Page - (000303) - in Knowledge and Networks

300 Data and Estimation Framework We used information on the technological content of patent documents to measure the ability of cities to renew and expand their knowledge base. Patents can be described as a bundle of different technologies, whereby each is identified by the technology codes classification used at the U.S. Patent and Trademark Office (Strumsky, Lobo, & van der Leeuw, 2012). The technology fields and their combi- nations found in locally produced patents can therefore provide a good description of a city’s knowledge base (Boschma, Balland, & Kogler, 2015; Kogler, Rigby, & Tucker, 2013). Importantly, this approach allows emphasis to be placed on the intrinsic recombinatorial nature of technical change and inventive processes (Fleming, 2001; Katila & Ahuja, 2002). We defined our dependent variable as the number of new pairs of technology codes (i.e., combinations) introduced in a city at time t, in which both technology codes are new to the city, in other words, no local patent had been classified in those fields before time t. Differently from other studies (e.g., Fleming et al., 2007), new combinations of technology codes previously used in a city’s patents have been excluded, because they simply recombine existing knowledge, with no renewal or expansion of a city’s knowledge base resulting. Still, our estimates are robust to alternative (less restrictive) measuring of the dependent variable (not shown for the sake of brevity). In order to compute the number of new combinations in each city, three filters have been applied. First, only cities showing persistent inventive activity (i.e., with a positive number of patents for each year in the period 1990–2004; i.e., 196 out of 370 cities) have been considered, so as to mitigate erratic patterns that may arise on a short time basis because of annual fluctuations and lumpiness in patent records. Second, new combinations have been identified at the technology group level, as it corresponds to the lowest hierarchical level in the International Patent Classification (IPC, 2014) adopted at the EPO.6 Importantly, the number of technology groups per patent class exhibits an extremely skewed distribution: 20 % of patents are classified in only one IPC technology group and, therefore, cannot lead to any new recombi- nation (i.e., they do not enter in the computation of the dependent variable); 95 % of all patents in the sample are classified in eight or less technology groups, and 99 % in fifteen or less groups, with the remaining 1 % of all patents being outliers classi- fied in a number of technology groups ranging from 15 to 63. It is quite obvious that the higher the number of technology groups the greater the number of new combi- nations. In order to mitigate the bias in the computation of the dependent variable due to the presence of such extreme observations, its construction is based on pat- ents classified in up to eight groups.7 6 Groups are next divided in subgroups; however, subgroups are nested into groups (i.e., their hier- archical level varies across groups) and therefore cannot be exploited in this study. Further details are available online, see IPC (2014). 7 The number of total patents in the sample is 504400. S. Breschi and C. Lenzi