Geographies of Knowledge Sourcing and the Complexity of Knowledge in Multilocational Firms

Abstract

The rise of the knowledge economy has placed innovation at the center of models of competitive advantage. Access to more valuable forms of knowledge remains contested as the geography of its production is uneven and as some knowledge assets are relatively immobile. Within this fractured knowledge landscape multilocational firms have clear advantages. They can exploit numerous localized pools of knowledge, they can shape the character of knowledge development in different places, and they have some control over who can tap local knowledge assets. Surprisingly, we still have little detailed knowledge of the technologies developed by multilocational firms across the sites where they are active. We augment the literature on multiunit firms on three fronts. First, we make use of the rich, technological information in patent data to show that multilocational firms operating research and development (R&D) units across US metropolitan areas produce different kinds of technological knowledge over space. Second, we provide quantitative evidence of geographic knowledge sourcing by linking the technologies produced within the R&D units of these firms to the knowledge stocks generated within the cities where they are located. Third, we report that as the number of R&D units within multilocational firms increase, so, up to a limit, the complexity of the knowledge those firms generate also increases. We show that these complexity gains are linked to the volume of knowledge sourced from local partners and to the integration of knowledge across units of the multilocational firm.

Key words:

• knowledge sourcing
• multilocational firms
• complex knowledge
• US cities

Access to knowledge is key to competitive advantage in the contemporary economy. However, the technological knowledge that underpins innovation is not an undifferentiated commodity. Complex knowledge, which combines numerous technological components and capabilities, is more valuable to firms and regions because it provides a more enduring source of competitive advantage (Kogut and Zander 1992). With invention broadly conceived as a process of recombining existing knowledge in new ways (see Schumpeter 1939; Arthur 2009), multilocational firms that can access different combinations of technologies may enjoy significant advantages over rivals who can search only a limited range of knowledge stocks.

The knowledge subsets that form the building blocks of valuable, new knowledge combinations are not ubiquitous (Balland and Rigby 2017). They are developed at specific times and in select locations by industry-specific clusters of firms and other organizations (Ellison and Glaeser 1997). Over rounds of investment, different knowledge types are mixed together in particular places; only some diffuse across space via networks and more ephemeral linkages that bind economic agents to one another (Bathelt et al. 2004; Maskell, Bathelt, and Malmberg 2006). The resulting knowledge landscape is highly differentiated with deep accumulations of more and less complex ideas found in a relatively small number of locations and much shallower deposits elsewhere (Balland et al. 2020).

Across this heterogeneous knowledge landscape, multilocational firms might develop advantages over single-plant firms. First, they can locate establishments in different knowledge environments to access multiple, localized forms of buzz and spillovers (Cantwell 2017). Second, embedding plants in different clusters may overcome constraints on effective knowledge flows imposed by different forms of proximity (Boschma 2005). Third, local embedding also gives multiplant firms the ability to shape the character of knowledge development over space as well as some control over who is able to exploit specific knowledge assets (Alcácer and Zhao 2012). Firm organization has evolved, at least in part, to take advantage of spatial variations in knowledge stocks, notwithstanding the difficulties of integrating activities with numerous local partners and across the heterogeneous units of the firm itself (see Tushman and O’Reilly 1996; Andersson, Forsgren, and Holm 2002).

Within economic geography, considerable attention has focused on knowledge sourcing and, in particular, whether firms access knowledge that is local or more distant (Bathelt et al. 2004; Tödtling, Lehner, and Trippl 2006; Trippl, Tödtling, and Lengauer 2009; Fitjar and Rodríguez-Pose 2011a). Yet, with only a few exceptions (see Wang 2015; Phelps and Fuller 2016; Li and Phelps 2019; Bathelt and Li 2020), the knowledge sourcing literature within economic geography has rarely explored how firm organization may be linked to sourcing strategies. In contrast, the international business and management fields offer a rich literature detailing how the structure of multinational enterprises (MNEs) has shifted to explore and exploit the uneven knowledge landscape across countries (Cantwell 1989; Kogut and Zander 1993; Cantwell and Mudambi 2005, 2011). Although economic geographers highlight the subnational heterogeneity of technological knowledge across cities and regions (Kogler, Rigby, and Tucker 2013), they have not shown that multilocational firms benefit from distributing their research and development (R&D) across locations within a single country. Furthermore, the specific kinds of knowledge that multiunit firms source from different locations, especially at the subnational level, is unknown beyond a small number of industries and aggregate technology types.

The primary value added by this article is a detailed exploration of the technological heterogeneity of knowledge generated within the R&D units of multilocational firms operating across metropolitan areas of the US. Through such analysis, we hope to draw together work in economic geography that celebrates the variation in knowledge assets over space with work in cognate fields that examines how multilocational firms structure their organization to exploit these geographic opportunities, following the calls of Cano-Kollman et al. (2016), Bathelt, Cantwell, and Mudambi (2018), and Mudambi et al. (2018). Our investigation focuses on three research questions. First, do multiunit firms generate different kinds of knowledge within the plants that they control? Second, is there evidence that these firms locate plants in particular US metropolitan areas to gain access to specific kinds of technological know-how? Third, are firms that distribute their R&D over multiple locations able to integrate different knowledge types and raise the complexity of the knowledge they produce?

The rest of the article is separated into four sections. A review of the knowledge sourcing literature within the fields of economic geography and international business, which helps situate the hypotheses we examine, is provided in the next section. This is followed by a brief overview of the data sources that underpin our research. Analysis of the research questions occupies the penultimate section. A brief conclusion ends the article where we highlight the connections between our results and related research within economic geography.

Literature Review

Within an increasingly integrated market economy, innovation has become critical to firm performance. Extending the resource-based view of the firm, knowledge-based visions tie the structures and routines of business organizations squarely to the search, production, exploitation, and safeguarding of technological know-how (Kogut and Zander 1992; Grant 1996). How do firms develop the technologies needed for innovation? An older vision of technological change focuses on cost reductions resulting from learning-by-doing and from dedicated processes of R&D that occur largely within the boundaries of the firm (Arrow 1962; Pavitt 1991). Newer visions of knowledge development see those boundaries as more porous, allowing the flow of technological ideas between interacting economic agents. Since Rosenberg (1982) highlighted the information exchange between producers and users of new technologies, research on external (to the firm) knowledge acquisition developed rapidly, focused on interactions of various sorts that occur between firms and their customers and suppliers, with universities and other research organizations (von Hippel 1988; Lundvall 1992). In Chesbrough’s (2003) open innovation framework, firms are nodes in heterogeneous networks of organizations that integrate external knowledge into their own routines of technology development and competition.

Localized Knowledge Sourcing

For economic geographers, the benefits of knowledge sharing across local communities of economic agents have long been seen as the foundation for industrial clusters or agglomerations. For Marshall (1920), knowledge sharing among networks of buyers and suppliers, the emergence of common knowledge atmospheres, the availability of deep pools of specialized workers and local knowledge spillovers sustain distinctive communities of economic agents. Jacobs (1969) imagines the benefits of cities in terms of local diversity in knowledge possibilities, consistent with the notion of recombinant invention. For Saxenian (1994) and Storper (1997), following Brown and Duguid (1991), place-specific communities of practice and institutional development foster collective learning (Lawson and Lorenz 1999) and shared capabilities that sustain more open forms of knowledge development and mobility. Through work on industrial districts and milieu (Camagni 1991) to the learning regions of Lundvall and Johnson (1994), the benefits of local knowledge sharing are celebrated. A key takeaway from much of this literature is that firms and other economic agents who are part of an innovative cluster are advantaged in ways that outsiders are not. Being there, as Gertler (2003) notes, smooths the sharing of ideas, whether they are generated by local buzz (Storper and Venables 2004), or delivered in the form of spillovers (Jaffe, Trajtenberg, and Henderson 1993), through worker mobility (Boschma, Eriksson, and Lindgren 2014) or more formal collaboration networks (Owen-Smith and Powell 2004; Giuliani 2013).

Amin and Cohendet (1999) raise important questions about the geographic scale at which knowledge sharing is considered effective, questioning the primacy of the local. Bathelt et al. (2004) extend these concerns, focusing on different processes of interaction within and between industrial clusters in their buzz and pipelines model. Combining arguments from the regional innovation systems literature (Asheim and Gertler 2005) and from knowledge-based ideas (Asheim and Coenen 2005; Martin and Moodysson 2011), recent work in economic geography explores the sources of external knowledge, the mechanisms through which that knowledge is obtained, and the geographic scale at which those mechanisms operate. Based on firm survey research, Tödtling, Lehner, and Trippl (2006) explore the nature of innovation and knowledge sourcing patterns across different sectors within the Austrian economy. They report significant differences between sectors with analytical and synthetic knowledge bases in terms of the reliance on internal (firm-specific) and external knowledge development, on external sources of new technology, the variety of sources used, and the geography of those sources. These ideas are extended in Trippl, Tödtling, and Lengauer (2009) who detail the range of external knowledge linkages found within the Vienna software industry. Here more formal R&D partnerships are reported to operate at local and national scales, whereas more informal knowledge networks are found at all spatial scales. Tödtling, Lehner, and Kaufmann (2009) show that incremental and more advanced innovations rely on specific kinds of knowledge interactions. Grillitsch and Trippl (2014) report similar findings in the Austrian automotive supply sector, and Isaksen and Trippl (2017) link regional innovation patterns to modes of innovation, regional innovation systems, and types of knowledge linkages. In related work, Fitjar and Rodríguez-Pose (2011a) report on the nature and geography of interfirm collaborations in Norwegian innovation, while Fitjar and Rodríguez-Pose (2011b) suggest that cognitive and organizational proximity with nonlocal partners are key to innovation in peripheral regions of Norway.

While this work is very rich in terms of detailing the geographic sources of knowledge tapped by different communities of economic agents, it has tended to ignore how firms might organize their structure to take advantage of the geographic unevenness in knowledge assets. Further, the knowledge sourcing literature provides few details of the types of knowledge firms obtain from their nonlocal partners and, thus, whether they source place-specific knowledge assets remains unclear. Here, we seek to complement the existing literature by exploring how multilocational firms obtain different kinds of technological knowledge across US urban areas. In these efforts we follow important new extensions in knowledge sourcing by Li and Bathelt (2018) and Bathelt and Li (2020) who outline the processes through which knowledge-seeking firms extend their geographic footprints, building the nonlocal linkages and networks that leverage technological capabilities found in different communities of actors (see also Cantwell and Iammarino 2005; Maskell, Bathelt, and Malmberg 2006; Phelps and Fuller 2016; Crescenzi and Gagliardi 2018).

Multilocational Knowledge Sourcing

The international business and management literatures have been much more focused on the evolution of firm organization to both exploit and explore knowledge opportunities in different locations. Early explanations for the existence of the MNE were related to the internalization of technology transfer and exploitation of the sunk costs of centralized R&D in different markets (Grant and Phene 2022). Bartlett and Ghoshal (1986) provide early challenges to this restricted vision of MNE knowledge creation. Consistent with knowledge-based visions of the firm (Grant 1996), new models of MNE organization highlight the geographic heterogeneity of knowledge (Cantwell 1989; Kogut and Zander 1993) and management structures that facilitate the acquisition and integration of that knowledge within the firm (Gupta and Govindarajan 1991; Andersson and Forsgren 1996; Birkinshaw and Hood 1998). For Foss and Pedersen (2002), Almeida and Phene (2004), and Cantwell and Mudambi (2011), the geographic expansion of MNE activity is explicitly understood as a strategy to access diverse knowledge assets and enhance opportunities for technological recombination.

Extensive reviews of the internationalization of R&D in MNE subsidiaries are provided by Niosi (1999), Kafouros, Buckley, and Clegg (2012), and Papanastassiou, Pearce, and Zanfei (2020). These articles make clear the growing international fragmentation of R&D and the differential use of subsidiaries to exploit, create, and augment firm knowledge assets. While most of the MNE literature examines the distribution of R&D activities at the international level, a series of articles ties the location of R&D subsidiaries to subnational centers of specialized knowledge production (Florida and Kenney 1994; Almeida 1996) and to well-connected global city-regions (Castellani et al. 2022) with high potential for agglomeration externalities (Alcácer and Chung 2014). Low-tech MNEs are more likely to avoid locating in dense business agglomerations, whereas R&D-intensive firms strategically target these locations (Chung and Alcácer 2002; Phelps 2008).

More explicit connections between subsidiary knowledge creation and the technological specialization of host economies are generated by Cantwell (1989). In subsequent work, Cantwell and Iammarino (2000) and Cantwell and Piscitello (2002) trace the correspondence of broad knowledge types generated by MNE subsidiaries and domestic firms across European countries and regions. However, Dunning and Narula (1995) report little evidence of the concentration of foreign R&D in the technologies where host economies exhibit comparative advantage, and Zander (1997) reports mixed findings on this question. In newer work, Phene and Tallman (2018) use detailed patent class data to explore how US semiconductor MNEs develop competence-exploiting and competence-exploring capabilities across their global subsidiaries. Along with Cantwell (2017) and Mudambi et al. (2018), Phene and Tallman (2018) call explicitly for more extensive and disaggregate comparisons of innovation activities in multilocational firms over space.

Outside case study work on selected MNEs, we do not know very much about the knowledge-sourcing activities of multilocational firms, foreign and especially domestic, operating within a single country. This is very much a missed opportunity within economic geography. While considerable work celebrates subnational differences in the structure of knowledge, to what extent individual firms separate their knowledge production efforts to take advantage of these differences is unclear. Do firms with multiple R&D locations produce similar or different kinds of knowledge over space and, if the latter, is the knowledge generated by firms in particular facilities linked to the broader knowledge stocks of the cities and regions where those facilities are placed? These questions prompt our first two hypotheses:

H₁: The individual R&D units within the multilocational firm produce different kinds of technological knowledge.

H₂: The R&D units of multilocational firms produce technology that is similar to that found in the cities where they are located.

Multilocational Firms, R&D Integration, and Knowledge Complexity

Mining diversified knowledge stocks through the geographic separation of R&D units may increase recombinant knowledge possibilities and thus potential gains in the volume and quality of innovative output. However, at this time, the empirical links between geographic knowledge sourcing and returns to the firm are less clear than expected. In a broad study of MNEs, Singh (2008) reports a negative relationship between the geographic dispersion of a firm’s R&D activities and the forward citation value of patents. He attributes this to the difficulty of integrating knowledge sourced from different countries. Lahiri (2010), Lecocq et al. (2012), and Scalera, Perri, and Hannigan (2018) take this issue to the intranational scale and report inverted U-shaped relationships between the geographic scope of R&D and innovation quality, noting increased costs as the number of innovating units increases. Reporting on manufacturing firms operating within Finland, Leiponen and Helfat (2011) find that multiple R&D locations raise the innovative output for Finnish firms, though that effect is limited to imitative technologies rather than those that are new to the market.

In the analysis below, we link geographies of knowledge sourcing to the complexity of technologies that multilocational firms generate. Here we build on Kogut and Zander (1993), Foss and Pederson (2002), and Cantwell and Mudambi (2011) who regard knowledge sourcing as a strategy to build competitive advantage through accessing diverse technological know-how and integrating that know-how in new knowledge possibilities. Building on the new literature in economic complexity (see Balland et al. 2022 for an overview), the complexity of knowledge is linked to the number of technological components from which it is assembled and by the difficulty of combining those components (Kogut and Zander 1992; Fleming and Sorenson 2001; Hidalgo and Hausmann 2009). Less complex forms of knowledge integrate relatively few technological components, the properties of which are broadly understood by economic agents. Consequently, less complex forms of knowledge are relatively ubiquitous and generate little by way of competitive advantage for their creators. More complex forms of knowledge integrate many more technological components, the properties of which are, generally, less well understood. These technologies are produced by a small number of firms and broader collectives of economic agents that in combination possess the diversified sets of specialized skills and capabilities from which complex knowledge is produced. The difficulty of generating complex knowledge means that it is relatively rare and thus provides a more enduring base for competitive rents. Alcácer and Zhao (2012) and Berry (2014) extend these claims to the context of multinational firms, arguing that the spatial dispersion and recombination of rare components within the organization can generate effective barriers against imitation.

At this time, there is a rapidly growing body of evidence, developed at the national and subnational level, that establishes a significant, positive relationship between the complexity of knowledge and growth (Petralia, Balland, and Morrison 2017; Balland et al. 2020; Mewes and Broekel 2022; Pintar and Scherngell 2022; Rigby et al. 2022; Li and Rigby 2023). Emerging links between knowledge complexity and performance also suggest that complexity is a useful indicator of competitive advantage within the firm (Audretsch and Belitski 2021). With the ability to access multiple sites of knowledge production, multilocational firms have considerable advantages over single plant firms in terms of accessing the varied components of more complex forms of knowledge, of embedding in the communities where such knowledge is produced, and learning how those different knowledge components may be integrated. However, the difficulties of managing such integration remain substantial as Berry (2023) makes clear. Fuller and Phelps (2018) argue that effective integration rests upon the firm’s orchestration capabilities, and its ability to manage the creation, configuration, and diffusion of dispersed assets. Integration mechanisms can be more or less tangible, from the establishment of shared values, culture, and institutional norms (Kogut and Zander 1993; Nohria and Ghoshal 1997), to more concrete mechanisms, such as the mobility of managers (Berry 2015) and inventors (Singh 2008; Castellani et al. 2022), as well as interunit collaboration (Berry 2014).

We revisit the relationship between multilocational firms, integration, and the complexity of knowledge in Hypotheses 3 and 4. In Hypothesis 3, we assert that as the firm expands its R&D activity geographically, it increases its recombination possibilities and thus the potential for developing more complex technologies. Hypothesis 4 claims that greater integration between subunits of the firm and between those subunits and economic agents in host cities supports the development of more complex technologies.

H₃: Multilocational firms that increase the number of their spatially distributed R&D units produce more complex knowledge.

H₄: Higher levels of R&D integration within the firm and between the firm’s R&D units and other economic agents within host cities are associated with more complex knowledge.

Data

Innovation activity in multilocational firms is explored using patent records from the US Patent and Trademark Office (USPTO) and data from the Bureau Van Dijk (BVD). The BVD data track connections between firms and establishments linked by ownership, while patent data capture the structure of technology developed within firms and cities. Granted utility patents are the focus of the analysis, disaggregated into 652 (four-digit) subclasses of the Cooperative Patent Classification (CPC). We use the four-digit CPC subclasses in our investigation because they link most closely to distinct technologies and because the four-digit subclass level of aggregation is the most widely used (Lobo and Strumsky 2019). Patent ownership is indicated by assignee information.

To determine the location of an invention, researchers generally use the address(es) of inventors rather than the assignee, since the latter often records the location of the headquarters (HQ) or the intellectual property department of an organization instead of the site of invention. USPTO records do not directly link inventors to an organization’s subunits. This is problematic in instances where inventors live in a geographic area different from that of the assignee listed on the patent record. To more accurately match the location of inventors and assignee organizations, geographic and ownership information for corporations with at least two R&D plants operating in the US were extracted from the BVD’s Orbis database. R&D units are defined as those that produce patents. Analysis in this article focuses only on these R&D units, and counts of plants within multilocational firms are based only on the different R&D facilities.

Orbis records were linked to USPTO patent assignee names using fuzzy matching based on a Jaro-Winkler algorithm. We further added information on majority-owned subsidiaries and tracked changes in ultimate ownership using merger and acquisition data from the Zephyr database. Patents from subsidiaries are attributed to a parent firm only in instances where the application year on the patent matches the period of actual subsidiary ownership.

To construct the final sample, patents were assigned to a specific establishment of an assignee if the location of the inventor coincided with the HQ or one of the non-HQ establishment (branch or subsidiary) locations of the assignee organization. In cases where a patent was assigned to firm X, and when the inventor on the patent was located in city A, if city A contained no establishment associated with firm X, the patent record was dropped from the analytical sample because we have no way of clearly linking the inventor to one of the R&D units of the firm. The analysis covers the years 2001–15. The geographic matching was performed across the 381 metropolitan statistical areas (MSAs) of the US. Research focused on firms (assignees) that patented from at least two establishments located in different metropolitan areas. In rare cases where a multilocational firm operated two branches in the same city in a given year, the data were combined to generate an aggregate unit. Patent counts were fractionally split across metropolitan areas when co-inventors were located in different MSAs, and in instances where only some inventor locations matched the establishments, the patent was also assigned fractionally to the matched locations. More information on data collection and development can be found in Appendix 1 in the online material.

The data cover 357,459 patents, 913 unique multilocational firms, and 5,357 establishments. These firms were responsible for 27 percent of all US utility patents granted from 2001 to 2015. Inventive intensity varies markedly across firm assignees, with some firms producing only two patents between 2001 and 2015, while the most inventive firm generated in excess of fifty thousand patents. On average, firms introduced patents from five locations, ranging from two to eighty-five, and generated patents in thirty-one different (CPC four-digit) technological classes. Finally, based on the location of the global HQ, 145 firms in our sample were identified as foreign.

Empirical Analysis

Testing Hypothesis 1: The Individual R&D Units within the Multiplant Firm Produce Different Kinds of Technological Knowledge

This first stage of the empirical analysis explores whether multilocational firms produce different technologies across the R&D establishments that they control. For all firm R&D locations, we collect detailed technology class data on patents generated. A standard test would be some variant of analysis of variance to explore how the distribution of patents across technology classes within plants varies about the firm mean distribution. In this case, patents are allocated across 652 (four-digit) classes of the CPC demanding use of multivariate analysis of variance (MANOVA). However, the large number of zeros in the data at the establishment by class level violates assumptions of MANOVA, and so a permutation based MANOVA (PERMANOVA) is employed. PERMANOVA is a nonparametric test that operates over a geometric partitioning of the variation in a data set that may be linked to multiple factors (Anderson 2017). In this case, the partitioning is based on a measure of distance between technologies generated within the establishments of the multiunit firm. Statistical inference is obtained by the permutation of patent observations across a firm’s plants, from which we obtain a distribution of randomly generated F-statistics. The null hypothesis is that observed differences between the technology centroids of a firm’s establishments are not statistically different from what would be observed if the firm’s patent data were randomly distributed across establishments. In all the statistical analysis, firm and establishment technology class count data were aggregated across three-year periods (2001–2003, 2004–2006, . . . 2013–15) to dampen annual fluctuations in patenting behavior. The temporal dimension in the data also provides a measure of the variance of technology within individual establishments.

To operationalize the PERMANOVA, the distance between technological classes is calculated from measures of the technological relatedness between classes (Breschi, Lissoni, and Malerba 2003; Kogler, Rigby, and Tucker 2013). Relatedness measures are generated between all technology classes based on the co-occurrence of those classes across USPTO patent records. Dissimilarity scores between all pairs of establishments within each multilocational firm are then calculated based on the inverse of the relatedness measures and from technology class data recorded on the patents each establishment produces. The dissimilarity scores are generated for each of the three-year time periods during which establishments produce patents.

The within-group variance for a given establishment is given by the technological dissimilarity within that plant between time periods, while the between-group variance is represented by the distance between the centroids of the establishments of a given multilocational firm. For the permutations, the model randomly assigns patent-class observations within a firm across its establishments, generating a new, simulated F-statistics for each of the random assignments. The observed F-statistic is compared to the distribution of simulated F-statistics to obtain a p-value for the PERMANOVA test statistic. To get a clearer picture of the extent to which firms differentiate their technological production across locations, separate PERMANOVA tests were performed for each individual firm. Because of repeated hypothesis testing an adjusted p-value is reported following the false discovery rate (FDR) approach.

Table 1 presents the results of the PERMANOVA tests by firm-size groups. The percentage of tests in each group yielding significant results at the 0.1, 0.05, and 0.01 confidence levels are shown. The right-hand side columns present the results with p-values adjusted for repeated hypothesis testing (FDR). The results are displayed only for firms where the overall number of patents allowed 999 permutations to be performed. Examining all firms produces no qualitative differences in results. In summary, using the adjusted p-values, slightly more than half of the firms in the sample produce different kinds of technologies across their establishments. However, there are important differences by firm size, measured by the number of locations across which they patent. Almost 70 percent of firms with more than two plants report significant differences in the type of patents produced within their plants, and this share exceeds 80 percent for firms with more than five locations.

Table 1 Summary of Firm-Level PERMANOVA Results

Firm Size # of Plants	# of Firms	# (%) of Firms with Significant Permanova (Permutations = 999)		# (%) of Firms with Significant Permanova Adjusted for FDR (Permutations = 999)
2	211	66 (31.3) 43 (20.4) 4 (1.9)	p < 0.1 p < 0.05 p < 0.01	37 (17.5) 13 (6.2) 0 (0)	p < 0.1 p < 0.05 p < 0.01
3–5	251	163 (64.9) 143 (57) 106 (42.2)	p < 0.1 p < 0.05 p < 0.01	137 (54.6) 114 (45.4) 63 (25.1)	p < 0.1 p < 0.05 p < 0.01
6–10	112	91 (81.2) 88 (78.6) 79 (70.5)	p < 0.1 p < 0.05 p < 0.01	85 (75.9) 81 (72.3) 68 (60.7)	p < 0.1 p < 0.05 p < 0.01
11–20	71	62 (87.3) 62 (87.3) 60 (84.5)	p < 0.1 p < 0.05 p < 0.01	62 (87.3) 60 (84.5) 59 (83.1)	p < 0.1 p < 0.05 p < 0.01
20	43	41 (95.3) 41 (95.3) 41 (95.3)	p < 0.1 p < 0.05 p < 0.01	41 (95.3) 41 (95.3) 40 (93)	p < 0.1 p < 0.05 p < 0.01
All firms	688	423 (61.4) 377 (54.8) 290 (42.5)	p < 0.1 p < 0.05 p < 0.01	362 (52.6) 309 (44.9) 230 (33.4)	p < 0.1 p < 0.05 p < 0.01

Notes: 224 tests are not reported because the number of permutations was lower than 999 and the test generated missing values.

The results generally support Hypothesis 1 because within most multilocational firms, technologies varied significantly between establishments. Figure 1 provides an illustrative example of technological differentiation across the R&D units of Ericsson, the Swedish telecommunications company. Ericsson generated patents in seven US establishments over the period 2001–15. Ordination techniques are used to visualize how the technologies produced differ between establishments located in different cities. From post hoc analysis, 50 percent of establishment pairs within Ericsson show significant differences in patenting class profiles. The Dallas and Durham establishments are quite distinct from other units within the firm (five of six of their pairwise tests are significant), while R&D establishments with fewer patents tend not to be distinctive. Three establishments are presented in Figure 1: Dallas, Durham (now closed), and Santa Clara. The fourth plot includes all seven R&D establishments. The three locations vary in terms of the technologies produced, as indicated by the locations of the centroids (larger shapes). The patterns for the Santa Clara establishment are less visible graphically, as the great majority of its patents are within the same class, and all overlap at the location of the centroid. Appendices 2 and 3 in the online material offer more analysis of the differences in technologies between R&D establishments within firms, and they provide additional case studies of individual firms.

Figure 1. Technology ordination plots, Ericsson.

Testing Hypothesis 2: The R&D Units of Multilocational Firms Produce Technology That Is Similar to That Found in the Cities Where They Are Located

The results above indicate that a large proportion of multilocational firms operating in the US produce different kinds of technologies across the establishments where they patent. However, these results alone do not confirm that corporations with multiple R&D locations are engaged in knowledge sourcing. To strengthen the argument for geographic knowledge sourcing, it would be useful to link the technology generated within a firm’s R&D units to that available in the cities where those units are located. An empirical test of this relationship is developed in the following way. Assume that a firm has three R&D plants located in cities A, B, and C. For this set of three plants and three cities, nine observations are added to a test data set. For three of these observations, where a plant is linked to the city where it is actually located, the dependent variable is coded 1. For the remaining six observations, plants are linked to cities where they are not located. For these observations, the dependent variable is coded 0. This data structure is repeated across time periods where all firms are active and for all other firms in the multilocational firm database. The binary nature of the dependent variable suggests a logit model, where we seek to predict the correct location of a firm’s plants using a single independent variable. That variable is a measure of the technological similarity of each plant–city pairing. The similarity index counts the number of technology classes in which a city and an establishment both exhibit revealed technological advantage (RTA). The observed similarity count is standardized with an expected value based on the probability of observing RTA technology matches, taking the form of a cosine index (see van Eck and Waltman 2009). Period fixed effects are added to the regression model to control for time-specific shocks and firm fixed effects control for unobserved firm heterogeneity. This model is not offered as an explanation of firm location, merely as a way of testing the significance of the correspondence between technologies produced within a firm’s establishments and across the cities where those firms are located. Note that in all comparisons of R&D plant and city technologies, the plant knowledge stocks are removed from the city prior to analysis.

Hypothesis 2 anticipates a positive coefficient for the standardized technological similarity index in the logistic regression. The results from this exercise are presented in Table 2. Firm and time-period fixed effects (FE) are employed in the logistic regression, and standard errors are robust. Note that clustering standard errors at the firm level does not substantially alter the results presented. The index of technological similarity between establishments and the locations where a firm’s R&D units are active has a positive sign and is statistically significant. This result provides additional support that the establishments of multilocational firms absorb and develop the technological know-how of the cities in which they are located, confirming Hypothesis 2. Whether or not these establishments are strategically located in different cities to access place-specific knowledge awaits further analysis. However, these results strongly suggest that possibility.

Table 2 The Similarity of Technologies in R&D Plants and Their Metropolitan Hosts

	Dependent Variable: Binary Plant–City
Tech similarity index	1.6144*** (0.1181)
Constant	-0.0154***
	(0.7133)
Observations	308,123
Pseudo R-squared	0.2047
Period fixed effects (FE)	Yes
Firm FE	Yes

Robust standard errors are in parentheses

*** p < 0.01, ** p < 0.05, * p < 0.1

Testing Hypothesis 3: Multilocational Firms that Increase the Number of Their Spatially Distributed R&D Units Produce More Complex Knowledge

Multilocational firms operating in the US tend to produce different kinds of technological knowledge across the R&D plants that they control, and the knowledge generated within those plants is related to the knowledge stocks of the cities where they are located. If multilocational firms are strategic in their knowledge sourcing activity, we might expect to see economic returns to their actions, returns that may highlight their ability to gather diverse knowledge assets and recombine them in productive ways. To date, research has produced mixed results on this question. In this subsection, we revisit the relationship between the value of knowledge produced by multilocational firms and the geographic extent of their R&D activity. We make two important changes to past work referenced earlier. First, rather than measuring the geographic dispersion of R&D activity by simply looking at the spatial distribution of inventors, we count the number of different establishments across which a firm distributes its knowledge production. This reduces noise in the dispersion measure that results from not linking inventors to a physical R&D unit of the firm. Second, rather than valuing patents by their forward citations, we capture the value of patents using measures of knowledge complexity. There is considerable noise in using forward citations as a measure of knowledge value (Bessen 2008). The complexity of knowledge provides an interesting new proxy for knowledge value as discussed above, and so we explore the utility of this measure in the analysis below.

The veracity of Hypothesis 3 is tested in a regression model that links the complexity of patents generated by multilocational firms to the number of R&D units that they control. The complexity of CPC patent classes, for the period 2001–15, is calculated using the method of reflections outlined in Hidalgo and Hausmann (2009). Complexity scores for individual patents are measured as weighted averages of the complexity values assigned to each of the classes listed on patents. We do not average patent complexity values at the firm level. Rather, our units of observation are individual patents produced by the multilocational firms in our sample. Focusing analysis at the patent level permits us to control for a series of patent characteristics that have been shown to influence patent values, as we report below.

The main explanatory variable in our model is the (log) number of establishments (ln # establishments) across which multilocational firms distribute their patent production. This correlates perfectly with the number of cities across which these firms produce patents. Rather than assume a monotonic relationship between the number of R&D units and the complexity of a firm’s patents, we allow that relationship to take a quadratic form by adding a squared-term on the (log) number of establishments.

Linkages between the R&D units of firms and the economic communities within which they are differentially embedded are captured by counts of the number of citations that those R&D plants make to patents generated by other firms within the same metropolitan area (# local citations external to firm). We anticipate that sourcing local knowledge from outside the firm will increase the complexity value of a focal firm’s patents, since it adds to the stock of technological information that the firm can process. To test Hypothesis 4, the integration of knowledge within the firm is captured in two ways. First, we count the number of citations that link patents produced in different R&D units within the firm in the three years prior to each of our three-year study periods (# citations between plants). Second, we count patents within the multilocational firm that are produced by co-inventors working in different R&D establishments within the firm in the three years before each study period (# co-invention patents).

The value of patents is often linked to the number of inventors, the number of technology classes, and the number of knowledge claims they each make (Breitzman and Thomas 2015). We include a series of additional variables in the model to control for these patent-level variables (# inventors, # technology classes, # knowledge claims, respectively). To these, we add dummies capturing whether the patent was developed in the HQ of the firm (HQ dummy) and whether the parent firm is foreign (if the global HQ is located outside the US) (Foreign firm dummy). To proxy for the local knowledge available to the firm through its network of establishments, we include a time-varying measure of the number of patents generated in the cities where the patent is developed (City size). Finally, to control for the volume of innovations in the firm and to more carefully isolate the role of the number of R&D units within the firm, we add a time-varying measure of firm size measured as the number of patents granted to the firm each period (Firm size). Firm, time period, and primary class fixed effects are added to control for period-specific shocks in the average value of complexity, heterogeneity across broad technological domains, and unobserved heterogeneity within the firm that does not vary over time. The results of the analysis are displayed in Table 3. We estimate a series of ordinary least squares regression models with standard errors that are robust to heteroscedasticity. Where patents are produced as a result of collaboration between a firm’s plants, they appear more than once in the data set. We add frequency weights to all patents so that each patent’s weight sums to one. Model 1 reports a significant positive relationship between the logarithm of the number of (R&D) establishments in the firm and the complexity value of a firm’s patents. The squared term on the establishment count variable is negative and significant. Thus, increases in the number of R&D units are associated with gains in the complexity value of a firm’s patents up to approximately five establishments, after which continued increases in the number of establishments cause the complexity value of patents to fall. The expected complexity value nevertheless remains higher than the estimated value at two establishments up to nine units. The relationship between establishment counts and patent complexity remain consistent as we add variables in successive models.

Table 3. The Number of R&D Establishments and the Average Technological Complexity of US Multilocational Firms

	Dependent Variable: Patent Complexity
	Model 1	Model 2	Model 3	Model 4
ln # establishments	0.0067***	0.0064***	0.0066***	0.0068***
	(0.0014)	(0.0015)	(0.0015)	(0.0015)
ln # establishments**2	-0.0022***	-0.0022***	-0.0020***	-0.0021***
	(0.0004)	(0.0004)	(0.0004)	(0.0004)
Integration
# co-invention patents				3.49E-06***
				(1.07E-06)
# citations between plants				-2.49E-07
				(4.94E-07)
# local citations external to firm				0.0006*** (0.0001)
Controls
# knowledge claims		0.0001***	0.0001***	0.0001***
		(0.0000)	(0.0000)	(0.0000)
# technology classes		-0.0032***	-0.0032***	-0.0033***
		(0.0002)	(0.0002)	(0.0002)
# inventors		0.0008***	0.0008***	0.0007***
		(0.0001)	(0.0001)	(0.0001)
HQ dummy			0.0003	0.0006
			(0.0004)	(0.0005)
Foreign firm dummy			-0.2810***	-0.2818***
			(0.0093)	(0.0093)
City size			0.0033***	0.0032***
			(0.0002)	(0.0002)
Firm size			-0.0008*	-0.0008*
			(0.0005)	(0.0005)
Constant	0.5751***	0.5803***	0.5469***	0.5476***
	(0.0075)	(0.0079)	(0.0048)	(0.0048)
Observations	446,164	446,164	446,164	446,164
R-squared	0.8015	0.8083	0.8084	0.8085
Year fixed effects	Yes	Yes	Yes	Yes
Primary class FE	Yes	Yes	Yes	Yes
Firm FE	Yes	Yes	Yes	Yes

Robust standard errors are in parentheses. Patent observations are frequency weighted so each patent has the same overall weight = 1 unit.

*** p < 0.01, ** p < 0.05, * p < 0.1

The patent level control variables in Model 2 all exhibit significant coefficients. More complex patents are associated with larger numbers of co-inventors and with the number of knowledge claims as we would anticipate. The negative relationship of the patent class count to complexity is surprising and likely explained by the fact that combining highly complex classes is extraordinarily difficult and not that common. Combining complex and less complex classes in similar fields is much more likely, but the lower complexity classes dampen the overall measure of patent complexity. In Model 3, the HQ dummy is not significant, while there is a relatively strong and negative coefficient on the foreign firm indicator, suggesting that these organizations tend to produce less complex technologies than domestic US firms. There is a significant, positive association between the volume of patenting activity in cities and more complex knowledge production. This confirms the increasingly well-known finding that larger cities produce more complex knowledge (Balland et al. 2020). Firm size, as measured by the overall number of patents produced within the firm, has a negative relationship with the average complexity of patents.

Testing Hypothesis 4: Higher Levels of R&D Integration within the Firm and between the Firm’s R&D Units and Other Economic Agents within Host Cities Are Associated with More Complex Knowledge

The local knowledge sourcing and firm integration variables are added in Model 4. Citations to knowledge produced by local economic actors external to the firm have a positive and significant impact on patent complexity, confirming the importance of external knowledge sourcing. In terms of firm-level integration, firms that have a history of producing patents with co-inventors that are drawn from different R&D establishments tend to produce knowledge that is significantly more complex than average. Firm integration as measured by citations between establishments had no significant influence on patent complexity. We assume that collaboration is a more direct and a more secure way of exchanging valuable knowledge within the firm as compared with citations.

Overall, Hypotheses 3 and 4 are generally confirmed. The number of R&D plants over which a firm distributes its knowledge production has a significant, positive relationship to the development of more complex technologies but only up to a certain point, after which the relationship turns negative. This finding is consistent with past work that highlights the difficulty of effectively integrating operations across establishments as multilocational firms expand the number of units they control. Explicit attempts at establishment integration, through collaboration for example, does appear to generate positive returns. And, finally, it is clear that more engagement with external knowledge producers in the different host communities across which the firm distributes its R&D establishments is positively related to the complexity of firm patents. Note that the results presented here are consistent with those presented by Frigon and Rigby (2023) for multiplant firms operating within Europe and for a sample of state-owned and private firms in China explored by Zhang (2024).

Discussion and Conclusion

This study bridges the literatures in economic geography and international business to examine geographic knowledge sourcing in multilocational firms. The literature on knowledge sourcing within economic geography largely portrays firms as fixed in space, benefiting from varied forms of local buzz and accessing nonlocal knowledge by partnering with economic agents elsewhere. Yet, many firms develop R&D facilities in different locations, at least in part, to access heterogeneous knowledge assets that are unevenly distributed across the economic landscape. Research in international business has long recognized how multinational firms engage in the development, dissemination, and appropriation of knowledge across borders. Our work shows that such activity is not confined to the international scale. Within the US, many firms, domestic and foreign, engage in knowledge production across numerous cities. Using detailed patent data, we report that for most of these firms, the nature of technologies produced in different locations varies significantly. That technological variability increases as the number of R&D units within the firm gets larger. Further analysis shows that the technological composition of patents produced in the R&D units of multilocational firms is related to the knowledge stocks of the cities where those units are located. This result supports claims of a systematic geography of knowledge sourcing by multilocational firms. We also add value to debates around the benefits of spatially distributed knowledge sourcing strategies by showing a positive relationship between the number of R&D facilities that multilocational firms operate and the complexity of the knowledge they produce, consistent with the claims of Kogut and Zander (1993), Foss and Pedersen (2002), and Cantwell and Mudambi (2011). However, our results also revealed that too many R&D facilities pose coordination problems for firms, potentially reducing gains from knowledge sourcing (see also Scalera, Perri, and Hannigan 2018; Berry 2023). Following Singh (2008), we show that these coordination costs might be lowered by greater knowledge integration across R&D units of the firm and with stronger local integration between the firm and external organizations.

While the discussion above focuses on the benefits of geographic knowledge sourcing for multilocational firms, related research in economic geography raises questions about the conditions under which host economies may benefit from the presence of such firms (Phelps 2008). How do multilocational firms impact regional economies, geographies of growth, and inequality (Iammarino and McCann 2013; Phelps, Atienza, and Arias 2018; Rodríguez-Pose 2018)? In terms of the spatial diffusion of skills and technologies, the agents of change literature suggests that multiplant firms play a lead role in distributing new capabilities across regions via their networks of establishments (Neffke et al. 2018; Elekes, Boschma, and Lengyel 2019; Lo Turco and Maggioni 2019; Crescenzi, Dyèvre, and Neffke 2022). However, to what degree these capabilities are broadly adopted by host-region firms, potentially leading to local forms of value capture, is largely unknown.

For economic geographers, much more work remains to explore the costs and opportunities of the integration or embedding of nonlocal economic actors; how these vary across firms, industries, technologies, and regions; and whether they differ between MNEs and domestic multiunit firms (see Bedreaga, Ortega Argilés, and McCann 2018). Of critical concern is whether the competitive advantages of regional economies, often tied to tacit forms of knowledge, can endure the absorption and diffusion of significant components of their knowledge assets by firms that are local and nonlocal at the same time. It is on these sorts of questions that insights from combining the international business and economic geography literatures promise the most leverage, drawing together scholarship on the organization of multilocational firms and on their connections to collaborators and competitors distributed within and between different countries.

Acknowledgments

The authors would like to thank the editor and the three anonymous referees whose comments significantly improved this manuscript.