CHANGES IN FIRM KNOWLEDGE COUPLINGS AND FIRM INNOVATION PERFORMANCE: THE MODERATING ROLE OF TECHNOLOGICAL COMPLEXITY

Changes in Knowledge Couplings and Innovation Performance 379 negatively moderated when domain complexity

ex ante , the performance implications of any is high.

combination of knowledge domains are initially unknown and can be discovered only through

experimentation. 2 Further, given the large number

THEORY AND HYPOTHESES

of potential interdependencies, firms have to make important choices about which to focus on. The

Firms and recombinant inventions

combinations that a firm decides to consider are governed primarily by the existing couplings

Technological inventions are generated by recom- between knowledge domains (Yayavaram and bining knowledge in novel ways (Fleming, 2001;

Ahuja, 2008). The coupling of two domains Kogut and Zander, 1992; Schumpeter, 1934). Each

indicates the extent to which the firm is likely to such combination represents an invention. Firms

combine knowledge elements from each in search- explore the “search space” of possible combina-

ing for new inventions. Couplings can vary from tions (i.e., the technological landscape) to iden-

strong (domains X and Y are always considered tify valuable inventions. A technological search

together) to weak (X and Y are considered together almost never starts ab initio. Rather, it typically

occasionally) to nonexistent (X and Y are always involves changing the configurations of an exist-

considered independently). 3 Because couplings ing combination to create new combinations, and

guide the combinations a firm considers, they then examining the many recombinations possible

play an important role in technological search. An to determine whether any will lead to an enhanced

example of a firm’s set of knowledge couplings performance outcome.

is provided in Figure 1. 4 In this figure, the nodes The eventual outcomes are determined by the

represent the firm’s knowledge domains, and node knowledge domains considered by the entity con-

size is a measure of domain size (i.e., the number ducting the search (in our case, a firm), the interde-

of the firm’s patents that belong to that domain). pendencies among those knowledge domains, and

Viewing a firm’s knowledge base as a network also the combinations considered. First, in the search

suggests that couplings can be interpreted as the process, a firm considers only a small subset of the

ties that the firm has between knowledge domains universe of potential knowledge domains. This set

and that the strength of a tie between two domains of domains is, by definition, the firm’s knowledge

represents the level of coupling between them. base. The number of domains that can potentially

It may be useful at this point to consider some

be included in the consideration set (the knowl- of the differences between the concepts of interde- edge universe) is the same across firms, but the

pendence and coupling. Interdependence is part of domains that are actually included differ by firm.

the natural world (Yayavaram and Ahuja, 2008). Second, search effectiveness is also strongly influ-

Since it is not possible for a firm to know all inter- enced by the interdependencies between knowl-

dependencies a priori , improving understanding of edge domains (Kauffman, 1993; Levinthal, 1997).

interdependencies is an important part of a techno- In the context of technological innovation, two

logical search. Coupling, in contrast, is a decision knowledge domains, X and Y, are interdependent

that belongs to the made world , that is, a decision when inventions that combine X and Y are more

that the search entity makes, either explicitly or likely to be successful than those that use X or Y

implicitly, while conducting the search. If a firm alone. Thus, inventors need to search both domains

to identify valuable XY configurations. When the two knowledge domains are independent, in con-

2 A firm can also engage in a cognitive search to generate maps

trast, a search in one can proceed independently of

(Fleming and Sorenson, 2004; Gavetti and Levinthal, 2000) that

that in the other.

can help to locate promising areas. However, the firm must

Third, the outcomes are affected by which

eventually engage in an experiential search to identify valuable outcomes (Gavetti and Levinthal, 2000).

combinations the search entity considers (Flem-

3 We define coupling pairwise in this paper. This definition,

ing, 2001; Stuart and Podolny, 1996). Although

however, does not preclude a firm from combining more than

interdependencies obviously influence which

two knowledge domains. In other words, if a firm has a high level

combinations are considered, they do not fully of coupling between domains X and Y and between domains Y

and Z, it is likely to combine all three domains in an invention.

determine these decisions. Since it is not possible

4 The approach used to construct these knowledge couplings is

to know all of the underlying interdependencies

discussed in the Methods section.

380 S. Yayavaram and W.-R. Chen

Figure 1. Knowledge couplings for Intel in 1996. The numbered nodes represent technology classes and the ties between nodes represent coupling between those classes, where the strength of a tie represents level of coupling and darker lines indicate stronger ties; the node size is a measure of the number of patents that belong to that technology class. Isolated nodes and very weak ties have been removed to clarify the illustration, which includes only those classes

that are related to the semiconductor industry 5

perceives X and Y to be interdependent (i.e., per- ceives that a combined search across X and Y can lead to valuable configurations), it may pro- vide for some level of coupling between these two domains. However, even if a firm perceives the two domains to be interdependent, it may choose not to couple them because it is impossible for it to consider all of the interdependencies in the techno- logical world. It may also decide that other interde- pendencies are more worthy of consideration than that between X and Y. Thus, the actual set of cou- plings that a firm chooses is likely to be a subset of the set of all potential couplings. Further, a firm’s couplings do not usually converge to the underly- ing interdependencies since provision of coupling between all interdependent domains makes the search process rigid (Yayavaram and Ahuja, 2008).

With respect to the technological search process, it is useful to distinguish between the invention,

5 Figure taken from Yayavaram and Ahuja, (2008). Decompos- ability in knowledge structures and its impact on the usefulness

of inventions and knowledge-base malleability. Administrative Science Quarterly 53(2): 333–362. Reprinted with permission from SAGE Publications.

the search entity, and the underlying technological environment. Our focus on knowledge couplings allows us to distinguish between the search space and the characteristics of the entity (in our case, the firm) conducting the search. In our concep- tualization, the characteristics of the search space are determined by interdependencies exogenous to the firm, whereas the choices of where to search (i.e., the consideration set) and which domains to couple or not couple are endogenous. As a point of clarification, our approach differs from that of Fleming and Sorenson (2001), whose focus is on the complexity of an invention. Ours is on the firm-level search process. By shifting our analy- sis to the firm level, we are able to examine the enablers of and hindrances to change that exist within a firm, not all of which exist for each indi- vidual invention. Further, we can examine features of a firm’s knowledge base, such as knowledge couplings, that exist at levels higher than that of an invention.

The coupling decisions within a firm are driven by the strategies and markets it pursues, the actions of its competitors, externally sourced knowledge,

Changes in Knowledge Couplings and Innovation Performance 381 social cognition (Barr, Stimpert, and Huff, 1992;

Kiesler and Sproull, 1982), social organization (Galunic and Rodan, 1998), and internal sociopo- litical processes (Lyles and Schwenk, 1992). Cou- plings are the outcomes of both deliberate and emergent activities within the firm. For example,

a coupling may be driven by a top manage- ment directive or emerge from the actions of researchers working independently. Couplings are embodied in routines, in beliefs about the under- lying interdependencies between domains, and in communication networks and the organizational structure (Yayavaram and Ahuja, 2008). Coupling between previously uncoupled domains X and Y is accomplished by changes in a firm’s inter- nal routines, communication patterns (e.g., the R&D staff in Research Unit 1, which special- izes in domain X, will now interact with the R&D staff in Research Unit 2, which special- izes in domain Y), organizational structure (the two research units may merge), or resource alloca- tion, effort, and attention (projects that involve X and Y will get more funding and top management attention).

In the next section, we consider the organiza- tional issues involved when couplings change, and develop hypotheses about what kinds of changes in a firm’s knowledge couplings lead to better innovation outcomes and how domain technolog- ical complexity may moderate the relationship between the two. We consider innovation qual- ity (i.e., the value or impact of a firm’s inven- tions) to be a more appropriate outcome variable than quantity (i.e., the number of inventions) in representing innovation performance. Innovation performance is enhanced when the firm gener- ates inventions that create economic value for the firm by resolving important techno-economic problems.

Changes in a firm’s knowledge couplings

A firm can change its knowledge base by chang- ing either the knowledge domains it covers or the couplings between those domains. Given the importance of recombination to the inno- vation process, we focus on how firms change the couplings between their knowledge domains while controlling for changes in domains per se that might also determine innovation outcomes. The need for a change in knowledge couplings arises from two factors. First, firms have to

change their technologies over time to keep pace with the changing environment— including the changing needs of customers, suppliers, and alliance partners—and their competitors. Second, the returns from an existing knowledge coupling decline over time because of the technological and social/psychological exhaustion of potential refinements (Fleming, 2001; Fleming and Soren- son, 2001). Technological exhaustion eventually occurs because there is a finite number of ways in which a given set of knowledge domains can

be recombined (Ahuja and Lampert, 2001; Kim and Kogut, 1996).

Changes in a firm’s knowledge couplings can

be disaggregated into five constituent types: (1) a firm can change the coupling among its existing knowledge domains; (2) when new knowledge is being acquired or developed organically, the firm may couple the new knowledge domains with existing domains; (3) when a firm discontinues the use of two or more knowledge domains (i.e., exits two or more technologies), it dissolves the couplings between those domains; (4) when a firm discontinues the use of particular knowledge domains, it dissolves the couplings between those domains and existing knowledge domains; and (5) the firm may couple new knowledge domains with one another (see Table S1 in the supporting information). For conceptual clarity, we focus on the first and second types of change (in which at least one of the knowledge domains is familiar to the firm) in the hypothesis section. These two types together account for nearly

70 percent of all change in couplings in our sample.

As an example of these two types of change, consider again Intel’s knowledge couplings in 1996 (see Figure 1). The increased coupling between class 257 (Active solid-state devices), class 174 (Electricity: conductors and insulators), and class 361 (Electricity: electrical systems and devices) from 1990 to 1996 is an example of coupling changes between existing knowledge domains. The increase in the level of coupling between these classes can be ascribed to an increase in the number of inventions targeting heat dissipation in microprocessors, which became a major issue for developers from the early 1990s onward as a result of the rapid increase in transistor density. An example of coupling between new and existing knowledge domains is Intel’s entry into technology class 348 (Television), a step it took

382 S. Yayavaram and W.-R. Chen to facilitate its entry into the videoconferencing

existing knowledge domains that were weakly market, and the subsequent couplings between

coupled before, the combinations that are being that class, which was new to Intel, and existing

considered are not entirely novel to the firm. As it classes 370 (Multiplex communications) and 375

already possesses the knowledge in question, the (Pulse or digital communications). Using network

factors that may hinder innovation, such as a lack terminology, a change in the level of coupling

of absorptive capacity or the “not-invented-here” between a pair of knowledge domains can be

syndrome, are absent. At the same time, by using interpreted as a change in the strength of the tie

its existing knowledge in novel ways, the firm can between those domains. It is important to note that

attempt novel modes of reasoning and adopt new in looking at changes in the strength of couplings,

problem-solving approaches and, thereby, over- rather than only at the addition or deletion of

come the “familiarity trap” (Ahuja and Lampert, couplings, we are looking at a more comprehensive

2001). The firm can thus generate combinations measure of change. In other words, changes in the

that, while being built on existing knowledge, strength of ties encompasses cases in which ties

provide the benefits of novelty and exploration. have been created or deleted, whereas the addition

However, the firm is also likely to face signif- or deletion of ties excludes all cases in which

icant costs in undertaking any coupling change, existing ties have become weaker or stronger.

whether it be the addition/deletion of couplings or To better understand how these different types

the strengthening/weakening of existing couplings. of change affect the innovation process, it is

Organizational learning is usually path dependent, important to look at both the costs and benefits

that is, firms typically evolve along paths on of change and whether the former outweigh the

which their technological expertise has played latter. It is not only a question of whether the

a critical role in the past (Cohen and Levinthal, benefits outweigh the costs, but also of identifying

1990; Dosi, 1988; Nelson and Winter, 1982). The the conditions under which they do so. We first

same applies to the combinations that the firm discuss the benefits and costs of change for the

is considering. Once a set of existing knowledge base case, and then examine how the technological

couplings has been developed, it is usually environment in which the firm operates may alter

stabilized in organizational routines and emerges the balance between the two.

as the dominant logic in guiding knowledge recombinations (Henderson, 1992). Owing to such routinization, the reconfiguration of existing

Coupling changes among existing knowledge

knowledge incurs costs and is likely to reduce the

domains

efficiency of knowledge integration (Grant, 1996).

A large-scale change may destroy the currently edge domains occurs when a firm introduces new

A change in the coupling among existing knowl-

shared assumptions of R&D staff, information couplings, discontinues old couplings, and weak-

filters, and problem-solving strategies within the ens or strengthens existing couplings among these

organization (Henderson and Clark, 1990). It is domains. We group all such changes into one cat-

difficult to undertake too many changes at once egory, as their effect on innovation outcomes is

because each requires changes in the way that likely to be the same. The mechanisms behind the

new relationships are perceived or accepted by the effects of increases and decreases in coupling on

organization and/or the innovation team. Dramatic innovation outcomes are the same, and hence these

changes can cause coordination problems in an increases and decreases should be grouped together

innovation community with a shared identity rather than considered separately.

regarding how organizational knowledge should Such changes in couplings can have a positive

be developed socially (Galunic and Rodan, 1998; impact on innovation performance for several

Kogut and Zander, 1992). In many cases, the reasons. By undertaking these changes, a firm is

coupling of knowledge domains may even be exploring new combinations or shifting its focus

hardwired in terms of geographic locations, away from one set of combinations to another. The

organizational divisions, and communication firm is already familiar with its existing knowl-

patterns. Since the search process is impaired, edge domains, and has a high level of expertise in

both the quantity and the quality (or value) of using them (Ahuja and Lampert, 2001). Further,

inventions are affected. Hence, the undertaking in the case of strengthening the coupling between

of major configuration changes in organizational

Changes in Knowledge Couplings and Innovation Performance 383 knowledge couplings is difficult, and can easily

satisfice in the search process and the outcomes lead to failure (Henderson and Clark, 1990).

of changes are uncertain, they may be content to While we argue that change has both benefits

stay at their existing position on the technological and costs, it is difficult to make a strong theoretical

landscape if they perceive it to be successful case for why benefits outweigh costs or vice versa.

enough. Thus, they are likely to resist change We therefore posit competing hypotheses.

unless their current performance does not meet the threshold of what they consider successful. Even

Hypothesis 1a: Change in coupling among when a firm sees a rival doing something different, existing knowledge domains has a positive effect

it may not accept that the rival’s approach is on firm innovation performance.

better since long-term performance implications of technologies are not immediately clear. Due to existing cognitive filters and expertise in these

Hypothesis 1b: Change in coupling among domains, R&D teams are less likely to accept existing knowledge domains has a negative

another approach as better especially when both effect on firm innovation performance.

domains are familiar. In the face of such resistance to change, reconfiguration of existing knowledge

Coupling between new and existing knowledge

domains will be poorly managed and thus unlikely

domains

to provide the desired benefits. When additional domains are considered in the

A firm may also alter its knowledge base by intro- search process, there may be greater awareness of ducing new technologies. Entering a new technol-

the need for change. A significant problem with ogy domain does not in itself guarantee success

existing couplings is that, owing to cognitive fil- because multiple factors may intervene. The value

ters, research teams may fail to recognize that of the new technology may be highly uncertain,

the domains with which they are familiar can be thereby increasing the possibility that the firm has

combined in novel ways. When domains are unfa- made the wrong choice. More importantly, the firm

miliar, in contrast, the combinations that are con- may lack the capability or necessary experience to

sidered would naturally be novel to the firm. Also, conduct a search in the new technological domain,

the presence of new domains may reveal inter- which may confer competitive advantage on other

dependencies of which the firm was previously more capable firms. Several benefits will accrue

unaware. When these additional interdependen- to the firm, however, if the new technology can

cies are considered, the firm may become aware

be coupled with existing knowledge. The num- of more valuable configurations in the immediate ber of combinations that the firm can examine

neighborhood. The current performance will then rises, as knowledge from the new domains can

be deemed unsatisfactory, and change will be seen potentially be combined with its existing knowl-

as desirable. While there may be disruptions due to edge. Such combinations involve less uncertainty

disagreements about which new domains to enter than entirely new combinations because the firm

and which search directions to pursue, most of the is familiar with at least part of the technology in

disagreements would occur before these decisions question, which can thus serve as a stepping stone

are made. Thus, once a firm decides on a particular to new technology entry, allowing the firm to grad-

course of action and change is ongoing, disruption ually build research capabilities in the new domain.

is less likely when some domains are new and, Further, coupling a new knowledge domain with

consequently, search is more effective with new existing domains creates linkages that cannot be

and existing domains than with existing domains easily matched by competitors.

alone.

Similar to the prior case of coupling changes Further, by coupling new and existing knowl- among existing knowledge domains, there are

edge, firms can engage in boundary-spanning costs associated with creating couplings between

exploration (Rosenkopf and Nerkar, 2001), which new and existing knowledge domains. However,

reinvigorates searches with existing knowledge as these costs are likely to be much lower in the case

new ideas are introduced and also enables the of coupling with new knowledge domains. To see

firm to explore new technological, market, or why, consider how the search process works with

geographic opportunities by leveraging its exist- existing knowledge domains. Since firms typically

ing strengths (Nerkar, 2003). The benefits of

384 S. Yayavaram and W.-R. Chen such new couplings go beyond the generation of

boundary-spanning exploratory inventions, since these couplings imply that the firm’s knowledge base itself has changed. Hence, coupling between new and existing knowledge domains can have

a lasting effect on search processes within the firm, unlike boundary-spanning exploration that, by itself, may or may not have such an effect. We thus expect more valuable inventions to arise when new knowledge is well integrated with exist- ing knowledge domains.

Hypothesis 2: Coupling between new and exist- ing knowledge domains has a positive influence on firm innovation performance .

Domain complexity

We argue that a firm’s technological environ- ment will determine the payoff from knowledge recombinations and thereby moderate the effects of search strategies on innovation outcomes. We focus on domain complexity, the contextual char- acteristic most often highlighted in the literature on innovation landscapes (Ethiraj and Levinthal, 2004; Levinthal and Warglien, 1999; Nickerson and Zenger, 2004; Rivkin, 2000). The degree of domain complexity is low when there are few interdependencies among the knowledge domains in a firm’s consideration set. In such situations, the firm needs to examine only a few combina- tions. Since there is less uncertainty about which combinations are valuable, a firm searching locally using its existing knowledge base will have fewer directions to pursue that will lead to enhanced innovation performance. A high degree of domain complexity, in contrast, leads to greater uncertainty about which combinations are valuable and how the search should progress. Further, the presence of

a large number of interdependencies implies that a change made to one domain during the search pro- cess may require changes in many other domains (Rivkin and Siggelkow, 2003). Discovery of most of the valuable combinations available requires the firm to follow a different search path (i.e., to “think outside the box”) and move to parts of the search space that it has not yet explored.

In our arguments for Hypothesis 1, we pointed out the benefits and costs of changes in the cou- plings among existing knowledge domains. The benefits of such changes are likely to be greater when the degree of domain complexity is high

for the following reasons. Changing the couplings among existing knowledge domains leads to new combinations or a different mix of combinations and can lead the firm away from its local neighbor- hood. Such changes, which usually reflect changes in the mental models of R&D researchers, chang- ing cognitive frameworks, and the approaching of problems from a fresh angle (Barr et al., 1992; Kiesler and Sproull, 1982), are more valuable in

a highly complex environment. Experiments with various recombinations of knowledge domains provide new starting points for further exploration, thereby helping the firm to avoid being stuck on

a competence peak (Ethiraj and Levinthal, 2004; Kauffman, 1993; Siggelkow and Rivkin, 2005). The resulting search diversity is preferable in an environment characterized by a high degree of complexity and multiple opportunities (Levinthal, 1997; Siggelkow and Rivkin, 2005). In simpler innovation contexts, in which the search space is devoid of peaks or has already been explored, there is less to be gained from reconfiguring existing coupling relationships. Hence, a change in cou- pling among existing knowledge domains is more useful when the underlying technologies are more complex.

Hypothesis 3: Domain complexity positively moderates the effect of change in coupling among existing knowledge domains on firm innovation performance (or attenuates the neg- ative effect of such a change) .

The complexity of the technological environ- ment is likely to have contrary effects with respect to couplings between new and existing knowledge domains. When the degree of domain complexity is low, a firm can evaluate the consequences of integrating unfamiliar new knowledge into its existing knowledge base without much difficulty. When it is high, however, knowledge domains simultaneously interact in multiple ways, and the relationships between existing knowledge domains and an unfamiliar new domain are difficult to understand in a timely fashion, and the effect of recombinations is not easy to predict (Flem- ing, 2001). Before a firm can fully grasp the benefits and constraints of combinations with a new knowledge domain, it may err in its under- standing of the relationships involved, thus result- ing in a search that takes it to parts of the

Changes in Knowledge Couplings and Innovation Performance 385 search space that are less promising than antici-

pated. In a highly complex environment, it may

be preferable for a firm to follow a less inte- grated approach, keeping highly interdependent and familiar domains together while leaving new and unfamiliar domains loosely coupled (Ethi- raj and Levinthal, 2004; Sanchez and Mahoney, 1996). We thus expect that the greater the domain complexity, the less effective the coupling of new knowledge domains with existing knowledge domains will be in generating valuable inventions.

Hypothesis 4: Domain complexity negatively moderates the effect of coupling between new and existing knowledge domains on firm inno- vation performance .

METHODS Sample

Our measures of changes in firms’ knowledge base and of contextual variables are based on the patents granted by the U.S. Patent and Trademark Office (USPTO). We tested our hypotheses on a longitudinal data set comprising all U.S. firms listed in Standard & Poor’s Compustat database from 1976 to 2004. Choosing firms from a wide range of industries made it possible for us to test the moderating effects of domain complexity. These firms were then matched to the assignee names on the patents issued by the USPTO via a procedure based on data from Professor Bronwyn Hall’s patent name-matching project (http://elsa. berkeley.edu/∼bhhall/pat/namematch.html). For all patent-related variables, we followed previous research in using the patent’s date of application, unless otherwise stated. Because we needed data from the six previous years to compute our measures of change, a firm entered our sample in year t only if it had patents in at least two years during the t -6 to t -4 period and in at least two years during the t -3 to t -1 period. We were thus left with a sample of 1,750 firms for which we were able to build change- and context-related measures.

Dependent variable

The dependent variable Firm innovation perfor- mance was measured as the number of citations

that a firm’s patents filed in year t had received in subsequent years up to 2004. The number of citations that a patent receives is generally accepted as a significant predictor of both its value (Hall, Jaffe, and Trajtenberg, 2005; Harhoff et al. , 1999) and technological impact. In our context, changes in coupling are expected to affect the quality of inventions that are generated and, thereby, affect the number of citations received. There could be concern that a firm may be erecting barriers to imitation when it attempts new couplings whether between existing knowledge domains or between existing and new domains, which then would affect the number of citations made by other firms. This is unlikely to be an issue in our context since what is new to a firm is not necessarily new to other firms.

Independent variables

We used the USPTO’s technology class data to build our measures of knowledge couplings. The USPTO assigns every patent it issues to one or more three-digit technology classes. We used the

classification that was current in 2005. 6 Unlike citations, which are prone to examiner bias, there should be few errors in class assignments because the USPTO takes adequate care when classifying patents. Following Fleming (2001) and Fleming and Sorenson (2001), we considered the technology classes assigned to patents as proxies for knowledge domains, and the co-listing of classes as indicative of a recombinant search process. 7

We assumed a firm’s knowledge base or patent portfolio at t to consist of all of the patents that it

6 Whenever the USPTO changes its technology classification system, it retroactively changes the class assignments for all

previous patents to maintain consistency at a particular point in time. Hence, our measures are unaffected by changes in the classification system.

7 Unlike Fleming and Sorenson (2001) who consider technolog- ical subclasses as domains, we chose to consider technology

classes as domains for the following reasons. Building coupling matrices using technological subclasses would result in sparse matrices and an overestimation of changes from one time period to another, as even a small inconsequential difference would count as a change. Coarser aggregation such as technological categories (with each category consisting of related technological classes), in contrast, would lead to an underestimation of change, since firms typically do not make broad changes at the techno- logical category level. Analysis at the technological class level avoids both extremes. Further, using subclasses would result in very large matrices that are difficult to handle computationally. We therefore chose to use classes for our coupling measure.

386 S. Yayavaram and W.-R. Chen had accumulated from t -3 to t -1. We used a three-

the data set. In setting a baseline, we needed year window to minimize the effects of yearly

to control for size because the median level of fluctuations in patent applications. The coupling

coupling can change with size. As the size of between technology classes j and k for firm i ,

a coupling matrix grows (i.e., as the number L i ,j − k , t − 3 to t − 1 , can be calculated as

of domains in a knowledge base increases), the number of possible couplings grows at a squared

rate. Naturally, the average coupling level between

n j + n k − n jk

any two domains can be expected to decline as the size of the coupling matrix increases. Our

where n j is the number of patents assigned to quartile ranking estimates took this effect into class j , n k is the number of patents assigned to

consideration by estimating the 25th, 50th, and class k , and n jk is the number of patents assigned

75th percentile values of coupling as a function to both classes (Yayavaram and Ahuja, 2008).

of the size of a firm’s patent portfolio and time, Coupling matrix L i ,j − k , t − 3 to t − 1 , consisting of

since there may be changes in the median level of L i ,j − k , t − 3 to t − 1 for all pairs of domains, repre-

coupling over time.

sents the firm’s knowledge base. The logic of this The coupling for each class pair can be placed measure is that a firm’s repeated recombination of

in one of four quartiles in both the earlier and later two domains implies that the two are strongly cou-

time periods. We classified an increase in coupling pled in its knowledge base. An important limitation

as significant when it changed from (1) a zero of this measure of revealed couplings is that it does

value 9 in the earlier time period to the second, not capture all of the couplings that were attempted

third, or fourth quartile in the later period, (2) from but did not result in patentable inventions.

the first quartile in the earlier time period to the To calculate the changes in a firm’s knowledge

third or fourth quartile in the later period, or (3) couplings, we compared each firm’s coupling

from the second to the fourth quartile. Likewise, a matrix for the t -6 to t -4 period with that for the

significant decrease was considered to occur when t -3 to t -1 period. Doing so precluded common

the coupling for a class pair changed from (4) patents between the two coupling matrices being

the fourth quartile to a zero value or the first or compared. We measured change in coupling as the

second quartile, (5) from the third quartile to a weighted number of technology class pairs that

zero value or the first quartile, or (6) from the had undergone a significant change in coupling

second quartile to a zero coupling. A coupling between the two time periods (Yayavaram and

change was then measured as the weighted number Ahuja, 2008). A significant change was defined

of technology class pairs that had undergone a as a change that exceeded one quartile. We

significant change in coupling between the two estimated the 25th, 50th, and 75th percentile values

time periods. The weight is equal to i + p j of coupling as a function of the size of the firm’s

i + p j )/ 2 where p i (resp. p j ) and p i (resp. p j ) patent portfolio and time.

represent the percentage of patents that belong We used a quartile ranking-based measure to

to technology class i (resp. j ) during the t -6 to avoid two kinds of bias in measuring change: (1)

t -4 and t -3 to t -1 periods, respectively. Change no change being recorded as a change because

8 of measurement errors, in couplings among existing knowledge domains and (2) coupling strength was measured as the change in coupling between

varying with the size of a firm’s knowledge base domains that were present in both time periods. (Yayavaram and Ahuja, 2008). We estimated the

Coupling between new and existing knowledge quartile rankings using all nonzero couplings in

domains was measured as the changes resulting from the addition of couplings between these

domains (see Table S1 for an illustration).

Consider an error of ε 1 in measuring coupling L during the first

time period and an error of ε 2 during the second. If there has

been no change in the true value of the coupling, the difference between the two time periods is

9 While estimating the quartiles, we considered only nonzero equal to

1 –L–ε 2 | , which is

couplings because there were a substantial number of zero this measure implies a change even though there has been no real

1 –ε 2 | . Because

1 –ε 2 | is greater than zero,

couplings. If these zero couplings were considered, the first change in the coupling value. We consider absolute values for

quartile would include only zero couplings. As we treated zero our measures since our theoretical arguments concern change,

couplings differently in estimating the quartiles, we also treated not simply increases or decreases in coupling.

them differently in considering significant changes in coupling.

Changes in Knowledge Couplings and Innovation Performance 387 Domain complexity depends on the number of

When measuring complexity, we must examine interdependencies between the domains in a firm’s

the time period during which the firm is under- consideration set. Since the exact level of inter-

going change. We therefore calculated an average dependencies is unknown, we use a proxy mea-

of this measure for each firm over the previous sure that is likely to be a close approximation of

six years that correspond to the two time periods the true level of interdependence. If two domains

for which we calculate change: are interdependent, a past search is likely to have combined searches across them, and the domains

C i , t −n would thus be co-listed on a patent. We used all

patents in the patent database to construct this mea- (4) sure of interdependence. Since this data is very

Although our measure of complexity is based comprehensive, past searches should be a reli-

on that of Fleming and Sorenson (2001), there able indicator of the presence of interdependence

is a difference in the way it is constructed between domains. While past searches may be less

owing to a difference in what is being measured. accurate about which two domains are interdepen-

They measure the interdependence of the domains dent, they would be more accurate at an aggregate

in a patent as the number of subclasses in a level about how many other domains are interde-

tk , whereas we stop at calculating E tk pendent with a focal domain. Accordingly, count-

for a technology subclass. A minor difference ing the number of other domains that have been

between the two measures is that Fleming and co-listed with the focal domain can be used to

Sorenson (2001) use data from the previous measure that domain’s interdependence, or what

10 years, whereas we use data from the previous Fleming and Sorenson (2001) call the potential for

five years to maximize the time period for our recombination (E tk ).

study. Also, using subclasses for the complexity Our estimate of E tk was based on the measure

measure and classes for the coupling measure does developed by Fleming and Sorenson (2001). To

not lead to bias because what is being measured maintain consistency with their measure, we mea-

is different in these two cases: one is a measure sured E at the technology subclass level rather

of the environment, and the other is a measure of

tk

than at the class level. Since we used all patents

coupling at the firm level.

in the patent database, it is an exogenous measure, We incorporated patent information from 1976 unlike coupling, which is endogenous to the firm.

to 2004. Because we require patent data from We calculated it as the number of other subclasses

the previous 11 years (five previous years for with which a subclass had been combined (i.e.,

the complexity-related variables and six previous co-listed in a patent) in the previous five years

years over which we are calculating change), the divided by the number of patents assigned to that

first year in our panel is 1987. Similarly, the need subclass in the same period:

to observe at least five years’ worth of forward citations led us to calculate the effect of inventions only for patents that had been granted in or before

count of subclasses previously 1999. Thus, we have an unbalanced panel data set combined with subclass k

E tk = for 1,750 firms over a 13-year period. count of previous patents in subclass k

We then calculated the domain complexity (C i,t )

Control variables

of the technological environment in which firm i

A firm with a large knowledge base can experiment engages in inventing during year t as a weighted

with more recombinations and thus may be more measure of the potential for recombination of

successful at technological search (Kogut and Zan- each of the technological subclasses in which

der, 1992). We therefore controlled for Size of the firm has patents. The weights (g itk ) for each

knowledge base , measured as the total number of subclass are the fraction of patents held by focal

patents for which the firm had applied (and subse- firm i in each technological subclass k . Thus,

quently been granted) in the previous three years. we have

Previous work suggests that technological diversi- fication can have a positive effect on innovative

Domain complexity = C i ,t =

g itk × E tk (3)

capabilities (Garcia-Vega, 2006; Quintana-Garc´ıa

388 S. Yayavaram and W.-R. Chen and Benavides-Velasco, 2008). In addition, it may

two classes because their degree of relatedness influence the level of coupling between classes

is known to be high (low). To control for this and the way in which such coupling changes over

effect, we created two control measures based time. We therefore controlled for Technological

on the level of relatedness between classes in a diversification based on the Herfindahl index of

firm’s knowledge base. The relatedness between concentration:

technology classes j and k , R j−k ,t − 3 to t − 1 , can

be calculated as

Technological diversification i ,t = 1−

f itk 2 ,

n jk

, (7) where f is the fraction of firm i ’s patents that are

in patent class k during the t -3 to t -1 period. We also controlled for Use of new knowledge

where n j is the number of patents in the patent domains , since adding a new domain to a knowl-

database assigned to class j , n k is the number of edge base may have an effect on the inventions that

patents assigned to class k , and n jk is the number of are generated (Ahuja and Lampert, 2001; Katila

patents assigned to both classes. In calculating this and Ahuja, 2002; Rosenkopf and Nerkar, 2001).

relatedness measure for each firm, we excluded Our measure for this variable was based on a firm’s

that firm’s patents.

technological position (Jaffe, 1986) in year t . Use Relatedness existing-existing was measured for of new knowledge domains was measured as the

each firm as the weighted sum of relatedness sum of the fractions for all technological classes

between two classes that (1) existed in both time k that were new to the firm’s knowledge base in

periods, and (2) were coupled with each other in year t compared with year t -3:

either the early or later period, but not in both. This measure thus allowed us to control for the relatedness of class pairs in which a “tie” had been

Use of new knowledge domains = i, t

f itk , (6)

either added or dropped between the two classes. The weights are the same as those used in the

where k is a technology class in which the firm has coupling change measures. Relatedness existing- patents in years t -3 to t -1, but not in t -6 to t -4.

new was measured for each firm as the weighted Because the number of citations that a patent

sum of relatedness between two classes, only one receives varies significantly across technology

of which existed in the first time period, whereas classes, a firm may receive more citations simply

both existed and were coupled in the second. because it generates patents that belong to more

These measures allowed us to control for changes popular classes. To control for such cross-class

that may have been driven by the exogenous differences, we created a Mean technology citation

technological environment. control , as used by Fleming and Sorenson (2001).

Assets (logged) were used as a measure of Firm For each patent with issue date t , we considered

size to control for the effects of scale and scope all technology classes to which the patent had been

on technology search and also for inertia in large assigned. For each such class, we calculated the

firms, which may make a knowledge base rigid. average number of citations that patents in that

Firm R&D intensity is a measure of the inputs to class had received in the five-year window up

the technology search process (Chen and Miller, to t (t-5.5 to t-0.5 ). We then weighted the term

2007). Firms that invest more in R&D generate for each technology class to which the patent had

more inventions, and it is thus necessary to control been assigned by the proportion of assignments

for this input measure. We measured R&D intensity to that class. Note that the window used to

as R&D expenditure divided by net sales. We consider the set of previous patents varied from

included Firm past performance as a control patent to patent. Finally, we summed this measure

variable because prior performance affects search for all patents for which a firm had applied in

behavior (Greve, 2003). Firm past performance each year.

was measured as a return on assets. Because we Changes in knowledge coupling may also occur

considered patents over the previous three years in response to the external environment that is

for each firm, we also used averages over the common to all firms in the population. A firm may

past three years for the preceding control variables. increase (decrease) the level of coupling between

Prior research shows that acquisitions can have

Changes in Knowledge Couplings and Innovation Performance 389

a significant effect on innovation performance variation in technological classes. Several of the (Ahuja and Katila, 2001). Hence, we included

strong correlations (such as Technological diver- Number of prior acquisitions as an additional

sification and Relatedness existing) are an effect control. Using data from the SDC Platinum

of knowledge base size. Larger organizations have database, we measured this variable as the number

greater scope for change than smaller organiza- of acquisitions made by the focal firm in the

tions, and hence it is important to control for previous three years. Year dummies were included

knowledge base size. 10 Table 2 presents the results to control for any time-varying effects, such as

of hypothesis testing for firm innovation perfor- trends in patenting or citation rates, and for the

mance. Model 1 in Table 2 shows the results for longer time period over which patents from earlier

the control variables. In Model 2, we introduce years could be cited compared to those from later

the Change in coupling among existing knowledge years.

domains variable, and the results indicate that it has a negative effect on innovation performance. Hypothesis 1b is thus supported. In Model 3, we

The model

include the interaction term with Domain com- Because our dependent variable is a count vari-

plexity . While the coefficient for the change vari- able with a high degree of variance relative to

able is negative and significant, the coefficient for its mean, we used negative binomial regression

the interaction term is positive and significant. In analyses with random effects. We chose to use

Model 4, we introduce the Coupling between new random effects rather than fixed effects because

and existing knowledge domains variable, and it our key independent measures are change-related

provides support for Hypothesis 2, which predicts measures. Fixed effects estimation is based on

that coupling between new and existing knowl- within-firm changes in variables, which in this

edge domains leads to inventions of greater value. study would imply changes in changes. Because

In Model 5, we include the interaction term with the changes that occur in a firm’s knowledge base

Domain complexity . While the coefficient of the may not vary sufficiently over time, fixed effects

change variable is positive and significant, the estimation may be inappropriate. One concern with

coefficient of the interaction term is negative and random effects models is the problem of unob-

significant at the 10 percent level. In Model 6, we served heterogeneity. To address this issue, we

include all of the variables. used the pre-sample fixed effects approach (Blun-

Since the coefficients of interaction terms in dell, Griffith, and Van Reenen, 1999). Pre-sample

nonlinear models do not represent the true interac- estimators address the problem of unobserved het-

tion effects (Hoetker, 2007; Wiersema and Bowen, erogeneity by generating an additional variable

2009), we use graphical analysis to interpret the from pre-sample data rather than using a within

interactions in Model 6. We plot the margins for estimator (Blundell et al., 1999). In this estimation,

the entire range of values of Change in coupling we used data from 10 years before the study period

among existing knowledge domains and two differ- to create a pre-sample value of the dependent vari-

ent values—high and low—of domain complexity able, Mean innovation performance T-1 to T-10 and

in Figure 2a. This figure shows that Change in then used this pre-sample variable as an addi-

coupling among existing knowledge domains and tional regressor. To build this variable, we took

innovation performance are negatively related at the mean of innovation performance for each

low levels of domain complexity and positively firm over this 10-year period prior to the study

related at high levels of domain complexity. We period.

plot the average marginal effects of Change in coupling among existing knowledge domains for various values of Domain complexity in Figure 2b.