Localization of collaborations in knowledge creation

Duranton and Overman (2005) and Nakajima et al. (2012) show that nearly half of the manufacturing industries in the UK and Japan tend to be localized when using distance-based methods.

Crescenzi et al. (2016) is an exception. They use inventor-level data and estimate the impact of geographic and other proximities on collaboration in the EU by regression approach. They identify the roles of organizational proximity and of complementarity between several proximity measures and geographical proximities with regard to collaboration. While they focus on estimating the impact of geographical proximity on the probability of collaborations between these inventors, we adopt nonparametric approach to identify the strength and the distance range of localizations of collaboration. Furthermore, as is explained later, we use a finer measure of location of inventions in which are actually occurred.

This concept is the same as the “control patent approach” that is often used in the literature on the localization of patent citations (e.g., Jaffe et al. 1993; Thompson and Fox-Kean 2005).

We use Japanese patent data. Japanese patent system is much different from other major patent systems, except the convention of inventor’s address registered, which is explained later. Indeed, the Japanese patent system has been harmonized to the US and the EU through the European Patent Convention in 1977 and the Patent Cooperation Treaty in 1978. The details of the comparison of the patent systems across these countries are summarized in Japan Patent Office (2017). For a brief survey of the Japanese patent system, see the annual reports from the Japan Patent Office (Japan Patent Office 2014). In addition, Maskus and McDaniel (1999) provide a helpful overview of the Japanese patent system.

One may raise concerns about the transfer of workers. Consider the case in which an inventor transfers from an establishment during the process of collaboration, and she registers her new establishment on the patent application form. In this case, the observed distance between the two establishments that are registered on the patent does not necessarily represent the actual collaborating distance. To check this issue, we estimate the frequency of the inventor’s transfer for the average duration of invention. To identify the inventor’s transfer by patent information, addressing the “same name problem,” i.e., that different researchers have same name, is necessary. Saito and Yamauchi (2015) proposed focusing on the unique names observed in telephone directories. Based on the methodology, we restrict samples to researchers who have “rare names,” and we define movers as researchers who published patents in at least two establishments. We find that 31.1% of researchers transfer their establishments in the 20 years from 1986 to 2005. Suzuki (2011) found that it takes a median of 18 months from starting a project to applying for a patent in Japan. Thus, we roughly estimate that 2.75% (= 1–(1–0.311)^18/12/20) of workers transferred establishments during the invention. On the other hand, share of researcher who registered at least one patent published with collaborators belonging in different establishments (which is “collaboration” in our definition in the analysis) in 18 months is 23.16%. Therefore, the magnitude of the pseudo-collaborations through worker transfer in our collaboration data is inconsequential.

A patent often has multiple IPCs. We use the primary IPC that is assigned to each patent.

For a robustness check, we conduct analyses using a more conservative definition for potential collaborating partners that restrict the establishments that have at least one collaboration experience. However, the results remained qualitatively unchanged. See Appendix 2.

One may argue that establishments can choose their locations by considering the expectations of future collaborations, and if this is true, these establishments choose their locations where many potential collaborators are located. In this case, the counterfactual collaborations will be more localized than the collaborations of the establishments that choose their location regardless of the existence of future collaborators. Counterfactual distribution may include the establishments proximity by the above motivation, and it is interpreted that our estimation strategy identifies the degree of localization of collaborations given in the location distribution of establishments made by the above location choice strategy.

We use great-circle distance, which is the shortest distance along the surface of the globe.

Silverman’s (1986) optimal bandwidth and Gaussian kernel function are used as default in the K-density methodology. To maintain comparability with the previous literature (e.g., Duranton and Overman 2005; Nakajima et al. 2012, Murata et al. 2014), we use the default setting.

Following Duranton and Overman (2005) and Nakajima et al. (2012), we set 180 km as the upper bound of our focus distance. The choice of the upper bound does not qualitatively change our results.

One may raise the concern that a significant number of firms might only have two establishments. In this case, potential collaborations are unique and there are no variations. If there is a collaboration between the two establishments, the actual collaboration always coincides with the potential one. However, there are a significant number of firms that have only two establishments and there is no collaboration within those firms. If distance impedes collaborations, two distant establishments tend not to collaborate within a firm; on the other hand, two close establishments tend to collaborate with each other. Thus, in firms that have only two establishments, we still have enough variations to identify the relationship between distance and collaborations. We find that within firms that have more than one establishment, 27% of firms have more than two establishments. Thus, the identification strategy explained above plays a substantial role in our estimation results.

To confirm the role of geographic proximity, we conducted interviews with firms that have experience with collaborative works. The interviews were conducted in March 2016 in a provincial city in Japan. Managers of these firms mentioned that geographical proximity promotes trust between firms through monitoring and spillovers of each firm’s information and facilitates collaboration between firms.

Using a firm database (Tokyo Shoko Research database), we confirm that this categorization has a strong positive correlation with firm employments (0.40) and sales (0.41). However, one may raise the concern that the “single-firm” includes large-sized foreign-owned companies that have only one research establishment in Japan. From the Survey of Trends in Business Activities of Foreign Affiliates by the Ministry of Economy, Trade, and Industry which is the complete survey of foreign-owned companies (companies for which 1/3 of their stock is owned by foreign companies or investors), we confirm there are just 3185 foreign-owned companies included in the survey in 1998. On the other hand, the number of “single firms” in our data is 46,904. In this sense, the effect of foreign-owned companies on our results is mostly negligible.