Professional Documents
Culture Documents
Kaplan 2014
Kaplan 2014
Kaplan 2014
Copyright © 2014 John Wiley & Sons, Ltd. Strat. Mgmt. J. (2014)
DOI: 10.1002/smj
S. Kaplan and K. Vakili
as citations as prior art by subsequent patents) forward citation counts to measure breakthroughs,
(Fleming, 2001; Gittelman and Kogut, 2003; Hall which capture the economic value of the patent.
et al., 2001; Singh and Fleming, 2010; Trajtenberg The relationship they test is represented by Path C
et al., 1997). The dynamics are represented in in Figure 1.
Figure 1 in Paths A and B. To date, however, these Specifically, research has suggested that recom-
studies have mainly looked at the effect of recom- bination is most likely to lead to higher citation
bination in patents (using a variety of measures) rates if the knowledge combined is technologi-
on subsequent citations to those patents. In other cally distant. The idea is that combining knowledge
words, rather than testing Paths A and B separately, from exploratory or long-jump search (Gavetti and
they have tested Path C in Figure 1. They find Levinthal, 2000; March, 1991) is more likely to pro-
that recombination is positively associated with duce inventions that break from the existing techno-
economic value but do not analyze directly the logical and scientific models and ultimately become
intervening step associated with the generation of highly cited (Phene et al., 2006; Rosenkopf and
novelty from those recombination processes. Nerkar, 2001; Trajtenberg et al., 1997). Similarly,
In using topic modeling to analyze breakthrough scholars have argued that highly cited patents are
patents, we can explore this relationship directly more likely to be combinations of not just distant but
by measuring the presence of novel ideas (as also diverse knowledge domains (Hall et al., 2001),
indicated by shifts in vocabularies in the patent where greater diversity (lower concentration) of
texts) and determining if this variable mediates the knowledge avoids intellectual lock-in. According
association between wide recombination and value to these theories, the positive impact of long-jump
(as indicated by forward citations) or if local search search on value occurs through the (untested) mech-
(narrower recombination) is more likely to lead anism of novelty. It is noteworthy that these schol-
to cognitive and economic breakthroughs. This ars have not claimed that novelty is the sole driver
approach will allow us to understand if there are of economic value nor that recombination only
any contradictions between the processes leading serves to generate novel ideas and has no direct
to novelty and those engendering economic value. effects on the degree of economic value created. For
We will accomplish this through a test of mediation example, combining diverse knowledge domains
(Baron and Kenny, 1986; Iacobucci, Saldanha, and might enlarge audiences for the innovation, increase
Deng, 2007; Zhao, Lynch, and Chen, 2010) so that the likelihood it will be found by inventors or patent
we can examine each of the paths and their joint examiners in a search for prior art, or broaden the
effects. In testing Paths C and C′ , we replicate prior network in which innovations would diffuse. Taking
innovation studies showing the association between these insights together, we hypothesize that:
wide recombination and economic value. In testing
Paths A (and A′ ) and B, we explore the relation-
Hypothesis 1a: A patent based on distant
ship between tension and foundational views in
and diverse recombinations is more likely to
producing novel ideas that should subsequently be
receive a higher number of citations than
associated with economic value.
a patent produced based on local search
(narrow recombination) (Path C in Figure 1).
Foundational vs. tension theories and economic
value
While the existing empirical evidence for break-
Tension assumptions about recombination (Har- through patents has overwhelmingly supported
gadon and Sutton, 1997; Weisberg, 1999) have the tension view of recombination, the founda-
dominated management scholarship on innovation. tional view would propose the reverse relationship
The view is that deep knowledge in a single or between recombination and measures of economic
small number of domains may lock inventors into value. Under this logic, inventors should need
one way of thinking and therefore block their to explore a relatively narrow domain in depth
ability to generate breakthroughs. Local search in order to know how to “defy the crowd” and
enables only narrow recombinations that produce “buy low and sell high” (Sternberg and Lubart,
incremental innovations. Therefore, to generate 1995; Sternberg and O’Hara, 1999). Inventors
breakthroughs, inventors must combine knowledge cannot see new sources of value without under-
from distant and diverse sources. These studies use standing what assumptions are behind the existing
Copyright © 2014 John Wiley & Sons, Ltd. Strat. Mgmt. J. (2014)
DOI: 10.1002/smj
The Double-Edged Sword of Recombination
sources of value, and these insights can only innovation literature on patents discussed above.
come from focused, local search. This view of Of course, it is a requirement of the U.S. Patent
creativity is consistent with ecological theories Office that every patent be novel to some extent,
that domain-spanning activities may suffer market though some inventions may be “improvements,”
penalties due to both deficiencies in production built upon existing technological trajectories, while
of innovations as well as problems of market others may be truly novel, introducing new techno-
reception. Recombinations of distant or diverse logical trajectories. Our concern here is with those
knowledge might get in the way of identifying that meet this latter standard. Thus, we hypothesize:
value because they would disperse effort and
distract from obtaining the incisive insight that Hypothesis 2: Patents that represent cognitive
comes from a deep appreciation of one domain breakthroughs (truly novel ideas) are more
(Hannan, Polos, and Carroll, 2007). Thus, wide likely to receive a higher number of citations
recombination could prevent the realization of than patents that do not (Path B in Figure 1).
economic value because it produces superficial or
incremental work. One might also infer that wide
Novelty has been portrayed in the innovation
recombination could compromise the realization of
literature as an (unmeasured) output of recombi-
economic value because market audiences penalize
offerings that span categories (Hsu, Kocak, and nation and an input to the creation of economic
Hannan, 2009; Rao, Monin, and Durand, 2005; value (citations). For example, Trajtenberg et al.
Ruef and Patterson, 2009; Zuckerman, 1999). That (1997: 29) claim that “synthesis of divergent ideas
is, wide recombination of inputs could potentially is characteristic of research that is highly original.”
lead to difficulties in classifying the innovation and Similarly, Phene et al. (2006: 370) suggest that
thus to penalties in the form of fewer citations over “knowledge that is technologically … distant pro-
time. We therefore offer a competing hypothesis vides the organization with an opportunity to make
(as represented in Path C′ ) to the recombination novel linkages.” These arguments are based in the
model of creativity in generating economic value: tension view of recombination, which assumes that
knowledge and creativity are opposing forces, such
that “knowledge may provide the basic elements …
Hypothesis 1b: A patent based on local search out of which are constructed new ideas, but in order
(narrower recombination) is more likely to for these building blocks to be available, the mortar
receive a higher number of citations than a holding the old ideas together must not be too
patent produced based on distant and diverse strong,” and too much knowledge of a domain can
recombinations (Path C′ in Figure 1). be habit forming and inertial (Weisberg, 1999: 226).
Recombination of distant or diverse knowledge
Foundational vs. tension theories and novelty can break these habits. In introducing a measure
of novel ideas, we can make the implicit model in
By calling out novelty as a separate creative output innovation studies explicit: novel ideas—what we
from the generation of economic value, we are able conceptualize as cognitive breakthroughs—are the
to interrogate existing research that has privileged
products of distant and diverse recombinations:
recombination processes as the source of innovative
breakthroughs. Implicit in the arguments made in
studies of breakthroughs is the idea that wide Hypothesis 3a: Patents produced through
recombination generates novel ideas (Path A in distant and diverse recombinations are more
Figure 1), which in turn are more likely to be cited likely to be cognitive breakthroughs (truly
as prior art by subsequent patents (Path B). novel ideas) than those that are not (Path A
With regard to the connection between novelty in Figure 1).
and economic value, while the creativity literature
makes it clear that not every truly novel idea will The foundational view makes the opposite claim
become valuable (Amabile, 1983; Sternberg, 1997; (Weisberg, 1999). Here, immersion in a particu-
Sternberg and O’Hara, 1999), they also indicate lar domain is required in order to produce novelty
that novelty will increase the probability that eco- (Csikszentmihalyi, 1996). Local search and nar-
nomic value can be obtained, all else equal. This rower recombinations based on deep knowledge
logic is consistent with the arguments made in the in one area enable the identification of anomalies
Copyright © 2014 John Wiley & Sons, Ltd. Strat. Mgmt. J. (2014)
DOI: 10.1002/smj
S. Kaplan and K. Vakili
that lead to new insights by exposing the ten- the topic. These originating documents can be seen
sions or challenges in the current ways of think- as cognitive breakthroughs. In our case, because
ing. This is consistent with the Kuhnian (Kuhn we study patents, we call these topic-originating
1962/1996) model in which paradigm shifts are trig- patents. Because topic modeling is a new method
gered by the accumulation of anomalies. Narrow in strategic management, we introduce it first here
but deep search leads to truly novel breakthroughs before delving, in the next section, into the empir-
in knowledge because it enables researchers to ical methods and measures of other variables that
identify “what rules to break” (Taylor and Greve, are more standard in the field. We explain how our
2006: 726). These findings are also consistent with method works and then show how we have imple-
work at the inventor level of analysis suggesting mented it in our sample of fullerene patents. For rea-
that specialization is important to push the frontier sons of space, we have left certain technical details
of knowledge outward as the “burden of knowl- (useful to those looking to implement topic mod-
edge” increases over time (Agrawal, Goldfarb, and eling in their own work) to an online supporting
Teodoridis, 2013; Conti, Gambardella, and Mariani, document that is available as a companion to this
2014; Jones, 2009). We therefore offer a competing article.
hypothesis (as represented in Path A′ ) to the tension Our methodological move is to treat the texts
model of creativity: of patents as representations of the inventive ideas
embodied in them. Bibliometric techniques to
understand the evolution of science and technology
Hypothesis 3b: Patents produced through
have a long tradition starting from the pioneering
local search (narrower recombination) are
work of de Solla Price (1965a,b). However, most
more likely to be cognitive breakthroughs
of the work to date has used citation analyses (e.g.,
(truly novel ideas) than those that are not
Dahlin and Behrens, 2005; Leydesdorff, Cozzens,
(Path A′ in Figure 1).
and Van den Besselaar, 1994; Meyer et al., 2004).
Text analysis has been much less frequent, and,
If developing truly novel inventions were the only until recently, the main uses of the texts were
mechanism through which recombination would counts, factor analyses, and co-word analyses
lead to higher economic value, we should expect of keywords (typically in the titles of papers
full mediation of Path C when introducing Paths A or patents) (Azoulay, Ding, and Stuart, 2007;
and B. However, there are reasons to expect that this Mogoutov and Kahane, 2007; Upham, Rosenkopf,
may not be the case (as outlined in the argument and Ungar, 2010; Yoon and Park, 2005). With the
for Hypothesis 1a). But, without a measure of novel increasing power of computation and availability of
ideas, prior scholars have not been able to tease texts in electronic form, scholars are exploring the
apart the effect of novelty from other effects of possibilities of more complete uses of texts, which
recombination. would therefore require unsupervised approaches.
The study reported here follows these recent
trends. It is premised on the idea that studying lan-
TOPIC MODELING OF PATENT TEXTS: guage in documents should provide a reading of
A MEASURE OF NOVEL IDEAS their cognitive content (Duriau, Reger, and Pfarrer,
2007; Whorf, 1956). In management studies, this
Crucial to our analysis is the introduction of topic idea has been adapted methodologically to use word
modeling as a way to create a new measure of nov- counts to represent themes (Abrahamson and Ham-
elty to contrast with existing citation-based mea- brick, 1997; Huff, 1990; Kaplan, 2008; Kaplan,
sures of economic value in patents. The intuition Murray, and Henderson, 2003). Where the concern
behind topic modeling as a method to identify is in identifying themes over large numbers of texts,
novel ideas is the following: the algorithm uses the topic modeling, a text analysis technique developed
co-location of words in a collection of documents in computer science, offers exciting potential (see
to infer the underlying (or latent) topics in those Blei, 2012, for an overview). The advantage of topic
texts and the weight of each topic in each individ- modeling over word counts and keyword analy-
ual document. We can then identify the documents ses is that it allows for polysemy—words can take
that are the originators of each topic by finding on different meanings depending on their contexts,
the earliest documents with a significant weight in and it is inductive—the scholar does not have to
Copyright © 2014 John Wiley & Sons, Ltd. Strat. Mgmt. J. (2014)
DOI: 10.1002/smj
The Double-Edged Sword of Recombination
specify categories a priori but can allow them to
emerge from the data. This is particularly important
given Kuhn’s (1962/1996: 205) argument that “pro- α θ z w β
Table 1. Descriptive statistics, total sample, topic-originating patents, top cited patents, and all others
Breakthroughs (top 5% cited) 0.048 0.111 0.042 0.069** 1.000 0.000 1.000**
(0.213) (0.315) (0.202) (p = 0.000) (0.000) (0.000) (p = 0.000)
# forward citations (five-year) 13.507 22.173 12.767 9.406** 91.982 9.559 82.422**
(21.678) (30.958) (20.536) (p = 0.000) (29.963) (11.106) (p = 0.000)
Topic-originating patent 0.079 1.000 0.000 1.000** 0.183 0.074 0.110**
(0.270) (0.000) (0.000) (p = 0.000) (0.389) (0.261) (p = 0.000)
Technological distance 0.405 0.366 0.409 −0.043+ 0.474 0.402 0.072*
(0.328) (0.331) (0.327) (p = 0.094) (0.325) (0.328) (p = 0.026)
Technological diversity 0.736 0.681 0.741 −0.059* 0.758 0.735 0.023
(0.323) (0.353) (0.320) (p = 0.018) (0.317) (0.323) (p = 0.466)
Ln(component familiarity) 4.661 4.337 4.689 −0.353** 4.735 4.658 0.078
(1.001) (1.090) (0.989) (p = 0.000) (0.834) (1.009) (p = 0.430)
Ln(combination familiarity) 0.306 0.264 0.309 −0.045 0.225 0.310 -0.084
(0.773) (0.709) (0.778) (p = 0.448) (0.687) (0.776) (p = 0.265)
Ln(cumulative combination) 0.409 0.370 0.412 −0.041 0.287 0.415 -0.127
(0.965) (0.871) (0.973) (p = 0.581) (0.819) (0.971) (p = 0.179)
Ln(average experience) 2.101 1.961 2.113 −0.152+ 2.158 2.098 0.059
(1.171) (1.272) (1.162) (p = 0.096) (1.197) (1.170) (p = 0.605)
Team 0.791 0.722 0.797 −0.075* 0.899 0.786 0.113**
(0.406) (0.449) (0.402) (p = 0.017) (0.303) (0.410) (p = 0.005)
Assigned 0.889 0.889 0.889 0.000 1.000 0.883 0.117**
(0.314) (0.315) (0.314) (p = 1.000) (0.000) (0.321) (p = 0.000)
No prior art 0.048 0.094 0.044 0.051** 0.027 0.049 -0.021
(0.214) (0.293) (0.205) (p = 0.002) (0.164) (0.216) (p = 0.308)
Ln(# prior art patents) 2.123 1.896 2.142 −0.246** 2.413 2.108 0.304**
(1.066) (1.198) (1.052) (p = 0.003) (1.238) (1.055) (p = 0.004)
Ln(# non-patent references) 1.498 1.641 1.486 0.156 2.360 1.455 0.905**
(1.301) (1.341) (1.297) (p = 0.124) (1.419) (1.280) (p = 0.000)
Ln(# claims) 2.835 2.725 2.844 −0.119* 3.133 2.820 0.313**
(0.764) (0.816) (0.759) (p = 0.045) (0.749) (0.762) (p = 0.000)
Ln(family size) 0.126 0.201 0.120 0.081** 0.386 0.113 0.273**
(0.331) (0.390) (0.325) (p = 0.002) (0.547) (0.311) (p = 0.000)
Means (standard deviations) of variables for each category of patents are shown. Column 5 reports the difference between columns 3 and
4 and column 8 reports the difference between columns 3 and 4 (p-value in parentheses).
**p < 0.01; *p < 0.05; +p < 0.10.
Copyright © 2014 John Wiley & Sons, Ltd. Strat. Mgmt. J. (2014)
DOI: 10.1002/smj
S. Kaplan and K. Vakili
This is the recommended approach for testing RESULTS
mediated relationships where there are more than
one independent variable (the case in our analysis) Tables 2 (for citation counts) and 3 (for break-
(Cho and Pucik, 2005; Iacobucci et al., 2007; Zhao throughs in the top 5% of citations) report the results
et al., 2010). Positive signs for our recombination of the structural equation models. Column 1 shows
variables as tested in Paths A, A × B (the indirect the direct effect of the independent variables mea-
effect of recombination on citations as mediated suring recombination processes and mediator mea-
by novelty), and C in Figure 1 would support the suring novel ideas on the dependent variable (Paths
tension view of creativity; negative signs (Paths C and B). Column 2 shows the direct effect of
A’, A’ × B, and C’) would be evidence for the recombination on novel ideas (Path A). Column
foundational view. This approach will also allow 3 identifies mediation effects in the analysis and
us to verify prior studies showing a direct, positive shows the indirect effect of recombination as medi-
association between distant and diverse recombi- ated by novel ideas (A × B). Column 4 represents
nation and citations (Path C). The advantage of the total effect of the independent variables and
SEM relative to running three separate regressions mediator on the dependent variables, taking into
(the traditional approach to testing mediation, account the direct and indirect effects ([A × B] + C).
according to Baron and Kenny, 1986) is that the
simultaneous equations control for measurement
Testing path C
errors that might lead to under- or overestimation
of mediation effects (Shaver, 2005). The indirect In testing Hypotheses 1a and 1b for Path C, we are
effect of one of the independent variables (IV) on replicating the prior studies on breakthrough inno-
the dependent variable (DV) through the mediator vations linking distant and diverse recombination
can be calculated by multiplying the estimated with subsequent citations (a measure of economic
direct effect of the IV on the mediator (Path A) value). Though the samples of the prior studies are
and the estimated direct effect of the mediator vastly different (in terms of numbers of observa-
on the DV (Path B). As a robustness check, we tions, time periods, technological arenas), we find
also conducted a mediation analysis with separate support for Hypothesis 1a in our fullerene patent
regressions for each path in Figure 1 and found dataset in terms of direction and, in one case, sig-
highly consistent results in both the effect size and nificance of effect for each of the dependent vari-
significance. ables (Model 1 in Tables 2 and 3). More distant
The nature of our dependent variables (one is a (technological distance) and diverse (technologi-
count and the other binary) and mediator variable cal diversity) combinations of knowledge are pos-
(also binary) places additional constraints on the itively associated with subsequent citations, though
SEM approach. Using a linear model with count and the effect is only significant for technological dis-
categorical dependent and mediator variables can tance in the patent count model (Table 2). Note also
lead to biased results. We therefore use the gener- that, as previous studies have found, other measures
alized SEM (GSEM) model introduced in Stata 13, sometimes linked with recombination—previous
which allows generalized linear response functions use of components and combinations, greater expe-
with count and binary outcomes. We use a negative rience of inventors, and invention in organizations
binomial function for regressions with count out- and teams—are positively associated with citation
comes and a probit function for regressions with rates. The significance of the effects are attenuated
binary outcomes. Employing a maximum likeli- when using the dummy variable for citation-based
hood estimator, GSEM provides consistent, effi- breakthroughs (in Table 3), but this is likely due to
cient, and asymptotically normal estimates for paths the reduction in variance in the dependent variable.
A, B, and C. We further use nonparametric boot-
strapping (with 1,000 replications) to adjust esti-
Testing mediation (Paths B and A)
mates for bias and to estimate the indirect effects
(A × B), total effects ([A × B] + C), their standard Confirming Hypothesis 2, Model 1 in both Tables 2
errors, and their confidence intervals. All the sig- and 3 shows that the measure of truly novel
nificance levels are determined by the bias-adjusted ideas (topic-originating patents) is strongly pos-
bootstrap confidence intervals (Efron and Tibshi- itively associated with subsequent citation rates.
rani, 1993; Mooney and Duval, 1993). This relationship (for Path B) is statistically and
Copyright © 2014 John Wiley & Sons, Ltd. Strat. Mgmt. J. (2014)
DOI: 10.1002/smj
The Double-Edged Sword of Recombination
Table 2. Tests of mediation (dv = citation counts, five-year window since grant date) (1991–2005)
Generalized structural equation model using bootstrapping (1,000 repetitions), bias-corrected coefficients, and robust standard errors.
**p < 0.01; *p < 0.05; +p < 0.10.
economically significant. Looking at Model 1 in patent is 0.031, which is substantially greater than
Table 2, a topic-originating patent is likely to the marginal effects of the recombination variables.
receive 1.4 times more citations than the average On the other hand, we do not see the positive
patent. Similarly, looking at Model 1 in Table 3, if a effect of distant and diverse recombination on nov-
patent is topic originating, the odds of it becoming elty anticipated by tension theories (Hypothesis 3a).
an economic breakthrough as measured by citation First, looking at the direct effect of recombination
rates increase by a factor of 1.7. In other words, variables on novel ideas in Model 2, we find that
holding all other variables at their means, the prob- technological distance and technological diversity
ability of gaining a breakthrough level of citations are negatively and significantly associated with
is 0.072 for topic-originating patents (those repre- topic-originating patents. That is, topic-originating
senting novel ideas) compared to 0.024 for other patents are not the result of the combination of dis-
patents. The marginal effect of topic-originating tant or diverse knowledge but are instead produced
patents on the likelihood of becoming a top-cited through local search (confirming the foundational
Copyright © 2014 John Wiley & Sons, Ltd. Strat. Mgmt. J. (2014)
DOI: 10.1002/smj
S. Kaplan and K. Vakili
Table 3. Tests of mediation (dv = dummy for citation-based breakthroughs, top 5 percent, five-year window since grant
date) (1991–2005)
Generalized structural equation model using bootstrapping (1,000 repetitions), bias-corrected coefficients, and robust standard errors.
**p < 0.01; *p < 0.05; +p < 0.10.
view of creativity as represented in Hypothesis 3b). Turning to Model 3, which is the test of media-
Of the recombination-related controls, only inven- tion, we do not find the complementary mediation
tor experience appears to have a significant associa- suggested by Hypothesis 3a, but instead a com-
tion with novelty. The positive and significant signs peting mediation relationship. Breakthrough novel
for experience on both value and novelty suggest ideas are associated with higher citation rates, but
that inventors’ previous experience in patenting distant and diverse recombinations do not appear to
may increase capabilities in recombination (as produce that novelty, and indeed pull in the opposite
suggested by the tension view) or deep knowledge direction. Because distant and diverse recombina-
in the field (as suggested by the foundation view), tions have a positive direct effect on citations but are
or both. Future research might explore the effect negatively associated with cognitive breakthroughs,
of experience on these two different creative their total effect (Model 4) is not statistically signif-
processes. icantly different from 0. This finding is the essence
Copyright © 2014 John Wiley & Sons, Ltd. Strat. Mgmt. J. (2014)
DOI: 10.1002/smj
The Double-Edged Sword of Recombination
of the double-edged sword of recombination. These for second generation technologies” (see also Fur-
competing mediation effects are statistically sig- man and Stern, 2011).
nificant based on multiple tests (Iacobucci, 2012; There are several reasons to believe that gen-
Kenny, 2013; MacKinnon and Dwyer, 1993).5 erating novel ideas may not automatically lead to
breakthrough levels of citations (which represent
Relationship between cognitive novelty economic value). A novel idea embodied in a patent
and economic value still depends on other factors to become known
and used, including the reputation and status of the
We find that patents that are especially novel are inventors (Azoulay, Stuart, and Wang, 2014; Mer-
also especially valuable. Following the methodol- ton, 1968), the distribution of the idea in the relevant
ogy introduced by Rysman and Simcoe (2008) to network (Singh, 2005), the match of the invention
evaluate patent citation patterns and rates adjusted with the environmental demand (Sørensen and
for confounding factors such as cohort effects, we Stuart, 2000), and the presence of complementary
found that topic-originating patents are more likely technologies (Rosenberg, 1996). In the absence of
to have higher citations than other patents both in such factors, a patent that represents a novel idea
their first generation and in their second generation may not gain traction. Similarly, not all highly
(that is, in patents citing patents that cite the focal cited patents represent novel breaks in knowledge.
patent) (results available from the authors). Patents with a broad scope and general claims,
On the other hand, this relationship is not per- patents inside patent thickets (dense networks of
fect. As mentioned above, only 20 of the 189 cogni- patents with overlapping claims), patents that make
tive breakthroughs (as measured by topic modeling) an original idea more understandable and usable,
are also economic breakthroughs (patents in the top or patents that distribute an idea strategically in a
5% of five-year forward citations—there are 109 network—all may lead to higher citations whether
of these in our dataset). Our method thus highlights or not the patent introduces a truly novel idea. By
the separate but interrelated nature of cognitive and adding a direct measure of novelty, our analysis is
economic breakthroughs. Those patents that repre- a first step in isolating the effects of novelty from
sent breakthrough levels of both novelty and value these social dynamics (often associated with recom-
are rare (less than 1% of our sample) but appear to bination processes) that should increase value.
have a greater impact on future innovation than any
other kind of invention. As a result, our approach
may offer empirical handholds for addressing ques-
tions of cumulative research, especially where, as DISCUSSION AND CONCLUSION
Scotchmer (1991: 39) suggests, a “first technology
has very little value on its own but is a foundation Our primary objective for this study was to
examine the double-edged sword of recombina-
5
In the first set of mediation estimations reported in Table 2,
tion in creating innovative breakthroughs. To do
the main dependent variable is a count variable (citation counts), this, we look at the effect of distant and diverse
whereas the mediator is a binary variable (a dummy for whether recombinations on two innovative outputs: nov-
or not the patent is topic originating). Hence, estimates associated
with Paths A and B in Figure 1 are calculated using different
elty and economic value. We introduce a new
response functions (probit vs. negative binomial). As a result, one method—topic modeling—for measuring the
might be concerned that the estimates for the total indirect effects novelty of ideas embedded in patent texts by
(A × B) and their standard errors might not be accurate. While
the properties of GSEM estimates and the bootstrapping method identifying those patents that originate new topics
should take care of this concern, we nevertheless performed a in a body of knowledge. This measure allows us
robustness test suggested by Iacobucci (2012) to examine the to distinguish inventions that are cognitively novel
significance of the indirect effects. Here, we find the z-statistic of
each indirect effect is significant, consistent with the results from in the Kuhnian sense—they introduce new lan-
the GSEM method. Furthermore, in the estimations where both the guage and therefore new ways of thinking—from
dependent variable and the mediator are dummies (as reported in inventions that are economically valuable (as
Table 3), we used the method proposed by Kenny (2013) (see also
MacKinnon and Dwyer, 1993) as a robustness check of our results. measured by the subsequent citations they receive).
In this method, the estimates for Paths A and B are first scaled to It also enables us to examine contrasting creative
similar levels and then their product is estimated using the delta processes—those based in either tension or foun-
method. Here, again, we find the indirect effects are significant.
Both robustness tests confirm the findings estimated by the GSEM dational assumptions—that contribute to novelty
method. and value.
Copyright © 2014 John Wiley & Sons, Ltd. Strat. Mgmt. J. (2014)
DOI: 10.1002/smj
S. Kaplan and K. Vakili
Implications for understanding breakthroughs and Greve, 2006; Weisberg, 1999). Theories of
recombination are based in the former, while our
Our approach takes seriously the idea put forth
results on the sources of cognitive breakthroughs
by Griliches (1990) and pursued in recent studies
are better explained by the latter: local search
(Alcácer and Gittelman, 2006; Alcácer, Gittelman,
to uncover anomalies is more likely to produce
and Sampat, 2009; Benner and Waldfogel, 2008;
breaks in the existing knowledge and language.
Hegde and Sampat, 2009; Jaffe, Trajtenberg, and
Identifying breakthroughs in knowledge using topic
Fogarty, 2000; Tan and Roberts, 2010) that patents modeling may help us develop further insights
should be assessed as historical documents pro- into Kuhn’s (1962/1996: 62) model of technolog-
duced by inventors, prosecuted by patent attorneys, ical change based on “the previous awareness of
and evaluated by patent examiners. An implication anomaly, the gradual and simultaneous emergence
is that it should be useful to analyze the texts in these of observational and conceptual recognition, and
patents, which is also consistent with the cogni- the consequent change of paradigm categories and
tive turn being made in studies of technology emer- procedures.”
gence and evolution (Kaplan and Tripsas, 2008). As an early foray into the use of a new method,
In doing so, we complement existing research this study is not surprisingly constrained by some
on technology evolution, in particular that which limitations. Most importantly, topic modeling is
draws on patent data to understand the sources of sensitive to the corpus of documents selected for
innovation. the analysis. Because the technique is based in the
The imperfect relationship between topic- generation of posterior probabilities, the identifica-
originating patents and those that receive high tion of topics and topic-originating patents will be
citations may indicate that there are different kinds affected by which documents are included in the
of breakthroughs, those that introduce truly novel analysis. This is in turn affected by which inven-
knowledge and those that are associated with tions are patented and by which documents the
economic value. Distinguishing between the novel researcher selects to include in the corpus. Patents
and the valuable (and understanding the sources are an imperfect source of information on new
of each) is quite important for several reasons. scientific and technological ideas. Not all inven-
Breakthroughs in knowledge mark the potential tions are patented (Griliches, 1990; Scherer, 1983).
origins of new technological paradigms. Further- We are therefore surely missing ideas and topics
more, identifying the patents that mark shifts that withered on the vine. This constraint is bal-
in knowledge may help us understand different anced by the rich bibliometric data that patents
mechanisms through which new ideas spread over provide, which allow the scholar to examine the
time and space and explain why some new ideas effects of citations, inventors, assignees, patent
become the wheels of economic fortune and some classes, and the like. We have also addressed poten-
simply grind to a halt after a few years (Podolny tial bias in our sample that pre-established clas-
and Stuart, 1995). sification systems create by using three different
By operationalizing the concept of novel ideas search methodologies to identify patents related to
implicit in many studies of the sources of inno- fullerenes.
vation, we are able to distinguish processes that
produce novelty from those that produce eco-
nomic value. The contrasting results for distant Implications for organizations
and diverse recombination are particularly striking. Our results are shaped by the possibility that the
They suggest that generating new topics requires processes we observe are endogenous to each other.
deep immersion in a narrower domain rather than We measure the impact of recombination on two
linking to more distant or diverse knowledge. On different aspects of innovation: novelty and value.
the other hand, patents that cite prior art from Assuming that individuals and organizations are
a wide range of patent classes are more applica- strategic in setting their goals, they simultane-
ble in a variety of domains and therefore more ously decide about how much novelty and value
likely to be cited in the future. These findings they should pursue in their innovative activities.
highlight the double-edged sword of recombination In other words, the decision to achieve a cer-
based on the tension and foundational models tain amount of economic value is simultaneously
of the role of knowledge in creativity (Taylor determined with the decision to achieve a certain
Copyright © 2014 John Wiley & Sons, Ltd. Strat. Mgmt. J. (2014)
DOI: 10.1002/smj
The Double-Edged Sword of Recombination
amount of novelty. As a result, what we observe Extensions of topic modeling as a tool in studies
in our data is the resulting outcome of such simul- of innovation
taneous decisions by individuals and organizations
In addition to identifying the sources and impacts
about how much effort to place on long-jump or
of breakthroughs, topic modeling may usefully
local search. In that sense, while pursuing nov-
contribute to other areas of research on science,
elty can lead to high citation rates (as we see
technology, and knowledge. For example, topic
in our results), pursuing economic value at the
modeling can allow us to analyze at a more
same time can influence the level of effort put
by individuals and organizations to achieve novel fine-grained level technological distance, ties, and
innovations. spillovers between firms and other entities. To
This endogeneity has an empirical implication. date, this research has primarily been conducted
The simultaneity and mutual dependency between through cross-citation analyses of the overlaps in
the two decisions directly influence the number of USPTO patent classifications among the patents of
valuable vs. novel innovations we observe in our different entities (e.g., Ahuja, 2000; Jaffe, 1986;
sample and also the size and significance of the Song, Almeida, and Wu, 2003) or citations between
regression coefficients. One can think of another entities (e.g., Henderson, Jaffe, and Trajtenberg,
equilibrium in which organizations would have 1998; Jaffe, Trajtenberg, and Henderson, 1993;
put much more effort in finding novel innovations, Mowery, Oxley, and Silverman, 1996). Scholars are
which could change the number of topic-originating increasingly raising concerns about the degree to
patents in our sample and consequently the size and which patent classifications are proxies of location
significance of the effects we find. In that sense, in technological space (Benner and Waldfogel,
what we measure in Paths A and B is influenced 2008) and about the noisiness of patent citations
by what we measure in Path C and vice versa (and as measures of knowledge flows (Alcácer and
this is precisely why we use the GSEM method- Gittelman, 2006; Duguet and MacGarvie, 2005;
ology). While this simultaneity and correlation Roach and Cohen, 2013). In the case of nan-
of outcomes can influence our estimated coeffi- otechnology, for example, ethnographic research
cients, our results nevertheless highlight important has found that knowledge flows are not fully
contrasting effects from distant and diverse captured by co-authoring and citations, where
recombination on novel and valuable innovative exchanging students and experimental materials,
outcomes. commenting on each other’s work, or participating
This endogeneity also has an organizational in problem-solving workshops were more powerful
implication. The results suggest that an effective and frequent mechanisms (Mody, 2011). Yet, we
innovation strategy needs to bridge between wide have lacked reasonable alternative quantitative
recombination and local search in order to facilitate measures for knowledge flows (Roach and Cohen,
the transformation of novel ideas into economically 2013).
valuable ones. By understanding the double-edged Topic modeling of patents may provide one solu-
sword of recombination in driving novelty and tion to augment existing approaches. Because topic
value, organizations can think about how to man- models produce a vector of weights of each topic
age these conflicts. The literature has not yet studied for each patent, one can evaluate the content of
the ways in which tension and foundational views ties using topics and the strength of ties using
of creativity interact. These views have been posi- weights. This approach may be a useful comple-
tioned as alternative theories of creativity rather ment to patent classes because it tracks the language
than as two processes operating simultaneously in of the actors rather than the classifications assigned
organizations. Our model might be consistent with by others. It also adds greater nuance than available
a variation-selection-retention view of innovation in current cross-citation approaches by, first, exam-
where variation (novelty) is produced by one set of ining the ideas directly rather than inferring them
processes while selection and retention of the most from citation ties and, second, allowing for the pos-
potentially valuable ideas are produced by another. sibility that connections among ideas occur even if
Future research could explore the organizational the patents are not cited.
design implications of the presence of these con- Topic modeling thus offers a new means of gener-
flicting effects, potentially across different stages of ating inductively classifications of ideas from texts,
innovation. which may be advantageous as we look beyond
Copyright © 2014 John Wiley & Sons, Ltd. Strat. Mgmt. J. (2014)
DOI: 10.1002/smj
S. Kaplan and K. Vakili
patents to other collections of documents. With REFERENCES
the burgeoning interest in classification and cat-
egorization (e.g., Kennedy, 2005; Lounsbury and Abrahamson E, Hambrick DC. 1997. Attentional homo-
Rao, 2004; Navis and Glynn, 2010; Pontikes, 2012), geneity in industries: the effect of discretion. Journal
topic models can identify themes or frames as of Organizational Behavior 18: 513–532.
they emerge and evolve over time (e.g., DiMag- Agrawal A, Goldfarb A, Teodoridis F. 2013. Does knowl-
edge accumulation increase the returns to collabora-
gio, Nag, and Blei, 2013; Ruef and Nag, 2014). tion? NBER working paper 19694, National Bureau
This approach has the distinct appeal of dispens- of Economic Research, Cambridge, MA. Available at:
ing with the requirement to use pre-established cat- http://www.nber.org/papers/w19694 (accessed 15 April
egories or to come up with ex post classification 2014).
systems. Instead, the data can speak for themselves, Ahuja G. 2000. Collaboration networks, structural holes,
and innovation: a longitudinal study. Administrative
thus allowing the researcher to observe paths that Science Quarterly 45(3): 425–455.
fall away as well as paths that become consolidated Ahuja G, Coff R, Lee P. 2005. Managerial foresight and
over time. As such, topic modeling could be vital attempted rent appropriation: insider trading on knowl-
in understanding the emergence and institutional- edge of imminent breakthroughs. Strategic Manage-
ization of new fields. We hope that our early foray ment Journal 26(8): 791–808.
into the application of topic modeling to social sci- Ahuja G, Lampert CM. 2001. Entrepreneurship in the large
corporation: a longitudinal study of how established
ence questions can instigate further explorations in firms create breakthrough inventions. Strategic Man-
these directions. agement Journal 22(6–7): 521–543.
Albert MB, Narin F, Avery D, McAllister P. 1991. Direct
validation of citation counts as indicators of industrially
ACKNOWLEDGEMENTS important patents. Research Policy 20(3): 251–259.
Alcácer J, Gittelman M. 2006. Patent citations as a measure
of knowledge flows: the influence of examiner citations.
We are grateful to Michael Lounsbury for the sug- Review of Economics and Statistics 88(4): 774–779.
gestion to study fullerenes and nanotubes within the Alcácer J, Gittelman M, Sampat B. 2009. Applicant and
broad domain of nanotechnology, Juan Alcácer for examiner citations in U.S. patents: an overview and
his help in developing the initial data set, and Hanna analysis. Research Policy 38(2): 415–427.
Amabile TM. 1983. The social-psychology of creativ-
Wallach for her early test-driving of topic modeling ity – a componential conceptualization. Journal of Per-
on these data. Grace Gui, Illan Kramer, Sara sonality and Social Psychology 45(2): 357–376.
Sojung Lee, Octavio Martinez, Shamsi Mohtashim, Amabile TM, Pillemer J. 2012. Perspectives on the social
Neal Parikh, Navid Soheilnia, Aaron Sutton, and psychology of creativity. Journal of Creative Behaviour
Xihua Wang provided much-appreciated research 46(1): 3–15.
assistance. Earlier drafts of this manuscript ben- Audia PG, Goncalo JA. 2007. Past success and creativity
over time: a study of inventors in the hard disk drive
efited from the comments of Ajay Agrawal, industry. Management Science 53(1): 1–15.
Christian Catalini, Kristina Dahlin, Andrea Fosfuri, Azoulay P, Ding W, Stuart T. 2007. The determinants of
Alfonso Gambardella, Constance Helfat, Nico faculty patenting behavior: demographics or opportu-
Lacetera, Anita McGahan, Andrea Mina, Jasjit nities? Journal of Economic Behavior & Organization
Singh, Kevyn Yong, two anonymous reviewers, 63(4): 599–623.
Azoulay P, Stuart T, Wang Y. 2014. Matthew: effect or
and participants in the DRUID conference 2012 fable? Management Science 60(1): 92–109.
(where it received the Best Paper Prize), EGOS Baron RM, Kenny DA. 1986. The moderator-mediator
Colloquium 2011, the West Coast Research Sym- variable distinction in social psychological research:
posium 2011, the slump_management research conceptual, strategic, and statistical considerations.
group, and in seminars at Rotman, the Yale School Journal of Personality and Social Psychology 51(6):
1173–1182.
of Management, Chicago Booth, University of Benner M, Waldfogel J. 2008. Close to you? Bias and
Illinois-Urbana Champaign, and Wharton. This precision in patent-based measures of technological
work is partially supported by the Mack Institute proximity. Research Policy 37(9): 1556–1567.
for Innovation Management at the Wharton School, Bessen J. 2008. The value of US patents by owner
University of Pennsylvania (where S. Kaplan is a and patent characteristics. Research Policy 37(5):
Senior Fellow) and the Canadian Social Sciences 932–945.
Blei DM. 2012. Probabilistic topic models. Communica-
and Humanities Research Council under grant tions of the ACM 55(4): 77–84.
#410-2010-0219. All errors and omissions remain Blei DM, Lafferty JD. 2007. A correlated topic model of
our own. science. Annals of Applied Statistics 1(2): 634–642.
Copyright © 2014 John Wiley & Sons, Ltd. Strat. Mgmt. J. (2014)
DOI: 10.1002/smj
The Double-Edged Sword of Recombination
Blei DM, Ng AY, Jordan M. 2003. Latent dirichlet Gavetti G, Levinthal D. 2000. Looking forward and
allocation. Journal of Machine Learning Research 3: looking backward: cognitive and experiential search.
993–1022. Administrative Science Quarterly 45(1): 113–137.
Bresnahan TF, Trajtenberg M. 1995. General purpose tech- Gittelman M, Kogut B. 2003. Does good science lead
nologies: engines of growth? Journal of Econometrics to valuable knowledge? Biotechnology firms and the
65(1): 83–108. evolutionary logic of citation patterns. Management
Campbell D. 1960. Blind variation and selective retention Science 49(4): 366–382.
in creative thought as in other knowledge processes. Griliches Z. 1990. Patent statistics as economic indicators
Psychological Review 67(6): 380–400. - a survey. Journal of Economic Literature 28(4):
Chang J, Boyd-Graber J, Wang C, Gerrish S, Blei D. 2009. 1661–1707.
Reading tea leaves: how humans interpret topic mod- Grimmer J, Stewart BM. 2013. Text as data: the promise
els. In Proceedings of Neutral Information Process- and pitfalls of automatic content analysis methods for
ing Systems. Vancouver, B.C., Canada. Available at: political texts. Political Analysis 21(3): 267–297.
http://goo.gl/QzeEv9 (accessed 15 April 2014). Grun B, Hornik K. 2011. Topic models: an R package
Cho HJ, Pucik V. 2005. Relationship between innovative- for fitting topic models. Journal of Statistical Software
ness, quality, growth, profitability, and market value. 40(13): 1–30.
Strategic Management Journal 26(6): 555–575. Guilford JP. 1967. The Nature of Human Intelligence.
Cockburn IM, Henderson RM. 1998. Absorptive capacity, McGraw-Hill: New York.
coauthoring behavior, and the organization of research Hall BH. 2002. A note on the bias of herfindahl-type
in drug discovery. Journal of Industrial Economics measures based on count data. In Patents, Citations, and
46(2): 157–182. Innovations, Jaffe A, Trajtenberg M (eds). MIT Press:
Conti R, Gambardella A, Mariani M. 2014. Learning Cambridge, MA; 454–459.
to be Edison? Individual inventive experience and Hall BH, Jaffe A, Trajtenberg M. 2001. The NBER patent
breakthrough inventions. Organization Science 25(3): citation data file: lessons, insights, and methodological
833–849. tools. NBER working paper 8498, National Bureau
Csikszentmihalyi M. 1996. Creativity: Flow and the Psy- of Economic Research, Cambridge, MA. Available at:
chology of Discovery and Invention. HarperCollins http://www.nber.org/papers/w8498 (accessed 15 April
Publishers: New York. 2014).
Dahlin KB, Behrens DM. 2005. When is an invention Hall BH, Jaffe A, Trajtenberg M. 2005. Market value and
really radical? Defining and measuring technological patent citations. RAND Journal of Economics 36(1):
radicalness. Research Policy 34(5): 717–737. 16–38.
Deng Z, Lev B, Narin F. 1999. Science and technology Hall D, Jurafsky D, Manning CD. 2008, Studying the
as predictor of stock performance. Financial Analysts histories of ideas using topic models. In Proceedings
Journal 53(3): 20–32. of the Conference on Empirical Methods in Natural
DiMaggio P, Nag M, Blei D. 2013. Exploiting affini- Language Processing. Honolulu, Hawaii. Avilable at:
ties between topic modeling and the sociological per- http://dl.acm.org/citation.cfm?id=1613763 (accessed
spective on culture: application to newspaper coverage 15 April 2014).
of government arts funding in the U.S. Poetics 41(6): Hannan MT, Pólos L, Carroll G. 2007. Logics of Organiza-
570–606. tion Theory: Audiences, Codes, and Ecologies. Prince-
Duguet E, MacGarvie M. 2005. How well do patent ton University Press: Princeton, NJ.
citations measure flows of technology? Evidence from Hargadon A, Sutton RI. 1997. Technology brokering and
French innovation surveys. Economics of Innovation innovation in a product development firm. Administra-
and New Technology 14(5): 375–393. tive Science Quarterly 42(4): 716–749.
Duriau VJ, Reger RK, Pfarrer MD. 2007. A content anal- Harhoff D, Narin F, Scherer FM, Vopel K. 1999. Citation
ysis of the content analysis literature in organization frequency and the value of patented inventions. Review
studies: research themes, data sources, and method- of Economics and Statistics 81(3): 511–515.
ological refinements. Organizational Research Meth- Hegde D, Sampat B. 2009. Examiner citations, applicant
ods 10(1): 5–34. citations, and the private value of patents. Research
Efron B, Tibshirani R. 1993. An Introduction to the Policy 105(3): 287–289.
Bootstrap. Chapman and Hall: New York. Helpman E. 1998. Diffusion of general purpose technolo-
Emma P. 2006. How to write a patent. IEEE Micro 26(1): gies. In General Purpose Technologies and Economic
143–144. Growth, Helpman E (ed). MIT Press: Cambridge, MA;
Ethiraj SK, Levinthal D. 2004. Modularity and innova- 85–117.
tion in complex systems. Management Science 50(2): Henderson R, Jaffe AB, Trajtenberg M. 1998. Universi-
159–173. ties as a source of commercial technology: a detailed
Fleming L. 2001. Recombinant uncertainty in technologi- analysis of university patenting, 1965–1988. Review of
cal search. Management Science 47(1): 117–132. Economics and Statistics 80(1): 119–127.
Furman JL, Stern S. 2011. Climbing atop the shoul- Hsu G, Kocak O, Hannan MT. 2009. Multiple category
ders of giants: the impact of institutions on cumu- memberships in markets: an integrative theory and two
lative research. American Economic Review 101(5): empirical tests. American Sociological Review 74(1):
1933–1963. 150–169.
Copyright © 2014 John Wiley & Sons, Ltd. Strat. Mgmt. J. (2014)
DOI: 10.1002/smj
S. Kaplan and K. Vakili
Huff AS. 1990. Mapping Strategic Thought. John Wiley March JG. 1991. Exploration and exploitation in organiza-
and Sons: Chichester, NY. tional learning. Organization Science 2(1): 71–87.
Iacobucci D. 2012. Mediation analysis and categorical Mei Q, Shen X, Zhai C. 2007. Automatic labeling of multi-
variables: the final frontier. Journal of Consumer Psy- nomial topic models. In Proceedings of the 13th ACM
chology 22(4): 582–594. SIGKDD International Conference on Knowledge Dis-
Iacobucci D, Saldanha N, Deng X. 2007. A meditation on covery and Data Mining. San Jose, California.
mediation: evidence that structural equations models Merton RK. 1968. The Matthew effect in science. Science
perform better than regressions. Journal of Consumer 159(3810): 56–63.
Psychology 17(2): 140–154. Meyer M, Pereira TS, Persson O, Granstrand O. 2004. The
Jaffe AB. 1986. Technological opportunity and spillovers scientometric world of Keith Pavitt. Research Policy
of research-and-development – evidence from 33(9): 1405–1417.
firms patents, profits, and market value. American Mody CC. 2011. Instrumental Community: Probe
Econonomic Review 76(5): 984–1001. Microscopy and the Path to Nanotechnology. MIT
Jaffe AB, Trajtenberg M, Fogarty M. 2000. Knowledge Press: Cambridge, MA.
spillovers and patent citations: evidence from a survey Mogoutov A, Kahane B. 2007. Data search strategy for
of inventors. American Economic Review Papers and science and technology emergence: a scalable and evo-
Proceedings 90(2): 215–218. lutionary query for nanotechnology tracking. Research
Jaffe AB, Trajtenberg M, Henderson R. 1993. Geo- Policy 36(6): 893–903.
graphic localization of knowledge spillovers as evi- Mooney CZ, Duval RD. 1993. Bootstrapping: A Nonpara-
denced by patent citations. Quarterly Journal of Eco- metric Approach to Statistical Inference. No. 94–95.
nomics 108(3): 577–598. Sage: Newbury Park, CA.
Jones BF. 2009. The burden of knowledge and the death Mowery DC, Oxley JE, Silverman BS. 1996. Strategic
of the renaissance man: is innovation getting harder? alliances and interfirm knowledge transfer. Strategic
Review of Economic Studies 76(1): 283–317. Management Journal 17: 77–91.
Kaplan S. 2008. Cognition, capabilities and incentives: Navis C, Glynn MA. 2010. How new market categories
assessing firm response to the fiber-optic revolution. emerge: temporal dynamics of legitimacy, identity,
Academy of Management Journal 51(4): 672–695. and entrepreneurship in satellite radio, 1990–2005.
Kaplan S, Murray F, Henderson R. 2003. Discontinuities Administrative Science Quarterly 55(3): 439–471.
Phene A, Fladmoe-Lindquist K, Marsh L. 2006. Break-
and senior management: assessing the role of recogni-
through innovations in the U.S. biotechnology industry:
tion in pharmaceutical firm response to biotechnology.
the effects of technological space and geographic ori-
Industrial and Corporate Change 12(2): 203–233.
gin. Strategic Management Journal 27(4): 369–388.
Kaplan S, Tripsas M. 2008. Thinking about technology:
Podolny JM, Stuart TE. 1995. A role-based ecology of
applying a cognitive lens to technical change. Research
technological change. American Journal of Sociology
Policy 37(5): 790–805. 100(5): 1224–1260.
Kennedy MT. 2005. Behind the one-way mirror: refrac- Pontikes EG. 2012. Two sides of the same coin: how
tion in the construction of product market categories. ambiguous classification affects multiple audiences’
Poetics 33(3–4): 201–226. evaluations. Administrative Science Quarterly 57(1):
Kenny D. 2013. Mediation with dichotomous outcomes. 81–118.
Working paper. Available at: http://davidakenny.net/ Ramage D, Rosen E, Chuang J, Manning C, McFarland
doc/dichmed.pdf (accessed 15 April 2014). DA. 2009. Topic modeling for the social sciences.
Kuhn TS. 1962/1996. The Structure of Scientific Revolu- In NIPS 2009 Workshop on Applications for Topic
tions. University of Chicago Press: Chicago, IL. Models: Text and Beyond. Whistler, Canada.
Kuusi O, Meyer M. 2007. Anticipating technologi- Rao H, Monin P, Durand R. 2005. Border crossing:
cal breakthroughs: using bibliometric coupling to bricolage and the erosion of categorical boundaries
explore the nanotubes paradigm. Scientometrics 70(3): in French gastronomy. American Sociological Review
759–777. 70(6): 968–991.
Lanjouw JO, Schankerman M. 2004. Patent quality and Roach M, Cohen WM. 2013. Lens or prism? Patent
research productivity: measuring innovation with mul- citations as a measure of knowledge flows from public
tiple indicators. Economic Journal 114(495): 441–465. research. Management Science 59(2): 504–525.
Leydesdorff L, Cozzens S, Van den Besselaar P. 1994. Rosenberg N. 1996. Uncertainty and technological change.
Tracking areas of strategic importance using scien- In The Mosaic of Economic Growth, Landau R, Taylor
tometric journal mappings. Research Policy 23(2): T, Wright G (eds). Stanford University Press: Stanford,
217–229. CA; 334–353.
Lounsbury M, Rao H. 2004. Sources of durability and Rosenkopf L, Nerkar A. 2001. Beyond local search:
change in market classifications: a study of the boundary-spanning, exploration and impact in the opti-
reconstitution of product categories in the American cal disc industry. Strategic Management Journal 22(4):
mutual fund industry, 1944–1985. Social Forces 82(3): 287–306.
969–999. Rothaermel FT, Thursby M. 2007. The nanotech ver-
MacKinnon DP, Dwyer JH. 1993. Estimating mediated sus biotech revolution: sources of productivity in
effects in prevention studies. Evaluation Review 17: incumbent firm research. Research Policy 36(6):
144–158. 832–849.
Copyright © 2014 John Wiley & Sons, Ltd. Strat. Mgmt. J. (2014)
DOI: 10.1002/smj
The Double-Edged Sword of Recombination
Ruef M, Nag M. 2014. The classification of organiza- Taylor A, Greve HR. 2006. Superman or the fantastic
tional forms: theory and application to the field of four? Knowledge combination and experience in inno-
higher education. In Remaking College: The Chang- vative teams. Academy of Management Journal 49(4):
ing Ecology of Higher Education, Stevens ML, Kirst 723–740.
MW (eds). Stanford University Press: Stanford, CA; Trajtenberg M. 1990. A penny for quotes: patent citations
147–160. and the value of innovations. RAND Journal of Eco-
Ruef M, Patterson K. 2009. Credit and classification: the nomics 21(1): 172–187.
impact of industry boundaries in nineteenth-century Trajtenberg M, Henderson R, Jaffe A. 1997. University
America. Administrative Science Quarterly 54(3): versus corporate patents: a window on the basicness of
486–520. invention. Economics of Innovation and New Technol-
Rysman M, Simcoe T. 2008. Patents and the performance ogy 5(1): 19–50.
of voluntary standard-setting organizations. Manage- Upham SP, Rosenkopf L, Ungar LH. 2010. Innovating
ment Science 54(11): 1920–1934. knowledge communities an analysis of group collabo-
Scherer FM. 1983. The propensity to patent. International ration and competition in science and technology. Sci-
Journal of Industrial Organization 1(1): 107–128. entometrics 83(2): 525–554.
Scotchmer S. 1991. Standing on the shoulders of US Patent and Trademark Office. 2005. Handbook of Clas-
giants – cumulative research and the patent-law. sification. United States Government Printing Office:
Journal of Economic Perspectives 5(1): 29–41. Washington, DC.
Shaver JM. 2005. Testing for mediating variables in Weisberg RW. 1999. Creativity and knowledge: a chal-
management research: concerns, implications, and lenge to theories. In Handbook of Creativity, Sternberg
alternative strategies. Journal of Management 31(3): RJ (ed). Cambridge University Press: Cambridge, UK;
330–353. 226–250.
Simonton DK. 1999. Creativity as blind variation and Whorf BL. 1956. Science and linguistics. In Language,
selective retention: is the creative process Darwinian? Thought, and Reality: Selected Writings of Benjamin
Psychological Inquiry 10(4): 309–328. Lee Whorf , Carroll JB (ed). MIT Press: Cambridge,
Singh J. 2005. Collaborative networks as determinants MA; 207–219.
of knowledge diffusion patterns. Management Science Wry T, Greenwood R, Jennings PD, Lounsbury M. 2010.
51(5): 756–770. Institutional sources of technological knowledge: a
Singh J, Fleming L. 2010. Lone inventors as sources of community perspective on nanotechnology emergence.
breakthroughs: myth or reality? Management Science Research in the Sociology of Organizations 29:
56(1): 41–56. 149–176.
de Solla Price D. 1965a. Is technology historically inde- Yoon B, Park Y. 2005. A systematic approach for iden-
pendent of science? A study in statistical historiogra- tifying technology opportunities: keyword-based mor-
phy. Technology and Culture 6: 553–568. phology analysis. Technological Forecasting and Social
de Solla Price D. 1965b. Little Science, Big Science. Change 72(2): 145–160.
Columbia University Press: New York. Zhao X, Lynch JG Jr, Chen Q. 2010. Reconsidering baron
Song J, Almeida P, Wu G. 2003. Learning-by-hiring: when and Kenny: myths and truths about mediation analysis.
is mobility more likely to facilitate interfirm knowledge Journal of Consumer Research 37(2): 197–206.
transfer? Management Science 49(4): 351–365. Zuckerman EW. 1999. The categorical imperative: secu-
Sørensen JB, Stuart TE. 2000. Aging, obsolescence, and rities analysts and the illegitimacy discount. American
organizational innovation. Administrative Science Journal of Sociology 104(5): 1398–1438.
Quarterly 45(1): 81–112.
Sternberg RJ. 1997. Thinking Styles. Cambridge University
Press: Cambridge, UK.
Sternberg RJ, Lubart TI. 1995. Defying the Crowd: Culti- SUPPORTING INFORMATION
vating Creativity in a Culture of Conformity. Free Press:
New York.
Sternberg RJ, O’Hara LA. 1999. Creativity and intel- Additional supporting information may be found
ligence. In Handbook of Creativity, Sternberg RJ in the online version of this article:
(ed). Cambridge University Press: Cambridge, UK;
251–272. Appendix S1. The double-edged sword of recombi-
Tan D, Roberts PW. 2010. Categorical coherence, clas- nation in breakthrough innovation.
sification volatility and examiner-added citations.
Research Policy 39(1): 89–102.
Copyright © 2014 John Wiley & Sons, Ltd. Strat. Mgmt. J. (2014)
DOI: 10.1002/smj