Local R&D Strategies and Multi-location Firms: The Role of Internal Linkages

Juan Alcácer Harvard Business School

Morgan Hall 227 Soldiers Field

Boston, MA 02163 (617) 495-6338

jalcacer@hbs.edu

Minyuan Zhao Ross School of Business University of Michigan 701 Tappan St., E4616 Ann Arbor, MI 48109

(734) 647-6978 myzhao@umich.edu

First version: January 2006

This version: December 2010

Local R&D Strategies of Multi-location Firms: The Role of Internal Linkages

ABSTRACT

This study looks at the role of internal linkages in highly competitive clusters. Besides serving as a mechanism to source knowledge, we argue that strong internal linkages also represent tighter control over local innovation and higher levels of technological interdependency across locations, which allow firms to increase internalization and reduce the risk of knowledge outflow to competitors. Our empirical analysis of the global semiconductor industry shows that the industry leaders tend to intensify their internal linkages across location in response to the presence of direct market competitors but not to the presence of innovators in the same technological field. In addition, we find that such internal linkages are associated with more intra-firm knowledge flow and lower knowledge flow to competitors in the same cluster. Our results suggest that research in cluster innovation should take into account multi-location firms with linkages across locations, and their different responses to technological competition and market competition in the cluster.

KEYWORDS: technology clusters, knowledge spillover, internalization, appropriability.

1. Introduction

Marshall (1920) suggests that clusters reduce cost to co-located firms by providing convenient access to

skilled labor, specialized suppliers, and knowledge spillovers. More recently, clusters are viewed as

ferments for innovation in high-technology industries (Saxenian, 1994) and source of competitive

advantage for firms. Locations with large number of firms and research institutions engaging in

innovative activities next to each other facilitate knowledge flow through frequent interpersonal

interactions (Jaffe et al. 1993; Audretsch and Feldman 1996) and labor mobility (Almeida and Kogut

1999), providing a fertile ground for exchanging knowledge.

However, clusters are also spaces where competition for ideas occurs. Porter (1998) emphasizes the

“vigorous competition among locally-based rivals”. Shaver & Flyer (2000) argue that not all firms would

benefit from co-locating; leading firms may even lose more than what they gain from a cluster. Looking

at knowledge spillovers, a key input for innovation, Alcácer & Chung (2007) show that leading firms may

shy away from manufacturing clusters. While firms and the R&D community at large gain from the

knowledge flow in clusters, unintended knowledge outflows to competitors can weaken the competitive

edge of leading innovators and compromise their ability to appropriate value from R&D. Yet in industries

such as semiconductors, leading firms still flock to clusters and innovate in them. What enables leading

firms in high-tech industries to benefit from clusters without risking their technological edge?

We address this question so we may refine our understanding of innovation in clusters. Specifically we

examine whether firms respond to local competitive environments by strengthening their internal

linkages, and whether stronger internal linkages effectively increase knowledge internalization and reduce

local knowledge spillover to local competitors.

In section 2.1 we review the positive and negative aspects that leading innovators face in clusters

according to previous research. More importantly, we bring the insight, absent in the literature, that

leading players in clusters are often geographically dispersed organizations with intra-firm ties across

multiple locations. Large multi-location firms are known for their ability to mobilize and integrate

knowledge on a global basis (Bartlett and Ghoshal 1990). Thus, to understand firms’ local R&D

strategies, we must recognize that a firm located in a particular cluster may also be part of a strategically

integrated internal network. The innovation strategy of IBM in Cambridge, Massachusetts, for example, is

intricately linked with the company’s eight other R&D labs and hundreds of facilities worldwide. How

firms organize their R&D activities internally will affect the extent to which they can appropriate value

from local innovation in clusters. After surveying the appropriation mechanisms suggested in the

innovation and strategy literatures in Section 2.2, we conclude that, by treating firms in abstract, these

literature streams tacitly assume that knowledge appropriation is location-free.

Our arguments on how internal linkages lead to appropriability in a cluster are present in section 2.3. We

recognize the dual nature of internal linkages. In one hand, internal linkages across a firm’s

geographically dispersed units can improve knowledge absorption and integration (Gupta and

Govindarajan 2000), increase the absorption of external knowledge at dispersed locations (Lahiri 2010),

facilitate the transfer of local knowledge back to the parent firm (Frost and Zhou 2005), and improve the

overall quality of innovation (Singh 2007). On the other hand, these studies mostly focus on knowledge

sourcing and value creation while overlooking the adverse effect of intensive local competition on

knowledge appropriation, a critical component of firms’ R&D strategies in highly competitive clusters.

We argue that internal linkages are a reflection of interdependence across locations and that this

interdependency can foster appropriation. Since specialized and co-specialized complementary assets are

critical to the value of an innovation (Teece 1986), a firm can minimize its loss from knowledge outflows

by strategically increasing the interdependence between geographically dispersed R&D activities. This is

in line with the disaggregation strategy suggested by Liebeskind (1996) – isolation of related knowledge-

production or knowledge-use processes – as a mechanism to protect against knowledge outflows and

Zhao (2006), who finds that firms tend to develop disaggregated components in high risk countries and

then integrate them at the firm level. Section 2.3 ends with the presentation of three propositions that

guide our empirical setting. These propositions zoom in on a particular type of internal linkage, cross-

cluster teams, that was identified as key in a set of interviews with managers in the semiconductor

industry.

Section 3 describes our empirical approach. The multi-dimensional relationships among local entities

(Cohen 1995, 230) allow us to separate the appropriation incentives from knowledge sourcing incentives

among multi-location firms. Firms in a technological cluster may share similar technological backgrounds

or even engage in patent races, but they do not necessarily compete in the same product market. Industry-

specific market information and other complementary resources reduce the risks associated with

knowledge exchanges, allowing symbiotic relationships to develop. If internal linkages are purely

mechanisms of knowledge sourcing, we should observe stronger internal linkages in clusters with greater

knowledge pools, e.g., a large number of neighboring firms in the same technological field. If internal

linkages serve as a hedge against knowledge outflows, they should be used more extensively when

neighboring firms share the same product market.

Our empirical setting is innovation in the semiconductor industry. Specifically we analyze three

innovation traits generated by the top 16 innovators in 25 clusters. Section 3.1 describes our data sources

and section 3.2 our definition of clusters. We depart from previous research that defines clusters in terms

of employment density reported in country-specific pre-determined geographic units. Instead, we identify

clusters’ contours using a mathematical algorithm whose input is patent inventor locations. Sections 3.3,

3.4, and 3.5 describe our models and variables.

Examining the largest innovating firms in the global semiconductor industry from 1998 to 2001, we find

supportive evidence in section 4 for both the knowledge sourcing and the value appropriation goals of

internal linkages, with much stronger results for the latter. Specifically, the leading firms are more likely

to use cross-cluster teams for their R&D projects when surrounded by direct competitors, even after

controlling for learning opportunities. We also find support for the effectiveness of this appropriation

strategy: technologies produced by cross-cluster teams are more likely to be transferred internally to other

locations of the same firm, but are less likely to be used by competitors in the same cluster. Thus, a

closely-knit internal R&D network can be a source of competitive advantage in crowded clusters.

Interestingly, while the knowledge sourcing effect is significant among the general population in the

cluster, the appropriation effect is significant only with market competitors, suggesting much targeted

appropriation strategies. This sheds light on the seemingly contradictory coexistence of knowledge

sourcing and knowledge appropriation in clusters: knowledge appropriation strategies targeted towards

direct market competitors do not prevent knowledge sourcing by other players in the community. These

findings are robust to extra tests described in section 5. Section 6 summarizes our results, contributions

and suggestions for future research.

Empirically, our study offers a new methodology to define clusters, specially designed to compare

clusters across countries. As a byproduct, the authors are making public location data for patents granted

in the U.S. from 1969 to 2010; data that has been cleaned and contains latitude and longitude information

for most inventors in the USPTO dataset. It also advocates for careful handling of patent data – for

example by using patent families instead of patents and controlling for examiner patent citations.

Theoretically, our study offers several contributions to the innovation and international business

literatures. First, it brings to the forefront a neglected dimension to the appropriation literature: location.

Actual knowledge outflows – hence the actions to prevent it – mostly happen at specific locations. It also

highlights the role of location in the interaction between firms’ technology and product-market strategies:

local knowledge spillover does not have to get in the way of market competition if the complementary

capabilities are spread across multiple locations. Second, our study challenges the cluster literature by

pointing out the need to look at multi-location firms, and their organization across locations, to fully

understand the dynamics of a specific cluster. As far as we know, we are also one of the first papers to

analyze the appropriation effect of firms’ internal linkages along with their knowledge sourcing effect.

While firms can source knowledge from a wide range of organizations in a technology cluster, the

strategically organized internal linkages also allow firms to protect their innovations from a targeted set of

market competitors. This in turn explains the paradox we propose at the beginning of the paper: how

leading firms in high-tech industries can benefit from clusters without risking their technological edge.

2. Theoretical Development

In this section, we first analyze the features of clusters and why appropriation strategies are particularly

important for firms surrounded by direct competitors. We further argue that, among the many

mechanisms discussed in the literature, firms’ internal linkages enable the multi-location firms to

integrate local R&D with complementary assets residing elsewhere in the world, hence reducing the

appropriation risk in clusters.

2.1. Clusters and Firm Heterogeneity

According to Porter (1998), clusters are a prominent feature in the landscape of every advanced economy.

Starting with seminal work by Marshall (1920), researches have shown that firms in an industry cluster

benefit from knowledge spillover across organizations, access to specialized labor, and access to

specialized intermediate inputs. Among the various activities along the value chain, R&D activities

benefit the most from local knowledge spillover, and thus show the highest level of concentration

(Audretsch and Feldman 1996; Alcácer 2006). Geographic proximity enables frequent interpersonal

interactions through existing social networks (Almeida and Kogut 1999) and local institutions (Gilson

1999; Stuart and Sorenson 2003), which facilitate the transfer of tacit knowledge in clusters.

Knowledge, however, flows in both directions. Knowledge flowing into the firm (knowledge inflow) is

likely to make R&D investments more productive, and thus raise the incentives to invest in R&D.

Meanwhile, knowledge flowing out of the firm (knowledge outflow) may hinder the firm’s ability to

appropriate value from its own innovations, thus lowering its incentive to conduct R&D in clusters

(Furman et al. 2006). In particular, losing knowledge to nearby competitors erodes the competitive edge

held by industry leaders.

These firms can move away from clusters to protect their cutting-edge technologies (Shaver and Flyer

2000), but this option may not be sustainable or desirable for two reasons. First, even if a leading firm

decides to locate apart, it has little control over the subsequent location decisions of competitors or the

emergence of new firms. To the extent that other firms have incentive to cluster around industry leaders,

geographic distance offers only temporary protection against knowledge outflow. Second, there may be

crucial resources in the cluster that the firm relies on, such as the talent pool from a local university.

Relocation would seriously compromise the firm’s long-term competitiveness in the industry. Hence,

protecting proprietary technologies from nearby competitors and appropriating value from innovation is a

strategic consideration leading firms cannot avoid.

One feature that industry leaders can take advantage of is their geographically dispersed, yet closely

integrated, innovation networks. The literature of clusters traditionally treats all local entities as stand-

alone organizations. As a result, interactions among local competitors have been examined without much

consideration of firms’ extended organization. At the same time, most of the leading firms in high-tech

industries are large firms with R&D activities in multiple locations – if not countries. As emphasized by

Pisano (2006), an industry’s methods of appropriation are created by the strategic decisions of firms in

that industry. Hence, the strategic allocation and integration of R&D activities by multi-location firms

will have important implications for firms’ interactions in clusters.

2.2. Appropriation Mechanisms

The innovation literature has discussed a wide range of strategies for value appropriation. In general,

these strategies fall into two broad categories: raising the barriers, or reducing the incentive of imitation.

Firms may raise the imitation barrier by maintaining physical distance away from potential imitators (e.g.,

Shaver and Flyer 2000). They may also implement organizational designs to ensure secrecy and manage

information access. For example, rules and procedures are often in place to control both virtual access

(e.g., password, color-coded databases) and physical access (e.g., USB drives, laptops, building security)

to information. In fact, in both the 1987 Yale Survey and the 1994 Carnegie Mellon Survey, secrecy has

been consistently identified as one of the most important mechanisms firms use to protect their R&D

investment across a wide range of industries (Levin et al. 1987, Cohen et al. 2000).

Legal devices such as patents, trade secrets, non-compete and non-disclosure clauses also effectively

increase imitation cost and deter information flows. Prior studies found that employees’ awareness of

trade secret handling procedures were positively related to the obligations they felt to protect trade secrets

(Hannah 2005). Even the corporate reputation for being “tough” in patent enforcement reduces the

knowledge spillover associated with employee mobility (Agarwal et al. 2009).

Alternatively, firms can reduce the incentive of imitation by making it less appealing to outsiders. For

example, technologies are valuable only when combined with the right complementary assets (Teece

1986, Anand and Galetovic 2004, Fosfuri et al. 2008), including physical assets, marketing and

managerial skills, brand names, know-how and technological capabilities. As a result, R&D is often firm-

specific in its intended use, leading to heterogeneity across firms in terms of R&D applications and better

appropriability of R&D returns by the innovating firms (Helfat 1994). Similarly, to reduce the negative

impact of knowledge outflow on firm performance, partners engaged in R&D alliances narrow the scope

of their alliance activities to pure R&D projects when they are competitors in final product and

geographic markets (Oxley and Sampson 2004). Such design reduces direct market competition and

hence increases the lead-time in the product market. Not surprisingly, the Carnegie Mellon Survey found

complementary capabilities and the subsequent lead time as another important mechanism of knowledge

appropriation (Cohen et al. 2000).

While the above mechanisms are important in various circumstances, most of them treat firms as an

abstract entity. However, knowledge flow is not location free; the actual spillovers – hence the actions to

prevent it – mostly happen at specific locations. In this light, the gap in the knowledge appropriation

literature is almost the opposite of that in the cluster literature: the appropriation literature discusses

firms’ strategic organization without paying much attention to the specific location characteristics, while

the cluster literature emphasizes the role of location but overlooks the complex internal organization of

firms. In the following discussion, we explore the interaction between firms’ strategic internal

organization and location characteristics, which we argue is an important aspect of learning and

appropriation in clusters. In particular, we focus on a specific type of internal organization: internal

linkages across locations.

2.3. Internal Linkages, Knowledge Sourcing and Knowledge Appropriation

Researchers have long recognized firms’ internal linkages as effective means of knowledge sourcing, the

absorption and integration of external knowledge. By establishing interactions across divisions or

distance, internal linkages facilitate the accumulation and integration of knowledge (Bartlett and Ghoshal

1990; Kogut and Zander 1993). Empirical evidence shows that strong internal linkages – evidenced by

collaboration among inventors across distances – are conducive to the absorption of external knowledge

(Lahiri 2010) and the knowledge flow from foreign subsidiaries to the parent companies (Frost and Zhou

2005). Furthermore, such linkages also affect innovation quality. Singh (2007) shows that geographic

dispersion of R&D, once accompanied by sufficient cross-regional ties among researchers from different

R&D units, is associated with an improvement in innovation quality.

We argue that, in addition to the knowledge sourcing effect, internal linkages are also means of

knowledge appropriation for multi-location firms. Specifically, we propose two mechanisms through

which internal linkages can promote internalization and reduce the risk of unintended knowledge outflow.

First, internal linkages allow firms to maintain effective control over innovative activities at various R&D

centers, which increases internal alignment of project progress across locations. Second, internal linkages

are associated with strong technological interdependence within the firm, making innovations more

valuable internally than to the outsiders. Below we explain these two mechanisms in detail.

Internal linkages as mechanism of control: Firms with strong internal ties can closely monitor the

progress of R&D activities at each location to make sure that it is in alignment with the firm agenda.

Examples of such internal ties include the participation of researchers from other locations in local R&D

projects (Nobel and Birkinshaw 1998) or the rotation of managers across units (Edstrom and Galbraith

1977), as both are considered ways to improve coordination and control in multinational organizations.

Our interviews with R&D managers in large multinational firms in the semiconductor industry also

suggest that, with frequent interactions among locations, valuable innovations can be promptly identified

and sometimes transferred to a safer location, often at the headquarters or the firms’ primary R&D

centers.

Internal linkages as mechanism of internal interdependence: Firms with strong internal ties can better

integrate innovation – wherever it emerges – with the complementary knowledge and resources within the

firm, leading to stronger competitive position in the product market. Modularity with firm-specific

interfaces is one example of such internal interdependence. Liebeskind (1996) proposes that firms can

isolate various components of the same product so that none of the project teams can reproduce the

product without the help of the others. In studies of multinational R&D strategies, Zhao (2006) and Zhao

and Islam (2006) find that firms with strong internal linkages are able to conduct R&D in environments

with weak IPR protection by substituting their internal organization for external institutions. With spatial

isolation, local innovations are less attractive to neighboring firms if these competitors do not have access

to the same set of complementary knowledge.

The importance of interdependence in knowledge appropriation has also been discussed in more general

settings. Using a theoretical model, Rajan and Zingales (2001) explain why flat hierarchies – in which all

division managers are required to collaborate with a central unit at the top – are ubiquitous in human

capital-intensive industries such as legal and consulting services. Because of the intangible nature of firm

resources, property rights protection is difficult to enforce. Yet, if the firm can increase everyone’s

dependence on the center office by controlling access to certain key resources, the risk of expropriation is

greatly reduced. In other words, the risk of knowledge outflows is reduced if the divisions of a firm are

highly dependent upon each other.

Based on the above discussion, we argue that a firm can appropriate more value from its local R&D with

the presence of stronger internal linkages with the rest of the firm. In the following analysis, we focus on

a specific form of internal linkages: cross-cluster teams. Although there are many other concrete

manifestations of internal linkages – from systematic meetings to enhanced communication mechanisms

across clusters – our interviews have identified cross-cluster teams as an important form of value

appropriation. More importantly, we are able to map cross-cluster teams to concrete data, and thus avoid

analyzing the appropriation mechanisms in abstract terms.

Cross-cluster teams need to overcome technological, organizational and geographical barriers (Frost et al.

2002) and are costly to manage (Doz et al. 2006). Firms will only implement such strategies when the

benefit outweighs the cost, namely when the threat to knowledge appropriation is high. Therefore, we

expect to observe more cross-cluster teams at locations with higher appropriation risks, e.g. in clusters

with a large number of direct competitors. Furthermore, if cross-cluster teams are effective in enhancing

knowledge appropriation, we would expect two effects at the same time. On the one hand, cross-cluster

teams facilitate the integration of local innovations with complementary knowledge in the firm.

Therefore, we would expect to see more intensive intra-firm knowledge flows across clusters with the

presence of cross-cluster teams. On the other hand, cross-cluster teams reflect internal interdependence,

increase the firm-specific nature of projects (i.e., stronger complementarities with firm-specific

resources), and hence raise the learning barriers faced by outsiders. Therefore, we would expect to see less

knowledge outflow to local competitors with the presence of cross-cluster teams.

Note that we are not dismissing the knowledge sourcing effect of the cross-cluster teams. In fact, we also

expect to see more intensive use of cross-cluster teams when there are more learning opportunities in the

cluster, where teams are used as tools of transferring locally absorbed knowledge back to the corporate

center. That said, we argue that firms’ strategic response to local appropriation risk is above and beyond

the knowledge sourcing effect: cross-cluster teams should be more prevalent when firms perceive higher

appropriation risk in the cluster, given the learning opportunities in the region. In the empirical analysis,

we simultaneously analyze the dual roles played by cross-cluster teams and attempt to tease out the

knowledge sourcing effect and the knowledge appropriation effect in multi-location firms.

3. Empirical Design

Our empirical design is conceptually based on the literature that looks at the interaction between a firms’

technology and product-market strategies. The multi-dimensional relationships among local entities

(Cohen 1995, 230) allow us to separate the appropriation incentives from knowledge sourcing incentives

among multi-location firms. Firms in a technological cluster may share similar technological backgrounds

or even engage in patent races, but they do not necessarily compete in the same product market. Industry-

specific market information and other complementary resources reduce the risks associated with

knowledge exchanges, allowing symbiotic relationships to develop. If internal linkages are purely

mechanisms of knowledge sourcing, we should observe stronger internal linkages in clusters with greater

knowledge pools, e.g., a large number of neighboring firms in the same technological field. If internal

linkages serve as a hedge against knowledge outflows, they should be used more extensively when

neighboring firms share the same product market.

Specifically, our empirical strategy follows three steps: (1) determining whether firms are more likely to

use cross-cluster teams in clusters with high levels of competitor firms, (2) testing whether innovations

created by cross-cluster teams are used by other locations within the firm – an indication of

internalization, and (3) determining whether those innovations associated to cross-cluster teams are less

cited locally by competitors – an evaluation of cross-cluster teams as mechanism to appropriate

knowledge.

Taken as a whole, the results from these three empirical analyses offer a thorough test of the propositions

introduced in section 2.3. Although most variables used in each step are common (for example

measurements that characterize the cluster competitive environment as described in section 3.5.1), there

are differences in terms of dependent variables (section 3.3) and its associated model-specific control

variables (section 3.5.1). Regardless of the specific analysis, all variables are calculated from the same

data sources, and for the sample (section 3.1), and use the same cluster definition (section 3.2).

3.1. Sample

Our empirical setting is the worldwide semiconductor industry from 1998 to 2001. We choose this

industry for several reasons. First, innovation is a key factor for success in semiconductors. Firms invest

relentlessly in R&D to introduce new products and improve production processes (Stuart 2000).

Moreover, semiconductor firms routinely patent their innovations, and patent data have been used to trace

the traits and geographic distribution of innovation. Second, the benefit of knowledge transfer between

firms has been shown to drive agglomeration in the industry (Saxenian 1994; Fleming et al. 2006). High

levels of geographic concentration also suggest that semiconductor firms have already developed

strategies to manage knowledge outflows. Finally, this is a truly global industry: leading firms operate at

multiple locations around the world, and there are significant differences between firms in terms of

product markets, R&D portfolios, positions in the value chain, and geographic locations. Firms range

from industry giants that participate in activities throughout the value chain to small enterprises that

specialize in design (known as fabless) or testing, and from large multinational firms to small local firms.

Other players, such as universities, national laboratories, and firms from other industries (e.g., aerospace

and chemicals) also conduct active R&D in semiconductors. Such heterogeneity allows us to identify the

effect of different local competitive environments on firms’ knowledge appropriation strategies and

allocation of R&D projects.

We piece together our dataset from several sources. First, we identify innovating firms using patent data

from the Derwent World Patent Index (DWPI), a well-recognized dataset that encompasses more than 30

million patent documents from 41 patent-issuing authorities worldwide, and we rely on Derwent’s

technological classification1

Many of these patents are linked to the same innovation, with exactly the same inventors, assignees and

abstracts. Multiple patents per innovation can occur either because patents are filed in multiple countries

or because an application in a given country spins out multiple patents. Failing to recognize multiple

patents associated to an innovation may lead to overestimate innovation output in a location – a given

innovation would be counted multiple times – or to underestimate its backward and forward citations.

Thus, we follow Gittelman and Kogut (2003) and use Derwent families of patents as our unit of analysis.

Derwent defines membership to a given family based on a common priority document

to obtain the universe of semiconductor patents. Patent data include

innovations that occur outside of the R&D facilities, thus are more inclusive than the number of labs or

the amount of R&D spending. Information from semiconductor patents applied between 1998 and 2001,

and granted between 2001 and 2003, results in a sample of 61,956 patents of which 28,334 were granted

in the U.S.

2. Each Derwent

family encompasses patents granted in all countries that are identical in terms of technology, inventors,

and locations, but differ in the scope of their claims. We restrict our sample to families that have at least

one American member – e.g. those families that have at least one patent granted in the U.S. – and build

forward and backward citation variables using only citations from and to American patents to avoid

biases that originate from citation standards and practices that vary across legal jurisdictions. The final

sample consists of 23,383 patent families whose assignees are American and foreign firms, universities,

and government- and industry-sponsored research labs. Patent families have an average of 2.6 foreign

patent members and 1.2 American patent members3

1 DWPI applies a consistent classification system to all patents. Classes used in this study are U11 (semiconductor materials and processes), U12 (discrete devices), U13 (integrated circuits) and U14 (memories, film and hybrid circuits). For more details, see http://scientific.thomson.com/support/patents/dwpiref/reftools/classification.

. For the 624 patent families with more than one

assignee, all assignees (and not only the first one) are considered.

2 The first member of the patent family to enter the DWPI database is referred to as the basic member. Basic members are identified by comparing the priority data on a newly received patent document with the priority data already in the DWPI database. If the priority data on the new document does not match the priority data of any previously processed document, then the new document is considered to be basic member and a new DWPI record is created with a unique family ID. When additional documents are subsequently received by Derwent that match the priority data of the basic members, then these are referred to as convention equivalents – based on the Paris convention agreement. In addition to using priority data comparisons, Derwent’s technology experts also examine patent documents to identify non-convention equivalent family members – applications filed after the 12-month convention period but that are related to a given patent in DWPI. 3 To be part of our sample, a patent family needs to have at least one American member.

We supplement this initial sample with directories of semiconductor plants, fabless companies, and

institutions behind scientific publications. Information about plants comes from the quarterly datasets of

the World Fab Watch provided by the Strategic Marketing Association, from 1998 to 2001. The datasets

encompass manufacturing facilities4

Because we treat every multi-unit firm as an integral entity, and because internal organization is a central

concept of this study, we put extra effort into identifying the ultimate parent for every entity in our

sample. First, for each year, we match the patent assignees, plants and fabless companies to firms in the

corresponding yearly Directory of Corporate Affiliations (DCA)

for a wide range of products: memories, microprocessors, generic

and specific chips, etc. Information on fabless companies is obtained from the Gartner Group’s annual

Directory of Fabless Semiconductor Companies for the same period. To assess the scientific activities in

the local community, we extract from ISI Web of Knowledge all journal publications in the sample period

that use “semiconductor” or “semiconductors” as part of their keywords. These four data sources provide

a comprehensive map of the industry at multiple levels: innovation (28,333 patent families), production

(974 plants), research (50,387 scientific publications), and development (549 fabless companies).

5

While we use data for all organizations to characterize local environments, our analysis of R&D

strategies focuses on 16 innovating firms, or the top 1% of the industry in terms of patent output

, an annual database that records

corporate ownership for more than 200,000 private and public firms worldwide. Second, for

organizations not identified in DCA, we search the Dun and Bradstreet Million Dollar Database to obtain

affiliation information. Finally, we check affiliation changes through SDC Platinum, company websites,

and various industry publications. The above steps map the 4,125 assignees in the sample to 2,217 unique

organizations. Fabless firms and manufacturing firms that do not own patents add 721 additional

organizations to our sample.

6

4 World Fab Watch data reports five types of manufacturing facilities: fabs, test facilities, assembly facilities, pilot fabs and test fabs. We use only fabs – manufacturing facilities that are fully operative.

. The

reason for focusing on these firms is that most of the semiconductor industry has the typical features of an

oligopoly industry, where the top 1% of firms represents 50% of the patent output and 40% of the plants

operating in this period. With the cost of developing new chips and building new manufacturing plants

running into the billions, there is a clear divide between industry leaders and everyone else – and the gap

is getting larger, according to IC Insights. Therefore, semiconductor industry leaders should have

qualitatively different innovation strategies than the thousands of industry followers. The composition of

the sample is similar to those in previous studies of the semiconductor industry (Stuart and Podolny 1996;

5 DCA is widely used by scholars in strategy and technology to identify corporate structures. Proquest, searched on November 1st, 2010, reveals that 190 papers in management journals report using DCA as data source. 6 The 16 firms are AMD, Intel, IBM, Texas Instruments, Hitachi, Matsushita, NEC, Siemens (including Infineon), Toshiba, Mitsubishi, Samsung, Micron, Fujitsu, TSMC, Hyundai, and STMicroelectronics.

Henisz and Macher 2004; Ziedonis 2004). As part of our robustness checks, we replicate our analyses

using an alternative sample composed of the top 30 firms, the top 5% of patent output, and obtain similar

results.

3.2. Cluster definition

Defining clusters is a crucial step of our empirical setup. Three elements must be specified in a cluster

definition: the variable that would be used to identify concentration levels of economic activity, the

geographic unit over which such a variable would be measured and the threshold concentration level

above which a location can be considered a cluster. Our definition of clusters departs from most previous

work in these three dimensions.

First, because we are interested in technology clusters and because agglomeration patterns for R&D are

different from those for production (Audrestch & Feldman 2003), we follow Alcácer (2006) and use the

geographic distribution of inventor activities instead of more conventional variables such as industry

employment, plant output or product sales.

Second, instead of relying upon predetermined administrative boundaries, such as states or metropolitan

areas, we apply a mathematical algorithm that uses latitude and longitude data to identify technological

clusters7

Third, we define clusters by the actual distribution of inventor locations, following a three-step approach

(see Appendix A for a more detailed explanation and comparison to alternative methods)

. We do this for two reasons. First, there is no single administrative unit defined across all

countries. We have to either focus on a specific country (e.g., the U.S.), which fails to capture important

features of global firms, or use a mix of different geographic units (e.g., states in the U.S., prefectures in

Japan, and provinces in Europe), which may create unexpected country biases. Second, technological

clusters do not necessarily follow predetermined administrative boundaries, which is clear after a quick

inspection of inventor locations in, for example, the northeastern U.S. or central Japan. One

administrative unit may encompass multiple clusters, while one technological cluster may expand across

several administrative lines.

8

7 Alternatively, we identify clusters using a traditional algorithm: hierarchical clustering with centroid linkages and re-estimated all models. The results from using this alternative definition are discussed in the robustness section and are similar in terms of magnitude, sign and statistical significance to those presented in this paper.

. In step one, we

identify the location of each element in the sample (i.e., a patent inventor, plant, fabless company, or

scientific publication) and match the locations to two comprehensive sources of geographic names. For

U.S. locations, we obtain latitude and longitude information for all 38,261 locations in the country from

the Geographic Names Information System (GNIS) of the U.S. Geological Survey. For foreign locations,

8 The analysis is replicated with a hierarchical clustering algorithm in robustness checks described in section 5.

we use the Geonet Names Server (GNS) of the National Geospatial Intelligence Agency. Besides its wide

coverage of 5.5 million location names worldwide, the GNS dataset uses phonetic variations to capture

spellings from a different alphabet (as in Asian countries) and from an alphabet with extra characters (as

in Scandinavian and Slavic countries). Ambiguous matches are checked manually by native residents

from various countries and areas. As a result, we are able to assign latitudes and longitudes to 61,385 of

the 61,461 foreign locations in the original sample. In the second step, we develop a mathematical

algorithm to identify geographic clusters using the latitude and longitude information. In the second step,

clusters are defined not only by the geographic distance among locations – as in many other traditional

clustering methods – but also by the variations in inventor density in neighboring areas. For example, a

rapid decrease in density may signal the end of a cluster, and a continuous level of inventor density may

signal a long or irregularly shaped cluster. Accordingly, the algorithm assigns two locations to the same

cluster if there is a continuity of high-density locations between them, despite their geographic distance.

In contrast, two locations separated by a stretch of low-density areas may be identified as two distinct

clusters, even if they aren’t far apart. Our clustering algorithm offers the additional advantage of having

the number of clusters emerge organically from the data, instead of being set arbitrarily ex ante. This

method produces 304 geographic units. In the final step, plants, fabless companies, and publications are

assigned to the geographic units defined from the patent data. In most cases, they fall within an existing

geographic unit. For each location that falls out of all existing units, we calculate its shortest distance to

them. The location is considered part of the closest cluster if the minimum distance is less than 15 miles9

Although the previous process generated 338 geographic units, not all of them qualify as clusters.

Numerous approaches have been used to determine whether a geographic concentration of economic

activity is large enough to be categorized as a cluster. For example, Ellison and Glaeser (1996) identify

clusters by visually inspecting a plot of their agglomeration measurement; Alcácer (2006) uses the upper

quartile of geographic patent concentration, and Delgado, Porter and Stern (2010) define clusters as those

in the top 20% of their specialization-size measurement. We take the top 25 locations

.

Otherwise, the unassigned locations are again clustered with the same algorithms as we use for the patent

locations. For the main sample, six and 28 geographic units were added by fabless and plant data,

respectively.

10

9 We also tried 20, 25, and 30 miles, with very similar outcomes.

which contain

84% of all patent families (88% of American patents), 57% of plants, 67% of fables and 76% of

publications.

10 We replicated our analysis using the top 50 and all geographic clusters identified where firms in our sample have some presence. The results are similar to those obtained by using the top 25 clusters. Results are available upon request to the authors.

Table 1 shows some descriptive statistics for the clusters used in our analysis.

3.3. Models

To identify firms’ strategic organization of R&D projects across clusters, we compare the technologies

developed in different local competitive environments, controlling for firm and patent characteristics.

Specifically we explore three dimensions of innovations; whether the innovation is associated to inventors

located across clusters, whether this type of innovation is internalized – cited as prior art by innovations

developed in other clusters, and whether the innovation is less cited by competitors in the same cluster.

Although the models estimated in each dimension differ in terms of dependent variables and some

independent variables, they all follow the following general structure:

DepVarfict = Cict + Xf + Yict + Zct + ζt +υi + τctry + γtech + εfict (1)

Where DepVarfict is one of three variables – cross_clusterfict , cross_cluster_self_citationfict , and

local_citation_by_competitorfict – (described in section 3.4) for steps 1, 2 and 3 respectively,

Cict is a vector of cluster-specific variables capturing the competitive environment faced by firm i

in cluster c and year t ( described in Section 3.5.1),

Xf is a vector of patent family-specific variables (described in Section 3.5.2),

Yict is a vector of firm-specific variables characterizing firm i in cluster c and year t (described in

Section 3.5.2),

Zct is a vector of location characteristics in year t (described in Section 3.5.1),

ζt, υi, τctry are γtech are four sets of dummy variables for year, firm, country and technology fixed

effects, respectively, and εict is the error term.

The estimation technique varies according the nature of the dependent variable. We use logit

estimation techniques for models with cross_clusterfict as a dependent variable given its binary

nature. Because cross_cluster_self_citationfict and local_citation_by_competitorfict are count

variables, we use negative binomial estimation techniques11

3.4. Dependent variables

.

Cross-cluster teams: Geographically dispersed R&D in a multi-location firm makes it more

difficult for local competitors to access the technology know-how residing in the firm's other

subsidiaries, thus reducing knowledge outflow. Teams spanning multiple clusters can also facilitate 11 We explored whether our data is better suited for negative binomial estimation than for Poisson estimation through a set of Hausman tests that compared both models. Our results, explained in more detail in sections 4.2 and 4.3, favor the use of negative binomial estimation.

the transfer of local know-how throughout the organization (Lahiri 2010). Thus, we define

cross_clusterfict as a binary equal to 1 if the patent family has inventors that are from at least two

different clusters.

Cross-cluster self-citations: A key concept in this study is the extent to which an innovation creates

value for the innovating firm. While there is no direct measure of value, technologies highly

dependent on internal resources are more likely to be utilized and further developed within the firm.

Trajtenberg et al. (1997) propose self-citations, defined as “the percentage of citing patents issued

to the same assignee as that of the originating patent,” to measure the “fraction of the benefits

captured by the original inventor.” Hall et al. (2005) also suggest that citations to patents belonging

to the same firm represent internalized knowledge transfers, bolstering the firm’s competitive

advantage. Hence, we use forward self-citations as a proxy for the value new technologies bring a

firm. Specifically, we define the variable cross_cluster_self_citationfict as the number of self-

citations received by patent family p by other patents whose assignee is firm i but where generated

in cluster other than cluster c. Because we are interested in firms as integrated organizations, any

citations among affiliated organizations are considered self-citations.

Local citations by competitors: If using cross-country teams leads to internalization and better

knowledge appropriation, we should observe that patents associated to cross-cluster teams are less

likely to be cited by competitors in the cluster. To test this proposition, we create a set of variables

that capture the number of citations made by competitors located in the same cluster. Because our

characterization of the cluster competitive environment is diverse, the count of local citations by

competitor varies accordingly. Thus we define local_citations_by_innovatorsfct as the total number

of citations to members in a family made by other assignees – regardless of its type;

local_citations_by_profitfct as the count of citations made by assignees classified as for profit

organizastions; local_citations_by_industryfct as the count of citations made by assignees in firm i’s

industry; local_citations_by_segmentfct as the count of citations made by assignees firm i’s product-

segment; and local_citations_by_competitorfct as the count of citations made by assignees identified

as competitors of firm i. Only citations made by patents originated in cluster c are counted. Section

3.5.1 describes how we classify assignees into different groups – profit, industry, segment and

competitor.

Two issues related to citation-based measurements are worth exploring further. First, citation measures

capture both the intensity and speed of citations. As our observation window ends in September 2010,

any citations that occur after that date are not included in the sample. Jaffe and Trajtenberg (2005) show

that the lag of forward citations peaks at around five years. The period to accumulate citations in our

sample ranges from seven to nine years, so our citations should represent the bulk of citations to be

received12

Second, a common critique of citation-based measurements is the unknown nature and extent of citations

imposed by patent examiners (Jaffe et al. 2000). Recent research reveals that examiner citations account

for 66% of all citations in an average patent, which may bias empirical tests (Alcácer and Gittelman

2006; Sampat 2009). To avoid this problem, our main models are estimated using citations listed by

assignees only. In our sample, about 30% of the patent families that receive at least one inventor citation

also have at least one self-citation. The number is 38% when both inventor and examiner citations are

considered. For robustness checks, we also repeat our analysis using all citations to a patent regardless of

their source.

. We also include time fixed effects (ζt) to account for variations in citations across cohorts and

to control for year-specific events that may affect patents applied in that year.

3.5. Independent variables

3.5.1. Characterizing cluster

Competition (Cict ): Firms sharing the same technological space have better absorptive capacity

for each other’s knowledge (Cohen and Levinthal 1990), so the likelihood of knowledge outflow

is larger in areas with many firms doing the same type of R&D. Meanwhile, such knowledge

outflow will only create high risks if the recipients are aiming for the same product market

(Dushnitsky and Shaver 2009). Therefore, we follow two dimensions – technology space and

product market – to characterize the competitive environment at the cluster-year level.

Along the technology space, competitors are defined generically as organizations that innovate in

the semiconductor field. The variable innovators represents the number of unique assignees with

semiconductor patents in a given cluster-year. We then classify assignees into two groups:

innovators_profit and innovators_nonprofit to capture the number of for-profit and nonprofit

assignees, respectively. In addition, we use the status information on patent applications to further

classify for-profit assignees into small or large entities, thus creating the variables

small_innovators and large_innovators. In the case of nonprofits assignees, we manually classify

them into three groups: universities (universities), government agencies (govt_innovators), and

other nonprofits such as research centers sponsored by industry associations (other_nonprofit).

Along the second dimension, competitors are defined as firms that share the same product-

market. For every focal firm in our sample, we rely on Hoover’s Online to identify its industry

12 We plotted the stream of citations for patents in our sample. We found that in fact citations peaked before September 2010.

(four-digit SICs), market segments within semiconductors13

Other cluster variables (Zct ): We complete the characterization of local innovation environments

with three more variables: plants_in_cluster, fabless_in_cluster, and publications_in_cluster,

which represent the numbers of plants, fabless companies and publications per cluster-year.

, and the names of direct competitors.

Then we count the number of for-profit assignees in the same industry (in_industry and

not_in_industry), in the same market segment (in_segment and not_in_segment), or on the list of

direct competitors (competitors and not_competitors). The self-reported competition data from

Hoover's serves our purpose well, since managers make strategic moves based upon perceived

competition in a technology cluster.

Zhao (2006) shows that internalization varies according to the property rights regime of a given

country. Thus we also control for variations in country-specific intellectual property right regimes by a

set of country dummies τctry.

3.5.2. Characterizing patent families (Xf)

As discussed in section 2.3, cross-cluster teams and internalization can be associated to knowledge

sourcing. To control for this alternative explanation, we create a set of dummy variables that indicate

whether any patent member of the focal family cites prior art granted to assignees in the same cluster14.

These variables are defined according to the competition measurement used in a specific model:

cites_localfc indicates if any member of the family cites at least one previous local patent (regardless of

the assignee type of such patent)15

Patents associated to cross-cluster teams that are more internalized or are less cited by local nearby

competitors may also have low intrinsic values. Through all specifications we control for innovation

quality through claimsf – the summation of patent claims across all American members of a patent

family (Lanjouw and Schankerman, 1999)

; cites_local_profitfc is equal to 1 for families with at least one patent

member citing an assignee that is classified as a for-profit entity; cites_local_industryfc if the citation is

to assignees in the semiconductor industry; cites_local_segementfc if there is any citation to assignees in

the same segment and cites_local_competitorfc if there is any citation to competitors.

16

13 Hoover’s reports 13 segments under semiconductors, including memory chips and modules, microprocessors, etc.

. Although our sample was drawn by sampling by

technology, our results may also be driven by technological differences within semiconductors;

therefore we add a set of dummy variables by technology classes (γtech ).

14 We are very thankful to an anonymous reviewer who suggested this approach to control for knowledge sourcing. 15 This variable indicates citations to any patent. 16 We use another alternative measurement for innovation quality, total number of countries where an innovation has been patented (Putnam, 1996), and obtain very similar results. Note that we cannot use the total number of citations as a measurement of quality since total citations is used as exposure variable in the negative binomial models.

Finally, some patent-family variables are relevant only for specific analyses. For example, in step 2 –

testing for cross-cluster self-citations, we control for the baseline propensity that a patent is more self-

cited because of sheer size of subsequent patents granted to a firm. Specifically we construct a patent-

variant variable patent_stockfit as the number of patents that firm i obtained between the time the first

patent within a focal firm was granted until September 2010. In steps 2 and 3 total_citationsft – defined

as total number of citations received by American patents in family f – is used as the exposure variable

for negative binomial models. Since the coefficients for exposure variables are forced to be 1, we are

essentially estimating the ratios of cross-cluster self-citations to total citations in step 2, and the ratio of

citations by local competitors to total citations in step 3.

3.5.3. Characterizing firms (Yict)

Technologies closely linked to manufacturing or product design may have different characteristics from

others. In addition to variables measuring local environment described in section 3.5.1, we use two

dummy variables, with_plantict and with_fablessict, to indicate whether a particular firm has plants or

fabless units in cluster c and year t. Both variables vary by firm-cluster-year and are used in steps 1 and

2.

Finally, Alcácer et al. (2008) show that firms differ significantly in their practices to cite prior art. As a

consequence, we also add firm fixed effects (υi) to control for these differences and any firm time-

invariant characteristic.

4. Empirical results

4.1. Are firms more likely to use cross-cluster teams when there are more competitors nearby?

The first step in our analysis is to determine whether innovations generated by cross-cluster-teams are

more likely to emerge in clusters with higher levels of competitor presence. The model to estimate this is

the following:

cross_clusterfict = Cict + Xf + Yict + Zct + ζt +υi + τctry + γtech + εfict (2)

Table 4 shows the results of estimating equation 2 using logit. Standard errors are clustered by geographic

cluster.

The positive coefficients of innovator_profit, large_firms, in_industry, in_segment and competitors

suggest that the presence of competing organizations increases the tendency to use cross-cluster teams.

For example, according to column 7, an increase in one standard deviation in the number of competitors

increases the likelihood of using a cross-cluster team (roughly six new competitors) by 48%17

Evidence of cross-cluster teams to channel local knowledge acquired in the cluster to distant units within

the firms is mixed. The dummy variable for backward citations is positive and significant only when it

implies citations to any type of assignees, but not citations to competitors or firms in the same segment or

industry. That is, knowledge sourcing from market competitors does not trigger the use of cross-cluster

teams, suggesting that “reverse knowledge integration” suggested by Frost and Zhou (2005), if it occurs,

is mostly from non-competitive institutions such as universities, research institutes and innovators that are

not direct competitors.

. Note that

the number of nonprofit innovators has no effect on the use of cross-cluster teams.

Two other variables deserve further comments. Having a plant in the cluster increases dramatically the

chances of having a cross-country team. This may reflect innovations in manufacturing that requires

coordination across production sites or specialized knowledge that resides in central R&D labs. Higher

quality innovations – those innovations with more claims – are also more likely to be associated to cross-

cluster teams. Given the higher level of coordination that cross-cluster teams require, it is not surprising

that firms are more inclined to use them when the potential benefits are higher.

Taken together these findings suggest that high levels of competitors in a cluster, not of other types of

innovators, are linked to cross-cluster teams and that cross-cluster teams seem to be associated to

appropriation rather than knowledge sourcing.

4.2. Are innovations created by cross-cluster teams more likely to be used at other locations within the

firm?

The second step in our analysis is to determine whether innovations generated by cross-cluster-teams are

more likely to be cited by innovations created by the same firm in other clusters. The model to estimate is

the following:

cross_cluster_self_citationfict = Cict + Xf + Yict + Zct + ζt +υi + τctry + γtech + εfict (3)

Table 4 shows the results of estimating equation 3 using negative binomial with total_citationsft as the

exposure variable. Because the dependent variable is the number of self-citations received by the focal

patent, and the exposure variable is the total number of forward citations, we are essentially examining

the patent’s cross-cluster self-citation ratio. OLS regressions with self-citation ratio as the dependent

variable produce very consistent results. Standard errors are clustered by geographic cluster. Note that

cross_clusterfict is now an independent variable.

17 The odd ratio value associated to the variable competitors in column 7 of Table 3 is 1.08.

Across specifications, innovations associated to cross-cluster teams are more internalized. For example,

model 7 on Table 4 indicates that cross-cluster team patents receive 0.5 more cross-cluster self-citations,

40% of the average cross-cluster self-citations in the sample. The result is robust, in magnitude and

significance, across specifications that vary in their characterization of competitors in the cluster.

The total number of innovators in the cluster does not seem to have a significant impact on

internalization. The effect of local competition emerges only when firms compete in the product market

and not when entities employ similar technology in different markets. Across various specifications of

local competitive environment, the coefficient on the number of local competitors is positive and

significant. The more market competitors there are in a cluster, the more likely firms are to self-cite

patents they develop there. Note that this effect is after controlling for innovations being developed by

cross-cluster teams; therefore other mechanisms that enhance internalization – mechanisms that we do not

measure directly – must be at play. To the extent that self-citations proxy for internalized value, this

finding supports the argument that in highly competitive environments, firms are more likely to share

technology development across the firm. Meanwhile, the presence of nonprofit innovators has little

impact on the degree of internalization. Without direct market competition, these nonprofit institutions

create a more open atmosphere in the local cluster.

Together, these findings suggest that firms do change the type of innovation performed depending on

local competitive environments beyond using cross-cluster teams. Innovation produced in clusters with a

strong presence of direct competition is more tightly intertwined with the firm’s internal knowledge base.

Note that there is little support to the argument that the internalization process across clusters is due to

knowledge sourcing: families that draw from assignees’ local innovation are not more likely to generate

cross-country self-citations. Not surprisingly, the coefficient patent_stock is positive and significant; the

larger the pool of later patents, the more likely that later citations are made to the focal patent. The

coefficient of with_plant is also positive and significant, indicating that technologies closely linked to

manufacturing processes are more firm-specific.

Note that the high cross-cluster self-citation ratios in competitive clusters are not due to the low intrinsic

value (small denominator) of these patents: we fail to reject the hypothesis that the coefficient for claims

is distinct from zero. Additionally, when running the same regressions in Table 5 but with total number of

citations instead of cross-cluster self-citations as the dependent variable, none of the coefficients

associated with competitive environments are significant.

4.3. Are innovations created by cross-cluster teams less used by competitors?

Our final step is to evaluate cross-cluster teams as mechanism to appropriate knowledge by determining

whether those innovations associated to cross-cluster teams are less cited locally by. Specifically, we

estimate the following equation:

local_citations_by_ENTITYfct = Cict + Xf + Yict + Zct + ζt +υi + τctry + γtech + εfict (3)

Table 6 shows the results of estimating equation 4 using negative binomial with total_citationsft as the

exposure variable. Standard errors are clustered by geographic cluster. Three issues in this table required

further discussion. First, as in Table 5, cross_clusterfict is an independent variable that characterizes a

patent family. Second, note that the dependent variable varies across columns depending on how the local

environment is characterized. In columns 1 and 2 the dependent variable is

local_citations_by_innovatorsfct, in columns 3 and 4 it is local_citations_by_profitfct, in column 5 it is

local_citations_by_industryfct, in column 6 it is local_citations_by_segmentfct and in column 7 it is

local_citations_by_competitorfct. Although all variables represent the same concept – citation by

assignees that may pose a threat – the definition of what type of assignee is considered varies to match the

characterization of competitive environment in the cluster.

Innovations associated to cross-cluster teams reduce the number of citations of local innovations

generated by competitors. For example, for model 7 in Table 6 – our more precise measurement of

competition – innovations associated to cross-cluster teams receive 0.56 less local citations by

competitors those who are not cross-cluster teams. This finding supports our hypothesis that cross-cluster

teams are an effective appropriation mechanism. If cross-cluster teams were solely associated to learning

and knowledge sourcing, we would not expect this result. This is after controlling for knowledge sourcing

through the backward citation variables, whose coefficients are positive and significant in models 1 and 2.

In Table 6, variables associated to different types of competition in a cluster (innovators,

innovators_profit, in_industry, in_segment, competitors) can be interpreted as capturing the risk set that

may drive local citations: the more assignees of a given type are in cluster, the more patents they can

obtain there and the more likely these patents cite patent families in our sample. Statistically-significant

positive coefficients would represent the sheer size effect described above; statistically-significant

negative coefficients would indicate that appropriation mechanisms, besides cross-cluster teams, are so

effective that citations by competitors decrease even when their sheer numbers increase. As expected,

most coefficients are positive and statistically significant suggesting that those other alternative

appropriation mechanisms may not be effective. However, the fact that for model 7 – a model with the

most accurate competition measurement – the coefficient is not significant and that these alternative

mechanisms have not been measured explicitly suggest that more research is required.

In terms of control variables, more valuable innovations are more likely to be cited by other assignees in

the cluster, being competitors or not.

5. Robustness checks

The above findings are consistent with our hypothesis that R&D projects in competitive clusters are more

internalized across clusters, and are less used by nearby competitors. Next, we conduct a series of

robustness tests using alternate samples, variable definitions, and estimation techniques.

First, Table 1 shows high positive correlations among variables that characterize clusters, namely number

of plants, fables, publications and assignees (after all, clusters by definition are locations with higher

levels of economic activity). Because multicollineartity increases standard errors it is less likely that

coefficients will be statistically significant. In other words, strong correlation among the cluster variables

works against us finding any significance for the local competitive effect, which we obtain anyway.

Nonetheless we re-estimated models in Tables 4, 5 and 6 in four different ways to minimize any effect of

collineartity. First, we orthogonalized variables in Cict and Zct for each model. Results were stronger in

terms of significance and magnitude. Second, we used one variable to characterize the cluster in each

model (innovators, innovators_profit, in_industry, in_segment, competitors for models 2 to 7

respectively). Although results were similar, their interpretation is less clear because each variable

encompasses multiple characteristics of the cluster. Third, we used cluster characteristic ratios instead of

individual variables. For example, instead of using competitors, not_competitors and

innovators_nonprofit from model 7 in any table, we created ratio_competitor = competitors/

not_competitors and ratio_nonprofit = innovators_nonprofit /innovators. Although these ratios are not

correlated and the results are similar, the coefficients are harder to interpret. Fourth, we repeated our

analysis with all locations identified in section 3.2 – not just the top 25 clusters. Including more

geographic units increase heterogeneity across clusters. As a result, the correlation between plants and

competitors decreases as more geographic units rich in manufacturing facilities but poor in innovation are

considered. In any instance, the results did not depart from those presented in section 4.

Second, we re-estimate all models with a different method to define clusters: hierarchical clustering with

centroid linkages. This method begins with each location as a separate group. Then two clusters with the

shortest Euclidian distance are combined into one, whose new geographic coordinates are the mean

longitude and latitude of all locations in the group. This process is repeated until a large hierarchical tree

is generated that includes all locations. We designate the number of clusters in each region to

accommodate a wide variation in local densities. The coefficients obtained with the hierarchical clustering

method are similar in sign, significance and magnitude to those in the previous tables.

Third, we repeat the analysis on self-citation ratios using both inventor and examiner citations. Recent

research suggests that high levels of examiner citations are associated with low quality patents (Alcácer

and Gittelman, 2006; Sampat, 2009). Therefore, including these citations adds a new set of observations –

patents whose citations are 100% examiner-imposed – that may represent inferior innovations. The results

using citations from all sources are similar in magnitude and sign too, but weaker in statistical

significance than those in Table 2.

Finally, we estimate the models extending our sample to the top 5% of semiconductor firms (30 firms)

while keeping 25 clusters and obtain similar results in terms of sign, magnitude and statically significance

to those of our baseline models.

6. Discussion

While geographic collocation has obvious benefits for firm innovation, it can also have serious

drawbacks. We explore how leading innovators can tap into technology clusters’ rich resources while still

appropriating value from their R&D investments. Our empirical findings suggest that the internal linkages

of multi-location firms play an important role in knowledge appropriation, even after the knowledge

sourcing opportunities are controlled for. Specifically, by increasing control and intra-firm

interdependency across locations, internal linkages facilitate knowledge internalization and at the same

time reduce knowledge flow to nearby competitors. We also find that firms’ strategic responses vary

depending upon the characteristics of local organizations. Firms tend to intensify their internal linkages

when neighboring firms share the same product market, but not when they overlap in the technological

space.

We believe that our study sheds light on some important aspects of location and innovation strategies. By

studying the interaction between firms’ strategic internal organization and location characteristics, we

contribute to the appropriation literature that treats firms as abstract entities, and to the technology cluster

literature that overlooks the complex internal organization of firms. We argue that a geographically

dispersed organization, if managed properly, can be a competitive advantage not only in knowledge

sourcing but also in knowledge appropriation. Therefore, we can achieve better understanding of the

dynamics in technology clusters by bringing multi-location firms and their strategic organization into the

picture.

Our findings also have important implications to both innovating firms and policy makers. For firms

making location decisions, this study shows that highly competitive technology clusters are not a

forbidden land for industry leaders. The risk of exposing certain technologies to local competitors is also

low if these technologies are highly dependent on internal resources residing somewhere else. For

policymakers eager to nurture local high-tech industries, our findings suggest that attracting leading

innovators to the location is only part of the job. Tax breaks and other incentives may influence where

R&D is conducted, but not how the R&D projects are actually organized. With local projects closely

intertwined with the firms’ global research agenda, the same R&D budget or R&D intensity may generate

very different knowledge outflows to the local community.

Admittedly, there remain several limitations in this study. For example, this is a one-industry study with a

small group of leading innovators in the industry; further analysis with more diverse contexts will make

the conclusions more generalizable. Also, we rely on the patent data to capture innovation and knowledge

flow, which leaves out other forms of knowledge and knowledge flow. Finally, there may be other

mechanisms at play that allow the industry leaders to alleviate appropriation concerns. For example, given

their common presence in the major clusters around the world, the leading semiconductor firms are in a

typical multi-market contact situation (Bernheim and Whinston 1998) in which competition may be

attenuated. It is impossible for us to outline all the other mechanisms, but the fact that we observe

stronger internal linkages in the presence of competitors – even with the possibilities of multi-market

contact and other attenuating mechanisms – makes our estimates conservative.

This study also points to several avenues for further inquiry. First, although the mechanisms discussed in

this paper are based on multi-location firms, the need to appropriate economic rents from proprietary

innovation applies to any firm or organization. Some aspects of internalization strategies are more

generally applicable, such as the separation of complementary components in a R&D project, but the

implementation of such strategies gets much harder without the geographic separation in multi-location

firms. Further research on internal organization and knowledge appropriation may explore other types of

strategic organization that are less location-specific.

Second, the strategies discussed in this study are based on a well-established set of internal routines and

organizational skills that facilitate the transfer and integration of geographically dispersed knowledge.

Essentially we are talking about the effect of internal linkages (an organization issue) on the assimilation

of R&D knowledge (a technological issue) in the face of competition (a product market issue). However,

organizations evolve slowly and internal linkages take time to develop, so they may not always be in sync

with changes in technological or competitive environments. Moreover, not every firm can manage the

internal linkages with enough efficiency or cost effectiveness. Hence, it is important to understand how

firm heterogeneity affects the applicability of these strategies.

Finally, our arguments revolve predominantly around competition and have excluded the possibility of

inter-organizational cooperation. However, there are frequent project collaborations, strategic alliances,

and industrial consortia among semiconductor firms, universities, and other research institutions.

Sometimes R&D is fragmented across the value chain and outsourced to specialized firms (Arora et al.

2001). In such circumstances, knowledge flow across organizational boundaries is necessary and even

desirable. Studies by Appleyard (1996) in the semiconductor industry and by Schrader (1991) in the

specialty steel and mini-mill industry identify information sharing even among employees of direct

competitors. All this limits firms’ abilities to exercise strict knowledge internalization. More research is

needed for us to better understand how firms protect and extract value from innovations developed within

permeable, changing, and diffuse firm boundaries.

7. References

Alcácer, J. 2006. Location choices across the value chain: How activity and capability influence co-location. Management Sci. 52(10) 1457–1471.

Alcácer, J., M. Gittelman. 2006. Patent citations as a measure of knowledge flows: The influence of examiner citations. Rev. of Econom. and Statist. Forthcoming.

Almeida, P., B. Kogut. 1999. Localization of knowledge and the mobility of engineers in regional networks. Management Sci. 45 905–917.

Anand, B.N., A. Galetovic. 2004. How market smarts can protect property rights. Harvard Bus. Rev., 82(12) 72.

Arora, A., A. Fosfuri, A. Gambardella. 2001. Markets for Technology: The Economics of Innovation and Corporate Strategy. The MIT Press, Cambridge.

Audretsch, D.B., M.B. Feldman. 1996. R&D spillovers and the geography of innovation and production. Amer. Econom. Rev. 86 630–640.

Bartlett, C.A. S. Ghoshal. 1990. The multinational corporation as an inter-organizational network. Acad. of Management Rev. 15 603–625.

Baum, J.A.C., H.A. Haveman. 1997. Love thy neighbor? Differentiation and agglomeration in the Manhattan hotel industry, 1898–1990. Admin. Sci. Quart. 42(2) 304–338.

Baum, J.A.C., O. Sorenson, eds. 2003. Geography and strategy. Advances in Strategic Management, Vol. 20. JAI/Elsevier, Oxford.

Bernheim, B. D., M. Whinston. 1990. Multimarket contact and collusive behavior. RAND J. Econom. 21(1) 1–26.

Buckley, P.J., C.M. Casson. 1976. The Future of the Multinational Enterprises. Macmillan: London.

Buenstrof, G., S. Klepper. 2004. The origin and location of entrants in the evolution of the U.S. tire industry. Papers in Economics and Evolution 2004-07. Max Planck Institute of Economics.

Cassiman, B., R. Veugelers. 2002. R&D cooperation and spillovers: Some empirical evidence from Belgium. Amer. Econom. Rev. 92(4) 1169–1184.

Chung, W., J. Alcácer. 2002. Knowledge seeking and location choice of foreign direct investment in the United States. Management Sci. 48 1534–1554.

Cohen, W.M., D. Levinthal. 1990. Absorptive capacity: A new perspective on learning and innovation. Admin. Sci. Quart. 24 128–152.

Cohen, W.M., R.R. Nelson, J.P. Walsh. 2000. Protecting their intellectual assets: Appropriability conditions and why U.S. manufacturing firms patent (or not). NBER Working Paper 7552.

Cohen, W.M. 1995. Empirical studies of innovative activity. In Handbook of the Economics of Innovations and Technological Change, Paul Stoneman, ed. Wiley-Blackwell, Oxford.

Cummings, J.N. 2004. Work groups, structural diversity, and knowledge sharing in a global organization. Management Sci. 50(3) 352–364.

Dierickx, I., K. Cool. 1989. Asset stock accumulation and sustainability of competitive advantage. Management Sci. 35 (12) 1504–1511.

Dushnitsky, G., J.M. Shaver. 2009. Limitations to interorganizational knowledge acquisition: the paradox of corporate venture capital. Strategic Management J. 30(10) 1045-1064.

Edstrom, A., J.R. Galbraith. 1977. Transfer of managers as a coordination and control strategy in multinational organizations. Admin. Sci. Quart. 22(2) 248-263.

Feinberg, S., A. Gupta. 2009. MNC subsidiaries and country risk: Internalization as a safeguard against weak external institutions. Acad. of Management J. 52(2), in press.

Fleming, L., C. King III, A. Juda. 2006. Small worlds and regional innovation. Organization Sci., forthcoming.

Fosfuri, A., M. Giarratana, A. Luzzi. 2008. The penguin has entered the building: The commercialization of open source software products. Organization Sci. 19 292-305.

Frost, T.S., Birkinshaw, J., Ensign, P.C. 2002. Centers of excellence in multinational firms. Strategic Management J. 23(11): 997-1018.

Frost, T.S., Zhou, C. 2005. R&D co-practice and 'reverse' knowledge integration in multinational firms. J. Intern. Bus. Stud. 36(6): 676-687.

Furman, J.L., M.K. Kyle, I. Cockburn, R. Henderson. 2006. Public and private spillovers, location, and the productivity of pharmaceutical research. Annales d’Economie et Statistique 79/80.

Gans, J., S. Stern. 2003. When does funding research by smaller firms bear fruit?: Evidence from the SBIR program.Economics of Innovation and New Technology, 12(4): 361-384.

Ghemawat, P., C. Thomas. 2008. Strategic interaction across countries and multinational agglomeration: An application to the cement industry. Management Sci. 54(12) 1980–1996.

Gilson, R.J. 1999. The legal infrastructure of high technology districts: Silicon Valley, Route 128, and covenants not to compete. New York University Law Rev. 74 575–629.

Gupta, A.K, V. Govindarajan. 2000. Knowledge flows within multinational corporations. Strategic Management J. 21 473-496.

Hall, B.H., A.B. Jaffe, M. Trajtenberg. 2005. Market value and patent citations. Rand J. of Econom. 36(1) 16–38.

Hannah, D.R. 2005. Should I keep a secret? The effects of trade secret protection procedures on employees’ obligations to protect trade secrets. Organization Sci. 16(1) 71–84.

Helfat, C.E., 1994. Firm-specificity in corporate applied R&D. Organization Sci. 5 173–184.

Henisz, W., J. Macher. 2004. Firm- and country-level tradeoffs and contingencies in the evaluation of foreign investment: The semiconductor industry, 1994–2002. Organization Sci. 15(6) 537–554.

Jacobs, J. 1969. The Economy of Cities. Random House, New York.

Jaffe, A.B. 1986. Technological opportunity and spillovers of R and D: evidence from firms’ patents, profits, and market value. Amer. Econom. Rev. 76(5) 984–1001.

Jaffe, A.B., M. Trajtenberg, R. Henderson. 1993. Geographic localization of knowledge spillovers as evidenced by patent citations. Quart. J. of Econom. 434 578–598.

Jaffe, A.B., M. Trajtenberg, M.S. Fogarty. 2000. The meaning of patent citations: report on the NBER/Case-Western Reserve Survey of patentees. NBER Working Paper 7631.

Kogut, B., U. Zander. 1992. Knowledge of the firm, combinative capabilities, and the replication of technology. Organization Sci. 3(3) 383–397.

Kogut, B., U. Zander. 1993. Knowledge of the firm and the evolutionary theory of the multinational corporation. J. Intern. Bus. Stud. 24 625–645.

Gittelman, M., B. Kogut. 2003. Does good science lead to valuable knowledge? Biotechnology firms and the evolutionary logic of citations patterns. Management Sci. 49(4) 366–382.

Lahiri, N. 2003. Knowledge spillovers: geography, technology and intra-firm linkages. Doctoral dissertation: Ross School of Business, University of Michigan.

Liebeskind, J.P. 1996. Knowledge, strategy, and the theory of the firm. Strategic Management J. 17(Winter Special Issue) 93–109.

Levin, R.C., A.K. Klevorick, R.R. Nelson, S.G. Winter. 1987. Appropriating the returns from industrial R&D. Brookings Papers on Economic Activity 3 783–820.

Mansfield, E., M. Schwartz, S. Wagner. 1981. Imitation costs and patents: An empirical study. Economic J. 91 907–918.

Marshall, A. 1922. Principles of Economics, 8th edition. MacMillan, London.

Nobel, R., J. Birkinshaw. 1998. Innovation in multinational corporations: control and communication patterns in international R&D operations. Strategic Management J. 19(5) 479–496.

Porter, M.E. 1998. On Competition. Harvard Business School Press, Cambridge, MA.

Rajan, R., L. Zingales. 2001. The firm as a dedicated hierarchy: A theory of the origins and growth of firms. Quart. J. of Econom. 116 (3) 805–851.

Sampat, B. 2009. When Do Applicants Search for Prior Art.? J. of Law and Econom. Forthcoming.

Sanna-Randaccio, F., R. Veugelers. 2002. Multinational knowledge spillovers with centralized versus decentralized R&D: A game theoretic approach. CEPR Discussion Paper 3151.

Saxenian, A. 1996. Regional Advantage. Harvard University Press, Cambridge, MA.

Shaver, J.M., F. Flyer. 2000. Agglomeration economies, firm heterogeneity, and foreign direct investment in the United States. Strategic Management J. 21 1175–1193.

Singh, J. 2005. Collaborative networks as determinants of knowledge diffusion patterns. Management Sci. 51(5) 756–770.

Swaminathan, A., J. Delacroix. 1991. Differentiation within an organizational population: Additional evidence from the wine industry. Acad. of Management J. 34(3) 679–692.

Stuart, T.E. 2000. Interorganizational alliances and the performance of firms: A study of growth and innovation rates in a high-technology industry. Strategic Management J. 21(8) 791–811.

Stuart, T.E., J.Podolny. 1996. Local search and the evolution of technological capabilities. Strategic Management J. 17 21–38.

Stuart, T.E., O. Sorenson. 2003. Liquidity events and the geographic distribution of entrepreneurial activity. Admin. Sci. Quart. 48 175–201.

Szulanski, G. 1996. Exploring internal stickiness: impediments to the transfer of best practices within the firm. Strategic Management J. 17 27–43.

Teece, D. 1986. Profiting from technological innovation: implications for integration, collaboration, licensing and public policy. Research Policy 15(6) 285–305.

Teece, D.J., G. Pisano, A. Shuen. 1997. Dynamic capabilities and strategic management. Strategic Management J. 18(7) 507-533.

Trajtenberg, M., R. Henderson, A. Jaffe. 1997. University versus corporate patents: a window on the basicness of invention. Economics of Innovation and New Technology 5(1) 19–50. Also in Patents, citations, and innovations, 2002, ed. A. Jaffe, M. Trajtenberg, p. 51-88. MIT Press, Cambridge, MA.

Vickers, J. 1986. The evolution of market structure when there is a sequence of innovations. J. of Industrial Econom. 35(1) 1–12.

Yoffie, D.B. 1993. Foreign direct investment in semiconductors. Foreign Direct Investment, ed. K.A. Froot, p. 197–230. University of Chicago Press, Chicago. Zhao, M. 2006. Conducting R&D in countries with weak intellectual property rights protection. Management Sci. 52(8) 1185–1199.

Ziedonis, R. H. 2004. Don’t fence me in: Fragmented markets for technology and the patent acquisition strategies of firms. Management Sci. 50(6) 804–820.

Table 2: Summary Statistics

Patent variablesvariable N mean sd min maxcross_cluster_team 7730 0.260543 0.43896 0 1cross_cluster_self_citations 7730 1.225226 5.679622 0 137local_citations_by_innovators 7730 2.355369 7.928347 0 188local_citations_by_profit 7730 2.354981 7.928405 0 188local_citations_by_industry 7730 0.431565 3.729852 0 188local_citations_by_segment 7730 0.439069 3.66472 0 188local_citations_by_competito 7730 0.186546 0.984877 0 37patent_stock 7730 17377.85 9046.405 4304 42096total_citations 7730 8.803105 16.08912 1 201claims 7730 21.03195 18.65432 1 521Cites backward locally? 7730 0.352264 0.477707 0 1Cites backward locally (profit) 7730 0.351876 0.477586 0 1Cites backward locally (indust 7730 0.101035 0.301395 0 1Cites backward locally (segme 7730 0.089651 0.285699 0 1Cites backward locally (compe 7730 0.06934 0.254048 0 1

Fimr-cluster variablesvariable N mean sd min maxwith plant 7730 0.161708 0.368206 0 1with fabless 7730 0.020699 0.142383 0 1in_industry 7730 19.0141 19.6924 0 73firms_not_industry 7730 42.1423 43.89665 2 168in_segment 7730 12.79405 14.04558 0 65firms_not_in_segment 7730 48.36235 50.09096 2 213competitors 7730 6.927943 6.200252 0 27firm_no_competitors 7730 54.22846 59.45098 2 234non-profit 7730 4.653299 4.398021 0 16

Cluster-variablesvariable N mean sd min maxplants 100 13.45918 12.89543 1 54fabless 100 15.9898 41.2894 0 206publications 100 136.4898 147.3925 1 754entities 100 38.43878 45.35838 4 253firms 100 35.09184 42.84185 3 240non-profit 100 3.346939 3.446504 0 16large firms 100 27.56122 30.50032 2 153small firms 100 7.530612 14.12259 0 90universities 100 2.030612 2.695755 0 14gov't entities 100 1.091837 1.385393 0 6other non-profit 100 0.22449 0.508294 0 2

Table 2: Correlation Matrix

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

1 cross_cluster_team 12 cross_cluster_self_citations 0.1923 13 total_citations 0.105 0.4695 14 Cites backward locally (competitor)? -0.0042 -0.0154 -0.0031 15 with plant -0.2972 -0.2883 -0.1182 -0.0032 16 with fabless -0.0242 -0.0267 -0.0184 0.0319 -0.0744 17 plants -0.1399 -0.096 -0.0449 0.2209 0.0302 -0.0782 18 fabless -0.0156 -0.0142 0.0393 0.0446 -0.2037 0.0169 0.5133 19 publications -0.122 -0.1103 -0.0909 0.2377 0.1632 -0.053 0.8023 -0.0213 1

10 claims 0.0959 0.1153 0.1248 0.0458 -0.162 -0.0396 -0.0605 0.0066 -0.0928 111 entities -0.05 -0.0488 0.0036 0.1571 -0.1005 -0.0704 0.8412 0.7814 0.5054 -0.038 112 firms -0.0552 -0.0508 0.003 0.1569 -0.0977 -0.0716 0.846 0.7923 0.4988 -0.0369 0.999 113 non-profit 0.0364 -0.0097 0.0105 0.1226 -0.117 -0.0373 0.5749 0.4414 0.4831 -0.0445 0.7798 0.7518 114 large firms -0.0796 -0.0654 -0.0154 0.1883 -0.0451 -0.0931 0.9236 0.6748 0.6519 -0.0505 0.9749 0.9755 0.7385 115 small firms 0.0067 -0.0105 0.0421 0.0626 -0.1942 -0.0132 0.5346 0.9092 0.0845 -0.0014 0.8799 0.8815 0.6518 0.7561 116 universities 0.0766 0.0428 0.0465 0.0599 -0.1402 -0.0555 0.3412 0.3435 0.2556 -0.0141 0.5823 0.5493 0.9202 0.5104 0.5388 117 gov't entities -0.0777 -0.1068 -0.0666 0.171 0.0137 0.019 0.6931 0.3348 0.6603 -0.0746 0.7003 0.6983 0.5653 0.7497 0.4684 0.2156 118 other non-profit 0.0251 -0.0341 -0.0211 0.0516 -0.0468 0.0149 0.1709 0.2129 0.1032 -0.0325 0.1766 0.1797 0.091 0.1709 0.1679 0.0022 -0.0531 119 competitors 0.1191 0.1355 0.0801 0.2094 0.0542 -0.071 0.8438 0.1743 0.8666 -0.0861 0.6304 0.628 0.5182 0.745 0.269 0.2985 0.6521 0.1253 120 firm_no_competitors -0.0462 -0.0398 0.0116 0.1448 -0.1095 -0.0686 0.8106 0.8234 0.4394 -0.0302 0.9954 0.9967 0.7445 0.9585 0.9082 0.5523 0.6737 0.1778 0.5627 1

Table 3: Logit estimates on cross-cluster teams

Table 4: Negative binomial on cross-cluster self-citations

Table 6: Negative binomial on local citations by competitors