Regional knowledge spillovers: a firm-based analysis of non-linear effects

ABSTRACT

This paper revisits the question of the role of knowledge externalities in firm productivity. It also addresses the overlooked issue of a plausible non-linear effect and differences among industries. Using a panel of Portuguese manufacturing firms, it finds that regional knowledge spillovers differ substantially across industries and they are non-linear, which is critical issue to promoting more assertive regional policies.

KEYWORDS:

• agglomeration economies
• regional knowledge spillovers
• total factor productivity
• low-, medium- and high-tech industries

JEL:

• D24
• L25
• O12
• R12

DISCLOSURE STATEMENT

No potential conflict of interest was reported by the authors.

ORCiD

Carlos Carreira http://orcid.org/0000-0002-4786-5605

Luís Lopes http://orcid.org/0000-0002-4846-3747

SUPPLEMENTAL DATA

Supplemental data for this article can be accessed https://doi.org/10.1080/00343404.2017.1360484.

Notes

1. As pointed out by Frenken et al. (2007, p. 686), it is important to distinguish diversity economies from urbanization economies. Diversity economies, i.e., variety of industries, are sometimes regarded as part of urbanization economies, i.e., variety of the local actors and infrastructure, as in Cingano and Schivardi (2004) and Martin et al. (2011), for example.

2. Other authors identify a fourth source: competition economies. However, in highly integrated markets such as those in European countries, the effect of local competition does not appear to be statistically significant. This is not so surprising since its marginal impact shrinks under high levels of competition (e.g., Cainelli et al., 2015; Carreira & Lopes, 2015; Martin et al., 2011).

3. It is assumed that each firm operates in a given industry j and is located in a given region r. Subscripts j and r are omitted to simplify the notation except when it causes ambiguity.

4. Since i-th firm employment is subtracted, the LOC variable is firm specific.

5. Saito and Gopinath (2009), Martin et al. (2011) and Combes and Gobillon (2015) use the log of the inverse Herfindahl–Hirschman index of regional employment shares of the industries different from j. However, the entropy index is a more standard measure of variety (Frenken et al., 2007). We have nonetheless estimated the model with the inverse Herfindahl–Hirschman index. The results do not significantly change (the results are available from authors upon request).

6. The knowledge produced by universities and other knowledge-generating organizations is also often measured by, inter al ia, the amount of money spent on R&D, the number of employees engaged in R&D activities, the number of students in higher degree establishments, the number of articles published in scientific and academic journals, or the number of patents (Audretsch et al., 2005; Baptista et al., 2011; Baptista & Mendonça, 2010; Cassia et al., 2009; Fritsch & Aamoucke, 2013; Fritsch & Slavtchev, 2007). However, most of these variables are highly correlated due to complementarity. Having a large number of students, for example, means a larger teaching staff and, therefore, a greater amount of R&D resources (Fritsch & Slavtchev, 2007). On the other hand, the number of R&D workers (or the R&D expenditures) is a more general and representative measure. We have nonetheless estimated the model with the number of students in higher degree establishments and the number of higher degree establishments, but the results were less significant (these results are available from authors upon request).

7. According to the European Monitoring Centre on Change, KIBS comprises the following CAE-rev2.1 divisions: CAE 72, computer and related activities; CAE 73, research and experimental development; and CAE 74, other business activities.

8. Some firms are the sole representative of their industry in their region. Thus, we added 1 to the specialization index to overcome the log-of-zero problem. Consequently, if , there are no localization economies. The entropy index does not require any transformation since it is already a weighted average of logs (i.e., ).

9. The non-linear effects could also be tested using quantile regression (e.g., Fritsch & Slavtchev, 2010). Because of the endogeneity problem in model (2), however, the ordinary quantile regression estimator is not a robust alternative. The instrumental variable quantile regression method is not efficient for a large number of endogenous variables either.

10. The U-test was performed using the Stata utest command (Lind & Mehlum, 2010).