The co-location of innovation and production in clusters

ABSTRACT

This paper quantifies the extent of co-location of innovation and production for industry clusters with varying knowledge intensity. If input-output, knowledge, and skill linkages are interdependent and geographically bounded, then we would expect innovation and production to be co-located in regional clusters. However, theory predicts that the degree of agglomeration benefits associated with co-location may vary across economic activities with different knowledge intensity. Using data from the U.S. Cluster Mapping Project, I develop measures of the co-location of innovation and production for 27 industry clusters, examining patterns across regions and over time (1998–2015) in the United States. I find that there is a significant co-location of innovation and production for many clusters, especially for those with higher knowledge intensity. This paper focuses on the Information Technology and Analytical Instruments cluster and the Automotive cluster to illustrate the co-location measures and the micro-geography of innovation and production.

KEYWORDS:

• Industry clusters
• co-location of innovation and production
• economies of agglomeration
• Automotive cluster
• Information and Technology cluster
• micro-geography

JEL CODES:

• R12
• R30
• O30
• L6
• L8

Acknowledgments

Chris Liu (especially), Maryann Feldman, Christoph Grimpe, Avi Goldfarb, Rich Bryden, Christian Ketels, Samantha Zyontz, Sarah Winton and reviewers have contributed very helpful suggestions. I thank Rich Bryden for great assistance with the data visualization. I am grateful for insightful comments by participants at Industry Studies 2017 Conference, the Microeconomics of Competitiveness Faculty Workshop at Harvard Business School, DRUID 2016 Conference, and the 4^th Global Conference on Economic Geography at the University of Oxford. Author contact information: Mercedes Delgado (Copenhagen Business School and MIT Innovation Initiative; [email protected]).

Disclosure statement

No potential conflict of interest was reported by the author.

Notes

¹ Delgado, Porter, and Stern (2016) offer a literature review on how to define clusters and detailed explanations of the BCD. The BCD is available at the USCMP website at http://www.clustermapping.us/content/cluster-mapping-methodology.

² Silverman (2002) assigns the patents to the set of industries whose goods and services are closer to the patent technology class. I use a revised version of Silverman’s (2002) algorithm developed by Delgado and Mills (2018).

³ The employment specialisation of a regional cluster captures various types of externalities (input–output, skills, knowledge, etc.). The patenting specialisation captures mainly knowledge externalities (Delgado, Porter, and Stern 2014).

⁴ A concern would be that employment may decline due to automation. One advantage of using LQ_Emp is that its distribution across regions for a cluster (say Automotive) would not be affected by a decline in the US-cluster employment per se. If regional clusters are using similar automation practices, the LQ_Emp distribution won’t change much. However, regions could vary in their degree of automation. If a given region uses new automation practices that significantly reduce the employment in the regional cluster, then LQ_Emp may decrease even if production has not changed. Since our co-location measures use many regions, it is less likely they will be affected by these cases.

⁵ In other words, we keep regions with low employment (below percentile 25^th value), but meaningful patenting (patent>10) and vice versa. Note that the 25^th percentile value is computed across region-clusters with non-zero employment.

⁶ The strong clusters are defined as those in regions (EAs or MAs) with high employment specialisation in the cluster in a particular year. They meet these criteria: Location Quotient (LQ) of cluster employment must be greater than the 75^th percentile (measured across all regions with non-zero employment in the cluster) and above one. To differentiate marginal cases in small regions, the shares of U.S. cluster employment and establishments must be greater than the 25^th percentile values.

⁷ I exclude micropolitan and rural areas because they account for a very small share of employment and patents.

⁸ These five clusters have low co-location scores also for the alternative measures reported on Table A4.

⁹ I use Hecker’s (2005) categorisation of STEM occupations and the BLS data (2009) to compute the STEM Intensity of U.S. clusters.

¹⁰ This cluster ranks 1–2 for DSC and DSP each year.