Sectoral shifts, diversification and regional unemployment: evidence from local labour systems in Italy

Abstract

Using Local Labour Systems (LLSs) data, this work aims at assessing the effects of sectoral shifts and industry specialization patterns on regional unemployment in Italy over the years 2004–2008. Italy represents an interesting case study because of the high degree of spatial heterogeneity in local labour market performance and the well-known North–South divide. Furthermore, the presence of strongly specialized LLSs (Industrial Districts, IDs) allows us to test whether IDs perform better than highly diversified urban areas thanks to the effect of agglomeration economies, or viceversa. Building on a semiparametric spatial auto-regressive framework, our empirical investigation documents that sectoral shifts and the degree of specialization exert a negative role on unemployment dynamics. By contrast, highly diversified areas turn out to be characterized by better labour market performances.

Appendix

Each univariate smooth term in models (1)–(3) can be represented as $f_{j} \left( {x_{j} } \right) = \sum\nolimits_{k = 1}^{{K_{j} }} {\beta_{jk} b_{jk} } \left( {x_{j} } \right)$, where the $b_{jk} \left( {x_{j} } \right)$ are known basis functions, while the $\beta_{jk}$ are unknown parameters to be estimated. One or more measures of ‘wiggliness’ $\beta_{j}^{'} {\mathbf{S}}_{j} \beta_{j}$, where ${\mathbf{S}}_{j}$ are positive semi-definite matrices, is associated with each smooth function. Typically, the wiggliness measure evaluates something like the univariate spline penalty $\int {f_{j}^{\prime \prime } \left( {x_{j} } \right)^{2} dx}$. The penalized spline base-learners can be extended to two or more dimensions so as to handle interactions. Specifically, smooth interaction terms are specified here using thin-plate regression splines (see Wood 2006).

Finally, it is important to mention that, like in parametric spatial lag models (SLM), also in semiparametric SLM the estimated coefficients of parametric terms as well as the estimated smooth effects of nonparametric terms cannot be interpreted as marginal impacts of the explanatory variables on the dependent variable due to the presence of a significant spatial autoregressive parameter $\left( \rho \right)$. For parametric SLM, LeSage and Pace (2009) have developed a method for correctly interpreting and summarizing these marginal effects. This method allows one to distinguish between direct and indirect effects, the sum of the two denoted as the total effect. In particular, they have shown that the average total effect (ATE) in the case of the SLM takes the simple form $ATE = \left( {1 - \hat{\rho }} \right)^{ - 1} \hat{\beta }$. Taking advantage of this result, we could apply a similar algorithm to compute the ATE of each univariate smooth term in the semiparametric SLM. Specifically, we have computed $\hat{f}_{j}^{ATE} \left( {x_{j} } \right) = \sum\nolimits_{k = 1}^{{K_{j} }} {\left( {1 - \hat{\rho }} \right)\hat{\beta }_{jk} b_{jk} } \left( {x_{j} } \right)$. However, such a transformation does not change at all the shape of the univariate terms plotted in the figures reported in the paper, which is exactly our main interest for the economic interpretation of the results. It only induces a change in the scale of the vertical axes (i.e. the scale of the smooth effect) in each plot (Table 6).

Table 6 Variables description and sources