Spatial agglomeration, innovation and firm survival for Italian manufacturing firms

ABSTRACT

The innovativeness of a firm not only improves its own survival chances but also can generate externalities on its neighbouring firms. We empirically examine the role of agglomeration economies in how innovativeness affects firm survival in southern Italy, using spatial weights to model spillovers. Spatial Durbin probit model estimates confirm that innovation is an important determinant of firm survival, not only for firms that are themselves innovative but also for those located close to other innovative firms. Adequate definitions of spatial scale and spatial weights are important. Spillover benefits are enhanced by agglomeration economies, but only at a very local scale.

KEYWORDS:

• firm survival
• spatial models
• innovation
• spillovers
• southern Italian SMEs

JEL:

• C21
• C41
• D22
• L20
• O3

ACKNOWLEDGEMENTS

The paper is dedicated to the fond and loving memory of Ornella Maietta, who died while the first draft of the paper was being prepared. The authors thank the editor-in-chief, two anonymous referees and the editorial board for their encouraging, constructive and challenging comments that helped us to improve the paper substantially. In particular, careful attention to survey sampling issues and reorganization of theory and empirical literature owes much to their helpful guidance. The paper has also benefited from comments made by the participants at the 60th Annual Conference of the Italian Economic Association (Società Italiana di Economia) in 2019. Together, the authors are grateful to Franco Giordano and Gianluigi Coppola for help with understanding the survey design, and Taps Maiti for discussions on subsampling and superpopulation models in survey sampling, as well as Roberto Basile, Sean Holly and Cristina Tealdi for many helpful comments and suggestions. The usual disclaimer applies.

DISCLOSURE STATEMENT

No potential conflict of interest was reported by the authors.

Notes

1 ICE is the Italian agency for international trade, which promotes internationalization of Italian firms.

2 One of the highest proportions in the whole of the European Union (European Commission, 2012).

3 Question: ‘Who were the principal partners with whom you implemented the innovation (max. three)?’

4 Question: ‘According to your innovation strategy for the future, what field of knowledge available in the University of Salerno do you consider of particular interest?’. Responses: chemistry, engineering, computer science, business, etc.

5 There were more patent activities in the chemistry department (11 patents out a total of 21 for the University of Salerno) and more contract research in the engineering and computer science departments, whereas spin-off creation was equally frequent in the chemistry and engineering departments (two out a total of six for the University of Salerno).

6 The Istituto Sperimentale per l’Orticoltura (Experimental Institute for Horticulture) is located in Pontecagnano Faiano, the Istituto Sperimentale per il Tabacco (Experimental Institute for the Cultivation and Transformation of Tobacco) is in Scafati, and the Science and Technology Park in Salerno.

7 For example, the spatial autoregressive (SAR) model has $\rho \ne 0$, but sets $\gamma =\lambda =0$. By contrast, the SLX model has $\gamma \ne 0$, but $\rho =\lambda =0$; and the spatial error model with $\lambda \ne 0$, but $\rho =\gamma =0$. One can have a combination of spatial effects as well. For example, the SDM sets $\lambda =0$, but allows $\rho$ and $\gamma$ to have non-zero effects.

8 The issue of selection bias in spatial modelling has attracted only limited attention, and specifically for spatial probit models there is no literature. We are grateful to an anonymous referee for encouraging us strongly to give this issue due consideration and rigorous analytical treatment.

9 We are grateful to the editor-in-chief and an anonymous referee for help with making sense of the scattered literature and developing a nuanced understanding of alternate methods, together with their underlying advantages and disadvantages.

10 We also attempted to estimate our models by ML, but faced prohibitively intensive computations given our large population (sample) size of 7248 firms and a relatively flat log-likelihood surface.

11 The log-likelihood and AIC are reported in Table 2 and Table A3 in Appendix A in the supplemental data online.

12 The distances were computed using the R package spdep, a collection of functions to create a spatial weights matrix (Bivand & Piras, 2015); in particular, we used the function dnearneigh, which identifies neighbours of data points by the Great Circle distance (km) between lower (zero) and upper band (10 and 50 km). For all the tests (LR, Geary, Moran and join count test for spatial association), we use routines in R.

13 The AIC is defined as 2k – 2L, where $k$ denotes the number of parameters in the model and L is the log-likelihood. It balances model complexity against model fit and is also related to the likelihood ratio test for nested models.