Location, Internationalization and Performance of Firms in Italy: a Multilevel Approach

Transcription

1 Location, Internationalization and Performance of Firms in Italy: a Multilevel Approach G.Giovannetti, G.Ricchiuti, M.Velucchi Firenze, Mar

2 The performance of firms in a globalized world depends on specific firms characteristics and on their flexibility to react to market changes. Successful competitors in an industry cluster in the same geographic areas Companies use advantages of localization to compete at a global level. Firms performance depend on firms specific characteristics ( such as propensity to export or technology level) context-related variables (such as infrastructures and commercial networks).

3 Multilevel Why Multilevel? The multilevel approach allows greater concordance between the theoretical views and the models employed for studying firms behavior; When complex structure of data exists, standard regression models (such as the Generalized Linear Models) are not adequate as they do not take into account the data (hidden) hierarchical structure. Standard regression models make unsuitable assumptions on the variance-covariance structure. They assume independence of the observations, while the performance of the firms working in the same province are likely to be positively correlated (Rabe-Hesketh and Skrondal, 2010); The standard one level approach to hierarchical data gives rise to biased estimates and standard errors (Aitkin and Longford, 1986; Burstein et al., 1978);

4 Multilevel Why Multilevel? A multilevel approach, instead, allows to take into account hierarchical levels in the data and obtain correct and efficient estimates (Maas and Hox, 2004; Snijders and Bosker, 1999), considering clustering as a characteristics of the data and not simply a temporary nuisance. Multilevel models assume a non homogeneous and not constant correlation structure at higher level.

5 Multilevel A simple multilevel model like y ij = α + βx ij + u i + e j where i : 1,..., n units are clustered in j : 1,..., k groups, the correlation between any two units i and j will be σ2 u corr(y ij, y i j ) = σu 2 + σe 2, i i thus allowing to better capture the variance of the system. A standard approach like OLS with clustered error, though considering that the correlation is not constant across the units, assumes an homogeneous correlation structure within each cluster. This gives biased and not consistent estimates in case of hidden hierarchical data for which the correlation structure is likely to vary across groups (clusters/levels). The consequence of the independence assumption is a poor quantification of uncertainty.

6 We aim at analyzing the performance of Italian firms separately accounting for the effect of individual and context characteristics. In our paper: individual characteristics are size, technology, R&D expenditures, internationalization mode context characteristics, at a province level, are infrastructures (seaports and airports) presence of industrial districts

7 The Model The multilevel approach allows to simultaneously model individual variables: X h,i,j h is the number of covariates and i is the firm in the j-th province context variables: Z k,j k is the number of covariates and j the province. The linear specification can be written as Y ij = α + r β h X hij + h=1 where i : 1,...n and j = 1,...p. s γ k Z kj + u j + e ij (1) u j N(0, τ 2 ) - the second level casual effects of the model- are the model s residuals for each province. Where u j and e ij : second and first level residuals, normally distributed. u j is the difference between the j-province and the total average. k=1

8 Estimation in three steps 1. Null Model: Y i,j = α + u j Where α is the average of the overall population, u j N(0, τ 2 ) is the deviation from the average for the j-th province. 2. L.R. test on τ 2 : if the null hypothesis (absence of a second level in the data) is rejected, then a territorial effect (at a province level) is present and a multilevel model is appropriate. 3. General Model estimation.

9 Match and Merge of AIDA, Capitalia (2003), ICE-Reprint: Capitalia only manufacturing all firms larger than 500 employees Stratified sample of firms with less than 500 employees AIDA (Bureau Van Dijk) Balance Sheet of all companies (sales>100,000 euro) manufacturing and service ICE-Reprint: manufacturing and service all affiliates > 2.5 mln euro sales non probabilistic sample of smaller firms (info from Camere di Commercio)

10 Also information on Infrastructures in Italy (ISTAT, 2005) Export levels (confidential ICE-ISTAT, 4-digits, by province) Technological Intensity (Pavitt taxonomy, Capitalia)

11 Info on Data Table: Descriptive Statistics Variables Average St. Dev. Min. Max Propensity to export (%) Average propensity to export (province, %) R&D on sales (%) Delocalization (%) Areas of Export per firm (number) Average Areas of Export per firm (province) Innovation (dummy) District (dummy) Seaport (dummy) Airport (dummy) Size (classes) Technological intensity (dummy) Note: number of observations included Size classes: class 1 (11-20); class 2 (21-50); class 3 (51-100); class 4 ( ), class 5 (> 250)

12 Step 1/2 Likelihood Ratio Test on Null Model: Rejection of the null presence of territorial effect (space for multilevel) Table: Likelihood Ratio Test Likelihood Ratio Test LR chi2(9) = p-value>0.001

13 The Whole Sample

14 Predictions and Residuals Based on the selected general model, we can derive the predicted propensity to export for each province and represent it on a map. Differences among provinces can also be analyzed by looking at the random effects (empirical Bayes residuals) of the model. These figures convey all the province-level factors that count. Residuals reveal an unexpected propensity to export, given the estimates of our model: Positive values: contextual factors improve firms propensity to export. Firms have higher propensity to export than predicted by the model including only context variables(red). Negative values: contextual factors reduce firms propensity to export. Firms have lower propensity to export than predicted by the model including only context variables(lilac). This suggests that firms in provinces with a favorable context (red) could take more advantage of it.

15 Propensity to export predicted by the selected model for each province (quantiles)

16 Bayes Empirical residuals (quantiles)

17 Multilevel Model used to capture the effect of individual and context related variables on firms propensity to export at a province level. Context variables (province related) strongly influence the firms propensity to export. Firms/province heterogeneity persists: Some firms in specific provinces succeed international markets even if their socio-economic context is not favorable. Some firms in specific provinces, instead, are not able to fully benefit from the dynamic, positive province externality. We run regressions on size sub-samples showing that small and large firms do not equally depend on the socio-economic context: small firms benefit from the social capital that spill over industrial districts while large firms propensity and performance strongly depend on their own investments in R&D and technology.