Technological relatedness: how do firms diversify their technology?

The principle of relatedness, which helps estimate the affinity between economic activities, suggests that countries, regions and cities tend to undertake new economic activities when they already perform related activities. This empirical principle has been confirmed for various dimensions—cities, regions, and countries—and their activities—developing new technologies, products, and industries. However, the technological diversification of firms is relatively unexplored. Is a firm more successful at entering a new technology when it has already accumulated related technologies? Here, we explore this issue using a unique dataset that contains firms’ patent data and financial and market information. In particular, we examine Korean firms listed on the Korean stock market that published patents at the patent offices in Korea, Europe and the United States from 1984 to 2014. We develop a technological relatedness measure to estimate whether a firm has already published patents with similar technologies. We find that firms are more likely to develop a new technology when they already have related technologies. Furthermore, We also check the robustness of this effect with varying the definition of proximity and by using propensity score matching. Our findings suggest that the effects of technological relatedness remain significant when varying the proximity and controlling for potential confounding effects. These findings extend the concept of relatedness to a firm’s technological diversification and show that the development of a firm’s technological knowledge is shaped by its technological relatedness.

1. https://www.kisvalue.com/.

2. https://www.epo.org/searching-for-patents/business/patstat.

3. The Korean stock market can be roughly divided into two types. The first type is the Korea Composite Stock Price Index that targets large companies. The second type targets small and medium-sized enterprises (SMEs), including the Korea Securities Dealers Automated Quotation, which focuses on promising SMEs based on advanced technology, and Korea New Exchange, a stock market for SMEs with lower listing thresholds. This study targets the manufacturing companies listed on the above three markets.

4. Although the KIS-value database was launched in 1985, it provides financial information on firms from 1980.

5. https://www.oecd.org/sti/inno/intellectual-property-statistics-and-analysis.

6. We corrected two types of main problems; errors of applicant name; and mismatches caused by historical changes in company name. For example about errors, the official KIS-name ‘SAMSUNG ELECTRONICS CO.,LTD.’ is observed with various forms in OECD HAN database, such as ‘SAMSUNGE ELECT CO LTD (Han-id: 2587222)’, ‘SAMSUNG ECTRONICS CO LTD (2626868)’, ‘SAMSUNG EECTRONICS CO LTD (2626869)’, ‘SAMSUNG EELCTRONICS CO LTD (2602262)’, ‘SAMSUNG EELECT CO LTD (2770031)’, ‘SAMSUNG EKECTRINICS CO LTD (2627307)’, ‘SAMSUNG EL ECTRONICS CO LTD (2770033)’, etc. (It might exist in more diverse forms within PATSTAT as OECD HAN database is the result of already harmonised works.). We collected all possible variations of each firm’s name and re-harmonized their multiple Han-ids into a single unique KIS-id after checking their presence. Next, for example about historical changes in company name, ‘Dongyang Magic’, a company provided rental services of electric home appliances, changed their company name as ‘SK Magic’ after being merged by ‘SK networks’ in 2016. We combined all varied applicant names of ‘Dongyang Magic’ (such as ‘TONGYANG MAGIC INC (3576792)’, ‘TONG YANG MAGIC CO LTD (3019027)’, ‘DONGYANG MAGIC CO LTD (814662)’, etc.) and re-harmonized them into a single unique KIS-id of ‘SK MAGIC INC.’, their official KIS-name.

7. From the research of Wong et al. (2022) with perspective of Regional Innovation System (RIS), we can verify the effect of ’dot-com bubble’ crises on patents and infer the same results at firm level. Before and after ’dot-com bubble’ shock, Wong et al. (2022) observed that there are differences in the rate of patenting activities at the Tiger group with 9 cities (including Seoul, Daejeon and Suwon of Korea). Following this relationship, there are many researches that observed the similar patterns in patent activity of Korea near ’dot-com bubble’ shock. The similar decreasing pattern of annual patents of Korean listed manufacturing firm can be found at the work of Kang et al. (2019), which is the outcome of using the independently constructed dataset. In addition, when we look into the patents applied at KIPO only or together with East Asia & Pacific regions, total number of patents applications recoiled for 2 years after 2000 (Fink, 2013; Olavarrieta & Villena, 2014).

8. A level of classification for measuring diversification is controversial and still remained unsolved. Since there is no review paper comparing explanatory power of different patent classification to our knowledge, we introduce a study of Beaudry and Schiffauerova (2009) that reviewed huge amount of results from different industrial classification level. Beaudry and Schiffauerova (2009) reasoned out a conclusion that selection of 3-digit industrial classification would be a threshold for examining the effects of agglomeration externalities on firm’s innovation. The reason we chose 3-digit patent classification to analyze the mechanism of how technological relatedness affects technological diversification is come from similar backgrounds.

9. When we calculate the proximity \(\phi _{\alpha ,\beta ,t}\), we aggregate the patents for five years from year t since the highest backward citation is achieved by patents published for fewer than five years (Hall et al., 2001), and technology \(\alpha \) at time t is developed based on the background knowledge accumulated from time \(t-5\) to time t

10. We could indirectly deduce the effective periods of previously developed technologies from the frequency of self-citation. Aksnes (2003) observed that the percentage of self-citations is attenuated by the year after publication. It means that the technologies already developed affect developing the next generation of new technologies for certain period. Although its effect often lasted up more than 15 years, we examined the 3 years before the time of observation, t as a rapid decrease in self-citation starts from 3 years after publication.

11. Box–Cox transformation is a data transform method that makes our data fit to normal distribution. All kinds of simple power transformation including the log or the square root cannot change our data to exact normal. To solve this problem, Box and Cox (1964) proposed a method to execute a range of power transforms lead to a normal distribution with no skew. The optimal type of power transforms, which can be ranged from reciprocal to none (log or square root transform are in between), is estimated from our data.

12. Samsung Electronics’ number of IPC assigned to their patents is 156,794 compared with 286,499 for the four major companies, which are Samsung Electronics, LG Electronics, Hyundai Motors, and Posco. The number of Korean IPC of patents granted by the KIPO and applied for or granted by the USPTO is 518,458.

This work was supported by the National Research Foundation of Korea (NRF) Grant funded by the Korea government(MSIT) (Grant Nos. NRF-2022R1A5A7033499 and 2017R1A2B4009376). We also wish to acknowledge the support from Inha University.

Authors and Affiliations

1. Technology Management, Economics and Policy Program, Seoul National University, Seoul, South Korea

   Seung Hwan Kim & Jeong-Dong Lee

2. Department of Economics, Inha University, Incheon, South Korea

   Bogang Jun

3. Department of Data Science, Inha University, Incheon, South Korea

   Bogang Jun

4. Research Center for Small Businesses Ecosystem, Inha University, Incheon, 22212, South Korea

   Seung Hwan Kim & Bogang Jun

Corresponding author

Correspondence to Bogang Jun.

Appendix 1: Review of the measurements

We reviewed three different kinds of methodologies used at Hidalgo et al. (2007), Leten et al. (2016), Ning and Guo (2022), and Pugliese et al. (2019). First, there are three methods to calculate proximity value between two different technologies. All the methods have pros and cons, respectively. Hidalgo et al. (2007) used conditional probability of RTA based co-occurrence, which is minimum probability of the pairwise conditional (in other words, joint) probability that both different technologies have RTA within the same firm. This joint probability presuppose the fact that the two different economic activities requires similar capability to develop both, if two different economic activities are often observed in the same organization. Therefore this method can be said to be rooted to the concept of capability while it can be criticized simultaneously for agnostic way of measurement, which is an indirect inference. Leten et al. (2016) and Ning and Guo (2022) got the proximity value based on patents’ citing-cited relationship. If an observed number of citing-cited relationship between two different technologies of different IPCs is larger than its expected number, the two different IPC assigned to different patents is considered similar. It can measure their similarity directly from a clear link, citing-cited relationship. However there could exist potential biasedness caused by deliberate patenting strategy of the firm (Corsino et al., 2019). Lastly, Pugliese et al. (2019) calculated a conditional probability of technology corresponding to the probability of random binomial case. The number of firms who have RTA at two different technologies is normalized by both the maximum ubiquity of two technologies and the diversity of firm’s technology portfolio. Same as the methodology of Hidalgo et al. (2007), this method is also capability-based measurement, but additionally considers normalization. However as Zaccaria et al. (2014) mentioned, additional efforts are needed to check a posteriori goodness of the result because the exponents of ubiquity and diversity is not joint probability, but conditional probability.

Next, we will examine three different density equations derived from each proximity value. Hidalgo et al. (2007) get the density from share of sum of all proximity values to sum of every proximity values centering around \(\alpha \). This method is a complete structural variable, in other words all elements except each firm’s \(U_{i,\beta ,t}\) are come from proximity value. As we only check whether the technology has RTA or not, this method is free from overstatement of relatedness. For example, when we consider the share of technology in each firm, if proximity value is less than 1 (which is not similar with each other), then the density value is decreased dramatically because of multiplication of two proper fractions. Although we add all weighted shares of technologies, the effects of unrelated technology is sharply diminished. Meanwhile the effect of relatedness becomes magnified from summation of proper fraction (share of technology) multiplying by improper fraction (high proximity value). However, the density suggested by Hidalgo et al. (2007) cannot reflect the degree of development as we substitute the RTA to 1, regardless of their RTA value when it is lager than 1. For the same, but opposite backgrounds, the density Leten et al. (2016) and Ning and Guo (2022) has a merit of considering the degree of development, whereas has a merit of potential overstatement. Lastly, we cannot compare Coherent Technological Diversification (CTD) directly with the aforementioned two methods. Because CTD is a averaged index of all technologies for the firm (that is (firm, year) combination), so the dimension is different with (firm, technology, year) combination.

All above explanations are summarized at Table 9.

Appendix 2: Correlation table

Since the correlation coefficient between \(Age_{i,t}\) and \(Sales_{i,t}\) is insufficiently low, we checked the Variance Inflation Factor (VIF) for all the pairwise variables. There was no result of VIF test with pair of variables that has a value higher than 2. The results of the variance inflation factor test of \(Age_{i,t}\) and \(Sales_{i,t}\) are 1.035 and 1.272, respectively (1.075 and 1.277 when we only consider the main results from 2004 to 2014). The correlation coefficient between \(\omega _{i,\alpha ,t}\) and \(Sales_{i,t}\) is also checked as the value is near 0.4. The results of the VIF test of \(\omega _{i,\alpha ,t}\) and \(Sales_{i,t}\) are 1.129 and 1.431, respectively (1.144 and 1.426 when we consider the regression input only from 2004 to 2014). From the above test results, we use all the covariables together (Table 10).

Appendix 3: Building a technology space for 2004–2014

To construct a representative technology space, patents from 2004 to 2014 are aggregated to generalize the relationship between two different technologies. If not, the annual trend of a certain technology section could be highlighted. After calculating the proximity, \(\phi _{\alpha , \beta }\) in Eq. 1, we build a proximity matrix \(\Phi \) for all pairs of technologies that have a 3-digit IPC level whose sub-sectoral number is 131. To visualize the representative technology space, we follow the methodology of Gao et al. (2021). The first step is building a maximum spanning network that connects all the nodes with the minimum number of links, as shown in Fig. 5A. The result includes 121 links connecting 122 nodes that maximize total proximity and ensure connectivity.

Process of constructing the technology space

Full size image

The second step is to build a maximum weighted network with links whose proximity is above a certain threshold value, \(\phi '\). We set \(\phi '\) as 0.4, meaning that another network consisting of 258 links with 85 nodes is created, as shown in Fig. 5B.

In the third step, we combine the above two networks to create a superposed network comprising 122 nodes with 300 links, as depicted in Fig. 5C. To improve the visualization, we use the ForceAtlas2 algorithm of Gephi (http://gephi.github.io) to locate the node with the better place to minimize the number of overlapping links and untangle dense clusters. Figure 5D illustrates the final composition of the technology space. Lastly, we add color to the nodes to distinguish the IPC category at the 1-digit level. The thickness and its color represent the link weight. The final output is Fig. 3A.

The hierarchically clustered matrix in Fig. 6A shows the colored matrix for all the proximities, \(\Phi \), of every pairwise IPC relationship. Figure 6B and C show this distribution graphically. When we consider all the pairwise relationships of two technologies (\({^{122}}C_{2}\)), its frequency appears as a log-normal distribution. After carrying out the four steps above, we find that the shape of the distribution of 300 links (i.e., only those proximity values used to construct the technology space) does not change substantially.

A Hierarchically clustered matrix of \(\Phi \), B the distribution of all the proximities, \(\Phi \), and C the distribution of \(\phi \) after conducting the above-described four steps

Full size image

Figure 3A shows the technology space of the Korean manufacturing industry for 2004–2014. In agreement with previous findings on the product space and industry space (Hidalgo et al., 2007; Gao et al., 2021), even in the technology space, we can also find similar results for less complicated technologies. As such, the IPC codes of Textiles, Paper (D), Fixed Constructions (E), and Human necessities (A) are located at the periphery of the technology space.

The general characteristics of the network are as follows. The node with the highest degree (i.e., the most connected node) is H02 called “Generation, Conversion, or Distribution Of Electric Power” in electricity (H). The links with the highest weight are B33–G06–H04 and C9–H01. The names of these IPC codes are (B33) Additive Manufacturing Technology; (C99) Subject Matter Not Otherwise Provided For In Chemistry/Metallurgy; (G06) Computing; Calculating; Counting; (H01) Basic Electric Elements; and (H04) Electric Communication Technique.

However, in the technology space at the firm level, we find a unique feature not available at the country or industry level. We can classify the whole into a continent and an island. At 11 o’clock, we identify a cluster that mainly consists of physics (G) and electricity (H). Both links with the highest weight mentioned above (B33–G06–H04 and C99–H01) are located on this island. These technologies are related to home electronic appliances, displays, and semiconductors, which are the main products of Korea’s manufacturing industry. As we mentioned at Sect. 5.1, this cluster is come from Samsung Electronics, which has a dominant effect on the technology space of Korea as it owns 30.24% of IPCs of all Korean patents. Also, as numerous cooperative companies or suppliers or customer of Samsung Electronics are existed and are intertwined within industry, the impact could be bigger when we add up all patents applied by other firms related with Samsung Electronics. Therefore, the clustering structure of technology space, what we named island, calls for caution in interpretation as it could be special case such as Korea whose economy heavily depends on few firms.

One concept suggested by Hidalgo et al. (2007) is ubiquity, or how many countries can produce a certain product. If the ubiquity of a product is low, only few countries can produce it (In other words, the product with low ubiquity is hard to produce). In the product space, those products positioned at the center have low ubiquity, and consequentially only a few countries occupy that space. However, in the technology space, complicated technologies (nodes) that have low ubiquity are found on the island (tied with the highest weighted links). The different shape between product space of countries and technology space of firms can be interpreted from nestedness (Laudati et al., 2022). The nestedness means that the developed economic agents diversify their economic activities and possess many of them, while developing economic agents only possess few ubiquitous products. In this context, country-product is nestedness as the countries that can produce complex products also have the capability to produce less complex products. Comparing to country-product relationship, firm-technology is in-block nestedness, which means being partitioned into specialized technologies, therewith, low technologies of ubiquity are found in general. This is because a firm only develops their technology necessary to create the products confined to their specialty, while a country produces wide range of products from simple to sophisticated (if possible). As a result, as we can see at Fig. 3B–D, technologies mainly consisting of the island of the technology space are occupied by certain firms (in this case, Samsung Electronics) and it means that the technologies owned by firms are localized compared with at the country level.

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.

Reprints and permissions