Home » 2014 Categories » 2014 Technical Presentation » Innovation in cities: Science-based diversity, specialization, and localized competition

Innovation in cities: Science-based diversity, specialization, and localized competition

By Olivia Nicolaus


This paper by Feldman and Audretsch attempts to address the question of “whether diversity or specialization of economic activity better promotes technological change and subsequent economic growth” (409).  It finds considerable support for diversity as a catalyst for innovation and little support for specialization (409)


A number of scholars including Krugman, Romer, and Lucas support the importance of concentration of people and firms as the most important factor for economic activity (410).  Concentration creates knowledge spillovers, which are the transmission of knowledge “through face-to-face interaction and through frequent contact” (411).  Scholars disagree over the significance of knowledge spillovers within and across disciplines, but it is widely accepted that physical proximity is key for the transmission of “sticky knowledge,” or that which is highly contextual (411).  Knowledge spillovers create increasing returns to scale within a geographically bounded space, primarily the relatively compact area of cities (410).

An important question to ask in relation to agglomeration economies is “does the specific type of economic activity undertaken within any particular geographic region matter?” (410). This opens the debate to two options: a geographic region that specializes in a particular industry, or a geographic region with diverse firms and economic agents.  In order to answer this question, Feldman and Audretsch attempt in this paper to classify the extent of diversity or specialization in geographic regions and then measure “how this composition influence innovation output” (410).

Connecting Innovation and Cities

In this paper Feldman and Audretsch use data from the United States Small Business Administration Data Base as a direct measure of innovative output.  This database is composed of product innovations, each with a four-digit standard industrial code (SIC).  Limitations to this data include the emphasis of product innovations over process innovations, variation in the quality of innovations, and the necessity to treat all innovations as homogenous (414).

By attributing each SIC to a Consolidated Metropolitan Statistical Area or Metropolitan Statistical Area, the researchers are able to rank cities in terms of gross quantity of innovation.  The results are exhibited in Table 1, which shows that the most innovative city in the United States in 1982 was New York.  It is also important to note the overwhelming source of innovation is urban areas, with less than 4% of all observed innovations occurring outside of metropolitan areas (415).  For reference, 70% of the population at this time lived outside of metropolitan areas (415).  The table also uses population statistics to provide a more accurate calculation of innovation, finding San Francisco with the highest innovation rate per capita (415).

Connecting industry clusters, academic departments, and geographical areas

Feldman and Audretsch then attempt to link “products on their closeness in technological space” (415).  To do so they utilize the relevance ranking scale in the Yale Survey of R&D managers to establish groupings between industries that share a common scientific base.  The results are shown in Table 2.

The researchers find that industries that rely on a “common science base” exhibit a tendency to cluster together geographically with regard to the location of employment and innovation (418).  This is the initial information that the researchers use to create a model for determining the effect of a variety of factors on the quantity of innovations in different locations.

Modeling Framework

Feldman and Audretsch establish the dependent variable of their analysis as the number of innovations attributed to a specific SIC industry in a particular city.  They isolate three explanatory variables: a measure of industry specialization, a measure for the presence of science-based related industries, and an index for localized competition (419).  The equation, mean, and standard deviation for these three variables are exhibited in Table 3.


Feldman and Audretsch’s results are shown in Table 4, titled Poisson estimation results for the Poisson regression estimation method.  This method was selected because to model count variable because “the dependent variable is a limited dependent variable with a highly skewed distribution” (420).  This means that the events represented by the data are somewhat rare.  This type of distribution can be used for cancer, cases, number of accidents, or number of bird sightings, but in this case is used for counts of product innovation (Schwartz 1466).

The first column (Model 1) provides results for the three independent variable measures (specialization, science-based related industries, localized competition).  For industry specialization, the negative and statistically significant coefficient suggests that cities that specialize in economic activity in a certain industry have a lower rate of innovative activity.  For science-based related industries, the positive and statistically significant coefficient means that innovative activity is correlated with a strong presence of complementary industries sharing a common science base.  Finally, the negative coefficient on the third variable, localized competition, suggests that innovative activity of an industry is actually associated with less localized competition.  To translate these correlations: the results provide support for diversity in spurring innovation as opposed to specialization spurring innovation (421).

Potential Concerns

There are a few potential drawbacks to using this model.  The first is of city size; that large cities might be expected to have more innovation purely as a result of advantages in total manpower and resources.  There may be a greater degree of economic activity and localized competition.  In the second column of Table 3 (Model 2), total employment is normalized and the results for the third variable change.  This new positive coefficient means that localized competition is, in fact, conducive to innovative activity.  The other two coefficients remain unchanged. Another concern with Model one is the variation in innovation across industries.  In the third and fourth columns (Models 3 and 4), the number of innovations recorded for the specific industry is controlled.  The basic results remain the same.

 Policy Implications and Importance

The answer to the debate of specialization versus diversity prompts two different policy implications.  If innovation is fostered more effectively in specialized economies, policymakers should “focus on developing a narrow set of economic activities within a geographic region” (410).  However, since the opposite is true and diversity prompts innovation, policymakers should attempt to “identify commonalities and foster diversity” within the geographic region (410).

The specialization versus diversity question draws parallels to two types of modern development: university research parks versus the traditional urban form.  According to this study, the diversity of work types that occurs in a traditional urban setting is more innovative and therefore more economically productive than a more focused research park.  If policymakers are purely hoping to pump out a vast quantity of innovation in the form of new products, they should focus primarily on developing a diversity of businesses and corporations in cities, and also figuring out ways to encourage face-to-face contact that is valuable to knowledge spillovers.

However, this research does not measure the quality of innovations, and thus should not be taken at face value.  Further research could incorporate the quality of innovations in the spatial analysis.  In addition, further research could measure how the degree of specialization within research parks affects the amount of innovation created.


Table 1: Counts of innovation normalized by population


Table 2: The Common science bases of industrial clusters


Table 3

Table 4: Poisson estimation results



Feldman, Maryann P., and David B. Audretsch. “Innovation in Cities:.” European Economic Review 43.2 (1999): 409-29. Print.

Schwartz, Carl J. “Poisson Regression.” Poisson Regression. Simon Frazer University, 7 June 2013. Web. 28 Mar. 2014. <http://people.stat.sfu.ca/~cschwarz/Stat-650/Notes/PDFbigbook-JMP/JMP-part025.pdf>.



Leave a comment

Your email address will not be published. Required fields are marked *