A Tale of Two Americas: the Evolution of Innovation Networks across US Cities

Transcription

1 A Tale of Two Americas: the Evolution of Innovation Networks across US Cities Alessandra Fogli Hyunju Lee For the latest version click here Abstract We analyze the geographic dimension of innovation. Innovation is the critical component of long term prosperity and it is unevenly distributed across US. Using data on 1.8 million US patents and their citation properties, we map the innovation network of all major US cities over the last three decades. We find that the innovation gap among cities, which was shrinking until 1980, has recently started growing, generating divergence. We develop a network model of cities that captures knowledge spillovers within and across industries as well as within and across cities, and calibrate it using information on the patterns of patents citations. We show that the IT revolution, by reducing the cost of information exchange across cities, induced an endogenous response of the network structure of cities. This change in network structure can explain a large part of the recent divergence in innovation patterns, and is consistent with a number of stylized facts about the evolution of US cities over the past thirty years. Keywords: Innovation, city divergence, network, patent Alessandra Fogli: Federal Reserve Bank of Minneapolis. afogli00@gmail.com. Hyunju Lee: Department of Economics, University of Minnesota and Federal Reserve Bank of Minneapolis. leex4557@umn.edu. The views expressed herein are those of the authors and not necessarily those of the Federal Reserve Bank of Minneapolis or the Federal Reserve System. 1

2 1 Introduction Americans are more divided than ever, gridlocked over social issues, race, gender and the economy. (The Associated Press, Divided America series 1 ) Many recent studies show that American cities are diverging. Housing price, labor income, wealth, education attainment, and many other economic measures indicate that some parts of the country is becoming more and more affluent, whereas the rest is left behind. (Diamond, 2016, Moretti, 2012) Jacobs (1969) call cities as primary economic organs, functioning as places where ideas are interchanged and growth is spurred. Jacobs emphasize the role of city as the melting pot, where diversities create new ideas. Glaeser et al. (1995) show that for cities between 1956 and 1987, knowledge spillover across industries rather than within is impoartant for economic growth, supporting Jacob s view on cities. We build on the theory of cities as a network where diverse ideas are linked and generate economic growth. This paper studies the network origins of city divergence, where the nexus lies in innovation. We first document that cities are diverging in innovation, especially from 1990s. We use 1.8 million U.S. patents and inspect the growth of patents in each city from 1975 to There was convergence of patent growth from 1980s to 1990s, but divergence takes off from 1990s to 2000s. Cities with higher patent per capita in 1990s experienced faster growth 2000s. Our hypothesis is that differences in city network are driving the divergence of innovation. We consider two dimensions of city network. First is internal network, where different technology fields within a city form a network. Second is external network, where network is formed across different cities. We empirically measure each network structure for 279 US metropolitan areas for 3 decades (1980s s) based on U.S. patents citation data. We find that first, more connected network structure predict strong growth of innovation in 1990s to 2000s. We summarize network structure by degrees, which is a measure of connectedness. Higher degree network is more connected to each other. Cties with lower 1 2

3 degree of either network drives the convergence of innovation growth in and cities with higher degree experience faster growth in innovation in , resulting in divergence. Second, we find that internal network is a push factor that predicts both convergence and divergence of innovation growth, whereas external network is the pull factor that works against the pattern of innovation growth over time. Internal degree of 1980s predict convergence of innovation growth, whereas higher external degree is associated with divergence. In the same token, cities with higher internal degree experience faster growth of innovation in 2000s, whereas higher external degree predicts the opposite direction of innovation growth. Literature review We study the impact of citation network to growth into two dimentions. First is netework among different technologies and fields. Acemoglu et al. (2016) use patent data of in order to measure the cross-field citation network, and find that growth in upstream field propell future innovation in linked fields. The second dimension is geographic links. Akcigit et al. (2017b) show that based on patent data between 1880 and 1940, more geographically connected states are more inventive. The observation of geographical link and patent goes back to Sokoloff (1988), who shows that proximity to waterways have a strong correlation with patents, using patent records between 1790 and Based on citation data, Mohnen and Belderbos (2013) describes matrices of inter-sectoral and international spillover weights in order to access Europe-wide impact of R & D policies. Freeman and Huang (2015) find that more diverse coauthor network leads to higher quality research, based on over 2.5 million scientific papers in Akcigit et al. (2017a) and Moser et al. (2014) both show that fields with more immigrants experienced faster growth, based on historical citation data of and , respectively. 3

4 2 Data In this section, we describe the data we used to show the divergence of innovation across cities. We first construct the data based on patent record from that is used to calculate network structure decribed below. Then, we show the convergence of innovation across US cities from 1980s to 1990s, which contrasts to the divergence in 1990s to 2000s. Patent citation across cities and industries We analyze the patent data from The main source of data on innovation comes from the US patent office (USTPO), where we use data by Li et al. (2014) and Hall et al. (2001). We construct a data set where we merge patent inventors location, technology field, and citation. We focus on the patents where at least one inventor has address in US metropolitan area. We define US metro following 2000 Census definition of Metropolitan Statistical Area (MSA), which are 279 metros. 2 We used Li et al. (2014) who has merged zip code of each inventor into their data, and did cross-walk based on zip code - county - MSA. 3 We construct two sets of data, in order to contruct cross-city citation network and cross-industry citation network within each city. In the first data, each observation is unique in (citing city citing patent cited city cited patent). In cases where there are multiple inventors who live in different metros for one patent, I make a Cartesian product of inventor and patent, for each citing and cited patent. That is, if there are two cities A and B where inventors of citing patent 1 live, and similarly cities C and D for cited patent 2, I will have 4 observations in total: (A 1 C 2 )... (B 1 D 2 ), alphabetically ordered. Each patent has on average 1.26 unique cities that inventors live. We contruct cross-city network by decades, where we have three from 1980 to Each 2 More precisely, we follow IPUMS definition of metros. In IPUMS, some of the MSAs are grouped together into bigger metro areas, which is in between PMSAs (Primary Metropolitan Statistical Area ) and CMSAs (Consolidated Metropolitan Statistical Area). 3 Zip code to county file is downloaded from SAS website: We used 2008 version. For county-msa cross walk, we used census definition for 2000 Census. 4

5 decade is a midpoint of citing patent application year. For example, in 1980 data include citing patents applied from 1975 to Also, we further restrict the sample to the citingcited pair within 10-year window. That is, absolute difference of application year of citing and cited patent should be less than or equal to 10 years to be included in the sample. In the constructed data, we have 590,465 observations in 1980, 2,440,086 in 1990, and 10,204,560 in In the second data, we select samples of citing and cited patent for each 279 metro. If any of the inventors in both citing and cited patent are living in metro A, the citing-cited pair is included in the network of metro A. Then, we merge technology fields of each citing and cited patent, so that we can construct the network of cross-industry citation. There are 37 different technology subcategories, which is then grouped into 6 categories. Following Acemoglu et al. (2016), we construct network of cross-cateogiry citation. What is different in this paper is that we add geographic dimension, by considering the network for each metro area as well as across them. Divergence of innovation We first document divergence of innovation across cities. Our primary measure of innovation in a city is number of patents per capita. Here, number of patents are weighted by citations, using index constructed by Hall et al. (2001). We find that from 1980s to 1990, there was a convergence of growth in innovations across cities. As column (1) of Table 1 shows, cities with 1% higher number of patents per capita in 1980s experienced slow growth of innovation in 1990s by 5.38%. In 1990s to 2000s, however, cities diverge in their innovation growth. Metros with higher initial innovation level takes off, growing by 8.74% more per 1% higher initial level. This pattern is consistent across different measures, such as number of citation weighted patents (Table 1 columns (3) and (4)), number of patents without citation weight, and weighting on population size (Appendix Table 8 and 9). We visualize the convergence and divergence in Figure 1, and label the cities with highest initial level in 1980 and 1990, respectively. In Table 2, we list the top three cities in 1980s and 1990s in the initial level of innovation, which is per capita patents. In 1980, Treton, NJ has the highest patents to population ratio, followed by Dutchess Co., NY, and San Jose, 5

6 Table 1: Divergence of innovation, Innovation measures Log patents per capita Log patents (1) (2) (3) (4) Log patents per capita (1980) ** (0.0267) Log patents per capita (1990) *** (0.0163) Log patents (1980) *** ( ) Log patents (1990) *** (0.0103) Constant *** 0.855*** 0.295*** (0.128) (0.0757) (0.0782) (0.0861) Observations Pop weight Yes Yes Yes Yes F-test Note: Log patents are number of patents weighted by HJT index (Hall et al., 2001). All regressions are weighted by total population of initial period (1980 for columns 1 and 3, 1990 for the rest). Standard errors are in parenthesis. *** p<0.01, ** p<0.05, * p<0.1 Table 2: Divergence of innovation, City Growth Initial level City Growth Initial level Trenton, NJ San Jose, CA Dutchess Co., NY Rochester, NY San Jose, CA Dutchess Co., NY Note: Growth is the log difference of patents per capita, and initial level is the patents per capita in initial decade (1980 for left two columns, and 1990 for right two columns). CA. Looking at their growth of per capita patents from 80s to 90s, San Jose, the city with lowest initial level, experienced the fastest growth among the three. On the other hand, in 1990s, cities with highest initial level of innovation grows the fastest. San Jose, which had the most per capita patent numer among all cities, experienced 131% growth in patents to population ratio. Strong divergence in recent years are captured by the positive slope in Figure 1, as well as in columns (2) and (4) in Table 1. 6

7 Figure 1: Divergence of Innovation,

8 3 Network Model Figure 2: Graphic representation of network In order to explain the forces of convergence and divergence in innovation across cities over time, we formulate the innovation network in the model. We build double-decker network, which has internal network within a city and external network across cities. This network frames patent citation data where we empirically analyze in the next section. In this section, we define each network structure and summary statistics that we use in the following section. Figure 2 shows the concept of network in a simplified way. In our setting, there are multiple cities and categories (cat in the medium pale blue circles), as well as sub-categoreis (subcat in small white circles) within each city. Categories are defined by USTPO in order to classify their patents. There are 6 categoreis (Chemical, Computers & Communications, Drugs & Medical, Electrical & Electronic, Mechanicla, and Others) and total of 37 sub-categoreis. In our sample, there are 279 metropolitan areas. We define internal network as a cross-sub-category network within a city. Each node is a sub-category, and links are citation between each nodes. Links are weighted by the number of citations. Since we observe direction of citation, the network is directed and weighted network. Let M be an internal network matrix of a city. Then, elements of matrix M is 8

9 m(i, j) i, j = 1,... 37, where m(i, j) is a number of citation from sub-category i to subcategory j. As in Acemoglu et al. (2016), knowledge flows from j to i. Since we want to study the effect of cross-industry network to the innovation growth through the internal network, we do not consider within-sub-category citations. Also, as noted by Acemoglu et al. (2016), within-field citation is the largest by the number of patents, network structure will be mostly shaped by within-sub-category citations if included. The same logic holds for the external network, which is defined in the following. The structure of external network is similar, and the only difference is that nodes are now cities instead of technology fields. Links are citations across different cities, regardless of the sub-categories. That is, even if the citation is from sub-category A to the same field A, but inventors of citing and cited patents are from different cities, the observation will be included in external network. We use three summary statistics to summarize a network structure. I describe the statistics for internal network, and analogous definition applies to external networks as well. The three measures are average path length, degree without citation number weight, and degree with weight. Average path length captures the length of shortest path on average from one node to another. Degree describes the number of nodes one node is connected to, and we measure by with and without citation weights on the link. With shorter average path length, and higher degree, a network is denser and nodes are well-connected. Formal definitions are in the following. 1. Average path length (avpath) Shortest path to get from subcategory i to j (p i j ) on average. avpath = 1 n i j i p i j n 1 (1) I calculate it by multiplying network structure. I have set the max length to be 37 for those would be otherwise infinite. 2. Degree, without weight (deg 01) 9

10 We count by in-degree, which is the number of citations each node gets. That is, degree of a node i is d01 i = #{j : m(j, i) > 0} (2) Density of a network is deg 01 = 1 n i j i d01 i n 1 (3) 3. Degree, weighted (deg w) We count the number of citations that each node got. d i = j i m(j, i) (4) Weighted density of a network is deg w = 1 n i j i d i n 1 (5) 4 Network and Innovation In this section, we analyze the internal and external network defined in the previous section using patent citation data from More specifically, we ask whether network structure is associated with the growth in innovation, and which network - external or internalis driving the convergence and divergence of innovation documented in the previous section. We calculate the three summary statistics (average path length, degree without citation weight, and degree with citation weight) for internal network of 279 metros and external network across cities in each decades for Summary statistics of the entire variables are in the Appendix. First, we find that network structure predict growth of innovation. Cities with lower degree 10

11 of either network drives the convergence of innovation growth in Cities with higher degree experience faster growth in innovation in , resulting in divergence. In Tables 3 and 4, we divide cities into two groups and show the result of innovation convergence and divergence. One is higher degree group (which is deg w defined in the previous section), those cities with higher than median degree, and the lower degree group. Table 3 is based on external network and Table 4 is based on internal network. In both tables, growth of innovation is negative for low degree groups over the entire period, and especially so in 1980s by external network and 1990s by internal network. Also, cities with degree, either internal or external, exprecienced strong growth in innovation from 1990s to 2000s, whereas there was no such divergence in 1980s to 1990s. Figures 3 and 4 show the patterns into scatter plots, where low degree group (grey, solid line) always experience convergence and high degree group (maroon, dash line) take off in 1990s s. Second, we find that internal network is a push factor that predicts both convergence and divergence of innovation growth, whereas external network is the pull factor that works against the pattern of innovation growth over time. In Tables 5 and 6, we control the innovation growth with internal and external network degree. In the last column of Table 5, we find that internal degree of 1980s predict convergence of innovation growth, whereas higher external degree is associated with divergence. In the same token, in Table 6, cities with higher internal degree experience faster growth of innovation in 2000s, whereas higher external degree predicts the opposite direction of innovation growth. 11

12 Table 3: Divergence of innovation by external network structure, Log patents per capita External network degree High Low High Low Log patents per capita (1980) *** (0.0463) (0.0365) Log patents per capita (1990) *** (0.0438) (0.0342) Constant * *** (0.271) (0.166) (0.239) (0.144) Observations Pop weight Yes Yes Yes Yes F-test Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Table 4: Divergence of innovation by internal network structure, Log patents per capita Internal network degree High Low High Low Log patents per capita (1980) (0.0465) (0.0384) Log patents per capita (1990) ** 0.107*** (0.0407) (0.0348) Constant ** *** (0.269) (0.174) (0.222) (0.147) Observations Degree Low High Low High Pop weight Yes Yes Yes Yes F-test Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 12

13 Figure 3: Divergence of innovation by external network structure, Note: 13

14 Figure 4: Divergence of innovation by internal network structure, Note: 14

15 Table 5: Divergence of innovation by internal/external network, Log patents per capita Baseline External Internal Both Log patents per capita ** (0.0215) (0.0329) (0.0358) (0.0359) Log external degree ** (0.0142) (0.0416) Log internal degree ** *** (0.0126) (0.0368) Constant ** 0.450*** (0.103) (0.178) (0.159) (0.184) Observations Adjusted R-squared Pop weight Yes Yes Yes Yes F-test Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Table 6: Divergence of innovation by internal/external structure, Log patents per capita Baseline External Internal Both Log patents per capita *** 0.101*** * (0.0210) (0.0298) (0.0319) (0.0269) Log external degree *** (0.0138) (0.0271) Log internal degree *** 0.298*** (0.0120) (0.0244) Constant 0.905*** 0.878*** 0.373** 0.907*** (0.0932) (0.168) (0.152) (0.136) Observations Adjusted R-squared Pop weight Yes Yes Yes Yes F-test Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 15

16 5 Conclusion This paper studies the network origins of city divergence, where the nexus lies in innovation. We first document that cities are diverging in innovation, especially from 1990s. There was convergence of patent growth from 1980s to 1990s, but divergence takes off from 1990s to 2000s. Cities with higher patent per capita in 1990s experienced faster growth 2000s. As of the interaction of cities and networks, we find that first, more connected network structure predict strong growth of innovation in 1990s to 2000s. We summarize network structure by degrees, which is a measure of connectedness. Higher degree network is more connected to each other. Cties with lower degree of either network drives the convergence of innovation growth in and cities with higher degree experience faster growth in innovation in , resulting in divergence. Second, we find that internal network is a push factor that predicts both convergence and divergence of innovation growth, whereas external network is the pull factor that works against the pattern of innovation growth over time. Internal degree of 1980s predict convergence of innovation growth, whereas higher external degree is associated with divergence. In the same token, cities with higher internal degree experience faster growth of innovation in 2000s, whereas higher external degree predicts the opposite direction of innovation growth. Future work will study more on the comparison of internal versus external network, inspecting the relation of network with industry structure and other city environments as well. 16

17 References Acemoglu, Daron, Ufuk Akcigit, and William R. Kerr (2016), Innovation network. Proceedings of the National Academy of Sciences, 113, , URL org/content/113/41/11483.abstract. Akcigit, Ufuk, John Grigsby, and Tom Nicholas (2017a), Immigration and the rise of american ingenuity. Technical report, National Bureau of Economic Research. Akcigit, Ufuk, John Grigsby, and Tom Nicholas (2017b), The rise of american ingenuity: Innovation and inventors of the golden age. Diamond, Rebecca (2016), The determinants and welfare implications of us workers diverging location choices by skill: The American Economic Review, 106, Freeman, Richard B. and Wei Huang (2015), Collaborating with people like me: Ethnic coauthorship within the united states. Journal of Labor Economics, 33, S289 S318, URL Glaeser, Edward L, JoséA Scheinkman, and Andrei Shleifer (1995), Economic growth in a cross-section of cities. Journal of monetary economics, 36, Hall, Bronwyn H, Adam B Jaffe, and Manuel Trajtenberg (2001), The nber patent citation data file: Lessons, insights and methodological tools. Technical report, National Bureau of Economic Research. Jacobs, Jane (1969), The economy of cities. Vintage. Li, Guan-Cheng, Ronald Lai, Alexander DAmour, David M Doolin, Ye Sun, Vetle I Torvik, Z Yu Amy, and Lee Fleming (2014), Disambiguation and co-authorship networks of the us patent inventor database ( ). Research Policy, 43, Mohnen, Pierre and Rene Belderbos (2013), Intersectoral and international r&d spillovers. Technical report, SIMPATIC Working paper 02 (Bruegel, Brussels). Moretti, Enrico (2012), The new geography of jobs. Houghton Mifflin Harcourt. 17

18 Moser, Petra, Alessandra Voena, and Fabian Waldinger (2014), German jewish émigrés and us invention. The American Economic Review, 104, Sokoloff, Kenneth L (1988), Inventive activity in early industrial america: Evidence from patent records, The Journal of Economic History. 18

19 A Summary statistics Table 7: Summary statistics Variable Mean Std. Dev. N ln int deg w ln int deg ln int deg w ln int deg ln int avpath ln int avpath ln ext deg w ln ext deg ln ext deg w ln ext deg ln ext avpath ln ext avpath ln r patw pop ln r patw pop ln r patw pop ln Npatent ln Npatent ln Npatent ln Npatent w ln Npatent w ln Npatent w pop tot pop tot pop tot

20 B Robustness Check Table 8: Log difference, Urban (1) (2) (3) (4) (5) (6) VARIABLES ln diff pat80 ln diff pat w80 diff ln r patw pop80 ln diff pat80 ln diff pat w80 diff ln r patw pop80 ln Npatent *** *** (0.0144) (0.0108) ln Npatent w *** *** (0.0143) (0.0108) ln r patw pop *** ** (0.0226) (0.0215) Constant 0.770*** 0.768*** *** 0.847*** (0.0914) (0.0912) (0.118) (0.0878) (0.0877) (0.103) Observations Adjusted R-squared F-test Pop weight Yes Yes Yes Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Table 9: Log difference, Urban (1) (2) (3) (4) (5) (6) VARIABLES ln diff pat90 ln diff pat w90 diff ln r patw pop90 ln diff pat90 ln diff pat w90 diff ln r patw pop90 ln Npatent *** *** (0.0147) (0.0118) ln Npatent w *** ** (0.0143) (0.0112) ln r patw pop *** 0.105*** (0.0222) (0.0210) Constant 0.454*** 0.199** 0.808*** 0.534*** 0.312*** 0.905*** (0.101) (0.0983) (0.108) (0.101) (0.0955) (0.0932) Observations Adjusted R-squared F-test Pop weight Yes Yes Yes Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 20