Why do some patents get licensed while others do not?

Why do some patents get licensed while others
do not?
Karen Ruckman1,
* and Ian McCarthy2
1
Beedie School of Business, Simon Fraser University, 8888 University Avenue, Burnaby, British Columbia
v5a 1s6, Canada. email: ruckman@sfu.ca and 2
Beedie School of Business, Simon Fraser University, 8888
University Avenue, Burnaby, British Columbia v5a 1s6, Canada. email: imccarthy@sfu.ca
*Main author for correspondence.
Abstract
To understand why some patents get licensed and others do not, we estimate a portfolio of firm- and
patent-level determinants for why a particular licensor’s patent was licensed over all technologically
similar patents held by other licensors. Using data for licensed biopharmaceutical patents, we build a
set of alternate patents that could have been licensed-in using topic modeling techniques. This pro-
vides a more sophisticated way of controlling for patent characteristics and analyzing the attractive-
ness of a licensor and the characteristics of the patent itself. We find that patents owned by licensors
with technological prestige, experience at licensing, and combined technological depth and breadth
have a greater chance at being chosen by licensees. This suggests that a licensor’s standing and or-
ganizational learning rather than the quality of its patent alone influence the success of outward
licensing.
JEL classification: L24, L65
1. Introduction
The global licensing market was estimated to be worth $200 billion in 2011 (Alvarez and Lopez, 2015) and licensing
royalty payments grew at a rate of almost 10% between 1990 and 2003 (Athreye and Cantwell, 2007). At the same
time, however, numerous patents remain unlicensed (Gambardella et al., 2007) and many firms have difficulty find-
ing licensing partners (Zuniga and Guellec, 2008; Kani and Motohashi, 2012). Reconciliation of this puzzle is central
to the call to better understand why firms choose each other for a licensing agreement (Kim and Vonortas, 2006a;
Arora and Gambardella, 2010a).
To address this research opportunity, we examine the impact of eight licensor characteristics and four patent
characteristics on the ability of a patent owner to be identified and selected by a buyer. In varying and overlapping
ways, the firm-level determinants reflect the search and learning capabilities of the licensor, as well as its visibility
and attractiveness. The patent-level determinants control for differences in patent quality. Three of the firm-level de-
terminants (licensor prestige, licensor technological depth, and licensor technology breadth) are new to the licensing
research literature and allow us to examine how the licensor characteristics act as signals that help licensees to iden-
tify, evaluate, and choose a licensor and its patent over licensors with similar patents. While the other determinants
have been examined in previous studies, we confirm or refute previous findings using a more sophisticated empirical
method of identifying alternate patents.
VC The Author 2016. Published by Oxford University Press on behalf of Associazione ICC. All rights reserved.
Industrial and Corporate Change, 2016, 1–22
doi: 10.1093/icc/dtw046
Original article
Industrial and Corporate Change Advance Access published November 29, 2016
byguestonNovember30,2016http://icc.oxfordjournals.org/Downloadedfrom

Examining which licensor characteristics appeal to licensees is important because it answers calls to incorporate
the demand side of licensing into the analysis of licensing activities (Arora and Gambardella, 2010b; Ceccagnoli and
Jiang, 2013). The prior literature on licensing is largely from a supply perspective, i.e., what determines the propen-
sity to engage in licensing-out. This research has mixed results; finding that small licensors (Gambardella et al.,
2007) and both small and large licensors (Zuniga and Guellec, 2008) are more likely to out-license technologies. The
rate of outward licensing is also positively associated with very small and very large licensors (Motohashi, 2008), cur-
rent patent stock and past history in licensing (Kim and Vonortas, 2006b), and negatively related to large market
shares (Fosfuri, 2006) and the disclosure of too much or too little voluntary information (Wuyts and Dutta, 2008).
Our analysis sheds light on the licensor- and patent-level characteristics that are associated with increasing licensing
probability by fixing the licensee perspective on only those patents that are technologically similar.1
Our study makes an empirical contribution by building on previous licensing likelihood studies and addressing
the problem of which licensor–licensee pairs to include in the analysis. In previous licensing probability studies (e.g.,
Kim and Vonortas, 2006a; Kim, 2009), all known licensors were paired with all known licensees. Indeed, this is also
the case for alliance probability studies such as Diestre and Rajagopalan (2012) and Rothaermel and Boeker (2008)
where all known alliance partners are paired. This method conflates the possible technological motivations for poten-
tial licensing agreements, leading to the overemphasis of licensing determinants as they were compared to licensing
pairs that would have no feasible reason to license with each other. In our study, we use a text analytics approach
called topic modeling (Kaplan, 2012; Venugopalan and Rai, 2014) to identify alternative patents that could have
been licensed-in but were not by examining the text in each patent’s abstract. This technique enables us to identify re-
markably similar patents for each licensed patent allowing the analysis to disregard the technological motivation for
the licensing agreement and instead focus on the attractiveness of the patent owner and the patent itself. In doing so,
we are able to incorporate the demand side of the licensing transaction which previously had been under-represented
in licensing literature (Arora and Gambardella, 2010b).
We examine the determinants of licensing using 93 licensing agreements in the worldwide biopharmaceutical sec-
tor in a panel data set spanning 1993–2007. We find that licensors with strength in technological prestige, licensing
experience, and combined technological depth and breadth are more likely to license-out their patents. The results
also indicate that newer and highly cited patents are more likely to be licensed. Together these results offer a deeper
understanding of licensing patterns and reveal the licensor characteristics that licensees prefer. This is inherently use-
ful on a managerial level for firms looking to license-out as well as in-license. For instance, a licensor cannot control
the attributes of a potential licensee but it can understand which of its own attributes will appeal to licensees. These
attributes could be developed and promoted to help make a licensor more visible, viable, and ultimately attractive to
potential licensees.
2. Determinants of licensing-out probability
Technology licensing involves the transfer of intellectual property (IP) between two parties: the owner of the IP (the
licensor) and the buyer (the licensee). The licensor seeks to extract value from it in the form of licensing revenues and
may be motivated to license-out their technologies because they lack the financial, physical, or intellectual resources
to commercialize it. The propensity to license-out is associated with very small and very large licensors (Gambardella
et al., 2007; Motohashi, 2008; Zuniga and Guellec, 2008), but also current patent stock and past history in licensing
(Kim and Vonortas, 2006b). Licensees, on the other hand, typically acquire patented technologies as a means of di-
versifying their technological assets or satisfying a particular internal technological gap. The motivation to do so may
be inversely related to the licensee’s level of internal research. Ceccagnoli et al. (2010) found that firms with higher
RD productivity and more complementary resources will be less likely to externally source their technology.
Licensing is a process where both parties want to pair with the best possible partner. Once one looks beyond the
technological motivation for a particular licensing agreement, the choice of a particular licensor’s patent to license-in
over other technically similar patents can be a complicated decision involving the ease of identification of the patent
and the licensor, the attractiveness of the licensor’s technical prowess bargaining strength, and projections about the
1 It is important to note that this analysis is not a matching study or an experiment of the actual patents that the licensees
chose between. The comparison of the alternate patents to the actual licensed patent allow us to ascertain the charac-
teristics that are correlated with the licensee choice.
2 K. Ruckman and I. Mccarthy

licensor’s future knowledge transfer capabilities. We now introduce determinants for this licensing probability and
discuss how each one affects the ability of a licensor to get its patent licensed. We focus on three determinants (licen-
sor prestige, licensor technological depth, and licensor technology breadth) that have not been previously examined
in the licensing probability or propensity literatures. We argue that these are important for understanding how licen-
sor visibility, credibility, and learning impact the probability a licensor’s patent will be licensed over technologically
similar patents held by other licensors.
2.1 Licensor prestige
Licensor prestige is the extent to which a licensor is viewed favorably by all licensees. Drawing on prior research that
shows that institutional prestige increases a licensor’s rate of licensing (Sine et al., 2003), there are four reasons why
licensor prestige makes a licensor more attractive as a licensing partner.
First, patents owned by licensors that are thought to have high prestige will be more attractive to licensees due to
a “halo effect” (Thorndike, 1920, p. 25). This effect is a confirmation bias about a firm’s reputation, which allows
firms to better attract resources and opportunities (Roberts and Dowling, 2002). For licensors, their reputation is
based on the quality of its key productive resources including, for example, star scientists in biotechnology firms
(Audretsch and Stephan, 1996) and the quality of patent stocks (Alcacer Gittelman, 2006; Marco, 2007). When li-
censees perceive licensors to be strong at one or more of these research indicators they transfer this positive percep-
tion to the strategic and commercial attractiveness of the licensors and their technology. For example, research
shows that the halo effect enables universities with strong research standings to license-out more than their lower
standing research counterparts (Sine et al., 2003).
Second, prestige increases a licensor’s visibility, which is the extent to which potential licensees will know about a
licensor’s patents. This phenomenon is explained using Merton’s (1968) seminal work on the Matthew effect (i.e.,
the rich get richer), which found that scientist prestige heightens the visibility of these scientists over less prestigious
scientists with the same quality of scientific research. Similarly, research on mergers and acquisitions shows that the
higher the average quality of a firm’s patent stock (as measured by forward citations), the greater the interest in
acquiring the firm (Grimpe Hussinger, 2008). In the licensing context, we suggest that strong prestige enhances
the visibility of a licensor, thereby increasing the likelihood that potential licenses will be aware of the licensor
(Lewin, 1935; Granovetter, 1985; Sine et al., 2003).
Third, licensor prestige increases the legitimacy of the claims in a licensor’s patent. That is, the more respected is a
licensor, the more likely licensees will view the licensor’s patent and technological claims as being credible. The cred-
ibility afforded by a licensor’s strong prestige will make a patent more attractive for licensing purposes. Conversely,
licensees will more likely doubt technological claims from less prestigious licensors and be less enticed to complete a
licensing agreement. This credibility effect of prestige has been shown in the placement of researchers into job pos-
itions (Flanagin and Metzger, 2007). The more prestigious a researcher’s training institute, the higher the quality of
the institute the researcher is placed in.
Fourth, licensees prefer to transact with more prestigious licensors as it increases their own prestige (Sine et al.,
2003). Individuals, teams, and firms can all build their own prestige by association, and in turn increase their stand-
ing, visibility, and credibility. For example, to compensate for liabilities of newness, start-up ventures often seek to
borrow reputation by association by forming a partnership with prestigious venture capital firms (Larson and Starr,
1993). This effect applied to licensing would mean prestigious licensors are more likely to attract potential licensees
interested in licensing their technologies and the benefits of association in doing so.
In summary, prestigious licensors will have a greater chance of licensing-out because licensees are more likely to
know about and be attracted to them due to the increases in licensor standing, visibility, credibility, and the benefits
by association (Sine et al., 2003).
2.2 Licensor technological depth
Licensor technological depth is the extent to which a firm has specialized in an area of technological knowledge and
is associated with post-contractual knowledge transfer and the ability of the licensor to convey knowledge effectively
to the licensee. Drawing upon the organizational learning literature, a firm’s depth of technological knowledge in the
licensed patent’s technological area is positively linked to knowledge transfer (Wu and Shanley, 2009) and particu-
larly in the biopharmaceutical industry, where expertise in disciplinary areas is crucial (Henderson and Cockburn,
Why some patents get licensed 3

1994). Technological depth acts as a signal to others of their technical expertise in a particular technical area (Arora
and Gambardella, 1990). Firms with strong technological depth are attractive to other firms in need of supplement-
ing their own knowledge through partnerships (Stuart, 1998; Baum et al., 2000) and will result in more firms willing
to partner with them (Ahuja, 2000).
Technological depth, like prestige, acts like a beacon, helping licensors to be identified. While searching patent
databases, licensees will encounter various patents in technological areas similar to those that are sought. Licensors
with technological depth in the sought area will be more likely to be encountered. Thus, greater technological depth
or expertise helps make licensors more noticed, noted, and selected by licensees. It is also likely that licensees per-
forming local searches (i.e., search for knowledge that is closely related to a firm’s preexisting knowledge base) may
already be familiar with the prospective licensor (Katila and Ahuja, 2002, Laursen, 2012). The patents, therefore,
owned by licensors with strong technological depth will likely be more known to prospective licensees than patents
owned by licensors with less technological depth.
2.3 Licensor technological breadth
Technological breadth refers to the variety or scope of technological knowledge areas a firm has explored (Wang
and von Tunzelmann, 2000; Katila and Ahuja, 2002; Lodh and Battaggion, 2015). As breadth of technological
knowledge has been found to be an important component of organizational learning (Zahra and George, 2002),
we assert that it is also a quality that makes licensors attractive to licensees. A licensor with a broad technological
knowledge base is better able to disseminate its technology to external parties. Cohen and Levinthal (1990: 136)
acknowledge that some fraction of a firm’s knowledge “must be fairly diverse to permit effective, creative utiliza-
tion of the new knowledge” by another firm. Similarly, Lane and Lubatkin’s (1998) research found that alliance
partners’ knowledge bases must be different in some way for firms to fully take advantage of learning opportuni-
ties. We argue that licensor technological breadth helps facilitate knowledge transfer to the licensee through the ef-
ficient transmission of knowledge, making the licensor more attractive for licensing agreements. This is consistent
with both Contractor (1981) and Teece (1977) who emphasized that suppliers’ knowledge transfer capability is an
important factor in lowering buyers’ integration costs and Ceccagnoli and Jiang (2013) who found that licensors
with strong technological breadth can more expertly communicate relevant knowledge to the licensee. This
endows licensors with a knowledge transfer ability that makes the patents owned by those licensors more likely to
be in-licensed by licensees.
2.4 Other established licensor determinants
Previous research on licensing propensity has focused on five other established firm-level determinants. As high-
lighted earlier, this literature has not examined these determinants by controlling with a set of technological equiva-
lent patents that could have been licensed. As this is the approach we take in this article, we can corroborate extant
results for these five determinants. We now introduce each one briefly here.
Licensor size is a simple indicator of a firm’s complementary assets. As we focus on biotechnology patents, firms
with more downstream assets (e.g., drug manufacturing facilities) are expected to be less inclined to license-out. This
determinant has been well established in its negative influence on firm licensing activities (Gambardella et al., 2007;
Kani and Motohashi, 2012) and participation in alliance agreements (Ahuja, 2000; Colombo et al., 2006). However,
in contrast, Kim (2009) and Kim and Vonortas (2006a) argue that large licensors are more likely to license-out to
better exploit their technologies.
A licensor’s experience at licensing is the second recognized determinant that we consider. Generally, firms ac-
quire knowledge, learn, develop capabilities, and get better at doing things with experience (Dodgson, 1993). In the
context of licensing, experience allows licensors to learn from past licensing decisions and apply that knowledge to
future choices. Licensor’s past history of licensing transactions (experience) has been shown to positively affect licens-
ing likelihood (Kim and Vonortas, 2006a,b; Kim, 2009).
The third established determinant, research intensity, is a proxy for the concentration of a biopharmaceutical
firm’s total RD inputs to the innovation process (Cohen and Levinthal, 1990). In any technology-based firm, the
amount of internal RD is expected to promote collaboration (Katila and Mang, 2003) and influence a firm’s motiv-
ation to license-out (Lane and Lubatkin, 1998; Zhang and Baden-Fuller, 2010). Licensor research intensity is ex-
pected to be a positive influence on licensing likelihood (Kani and Motohashi, 2012).

Research age is the fourth established determinant we consider. It is the length of time a licensor has been actively
researching and is a proxy for the licensor’s legitimacy, power, and standing. For biopharmaceutical firms it has been
found that there is a greater chance of alliance participation when partners are technologically similar (Rothaermel
and Boeker, 2008).
The fifth established determinant is the past relationships between a licensor and licensee. The greater the past re-
lationships, the more familiar a licensor and licensee should be with each other. This is expected to greatly reduce the
transaction costs associated with licensing and has been found to increase licensing likelihood (Kim and Vonortas,
2006a; Kim, 2009).
3. Empirical section
3.1 Model
Following Ruckman (2005) and Colombo et al. (2006), the model of the licensee’s choice decision is:
yÃ
ijpt ¼ V½b0
xijpt þ eijptŠ where
b0
xijpt ¼ b0
1xit þ b0
2xjt þ b0
3xpt
(1)
yÃ
ijpt is the unobserved variable that captures licensee i’s decision to enter into a licensing agreement at time t of patent
P owned by potential licensor j. The vector of explanatory variables (xijptÞ includes licensee-, licensor-, and patent-
specific characteristics. The choice to license patent P occurs when the valuation of the agreement exceeds all others
in the same choice set P. The valuation is determined at the time of the licensing agreement and thus is calculated as
the present value of all future revenue streams arising from the agreement less costs (Kim and Vonortas, 2006b). The
observed outcome (yijptÞ is defined as follows:
yijpt ¼ 1 if V b
0
xijpt þ eijpt
h i
V b
0
xijp0t þ eijp0t
h i
8p0
2 P; 0 otherwise: (2)
All agreements in this model are exclusive contracts between one licensee and one licensor reflecting accepted prac-
tice in the biopharmaceutical sector. The model structure has each known licensee choosing a patent to be licensed
among all technologically similar and available patents at the time t of the agreement. As such, the pairings are licen-
see–patent pairs, not licensee–licensor pairs.
The model fixes the licensee decision across all alternative patents. The approach, methodology, and the inter-
pretation of the results differ from existing licensing probability studies (Kim and Vonortas, 2006a; Kim, 2009)
which pair together all known licensors and licensees to form a matrix of potential dyads. The firm-level attri-
butes of the successful dyads in those studies are interpreted as contributing to the likelihood of an agreement
compared to all other potential agreements, many of which would involve technologies owned by licensors that
the licensees would have no interest in acquiring. This results in the over-emphasis and mis-interpretation of firm-
and patent-level characteristics contributing to the likelihood of an agreement. For instance, suppose the group of
patents that were being chosen between all had similar licensor characteristics (e.g., prestige). In that case, the pre-
vious technique would have inaccurately found those characteristics to be significant because it would have com-
pared the chosen patents to all other patents which would have a variance of those characteristics. Our analysis,
on the other hand, would find those determinants to be insignificant because they were not defining features of
the chosen patent compared to its alternatives. In our model, the patents in each choice set are technologically
close substitutes, making the potential licensee–patent pairings in our model realistic for the technological needs
of each licensee. Effectively, this allows a careful examination of the demand for licensing because any licensee
motivations to do with economies of scope or strategic fit of a particular technology are already accounted for. As
such, the patent owner and patent attributes for each patent are observed by the licensee in each choice group and
those attributes are interpreted to be attractive to the licensee or unattractive solely on the basis of those
attributes.
As the dependent variable is binary, the estimation will be in the logistic functional form, however, as there are
separate choice groups for each known agreement, there are two choices for the estimation method: fixed effects
method (called conditional logit) or random effects method. Conditional logit uses fixed effects to account for each

choice group2
and, in doing so, all the variables that do not vary within a group drop out of the regression
(Ruckman, 2005; Palomeras, 2007). In our model, this would be all the licensee-specific characteristics. If random ef-
fects for each choice group are used instead of fixed effects, all group-unvarying variables are retained but will intro-
duce bias in estimates (Allison, 2009; Bell and Jones, 2015). The Hausman test is a common method of determining
the best model and it does not reject the hypothesis that random effects are a consistent and efficient estimator of the
true parameters (Chi-squared (12 d.f.) ¼ 15.6, P-value ¼ 0.21), indicating that neither model is preferred over the
other. However, Clark et al., 2015 find that fixed effects is the preferred model over random effects for data sets
with within group standard deviation around 0.2 (ours is 0.18 while the standard is 1.0). We, therefore, present our
analysis using conditional logit as the default model; however, we will also show results using random effects for
comparison and robustness.
3.2 Identification of alternative patents
To identify the patents that are similar to the licensed patent, we employ text analytics which has been used recently
in the context of patent classification (Bergmann et al., 2008; Wu and Huang, 2010; Gerken and Moehrle, 2012).
Firms externally source technology sometimes to fill a particular internal technological gap (Higgins and Rodriguez,
2006; Danzon, Epstein and Nicholson, 2007) and it would be logical that these firms would consider various patents
that fit the same technological profile and satisfy the same technological need. Patents are assigned classification
codes which aim to categorize the technology and group together similar patents. However, the patent classification
codes are very broad and subjective due to variations in judgments of the patent examiners (Venugopalan and Rai,
2014) leading any data based on them to be inaccurate (Dahlin and Behrens, 2005). We follow methods in Kaplan
(2012) and Venugopalan and Rai (2014) which use the text analytics sub-branch called “topic modelling” to narrow
down a broad list of patents to those that are most similar based on the text and terms found in their abstracts.
The topic modeling technique is a powerful and innovative way of dealing with large textual databases, enabling re-
searchers to effectively sift through large sets of text to identify meaningful similarities (Thomas et al., 2014). This ap-
proach has two advantages. First, it allows us to identify a very specific set of alternative patents that could have been
licensed but were not. This enables us to narrow down the potential licensors (i.e., the supply side firms) to be only
those that own the patents in a choice set.3
Second, the topic modelling technique and the resulting conditional logistic
analysis are performed at the patent-level. This makes it possible to control for patent-level characteristics that contrib-
ute to the likelihood of an agreement, especially those associated with quality (Leone and Reichstein, 2012).
Topic model algorithms observe the collection of patents and reverse engineer the process by identifying a set of
topics and a distribution of the patents over the topics that could have generated these patents. We used the
MALLET toolkit (McCallum, 2002) to identify 500 distinct topics across the entire pool of patents4
and the
Kullback–Leibler distance measure (Bigi, 2003; Pinto et al., 2007) to determine the similarity of the distribution of
each patents’ topics to each licensed patent. Each choice set is composed of the 19 most similar patents to the licensed
patent plus the licensed patent itself.5
The topic modeling technique is described in detail in Appendix A.
There is a time-based limit to the value of patents leading licensees to discount old patents and instead consider
relatively recent patents for possible in-licensing (Gans et al., 2007). Therefore, the set of patents that are available to
license is further limited to those that were granted during the 5 years before the actual licensing agreement.
Furthermore, we limit the set to only those patents that are owned by firms that have out-licensed one of their patents
at least once before the time of the licensing agreement,6
indicating that the licensor is an active out-licensor and that
the other patents possessed by the firm are likely candidates for out-licensing as well.
2 Conditional logit directly accounts for heterogeneity between choice sets and latent individual licensee characteristics
that could affect the licensee’s valuation of a particular patent (like economies of scope or product gap).
3 That is not to say that the patents that the topic modeling technique identifies were the actual patents that the licensees
considered for an agreement. This analysis is not an experiment and was performed entirely ex poste.
4 We offer robustness checks using 200 topics in Table A1.
5 Three choice sets had less than 20 patents in total. The rest had substantially more than 20. There are robustness
checks on choice group sizes of 10 and 30 in Table A1.
6 This technique is similar to the grouping of licensing firms performed in Kani and Motohashi (2012). Those that have
licensed out are considered to be active licensors and it is assumed that all of their patents are likewise willing to be
out-licensed. Table A1 shows a robustness check that disregards potential licensors’ licensing history.

The appropriateness of a binary choice to model licensing agreements depends in large part on the constraint of
the choice by the focal firm (Mindruta et al., 2016). There are two criteria for the applicability of a discrete choice
model over a matching model. First, licensees cannot face competition for preferred partners. Of our 93 agreements,
four of them have alternate patents that were licensed by other firms. In addition, all four of these patents were
licensed in separate choice groups, suggesting that the little competition that existed was very dilute. Second, licensees
must have all the contracting power in a potential partnership. Our approach to our research problem assumes that
licensees predominantly have the contracting power and choose across licensors. Previous alliance research in the bio-
pharmaceutical industry suggests this is highly plausible (Rothaermel and Boeker, 2008; Diestre and Rajagopalan,
2012). Licensors are often smaller biotechnology firms focused on developing technologies to be licensed, while the
licensees tend to be larger pharmaceutical firms that in addition to developing their own technologies will acquire
technologies that they then manufacture and commercialize. These differences in firm size and scope produce a con-
tracting power disparity where large licensees more commonly choose among several potential smaller licensors, un-
like the opposite (Gambardella and Panico, 2014). To ensure the power dynamic is in the hands of the licensees, we
have excluded licensors whose size is larger than the licensee in their choice group in our default results.7
In addition,
we explore whether the results differ when the contracting power increases for licensees by presenting results for
when the licensors’ sizes are less than 75% and 50% the size of their licensees. Given that the vast majority of the li-
censees do not face rivalry constraints in choosing their preferred patent and that the power dynamic has been
ensured to be with the licensees, a discrete choice analysis will be an appropriate analytical tool.
3.3 Data
Our data are for licensing agreements in the biopharmaceutical industry and were gathered from the RECAP data-
base by Deloitte. RECAP is acknowledged to be the most complete database of the world’s biopharmaceutical part-
nerships (Schilling, 2009) and it has been used as a source of data in recent studies (Kollmer and Dowling, 2004;
Wuyts and Dutta, 2008). An initial search of non-university8
licensing agreements involving patents between 1993
and 2008 yielded 297 patents. After data matching across three databases, the default sample was reduced to 93
agreements.9
The RECAP database also provided information about the past relationship history of each partner.
Accounting data (revenues and RD expenditures) for the firms were gathered using Compustat and filings with the
US Securities and Exchange Commission. The data used to create the patent-based variables came from the NBER
US patent citation data file (Hall et al., 2001) and its 2006 update. For patents granted after 2001, we performed dir-
ect searches on the US Patent and Trademark Office patent database to compile accurate forward citations (for 5
years after their granting year).
The statistics of variables are in Table 1 are separated into summaries for all the patents in the choice sets (n ¼
1848) and the licensed patents themselves (n ¼ 93). Interestingly, on average, it appears that successful licensors have
lower depth, lower breadth, and lower experience than the pool of licensors. The correlations of variables are found
in two tables. Table 2a displays the correlation for all patents in all choice sets and Table 2b for just the licensed pa-
tents. The correlations for the successful licensors do not appear to be remarkably different than the pool of licensors.
We will now discuss each variable’s measurement.
Dependent variable: Chosen is a binary variable measured as 1 if patent is chosen by licensee for a licensing agree-
ment, 0 otherwise.
Licensor variables: Technological prestige reflects the licensor’s reputation for high-quality technologies. It is
measured as a licensor’s average number of non-self forward citations on patent stock within 5 years of each patent’s
granting. In as much as forward citations are correlated with value and importance of the technology (Harhoff et al.,
2003; Lanjouw and Schankerman, 2004; Ceccagnoli et al., 2010), average firm-level forward citations reflect the
7 A robustness check in Table A1 includes all licensors, regardless of their sizes relative to their licensees.
8 Agreements involving university licensors were not included in the data set as university technology transfer offices
have differing structures with different motivations (Markman, Phan, Balkin and Gianiodis, 2005) which may not emulate
the other profit-driven firms in the sample.
9 The resulting data set limits the licensors in terms of company size (very small firms with no financial history are
dropped), patenting history (licensors must have patented at least once), and licensing history (licensors much have
had out-licensing history to be included). The average values of all variables before and after data arrangement are not
significantly different: all but two variables are statistically equal by Welch’s t-test.

research reputation and technological legitimacy of a licensor. It is a firm-level characteristic that is particular to the
licensing sector but could be applied to other technology-based but non-patenting industries by measuring references
in industry reviews or technology awards. Leone and Reichstein (2012) use a similar measurement as a control vari-
able when determining the speed of licensee post-contractual invention but do not explore it deeply or hypothesized
about it. Kani and Motohashi (2012) also use the same measurement as a reflection of the potential demand for a li-
censor’s technology, although they do not find that it is significantly related to licensing propensity.
Depth is the number of patents granted to a firm in same International Patent Categories (IPC) as the licensed pa-
tent during the 5 years before the agreement, logged. This variable indicates extent of a firm’s existing knowledge in
a particular technological area before a potential licensing agreement takes place and is similar to variables used in
Leone and Reichstein (2012) (“unfamiliarity”) and Sakakibara (2010) (“proximity to licensed patent technology”).
Table 1. Statistics of variables
Variable Measure Mean Std. Dev. Min Max
Dependent variable:
1 Chosen 1 if patent is chosen by licensee for a
licensing agreement, 0 otherwise
0.05 0.2 0 1
Licensor variables:
2 Licensor technological
prestige
Licensor’s average number of non-self
forward citations on patent stock
within 5 years after patent granted
All patents
Licensed patents
4.5 6.5 0 63.3
8.9 15.2 0 63.3
3 Licensor depth Number of patents granted to licensor in
same IPC as licensed patent during
5 years before agreement, logged
All patents 2.3 1.4 0 6.3
Licensed patents 0.4 0.7 0 3.3
4 Licensor breadth Number of different IPC classes in licen-
sor patents that were granted within
5 years before agreement
All patents 8.1 10.3 0 115
Licensed patents 2.2 2.5 0 10
5 Licensor size Licensor revenues in millions of $US,
logged
All patents
Licensed patents
4.2 3.1 –6.2 10.6
2.1 3.3 –6.2 9.7
6 Licensor experience Number of license agreements during
5 years before agreement involving
the licensor
All patents 15.1 14.9 0 94
7 Licensor research
intensity
Licensor RD expenditure divided by
licensor revenue
All patents
Licensed patents
14.7 128.3 0 4146.8
21.5 58.7 0.01 389.2
8 Licensor research age Year of agreement minus year that licen-
sor ﬁrst applied for a patent
All patents 10.6 6.8 0 31
9 Past relationships Number of relationships between licen-
see and licensor during 5 years before
agreement
All patents 0.4 3.4 0 57
Patent variables:
10 Patent complexity Number of technological claims made by
the patent
All patents
Licensed patents
18.9 18.5 1 201
24.3 21.3 1 104
11 Patent age Year of agreement minus year patent
was granted
All patents
Licensed patents
2.6 1.4 1 5
1.2 0.6 1 4
12 Patent scope Number of IPC categories listed on the
patent
All patents
Licensed patents
2.5 1.5 1 17
2.3 1.4 1 7
13 Patent citations Number of backwards citations listed on
the patent
All patents 11.4 20.9 0 211

Breadth is the technological scope of a firm’s past patenting efforts. The broader a firm’s patent stock, the more
diverse is its own internal research spectrum and the better it is able to convey specific knowledge to third-parties. It
is measured as the number of different IPC classes the firm patented in during the 5 years before the agreement. This
particular measurement of technological breadth follows Zhang and Baden-Fuller (2010) and Leone and Reichstein’s
(2012) “technological diversity” variable.10
Drawing on prior studies on the motivations and capability to license, Size is measured as firm revenues in mil-
lions of $US (logged); Experience is the number of licensing agreements a licensor has been involved with during the
5 years before agreement. As it has been shown that firms use the same internal division of experts to facilitate both
inward and outward licensing agreements, the experience variable combines historical experience in both types of
agreements. Lowe and Taylor (1998) found that firms that are proficient at inward licensing are often proficient in
outward and vice versa, due to shared resources. We also measure Research intensity as research and development
(RD) expenditure divided by revenues. As this variable is broad in its measurement, its impact in this analysis is
limited. Research age is measured as the year of the licensing agreement minus the year that the licensor first applied
for a patent (Leone and Reichstein, 2012); Past relationships is the number of relationships between the licensor and
Table 2b. Correlation of Variables, licensed patents (n ¼ 93)
Variable 2 3 4 5 6 7 8 9 10 11 12 13
2 Licensor technological prestige 1
3 Licensor depth 0.28 1
4 Licensor breadth 0.28 0.49 1
5 Licensor size 0.34 0.37 0.28 1
6 Licensor experience 0.3 0.16 0.19 0.68 1
7 Licensor research intensity –0.15 –0.16 –0.21 –0.54 –0.16 1
8 Licensor research age –0.15 –0.09 –0.01 –0.08 –0.13 –0.01 1
9 Past relationships 0.12 –0.09 –0.07 0.03 –0.02 –0.07 0.17 1
10 Patent complexity 0.12 –0.13 –0.12 0.09 0.08 –0.04 0.21 0.13 1
11 Patent age 0.14 0.28 0.14 0.31 0.28 –0.1 0.16 0.06 –0.01 1
12 Patent scope 0.03 0.14 –0.04 0.04 0.15 –0.01 0.03 –0.08 –0.01 0.12 1
13 Patent citations 0.52 0.07 0.06 –0.03 –0.14 0.01 –0.18 –0.06 0.18 –0.05 –0.16 1
Table 2a. Correlation of variables, all patents (n ¼ 1848)
Variable 1 2 3 4 5 6 7 8 9 10 11 12 13
1 Chosen 1
2 Licensor technological prestige –0.16 1
3 Licensor depth –0.32 –0.15 1
4 Licensor breadth –0.13 –0.07 0.24 1
5 Licensor size –0.16 –0.06 0.32 0.47 1
6 Licensor experience –0.08 –0.04 0.3 0.18 0.63 1
7 Licensor research intensity 0.01 0.01 –0.06 –0.05 –0.25 –0.08 1
8 Licensor research age 0.04 –0.14 0.19 0.11 0.37 0.29 –0.04 1
9 Past relationships –0.01 –0.02 –0.004 0.01 0.14 0.22 –0.01 0.02 1
10 Patent complexity 0.07 0.14 0.04 –0.03 –0.10 –0.09 0.01 –0.07 –0.02 1
11 Patent age –0.23 –0.08 0.11 –0.01 0.12 0.09 –0.01 0.26 –0.01 –0.05 1
12 Patent scope –0.02 –0.04 0.01 –0.02 –0.02 0.06 0.02 –0.04 –0.02 0.06 –0.01 1
13 Patent citations 0.18 0.28 –0.01 –0.06 –0.14 –0.13 0.01 –0.09 –0.02 0.17 –0.09 –0.04 1
10 We also perform a robustness check where this variable is measured as (1 – Herfindahl index of the IPC’s that a firm
has patented in). The results (not shown) are very similar.

licensee during the 5 years before agreement (Vanhaverbeke et al., 2002). The types of relationships are broadly
defined and can include licensing, alliances, or co-development agreements.
Patent variables: Patent complexity is the number of technological claims made by the patent (Lanjouw, 1999;
Palomeras, 2007; Sakakibara, 2010). Patent age is measured as the year of licensing agreement minus the year the pa-
tent was granted (Sakakibara, 2010) and is an indicator of remaining patent life. Patent scope is the number of IPCs
listed on the patent (Gambardella et al., 2007; Sakakibara, 2010). Patent citations is the number of backward cit-
ations listed on the patent (Palomeras, 2007) which captures the innovativeness of the technology.
3.4 Results
Table 3 reports the conditional logit regression results for the licensee choice of a patent to in-license. Given the non-
linear nature of the dependent variable (Ai and Norton, 2003), the interactions were interpreted as multiplicative ra-
ther than additive (Buis, 2010). Thus, odds ratios are reported which is equivalent to the exponent of the coefficients
of the respective variables.11
For instance, the estimated odds ratio for licensor size in Column 1 of Table 3 is 0.78.
This is interpreted as follows: as licensor size increases by 1 unit ($10 million), the likelihood of an agreement de-
creases by 22%.
Columns 1–3 in Table 3 display the results for the licensor explanatory, licensor control, and patent variables, re-
spectively. Column 4 shows the regression with all the variables included and Column 5 includes an interaction term
of licensor depth and breadth. The results indicate that licensors with strong technological prestige are more likely to
be identified and have their patents chosen for a licensing agreement. A licensor’s research reputation and techno-
logical legitimacy can act as a beacon to increase the chances that its patent will be chosen by a licensee. Consistent
with Kim, 2009; Kani and Motohashi, 2012; Kim and Vonortas, 2006b, patents owned by licensors that have exten-
sive licensing experience are more likely to be licensed. The identification and communication afforded by experi-
enced licensing teams makes a licensor more attractive for a licensing partnership.
Unexpectedly, patents owned by licensors with strong knowledge depth in the patent’s technological area or with
strong knowledge breadth were less likely to be chosen for a licensing agreement. This suggests that licensors with
strong knowledge depth or breadth are unattractive to potential licensees when compared to other licensors with
similar technology and less depth or breadth. Although these results were unexpected, it bears noting that knowledge
depth (when measured as general patent stock) has had inconsistent results in the alliance literature. Colombo et al.
(2006) and Rothaermel and Boeker (2008) found patent stock to be positively related to alliance formation, while
Stern, Dukerich and Zajac (2014) found it to be the opposite. The search literature may offer an explanation for the
negative result found in this study. Katila and Ahuja (2002) found that too much depth will negatively impact inno-
vativeness because building on the same knowledge is subject to diminishing returns, and reusing existing knowledge
can make an organization rigid and become trapped in an existing technological competency. They also found that
too much breadth (or “search scope”) reduces innovativeness because the firm’s ability to respond to new informa-
tion correctly reduces and eventually the costs of integration exceed the benefits of acquiring new knowledge
(Laursen and Salter, 2006). It appears that these findings are mirrored in our study in that licensees are less likely to
choose patents owned by licensors with strong knowledge depth or breadth (on their own) perhaps due to concerns
about their internal rigidity and thin spread of research resources.
In an attempt to better understand the perplexing results for licensor depth and breadth, we add an interaction
term composed of the two variables into the analysis. We find that patents owned by licensors with both depth and
breadth are more likely to be chosen for a licensing agreement. Using the results in Column 5, the negative effect of li-
censor breadth (À23%) is reversed by 9% when paired with a licensee whose average number of citations in the
same IPC as the licensed patent increased by 10. These results are consistent with results in the search literature
(Katila and Ahuja, 2002) which found that firm depth and breadth are positively related to firm innovativeness
through the resulting knowledge uniqueness and absorptive capacity afforded by the mutual reciprocation of the two
combined characteristics. In the same way that licensor depth and breadth increase organizational learning and firm
innovativeness, they would also combine to increase a licensor’s ability to license-out. Technological depth and
technological breadth interact with each other to produce a combined balance of knowledge that suits exploratory
and problem-solving technology transfer situations (Katila and Ahuja, 2002). This complementary interaction allows
11 For similar empirical approaches see: Herrmann, Kundisch and Rahman (2014) and Foss, Lyngsie, and Zahra (2013).

Table3.Conditionallogitregressions
VariablesLicensorssmallerthanlicenseesa
Licensorssmallerthan
75%sizeoflicenseeb
Licensorssmallerthan
50%sizeoflicenseec
Column1Column2Column3Column4Column5Column6Column7
Licensortechnologicalprestige1.07(0.016)***1.06(0.024)**1.06(0.025)***1.04(0.034)***1.08(0.048)*
Licensordepth0.14(0.035)***0.14(0.044)***0.10(0.036)***0.08(0.040)***0.02(0.014)***
Licensorbreadth0.73(0.050)***0.83(0.071)**0.77(0.079)**0.74(0.091)**0.73(0.094)**
Licensorbreadth*Licensordepth1.09(0.043)**1.16(0.075)**1.30(0.083)***
Licensorsize0.56(0.039)***0.85(0.092)0.84(0.094)0.75(0.105)**0.72(0.128)*
Licensorexperience1.02(0.015)1.04(0.021)**1.04(0.020)**0.94(0.042)1.02(0.049)
Licensorresearchintensity0.99(0.002)**0.99(0.005)0.99(0.006)0.99(0.007)1.00(0.009)
Licensorresearchage1.10(0.020)***1.11(0.031)***1.11(0.032)***1.15(0.039)***1.18(0.050)***
Pastrelationships0.94(0.131)0.97(0.131)0.98(0.130)1.02(0.258)2.57(1.306)*
Patentcomplexity1.01(0.005)1.01(0.011)1.01(0.011)1.01(0.013)0.98(0.016)
Patentage0.22(0.045)***0.30(0.081)***0.31(0.082)***0.11(0.055)***0.09(0.058)***
Patentscope0.94(0.065)0.89(0.101)0.88(0.102)0.78(0.132)0.65(0.143)**
Patentcitations1.02(0.005)***1.02(0.007)***1.02(0.008)***1.02(0.010)**1.03(0.009)***
Statistics
Log–Likelihood:–115.2–220.4–192.0–65.1–63.7–47.1–34.6
WaldChi-squared:325.5(3d.f.)***115.1(5d.f.)***171.8(4d.f.)***425.7(12d.f.)***428.4(13d.f.)***349.2(13d.f.)***318.9(13d.f.)***
PseudoR-squared:0.590.210.310.770.770.790.82
Numberofchoicegroups:93939393937769
Numberofobservations:1848184818481848184814521276
***indicatessigniﬁcanceat1%,**5%,*10%.Oddsratiosreported.
a:Meansizeoflicensees:7.8,meansizeofpotentiallicensors:4.3,meansizeofchosenlicensors:2.1.
b:Meansizeoflicensees:7.9,meansizeofpotentiallicensors:2.9,meansizeofchosenlicensors:1.1.
c:Meansizeoflicensees:8.0,meansizeofpotentiallicensors:1.8,meansizeofchosenlicensors:0.7.

firms to create new, unique combinations of knowledge that can be developed and eventually commercialized
(Winter, 1984: 293). Thus, the mutual reciprocity afforded by licensor technological depth and breadth make them
attractive to licensees who wish to learn from a licensor. In sum, licensor technological depth and breadth combine
to produce an ability where the licensor has a balance of both deep and broad knowledge bases that help to expertly
and efficiently convey ideas.
The results indicate that patents owned by licensors that have been patenting for a long time (consistent with
Ceccagnoli and Jiang, 2013) are more likely to be licensed. We find that newer patents with more patent life remain-
ing have a higher chance of being licensed which is in direct contradiction of Ceccagnoli et al. (2010). Perhaps more
surprisingly, we do not find any significant connection between two parties having a past relationship and the likeli-
hood of a new agreement. This goes against the familiarity principle that has been empirically supported in licensing
studies (Kim and Vonortas, 2006a; Kim, 2009) as well as alliance studies (Diestre and Rajagopalan, 2012, Stern
et al., 2014). The contradictory results are perhaps explained by the difference in methodologies, where in the previ-
ously used pairwise probability method, some partner characteristics in successful alliances were artificially elevated
in significance as they were being compared to a large number of unrealistic alliance partners. In terms of patent-
level characteristics, we find that those patents with more extensive backwards citations, reflecting innovativeness
(consistent with Palomeras, 2007), are more likely to be licensed. Gambardella et al. (2007) found that patents with
complexity and scope are more likely to be licensed; however, we find no such significant relationship.
The impact of licensor size on the likelihood of a licensing agreement has had inconsistent results in the literature.
Kim (2009) and Kim and Vonortas (2006a) found that larger licensor size leads to higher chances of an agreement,
while Gambardella et al. (2007) and Kani and Motohashi (2012) find the opposite. At first, it may appear that our re-
sults uphold a negative relationship between licensor size and licensing likelihood; however, since our data set has
been selected on the basis of the size relationship between licensee and licensor, we cannot draw too much conclusion
from these results. However, the inconsistent results in the literature suggest a complex relationship that we choose
to examine more closely. Columns 6 and 7 replicate the model on a smaller sample size where a licensee size relative
to potential licensors in each choice group gets progressively larger. Increasing the relative power position of the li-
censee over its prospective licensors ensures that licensees face little constraints in their decisions and that a discrete
choice model is the appropriate analysis tool for our context (Mindruta et al., 2016). It will also allow us to detect
any changes in the attractiveness of licensor attributes as the licensee gains more negotiation strength. Column 5 (and
all the other results report in this study unless explicitly stated otherwise) involve licensors smaller in size than the li-
censee in their choice group. Column 6 includes only licensors that are smaller than 75% of the size of the licensees
in their choice group and Column 7 are for 50% the size. Licensor research age, the combination of licensor breadth
and depth and patent age and backwards citations largely retain their appeal as licensee size relative to licensor size
increase. Surprisingly, as the size of licensees increase relative to its potential licensors, licensees prefer to license pa-
tents owned by increasingly smaller firms and licensees do not appear to find licensor experience or technological
prestige as attractive as they did when the licensors were larger. The past relationship between licensor and licensee
weakly becomes a determinant of licensing likelihood and licensees increasingly prefer the scope of the patent to be
narrow. To sum up, as the size-power dynamic increase in the favor of licensees, they prefer to choose patents and li-
censors that are technologically focused with few patents in their area, yet have been researching for a long time. It
appears that when licensees are arguably unrestricted with patent choice, they look for patents that will satisfy one
particular technical need that is owned by a small but familiar licensor, thus reducing the transaction costs associated
with negotiations and probably subsequent payouts.
We offer the model estimated using random effects instead of the fixed effects used in conditional logit as a ro-
bustness check in Table 4.12
One of the advantages of the random effects model is that it easily allows for multiple
layers of random effects. Columns 1 and 2 include random effects for each choice group and Column 3 adds a layer
of random effects on each individual patent in each choice set to account for latent characteristics such as patent
quality. The results are robust to the conditional logit results found in Table 3 and remain remarkably consistent,
indicating the lack of latent heterogeneity in the data set.
12 Although a random effects model allows us to include choice-group-specific variables (such as licensee characteris-
tics), we choose to estimate the same variables in the condition logit analysis to be consistent and comparable across
both results. Also, the licensee variables do not vary enough across observations to make them reliably estimated.

4. Discussion section
In this article, we explored what determines the likelihood a licensee will choose to in-license a specific patent over
other similar patents. From the perspective of licensing propensity, our arguments and results add to prior research
focused on how licensor motivations (Arora and Fosfuri, 2003; Fosfuri, 2006; Gambardella et al., 2007; Motohashi,
2008) and licensor capabilities impact their tendency to license-out (Kim and Vonortas, 2006b; Motohashi, 2008;
Wuyts and Dutta, 2008; Kani and Motohashi, 2012). By considering the licensor characteristics that lead licensees to
choose certain licensors, we make a contribution that also incorporates the demand side of licensing, a perspective
that few researchers have studied. We now explain how the findings from our study help scholars and managers bet-
ter understand why some patents are selected for a licensing agreement.
4.1 Implications for research
Our findings add to the economics- and capability-based arguments as to why licensors and their patents get chosen.
We demonstrate support for the view that licensor prestige increases the likelihood a licensor and its patent will be
chosen over other licensors with similar patents. This suggests that licensors emit reputational signals to licensees and
that these create what organizational theorists call a “status-based model” of selection (Podolny, 1993). In the con-
text of technology licensing, this implies that the market for patented technologies is not just a matter of predicting
future revenue streams. It is also influenced by a licensor’s standing in the market, as derived by quality of their prior
research outputs. This is interesting as the finding suggests that the standing or prestige of the licensor rather than the
quality of its patent alone influences its success in licensing-out.
A second major implication of our study is the provision of an organizational learning explanation for
licensing-out success. Our results provide support for the view that experience enhances a licensor’s ability to
license-out. This contributes to and extends the concept that licensing-out is a capability where effective licensors
accumulate the specific knowledge, people, and routines required to succeed at licensing (Fosfuri, 2006).
However, we suggest these resources and their advantages are not only important for being able to license-out but
they also play a role in the demand side of licensing by helping licensors to be identified and evaluated by licensees.
Table 4. Mixed logit regressions
Variables Column 1 Column 2 Column 3
Licensor technological prestige 1.03 (0.015) ** 1.04 (0.015) *** 1.04 (0.016) **
Licensor depth 0.19 (0.039) *** 0.12 (0.032) *** 0.12 (0.028) ***
Licensor breadth 0.75 (0.046) *** 0.70 (0.054) *** 0.70 (0.052) ***
Licensor breadth * Licensee depth 1.12 (0.034) *** 1.12 (0.032) ***
Licensor size 0.85 (0.065) ** 0.86 (0.065) * 0.86 (0.068) *
Licensor experience 1.05 (0.017) *** 1.05 (0.017) *** 1.05 (0.017) ***
Licensor research intensity 0.99 (0.002) 0.99 (0.002) 0.99 (0.002)
Licensor research age 1.12 (0.025) *** 1.13 (0.026) *** 1.13 (0.027) ***
Past relationships 0.92 (0.155) 0.92 (0.074) 0.92 (0.151)
Patent complexity 1.00 (0.008) 1.00 (0.009) 1.00 (0.009)
Patent age 0.23 (0.055) *** 0.22 (0.055) *** 0.22 (0.051) ***
Patent scope 0.98 (0.092) 0.98 (0.095) 0.98 (0.097)
Patent citations 1.02 (0.006) *** 1.02 (0.006) *** 1.02 (0.006) ***
Constant 2.01 (1.021) 2.62 (1.395) * 2.61 (1.432) *
Statistics
Log-likelihood: –145.5 –142.7 –145.7
Wald Chi-squared: 131.1 (12 d.f.) *** 118.4 (13 d.f.) *** 167.1 (13 d.f.) ***
Random effects: choice sets yes yes yes
Random effects: patents no no yes
Odds ratios reported.
***indicates signiﬁcance at 1%, **5%, *10%.
Number of observations: 1848. Number of choice groups: 93.

Third, our study has theoretical implications for understanding how technological search and technology transfer
impact the technology transfer element of a licensor’s ability to license-out. There is a well-established optimal licens-
ing contract literature (Gallini and Wright, 1990; Beggs, 1992) of which some studies are concerned with how know-
ledge transfer affects the terms in a licensing contract (Arora, 1995; Macho-Stadler et al, 1996; Hegde, 2014). These
studies focus on the knowledge transfer terms in the agreement contracts and do not consider, to any extent, the char-
acteristics of the firms (other than basic size, previous history, and age). Our finding that licensor technological
breadth and depth combine to create an environment for effective knowledge transfer contributes to this literature. It
shows how this potential knowledge transfer would affect the likelihood of an agreement occurring in the first place.
This is an important implication as firms can vary significantly in technological depth and technological breadth, and
thus being able to appropriately balance depth and breadth can be vital to licensing success.
Finally, our empirical methodology is also an important contribution. We group each licensee with the patent it
licensed and used topic modeling techniques to identify similar licensable patents. This allows for tighter and more
controlled analysis than previous models which enlisted a simple binary analysis of the pairings of all known licensees
and licensors without regard to technological motivation (e.g., see Kim and Vonortas, 2006a, in the licensing context
and Colombo et al., 2006, in the alliance context). By identifying technologically similar patents, we are able to in-
corporate the demand side of the licensing transaction which previously had been under-represented in licensing lit-
erature (Arora and Gambardella, 2010b). Furthermore, the use of a new methodology encourages us to examine the
significance of variables previously thought to influence licensing probability using prior methodologies. In particu-
lar, we could not confirm that a previous relationship between a licensor and licensee would contribute to their
licensing likelihood, despite the findings from previous studies (Kim and Vonortas, 2006a; Kim, 2009). We did, how-
ever, confirm existing evidence that licensors with long patenting history (Ceccagnoli and Jiang, 2013) and patents
with extensive backward citations (Palomeras, 2007) are more likely to be licensed.
4.2 Implications for practice
A better understanding of why a particular patent is licensed over other technologically similar patents also has im-
portant implications for managers. Our findings suggest a few situations when managers’ attention to their own char-
acteristics can impact licensing-out success. First, building and maintaining licensor prestige can be a source of
licensing-out advantage for licensors. This is because licensees are not just making licensing decisions based on qual-
ity of the patented technology, but also on the standing of and technological fit with the licensor. Given that the costs
and rewards of licensing can be significant, this finding highlights the importance for licensors to consider the latent
value that their prior innovation activities signal to prospective licensees. In addition to producing innovations to en-
hance a product portfolio and compete in markets, the quantity and quality of the associated research resources can
help to attract future licensing partners.
Second, licensing experience increases a licensor’s ability to license-out. Learning-by-doing builds over time and a
history of licensing allows firms to acquire knowledge and capabilities that can be utilized in future licensing transac-
tions. It may also be possible that proactive licensing-out also involves licensors making themselves known to firms
who are active in an appropriate technology field (Solinas, 2015). Driven by such strategies, experienced licensors
will have more resources and expertise dedicated to outward licensing, which increases their ability to be known to
and selected by potential licensees. As the market for technology becomes more extensive, competitive, and litigious,
the costs of searching for, evaluating, and acquiring technologies are also increasing. An awareness of the efficiencies
that can be gained via experience, will help smooth the trade in technology.
Finally, balancing licensor technological depth and breadth is important for licensing-out likelihood. Individually,
high levels of the two qualities are found to make licensees less likely to choose a licensor’s patent for an agreement.
Depth is usually thought to be an attractive quality of an alliance partner (Ahuja, 2000); however, in the licensing
context, depth reflects research rigidity and evidence of reusing existing knowledge. Similarly, breadth on its own
demonstrates a licensor’s reduced reliability from spreading its research resources too thin. The results suggest that li-
censees anticipate the lack of innovative output which occurs at high levels of depth and breadth individually (Katila
and Ahuja, 2002) and choose licensing partners to avoid these issues. That said, the combination of licensor depth
and breadth together is an attractive quality to licensees. Licensor technological depth and breadth balance the know-
ledge processes for technology transfer transactions and reduce the negative effects of the individual depth and
breadth qualities. The two qualities combined act as an attractive signal to prospective licensees that a licensor

possesses the efficient ability to convey and transmit technical ideas. The post-transaction knowledge transfer is a
crucial activity between the two licensing parties (Ceccagnoli and Jiang, 2013) and the attractiveness of both depth
and breadth together highlights that licensees are keenly aware of the potential for knowledge transfer when making
their licensing choice.
4.3 Limitations and future research
The demand focus of our study and model might imply that licensing is solely about licensees choosing patents but,
in reality, licensors choose licensees as much as licensees choose licensors. Although we account for this in the struc-
ture of the model, the empirical analysis does not allow for direct verification of it. The formation of a choice set of
potential licensees and potential licensor which eventually lead to one particular agreement would be exceedingly dif-
ficult.13
It would be challenging to identify a realistic set of licensees that were interested ex ante in one particular
type of technology; the type that was eventually actually licensed.
The determinants were tested on one single industry which is beneficial in that it inherently avoids the complica-
tion of between industry differences that could affect partner choice; however, it also lacks empirical verification in
other industries. Future research could address this gap, especially focusing on other technology-based industries like
chemicals and software. Furthermore, technology licensing is just one mechanism by which firms transfer external
knowledge. Consequently, our arguments and approach could also be applied to other forms of technology transfer
including the sale of technology (Caviggioli Ughetto, 2012), inter-firm alliances (Eisenhardt Schoonhoven,
1996; Gulati et al., 2009), the acquisition of firms (Hayward, 2002; Nadolska Barkema, 2014) and the role of pa-
tent intermediaries: organizations that facilitate the exchange of IP (Hagiu Yoffie, 2013).
Another limitation of this study is that it does not examine the subsequent success of the licensing deal. The mere
attainment of a licensing agreement is only one factor of success in the licensing process. Future research could follow
Sakakibara (2010), who made a first step in this direction when she investigated the effect of individual characteris-
tics of licensors and licensees on royalty rates and lump sum payments made to licensors.
5. Conclusion
In this article, we set out to understand why some patents get licensed while many others do not. Our fundamental
conclusion is that beyond the technological motivation to license-in a patent and the quality of the patent, there are
firm-level qualities that can distinguish a licensor from others and these can influence the likelihood that a particular
patent will be licensed. At a minimum this incorporation of firm-level status and learning effects allude to the import-
ance of sociological mechanisms on licensing probability.
References
Arora, Ashish and Alfonso Gambardella (1990), ‘Complementarity and external linkages: the strategies of the large firms in biotech-
nology,’ The Journal of Industrial Economics, 361–379.
Ahuja, G. (2000), ‘The duality of collaboration: inducements and opportunities in the formation of interfirm linkages,’ Strategic
Management Journal, 21, 317–343.
Ai, C. and E. Norton (2003), ‘Interaction terms in logit and probit models,’ Economics Letters, 80(1), 123–129.
Alcacer, J. and M. Gittelman (2006), ‘Patent citations as a measure of knowledge flows: the influence of examiner citations,’ Review
of Economics and Statistics, 88(4), 774–779.
Allison, P. D. (2009), Fixed Effects Regression Models, Vol. 160. SAGE publications.
Alvarez, R. and R. A. Lopez (2015), ‘Foreign technology acquisition and changes in the real exchange rate,’ The World Economy,
38(4), 613–628.
Arora, A. (1995), ‘Licensing tacit knowledge: intellectual property rights and the market for know-how,’ Economics of Innovation
and New Technology, 4(1), 41–60.
Arora, A. and A. Fosfuri (2003), ‘Licensing the market for technology,’ Journal of Economic Behavior and Organization, 52(2),
277–295.
13 For a different approach to this problem using a matching model, see Mindruta et al. (2014). The pair-wise matching al-
gorithm works in the context of licensing but would not address the objective of determining the characteristics of the
licensor and patents which are successful compared to other feasible matches.

Arora, A. and A. Gambardella (2010a), ‘The Market for Technology,’ in B. Hall and N. Rosenberg (eds), Handbook of the
Economics of Innovation, Vol. 1. North Holland, pp. 641–678.
Arora, A. and A. Gambardella (2010b), ‘Ideas for rent: an overview of markets for technology,’ Industrial and Corporate Change,
19(3), 775–803.
Athreye, S. and J. Cantwell (2007), ‘Creating competition? Globalisation and the emergence of new technology producers,’ Research
Policy, 36 (2), 209–226.
Audretsch, D. and P. Stephan (1996), ‘Company-scientist locational links: the case of biotechnology,’ American Economic Review,
86 (3), 641–652.
Baum, J. A. C., T. Calabrese and B. Silverman (2000), ‘Don’t go it alone: alliance network composition and startups’ performance in
Canadian biotechnology,’ Strategic Management Journal, 21, 267–294.
Beggs, A. (1992), ‘The licensing of patents under asymmetric information,’ International Journal of Industrial Organization, 10,
171–191.
Bell, A. and K. Jones (2015), ‘Explaining fixed effects: random effects modeling of time-series cross-sectional and panel data,’
Political Science Research and Methods, 3(01), 133–153.
Bergmann, I., D. Butzke, L. Walter, J. Fuerste, M. Moehrle and V. Erdmann (2008), ‘Evaluating the risk of patent infringement by
means of semantic patent analysis: the case of DNA chips,’ RD Management, 38 (5), 550–562.
Bigi, B. (2003), Using Kullback-Leibler Distance for Text Categorization. Springer Berlin Heidelberg: Berlin.
Bigi, B., Y. Huang and R. D. Mori (2003), ‘Vocabulary and language model adaptation using information retrieval,’ in Proceedings
of the ECIR-2003, Vol. 2633. Springer-Verlag, pp. 305–319.
Blei, D. (2012), ‘Probabilistic topic models,’ Communications of the ACM, 55(4), 77–84.
Blei, D., A. Ng and M. Jordan (2003), ‘Latent dirichlet allocation,’ The Journal of Machine Learning Research, 3, 993–1022.
Buis, M. (2010), ‘Stata tip 87: interpertation of interactions in non-linear models,’ The Stata Journal, 10(2), 305–308.
Carpineto, C., R. D. Mori, G. Romano and B. Bigi (2001), ‘An information-theoretic approach to automatic query expansion,’ ACM
Transactions on Information Systems, 19(1), 1–27.
Caves, R., H. Crookell and J. Killing (1983), ‘The imperfect market for technology licenses,’ Oxford Bulletin of Economics and
Statistics, 45(3), 249–267.
Caviggioli, F. and E. Ughetto (2013), ‘The drivers of patent transactions: corporate views on the market for patents,’ RD
Management, 43(4), 318–332. Available at SSRN 2144107. Retrieved from http://papers.ssrn.com/sol3/papers.cfm?abstract_
id¼2144107.
Ceccagnoli, M., S. Graham, M. Higgins and J. Lee (2010), ‘Productivity and the role of complementary assets in firms demand for
technology innovations,’ Industrial and Corporate Change, 19(3), 839–869.
Ceccagnoli, M. and L. Jiang (2013), ‘The cost of integrating external technologies: supply and demand drivers of value creation in the
markets for technology,’ Strategic Management Journal, 34, 404–425.
Clark, Tom S. and Drew A. Linzer (2015), ‘Should I use fixed or random effects?’ Political Science Research and Methods, 3(2),
399–408.
Cohen, W. and D. Levinthal (1990), ‘Absorptive capacity: a new perspective on learning and innovation,’ Administrative Science
Quarterly, 35(1), 128–152.
Colombo, M., L. Grilli and E. Piva (2006), ‘In search of complementary assets: the determinants of alliance formation of high-tech
start-ups,’ Research Policy, 35, 1166–1199.
Contractor, F. (1981), International Technology Licensing: Compensation, Costs, and Negotiation. Lexington Books: Lexington, MA.
Dagan, I., L. Lee and F. Pereira (1999), ‘Similarity-based models of word cooccurrence probabilities,’ Machine Learning, 34(1),
43–69.
Dahlin, K. and D. Behrens (2005), ‘When is an invention really radical?: defining and measuring technological radicalness,’ Research
Policy, 34, 717–737.
Danzon, P., A. Epstein and S. Nicholson (2007), ‘Mergers and acquisitions in the pharmaceutical and biotech industries,’ Managerial
and Decision Economics, 28, 307–328.
Diestre, L. and N. Rajagopalan (2012), ‘Area all ‘sharks’ dangerous? New biopharmaceutical ventures and partner selection in RD
alliances,’ Strategic Management Journal, 33, 1115–1134.
Dodgson, M. (1993), ‘Organizational learning: a review of some literatures,’ Organization Studies, 14(3), 375–394.
Eisenhardt, K. and C. Schoonhoven (1996), ‘Resource-based view of strategic alliance formation: strategic and social effects in entre-
preneurial firms,’ Organization Science, 7(2), 136–150.
Flanagin, A. J. and M. J. Metzger (2007), ‘The role of site features, user attributes, and information verification behaviors on the per-
ceived credibility of web-based information,’ New Media and Society, 9(2), 319–342.
Fosfuri, A. (2006), ‘The licensing dilemma: understanding the determinants of the rate of technology licensing,’ Strategic
Foss, N., J. Lyngsie and S. Zahra (2013), ‘The role of external knowledge sources and organizational design in the process of oppor-
tunity exploitation,’ Strategic Management Journal, 34(12), 1453–1471.

Gallini, N. and B. Wright (1990), ‘Technology transfer under asymmetric information,’ Journal of Economics, 21(1), 147–160.
Gambardella, A., P. Giuri and A. Luzzi (2007), ‘The market for patents in Europe,’ Research Policy, 36, 1163–1183.
Gans, Joshua S., David H. Hsu and Scott Stern (2008), ‘The impact of uncertain intellectual property rights on the market for ideas:
Evidence from patent grant delays,’ Management Science, 54(5), 982–997.
Gambardella, Alfonso and Claudio Panico (2014), ‘On the management of open innovation,’ Research Policy, 43(5), 903–913.
Gerken, J. and M. Moehrle (2012), ‘A new instrument for technology monitoring: novelty in patents measured by semantic patent
analysis,’ Scientometrics, 91, 645–670.
Granovetter, M. (1985), ‘Economic action and social structure: the problem of embeddedness,’ American Journal of Sociology, 91,
481–510.
Grimpe, C. and K. Hussinger (2008), ‘Pre-empting technology competition through firm acquisitions,’ Economics Letters, 100(2),
189–191.
Gulati, R., D. Lavie and H. Singh (2009), ‘The nature of partnering experience and the gains from alliances,’ Strategic Management
Journal, 30(11), 1213–1233.
Hagiu, A. and D. Yoffie (2013), ‘The new patent intermediaries: platforms, defensive aggregators, and super-aggregators,’ Journal of
Economic Perspectives, 27(1), 45–65.
Hall, Bronwyn H., Adam B. Jaffe and Manuel Trajtenberg (2001), The NBER patent citation data file: Lessons, insights and methodo-
logical tools. No. w8498. National Bureau of Economic Research.
Harhoff, D., M. Scherer and K. Vopel (2003), ‘Citations, family size, opposition and the value of patent rights,’ Research Policy, 32,
1343–1363.
Hayward, M. (2002), ‘When do firms learn from their acquisition experience? Evidence from 1990 to 1995,’ Strategic Management
Journal, 23(1), 21–39.
Hegde, D. (2014), ‘Tacit knowledge and the structure of license contracts: evidence from the biomedical industry,’ Journal of
Economics and Management Strategy, 23(3), 568–600.
Henderson, R. and I. Cockburn. (1994), ‘Measuring competence? Exploring firm effects in pharmaceutical research,’ Strategic
Herrmann, P., D. Kundisch and M. Rahman (2014), ‘Beating irrationality: does delegating to IT alleviate the sunk cost effect?’
Management Science Articles in Advance, 61, 831–850.
Higgins, M. J. and D. Rodriguez (2006), ‘The outsourcing of RD through acquisition in the pharmaceutical industry,’ Journal of
Financial Economics, 80, 351–383.
Kani, M. and K. Motohashi (2012), ‘Understanding the technology market for patents: new insights from a licensing survey of
Japanese firms,’ Research Policy, 41, 226–235.
Kaplan, S. (2012), ‘Identifying breakthroughs: using topic modeling to distinguish the cognitive from the economic,’ Academy of
Management Proceedings, 2012(1), 1–1.
Katila, R. and G. Ahuja (2002), ‘Something old, something new: a longitudinal study of search behavior and new product introduc-
tion,’ Academy of Management Journal, 45(6), 1183–1194.
Katila, Riitta and Paul Y Mang (2003), ‘Exploiting technological opportunities: the timing of collaborations,’ Research policy, 32(2),
317–332.
Kim, Y. (2009), ‘Choosing between international technology licensing partners: an empirical analysis of U.S. biopharmaceutical
firms,’ Journal of Engineering and Technology Management, 26, 57–72.
Kim, Y. and N. Vonortas (2006a), ‘Technology licensing partners,’ Journal of Economics and Business, 58(4), 273–289.
Kim, Y. and N. Vonortas (2006b), ‘Determinants of technology licensing: the case of licensors,’ Managerial and Decision Economics,
27, 235–249.
Kollmer, H. and M. Dowling (2004), ‘Licensing as a commercialisation strategy for new technology-based firms,’ Research Policy,
33(8), 1141–1151.
Lane, P. and M. Lubatkin (1998), ‘Relative absorptive capacity and interorganizational learning,’ Strategic Management Journal, 19,
461–477.
Lanjouw, Jean O (1999), ‘Stylized Facts of Patent Ligation,’.
Lanjouw, J. and M. Schankerman (2004), ‘Patent quality and research productivity: measuring innovation with multiple indicators,’
Economic Journal, 114, 441–465.
Larson, P. and J. Starr (1993), ‘A network model of organizational formation,’ Entrepreneurship: Theory and Practice, 17(2), 5–16.
Laursen, K. (2012), ‘Keep searching and you’ll find: what do we know about variety creation through firms’ search activities for in-
novation,’ Industrial and Corporate Change, 21(5), 1181–1220.
Laursen, K. and A. Salter (2006), ‘Open for innovation: the role of openness in explaining innovation performance among UK manu-
facturing firms,’ Strategic Management Journal, 27(2), 131–150.
Leone, M. and T. Reichstein (2012), ‘Licensing-In Fosters Rapid Invention! The effect of the grant-back clause and technological un-
familiarity,’ Strategic Management Journal, 33, 965–985.

Lewin, K. (1935), Dynamic Theory of Personality. McGraw-Hill: New York.
Lodh, S. and M. R. Battaggion (2015), ‘Technological breadth and depth of knowledge in innovation: the role of mergers and acquisi-
tions in biotech,’ Industrial and Corporate Change, 24 (2), 383–415.
Lowe, J. and P. Taylor (1998), ‘R D and technology purchase through licence agreements: complementary strategies and comple-
mentary assets,’ R D Management, 28(4), 263–278.
Macho-Stadler, I., X. Martinez-Giralt and J. Perez-Castrillo (1996), ‘The role of information in licensing contract design,’ Research
Policy, 25(1), 43–57.
Marco, A. (2007), ‘The dynamics of patent citations,’ Economics Letters, 94(2), 290–296.
Markman, G., P. Phan, D. Balkin and P. Gianiodis (2005), ‘Entrepreneurship and university-based technology transfer,’ Journal of
Business Venturing, 20, 241–263.
McCallum, A. (2002), MALLET: a machine learning for language toolkit. http://mallet.cs.umass.edu.
Merton, R. (1968), ‘The Matthew effect in science,’ Science, 159, 56–63.
Mindruta, D., M. Moeen and R. Agarwal (2016), ‘A two-sided matching approach for partner selection and assessing complementar-
ities in partners’ attributes in inter-firm alliances,’ Strategic Management Journal, 37 (1), 206–231.
Motohashi, K. (2008), ‘Licensing or not licensing? An empirical analysis of the strategic use of patents by Japanese firms,’ Research
Policy, 37, 1548–1555.
Nadolska, A. and H. Barkema (2014), ‘Good learners: how top management teams affect the success and frequency of acquisitions,’
Strategic Management Journal, 35, 1483–1507.
Palomeras, N. (2007), ‘An analysis of pure-revenue technology licensing,’ Journal of Economics and Management Strategy, 16(4),
971–994.
Pinto, D., J. -M. Benedı and P. Rosso (2007), ‘Clustering narrow-domain short texts by using the Kullback-Leibler distance,’
Computational Linguistics and Intelligent Text Processing. Springer Berlin Heidelberg, pp. 611–622.
Podolny, Joel M A (1993), ‘Status-based model of market competition,,’ American journal of sociology, 829–872.
Roberts, Peter W. and Grahame R. Dowling (2002), ‘Corporate reputation and sustained superior financial performance,’ Strategic
management journal, 23 (12), 1077–1093.
Rothaermel, F. and W. Boeker (2008), ‘Old technology meets new technology: complementarities, similarities and alliance forma-
tion,’ Strategic Management Journal, 29, 47–77.
Ruckman, K. (2005), ‘Technology sourcing through acquisitions: evidence from the US drug industry,’ Journal of International
Business Studies, 36, 89–103.
Sakakibara, M. (2010), ‘An empirical analysis of pricing in patent licensing contracts,’ Industrial and Corporate Change, 19(3),
927–945.
Schilling, M. (2009), ‘Understanding the alliance data,’ Strategic Management Journal, 30(3), 233–260.
Sine, W., S. Shane and D. Gregorio (2003), ‘The halo effect and technology licensing: the influence of institutional prestige on the
licensing of university inventions,’ Management Science, 49(4), 478–496.
Solinas, G. (2015), ‘Licensing teams: an organizational approach to technology licensing proces,’ Economics, 27(4), 235–249.
Stern, I., J. M. Dukerich and E. Zajac (2014), ‘Unmixed signals: how reputation and status affect alliance formation,’ Strategic
Management Journal, 35(4), 512–531.
Stuart, T. E. (1998), ‘Network positions and propensities to collaborate: an investigation of strategic alliance formation in a high-
technology industry,’ Administrative Science Quarterly, 43(3), 668–698.
Teece, D. (1977), ‘Technology transfer by multinational firms: the resource cost of transferring technological know-how,’ Economic
Journal, 87, 242–261.
Thomas, S., B. Adams, A. Hassan and D. Blostein (2014), ‘Studying software evolution using topic models,’ Science of Computer
Programming, 80, 457–479.
Thorndike, E. L. (1920), ‘A constant error in psychological ratings,’ Journal of Applied Psychology, 4, 25–29.
Vanhaverbeke, W., G. Duysters and N. Noorderhaven (2002), ‘External technology sourcing through alliances or acquisitions: an
analysis of the application-specific integrated circuits industry,’ Organization Science, 13(6), 714–733.
Venugopalan, S. and V. Rai (2014), ‘Topic based classification and pattern identification in patents,’ Technological Forecasting and
Social Change, 94, 236–250.
Wang, Q. and N. von Tunzelmann (2000), ‘Complexity and the functions of the firm: breadth and depth,’ Research Policy, 29(7–8),
805–818.
Winter, S. (1984), ‘Schumpeterian competition in alternative technological regimes,’ Journal of Economic Behavior and
Organization, 5, 287–320.
Wu, C. -H. and T. Huang (2010), ‘Patent classification system using a new hybrid genetic algorithm support vector machine,’ Applied
Soft Computing, 10 (4), 1164–1177.
Wu, J. and M. Shanley (2009), ‘Knowledge stock, exploration, and innovation: research on the United States electromedical device in-
dustry,’ Journal of Business Research, 62, 474–483.

Wuyts, S. and S. Dutta (2008), ‘Licensing exchange—insights from the biopharmaceutical industry,’ International Journal of
Research in Marketing, 25, 273–281.
Zahra, S. and G. George (2002), ‘Absorptive capacity: a review, reconceptualization, and extension,’ Academy of Management
Review, 27(2), 185–203.
Zhang, J. and C. Baden-Fuller (2010), ‘The influence of technological knowledge base and organizational structure on technology col-
laboration,’ Journal of Management Studies, 47(4), 679–704.
Zuniga, M. P. and D. Guellec (2008), Survey on Patent Licensing: Initial Results from Europe and Japan. OECD: Paris.
Appendix A: Robustness checks
Table A1 reports seven robustness checks on the model. Each changes one aspect of the full model found in Column
5 in Table 3. The results remain remarkably robust across all robustness checks. Columns 1 and 2 change the num-
ber of patents in the choice groups to 10 (Column 1) or 30 (Column 2). Although there is a change in the number of
observations, the results remain fairly robust. Column 3 disregards potential licensor out-licensing history by ex-
panding the pool of available patents to those owned by all potential licensors, even those without any licensing ex-
perience. The relaxation of this assumption inflates the choice set of available patents to include those owned by
potential licensors who had not licensed out any of its patents in the previous 5 years. There is no way to discern
whether the patents owned by those licensors were intended for out-licensing or not. Column 4 does not require the
size of a potential licensor to be smaller than the licensee it its choice group. It appears that licensees prefer smaller
licensors when licensor size is unrestricted. Column 5 used the 200 topic model to produce the patents in the choice
groups.
Column 6 changes the choice sets to include only the patents owned by the chosen licensor. This includes all pa-
tents, regardless of the technological area, but only those that were applied for within 5 years before the agreement
date. As the firm-level licensor characteristics do not vary between groups, they drop out of the regression. This is a
robustness check deliberately crafted to test whether licensees were choosing licensors first and then picking the best
of their patents. Column 7 changes the choice sets to include all available biotechnology patents (identified by the
four-digit IPC classification used to initially form the pool of patents in the topic modeling technique outlined in
Appendix B). This robustness check is intended to confirm whether for the filtering procedure used to identify the
technology, similar patents for each licensed patent is more effective than simply using patent-level characteristics
which is why only the patent-level characteristics are included. Again, only those patents that were applied for within
the 5 years preceding the agreement were included. A comparison of the log-Likelihoods and the R-squared statistics
to those for Column 1 of Table 3 confirm that the topic modeling filtering approach adds value and that the model
where licensees first choose patents in a particular technical area before licensors is superior.
Appendix B: Topic modeling
Topic modeling was used in this study to narrow down the choice sets of patents. It was necessary because the IPC
associated with patents are too broad and subjectively determined by patent examiners (Venugopalan and Rai,
2014). Text analytics offers a more sophisticated method of classifying patents by comparing patents’ texts (Wu and
Huang, 2010; Gerken and Moehrle, 2012). Recently, Kaplan (2012) and Venugopalan and Rai (2014) applied the
sub-branch “topic modelling” to patent classification.
Topic modeling enables us to organize and summarize electronic archives at a scale that would be impossible by
human annotation (Blei, 2012). The topic modeling method observes the words in a set of “documents” (in our case,
patent abstracts) and seeks to find the topics that could have generated the observed documents (Venugopalan and
Rai, 2014). The topic modeling approach used in our study is the simplest one: latent Dirichlet Allocation (LDA).14
LDA determines the meanings of words by looking at co-presence with other words in the same document or block
of text (Blei, Ng Jordan, 2003). The same word may have different meanings depending on its association with
other words in a document (Kaplan, 2012). Topic models surmise a set of topics and identify the words associated
with each topic. As each document comprises different topics, the topic modeling algorithm additionally infers the
proportion of each topic that constitutes the document (Venugopalan and Rai, 2014).
14 For more formal information about the LDA model, see Appendix B in Venugopalan and Rai (2014), Blei, Ng and Jordan
(2003), or Blei (2012).

TableA1.Conditionallogitregressions,robustnesschecks
VariablesColumn1Column2Column3Column4Column5Column6Column7
10patentsineach
choicegroup
30patentsineach
choicegroup
Licensorsdonothave
tohaveprevious
out-licensing
history
Licensorsnot
restricted
bysizerelative
tolicensee
Choicesets
madeusing
200topicmodel
Choicesetsareall
patentsownedby
chosenlicensor
Choicesetsare
allbiotech
patentsavailable
Licensortechnologicalprestige1.08(0.034)**1.07(0.023)***1.06(0.023)***1.09(0.026)***1.09(0.033)***
Licensordepth0.11(0.047)***0.10(0.034)***0.14(0.042)***0.26(0.062)***0.08(0.031)***
Licensorbreadth0.73(0.094)**0.78(0.067)***0.76(0.053)***0.73(0.075)***0.77(0.078)**
Licensorbreadth*Licensordepth1.09(0.052)**1.08(0.038)**1.01(0.028)*1.06(0.027)**1.09(0.036)**
Licensorsize0.94(0.139)*0.81(0.077)**0.86(0.066)**0.75(0.077)***0.95(0.116)
Licensorexperience1.03(0.026)1.05(0.018)***1.05(0.020)***1.03(0.018)1.01(0.021)
Licensorresearchintensity1.00(0.006)1.00(0.005)1.00(0.003)1.00(0.006)1.00(0.007)
Licensorresearchage1.11(0.050)***1.14(0.031)***1.11(0.025)***1.11(0.031)***1.11(0.034)***
Pastrelationships0.94(0.542)0.98(0.143)1.08(0.111)1.01(0.092)1.03(0.184)
Patentcomplexity1.00(0.014)1.01(0.010)1.00(0.009)1.00(0.009)1.01(0.011)0.99(0.006)1.01(0.002)***
Patentage0.35(0.098)***0.26(0.071)***0.24(0.062)***0.29(0.082)***0.24(0.074)***0.66(0.049)***0.79(0.045)***
Patentscope0.86(0.128)0.89(0.098)0.97(0.100)0.88(0.101)0.92(0.128)1.02(0.084)1.20(0.053)***
Patentcitations1.02(0.009)***1.02(0.006)***1.02(0.006)***1.02(0.007)***1.02(0.008)***1.02(0.005)***1.01(0.001)***
Statistics
Log-likelihood:–45.4–78.9–55.2–72.4–56.6–283.9–1576.5
WaldChi-squared:337.5(13d.f.)***468.5(13d.f.)***500.7(13d.f.)***465.3(13d.f.)***442.6(13d.f.)***74.3(4d.f.)***69.6(4d.f.)***
pseudoR-squared:0.790.750.770.760.790.110.02
Numberofobservations:9302715251320301848609214,843,384
Numberofchoicegroups:939312810293117140
***indicatessigniﬁcanceat1%,**5%,*10%.Oddsratiosreported.
PotentiallicensorsaresmallerthanlicenseeinchoicesetforallcolumnsexceptColumns4,6,and7.

The first step in this method was identifying the relevant patents for which the topic modeling approach will be
applied. The pool of patents had to be from the broadest technological area and yet still pertain to biotechnology.
We noted the four-digit IPC of the licensed patents and then gathered all the patents in the NBER database in those
same categories (A01N, A61B, A61F, A61K, A61M, A61N, A61P, B05B, C07C, C07D, C07H, C07K, C12N, C12P,
C12Q, F21V, G01N, Q61K). There were over 90,000 patents that were granted between 1988 (5 years before first
agreement in our data set) and 2007 in these 18 categories.
The abstracts for all the relevant patents were assembled and the topic modeling software (MALLET) identified an
initial set of topics.15
The top words associated with the initial set of topics showed a number of “stop words” (such
as “the,” “or,” “and”) which were removed in subsequent analyses. MALLET requires that we pre-set the number of
topics in the database. We experimented with 100, 200, and 500 topics. After doing some careful comparison of
Table B1. Abstracts of sample patents
Patent
number
Patent title Patent abstract % topic 60 % topic 112
Licensed
patent
6623926 Methods for producing
5’-nucleic acid-protein
conjugates
Disclosed herein is a method for generating a 5’-
nucleic acid-protein conjugate, the method
involving: (a) providing a nucleic acid which
carries a reactive group at its 5’ end; (b) pro-
viding a non-derivatized protein; and (c) con-
tacting the nucleic acid and the protein under
conditions which allow the reactive group to
react with the N-terminus of the protein,
thereby forming a 5’-nucleic acid-protein con-
jugate. Also disclosed herein are 5’-nucleic
acid-protein conjugates and methods for their
use.
51 27
Most
similar
patent
6916632 Methods and reagents
for molecular cloning
The present invention provides compositions,
methods, and kits for covalently linking nu-
cleic acid molecules. The methods include a
strand invasion step, and the compositions
and kits are useful for performing such meth-
ods. For example, a method of covalently
linking double stranded (ds) nucleic acid mol-
ecules can include contacting a first ds nucleic
acid molecule, which has a topoisomerase
linked to a 3’ terminus of one end and has a
single stranded 5’ overhang at the same end,
with a second ds nucleic acid molecule having
a blunt end, such that the 5’ overhang can hy-
bridize to a complementary sequence of the
blunt end of the second nucleic acid molecule,
and the topoisomerase can covalently link the
ds nucleic acid molecules. The methods are
simpler and more efficient than previous
methods for covalently linking nucleic acid
sequences, and the compositions and kits fa-
cilitate practicing the methods, including
methods of directionally linking two or more
ds nucleic acid molecules.
84 11
15 We used MALLET software as it is actively updated and used in current literature (for examples, which has used
MALLET: Venugopalan and Rai 2014 and Thomas, Adams, Hassan and Blostein 2014).

similar patents, we found that 100 and 200 topics were far too broadly defined and not useful for our purposes. The
500 topic model performed well on random checks of the patents. We used the 500 topic model as the default for our
data analysis; however, we offer the 200-topic model as a robustness check in Table A1.
The algorithm then produces a vector of weights of each topic in each patent abstract. Patents may contain several
topics, though of different weights. We used the KL divergence formula (Bigi, 2003; Pinto et al., 2007) to determine
the similarity of each licensed patent to each possible patent from the relevant pool determined earlier. The KL diver-
gence of the probability distributions P, Q on a finite set X is defined as shown in the following equation16
:
DKL PjjQð Þ ¼
X
x2X
P xð Þ log
P xð Þ
Q xð Þ
:
The KL distance is symmetric and is derived from the KL divergence:
DKLD PjjQð Þ ¼ DKL PjjQð Þ þ DKL QjjPð Þ:
The KL distance measures how similar each patent’s distribution of applicable topics are to each other and, in our
data set, ranges from close to 0 (very similar) to 15 (very different). Table B1 shows the abstracts of a sample licensed
patent and its most similar patent along with the percentage distribution in their two dominant topics. Table B2 dis-
plays the top 20 terms associated with the two topics.
Each choice set is composed of the top 19 most similar patents to the licensed patent in the choice set using the KL
distance metric and the licensed patent itself, totaling 20 patents. We offer robustness checks where choice sets have
10 or 30 patents each in Table A1.
Table B2. Top 20 terms associated with topics 60 and 112
Topic 60 Nucleic Invention Acid Proteins Molecules Methods Vectors Cells Sequences Expression
Polypeptides Host Isolated Recombinant Provided Encoding Acids Compositions Disclosed Antibodies
Topic 112 Nucleic Acid Amplification DNA Target Sequence Reaction Primer Polymerase Invention
Methods Sequences Sample Method Detection Primers Chain Nucleotide Sequencing Present
16 KL distance has been used in many natural language applications like query expansion (Carpineto, Mori, Romano and
Bigi 2001), language models (Bigi, Huang and Mori 2003), and statistical language modeling (Dagan, Lee and Pereira
1999).

Why do some patents get licensed while others do not?

More Related Content

Viewers also liked

Similar to Why do some patents get licensed while others do not?

More from Ian McCarthy

Recently uploaded

Why do some patents get licensed while others do not?