Are innovative regions more resilient? Evidence from Europe in 2008–2016

Abstract

This paper studies the relationship between the innovation performance of European regions and their resilience. By exploiting a novel dataset that includes patents and trademarks at the regional (NUTS2) level for the 2008–2016 period, the paper addresses two research questions: (1) are innovative regions more resilient? (2) which type of innovation is more conducive to resilience? We frame the relationship between resilience and innovation within the Schumpeterian notion of innovation as a ‘creative response in history’. Overall, we find that a stronger performance in innovation is associated with a better performance in employment both during and in the aftermarket of the 2008 financial crisis. We argue that learning capabilities built over time by regions make them more effective in adapting and recovering during major shocks. While the crisis may have created an opportunity for less developed regions to move ahead, this opportunity has in fact been grasped mainly by those already having a strong regional system of innovation in place.

Appendices

Appendix 1

See Tables 4 and 5.

Table 4 Correlation table and descriptive statistics

Table 5 Countries involved in the study

Appendix 2: Controlling for spatial correlation

A common feature of regional performance is the presence of spillovers, mostly occurring among neighbour regions.³ Spillovers can arise from various factors, such as knowledge flows, inter-regional trade and other linkages among the different regional economic systems. Hence, the performance of a region is likely to be affected by that of neighbour regions. It can also be affected by regions that are further away as for instance through trade, the operations of multinational corporations or international collaborations. The literature has consistently found that the effect of spillover tends to decay with distance, and therefore proximity matters (Jaffe et al. 1993; Iammarino and McCann 2006). Actually, Bottazzi and Peri (2003) suggest that due to the tacit nature of knowledge spillover can be very much localized, according to their estimations, on average, spillover operates on a range of about 300 km.

As a result, we expect that the resilience of a region—hence both SI and RI—isaffected by the resilience of the other regions, with a positive effect that is maximum for continuous regions and decreases with distance.

To identify clusters of high or low resilience we have carried out a Local Indicator of Spatial Association (LISA). LISA allows to assess the similarity of each observation (region) with that of its surroundings. In this way we can identify patterns of spatial clustering for the resilience values.

The LISA identifies the basic regional patterns both for the Sensitivity Index and the Response Index. In Fig. 4, we colour only the values with a significance level of 0.05. It is possible to note that some regions in United Kingdom, Germany and Austria (high resilience) show highly significant local spatial correlations, as well as Greece, Bulgaria and Romania (low resilience).

Fig. 4

LISA map by SI (left map) and RI (right map) for the European regions. Source: Authors’ elaboration. Regions are split in four clusters (quintiles, 25%); a darker blue indicates a stronger positive effect of local spatial correlation. In white the no statistical significant regions and missing data

As Cainelli et al. (2019b) suggest, there are several spatial models available (e.g. SAR model, SEM, Spatial Durbin Model (SDM) etc.). In our paper we want to know whether the dependent variables are related to those of the neighbour clusters, and we assume that the effects of the spatial lag of the dependent variables are linear and constant across observations. This brings us to consider the spatial autoregressive model (SAR), where the outcomes of a region are affected by the outcomes of “nearby” regions.

In order to take into account this feature this section presents spatial autoregressive model (SAR) based on the same model as for the estimates of Tables 1 and 2. The spatial model allows controlling for the effect of similar state in the SI and RI indexes in a given region by neighbours regions through the matrix of contiguity where the centroid of polygons (polygons are the regions) is the reference point (latitude and longitude) to calculate the geographical distance among regions. We have thus generated a matrix of spatial weights based on the distances between points obtaining the lagged dependent variable in space. In the SAR model, y is a function of observable characteristics Xβ, the spatial lags of the dependent variable ρWy and unobservable characteristics ε, producing a spatial regression relationship:

$$Y = \alpha + \rho Wy + BX^{\prime} + \varepsilon$$

with W that represents the matrix of spatial weights. This model indicates that a region derives an advantage in terms of SI and RI that reflects a linear combination of the resilience (namely RI and SI) of the neighbour regions; B captures the effect of regional characteristics and p represents the effect of the resilience of neighbour regions (conditional on observed regional characteristics).

Tables 6 and 7 report the estimates using the same specification of the results reported in the main text and adding the spatial correlation control for SI and RI respectively. Both estimates report a strong and positive spatial correlation effect, thus confirming the important role played by proximity with other regions to explain the resilience of region i.

Table 6 Same estimates as for Table 1—sensitivity index controlling for spatial correlation

Table 7 Same estimates as for Table 2—response index controlling for spatial correlation

To test the robustness of results obtained with the SAR model, we run the estimations also using a spatially auto-correlated error model (SEM). SEM drops the assumption that outcomes are affected by spatial lags of the output variable and instead assume a SAR-type spatial autocorrelation in the error process. The SEM model (Tables 8 and 9), applied to the resilience variables, shows that our results are robust and virtually unchanged.

Table 8 Sensitivity index controlling for spatial correlation (SEM)

Table 9 Response Index controlling for spatial correlation (SEM)

Turning to our variables of interest, Table 6 shows that patents are still positively correlated with SI, but this is no longer the case for trademarks. Table 7 shows that both patents and trademarks remain strongly and positively correlated with RI. To sum up, the results in this section confirm that more innovative regions are more resilient, particularly when considering technological innovation. Service innovation, most notably KIBS, is not associated to resilience during the crisis, while it is positively associated to resilience in the aftermath of the crisis.