Close together or far apart? The geography of host-country knowledge sourcing and MNCs’ innovation performance

ABSTRACT We investigate the influence of different host-country knowledge sources on the quality and generality of multinational corporation (MNC) innovation. We suggest that the quality of MNC innovation increases with the geographical distance from local industrial organizations. We also argue that the generality of MNC innovation increases with the distance from local research institutions, but decreases after a certain threshold. We explain these relationships based on the different cognitive and geographical distances separating MNC innovation activities from industrial organizations and research institutions. We test our arguments on a sample of US Patent and Trademark Office (USPTO) patents developed in the United States by foreign semiconductor MNCs.


INTRODUCTION
Innovation within multinational corporations (MNCs) increasingly relies upon knowledge sourced outside the firm's home country (Cantwell & Iammarino, 2003). To secure inflows of relevant knowledge, MNCs strive to create global linkages in the form of pipelines (Bathelt et al., 2004). These linkages are especially valuable when they target large countries, such as the United States, where different high-technology agglomerations exist within the same industry. In these contexts, the set-up of local innovative activities enables the establishment of systematic communication channels with knowledge sources in more or less proximate locations within the host country's national innovation system (NIS) (Lundvall, 1992).
Even within the same host country, external knowledge sources are not all alike (Lundvall, 1988). Rather, they differ along various dimensions of proximity/distance to the MNC's local innovation activities (Boschma, 2005). Within a host country, different local actors develop different types of knowledge that MNCs can source (Alcácer & Chung, 2007). Like the MNC, host-country industrial organizations (IOs), such as private firms, tend to develop applied knowledge. Instead, local research institutions (RIs), such as universities and government laboratories, typically engage in the generation of basic knowledge with which the MNC is less familiar (for a review, see, e.g., Alcácer & Chung, 2007). While the cognitive distance between MNCs and IOs is limited, as both develop applied knowledge, MNCs sourcing basic knowledge from RIs have to bridge higher cognitive barriers. However, they can also gain from the synergies that arise when combining bodies of knowledge that differ in nature. Thus, the absorption and use of these two types of knowledge provide different opportunities and challenges to the MNC due to the different cognitive distances between source and recipient (Nooteboom et al., 2007). 1 Knowledge sources in a host country can be located at different geographical distances from the MNC's local innovation activities. Geographical distance from a hostcountry knowledge source generates assimilation costs because knowledge spillovers decay rapidly with distance (Bode, 2004;Paci & Usai, 2009;Rodríguez-Pose & Crescenzi, 2008), making effective absorption channels more difficult to establish and maintain. At the same time, geographical distance from host-country knowledge sources may have positive effects, by offering opportunities for novel knowledge combination, which arise from searching for outside-of-the-box knowledge and ideas available in close proximity (Cantwell & Santangelo, 1999;Crespo & Vicente, 2016;McCann, 1995;Singh, 2008). Thus, there are countervailing forces associated with the type (Alcácer & Chung, 2007) and geography (Iammarino & McCann, 2013) of knowledge sources in a host country, with different implications for MNCs' innovation performance.
The issue is especially salient for foreign MNCs. They leverage their innovative activities to embed themselves in the host country's NIS and overcome their liability of outsidership by establishing formal and informal local linkages (Johanson & Vahlne, 2009), which would otherwise have been difficult to establish and maintain. The role of the geography of different host-country knowledge sources in MNC innovation performance has been studied in relation to their collocation/non-collocation with the MNC's local innovative activities (Phene & Almeida, 2008). We lack an understanding of how the type and geography (not collocation) of knowledge sources in a host country influence MNC innovation performance (Iammarino & McCann, 2013).
To address this issue, we investigate the effect of the geographical distance separating the MNC's local innovative activities from the host-country IOs and RIs, from which the MNC sources knowledge, on the MNC's innovation performance. To account for the heterogeneous and multifaceted nature of innovation performance, we unpack MNCs' innovation performance into the quality and generality of innovation. These are distinct dimensions that contribute to the MNC's development. The quality dimension is associated with the commercial success of a firm's innovation, and the generality captures its applicability across diverse technological fields (Trajtenberg et al., 1997).
We consider MNC-specific benefits and costs that overlay the spillover-related centripetal forces and congestion-based centrifugal forces that have traditionally been studied. Specifically, drawing on research on economic geography (Iammarino & McCann, 2013;Simmie, 1998) and external knowledge absorption (Cohen & Levinthal, 1990;Nooteboom et al., 2007;, we argue that the geographical distance of the MNC's local innovation activities from hostcountry industrial knowledge sources has a positive effect on the quality of MNC innovation. This is because the costs associated with sourcing this type of knowledge increase at a slower pace, as the distance increases, than do the benefits. Although knowledge spillovers decay with distance, sourcing applied knowledge at a close proximity to the MNC raises the risks of unintended knowledge outflows to nearby IOs (Alcácer & Chung, 2007;Cantwell & Santangelo, 2002;Iammarino & McCann, 2013;Simmie, 1998). To limit these risks, the MNC can leverage its local innovation activities to assimilate applied knowledge across space, and access ideas that are progressively more diverse than those available at a close proximity (Cantwell & Santangelo, 1999;Crespo & Vicente, 2016;Singh, 2008). An MNC's familiarity with applied knowledge reduces the cognitive gap separating source and recipient, and the costs of absorption , thus fostering the quality of MNC innovation.
Instead, the relationship between the geographical distance of the MNC's local innovation activities from RIs, and the generality of MNC innovation, is an inverted 'U' shape. This is because the costs associated with sourcing this type of knowledge increase, as the distance increases, at a slower pace than do the benefits, but only up to a point. As the distance from RIs increases, the costs of establishing and maintaining absorption channels (such as formal and informal linkages) increase, due to the spatial decay of spillovers. However, the benefits rise faster due to the synergies that the MNC can exploit by combining its own knowledge with the basic knowledge (Nooteboom et al., 2007) of RIs located outside of the box available to its nearby peers (Cantwell & Santangelo, 1999;Crespo & Vicente, 2016;Singh, 2008). As the distance between these sources and the MNC's local innovation activities increases further, the costs of sourcing basic knowledge increase faster than the benefits. The investment required to bridge the cognitive gap across larger geographical distances becomes prohibitive, yielding a poor understanding and ultimately detrimental effects on the generality of MNC innovation (Cohen & Levinthal, 1990;. We test and find support for our arguments using patents granted by the United States Patent and Trademark Office (USPTO) to a sample of foreign semiconductor MNCs for research developed in the United States between 1985 and 2000. This offers an appropriate setting not only because the semiconductor industry originated in this country in the late 1940s, but also because the United States is home to several semiconductor clusters (e.g., Silicon Valley, Dallas, New York) scattered across different states and characterized by location-specific agglomeration dynamics, labour markets and knowledge infrastructures (Almeida, 1996). Thus, MNCs' local innovative activities might be expected to source knowledge from different host-country regions located at different geographical distances within the NIS.
Our study contributes to the relatively recent stream of research that seeks to integrate economic geography, international business and innovation (Beugelsdijk et al., 2010;Mudambi et al., 2018). In relation to the phenomenon of international knowledge sourcing, we shed light on the joint consideration of geography, which has so far been investigated in terms of collocation, and cognitive distance. We also add to the international knowledge-sourcing literature by exploring the heterogeneous relationships between different types of local knowledge sources and distinct dimensions of innovation performance. phenomenon by combining the resource-based view with transaction-costs perspectives (e.g., Gooris & Peeters, 2016;Iammarino & McCann, 2013;Martinez-Noya et al., 2012). In their quest for resources and capabilities that are not available internally, firms often look beyond their national borders, trying to limit the transaction costs that could arise from operating in foreign countries (Martinez-Noya et al., 2012;Martínez-Noya & García-Canal, 2018). The MNC's innovative activities in a host country are, for it, a 'gateway' to understanding the knowledge-sourcing opportunities available in the host NIS . They help them learn how to effectively recombine local knowledge with relevant internal knowledge to improve the MNC's innovation performance (Almeida & Phene, 2004;. This is especially the case in large host countries such as the United States, where the establishment of local innovative activities provides access to a variety of knowledge sources, more or less geographically proximate, within the host NIS (Carlsson, 2006).
Thus, the ability of the MNC to source local knowledge and translate it into an innovation outcome depends, among other things, on the cognitive and geographical distance between the host-country source and the MNC recipient (Delgado-Márquez et al., 2018;Fitjar et al., 2016). In what follows, we discuss the structure of MNC-specific costs and benefits associated with cognitive and geographical distance. These benefits and costs overlay the spillover-related centripetal forces and congestionbased centrifugal forces that have traditionally explained agglomerations (Fujita & Thisse, 2002;Marshall, 1890). The explanation is twofold. First, MNCs typically have more to lose from outgoing knowledge leakages, and less to gain from incoming knowledge leakages, than domestic companies because of their technology-related ownership advantage (Iammarino & McCann, 2013). Second, they have superior recombination capabilities across space (Kogut & Zander, 1993).

The role of cognitive distance
In the context of host-country knowledge sourcing, cognitive distance is strongly influenced by the type of knowledge that host-country sources develop. RIs produce mainly basic knowledge that focuses on scientific questions and has an impact across a broad range of fields, but rarely generates immediate commercial applications (e.g., Kuhn, 1962). The research outcome of IOs is typically applied knowledge that aims to solve technical problems and is closer to the market because it yields immediate and likely returns (Kuhn, 1962). Like other IOs, MNCs are more familiar with applied (versus basic) knowledge (Delgado-Márquez et al., 2018). Familiarity with applied knowledge creates a common cognitive background between the MNC and host-country IOs. It enhances the absorptive capacity of both parties (Cohen & Levinthal, 1990) and, thus, raises the risk of unintended knowledge outflows between them (Iammarino & McCann, 2013). On the other hand, MNCs are less familiar with the basic knowledge developed by RIs, knowledge which can only be converted into useful technology through major efforts (Pavitt, 2001), especially when sourced across distance. To effectively use the outcomes of the RIs' innovation processes, MNCs need to bridge the cognitive gap between basic and applied knowledge (Delgado-Márquez et al., 2018). At the same time, by sourcing knowledge from RIs, the MNC benefits from the synergies that derive from the combination of its own body of applied knowledge with the basic knowledge developed by RIs. These benefits are greater the more the MNC sources knowledge and ideas different from those available to other companies located in close proximity to it. By doing so, the MNC increases its ability to pursue previously unexplored knowledge recombination opportunities.

The role of geographical distance
Within a host country, knowledge sources also differ in relation to their geographical distance from the MNC's local innovation activities. Since knowledge spillovers decay with distance (Storper & Venables, 2004), assimilating knowledge located at close quarters and in distant locations respectively encompass different costs and benefits, with countervailing forces at play.
Geographical distance from host-country sources raises knowledge assimilation costs. However, it offers opportunities for novelty and recombination (Cantwell & Santangelo, 1999;Singh, 2008). Due to common local factor conditions (McCann, 1995), proximate firms tend to access the same local resources and face the potential exhaustion of their efficient use (Crespo & Vicente, 2016). Instead, sourcing distant knowledge within the host country allows the MNC to recombine its own technology with knowledge and ideas different from those available to its nearby peers. The greater the geographical separation between contexts, the more diverse the evolution of contexts' technological trajectories (Anselin et al., 1997). Thus, as the geographical distance between the MNC's local innovative activities and the host-country knowledge source increases, the knowledge and ideas that the MNC can access gradually become more diverse from those developed in the neighbouring contexts (Scalera et al., 2018).

THE EFFECT OF DIVERSE HOST-COUNTRY KNOWLEDGE SOURCES ON MNC INNOVATION PERFORMANCE
The MNC's effort to effectively source knowledge from different host-country actors is aimed at improving its innovation performance, which is a complex and multidimensional concept as reflected in the outcomes of the research and development (R&D) process (e.g., Trajtenberg et al., 1997). Innovation performance can be unpacked into quality and generality (Trajtenberg et al., 1997). Quality of innovation denotes the extent to which an invention has spurred future technical advances. It is associated with the monetary value of the innovation and contributes to the firms' market value (Hall et al., 2005;Trajtenberg, 1990). Generality of innovation denotes the degree to which follow-up inventions that build upon the focal innovation are spread across diverse technologies (Trajtenberg et al., 1997). This aspect of innovation performance captures the applicability of the innovation across scientific and technological fields.
Variation in MNCs' innovation performance has been documented mainly in terms of size. More qualitative measures of innovation performance have also been adopted, and related to the firm's ability to source hostcountry external knowledge (e.g., Phene & Almeida, 2008;Scalera et al., 2018). The effect of geographical distance between source and recipient on different dimensions of a firm's innovation can be both linear and nonlinear (Boschma & Martin, 2010;Fitjar et al., 2016;Fitjar & Gjelsvik, 2018). Arguably, it depends on the type of the source within the host country .
We relate knowledge sourcing from IOs to the quality of MNC innovation because IOs traditionally focus on the refinement and specialization rather than the broad applicability of the knowledge developed. This knowledge is less likely to influence the generality of innovation. Instead, we relate knowledge sourcing from RIs to the generality of MNC innovation. The basic knowledge these local actors develop is likely to influence a variety of fields, but is seldom directly useful for innovation activities aimed at generating immediate returns and higher market value.

Knowledge sourcing from host-country IOs
The knowledge of host-country IOs has an applied nature and is a critical input to the innovation processes of MNCs (Phene & Almeida, 2008). It can offer new ideas and expertise that contribute to the solving of particular technical problems and facilitate commercial success, improving the quality of firm innovation.
MNCs share a common cognitive background with the host-country IOs because of their familiarity with the type of knowledge these organizations produce. The common cognitive background enhances their reciprocal absorptive capacity (Cohen & Levinthal, 1990). This facilitates knowledge sourcing, especially when the MNC's local innovative activities and host-country IOs are located in close proximity to one another. It also raises the risk of unintended knowledge outflows that would be more dangerous for the MNC than for the host-country IOs because the MNC will typically have more to lose and less to gain (Iammarino & McCann, 2013). As the geographical distance between the MNC's local innovative activities and host-country IOs increases, the knowledge spillovers decay. Thus, formal and informal linkages need to be established for the applied knowledge of IOs across space within the NIS to be absorbed and understood effectively. Due to the familiarity with the type of knowledge that host-country IOs develop, the costs of establishing and maintaining these linkages will arguably be limited. For instance, leveraging their common cognitive background, the MNC's local knowledge workers will likely need only a few trips to meet their peers in faraway IOs in order to form and implement a knowledgebased collaboration.
Reaching out to grasp diverse ideas and expertise developed in distant host-country locations opens up opportunities for novelty (Singh, 2008). These opportunities increase with the geographical distance between source and recipient. The MNC can progressively search further outside of the box of the applied knowledge available to its nearby peers, for increasingly more diverse ideas. Opportunities for novelty and idea recombination will on the other hand be more limited when the MNC sources knowledge from geographically proximate IOs (McCann, 1995). Because geographical proximity eases localized spillovers (Jaffe et al., 1993;Paci & Usai, 2009) and strategic technological partnerships (Narula & Santangelo, 2009), geographically close organizations are likely to develop similar knowledge outputs (Cantwell & Santangelo, 2002;McFadyen & Cannella, 2005) and have to face the exhaustion of the efficient use of local resources and knowledge recombination opportunities, along with the threat of imitation by proximate firms.
In sum, we expect that the costs of sourcing knowledge from host-country IOs increase at a slower pace as distance increases than do the benefits. As a result, the quality of MNC innovation will increase with the geographical distance from host-country IOs' knowledge sources.

Knowledge sourcing from host-country RIs
MNCs are eager to rely on host-country RIs to access basic knowledge (Anselin et al., 1997), which is critical to increasing the generality of their innovation. The knowledge RIs develop focuses on scientific questions, elucidates general laws and has wide applicability across a broad range of fields. It is also substantially different from the applied knowledge that the MNC develops. This dissimilarity creates opportunities for synergies between the two types of knowledge (Agarwal & Ohyama, 2013) because it facilitates novel knowledge combinations (Kogut & Zander, 1992). At the same time, the dissimilarity between the knowledge developed by the MNC and that developed Close together or far apart? The geography of host-country knowledge sourcing and MNCs' innovation performance by host-country RIs yields a cognitive gap between them (Nooteboom et al., 2007), such that rich and extensive formal and informal relational linkages are needed to transfer and assimilate knowledge.
As the geographical distance from host-country RIs increases, so does the scope for synergies between the applied knowledge developed by the MNC and the basic knowledge developed by the host-country RIs. This knowledge is dissimilar from the type of knowledge the MNC traditionally develops, and, as the distance increases, it becomes increasingly different from the basic knowledge available in close proximity to the MNC's local innovation activities, which can also be accessed by other nearby firms (McCann, 1995). Reaching out to increasingly distant knowledge allows the MNC to progressively search outside of the box of the basic knowledge available to its nearby peers (Audretsch & Feldman, 1996;Crespo & Vicente, 2016) and amplifies the potential for developing very general inventions.
However, geographical distance also exacerbates the costs of bridging the cognitive gap in order to access basic knowledge. While knowledge sourcing from proximate host-country RIs may rely on knowledge spillovers, sourcing basic knowledge across greater and greater distances becomes increasingly costly. When host-country RIs are not too distant from the MNC's local innovation activities, the MNC can afford to create and maintain the formal and informal relational linkages required to figure out the usefulness of the knowledge these RIs develop, for its own innovation activities . For instance, the MNC can bear the costs of having its local engineers travel on a regular basis to meet colleagues working for these RIs.
When the geographical distance between the MNC's local innovation activities and the host-country RIs increases further, establishing and sustaining formal and informal linkages becomes too costly (Fitjar & Gjelsvik, 2018). This is because relational proximity is facilitated by close geographical proximity (Morgan, 2004), which nurtures trust and eases interaction (Bathelt & Li, 2020), unless inappropriate institutions for intellectual property rights' protection hinder the development of relational assets Wang & Lin, 2013). 'Relational knowledge is contained "in the air", as part of the "local buzz", and can be utilized by "being there"' (Bathelt & Li, 2020, p. 5). This applies to formal and informal linkages. Formal linkages such as contract research, consulting or partnership with faraway hostcountry RIs require the MNC's local inventors to travel long distances to effectively evaluate and implement potential research collaborations, and to translate basic knowledge into useful inputs for the MNC's innovation activities. The considerable travel and opportunity costs associated with these transfers are then likely to discourage formal linkages. Establishing and maintaining informal linkages would be even more prohibitive if large distances were involved. Social networks, for instance, typically require frequent and regular contact, which is costly to establish and maintain across large distances (Saxenian, 1990). Moreover, social networks tend not to cross community boundaries (Sytch & Tatarynowicz, 2014). Thus, networks between corporate inventors and academic scientists are rare and less likely to occur. As a last option, MNCs might try to benefit from the mobility of academic inventors, who could be hired with the aim of gaining access to their tacit understanding of the basic knowledge that MNCs are willing to source. Yet, academic scientists tend to be reluctant to relocate in space, having a preference 'to enter into contractual agreements with firms located within commuting distance of their university, where they tend to retain affiliation for reasons of reputation' (Breschi & Lissoni, 2009, p. 5). In the absence of effective formal and/or informal channels for knowledge absorption, the attempt to source knowledge from hostcountry RIs that are located far away from the MNC's local innovation activities may generate a poor understanding of the knowledge developed by geographically distant RIs and, thus, be detrimental for the generality of the MNC innovation.
In sum, as the geographical distance between the MNC's local innovation activities and host-country RIs increases, initially, we expect that the benefits in terms of synergies and potential knowledge combinations will offset the costs of bridging the cognitive gap between source and recipient, thus increasing the generality of MNC innovation, but only up to a point. After that point, as the geographical distance between source and recipient further exacerbates the costs of linking to the knowledge source, the MNC will likely fall short of creating effective knowledge absorption channels, and the knowledge sourced from faraway RIs will ultimately damage the generality of its innovation, resulting in a non-linear effect of geographical distance on the generality of the MNC's innovation.

Data
We analyse a sample of USPTO patents granted to foreign semiconductor MNCs for research developed within the United States over the years 1985-2000, and link them to their backward citations to trace knowledge sourcing.
The semiconductor industry is an appropriate empirical setting since it is one of the most technology-intensive industries. In the period of analysis, non-US firms started operating in the United States (Phene & Almeida, 2008) with the strategic intent to assimilate local knowledge from a variety of actors and thereby reap the benefits of the substantial technological development of the industry in the country (Almeida, 1996). The focus on a single foreign country enables us to hold constant cultural and language differences (Phene et al., 2006) that might influence knowledge sourcing among distant actors. Despite advances in communications and computing in the period of analysis, interaction across space remains critical (Leamer & Storper, 2001), especially in high-tech industries where 'the very tacitness of the knowledge creates conditions for market failure and increases the difficulties of arm's-length interactions' (Nachum & Zaheer, 2005, p. 751). In such contexts, interactions are also facilitated by an appropriate institutional environment for intellectual property rights' protection, which enables the development of linkages and knowledge circulation (e.g., Wang & Lin, 2013).
The use of patents to study knowledge sourcing has some limitations, but also many advantages. In the US patent system (Almeida, 1996), it is mandatory to include a list of citations of other patents in the patent document. Such lists identify the technological antecedents to the specific innovation in question. These antecedents represent the knowledge inputs that innovators combine to generate new technology, and allow the tracing of the geographic origin of the knowledge sources utilized for innovation. While patents represent the codified component of knowledge, the codified and tacit components of knowledge tend to be correlated and complementary (Mowery et al., 1996). To ensure that technology can be assimilated effectively, recipients need access to both the codified part of knowledge and the related tacit know-how (Kogut & Zander, 1993). Thus, tracing the flow of codified technology through patent citations helps map the movement of the underlying tacit knowledge, as recognized by the established empirical literature's use of patent citations to infer knowledge flows (Jaffe et al., 1993).
The use of US patents is especially appropriate in the context of semiconductor companies, which rely extensively on patents for strategic reasons (Hall & Ziedonis, 2001). Also, any major semiconductor MNC, irrespective of its country of origin, is likely to hold a comprehensive portfolio of US patents that reflects its innovation activities (Almeida & Kogut, 1999). Finally, the use of patents from a common foreign country (i.e., the United States) limits potential home-country biases (Soete & Wyatt, 1983).
To identify our sample, we compiled a list of non-US MNCs carrying out innovation activities in the United States, drawing on Gartner Dataquest. Combining information from the USPTO, the National Bureau of Economic Research (NBER) US Patent Citation Data File  and the Harvard Patent Dataverse (Li et al., 2014), we identified all utility patents that (1) had been assigned to our sample MNCs, (2) whose first inventor had a US-based address between 1985 and 2000, and (3) whose technological class belonged to the set of Derwent patent classes included in the section 'Semiconductors and Electronic Circuitry' (Alcácer & Zhao, 2012). This process led us to focus on 11 MNCs 2 and a total of 1788 patents that represent our level of analysis.

Measures 4.2.1. Dependent variables
The quality of innovation is measured as the number of forward citations received by a focal patent (Hall & Ziedonis, 2001). The more a patent is used in subsequent innovations, the higher its importance and economic value (Hall et al., 2005;Lanjouw & Schankerman, 1999). This measure is correlated with the 'private value of the patentable invention' (Hall & Ziedonis, 2001, p. 123). It indicates economic success since the subsequent patents citing a focal innovation 'are the result of costly innovation efforts undertaken mostly by profit-seeking agents' (Trajtenberg, 1990, p. 174). Firms whose patents are more cited enjoy, on average, higher stock market values  and experience economic benefits (Trajtenberg, 1990), which can be used to sustain future innovative endeavours.
To address potential 'right-truncation' problems (i.e., older patents are likely to be more cited), we restrict forward citations to those occurring within three years of the focal patent's application date (Sterzi, 2013). We chose this timeframe because the citations a patent receives in the very early stage of its life are more strongly correlated with its economic value, as they signal 'the presence of others working in a similar area, and thus that the area is expected by others to generate economic value' (Lanjouw & Schankerman, 1999, p. 7). This citation premium shrinks, and eventually disappears, when considering longer time intervals (Sterzi, 2013).
Following previous studies (Trajtenberg et al., 1997), we measure the generality 3 of MNC innovation as follows: where s ij indicates the percentage of citations received by patent i that belong to the technological class j, out of n i technological classes. Patents that are cited by subsequent patents that belong to a wide range of technologies have an influence on a more diverse set of technological fields. Following previous studies (Singh, 2008), we consider 36 two-digit technological classes proposed by . By definition, this measure can be calculated only for those patents with at least one forward citation (Mowery & Ziedonis, 2002), thus reducing the sample.
As the generality of MNC innovation is meant to capture the extent to which an innovation is able to instigate successive inventions across a wide range of technological fields, a sufficiently long citation lag needs to be allowed in order for us to observe a more comprehensive set of followup inventions potentially building on our sample patents (Czarnitzki et al., 2012). Thus, to compute this variable, we consider forward citations occurring within five years of the focal patent application date. Also, the Generality measure may be biased because the more citations a patent receives, the greater the likelihood that these citations will be spread across a larger number of technological classes . To correct for this bias, we follow . We multiply the Generality index by n/(n − 1), where n is the number of forward citations that a focal patent has received during the first five years following its filing date.

Independent variables
As a proxy for the extent to which the MNC's local innovation activities draw on knowledge from geographically distant actors within the host country, we follow extant Close together or far apart? The geography of host-country knowledge sourcing and MNCs' innovation performance research and look at our focal patents' backward citations (Almeida & Phene, 2004;Singh, 2008). We proceed in four steps.
First, we identify all the citations of patents that are not assigned to the MNC, so as to focus on 'external' knowledge sources, and retain only those patents whose first inventor was located in the United States, as we are interested in the MNC's knowledge sourcing within the host country.
Second, we distinguish between host-country knowledge sources that are industrial organizations (IOs-KS) and those that are research institutions (RIs-KS) using the NBER assignee classification . This allows us to differentiate between knowledge sources that expose MNCs to different degrees of cognitive distance. To confirm that this distinction reflects the different natures of the knowledge developed by these two types of sources, we test for statistical differences in the numbers of citations of scientific papers by each of these sources, based on the argument that scientific citations are typically the output of basic research, thus an indicator of the basicness of the resulting invention (Trajtenberg et al., 1997). The result of the t-test suggests that patents assigned to RIs cite, on average, a higher number of scientific publications than patents assigned to IOs (t ¼ 11.226, p > 0.01), 4 thus confirming the greater 'basicness' of the knowledge developed by RIs compared with that developed by IOs.
Third, for each of the two sets of backward citations, we develop measures of the extent to which they are geographically distant from the MNC's innovation activities within the United States, based on the address of the focal patent's first inventor. We consider geographical distance across three geographical units (i.e., US metropolitan statistical area/consolidated metropolitan statistical area (MSA/CMSA), 5 state and the entire United States) based on the evidence of localized spillovers at the level of each unit, as well as the existence of border effects across these units (Singh & Marx, 2013). We assign each focal patent to an MSA/CMSA using concordance data made available by Thompson (2006) 6 to map the first inventors' US addresses to metropolitan areas. For US addresses that cannot be matched to metropolitan areas, we follow the standard procedure of defining a 'phantom area' per each state (Thompson, 2006). We divide the backward citations into three groups, depending on whether the cited patent's first inventor was located in the same MSA/ CMSA as the focal patent's first inventor, in the same US state as the focal patent's first inventor (excluding inventors located in the same MSA/CMSA) or elsewhere in the United States. For each of the three geographical groups of backward citations, we calculate the average distance between the focal patent's first inventor and the knowledge sources in that group. In particular, we geocode the locations of both focal and cited patents and calculate the travel time between them when driving (Lublinski, 2003). We use travel time rather than standard geographical distance in miles/kilometres to account for the frictions associated with geographical distance (e.g., preparation time and inefficiencies of transport infrastructure), which generate additional costs that make contacts (or the perception of them) between locations less convenient (Boeh & Beamish, 2012).
In the last step, we build our measure of geographical distance from the host-country knowledge sources (Geodistance) as follows: where x denotes the focal MNC's innovation activities in the United States; and y either the IO or RI knowledge sources. If y represents the IOs, then the variable denotes the variable Geographical distance from IOs-KS. If it represents the RIs, then it denotes the variable Geographical distance from RIs-KS. AD M , AD S and AD C are the average distances between the focal MNC's US innovation activities and the local sources located in the same MSA as the focal MNC's innovation activities (M), in the same state (but not in the same MSA) as the focal MNC's innovation activities (S) and elsewhere in the United States (C), respectively. To disentangle situations where AD M is 0 because there are no MSA-based knowledge sources from situations where all MSA-based knowledge sources share the same location as the focal patents, leading to a travel distance equal to 0, we rescale each average distance to 1. In addition, each of these distances is weighted by the intensity of the knowledge sourcing from each of the three geographical units (p iM , p iS and p iC ). For those observations corresponding to focal patents that report no US-based backward citations, we set both independent variables to 0 and define a dummy variable 'No US backward' that equals 1, to account for these special cases (Singh, 2008).

Controls
To rule out the effect of non-host-country-specific knowledge sources on the quality and generality of MNC innovation, we include a number of controls in our models (Phene & Almeida, 2008). Previous research has suggested that the knowledge sourced from the headquarters and other subsidiaries can influence the performance of MNCs' local innovative activities (Almeida & Phene, 2004). Thus, we account for the importance of relying on within-MNC Internal technology knowledge sourcing using the total number of citations the focal patent had, which referenced patents assigned to the MNC. Beyond within-MNC internal knowledge sourcing, subsidiaries may also assimilate knowledge from other external sources outside the host country (Phene & Almeida, 2008). To account for the potential role of these sources, we include the total number of citations of the focal patent which referenced patents assigned to organizations external to the MNC and whose first inventor was located outside the United States (Non-US knowledge sourcing). Previous research has emphasized the role of social and professional linkages as conduits for knowledge 376 Alessandra Perri and Grazia D. Santangelo transmission (Agrawal et al., 2006). To exclude the possibility that our results are driven by such relational proximity (Breschi & Lissoni, 2009), we control for the coinventor relationships between our focal patents' inventors and the inventors of the cited patents (Rosenkopf & Nerkar, 2001). We first count all those focal-to-backwardpatent pairs featuring at least one common inventor, using disambiguated inventor information from the Harvard Patent Dataverse (Li et al., 2014). Then, we divide this number by the total number of relevant cited patents, to obtain an average value of the co-inventor relationships between our focal patents and the respective cited patents (Co-inventor). Finally, because collaborative R&D efforts may leverage synergic knowledge bases, thus generating better innovative outcomes (Belderbos et al., 2014), we include a dummy that takes the value 1 if the focal patent has more than one assignee (Co-assigned), to control for innovations that result from R&D collaborations between MNCs and external partners. Table 1 presents the descriptive statistics and bivariate correlations for all variables included in our models. The correlation coefficients do not raise collinearity concerns.

Model
To account for the distribution of our dependent variables, we estimate negative binomial regression models to explore the role host-country knowledge sourcing from IOs plays in the MNCs' innovation quality (3), and twosided Tobit regression models to investigate the role host-country knowledge sourcing from RIs plays in the generality of MNCs' innovation (4): Generality i = a 0 + a 1 Geodistance IOi + a 2 Geodistance 2 IOi + a 3 Geodistance RIi + a 4 Geodistance 2 RIi + bControls + 1 i (4) where the dependent variables are respectively the Quality and Generality of patent i, and Geodistance IOi and Geodistance RIi are the two independent variables, for which we also include the quadratic term. 7 Both equations include the controls described above. Finally, we cluster standard errors by MNC to control for possible non-independence among patents developed within the same MNC. Table 2 reports the findings for the quality of MNC innovation.

RESULTS
In model 1, we test our baseline in which all controls are included. In model 2, we introduce Geographical distance from IOs-KS, which, as expected, has a positive (0.049) and statistically significant (p ¼ 0.031) coefficient. In model 3, we include the squared term of Geographical distance from IOs-KS to empirically rule out a curvilinear effect on the Quality of MNC innovation. In this model, the coefficient of Geographical distance from IOs-KS remains positive (0.051) and statistically significant (p ¼ 0.022), while the coefficient of the squared term is not significant. This supports our expectation that the quality of MNC innovation will increase with the geographical distance of the host-country IOs from which the MNC's local innovation activities source knowledge. In all models, Geographical distance from RIs-KS is also non-significant. Among the controls, Co-inventor and Co-assigned have positive and significant effects (p ¼ 0.026 and 0.000, respectively) on the Quality of MNC innovation in all models. In contrast, Internal knowledge sourcing and Non-US knowledge sourcing do not seem to have a significant effect on the Quality of MNC innovation. However, their signs are consistent with extant research. Table 3 presents the findings for the Generality of MNC innovation.
In model 1, we test our baseline in which all controls are included. In model 2, we introduce Geographical distance from RIs-KS, which has a positive (0.047) and statistically significant (p ¼ 0.000) coefficient. In model 3, we include the squared term of Geographical distance from RIs-KS. The coefficient of the linear term remains positive (0.154) and statistically significant (p ¼ 0.001) and the coefficient of the squared term is negative (−0.040) and statistically significant (p ¼ 0.014). This supports our expectation that the generality of MNC innovation increases initially with the distance from local RIs but then decreases after a threshold. In particular, travelling further, up to 9.14 hours, to RIs increasingly promotes innovation generality, which then reduces progressively under additional increases in travelling times. 8 Among the controls, Co-assigned has a positive and statistically significant (p ¼ 0.000) coefficient in all models. None of the other controls is significant, but their signs are overall consistent with the existing literature's predictions. In the appendix, we discuss a number of robustness checks that have been performed to account for, among other things, additional factors that may potentially influence MNC innovation performance but could not be added to our main estimations due to constraints in model convergence.

DISCUSSION AND CONCLUSIONS
Host-country knowledge sourcing is critical for MNC innovation performance. Yet, different knowledge sources in a host country can be beneficial to specific dimensions of MNC innovation performance, depending on the cognitive and geographical distance between the local knowledge sources and MNC innovation activities. Our study finds that the quality of MNC innovation benefits from knowledge sourced from faraway IOs. This is consistent with previous research suggesting that technologically advanced MNCs avoid host regions with a high concentration of industrial activity (Alcácer & Chung, 2007). Moreover, we show that the generality of MNC innovation benefits from not-too-distant RIs. While this idea of an 'optimal' distance has been explored in the context of innovation networks (Boschma & Martin, 2010;Fitjar et al., 2016), we add to this literature by illustrating that sourcing knowledge from RIs that are not too close, but neither too distant, is conducive to greater generality of innovation. Our theorizing implies that, if host-country RIs and IOs are collocated, MNCs will locate themselves as far away as possible from both to maximize the net benefits. In addition, if the costs of local agglomeration factors (e.g., wages) fall as MNCs move away from the joint location of RIs and IOs, then this will even further enhance the push of MNC activities away from both the RIs and IOs.
Our study contributes to the recent stream of research that seeks to integrate economic geography, international business and innovation (Beugelsdijk et al., 2010;Mudambi et al., 2018). In relation to international knowledge sourcing, we suggest that geographical distance is a continuous concept within the boundaries of specific host countries. In the research investigating external knowledge sourcing by MNCs, the geographical distance between the MNC and the external sources has been accounted for in terms of collocation of source and recipient Perri & Santangelo, 2017). Yet, looking at the geography of host-country sources mainly in terms of binary discontinuity fails to account for a number of situations falling in the range between 0 and 1 (Beugelsdijk & Mudambi, 2013). Knowledge spillovers progressively decay with distance and, thus, the novelty of ideas that can be sourced increases with the geographical distance between the source and recipient (Anselin et al., 1997). As a result, reaching out to progressively distant knowledge sources enables firms to access increasingly novel ideas, expanding the opportunities for recombination. Yet, the geographical distance from the source has a differentiated impact on different dimensions of innovation performance. We suggest that this impact may be linear or curvilinear, depending on the cognitive distance between source and recipient, and the dimension of innovation performance.
The study also provides a more nuanced understanding of MNCs' knowledge-sourcing strategies, by relating the multidimensional nature of innovation performance to the diversity of host-country knowledge sources (Scalera et al., 2018). Exploring the contribution of diverse hostcountry knowledge sources to MNC innovation performance, we show that the heterogeneity of host-country knowledge sources enables diverse sourcing strategies, which show varying sensitivity to the type and geography of the source. Also, our study points to the fact that different host-country knowledge-sourcing strategies result in different innovation outcomes in terms of greater value and greater applicability. By taking into account the multidimensional and complex nature of innovation, we are able to unveil the diverse benefits MNCs can reap from specific knowledge-sourcing strategies.
Our findings provide interesting insights for national and local policies aimed at attracting and maintaining the innovation activities of advanced foreign MNCs. First, policymakers should recognize the different advantages MNCs can obtain when sourcing from IOs and RIs, and tailor attraction policies across subnational locations accordingly. Second, national policymakers should be aware of the potentially harmful effect that a very high geographical distance from RIs could have on the generality of MNC innovation, and take actions to boost RIs in regions where the attraction of foreign MNCs may benefit the initiation of virtuous cycles (Mudambi & Santangelo, 2016). They may, for instance, promote and also fund initiatives of virtual or temporary collocation, such as platforms, conferences and workshops. At the same time, these initiatives should be tailored to the local sectoral specialization. In industries whose knowledge base is highly tacit, the proposed solutions might be insufficient to remedy geographical distance, and policies aiming at a more permanent attraction of foreign MNCs would be required. Data limitations in our analysis offer opportunities for future research. First, we distinguish different types of host-country knowledge sources, but we are not able to measure their knowledge content. Second, we are not able to observe the formal and informal absorption channels of knowledge sourcing nor to account for changes in MNCs' formal R&D organization. More informative datasets supplementing patents with survey data may

DISCLOSURE STATEMENT
No potential conflict of interest was reported by the authors.

NOTES
1. We recognize that IOs do perform basic research, but this activity remains residual and unintentional (Rosenberg, 1990). 2. While this is a small number of MNCs, it is in line with the market structure of the semiconductor industry, which is concentrated in all its submarkets (Gruber, 2000), as well as with our objective to study the innovation performance of companies that can afford to develop research and development (R&D) activities in foreign countries. 3. The generality measure differs from other typical indexes used to qualify patents, such as technological relatedness or complexity. In fact, technological relatedness is a relative measure that captures the degree of technological overlap between two patents or patent portfolios, while technological complexity focuses on the technological classes of a specific patent, and is meant to assess the ease with which these classes have been recombined with each other in the past. 4. We also perform a similar test by using the share of scientific citations out of the total number of citations (including both patent and scientific citations), rather than the absolute number of scientific citations, and the results are robust (t ¼ 9.444, p > 0.01). Data on scientific citations come from the 'USPTO Patent and Citation Data', published by Sampat Bhaven within the Harvard Dataverse programme (2011, hdl:1902. 5. The US Census defines MSAs as areas composed of a densely populated nucleus, aggregated with those adjacent communities featuring high levels of economic and social linkages with that nucleus (assessed, among other things, via the analysis of commuting patterns). MSAs with a population of at least 1 million inhabitants can be recognized as CMSAs if they demonstrate specific levels of economic and social integration. Our empirical analysis includes 53 MSAs/CMSAs. 6. The data are available at http://www.peterthompson. gatech.edu/. 7. We standardize the independent variables before squaring them. 8. This is the travel time associated with the turning point, which is equal to 1.938 and lies within the range of the variable distribution and within 2 SD above the mean. The travel time associated with the turning point has been computed by accounting for the fact that the independent variable has been standardized and rescaled.