPTO.

Dependent variable:
Forward Citations Delta in Forward Citations
(1) (2)
Female Lead Inventor 0.499 −1.915∗∗∗
(0.470) (0.440)

Agriculture,Husbandry,Food× Female Lead Inventor 0.204 −1.885

(0.674) (1.176)

Amusement Devices× Female Lead Inventor 1.488 −3.426∗∗∗

(2.544) (1.048)

Apparel & Textile× Female Lead Inventor 2.114∗∗∗ −0.879

(0.662) (0.522)

Coating× Female Lead Inventor −1.591 −2.853∗∗

(0.967) (1.113)

Communications× Female Lead Inventor −2.118∗∗∗ −5.103∗∗∗

(0.528) (1.794)

Computer Hardware & Software× Female Lead Inventor −1.807∗∗ −4.650∗∗

(0.883) (1.819)

Computer Peripherials× Female Lead Inventor 0.546 −2.940∗∗∗

(0.955) (0.944)

Drugs× Female Lead Inventor −4.586∗∗∗ −3.945∗∗∗

(1.045) (0.870)

Earth Working & Wells× Female Lead Inventor −2.061∗∗∗ −2.134∗∗

(0.637) (0.823)

Electrical Devices× Female Lead Inventor −1.157∗∗ −1.314

(0.473) (0.890)

Electrical Lighting× Female Lead Inventor −2.045∗ −0.249

(1.195) (0.494)

Electronic business methods and software× Female Lead Inventor −8.866∗∗ −8.702∗
(3.623) (4.926)

Furniture,House Fixtures× Female Lead Inventor −0.214 0.207

(0.792) (0.482)

Gas× Female Lead Inventor 0.102 −1.773

(1.447) (1.397)

Genetics× Female Lead Inventor −8.841∗ −8.189

(4.996) (6.230)

Heating× Female Lead Inventor −1.353 1.241∗∗

(1.130) (0.585)

Information Storage× Female Lead Inventor −2.166∗∗ −0.798

(0.879) (0.948)
68
continued on next page...

Electronic copy available at: https://ssrn.com/abstract=4269703

 TABLE IA4: CITATION BY PATENT NBER SUB CATEGORY (CONTINUED)

Dependent variable:
Forward Citations Delta in Forward Citations
(1) (2)
Mat. Proc & Handling× Female Lead Inventor 0.259 −0.341
(0.665) (0.427)

Measuring & Testing× Female Lead Inventor −0.728 −0.553

(0.513) (0.887)

Metal Working× Female Lead Inventor 0.509 −0.217

(1.003) (0.586)

Miscellaneous (Chemical)× Female Lead Inventor −1.127∗ −1.556∗

(0.582) (0.776)

Miscellaneous (Drgs&Med)× Female Lead Inventor −5.059 −10.416∗∗∗

(3.144) (1.845)

Miscellaneous (Elec)× Female Lead Inventor −0.584 −1.058

(0.592) (0.976)

Miscellaneous (Mech)× Female Lead Inventor −1.440 −2.400∗∗

(1.775) (1.045)

Miscellaneous (Others)× Female Lead Inventor −0.532 −1.303

(0.927) (0.795)

Motors & Engines + Parts× Female Lead Inventor 0.0002 0.483

(0.662) (0.361)

Nuclear & X-rays× Female Lead Inventor −2.954∗∗∗ −0.485

(0.961) (0.606)

Optics× Female Lead Inventor −2.472 −2.469∗∗

(1.711) (1.099)

Organic Compounds× Female Lead Inventor −1.812 −0.415

(1.354) (0.608)

Pipes & Joints× Female Lead Inventor −4.125∗∗ −0.750

(1.746) (0.600)

Power Systems× Female Lead Inventor −0.896 −0.436

(0.565) (0.630)

Receptacles× Female Lead Inventor −2.042 −1.921∗∗∗

(1.275) (0.524)

Resins× Female Lead Inventor −1.825∗∗ −2.785∗∗∗

(0.824) (0.924)

Semiconductor Devices× Female Lead Inventor −2.144∗∗ −3.288∗∗∗

(0.891) (1.155)

Surgery & Med Inst.× Female Lead Inventor −7.317∗∗∗ −16.319∗∗∗

(1.159) (2.204)

Transportation× Female Lead Inventor −0.866∗ 0.443

(0.438) (0.611)

Patent Subcategory (NBER) FE Yes Yes

Examiner x Customer FE Yes Yes
Examiner Art Unit FE Yes Yes
Patent Issue Year FE Yes Yes
Observations 69 436,458 436,458
Adjusted R2 0.299 0.283

Electronic copy available at: https://ssrn.com/abstract=4269703

 TABLE IA5: ROBUSTNESS TO SAMPLE SELECTION

This table establishes robustness of our baseline specification of Panel B of Table 3. Panel A uses a single-
author patent while Panel B uses both single-author patents and patents where all inventors share the
same gender. The sample covers patents issued from 1976-01-01 through 2021-12-31. Standard errors are
clustered at the patent customer and patent issue year level. ***, **, * denote significance at the 1%, 5%,
and 10% level, respectively. Data source: USPTO.

Panel A: Single Author

Delta in Forward Citations
(1) (2) (3) (4)
Lead Female Inventor −3.431∗∗∗ −3.376∗∗∗ −3.234∗∗∗ −2.764∗∗∗
(0.614) (0.598) (0.628) (0.823)

Intercept 1.456∗∗∗
(0.261)

Customer FE No No Yes No
Examiner FE No No Yes No
Examiner x Customer FE No No No Yes
Art Unit FE No Yes Yes Yes
Patent Issue Year FE No Yes Yes Yes
Observations 124,280 124,280 124,280 124,280
Adjusted R2 0.002 0.015 0.138 0.421

Panel B: Inventors Same Gender

Delta in Forward Citations
(1) (2) (3) (4)
Lead Female Inventor −4.042∗∗∗ −3.984∗∗∗ −3.931∗∗∗ −3.769∗∗∗
(0.338) (0.323) (0.296) (0.163)

Intercept 1.725∗∗∗
(0.239)

Customer FE No No Yes No
Examiner FE No No Yes No
Examiner x Customer FE No No No Yes
Art Unit FE No Yes Yes Yes
Patent Issue Year FE No Yes Yes Yes
Observations 129,298 129,298 129,298 129,298
Adjusted R2 0.003 0.017 0.211 0.466

70

Electronic copy available at: https://ssrn.com/abstract=4269703

 TABLE IA6: BIAS IN FORWARD CITATION, USING LONGFORMER

This table replaces the BERT model using Longformer to study the full text of patents. The sample covers
patents ranging from 4800000 through 5299999. These range roughly from 1986 until 1994. Standard
errors are clustered at the patent customer and patent issue year level. ***, **, * denote significance at the
1%, 5%, and 10% level, respectively. Data source: USPTO.

Delta in Forward Citations

(1) (2) (3) (4)
Lead Female Inventor −1.961∗∗∗ −2.108∗∗∗ −2.198∗∗∗ −1.993∗∗∗
(0.703) (0.548) (0.220) (0.201)

Intercept 4.767∗∗∗
(0.471)

Customer FE No No Yes No
Examiner FE No No Yes No
Examiner x Customer FE No No No Yes
Art Unit FE No Yes Yes Yes
Patent Issue Year FE No Yes Yes Yes
Observations 37,020 37,020 37,020 37,020
Adjusted R2 0.001 0.020 0.255 0.292

71

Electronic copy available at: https://ssrn.com/abstract=4269703

 TABLE IA7: PLACEBO, RANDOM SAMPLE

This table presents a placebo test by randomizing the gender of patents and re-running our C-BERT
approach to establish the effects are not an artifact of C-BERT. Panel A uses the number of forward
citations as the dependent variable, while Panel B uses bias computed using C-BERT. The sample covers
patents issued from 1976-01-01 through 2021-12-31. Standard errors are clustered at the patent customer
and patent issue year level. ***, **, * denote significance at the 1%, 5%, and 10% level, respectively. Data
source: USPTO.

Panel A: Forward Citations

Forward Citations
(1) (2) (3) (4)
Lead Female Inventor 0.022 −0.027 −0.066 −0.106
(0.154) (0.145) (0.106) (0.096)

Intercept 19.918∗∗∗
(3.146)

Customer FE No No Yes No
Examiner FE No No Yes No
Examiner x Customer FE No No No Yes
Art Unit FE No Yes Yes Yes
Patent Issue Year FE No Yes Yes Yes
Observations 471,461 471,461 471,461 471,461
Adjusted R2 −0.00000 0.109 0.202 0.304

Panel B: Delta in Forward Citations

Delta in Forward Citations
(1) (2) (3) (4)
Lead Female Inventor −0.042 −0.036 −0.058 −0.214
(0.202) (0.208) (0.212) (0.250)

Intercept −3.626∗∗∗
(0.537)

Customer FE No No Yes No
Examiner FE No No Yes No
Examiner x Customer FE No No No Yes
Art Unit FE No Yes Yes Yes
Patent Issue Year FE No Yes Yes Yes
Observations 471,461 471,461 471,461 471,461
Adjusted R2 −0.00000 0.025 0.051 0.221

72

Electronic copy available at: https://ssrn.com/abstract=4269703

 TABLE IA8: BIAS IN FORWARD CITATION, USING SCIBERT

This table replaces the BERT model with SciBERT. The sample covers patents issued from 1976-01-01
through 2021-12-31. Standard errors are clustered at the patent customer and patent issue year level. ***,
**, * denote significance at the 1%, 5%, and 10% level, respectively. Data source: USPTO.

Delta in Forward Citations

(1) (2) (3) (4)
Lead Female Inventor −7.524∗∗∗ −7.693∗∗∗ −7.584∗∗∗ −6.580∗∗∗
(1.536) (1.458) (1.514) (2.115)

Intercept 0.680∗∗∗
(0.141)

Customer FE No No Yes No
Examiner FE No No Yes No
Examiner x Customer FE No No No Yes
Art Unit FE No Yes Yes Yes
Patent Issue Year FE No Yes Yes Yes
Observations 602,974 602,974 602,974 602,974
Adjusted R2 0.012 0.033 0.064 0.195

73

Electronic copy available at: https://ssrn.com/abstract=4269703