Relaedness and Technological Change in Ciies: The rise and fall of echnological knowledge in U.S. meropolian areas from 1981 o 2010 Ron Boschma *, Pierre-Alexandre Balland, Dieer Franz Kogler β, * CIRCLE, Lund Universiy, Sölvegaan 16, S-22100 Lund, Sweden Deparmen of Economic Geography - URU, Urech Universiy, Heidelberglaan 2, 3508 TC Urech, The Neherlands Ron.Boschma@circle.lu.se Deparmen of Economic Geography - URU, Urech Universiy, Heidelberglaan 2, 3508 TC Urech, The Neherlands p.balland@uu.nl β School of Geography, Planning & Environmenal Policy, Universiy College Dublin, Newman Building, Belfield, Dublin 4, Ireland dieer.kogler@ucd.ie 1
Relaedness and Technological Change in Ciies: The rise and fall of echnological knowledge in U.S. meropolian areas from 1981 o 2010 Absrac This paper invesigaes by means of USPTO paen daa wheher echnological relaedness a he ciy level was a crucial driving force behind echnological change in 366 U.S. ciies from 1981 o 2010. Based on a hree-way fixed effecs model, we find ha he enry probabiliy of a new echnology in a ciy increases by 30 percen if he level of relaedness wih exising echnologies in he ciy increases by 10 percen, while he exi probabiliy of an exising echnology decreases by 8 percen. Keywords: relaedness, echnological change, urban diversificaion, U.S. ciies, echnology space JEL codes: O33, R11, L65, D83 2
1. Inroducion In evoluionary hinking, knowledge producion is ofen depiced as a cumulaive, pahdependen and ineracive process (Akinson and Sigliz, 1969; Dosi, 1982; Nelson and Winer, 1982). Because of uncerainy, agens draw on knowledge acquired in he pas, which provides opporuniies bu also ses limis o wha can be learned (Heiner, 1983; Cohen and Levinhal, 1990). This happens a he organizaional level, where knowledge accumulaes wihin he boundaries of he firm, bu also a he level of erriories, as demonsraed by he cumulaive and ofen persisen naure of echnological specializaion in counries and ciies (Archibugi and Piana, 1992; Lundvall, 1992; Malmberg and Maskell, 1997; Canwell and Verova, 2004, Sonn and Sorper, 2008). More recenly, research effors have been direced owards he process of geographical diversificaion. Scholars like Hidalgo e al. (2007), Hausmann and Klinger (2007) and Hausmann and Hidalgo (2010) have argued ha he exising se of capabiliies in a counry deermines which new indusries will be feasible, and mos likely, o develop in he fuure. By analyzing dynamics in he expor porfolios of counries, Hausmann and Klinger (2007) showed ha counries predominanly move ino new expor producs ha are relaed o heir curren expor baske. Neffke (2009) suggess ha capabiliies may no move wih ease also wihin counries, and herefore regions are considered o possess specific capabiliies ha define which new indusries are more likely o emerge and develop in he fuure. Sudies by Neffke e al. (2011) and Boschma e al. (2013) and Esslezbichler (2013) on he long-erm indusrial evoluion of regions found ha a new indusry is more likely o ener a region when i is relaed o oher indusries already in place, and ha an exising indusry had a higher probabiliy o exi a region when i was no, or poorly, relaed o oher indusries already presen in he region. However, hese sudies neglec some of he imporan feaures of ciies ha may affec diversificaion as emphasized in he agglomeraion economies lieraure, such as urban densiy and echnological specializaion. Moreover, hese sudies make he claim ha regional diversificaion is driven by echnological relaedness a he ciy level, bu hey analyze his process in erms of indusrial dynamics (i.e. he rise and fall of indusries in regions). However, analyzing relaed diversificaion in ciies by means of echnological dynamics (i.e. he rise and fall of echnological knowledge in ciies) would make a more direc link beween urban diversificaion and he underlying echnological naure of relaedness in ciies. Therefore, insead of focusing on indusrial dynamics in regions, we focus in his paper on echnological knowledge dynamics in ciies, and we analyze wheher he rise and fall of echnological knowledge is shaped by he exising knowledge base of ciies. We draw on he agglomeraion economies and he innovaion sudies lieraures (in paricular on sudies ha discuss he economic performance and echnological diversificaion of urban cenres) o explain echnological change in ciies. Our main claim is ha ciies are more likely o diversify ino new echnologies ha are relaed o heir exising local se of echnologies. In his respec, we no only poin ou ha ciy characerisics drive he process of diversificaion, bu also he overall se of echnologies ha are presen in ciies. Based on paen daa from he Unied Saes Paen and Trademark Office (USPTO), we invesigae he long-erm evoluion of he paen echnology class porfolios of 366 U.S. ciies for he period 1981-2010. Firs, we consruc a so-called echnology space in which we measure he degree of relaedness beween 438 echnologies (main paen classes). Then, we deermine he relaedness beween new and disappearing echnologies and he se of pre-exising echnologies in ciies. Finally, we esimae a hree-way fixed-effecs model by using linear probabiliy OLS regression. The resuls indicae ha echnological relaedness a he ciy level was a crucial driving force behind echnological change in U.S. ciies over he las 30 years. 3
The srucure of he paper is as follows. Secion 2 ses ou he main heoreical ideas on echnological knowledge dynamics a he urban scale. Secion 3 describes he daa, and Secion 4 oulines he mehodology. We explain he way relaedness beween paen echnology classes was defined, and how we assess he impac of echnological relaedness a he ciy level on he rise and fall of paen classes in U.S. ciies. Secion 5 presens he findings, and he final secion provides a discussion and concluding remarks. 2. Technological change and relaed diversificaion in ciies Ciies are engines of invenion and economic growh (Hall, 1998; Beencour e al., 2007). In he las decade, scholars have been engaged in research o deermine wheher Jacobs or Marshallian exernaliies affec urban invenion raes (see e.g. Feldman and Audresch, 1999; Paci and Usai, 1999; Ejermo, 2005; O Huallachain and Lee, 2010). Concisely, Jacobs exernaliies are associaed wih an urban srucure composed of a variey of echnologies ha spark creaiviy, enable he cross-ferilizaion of ideas among secors, and hus generae more invenions. By conras, Marshallian exernaliies are cos-reducing exernaliies, in which he echnological specializaion of a place enables he beer maching of skilled labour and inpuoupu ransacions, and more effecive learning by means of knowledge spillovers. Generally, empirical sudies repor raher inconclusive resuls concerning he quesion wheher echnological specializaion or diversiy leads o higher invenion raes (Beaudry and Schiffauerova, 2009). Paci and Usai (1999) showed in a sudy on 784 Ialian local labour sysems for he period 1978-1995 ha paening aciviy is enhanced boh by indusrial specializaion and diversiy. Auan-Bernard (2001) found ha echnological specializaion promoed paen aciviy in French regions. Ejermo (2005) found a posiive relaionship beween echnological specializaion (as proxied by paen similariy) and paen produciviy in Swedish labour marke regions, while O huallachain and Leslie (2007) and Lobo and Srumsky (2008) found a posiive relaionship beween per capia paen raes and paen specializaion in U.S. ciies. In heir sudy on he invenion porfolios and he paening inensiy of U.S. ciies, O huallachain and Lee (2010) showed ha urban invenion raes are affeced by echnological specializaion and diversiy, and ha he mos invenive ciies have deep specializaions in differen echnologies. While hese sudies on urban specializaion versus diversiy have led o valuable insighs, hey end o rea he echnological or indusrial srucure of ciies as given, as if hey remain he same, while in realiy, hose urban srucures change over ime. Moreover, mos of hese sudies (he 2005 sudy of Ejermo being a noable excepion) do no fully characerize he underlying knowledge sock in ciies, and hus he naure of associaion beween he echnology/indusry classes found in ciies remains largely unspecified. Frenken e al. (2007) and Neffke (2009), among ohers, have argued ha he echnological relaedness or coherence beween indusries in ciies is crucial in his respec, as relaedness deermines learning poenials beween echnologies and indusries in ciies. Only recenly, sudies have aken a more dynamic approach on he echnological and indusrial srucures of erriories, and have combined ha wih a relaedness perspecive. In he pas, sudies have shed ligh on he cumulaive and persisen naure of echnological specializaion in counries (Archibugi and Piana, 1992; Lundvall, 1992; Canwell and Verova, 2004), bu recenly research effors are also frequenly direced owards he process of erriorial diversificaion (Kogler e al., 2013). Hidalgo e al. (2007) and Hausmann and Hidalgo (2010) argue ha exising capabiliies in counries affec heir possibiliies o develop new indusries, and ha hese capabiliies are no inernaionally radable. Hausmann and Klinger (2007) 4
demonsrae ha counries end o expand heir expor aciviies by moving ino expor producs ha are relaed o heir presen expor porfolio, and ha counries wih a wide range of relaed expor producs have more opporuniies o deploy heir capabiliies ino new relaed expor producs. Economic geographers have claimed ha he urban or regional scale migh be even more imporan for his process of relaed diversificaion (Boschma e al., 2013), as many capabiliies do no move easily wihin counries as well (Neffke 2009). In his conex, Maskell and Malmberg (1999) poin o he significance of localized capabiliies, which are associaed wih a paricular local knowledge base and insiuional conex. As regions accumulae specific compeences, hese offer addiional learning opporuniies for local organizaions and lower search coss for new knowledge in similar fields. Consequenly, search behavior for new knowledge ends o be myopic and localized, boh in cogniive and geographical erms (Lawson, 1999; Maskell and Malmberg, 2007). Such geographically localized learning is embedded in local insiuions, like social convenions ha creae muual undersanding beween local agens and make hem inerac and learn (Sorper, 1995; Gerler, 2003). These localized capabiliies are regional inangibles asses wih a high degree of aciness ha are difficul o replicae in oher places. Only very recenly, here is a growing awareness ha hese geographically localized capabiliies also operae as a key source of echnological diversificaion (Boschma and Frenken, 2011). Technological diversificaion is accompanied wih high risks and swiching coss because he capabiliies of firms and heir embeddedness in he local environmen clearly limi he possibiliies o move in compleely differen echnology secors and markes. Therefore, usually when firms diversify ino new echnologies and producs, hey will say close o heir exising capabiliies (Penrose, 1959; Teece e al., 1994; Anonelli, 1995; Breschi e al., 2003; Pisciello, 2004), and remain in he same locaion where hey can more easily draw on relaed capabiliies (Frenken and Boschma, 2007; Buensorf and Guenher, 2011). There is srong evidence from longiudinal sudies on indusries ha many successful enrepreneurs in new indusries do exploi regional compeences hey previously acquired in echnologically relaed indusries (Klepper, 2007; Buensorf and Klepper, 2009), in paricular during he infan sage of he indusry (Boschma and Wening, 2007). I is also likely ha new indusries recrui skilled labour from local relaed indusries and benefi from ha, as he local supply of relaed skills enables easier maching of labour and enhances learning processes (Eriksson, 2011). A large body of descripive sudies has demonsraed ha new local indusries are indeed rooed in relaed regional aciviies (see e.g. Chapman, 1991; Glaeser, 2005; Bahel e al., 2011). Recenly, more quaniaive sudies (Neffke e al., 2011; Boschma e al., 2013) have focused on his process of relaed diversificaion in a large number of regions in counries like Sweden and Spain over a long period of ime. These sudies found sysemaic evidence ha new indusries are more likely o ener a region when hese are echnologically relaed o oher indusries in ha region. Anoher ineresing finding was ha an exising indusry has a lower probabiliy o exi a region when ha indusry is echnologically relaed o oher indusries in he region. This laer finding is as expeced, considering ha hese indusries are more cenrally posiioned and more fully embedded in he local neworks of relaed indusries, and because hey are beer capable of securing heir vesed ineress hrough heir srong ies wih oher local sakeholders, including policy makers (Hassink, 2010). In essence, high coss preven regions o build compleely new indusries from scrach and o abandon exising indusries ha are deeply rooed locally. Thus i is no surprising ha empirical sudies find ha he rise and fall of indusries in regions is subjec o a pah-dependen process, 5
which is driven by he degree of echnological relaedness wih oher local indusries. This also explains why he indusrial srucure in regions is mos likely echnologically cohesive; somehing ha ends o persis over ime despie he fac ha indusries come and go (Rigby and Esslezbichler, 1997; Neffke e al., 2011; Rigby, 2012; Esslezbichler 2013). I is no he lack of indusrial dynamics, bu precisely he acual occurrence of (quie regular paerns of) indusrial dynamics ha makes he echno-indusrial srucure of regions raher cohesive. This is mainly due o he exi or loss of exising indusries, which ends o lower variey bu increase relaed variey in regions, as more unrelaed indusries are more likely o disappear. Alhough he enry of new indusries injecs new variey ino regions, his will no concern jus any indusry, bu raher indusries ha are echnologically relaed o oher regional indusries (Neffke, 2009). The sudies on relaed diversificaion briefly menioned above have focused on indusrial dynamics in regions, and how he degree of echnological relaedness wih exising regional indusries impaced on he rise and fall of indusries. While providing imporan insighs, hese sudies on regional diversificaion also suffer from wo heoreical shorcomings. Firs, hey have negleced feaures of ciies like populaion densiy and echnological specializaion of ciies ha may affec he diversificaion process, as emphasized in he agglomeraion economies lieraure. Second, alhough hese sudies argue ha regional diversificaion is driven by echnological relaedness a he regional scale, hey analyze his process by means of he rise and fall of indusries in regions. However, i would be more plausible o analyze relaed echnological diversificaion in ciies by means of echnological dynamics insead of indusrial dynamics, as a sudy on he rise and fall of echnological knowledge in ciies would esablish a more direc link beween relaed urban diversificaion and is underlying echnological naure. Therefore, his paper analyzes wheher he rise and fall of invenions, or more precisely, he enry and exi of paen echnology classes in ciies is condiioned by he exising se of echnological knowledge in hese meropolian areas. I is expeced ha echnological relaedness a he urban level is a key driving force behind echnological change in ciies. In addiion, i is also expeced ha also oher more general feaures of ciies such as populaion densiy and echnological specializaion migh affec he process of echnological diversificaion in ciies. To es his, we invesigae he evoluion of paen porfolios in 366 U.S. ciies for he period 1981-2010. 3. Consrucing he daase Paens and paen saisics encompass an incredible wealh of informaion wih he poenial o faciliae a muliude of approaches in he invesigaion of knowledge creaion and diffusion processes (Scherer, 1984; Griliches, 1990; Jaffe and Trajenberg, 2002). Paen saisics are considered a noisy indicaor when uilized as an overall measure of economic and invenive aciviy mainly because paened invenions do no represen all forms of knowledge producion wih an economy and hus cerainly do no capure all produced knowledge (Pavi, 1985; Griliches, 1990). Neverheless, if he focus is on economic valuable echnical knowledge ha perains o invenions of uiliy, paens provide an excellen opporuniy for he sudy of echnological change. For example, more recenly paen daa have been uilized o sudy he evoluion of echnologies by aking advanage of he largely unexploied informaion of echnology classes ha are lised in paen documens (Fleming and Sorenson, 2001; Nesa, 2008; Quararo, 2010; Srumsky e al., 2012). Following his lead, he aim of our analysis is o exend his approach and empirically es how he presence of, and relaedness among paen classes shapes echnological change in an urban seing. Specifically, he goal is o ouline a 6
model ha describes he emergence, as well as he exi, of new echnologies in U.S. meropolian areas from 1981 o 2010. There are several paen daabases ha are publicly available for research purposes. Two prominen examples are he Paen and Ciaions Daa of he Naional Bureau of Economic Research (Hall e al., 2001) and he Paen Nework Daaverse from he Insiue for Quaniaive Social Science a Harvard Universiy (Lai e al., 2011). The Unied Saes Paen and Trademark Office (USPTO) served as original daa source for boh of hese, and furher provides supplemen informaion, i.e. he USPTO Harmonizaion of Names of Organizaions Daa File (USPTO, 2010), ha allowed for an exension of hese daabases, which are uilized in he presen sudy. To faciliae an analysis peraining o echnological change in U.S. ciies, individual paen documens were allocaed o one of 949 Core Based Saisical Areas (CBSAs) based on he firs invenor s residency record (OMB, 2010). For more recen records, his was an efforless ask due o he availabiliy of ZIP codes. However, for some of he older records in he available paen daabases, i was necessary o use he geographical correspondence engine available hrough he Missouri Census Daa Cener, in order o link invenor records o heir respecive ciies of residency a he ime he invenion was developed. I was hen deemed o be reasonable o only focus on he 366 Meropolian Saisical Areas (MSAs) raher han he whole populaion of ciies ha also includes micropolian areas, simply because MSAs house well over 90% of all paened uiliy invenion in he U.S. over he pas hree decades. In order o produce resuls ha reflec he real iming of invenive aciviy, and hus he enry and exi of echnologies in ciies, all indicaors ha were developed in he daase subsequenly are based on he applicaion raher han on he gran daes lised in he original paen documens. This is also mainly o avoid ime skewed resuls due o he coninuously increasing ime-lag from he dae of invenion and filing o he gran dae over 30-year ime frame. Paens are classified ino one or more disinc echnology classes ha reflec he scope of he approved claims lised in a paen documen. Based on he available daa, here are 438 main paen classes, i.e. echnology codes ha uiliy paens have been assigned o by he USPTO over he pas hree decades. I should be noed ha his refers o he number of paen specific codes ha were up o dae by end of 2010 raher han all codes ha he USPTO has ever uilized since i was esablished more han 200 years ago. In essence, he USPTO redefines classes, adds new ones, and someimes, alhough rarely, even abandons exising ones, on an ongoing basis. All of his is documened in he monhly classificaion orders ha are issued by he USPTO, indicaing changes o he definiion of he classificaion sysem a a given ime. The advanage of his coninuous process is ha i provides a consisen se of echnology classes ino which paens are placed, somehing essenial in an invesigaion ha relies on daa colleced over prolonged ime period. Srumsky e al. (2012) provide a deailed accoun of paen echnology codes and how hese should be inerpreed. The number of 438 main paen classes uilized in he presen analysis also corresponds o he number used in oher relevan sudies, including Rigby (2013) and Kogler e al. (2013). The spaial and echnology codes ha have been consruced and idenified in he paen daabase briefly oulined here will serve as poin of deparure for he analysis ha follows. 4. Indicaors of relaedness and economeric model As explained in Secion 2, we expec relaedness o be a major driving force behind echnological change in ciies over ime. We argue ha new echnologies emerge from he recombinaion of exising echnological knowledge, leading o he diversificaion of ciies ino echnological aciviies ha are relaed o heir specific knowledge srucure. To es hese 7
predicions, we follow he mehodology developed and applied in recen sudies by Hidalgo e al. (2007), Neffke e al. (2011), Rigby (2013) and Boschma e al. (2013). Firs, we consruc a so-called echnology space in which we measure he degree of relaedness beween all echnologies. Second, we deermine he relaedness beween new and disappearing echnologies and he pre-exising echnological knowledge srucure of ciies, which we define as relaedness densiy (densiy of relaed echnologies). We use USPTO paen daa o regress he enry and exi of echnological aciviies in U.S ciies during he period 1976-2010 on he degree of echnological relaedness of hese aciviies wih he exising ones in ciies. 4.1. The echnology space To measure he relaedness beween echnologies and draw he resuling echnology space, we follow he produc space framework proposed by Hidalgo e al. (2007). The produc space is a nework-based represenaion of he economy, where he nodes define produc caegories and he ies beween hem indicae heir degree of relaedness. The main idea developed by Hidalgo and his colleagues o capure relaedness is o look a how ofen wo producs are expored by counries. Two producs are hen considered o be relaed if hey are co-expored by many counries, because hey are assumed o require he same capabiliies following he reasoning oulined in deail above. Using his framework, we consruc he echnology space, which is a nework-based represenaion of he relaedness beween all he echnologies ha can be found in he domesic paen porfolio of he Unied Saes from 1976 o 2010. In his n*n nework, each node i (i =1,,n) represens a specific echnological class. Applying he measure o he 3-digi USPTO main paen classes (Hall e al., 2001), we are able o map he degree of relaedness beween 438 differen echnological classes. For insance, one of he bigges nodes in his nework represens he echnological class 800 ("mulicellular living organisms"), which is a sub-caegory of he bioechnology class 1. To compue he degree of relaedness beween all hese 438 echnologies and draw he resuling nework, we focus on how ofen wo echnologies are found wihin he same U.S. ciy, defined as a Meropolian Saisical Area (MSA). 2 The relaedness ϕi, j, beween each pair of echnologies i and j is compued by aking he minimum of he pair-wise condiional probabiliies of ciies paening in one echnological class i given ha hey paen in anoher echnological class j during he same period. To avoid he noise induced by negligible paening aciviy, we only consider he primary echnological classes lised on paen documens in which ciies have a revealed comparaive advanage (RCA), as proposed by Hidalgo e al. (2007). { ( RCAx RCAx ), P( RCAx RCAx )} i, j, = min P i, j, j, i, ϕ (1) where a ciy c has a RCA in echnology i in ime if he share of echnology i in he ciy's echnological porfolio is higher han he share of echnology i in he enire U.S. paen porfolio. More formally, RCA c, ( i) = 1 if: c paens paens c, c, ( i) / ( i) / i paens c, ( i) paens c i c, ( i) > 1 (2) 8
We compue he relaedness ϕi, j, beween each pair of echnologies i and j for 7 periods of 5- years, from 1976 o 2010. As presened in Table 1, he correlaion beween hese 7 marices of relaedness is very high, indicaing ha echnological change is a slow, gradual, and pah dependen process. Figure 1 provides a visual impression of he echnology space based on he average degree of relaedness for he enire period 1976-2010. From his graph, i is clear ha he differen echnological classes end o form inerconneced groups ha closely corresponds o he classificaion 3 in six main echnological areas (Mechanical, Chemical, Drugs and Medical, Elecrical and Elecronic, Compuers and Communicaions, Ohers) as proposed by Hall e al. (2001). Table 1. Change in relaedness beween echnologies (1976-2010) 1976-1980 1981-1985 1986-1990 1991-1995 1996-2000 2001-2005 2006-2010 1976-1980 1.000 - - - - - - 1981-1985 0.800 1.000 - - - - - 1986-1990 0.750 0.798 1.000 - - - - 1991-1995 0.725 0.768 0.836 1.000 - - - 1996-2000 0.711 0.740 0.802 0.852 1.000 - - 2001-2005 0.690 0.720 0.782 0.825 0.868 1.000-2006-2010 0.597 0.638 0.688 0.702 0.735 0.773 1.000 As a robusness check, however, we will also verify wheher he economeric resuls hold for alernaive measures of relaedness. Firs, we consruced similar nework marices of he echnology space by measuring relaedness hrough normalized co-occurrences. Following i Cancho and Solén (2001) in he conex of words co-occurence wihin senences, we consider ha echnological classes are relaed when hey co-occur more han one can expec by chance, ha is, when he observed co-occurrences are higher han he expeced values based on probabiliy calculus. Second, we consruced a relaedness marix where echnological classes are relaed if hey are lised in he same (2 digis) sub-caegory 4 of he paen classificaion proposed by Hall e al. (2001). 9
Figure 1. The U.S. echnology space (1976-2010) Noes: The "echnology space" is consruced in a similar fashion han he "produc space" proposed by Hidalgo e al. (2007). Each node (n=438) represens a paen echnology class (see Hall e al., 2001), and he links beween hese paen classes indicae heir echnological relaedness. 10
4.2. Relaedness densiy of U.S. ciies To analyze how relaedness influence echnological change wihin ciies, we need o consruc a ciy-echnology level variable 5 ha indicaes how close a poenial new echnology is o he exising echnological porfolio of a given ciy. This idea is operaionalized by he densiy index (Hidalgo e al., 2007), which measures in our case he proximiy of a new echnology o he exising se of echnologies in a given ciy. More formally, he densiy around a given echnology i in he ciy c in ime is compued from he echnological relaedness 6 of echnology i o he echnologies in which he ciy c has a RCA in ime, divided by he sum of echnological relaedness of echnology i o all he oher echnologies in he U.S. in ime : RD i ij, c, j c, j i ϕij j i = ϕ 100 (3) By consrucion, his relaedness densiy variable lies beween 0 % and 100 %. A ciyechnology level densiy equal o 0 % indicaes ha here is no echnology relaed o echnology i in he ciy c, while a value of 100 % indicaes ha all he echnologies relaed o echnology i belong o ciy c's echnological porfolio. To ake a concree example, he densiy around he echnological class "Semiconducor Devices" (class 438) in he Alana meropolian area 7 was equal o 52% during he period 1981-1985. In fac, he echnological class "Semiconducor Devices" was relaed o 34 oher classes in oal, and 18 of hese 34 classes belonged o Alana's echnological porfolio a ha ime. In he nex period (1986-2010), class 438 acually emerged in Alana, which follows he expecaion ha he densiy of relaed echnologies shapes echnological change in ciies. As figure 2 shows, a srong and posiive relaionship seems o exis beween relaedness densiy and he emergence of new echnologies in ciies. To draw figure 2, we ploed he average densiy values and he rae of enry of new echnologies for each of he 366 ciies. The rae of enry is given by he sum of enries of new echnologies in a ciy from 1986 o 2010, divided by oal number of possible enries during his period. For insance, one of he meropolian area wih he highes rae of enry is he Greaer Boson area 8. On average, a each of he six 5-year periods, 20 new echnologies enered, while 80 could have enered. This indicaes a probabiliy of enry of abou 25% while he average rae of enry in American ciies is abou 9% 9. The average densiy around hese poenial echnologies was also abou 25%. Figure 2 shows ha here is a very high posiive correlaion beween he level of relaedness densiy and he probabiliy of enry. Ciies ha have a more diverse echnological porfolio and ha have compeences in core echnologies (i.e. a he core raher han a he periphery of he echnological space displayed in figure 1) seems o renew hemselves more quickly over ime. Bu of course, he Greaer Boson area is a rich meropolian area, which hoss very producive research universiies, and also scores high in erm of human capial (Glaeser, 2005). Therefore, o es he relaionship beween relaedness and echnological change a he ciy level, one has o conrol for ciy and echnology ime varian and ime invarian characerisics. This is he purpose of he economeric framework presened furher below. 11
Figure 2. Relaedness and echnological change in U.S. ciies (1981-2010) Noes: Each do represens one of he 366 U.S. ciies (MSA). The rae of enry represens he number of new echnologies ha enered a ciy s echnology space divided by he oal number of possible enries. Average densiy is he average percenage of relaed echnologies in he ciy. 4.3. Economeric specificaions We wan o esimae how relaedness influences echnological change a he ciy level. Therefore, we regress he emergence of new echnologies on heir degree of relaedness wih he echnological porfolio of ciies, which is capured by he relaedness densiy variable. The basic economeric equaion o be esimaed can be wrien as follows: Enry i, c, = β 1Densiyi, c, 1 + β 2Ciyc, 1 + β3technoi, 1 + φc + ψ i + α + ε i, c, (4) where he dependen variable Enry 1 if a echnology i ha did no belong o he i, c, = echnological porfolio of ciy c in ime -1 eners is echnological porfolio in ime, and 0 oherwise. The key explanaory variable Densiy i, c, 1 indicaes how relaed he poenial new echnology i is o he pre-exising echnological se of capabiliies of ciy c. Ciy is a vecor ha summarizes a range of observable ime-varying ciy characerisics 10. c, 1 We consruced variables like he number of employees in a ciy (employmen), he number of inhabians by square meers (populaion densiy) 11, he raio of invenors o employees (invenive capaciy), he growh rae of he number of invenors (MSA echnological growh rae), and he economic wealh of a ciy (income per employee) 12. The variable echnological specializaion of ciy c has been measured by he average locaion quoien weighed by he number of paens: 12
Pci Specializa ionc = LQci (5) i P c where Pci denoes he number of paens of ciy c in class i, Pc he oal number of paens of ciy c, and LQ he locaion quoien of echnology i in ciy c. Techno is a vecor ha summarizes a range of observable ime-varying echnology i, 1 characerisics. Firs, we ake he oal number of invenors (Nb. Invenors) compued a he echnology level o conrol for echnology size, as echnologies ha involve many invenors are more likely o ener any ciy by chance. We included wo variables o accoun for echnology age and he expansion/exracion of echnological opporuniies, by using he year in which a echnological class has been officially esablished by he USPTO (dae esablished), and he growh rae of he number of invenors paening in a given echnology class (Tech. Class growh rae). As concenraion of invenive aciviies could be more conducive o relaed echnological diversificaion, we also measured he average locaion quoien weighed by he number of paens (echnological concenraion): Pci Concenra ioni = LQci (6) c P i where Pci denoes he number of paens of class i in ciy c, Pi he oal number of paens of class i, and LQ he locaion quoien of echnology i in ciy c. Finally, φ c is a ciy fixed effec, ε is a regression residual. i, c, ψ i is a echnology fixed effec, α is a ime fixed effec, and Therefore, he baseline economeric model used is a hree-way fixed-effecs model, o ake ino accoun omied variable bias a he ciy and echnology levels, assuming ha hese omied variables are consan over ime. We esimae equaion (4) by using a linear probabiliy (OLS) regression 13. φ c, ψ i and α fixed effecs are direcly esimaed by including dummy variables for each ciy, echnology and ime period ha compose our panel. As exensively discussed by Wooldridge (2003) and Cameron e al. (2011) sandard errors should be adjused for clusering when he errors are correlaed wihin groups of observaions, such as ciies and echnologies in our case. Therefore, all he regression resuls presened in his paper are clusered a he ciy and echnology level 14. Our panel consiss of daa for 366 ciies (MSAs), 438 echnologies (paen echnological classes a he 3-digis level) over he period 1976-2010. We average he daa 15 over non-overlapping five-year periods (1976-1980,..., 2006-2010), denoed by. To avoid poenial endogeneiy issues, all he independen variables are lagged by one period 16, so ha we have 6 observaions per ciy-echnology pair, resuling in a balanced panel wih 961,848 observaions 17. Table 2 provides some summary saisics of he variables used in he economeric analysis 18. 13
Table 2. Summary saisics Variables Obs Mean Sd. Dev. Min Max Enry 748,458.092585.2898502 0 1 Exi 213,390.3654717.4815633 0 1 Relaedness densiy [Hidalgo e al.] 961,848 21.58084 29.81881 0 100 Relaedness densiy [Co-occurrence] 961,848 32.84199 28.20259 0 100 Relaedness densiy [Hall e al.] 961,848 29.96201 28.47284 0 100 Employmen [ciy] 961,848 269896 648163.6 2630.2 8538557 Populaion densiy [ciy] 961,848 239.9645 289.9026 3.59798 2790.44 Invenive capaciy [ciy] 961,848.19754.21342 0 3.5763 Technological specializaion [ciy] 913,668 15.80558 13.30246 1.30188 63.62291 MSA echnological growh rae [ciy] 904,376.0794336.3963024 -.8743017.962963 Income per employee [ciy] 956,592 29.05074 10.86576 10.374 93.57 Nb. Invenors [echno] 961,848 1542.409 2484.688 0 27984 Technological concenraion [echno] 913,902 9.756013 11.46718 1.388236 68.73915 Dae esablished [echno] 961,848 1955.249 32.97904 1899 2009 Tech. Class growh rae [echno] 896,700.0948035.4865253 -.9944946 2.333333 Noe: In he economeric esimaions presened in he paper, employmen, income per employee and Nb. Invenors have been log-ransformed. 5. Empirical resuls In his secion we presen he economeric resuls of he impac of relaedness a he ciy level on echnological change in U.S. ciies from 1981 o 2010. We analyze he probabiliy of enry, bu also he probabiliy of exi of paen echnology classes in meropolian areas. 5.1. Do ciies diversify ino relaed echnologies? Table 3 presens he resuls for he esimaion of equaion 4. The baseline model (model 1) regresses he enry of a given echnology in a given ciy on he densiy of links around his echnology in his ciy (lagged by one period). Column 1 presens esimaion resuls from pooled OLS 19, while column 5 provides coefficien esimaes from he hree-way fixed effecs model (F.E.) wih all he ciy and echnology variables. In all he differen specificaions, relaedness densiy has a posiive and significan effec. I indicaes ha relaedness densiy has been a crucial driving force behind echnological change in U.S. ciies for he las 30 years. Relaedness densiy is no only saisically, bu also economically significan. If he level of densiy for a given echnology in a given ciy increases by 10 percen 20, he probabiliy of enry of his echnology in his ciy during he nex period increases by abou 55 percen (0.051/0.092) in he simples specificaion (Table 3; column 1). The economic impac of relaedness densiy remains sable across he differen economeric specificaions. 14
Dependen variable is: Enry Relaedness densiy -1 Table 3. Emergence of new echnologies in U.S. ciies (1981-2010) Model 1 Rel. densiy Model 2 Ciy variables Model 3 Tech. variables Model 4 Full model Model 5 Full model (F.E.) 0.00515979 ** 0.00373407 ** 0.00271463 ** (0.00012770) (0.00014135) (0.00016884) Log (Employmen) -1 Populaion densiy -1 Invenive capaciy -1 Tech. Specializaion -1 MSA growh rae -1 Log (Income per employee) -1 Log (Nb. Invenors) -1 Tech. concenraion -1 Dae esablished -1 Tech. growh rae -1 Consan 0.04934166 ** 0.03611889 ** 0.04633250 ** (0.00286818) (0.00247147) (0.00782869) 0.00001106 0.00002520 ** -0.00021341 ** (0.00000997) (0.00000843) (0.00003836) 0.07718815 ** 0.03883926 ** -0.08487966 ** (0.01294204) (0.0078352020) (0.01505564) -0.00089296 ** -0.00047160 ** 0.00005120 (0.00011548) (0.00009315) (0.00011022) 0.04443962 ** 0.04032813 ** 0.00865397 ** (0.00355534) (0.00353667) (0.00298386) -0.07584685 ** -0.10127439 ** 0.00368879 (0.00441610) (0.00538561) (0.01663469) 0.02658895 ** 0.02324554 ** 0.00159990 (0.00197752) (0.00183672) (0.00246612) -0.00102840 ** -0.00010693 0.00041990 * (0.00014936) (0.00011541) (0.00016760) -0.00056684 ** -0.00042520 ** -0.00330620 ** (0.00007012) (0.00005456) (0.00017699) 0.01423964 ** 0.02183910 ** 0.01141729 ** (0.00233334) (0.00285492) (0.00260757) 0.09258502 ** 0.09296771 ** 0.09019069 ** 0.08909252 ** 0.11108572 ** (0.00194271) (0.00378306) (0.00398429) (0.00183778) (0.01040890) Ciy F.E. No No No No Yes Technology F.E. No No No No Yes Period F.E. No No No No Yes R 2 0.11 0.04 0.02 0.13 0.16 N 748,458 653,660 656,618 572,550 572,550 Noes: The dependen variable enry = 1 if a given echnology (n = 438) eners in he echnological porfolio of a given ciy (n = 366) during he corresponding 5-years window (n = 6), and 0 oherwise. The independen variables are cenered around heir mean. Coefficiens are saisically significan a he p <0.05; and p <0.01 level. Heeroskedasiciy-robus sandard errors (clusered a he ciy and echnology level) in parenheses. In order o verify ha our resuls are no affeced by omied variables bias, we conrol for imporan ciy and echnology characerisics. A second model (model 2), repored in column 2 includes variables ha capure he heerogeneiy of ciies and ha migh influence echnological change. As expeced, he economic size of ciies (employmen), he raio of invenors o employees (invenive capaciy) and he growh rae of he number of invenors (MSA echnological growh rae) play a posiive and significan role on he enry of new echnologies. Populaion densiy has also a posiive impac, bu he coefficien is no saisically significan. Our resuls also confirm he idea ha ciies characerized by a very specialized echnological srucure (echnological specializaion) are less prone o echnological change. A more counerinuiive finding, however, is he negaive role played by he economic wealh of he ciy (income per employee). I migh be explained by he fac ha once one included he variables discussed above, income per employee does no reflec he invenive capaciy of ciies. A hird model (model 3), repored in column 3 includes variables ha capure he heerogeneiy of echnological classes ha migh influence heir general enry in ciies. I is no surprising ha 15
large echnological classes, i.e. wih a large pool of invenors are more likely o ener in any U.S. meropolian area (Nb. Invenors), especially if he producion of knowledge in his echnological class is growing (Tech. Class growh rae). On he conrary, older echnologies (dae esablished) and echnologies ha are very much concenraed in space are significanly less likely o be developed by many differen ciies in he fuure. Overall, almos all he variables explain an imporan par of he variaion in erms of enry of new echnologies, and herefore hey are imporan predicors of echnological change. In he full model specificaion (model 4) we esed wheher he effec of relaedness densiy was affeced by hese imporan feaures of echnologies and ciies. Column 4 presens esimaion resuls from a pooled OLS, while column 5 presens he complee esimaions resuls from equaion 4, i.e. including relaedness densiy, ciy and echnology ime-varying variables, and fixed effecs for ciies, echnologies and ime. Ineresingly, he coefficien for densiy remains highly significan, bu is magniude slighly decreases wih he addiion of hese conrol variables. When fixed effecs are included, he rae of enry increases by approximaely 30 percen for a 10 percen increase in he level of densiy in ciy-echnology pairs 21 (0.027/0.089). 5.2. Robusness analysis In Table 4, we presen alernaive economeric specificaions o es he robusness of he relaionship of ineres, i.e. he effec of relaedness densiy on echnological change in ciies. We run hree differen ses of robusness checks: (i) using alernaive measures of echnological relaedness as independen variables, (ii) excluding observaions wih exreme values (i.e. ouliers), and (iii) using alernaive economeric mehods o he linear probabiliy model. The resuls repored 22 in Table 4 shows ha he posiive and significan impac of relaedness densiy on he probabiliy of enry is robus o hese alernaive specificaions. Firs, we verify ha our resuls are no driven by he echnological relaedness measure we used (Table 4.; col. 1 o col. 2). Therefore, we esimaed equaion 4 (see specificaion in Table. 3; col. 5) by using he Hall e al. (2001) paen classificaion, and a normalized co-occurrence analysis. The coefficien 23 on relaedness densiy is smaller in hose alernaive specificaions, bu remains saisically and economically significan. When densiy increases by 10 percen, he probabiliy of enry raises by approximaely 15 percen using he Hall e al. (2001) classificaion, and 20 percen using co-occurrences analysis. 16
Table 4. Enry of new echnologies in U.S. ciies - Robusness check Dependen variable is: Enry Alernaive relaedness measures Model COOC [Fixed Effecs] Model USPTO [Fixed Effecs] w/o op densiy # Ouliers analysis w/o op ciies w/o op echno GLM specificaions Logisic regression Probi regression Densiy -1 0.00224635 ** 0.00264742 ** 0.00239119 ** 0.0216442 ** 0.0125646 ** [baseline] (0.00016733) (0.00018286) (0.00017202) (0.0002433) (0.0001334) Densiy -1 0.00184525 ** [COOC] (0.00016940) Densiy -1 0.00142651 ** [USPTO] (0.00014815) Ciy conrols Yes Yes Yes Yes Yes Yes Yes Tech. conrols Yes Yes Yes Yes Yes Yes Yes Ciy F.E. Yes Yes Yes Yes Yes No No Technology F.E. Yes Yes Yes Yes Yes No No Period F.E. Yes Yes Yes Yes Yes Yes Yes R 2 /Pseudo R 2 0.15 0.15 0.11 0.15 0.14 0.19 0.19 N 572,550 572,550 495,077 515,350 514,091 572,550 572,550 Noes: The dependen variable enry = 1 if a given echnology (n = 438) eners in he echnological porfolio of a given ciy (n = 366) during he corresponding 5-years window (n = 6), and 0 oherwise. Coefficiens are saisically significan a he p < 0.05; and p <0.01 level. Heeroskedasiciy-robus sandard errors (clusered a he ciy-echnology level for he logisic and probi regression; clusered a he ciy and echnology level in all oher regressions) in parenheses. # The op 10% of he ciy-echnology pairs wih he highes densiy are dropped. The op 10% of he ciies ha experienced he highes number of echnology enry are dropped. The op 10% of he echnologies ha enered ciies he mos frequenly are dropped. Second, we check ha our resuls were no driven by exreme values a he op decile level (ab 4.; col. 3 o col. 5). Using our baseline measure of echnological relaedness (Hidalgo e al., 2007) we esimaed equaion 4 (see specificaion in ab. 3; col. 5) by removing he op 10% of he ciy-echnology pairs wih he highes densiy (ab 4.; col. 3), by removing he op 10% of he ciies ha experienced he highes number of echnology enry (ab 4.; col. 4) and finally by removing he op 10% of he echnologies ha enered ciies he mos frequenly (ab 4.; col. 5). None of hese alernaive specificaions wih alered daa samples seem o subsanially affec he saisical or economic effec of relaedness densiy. Third, we esimaed equaion 4 wih generalized linear models, i.e. logi and probi (ab. 4; col. 6 and col. 7). In he paper, we focused on linear probabiliy models bu since he dependen variable (enry) is dichoomous, we also check ha our resuls are robus o radiional GLM specificaions. The las wo columns in Table 4 show resuls from logisic and probi regressions and hey confirm he posiive and significan impac of densiy of relaed echnologies on he probabiliy of enry. The alernaive measures of echnological relaedness, bu also esimaes from daa samples wihou exreme values and alernaive economeric models all suppor our key findings. On op of ha we also esimaed models by using densiy from weighed relaedness marices, by using all classes lised on paen documens o consruc he relaedness space and he corresponding enry variables, by using invenor shares o localize paens insead of he primary invenor, and by consraining he enry evens o echnologies in which ciies have a comparaive advanage 24. 17
These addiional analyses do no affec he resuls presened here and sugges ha our saisically and economically significan posiive relaionship beween relaedness densiy and he probabiliy of enry is robus o several key economeric specificaions. 5.3. Does relaedness densiy preven he exi of echnologies? Bu echnological change is no only abou enry of new echnologies wihin ciies. In fac, echnological change can be undersood as a process of creaive desrucion in which he exi of exising echnologies also conribues o change he echnological landscape of ciies. Table 5 repors esimaion resuls where he dependen variable "exi" is used insead of he dependen variable "enry". The resuls indicae ha relaedness densiy has a negaive and significan impac on he exi of echnologies. If he level of densiy for a given echnology in a given ciy increases by 10 percen, he probabiliy of exi of his echnology in his ciy during he nex period decreases by abou 8 o 17 percen, depending on he economeric specificaions. The resuls concerning ciy and echnology characerisics are also consisen wih our expecaions. Ciies wih a high economic poenial are more likely o preven he exi of echnologies, while economically imporan echnologies are less likely o exi in all he ciies. Wha should be noiced, however, is ha he relaive economic imporance of relaedness densiy compared o ciy and echnology characerisics seems o be smaller o explain variaions in he exi han he enry of echnologies. 18
Table 5. Exi of echnologies in U.S. ciies (1981-2010) Dependen variable is: Exi Relaedness densiy -1 Log (Employmen) -1 Populaion densiy -1 Invenive capaciy -1 Tech. Specializaion -1 MSA growh rae -1 Log (Income per employee) -1 Log (Nb. Invenors) -1 Tech. concenraion -1 Dae esablished -1 Tech. growh rae -1 Model 1 Rel. densiy Model 2 Ciy variables Model 3 Tech. variables Model 4 Full model Model 5 Full model (F.E.) -0.00646272 ** -0.00384300 ** -0.00287999 ** (0.00013398) (0.00022311) (0.00021200) -0.10857437 ** -0.06943327 ** -0.08359852 ** (0.00614202) (0.00626204) (0.01651044) -0.00003837 * -0.00006553 ** -0.00011335 * (0.00001950) (0.00001364) (0.00004718) -0.16931248 ** -0.11188970 ** -0.02739567 ** (.05336078) (.02941733) (.00841076) 0.00437970 ** 0.00180088 ** -0.00056492 (0.00061919) (0.00040634) (0.00042826) -0.16187457 ** -0.15036339 ** -0.01352593 (0.00828661) (0.00882790) (0.00966971) 0.22689471 ** 0.31767021 ** 0.04962913 (0.01306891) (0.01236049) (0.03082188) -0.04660531 ** -0.09098814 ** -0.05541312 ** (0.00593058) (0.00406299) (0.00624571) 0.00418752 ** -0.00137922 ** -0.00200006 ** (0.00047497) (0.00043938) (0.00058743) -0.00018739 0.00022470 * 0.00233776 ** (0.00011545) (0.00010297) (0.00022809) -0.06741134 ** -0.05102216 ** -0.01451667 * (0.00590281) (0.00701074) (0.00652948) Consan 0.36547167 ** 0.36534949 ** 0.36590647 ** 0.36470402 ** 0.54798934 ** (0.00609779) (0.00841466) (0.01460965) (0.00426248) (0.03747886) Ciy F.E. No No No No Yes Technology F.E. No No No No Yes Period F.E. No No No No Yes R 2 0.19 0.15 0.03 0.25 0.30 N 213,390 202,584 201,286 191,313 191,313 Noes: The dependen variable enry = 1 if a given echnology (n = 438) exis he echnological porfolio of a given ciy (n = 366) during he corresponding 5-years window (n = 6), and 0 oherwise. The independen variables are cenered around heir mean. Coefficiens are saisically significan a he p < 0.05; and p <0.01 level. Heeroskedasiciy-robus sandard errors (clusered a he ciy and echnology level) in parenheses. 6. Discussion and Concluding Remarks In his paper, we found evidence ha he rise and fall of echnological knowledge, as proxied by he enry and exi of paen echnology classes in ciies, is condiioned by he exising echnological knowledge base of ciies. Analyzing he long-erm evoluion of paen porfolios of 366 U.S. ciies during he period 1976-2010, we found ha a new echnology is more likely o ener a ciy when echnologically relaed o oher echnologies in ha ciy, and an exising echnology had a higher probabiliy o exi a ciy when i was no, or poorly relaed, o oher echnologies in ha ciy. These resuls indicae ha echnological relaedness was a driving force behind echnological change in U.S. ciies in he las 30 years, and ha he long-erm evoluion of he echnological urban landscape is subjec o pah dependency. These findings call for furher research. As new echnologies emerge sysemaically in ciies wih relaed echnologies, his suggess ha new echnologies are all recombinaions of exising echnologies. While his migh be rue for a large fracion of new echnologies, i is no 19
necessarily rue for all of hem. In fac, some new echnologies (paens) are ruly novel, wih few or no relaed echnologies on which hese buil during heir ime of emergence (Dahlin and Behrens, 2005; Casaldi and Los, 2007; Kraff e al., 2011). From a geographical perspecive, i would be ineresing o invesigae where radical echnologies and new echnological rajecories come ino being. For insance, do hese need highly diversified ciies, insead of echnologically specialized ciies (Duranon and Puga, 2000)? Anoher issue o be aken up in fuure research is wheher new paens acually build on relaed paens in ciies. By looking a he se of (relaed) paens a he ciy level, we did no invesigae he exen o which a new paen ha is new for a ciy acually cies oher paens in relaed echnology classes from he same ciy. This would provide evidence a he level of paens (raher han a he level of ciies) ha invenion aciviy acually builds on relaed knowledge a he ciy level. This would also shed ligh on he imporance of knowledge flows from oher ciies, as paens migh draw on and cie relaed paens from oher ciies. Anoher issue is he selecion of he relaedness indicaor o sudy urban diversificaion. We made use of cooccurrence analysis based on he frequency of combinaions of paen classes wihin he same ciies. Oher scholars like Leen e al. (2007), and Van der Wouden (2012) and Rigby (2013) have used alernaive indicaors o measure relaedness beween paen classes, such as paen ciaions. However, recen sudies like Rigby (2013) have shown ha findings are unlikely o change when using such alernaive measures of echnological relaedness. As a robusness check, we made use of wo alernaive measures of relaedness (i.e. normalized co-occurrences and paen classes belonging o he same 2 digi caegory), and our findings wih respec o relaedness densiy remained saisically significan. Oher sudies on regional diversificaion have used oher measures of relaedness, based on he inensiy of inpu-oupu linkages beween indusries (Esslezbichler 2013), or based on co-occurrence analysis of produc caegories eiher wihin plans (Neffke e al., 2011) or wihin counries (Boschma e al., 2013). Alhough hese sudies are very differen in erms of heir relaedness measure, he use of spaial unis and mehodologies, he ime period covered, and he selecion/measuremen of he dependen and independen variables, hey also found evidence of relaedness a he regional scale driving regional diversificaion. A final issue deserves aenion in fuure research. In his paper, we explored he exen o which he enry of a new echnology depends on oher echnologies o which i is relaed a he ciy level. However, we did no explore oher dimensions ha migh be crucial in he process of echnological diversificaion, such as insiuional precondiions a he ciy level (Srambach, 2010). In fac, such a sudy would shed ligh on he exen o which relaed echnologies draw on and require similar ses of urban insiuions, which could provide an addiional explanaion for he fac ha relaed echnologies end o benefi from each oher s co-presence in ciies. 20