6. Location of CCIs: Innovation and Filière Behind the Scenes

https://​www.​theguardian.​com/​travel/​2011/​jul/​24/​france-mougins-art-picasso. Accessed on 20.09.2022.

Besides Eurostat, robustness checks have been performed using Cambridge Econometrics data (ARDECO) for regional employment. No major changes emerge.

To cope with this issue, a NUTS3 weight has also been considered: younger regions receive a premium compared to older ones, using the median age of the NUTS3 region r w.r.t. the median age of the NUTS2 region s containing it. The indicator would result as: \(Tech.\,retail_{r} = onlinesales_{s} *\left( {median\,age_{r \in s} /median\,age_{s} } \right)^{ - 1}\). Although this choice does not find unanimous support in the literature as older people have a larger purchasing capacity also online (Sorce et al. 2005) and they are becoming more and more engaged in e-commerce (Lian and Yen 2014), younger people still represent the most active category. Furthermore, the indicator aims at measuring the pervasiveness of technology in retail as a measure of acceptance by consumers, as most tech-oriented consumers are expected to influence the innovativeness of firms in this segment (Pantano and Di Pietro 2012). For this reason, the “premium” generated by a relatively younger society is coherent due to the kind of innovations that market-oriented sectors generate such as digital signage, mobile apps, and ubiquitous computing (Pantano 2014). The results, available upon request, do not substantially differ from those obtained using the “original” variable and, for this reason, they are not shown here.

Technically, in any classical regression model, the OVB results if the error term produces some noise that is correlated with the regressor of interest, creating possible misunderstandings. In formulas, considering a classical OLS regression: \({\varvec{y}} = X{\varvec{\beta}} + {\varvec{\varepsilon}}\), the OVB arises if one of the Gauss–Markov assumptions fails. More specifically, the assumption that does not hold in the presence of an OVB is the strict exogeneity: \(E[{\varvec{\varepsilon}}|X] = \varvec{0}\). This assumption is important because if it fails, the interpretation of both the sign and the magnitude of the results is potentially misleading.

Variables obtained through the patents in Orbis refer to 2016 as they are cumulated values. Sections 5.​1 and 5.​3 presented in detail the data collection process and the pros and cons of the dataset.

Though this procedure is controversial in literature (Freckleton 2002), residuals of a first-stage regression are often used for the purpose of controlling for unwanted effects in multivariable framework. In this way, the effect of education and wealth on technology adoption are “cleaned” and only the “true” adoption of technology in retail remains.

In any regression output that includes interaction terms, the coefficient of the interaction between the dummy and the variable of interest shows the difference in slope between the two groups (dummy = 0 vs. dummy = 1). Hence, the actual coefficient of the variable of interest in the group with dummy = 1 is the sum between the coefficient of the variable not interacted and the coefficient of interaction. In formulas, given a regression model \(y_{i} = \alpha + \beta x_{i} + \gamma d_{i} + \vartheta x_{i} * d_{i} + \varepsilon_{i}\), \(\beta\) represents the slope of the regression line if \(d_{i} = 0\) while \(\beta + \vartheta\) is the slope if \(d_{i} = 1\). However, the significance of the coefficient of the interaction refers to the difference, not to the magnitude.

As education levels are potentially correlated with population and GDP per capita, all typical of urban settings, an additional robustness check has been performed. All these three variables have been included one at time in the regressions (1)–(8). The results do not present major changes, except for some cases. First, in the Short filière group, GDP per capita undoes the effect of education levels. Basically, including education (but also population) and excluding GDP per capita, the level of education becomes significant. All the other trials do not lead to changes in these three variables.

AMEs are available upon request.

In general, there is not a universal rule for the interpretation of the VIFs. In any case, statistical analyses use as a rule of thumb VIF = 1: not correlated; 1 < VIF < 5: moderately correlated; VIF ≥ 5: highly correlated.