Structure of a firm’s knowledge base and the effectiveness of

technological search

Sai Yayavaram

National University of Singapore NUS Business School

BIZ 1 Building, 1 Business Link Singapore 117592

(65) 6874-3154 bizsy@nus.edu.sg

Gautam Ahuja University of Michigan

701 Tappan Street, Ann Arbor, MI 48109 (734) 763-1591

gahuja@bus.umich.edu

September 30, 2005

Structure of a firm’s knowledge base and the effectiveness of

technological search

ABSTRACT

In this study, we address the issue of how firms should structure their knowledge bases. The structure of a

firm’s knowledge base refers to the pattern of combinative relationships or couplings between the

elements of the knowledge base. We make a distinction between coupling, a search related decision made

by a firm on which knowledge elements should be combined and interdependence, the inherent

relationship between knowledge elements. We argue that this distinction provides additional insights into

the factors that affect technological search. Specifically, we investigate the proposition that, in contexts

where interdependencies are pervasive, a nearly decomposable structure increases the usefulness of

inventions by mitigating the effects of computational complexity associated with technological search and

improving our knowledge about interdependencies. We also investigate the proposition that a nearly

decomposable structure makes the knowledge base malleable by increasing the absorptive capacity of the

firm. We find support for our hypotheses in the context of the global semiconductor industry.

2

Introduction

“The aim of science is not things themselves…but the relations among things; outside these

relations there is no reality knowable” - Henri Poincare in Science and Hypothesis (1905: xxiv)

Questions of structure and design have occupied a central role in organization theory. For instance, in the

past and on a continuing basis, organization theorists have wrestled with problems of organization design

and structure and tried to understand the implications of alternate organizational structures for

organizational outcomes; in the process creating a wealth of understanding of how organizations function.

More recently scholars have focused on developing theories of network structure and problems of

network design to improve our understanding of how intra and inter-organizational networks function and

evolve, again a question of central import. In a related fashion, the significantly increased importance of

knowledge as a central asset of firms and technology (and its dynamism) as a significant arbiter of firm

fortunes suggests that a third such “almost paradigmatic” question seems to be emerging: how should

organizations structure their knowledge-bases?

This question of knowledge-base architecture or design is important from both conceptual and

practical perspectives. Conceptually, prior research suggests that organizational knowledge may hold the

key to organizational adaptation in a technologically dynamic environment, yet there is limited work

identifying the core attributes on which organizational knowledge-bases might differ and the implications

of these differences (Ahuja and Katila 2001, Rosenkopf and Nerkar 2001). Identifying such attributes

and studying their implications is key to improving understanding of why organizational responses differ

in the face of change and are differentially successful in their final outcomes with respect to it. In

practical terms the central and ascendant role of knowledge in organizations specifically, and society

more generally, suggests that the ability to prosper for not just firms but economies may be in part be

linked to their ability to develop new knowledge or adapt existing knowledge to changing conditions.

In this study, we take one cut on this question, trying to address the issue of how firms should structure

their knowledge bases (Schilling and Phelps 2004). We identify one basic attribute of knowledge-base

3

structure, its internal connectivity in terms of ties between knowledge elements, in the context of one

industrial environment, one characterized by high dynamism and pervasive interdependence, and explore

the implications of this knowledge base structure for two important outcomes, the utility of inventions

generated by the knowledge-base and the knowledge-base’s own malleability or capacity for change.

A firm’s knowledge base has often been conceptualized as a set of elements or components, for

instance, by placing the firm’s patents (Jaffe 1989, Ahuja and Katila 2001, Rosenkopf and Nerkar 2001),

its R&D expenditures (e.g., Helfat 1994) or its human resources (e.g., Chang 1996) into various

categories, where each category represents an element of knowledge. In this study we begin by suggesting

that it is more appropriate to conceptualize a firm’s knowledge base as a network of knowledge elements

(Schilling and Phelps 2004), where the knowledge of a firm is embodied not just in the elements

themselves (the nodes), but also in the combinative relationships or couplings (ties) between these nodes.

We argue that the structure of the network of knowledge elements can guide the process of

recombinatory search for new inventions, and thus directly affect the utility of such inventions. Further,

prior research has pointed out that firms may find it difficult to learn novel architectural knowledge i.e.

knowledge about the links between product components (Henderson and Clark 1990) and this can lead to

failures in adaptation. We suggest that by choosing an appropriate structure or pattern of relationships

between knowledge elements firms can increase the absorptive capacity that pertains to knowledge about

relationships between elements. This in turn can increase the malleability or capacity for change in the

knowledge base itself.

We build on past research on adaptive complex systems (Kauffman 1993, Rivkin 2000, Fleming

and Sorenson 2001), complex networks (Barabasi 2002, Watts and Strogatz, 1998) and the literature on

technology search. In the literature on complex adaptive systems the notion of interdependence plays an

important role (Kauffman 1993, Levinthal 1997). Kauffman (1993) finds that as the number of

interdependencies in a system increases, it becomes more difficult to identify a configuration of

constituent elements of that system that is valuable. Hence, he suggests that a system with fewer

interdependencies is more conducive to adaptive evolution. Building on these ideas other researchers have

4

argued that in the presence of high levels of interdependence it is difficult for firms to understand the true

nature of relationships and convert their own experience into learning (Sorenson 2003), or for other firms

to imitate them successfully (Rivkin, 2000). Similarly, in the technology search literature prior research

has established that technologies are likely to evolve faster and more usefully if they are based on the

recombination of elements of intermediate interdependence (Fleming and Sorenson 2001). Further, an

intermediate level of interdependence between knowledge elements will also increase the advantages of

internal transfer of knowledge over transfer across boundaries, whether these boundaries are at the firm,

geography or technology level (Sorenson, Rivkin and Fleming 2003). Thus the notion that

interdependence between knowledge elements has important innovation consequences is established.

We build on this work and take it a few steps further. First, we make a distinction between

interdependence, and the related, but distinct, concept of coupling. Conceptualizing a firm’s knowledge as

a network of elements we argue that two types of relationships or ties can exist between knowledge

elements: interdependence and coupling. Interdependence is the degree to which two elements are related

to each other in the natural world. Thus, element a may be related to element b such that any actions on a

have an effect on the contribution or performance of b. It is a state of nature and not necessarily known a

priori. On the other hand, coupling is the extent to which search across two elements is combined by an

entity (e.g., a firm) that is conducting the search. Thus, it is the human decision on whether the elements a

or b will be jointly considered or used by the decision-maker. For instance, the decision-maker might

decide to always use both elements together and only look at them jointly (strong coupling), or she might

always consider them independently (uncoupled). From a network perspective, coupling of two

knowledge elements corresponds to forming a tie or combinative relationship between them. To

summarize, we suggest that the decision on which elements should be strongly coupled and which

elements should be weakly coupled or not coupled belongs to the made world or the artificial world

(Simon, 1996), the world of human designed artifacts. In contrast, interdependence exists in the natural

world, the world of laws of nature. Mostly, prior research has not distinguished between interdependence

5

and coupling. In this paper, we build on Fleming and Sorenson’s (2001) observation that they are actually

distinct concepts.

Recognizing that couplings reflect organizational decisions rather than a state of nature we argue

that even in a given technological environment with a set of common underlying interdependencies,

coupling choices can vary across firms. The choices that any firm makes on coupling are naturally driven

by the firm’s current understanding of the interdependencies between knowledge elements; hence

coupling and interdependence are related. However, even though the pattern of couplings between

knowledge elements needs to be similar to the pattern of interdependencies between the knowledge

elements for effective search, the match between a firm’s coupling choices and the underlying

interdependencies cannot and need not be exact for multiple reasons.

First, in a general sense, if interdependence is defined as the relationship between elements that

exists in the natural world, it is reasonable to assume that our knowledge about interdependencies is not

complete and will not be so in the near future. It follows then that one cannot achieve an exact match

between coupling and interdependence. Second, in a situation where interdependencies between

knowledge elements are pervasive, the context that we examine in this paper1, it is impractical for a firm

to conduct a search across all elements and consider interdependencies among all knowledge elements.

To avoid the complexity that results from coupling search across all elements, the firm has to make a

choice about what interdependencies it wants to focus on and ignore the rest. Since many real world

problems do not have clear or natural boundaries, it is up to the decision maker to define such boundaries

for search and ignore interdependencies that exist across the boundary of the problem as defined.

Consequently, while the patterns of interdependence may be the same (or given) for all firms, their

choices on coupling may differ. Some firms may repeatedly couple one set of elements with another

1 It should be noted that we are referring to interdependencies between knowledge elements and not product elements. Even though we can broadly decompose technological knowledge into broad domains such as electrical, mechanical, chemical, and biological and so on, within each domain relationships between elements are quite pervasive. Further, extensive relationships exist across domains as well as in electro-mechanical and bio-chemical. As such, contexts where interdependencies are pervasive are not uncommon.

6

while others may rarely do so. Thus, the coupling matrix for a given firm may be quite distinct from the

interdependence matrix facing all firms or the coupling matrices of other firms.

The level of coupling between two knowledge elements indicates the strength of the tie between

those elements and can be used to identify cluster or module membership in that organizational

knowledge-base. Essentially, all elements that have high coupling with each other can be identified as a

cluster. A level of coupling that is greater than zero between two clusters indicates that there is some

integration across the two clusters. Thus, the notion of coupling can be used to delineate a continuum of

possible structures, from complete modularity (i.e. no across cluster couplings exist) to complete

integration (i.e. no clusters exist), and enables us to study the effects of variation in organizational

knowledge-base structure from highly modular to highly integrated.

Cross-sectional variations in coupling choices across firms are meaningful in that they imply

variations in the structure of firms’ knowledge-bases, and variations in knowledge-base structure (we

argue subsequently) imply variations in the utility of the inventions generated from these knowledge-

bases. However, inter-temporal variations in a firm’s coupling choices are also meaningful and important

in their own right. A firm’s coupling of certain elements of knowledge reflects its implicit or explicit

assumptions about their underlying (but possibly unconfirmed or unknown) interdependence. Cumulated

over time and across elements these choices represent the firm’s knowledge architecture – the assemblage

of its beliefs on what elements can be fruitfully combined and those that it believes are disconnected (see

Baldwin and Clark, 2005 for a related exposition of design architectures). Repeated coupling (or

decoupling) of two elements implies reinforcement or institutionalization of a given belief. Indeed, over

time an organization’s information filters, communication channels and design strategies reflect its

knowledge architecture and become encoded into its structure (Henderson and Clark, 1990).

Yet, the very reinforcement that institutionalizes these beliefs and makes their usage efficient and

“taken for granted” may limit the evolution of the organization’s knowledge. For an organization’s

knowledge to evolve it is critical that its knowledge-base have the capacity to change the structure of its

couplings. As new technological possibilities emerge, new interdependencies are discovered, and old ones

7

rendered obsolete, the structure of existing couplings can become ineffective. The theory of absorptive

capacity suggests that knowledge-absorption depends on the possession of prior related knowledge and

the ability to form associative bonds between new and old knowledge (Cohen and Levinthal 1989). To the

extent that coupling structures in an organization become calcified, the ability of the organization’s

knowledge-base to evolve gets restricted. Indeed, from an organization’s perspective the discovery of new

interdependencies itself depends on organizations being able to experiment with changes in their coupling

structure. When the match between interdependencies and couplings is exact, search is essentially

restricted to the exploitation of current knowledge. A divergence between interdependencies which are

well understood and couplings is necessary to explore new combinations of knowledge elements.

Knowledge-bases that are exploring, evolving and absorbing new knowledge should demonstrate changes

in their patterns of coupling over time. Thus, in addition to the cross-sectional variation in knowledge-

base structures across firms, the intertemporal changes in an organization’s own knowledge base are also

significant. We describe a knowledge-base’s capacity for change as the malleability of the knowledge–

base.

Thus, in this paper, we highlight two implications of this coupling-interdependence distinction for

firm technology strategy. First, we propose that firms can address the problem of computational

complexity that emerges in a recombinatory invention process by choosing an appropriate structure for

their knowledge-bases. We focus on decomposability in couplings (and not interdependencies) and argue

that knowledge-bases that are nearly decomposable (Simon 1962), relative to more integrated or more

modular knowledge-bases, can lead to outcomes such as enhanced innovation quality in contexts where

the interdependence matrix itself is not decomposable. We suggest that having a nearly decomposable

structure for a knowledge base increases the effectiveness of search even for problems (i.e.

interdependencies) that do not have a decomposable structure. In the second contribution of this paper we

develop arguments on how coupling plays an important role in improving our knowledge of

interdependencies, which in turn leads to changes in coupling over time. Thus, in this study, we examine

8

a) the effect of knowledge-base structure on the quality of inventions generated from it, and b) the role of

knowledge-base structure in making the structure itself malleable, i.e. capable of change.

A Model of Technological Search

A technological invention can be seen as the outcome of a recombination of existing knowledge elements

(Schumpeter 1939, Fleming 2001). For example, a new type of storage device that is being developed by

IBM is based on punching holes using a scanning tunneling microscope which was invented in 1981 and

whose initial purpose was to produce images of single atoms (Chang 2002). Like computer punch cards,

an early form of computer storage, this new system also stores information in holes, though of a much

smaller size. Thus, recombining one element of knowledge, holes to store data, with another element,

scanning technology – originally meant to produce images, resulted in a new invention.

However, the process of creating inventions is not simply a matter of recombining knowledge

elements. For even small knowledge bases the large number of potential combinations of existing

elements can lead to a combinatorial explosion of the space of possible inventions (Fleming and Sorenson

2001). For instance, even with 100 elements, taking just the simplest combinations of 2 elements at a time

can lead to over 4,500 possible combinations. Considering other sets of combinations (3 elements at a

time etc.) worsens the magnitude of this problem. We argue that coupling or combinative relationships

between knowledge elements can lead to useful inventions by simplifying the problem of searching in this

large design space.

In a recombinant model of innovation search for useful inventions can be represented as the

problem of identifying combinations or configuration of knowledge elements that lead to higher payoffs

or utility (Kauffman et al 2000, Fleming and Sorenson 2001)2. Starting with a vector that represents a

given configuration of N knowledge elements, a firm can generate a new configuration by changing the

2 There is one important difference between Kauffman et al’s (2000) model and Fleming and Sorenson’s (2001) approach. Fleming and Sorenson (2001) assume that a technology landscape is associated with each set of components. Addition of another component shifts the search to a different landscape. Kauffman et al (2000) define a landscape as the space of technological possibilities where each point represents an invention. In this paper, we follow Kauffman et al’s approach to the problem.

9

states of m elements, where can range from 1 to N. For example, when a new microprocessor is

introduced, all the other components in a computer including software, memory, storage capacity and

interfaces need to undergo a change. When is small, the move can be regarded as local and the new

configuration chosen for trial is located close to the original configuration on the technology landscape;

when is large, the move is a long jump and the new configuration is at a distance. Prior research

suggests that, typically, m is small as firms restrict themselves to local search (Helfat 1994, Stuart and

Podolny 1996). It is important to note that, in addition to a long jump, a series of local moves can also

lead to exploration of distant configurations.

m

m

m

After generating a new configuration, the firm can either accept or reject this change using a

specified criterion, for instance, that the value or utility of the new configuration should be higher than the

value or utility of the old configuration. After making an initial change, the firm can consider another

configuration again. In a world of interdependent knowledge elements, when the state of an element is

changed, it affects the payoffs of other elements that are dependent on the first element. The firm can

possibly move to a better configuration when changes are made in the dependent elements that are

congruent with changes in the first element. This process can continue further where changes in the

second element leads to changes in a third element and so on, setting off the firm on what has been called

an adaptive walk (Kauffman et al 2000). This adaptive walk ends when the firm reaches a useful

invention. When the search process leads to a new configuration that is close to the original configuration,

it can be characterized as exploitation and when the search process leads to a new configuration that is

distant or involves boundary spanning (Rosenkopf and Nerkar 2001), it can be characterized as

exploration. Thus, to increase exploration it is important to ensure that such adaptive walks are not cut

short.

High levels of interdependence in a system can severely curtail the length of adaptive walks as

well as the likelihood of finding highly useful new configurations as managing the complexity in such

systems is prohibitively difficult (Kauffman 1993). This has led researchers to suggest that

interdependencies should be tuned based on the outcome that is desired. Interdependence should be

10

reduced to achieve robust designs and increased to achieve higher levels of exploration (Levinthal and

Warglien, 1999). We depart from this logic and argue that it is more appropriate to accept

interdependence as something that is given and cannot be changed (as Levinthal and Warglien (1999:

343) themselves note), i.e. it represents a state of nature. Further, it is often likely that interdependence

will not be known a priori and uncovering interdependencies or relationships that exist in the natural

world can itself be thought of as a key goal of the search process.

Thus, while interdependence is not necessarily under the control of the organization, its coupling

choices are. The organization can decide which of a set of elements it will choose to consider jointly, and

which it will prefer to treat as disconnected. In the context of our model of recombinant search we relate

the notion of coupling to the criterion that is used for accepting or rejecting a new configuration that is

generated in the search process. One can conceive of various levels of aggregation in determining the net

effect of a change in a configuration of elements. For instance, at one level one could argue that any

change in the state of an element is accepted only if it increases the value of the system as a whole. An

alternative criterion for guiding the search could be to accept a new configuration based on the payoff of

an element or a group of elements rather than that of the entire system. In such a context coupling can be

defined as the weight attached to the value contributed by another element in deciding the state of the

focal element3. Or alternately, coupling is the extent to which the decision-maker considers the

contributions of another element in deciding the state of the focal element i.e. it is the extent to which

search is combined across the two elements. When coupling between elements is high or close to 1, the

search for improved value at one element is closely tied to the search for improved value at the other

element, i.e. the decision maker treats the two elements jointly. In contrast, when coupling between two

elements is very low or zero then the search for higher value at one element is independent of the search

3 Rivkin and Siggelkow (2003) also define a similar search criterion, but use it to model incentives. Instead of L, they define a parameter INCENT. When INCENT = 0, a manager in charge of a group considers the payoff for only his or her group. When INCENT =1, a manager gives equal weight to the rest of the system as well.

11

for higher value at the other element. For example, the development of software can be coupled to a

specific operating system or it can be decoupled from the operating system.

We focus on systems that have pervasive and non-decomposable interdependencies and the role

that one type of structure of couplings viz., a nearly decomposable structure (Simon 1962) plays in such

systems. A system is said to be nearly decomposable if it consists of clusters that have a large number of

interactions within themselves along with weak interactions across clusters. The literature on modularity

(Baldwin and Clark 2000, Schilling 2000) and the literature on loose coupling (Weick 1976) suggest that

a structure that is nearly decomposable has certain desirable properties such as potential for

recombination, persistence and adaptability. We argue that near decomposability in the structure of a

firm’s knowledge base leads to the appropriate balance between exploitation and exploration (March

1991, Levitan et al 1999) that is required for effective technology search. It should be noted that here we

are referring to near decomposability in the coupling matrix that represents a firm’s knowledge and not to

near decomposability in the interdependence matrix and also not to near decomposability in products or

organizational units. We next develop hypotheses that elaborate on the role of near decomposability in

coupling, which we then empirically test in the context of the global semiconductor industry4.

Hypotheses

Structure of a firm’s knowledge base

Near decomposability improves the outcomes of the search process for the following reasons. First,

modularity implies that the firm is no longer searching across all its elements simultaneously. The search

process can be divided across various clusters, which has the advantages of both simplicity within a

cluster and simultaneity across clusters. Dividing the knowledge base into clusters can also be seen as the

process of differentiation (Lawrence and Lorsch 1967) or using the splitting operator (Baldwin and Clark

2000). It is easier to search for good configurations for the elements within the cluster because of the

reduced number of elements under consideration.

4We present the results from the simulation study of the above model in a related paper.

12

Second, while splitting the system into clusters has a large number of advantages, prior theory

strongly suggests that differentiation should be accompanied by integration (Lawrence and Lorsch 1967,

Baldwin and Clark 2000). This can be accomplished by providing low coupling or integration5 between

clusters. This makes the matrix of couplings between the knowledge elements nearly decomposable –

coupling within clusters will be high compared to the coupling or integration between clusters. When

integration exists between clusters of elements then the search for better configurations in one cluster is

partly tied to the search for configurations in the other cluster. This low to intermediate level of

integration has a very important advantage of leading the firm to exploration, as discussed next.

The coupling of search across two clusters implies that when inventors consider any changes in

their own cluster, they take into account the effects on the other cluster as well. The negative side of

considering the effect on elements in the other cluster is that inventors would then tend to make only

those changes in their own cluster that will have a positive effect on the second cluster as well. This

would reduce the number of possible moves they consider because many moves that have a positive effect

on the first cluster may have no effect or even a negative impact on the second cluster. Consequently, they

tend to resist exploratory moves that will take them away from their current locally optimal configuration.

Modularity fosters exploration by reducing the threshold to be considered in adopting changes:

improvements in only the local cluster may be sufficient to induce a move from the firm’s current

configuration, relative to an integrated structure where the threshold for moving may entail simultaneous

improvements in two or more clusters. While it is possible that some moves in one cluster can have a

negative effect on the elements in the second cluster, the advantage is that the firm moves away from its

current configuration. The changes made in the first cluster can then lead to changes in the second cluster

and these changes can result in further changes in other clusters and so on. The adaptive walks then span

the boundaries of clusters and such boundary spanning leads to exploration (Rosenkopf and Nerkar 2001).

So, firms can explore distant neighborhoods even through local search when modularization is present.

5 We use the term integration to refer to coupling between clusters of elements.

13

When modularization is absent, local search pulls the firm back to the starting locally optimal

configuration, its competency trap (Levitt and March 1988). However, when modularization is complete

and there is no integration between clusters, any changes made in one cluster may have an effect on

elements in other clusters, but technical personnel in the other cluster may not be aware of these effects

and the opportunity for creating inventions is wasted6.

In a nearly decomposable structure it is possible to balance the harmful effects of complete

modularity or complete integration. When an intermediate level of integration between clusters exists

changes in one cluster that are highly deleterious to other clusters will not get adopted. Further, at such

intermediate values of integration between clusters, decision-makers are aware of the interaction effects

across clusters (Glassman 1973) but can strike a balance between ignoring the effects that occur in other

clusters and giving them too much importance. So, changes that cause the firm to move away from its

current configuration to new neighborhoods are still accepted. Thus, at an intermediate level of

integration between clusters, it is possible to engage in both exploitation and exploration. When both

exploitation and exploration exist, the outcomes of the search process are superior (March 1991) i.e. more

useful inventions are generated. Further, joint search between the two clusters also creates the absorptive

capacity that is required for transfer of knowledge from one cluster to another.

Third, as discussed before, firms are not aware of the true underlying interdependences and an

important aim of technology search is to uncover these interdependencies. Splitting into clusters has the

advantage that it makes the task of uncovering interdependencies easier. Splitting into clusters is

beneficial even when interdependencies are pervasive because it is easier to observe the relationships

between a set of variables when the effects of other variables are not considered. Research in cognitive

psychology shows that restricting the scope of a problem can make it easier to notice the unexpected

(Goldenberg et al 1999) and uncover the “true” underlying relationship within the cluster. Obviously, the

“true” relationships between the entire set of variables would be better understood if it were possible to

6 For example, a new technology developed in one cluster may have a new use in another cluster but without integration between the two clusters knowledge transfer will not take place.

14

handle all the variables simultaneously. However, computational complexity implies that inventors have

to set a boundary at some stage or the other and it is this boundary that results in the formation of clusters.

Integration between clusters also has a positive effect on discovering the “true” level of

interdependencies. This would happen if an interaction between two clusters previously assumed to be

weak turns out to be more important. A firm is more likely to be aware of this interaction and have the

absorptive capacity to acquire knowledge of this interaction if it has provided for this interaction in its

coupling matrix. Hence, coupling between clusters makes it easier to uncover interdependencies that exist

across clusters and thus ensures that the firm is able to redefine the partitioning of its knowledge system

when necessary. Thus, firms that are able to specialize and develop knowledge within clusters, and use a

few linkages to integrate knowledge across clusters are likely to be more successful at generating useful

inventions; such firms are better able to balance exploration and exploitation, and more likely to have a

better understanding of interdependencies.

If the knowledge-bases of all firms have a nearly decomposable structure and if all achieve the

optimal level of coupling, then there would be no performance differences between firms with respect to

their technological search. However, firms are likely to err on both sides and choose higher than optimal

levels of differentiation or integration. Consequently, across the range of firms, one can expect to see

firms that have got it right (and reap the benefits in terms of generating useful inventions) and others that

have too much or too little integration. This leads to the following hypothesis:

Hypothesis 1: The usefulness of a firm’s inventions is related to the level of integration between its clusters of knowledge in a curvilinear (inverted-U shaped) manner.

Changes in a firm’s knowledge base

Prior research shows that firms engage in local search in the space of possible states of the elements that

they recombine (Stuart and Podolny 1996). For instance, Henderson and Clark (1990) point out that

changing the architectural knowledge of a firm is difficult because it is not easy to dramatically modify

the organization systems and information channels that embody the firm’s architectural knowledge.

Extending this logic to the structure of a knowledge base suggests that firms are likely to search locally

15

for new structures and restrict their search to the local neighborhood of the current structure of their

knowledge base. However, we argue here that near decomposability in the structure can increase the

malleability of the structure and provide the capacity for greater change.

Firms are likely to change the couplings they provide between knowledge elements in response to

changes in perceived interdependence or in the process of searching for previously unknown

interdependencies. Such changes are more likely to be an outcome of distant search rather than local

search. A key rationale (and advantage) of local search is that it covers the terrain the firm knows best,

one where the elements and their interdependencies are relatively well known and understood by the firm.

Naturally, the scope for discovering new interdependencies in this terrain is limited. In contrast, distant

search or exploration occurs when search cuts across cluster boundaries and leads to changes in the states

of elements across clusters. Some of these distant searches would be based on knowledge of

interdependencies that is relatively less complete. Hence, exploration has the potential for uncovering

new interdependencies, and thus the potential for restructuring of the firm’s knowledge base.

However, simply occasioning upon new interdependencies is not sufficient for the restructuring

of the knowledge base. For new knowledge to become an actionable and useful component of a firm’s

knowledge base, the said knowledge needs to be integrated and absorbed into the organization’s

knowledge base (Cohen and Levinthal, 1989). Thus, restructuring of an organization’s knowledge base

entails both, discovering new interdependencies, and having the absorptive capacity to recognize,

assimilate and eventually use this new knowledge in conjunction with the firm’s existing knowledge.

Recognizing these dual requirements suggests that firms with nearly decomposable knowledge structures

are likely to have a greater ability to restructure their knowledge bases than firms that have knowledge

bases that are completely modular or integrated.

The ability of firms with knowledge bases that have a highly modular structure to integrate

knowledge of new interdependencies into their exiting knowledge bases is likely to be weak. Absent the

mechanisms of cross cluster integration their absorptive capacity for this knowledge is likely to be lower.

Firms that have high integration between clusters can achieve the integration required to absorb

16

knowledge, but given their highly integrated structures they are less likely to foster the exploration that

results in the discovery of new interdependencies. To conclude, firms that have an intermediate level of

integration between their clusters are more likely to identify new interdependencies and integrate them

into their existing knowledge bases, suggesting that the ability of firms to accomplish changes in their

structure is also likely to vary with the integration between a firm’s knowledge clusters.

Hypothesis 2: The extent of changes that occur in the structure of a firm’s knowledge base is related to the level of integration between its clusters of knowledge in a curvilinear (inverted U-shaped) manner.

Methods

Sample

The empirical setting for this study is the worldwide semiconductor industry in the period 1984 to 1994.

A number of reasons motivate the choice of semiconductors as the setting for the study. First, the high

R&D intensity of the semiconductor industry implies that technology search is of considerable

importance in this industry. Second, the industry is also characterized by incessant technology change.

Since in Hypothesis 2 we examine the changes in a firm’s knowledge base, an industry that has witnessed

considerable technological change is an appropriate setting for the study.

The key dependent and independent variables are based on patents granted by the US patent

office (USPTO) and the citations made and received by these patents. While patent and citation based

measures have certain limitations, a number of studies have demonstrated their validity as measures of

invention (for example, Hall et al 2000). Further, since this study is limited to a single industry,

differences in patenting and citation rates across industries, a key problem with using patent data (Hall et

al 2001) is not present here.

Each patent provides information on the firm (or inventor) to which the patent has been assigned,

its year of application, and citations to previously granted patents. Further, USPTO assigns each patent to

a 3-digit technology classes. There are about 400 technology classes to which a patent can be assigned.

17

This classification system is being continuously updated as the technologies continue to evolve. We used

the classification that was current on 31st December 1999 (Hall et al 2001).

As our study used patents for building key variables, firms that did not patent were excluded. We

identified 115 semiconductor firms across North America, Europe and Japan. Patents belonging to these

firms and their subsidiaries were identified. Information on subsidiaries was collected from Who Owns

Whom Corporate Directories. Data on control variables was obtained from Compustat, Japan Company

Handbooks and annual reports. Lack of data on firm size, performance and research intensity reduced the

sample to 84 firms. As discussed later, we also ran all our models on the entire sample of 115 firms while

excluding data on size, performance and research intensity.

Measures

Usefulness of inventions. Usefulness of inventions is measured as the number of citations that the firm’s

patents in year receive in subsequent years till 1999. Consistent with the literature, we use the date of

application for a patent and not the date when it was granted. The number of citations that a patent

receives is a significant predictor of the value of a patent (Harhoff et al 1999, Hall et al 2001). We control

for number of patents that the firm has received in year t as the number of citations received is highly

correlated with the number of patents.

t

Independent variables. Several studies (for example, Jaffe et al 1993, Stuart and Podolny 1996, Katila

and Ahuja 2002) have shown that citations to patents can be used to study the search behavior of firms.

We follow a similar approach and examine the citations made by the firm’s patents. The coupling

between the knowledge elements of a firm’s knowledge base can be approximated by the recombinations

that exist in its patent portfolio. The underlying assumption is that if firms jointly search across two

elements, that is, when the two elements are coupled, they are likely to generate an invention that

repeatedly recombines these two elements. Reversing the logic, we consider that a repeated

recombination of two elements by a firm can be taken as an indicator that the two elements are highly

coupled in that firm’s knowledge-base.

18

A firm’s knowledge base or patent portfolio at is assumed to consist of all the patents that the

firm has accumulated during to

t

3−t 1−t years. To ease calculations, the technology classes that a firm

cites (i.e., the classes to which cited patents belong) are assumed to be the elements in the firm’s

knowledge base. The knowledge base in year (comprising of patents from t 3−t to t ) is used in

generating inventions or patents in year . The coupling between technology classes and k for firm i ,

can be calculated as7

1−

jt

13,, −−− ttotkjiL

13,, −−− ttotkjiL = )( cba

a++

where is the number of patents in which the two classes have been cited together, b is the number of

patents in which class is cited, but class

a

i j is not cited and c is the number of patents in which class j

is cited, but class is not cited. The coupling matrix, L i consisting of for all

pairs of elements represents the structure of the firm’s knowledge base.

i 1,, −− ttokj 3−t 13, −− ttotiL ,−kj

Figure 1 provides an illustration of the coupling matrix viewed as a network for a single firm

(Intel) in 1990. The density of this network is 0.15 while the density of a similar network (i.e. with the

same set of nodes or semiconductor related technology classes that are present in Intel’s knowledge base)

for the sample of firms as a whole is 0.47 and for the entire patent database is 0.81, providing some

evidence that interdependencies are pervasive in the semiconductor industry while couplings are not

pervasive when one considers technology classes as elements. As can be seen from Figure 1, the structure

of Intel’s knowledge base does exhibit near decomposability especially when we consider the strong ties

(thick lines in the figure)8.

To show that firms differ in how they structure their knowledge bases, we compared each firm’s

knowledge-base structure with that of the knowledge-base structure of the entire industry because any

firm differences that exist will be reflected in differences between individual firms and the industry. We

7 This ratio is known as Jaccard’s coefficient in cluster analysis (Everitt 1993: 41) 8 The empirical method for classifying ties as strong or weak is discussed later in the paper.

19

counted the number of mismatches assuming that a mismatch exists when a firm provides strong coupling

between two technology classes while the coupling between these technology classes is weak at the

industry level or vice versa. Given that firms differ in the technology classes that they use in a significant

manner, it follows that there will be many instances in which coupling between two technology classes

provided by a firm will be zero while coupling at the industry level is nonzero and hence our assertion

that firms differ will be trivially true. To make a more stringent claim, we consider only those technology

class pairs for each firm for which the firm has a nonzero coupling. Match between firm and industry is

measured as the number of technology class pairs for which there is a match between firm and industry

divided by the total number of nonzero couplings that exist in the firm’s knowledge base. Results show

that this measure is less than 0.7 for about 75% of the firms in the sample9.

We performed a similar analysis by comparing the knowledge-base structures of all firms with

the knowledge-base structure of a single firm (Intel). This comparison was performed only for those

technology classes for which both the firm and Intel had non-zero couplings. Again, we found that the

match measure was less than 0.7 for about 70% of the firms in the sample. These findings suggest that

firms do differ from each other in terms of how they couple technology classes and that coupling involves

choices made by the firm and is not entirely driven by technological imperatives or interdependencies

between elements. We next describe a procedure for quantifying the intuitive notion of near

decomposability which is measured here as the level of integration between clusters.

*********** Insert Figure 1 about here *****************

Our measure of integration between clusters in the knowledge base is a modification of clustering

coefficient, a concept used in the literature on complex networks (Watts and Strogatz 1998, Barabasi

2002). Clustering coefficient for an element or a node (i.e. a technology class) with ties is defined as ik

2)1( −×

=ii

ii kk

nCC ,

9 It should be noted here that we are looking only at patents that belong to semiconductor related technology classes.

20

where is the number of ties between the neighbors of node . The denominator is the maximum

number of ties that are possible between the k neighbors of node i and the numerator is the actual

number of ties that exist. Clustering coefficient for the network, , is CC averaged over all nodes.

in ik i

CC

i

i

As defined in the current literature, the definition of clustering coefficient does not take the

strength of a tie into account. The measure of integration between clusters that we built for this study is a

modification of clustering coefficient that takes into account the strength of ties between nodes. First, we

classify ties between nodes as strong or weak ties based on whether tie strength is greater or less than a

prescribed cutoff value. We will discuss shortly how this cutoff value is determined. Second, we identify

those nodes that can be considered as within cluster neighbors and across cluster neighbors for each node

in the network by a method that we discuss below. Then, for each node we calculate its integration with

neighboring nodes outside its cluster as

gggwhnIntegratio node

+−×+

=

2)1(

where is the number of neighboring nodes which are outside the focal node's cluster, is the number

of ties between neighboring nodes that are outside the focal node's cluster and

h w

g is the number of all

nodes to which focal node is connected such that 2

)1( −× gg is the maximum possible number of ties

between the nodes to which focal node is connected10. Finally, integration between clusters for the entire

network is measured as a weighted sum of the integration for each node, with the percentage of patents

that cite each node (i.e., technology class) as the weight.

We used the following procedure to determine the cutoff value for classifying ties as strong or

weak. Using a single value of coupling as a cutoff across all patent portfolio sizes is not appropriate as the

median level of coupling (among all couplings greater than zero) in a patent portfolio depends on the size

of the patent portfolio to a large extent. This decrease in the median value can be a true effect of size,

10 Integration for a node is assumed to be zero when it has no neighbors.

21

which implies that in larger networks nodes are relatively weakly connected as compared to smaller

networks. On the other hand, it can be an artifact of the measurement process. For each firm’s knowledge

network, one can assume that each class pair i-j will contain a certain proportion (pij) of the total number

of patents in the patent portfolio. So, the number of patents in each class pair can be calculated as number

of patents in patent portfolio × pij. This number will be small if number of patents in patent portfolio is

small or if pij is small. Since patents are count variables, any value less than 1 will imply that no patents

will belong to that class pair. Consequently, the number of nodes in a firm’s knowledge network will be

small when number of patents is small, even when pij is independent of size. In contrast, when the size of

the patent portfolio is large, the number of nodes is much larger. As the number of nodes increases, the

number of ties for each node increases, which in turn decreases the value of coupling between any two

nodes.

To remove this effect of size, we estimated the median value of coupling as a function of the size

of the patent portfolio and time and then assumed that this median value is the cutoff that is comparable

across patent portfolios of different sizes for classifying ties into strong and weak ties. A log-log or power

law relationship11 proved to be a better fit than linear, exponential or log relationship. The estimating

equation is

Log (median value) = 0.30-0.004×(year-1974)-0.53×log (number of patents in patent portfolio) (R-squared=0.93)

We used two alternative methods to identify which neighboring nodes can be considered to be

within cluster for each node. In the first method, we considered two nodes to be within the same cluster if

one of the following conditions was satisfied: 1) they have a strong tie between them and there is at least

one common node to which both are tied12, 2) they have a strong tie between them and both nodes do not

have any other ties at all., 3) they have a weak tie and there exists at least one node to which both are

11 The power law relationship suggests the intriguing possibility of scale-free behavior (Barabasi, 2002) and possibly, alternative explanations for the relationship between median value of coupling and size of patent portfolio. 12 A strong tie between two nodes that have neighbors but not any common neighbors is still considered as an across cluster tie.

22

strongly tied. All other nodes which are not in the same cluster as the focal node are considered to be

outside the focal node's cluster.

In the second and simpler method, we consider two nodes to be in the same cluster only if they

have a strong tie. We used the second method to test the sensitivity of the results to the first

conceptualization. It should be noted that in both these approaches, cluster relationships are not transitive.

That is, if node A and node B are in the same cluster and node B and node C are in the same cluster, then

it does not follow that node A and node C are in the same cluster. So, this measure results in "fuzzy"

clusters and avoids the problems associated with identifying precise cluster boundaries [Everitt 1993].

Extent of change in structure. For calculating change, our second dependent variable, we

compared the coupling matrix for each firm at time t with the coupling matrix at time t . This ensured

that there are no common patents between the two coupling matrices that are being compared. Change in

structure was measured as the weighted number of technology class pairs that had a significant change in

coupling between two time periods. It is important to consider only a significant increase or decrease in

coupling and not all increases or decreases so that minor random fluctuations do not influence the

measure of change. A significant change in coupling was defined as change in coupling that exceeds one

quartile. We estimated the 25th percentile and the 75th percentile value of coupling as a function of the size

of the patent portfolio and time.

3+

The coupling for each class pair can be placed in one of four quartiles in both the previous and

later time periods. We consider that a significant increase in coupling has occurred when the coupling for

a class pair changes from a) first quartile in the previous time period to the third or fourth quartile in the

later period or b) from second to fourth quartile; and a significant decrease has occurred when the

coupling for a class pair changes c) from fourth to first or second quartile or d) from third to first quartile.

Change in structure was measured as the weighted number of technology class pairs that had a significant

change in coupling between two time periods. The weight is equal to 22

''jiji pppp +

++ where )( ji pp

23

and represent the percentage of patents that belong to a technology class i at time t and )( ''ji pp )( j 3+t

respectively. We logged this variable as it can take on only positive values.

Control Variables. Citation frequency varies across classes independent of the usefulness of the

invention. This variance may be due to factors such as patenting propensity and technological

opportunity, which may vary across technological fields. So, a firm that patents mainly in classes that

have high citation rates can get spuriously high citation rates for its inventions. To control for this effect, a

technology mean control variable was defined. For each technology class, the average citation rate of all

patents in that technology class was calculated for each year (Fleming and Sorenson 2001). For each firm,

the technology mean control variable is defined as citation rate for each class × number of the firm’s

inventions that belong to that technology class in that year summed over all technology classes.

Number of firm employees is used as a measure of firm size. This variable would control for the

effects of scale and scope on technology search (Henderson and Cockburn 1996). It would also control

for inertia in large firms that may make a knowledge base rigid. The firm’s R&D intensity is a measure of

the inputs to the technology search process. Firms that invest more in R&D generate more inventions and

hence it is necessary to control for this input measure. R&D intensity is measured as R&D expenditure

divided by net sales. Both positive and negative effects of product diversification on the search process

have been discussed in previous literature. Product diversification can have a positive effect as it increases

the opportunities for using knowledge internally (Kamien and Schwartz 1982). It may also have a

negative effect because the top management team in a diversified firm may have a poor understanding of

R&D and therefore is less likely to invest in R&D (Hoskisson and Hitt 1988). It is measured as a dummy

variable with a value of 1 if a firm had businesses other than semiconductors. Again, both positive and

negative effects of firm performance have been hypothesized previously. While profitable firms may have

the slack to pursue exploration, they may have fewer incentives as compared to less profitable firms.

Similarly, firm performance may affect the malleability of a knowledge base in both positive and negative

ways. Differences may exist across countries in research productivity and patenting propensity. This was

24

controlled for by including a non-American dummy variable that was set to 1 for firms whose country of

origin was not in North America. Year dummies were included to control for any industry wide time

varying effects as well as any trends in patenting rates and citation rates.

Statistical analysis

Since the dependent variable for Hypothesis 1 is a count variable that has high variance relative to mean,

we used negative binomial regression analysis (Cameron and Trivedi 1986). For Hypothesis 2 we used a

linear regression model. For both the analyses we used fixed effect models as the data has a panel form.

Results

Table 1 provides descriptive statistics and correlations for all variables. Table 2 presents the results of the

hypothesis testing for the number of citations. Model 1 in Table 2 presents the results for the control

variables. Model 2 adds the variable integration between clusters and Model 3 adds the squared term.

Model 3 indicates support for Hypothesis 1 that integration between clusters of a firm’s knowledge base

has a curvilinear relationship with the number of citations received by a firm’s patents. The estimated

coefficients indicate that the maximum value is reached at a value (-1.57/(2*-2.75)=0.28) that is within

the observed range of the variable (0 to 1). Among the control variables, the number of patents and

research intensity are significant predictors of the number of citations as expected. Further, firm size and

firm performance increase the number of citations received while product diversification has a negative

effect. The year dummies showed that the number of citations increased in the early part of the time

period (1986 to 1988) for the study. This reflects the general trend in an increase in number of patents

made and citations received. In results not reported here, we counted only those citations that a firm

received from other firms and removed self-citations. Results were nearly identical. Using an entropy

measure of product diversification (Palepu 1985) instead of the dummy variable made no appreciable

difference to the results.

Table 3 presents the results for the change that occurs in the structure of a firm’s knowledge base.

Model 1 in Table 3 presents the base line case with control variables. Model 2 adds the variable

25

integration between clusters and Model 3 adds the squared term. Model 3 indicates support to the

hypothesis that the extent of change that occurs in the structure of a firm’s knowledge base is

curvilinearly related to the integration between clusters. The estimated coefficients indicate that the peak

is reached at a value of .32, which is within the observed range of integration between clusters. We also

ran all the models for both the hypotheses without including size, which allowed us to use the entire

sample of 115 firms. Results were largely unchanged. The alternative measure of integration between

clusters described in the methods section also led to the same results.

***************Insert Tables 1, 2 & Table 3 about here***********

Discussion

In this study, we addressed the question of how firms should structure their knowledge bases. We

investigated the continuum of structures from modular to integrated and examined the relationship

between the structure of a knowledge base and two key characteristics associated with the knowledge-

base, the utility of inventions generated from the knowledge-base as well as the malleability of the

knowledge base itself. The results of this study suggest that alternative forms or structures of knowledge-

bases have implications for the search process and the invention outcomes from those search processes.

Specifically, the study demonstrates that in the context of a dynamic technological environment

characterized by pervasive interdependencies across knowledge elements, nearly decomposable

structures, with dense ties within knowledge clusters and few ties between knowledge clusters, can lead to

a balance between exploitation and exploration and thus improve the search process relative to very

tightly integrated knowledge-bases or relatively modular knowledge bases, which have a tendency to limit

exploration. This balance between exploitation and exploration enables nearly decomposable knowledge

structures to generate more useful innovations.

Further, this study also introduces the notion of malleability as a relevant attribute of a knowledge

base and establishes that nearly decomposable knowledge-bases are more malleable, relative to tightly

integrated or relatively modular knowledge-bases. Problems of search and knowledge recombination are

26

inherently dynamic. A knowledge-base structure that is effective at one point of time may become less so

over time. New knowledge elements are discovered that need to be integrated within the knowledge-base,

new interdependencies may be discovered that require the updating of existing coupling relationships, or

changes in the external environment or sources of innovation may render some knowledge couplings

irrelevant, necessitating their elimination. Hence, it is important that the structure of combinative

relationships in the knowledge-base be adaptive with respect to these needs. This notion of the degree to

which a knowledge-base is amenable to change over time is what we describe as the malleability of the

knowledge-base.

It has been argued previously that firms are constrained to local search due to the cognitive

limitations of managers and lack of absorptive capacity that is required for long jumps; such local search

can lead to inertia. In this study we argue that firms can engage in search in distant neighborhoods or

exploration through a series of local moves. In the absence of near decomposability, these local moves

fail to take the firm away from its locally optimal configuration. However, an intermediate level of

integration between clusters ensures that local moves at the level of a cluster sometimes lead to adaptive

walks that span cluster boundaries and thus lead to exploration. While making such exploratory moves, a

firm improves its knowledge of interdependencies and thus is able to bring about changes in its

knowledge base. These mechanisms enable the identification of new interdependencies and facilitate

changes in the knowledge-base over time. Thus, nearly decomposable knowledge-bases are relatively

more malleable, enhancing absorptive capacity and permitting non-local search.

This inquiry is of theoretical significance from the perspectives of the nascent but fast developing

literatures on technology search and recombinant invention. The concept of a knowledge-base, or “what

the organization knows”, is central to such theories (Grant, 1996; Kogut and Zander, 1992). This study

suggests that from the perceptive of conceptualizing knowledge-bases, it is better to conceive of them as

networks of elements, where even the ties between knowledge elements are important, rather than as

simple sets where just the elements of the set matter. Thus, assessing the productivity of organizational

knowledge bases is not just a matter of delineating the elements of an organization’s knowledge-base or

27

“what the organization knows”. It is also important to understand how this knowledge has been structured

by the organization – what elements of knowledge are linked to what others for this particular

organization; when the organization is confronted with a search problem what combinations of

knowledge elements are simultaneously brought to bear on the issue, and what elements are jointly or

individually excluded. This notion of the criticality of knowledge architecture can be seen to be closely

linked to the idea of design architectures that have been regarded as critical to understanding innovation

but have hitherto been understudied (Baldwin and Clark 2005).

The notion of an organizational knowledge-base as a network of knowledge elements where the

ties between elements are important, rather than treating a knowledge-base simply as a set, is important

for multiple reasons. First, from the perspective of the technology search literature this conception

presents one resolution to an existing dilemma in the theory of recombinant invention. Past research has

argued that invention is often the result of a process of recombination of elements that constitute the

knowledge-base of the inventing entity (Ahuja and Katila, 2001; Fleming, 2001; Schumpeter, 1939). This

recombinant notion of invention however runs into the roadblock of computational complexity – even

moderately sized knowledge bases can have extremely large recombinant possibilities. One way of

solving this problem is presented in this study – firms can use coupling, the forming of associative

relationships between knowledge elements, to reduce the number of recombinant choices, and hence

simplify the combinatorial complexity of the problem. Consistently combining some elements and

thereby treating them as one reduces the number of combinations to be explored significantly.

Second this notion that organizational knowledge consists of nodes and ties raises the important

possibility that organizations can learn even without adding new elements to their knowledge-base. Thus

learning can occur without new knowledge being added to the organization, but simply through a

rearrangement of ties between existing elements of knowledge. This notion closely parallels a similar

notion emerging in the economics literature, that of fact free learning (AER, 2005). Thus organizational

knowledge can emerge through the addition of new elements or the addition of new ties, suggesting that

the exploration of new knowledge domains is perhaps not the only way to expand organizational

28

knowledge. Exploring the implications and relationship between these two knowledge enhancing

strategies, changing ties versus adding nodes, would be a fruitful direction of further research.

A third implication of this node-tie or network conceptualization of a knowledge-base emerges

from the notion of malleability that we develop in the paper. The idea that incumbents often find

architectural innovation difficult is well established in the literature (Henderson and Clark, 1990). Yet

some incumbents are able to make these transitions and survive architectural change while others are

unable to do so. Variations in the degree of malleability of organizational knowledge bases could be one

explanation for why some firms handle architectural innovations in products better than others. As

Huber (1991) notes, organizational learning occurs when there is a change in the repertoire of possible

responses from an organization. Thus changes in the structure of ties in a knowledge-base or knowledge-

base malleability may hold the key to understanding why some firms are able to draw upon enhanced

repertoires for reaction relative to other firms. Developing and testing the relationship between the

malleability of knowledge architectures and incumbent survival in the face of architectural change in the

product may be another useful direction of future research.

This study also benefited from and contributes to the nascent stream of literature using a complex

adaptive systems approach to understanding complexity in organizational contexts. First, we emphasize

the difference between coupling and interdependence at a conceptual level. Interdependence is what exists

in nature and needs to be determined. Coupling involves human decision making and has an important

role in improving our knowledge about interdependencies including discovering new ones. This

distinction between coupling and interdependence is also provided empirical teeth by our data. If coupling

were entirely decided by interdependence between knowledge elements, and there was no distinction

between the two, then all firms in the same industry should exhibit the same pattern of couplings. Yet we

observe that within the context of the semiconductor industry, individual firms vary considerably in their

decision to couple technologies and thus even given the same technological context the actual patterns of

combinative structuring differ across firms.

29

Our second contribution to the complex adaptive systems literature is that we explicate the role of

modularity in technology search. Buffering (Thompson 1967) or sealing off one module from another is

usually seen as a useful outcome of modularization because such sealing off is thought to prevent the

spread of problems (Weick 1976). For example, in software design, modularity has been seen as very

desirable because it reduces ripple effects. Similarly, modularization has been extolled in the product

design literature due its ability to limit ripple effects (Baldwin and Clark 2000). In this study, we argue

that choosing independent modules and avoiding ripple effects is not necessarily desirable in technology

search. An intermediate level of integration between modules initiates boundary spanning adaptive walks

that lead to exploration. While modularity can lead to exploration within a module, it limits exploration

across modules. Thus we join other recent studies that identify the dangers of excessive modularity such

as loss of competitive advantage (Fleming and Sorenson 2001) and losses due to the non-utilization of

synergistic potential (Schilling 2000).

In conclusion we note that the problem of structuring organizational knowledge represents a

significant frontier for organizational research, with immense and exciting possibilities. The theoretical

significance of this problem domain is only accentuated by the criticality, magnitude and centrality of the

role of knowledge in the modern corporation. We believe this study scratches but the surface of this

issue, limited as it is to one industrial context, that characterized by high dynamism and interdependence.

How appropriate knowledge structures vary across different industrial contexts, for instance contexts

characterized by more or less interdependence or dynamism, is an issue open for further study. Similarly,

understanding the relationship between knowledge architectures and other organizational outcomes of

interest, beyond invention utility and knowledge-base malleability is another potential direction of fruitful

research.

30

References

Ahuja, G., R.Katila. 2001. Technological acquisitions and the innovation performance of acquiring firms: a longitudinal study. Strategic Management Journal 22 197-220. Baldwin, C., K.Clark. 2000. Design rules: The power of modularity. MIT Press, Cambridge, MA. Baldwin, C., K.Clark. 2005. Designs and design architecture: The missing link between 'knowledge' and the 'economy’. Harvard Business School Working Paper No. 05-052; Harvard NOM Working Paper No. 05-01. Barabasi, A. 2002. Linked: The new science of networks. Perseus Publishing, Cambridge, MA. Cameron, A., P.Trivedi. 1986. Econometric models based on count data: Comparisons and applications of some estimators and tests. Journal of Applied Econometrics 1 29-53. Chang, S.J. 1996. An evolutionary perspective on diversification and corporate restructuring: Entry, exit, and economic performance during 1981-89. Strategic Management Journal 17 587-611. Chang, K. 2002. A New System for Storing Data: Think Punch Cards, but Tiny. New York Times. June 11, 2002. Cohen, W.M., D.A.Levinthal. 1990. Absorptive capacity: A new perspective on learning and innovation. Administrative Science Quarterly 35 128-152. Everitt, B.S. 1993. Cluster Analysis (3rd ed.). Edward Arnold, London. Fleming, L. 2001. Recombinant uncertainty in technology search. Management Science 47 117-132. Fleming, L., O.Sorenson. 2001. Technology as a complex adaptive system: Evidence from patent data. Research Policy 30 1019-1039. Glassman, R.B. 1973. Persistence and loose coupling in living systems. Behavioral Science 18 83-98. Goldenberg, J., D.Mazursky, S.Solomon. 1999. Toward identifying the inventive templates of new products: A channeled ideation approach. Journal of Marketing Research 36 200-210. Grant, R.M. 1996. Toward a knowledge-based theory of the firm. Strategic Management Journal 17: 109-122. Hall, B.H., A.Jaffe, M.Trajtenberg. 2000. Market value and patent citations: A first look. Working Paper No. 7741, NBER Working Paper Series. Hall, B.H., A.Jaffe, M.Trajtenberg. 2001. The NBER patent data citation file: Lessons, insights and methodological tools. Working Paper No. 8498, NBER Working Paper Series. Harhoff, D., F.Narin, F.M.Scherer, K.Vopel. 1999. Citation frequency and the value of patented innovations. The Review of Economics and Statistics 81 511-515. Helfat, C.E. 1994. Evolutionary trajectories in petroleum firm R&D. Management Science 40: 1720-1747.

31

Henderson, R.M., K.B.Clark. 1990. “Architectural innovation: The reconfiguration of existing product technologies and the failure of established firm”. Administrative Science Quarterly 35 9-30. Henderson, R., I.Cockburn. 1996. Scale, scope, and spillovers: The determinants of research productivity in drug discovery. Rand Journal of Economics 27 32-59. Hoskisson, R., M.Hitt. 1988. Strategic control systems and relative R&D investment in large multiproduct firms. Strategic Management Journal 9 605-621. Jaffe, A.B. 1989. Characterizing the “technological position” of firms, with application to quantifying technological opportunity and research spillovers. Research Policy 18 87-97. Jaffe, A.B., M.Tratjenberg, R.Henderson. 1993. Geographic localization of knowledge spillovers as evidenced by patent citations. The Quarterly Journal of Economics 108 577-598. Kamien, M., N.Schwartz. 1982. Market structure and innovation. Cambridge University Press, New York. Katila, R., G.Ahuja. 2002. Something old, something new: A longitudinal study of search behavior and new product introduction. Academy of Management Journal 45 1183-1194. Kauffman, S. 1993. The origins of order: Self-organization and selection in evolution. Oxford University Press, New York. Kauffman, S., J.Lobo, W.G.Macready. 2000. Optimal search on a technology landscape. Journal of Economic Behavior & Organization 43 141-166. Klevorick, A., R.Levin, R.Nelson, S.Winter. 1995. On the sources and significance of interindustry differences in technological opportunities. Research Policy 24 185-205. Kogut, B., U.Zander. 1992. Knowledge of the firm, combinative capabilities, and the replication of technology. Organization Science 3 383-397. Lawrence, P.R., J.W.Lorsch. 1967. Organization and environment: Managing differentiation and integration. Harvard Business School Press, Boston, MA. Levinthal, D.A. 1997. Adaptation on rugged landscapes. Management Science 43 934-950. Levinthal, D.A., M.Warglien. 1999. Landscape design: Designing for local action in complex worlds. Organization Science 10 342-357. Levitan, B., J.Lobo, S.Kauffman, R.Schuler. 1999. Optimal organizational size in a stochastic environment with externalities. Working Paper No. 99-04-024, The Santa Fe Institute. Levitt, B., J.G.March. 1988. Organizational learning. In W.R. Scott (Ed.), Annual Review of Sociology 14 319-340 March, J.G. 1991. Exploration and exploitation in organizational learning. Organization Science 10 342-357.

32

Palepu, K. 1985. Diversification strategy, profit performance and the entropy measure. Strategic Management Journal 6 239-255. Rivkin, J.W. 2000. Imitation of complex strategies. Management Science 46 824-844. Rivkin, J.W., N.Siggelkow. 2003. Balancing search and stability: Interdependencies among elements of organizational design. Management Science 49 290-311. Rosenkopf, L., A.Nerkar. 2001. Beyond local search: Boundary-spanning, exploration, and impact in the optical disc industry. Strategic Management Journal 22 287-306. Schilling, M.A. 2000. Toward a general modular systems theory and its application to interfirm product modularity. Academy of Management Review 25 312-334. Schilling, M.A. and Phelps, C. 2004. Small-world networks and knowledge creation: Implications for multiple levels of analysis. Working Paper. Schumpeter, J. 1939. Business cycles. McGraw-Hill, New York. Simon, H.A. 1962. The architecture of complexity. Proceedings of the American Philosophical Society, 106 467-482. Simon, H.A. 1996. The Sciences of the Artificial (3rd Edition). The MIT Press, Cambridge, MA. Sorenson, O. 2003. Interdependence and adaptability: Organizational learning and the long term effect of integration. Management Science 49 446-463. Sorenson, O., J.W.Rivkin, L.Fleming. 2003. Complexity, networks and knowledge flow. Working Paper. Stuart, T.E., J.M.Podolny. 1996. Local search and the evolution of technological capabilities. Strategic Management Journal 17 21-38. Thompson, J.D. 1967. Organizations in action. McGraw-Hill, New York. Watts, D., S.H.Strogatz. 1998. Collective dynamics of small-world networks. Nature 393 440-442. Weick, K.E. 1976. Educational organizations as loosely coupled systems. Administrative Science Quarterly 21 1-18.

33

Table 1 Descriptive Statistics and Correlationsa,b

Variable Mean s.d. Min. Max. 1 2 3 4 5 6 7 8 9 10

1. Number of citations 684.19 1242.33 0 7633

2.

Number of patents 85.71 155.88 0 1042

0.88*

3. No. of patents in patent portfolio 202.16 369.01 1 2206

0.85* 0.97*

4. Integration between clusters 0.25 0.21 0 1

-0.31* -0.32* -0.33*

5. Technology control 671.92 1189.04 0 6569

0.96* 0.93* 0.90* -0.32*

6. Research intensity 0.14 0.47 0 7.97

-0.06 -0.06 -0.06 0.08 -0.06

7. Firm size (log) 8.51 2.40 3.09 13.29

0.60* 0.60* 0.61* -0.32* 0.63* -0.21*

8. Firm performance -0.07 0.94 -15.5 0.33

0.06 0.06 0.05 -0.03 0.06 -0.97* 0.19*

9. Product diversification .44 0.50 0 1

0.24* 0.34* 0.35* -0.22* 0.33* -0.15* 0.59* 0.09*

10. Non-American firm 0.27 0.44 0 1

0.21* 0.34* 0.34* -0.12* 0.33* -0.10* 0.33* 0.06 0.67*

11. Change in structure 24.47 21.53 0 179.2

0.27* 0.28* 0.27* -0.27* 0.24* -0.07 0.25* 0.09* 0.05 -0.05a n=639 firm years b All independent variables except for number of patents are lagged by one year * p<.05

34

Table 2 Results of Negative Binomial Regression Analyses for Number of Citationsa

Variable Model 1 Model 2 Model 3

Constant -3.97*** -3.87*** -3.85*** 0.30 0.33 0.33

Integration between clusters -0.16 1.57** 0.20 0.59

Integration between clusters squared -2.75** 0.89

Number of patents 6.4E-04** 6.4E-04** 6.9E-04** 2.1E-04 2.1E-04 2.1E-04

Research Intensity 2.32*** 2.30*** 2.50*** 0.51 0.51 0.50

Firm size, logged 0.52*** 0.51*** 0.49*** 0.03 0.03 0.03

Firm performance 0.99*** 0.98*** 1.09*** 0.24 0.24 0.24

Product diversification -0.13 -0.12 -0.16† 0.09 0.09 0.10

Non-American 0.48** 0.46** 0.47** 0.17 0.18 0.17

Technology control 2.5E-04*** 2.5E-04*** 2.5E-04*** 3.3E-05 3.3E-05 3.3E-05

Number of observations 617 617 617

Number of groups 82 82 82

Log Likelihood -2849.77 -2849.48 -2844.24

a All models include year dummies for which results have not been displayed. † p <.10 * p <.05 ** p <.01 *** p <.001

35

Table 3 Results of Regression Analyses for Change in Structurea

Model 1 Model 2 Model 3

Constant 1.41* 1.51* 1.55* 0.62 0.63 0.62

Integration between clusters -0.14 1.66** 0.16 0.48

Integration between clusters squared -2.59*** 0.65

Number of patents -1.3E-04 -1.4E-04 -4.7E-05 1.6E-04 1.6E-04 1.6E-04

Research Intensity -2.69** -2.72** -2.25* 0.97 0.97 0.96

Firm size, logged 0.12† 0.11† 0.09 0.06 0.06 0.06

Firm performance -1.18** -1.19** -0.99* 0.43 0.43 0.42

Product diversification 0.14 0.14 0.09 0.17 0.17 0.16

Non-American 0.41 0.39 0.40 0.79 0.79 0.77

Number of observations 527 527 527

Number of groups 84 84 84 a All regressions include year and class dummies for which results have not been displayed. † p <.10 * p <.05 ** p <.01 *** p <.001

36

Figure 1 Coupling matrix visualized as a network for Intel in 1990a,b

a Nodes represent technology classes, ties between nodes represent coupling between technology classes and strength of a tie represents level of coupling; darker lines indicate stronger ties; size of node is a measure of the number of patents that cite that technology class. b Isolates have been removed for illustration purposes.

37