Evaluating the implementation of Smart Specialisation policy

ABSTRACT The Smart Specialisation Strategy (S3) is at the core of the 2014–20 European Cohesion Policy, supporting regions to identify the technologies and economic sectors that might comprise sustainable growth paths. This paper provides an early attempt to assess empirically, for the whole European Union, whether the choices made by regions in selecting S3 target sectors are consistent with their current or potential specialization patterns. Results show only a few regions selected S3 paths rooted in both their current specializations and related activities, most of them prioritized different combinations of unspecialized or unrelated sectors, thus limiting the growth potential of their S3 policy choices.


INTRODUCTION
The European Union (EU) assigned a central role to the Smart Specialisation Strategy (S3) within the Europe 2020 development agenda promoting smart, sustainable and inclusive growth (European Commission, 2012). The Regional Operational Programme 2014-20, and especially the European Regional Development Fund (ERDF), incorporated S3 policy in their agendas, devoting significant financial resources to implement the new approach envisaged by the programme. Over the period 2014-16, all EU regions defined their S3 policy priorities after prolonged negotiations with local stakeholders and European Commission (EC) officers. These priorities are currently being implemented through public calls and other administrative procedures.
Almost immediately, operationalization of the S3 programme, and its 'bottom-up' process of identifying regional targets of economic transformation through an 'entrepreneurial discovery logic', were criticized (see the recent review by Aranguren et al., 2019). Early concerns focused on the risks of ineffective implementation, especially in peripheral regions that face additional developmental constraints (Boschma, 2014;Iacobucci & Guzzini, 2016;Morgan, 2015). Quality of governance questions, weak regional innovation systems, the lack of capacity in specific knowledge-based sectors and concerns with the potential integration of local markets into global value chains have been highlighted as potential policy limitations (Capello & Kroll, 2016). Broader issues with the appropriate spatial scale of policy actions, of regional 'lock-in' and the complex interplay between tangible and intangible knowledge production assets, and their territorial distribution, have also been raised. Hassink and Gong (2019) remain ardent sceptics, prompting renewed defence by Foray (2019).
It is important to highlight that the current debate around S3, although very intense, has remained mostly speculative, with limited evidence-based analysis. Only recently a few studies have started assessment of the implementation of S3, focusing on the adherence of actual policy actions to the conceptual framework of Smart Specialisation directly. Among others, D' Adda et al. (2020) for Italian regions, Gianelle et al. (2020b) for the Italian and Polish cases, Trippl et al. (2020) for a set of 15 regions in 14 countries, Di Cataldo et al. (2021) and Deegan et al. (2021) for a set of the EU regions, and Biagi et al. (2021) for S3 in the tourism sector. None of these papers covers the entire geographical domain that S3 targets or evaluates the adherence of the actual strategies to the EC guidelines.
In this paper we address the partial lack of evidencebased systematic and comprehensive analysis on S3 operationalization for the EU as a whole. We propose a novel empirical framework to examine the coherence of regional policymakers' choices with the theoretical foundations of S3 and related EC recommendations for defining Smart Specialisation strategies. It is clearly stated that S3 should prioritize domains, and economic activities that 'build on each country''s/region's strengths, competitive advantages and potential for excellence ' (European Commission, 2012, p. 8). In order to test whether and to what extent the practice of S3 policy is coherent with these recommendations, we build an empirical representation of the policy choices selected by all EU regions. For each region, we match the S3 sectors targeted to their current pattern of economic specialization and to the potential for local knowledge-driven growth identified by the relatedness of the S3 targets to the existing economic base. We then compute a composite S3 policy coherence indicator to assess, through econometric analysis, whether the S3 selection process is associated with the institutional, economic and structural characteristics of EU regions overall.
Results show that regions across the EU have identified S3 priority sectors that vary significantly in terms of their connection to host economies. Relatively few regions have chosen Smart Specialisation 'by the book', targeting sectors in which they already have competitive advantage or in which they show clear potential to develop such. Most regions have chosen to distribute S3 funds across industries in which they exhibit only tangential evidence of building new growth trajectories on existing sets of capabilities. These findings sum to a rather grim diagnosis of the likely success of the S3 policy programme. Finally, estimation results indicate that S3 choices are robustly associated with the quality of local governments. Highquality institutions have focused S3 policy choices on EC guidelines by promoting regional strategies more closely related to current patterns of specialization or to sectors closely related to existing capabilities.
The rest of this paper is organized as follows. The next section provides a brief overview of the Smart Specialisation literature, highlighting recent evaluation of the policy rollout and the model of regional branching and related diversification upon which the programme rests. The core research questions that we examine are derived from this discussion. The third section outlines the construction and description of data with a special focus on identification of S3 target sectors. The first and second research hypotheses are analysed in the fourth and fifth sections, respectively. In the sixth section we combine the analysis of our two research hypotheses and employ an econometric model to link the choice of S3 target sectors to the institutional and economic characteristics of EU regions. The last section provides some concluding remarks and policy implications.

LITERATURE REVIEW AND THEORETICAL BACKGROUND
The S3 programme represents a radical shift in EU Cohesion Policy, a clear break with the industrial policy of the past few decades and an embrace of place-based reforms that target national and regional economic development strategies and allied institutional reforms (Barca, 2009;McCann & Ortega-Argiles, 2013). Envisaged largely as a 'bottom-up' initiative, the S3 platform seeks both to renew and to widen the knowledge and industrial base of regions, at different spatial scales, leveraging existing capabilities to support innovation and new trajectories of growth (Kroll, 2015). This Smart Specialisation agenda emerged out of the Knowledge for Growth Expert Group (Foray et al., 2009) that emphasized 'entrepreneurial discovery' to identify those research domains, sectors and institutional structures in which regions possessed existing strengths, alongside a vision that that long-run competitive advantage would result from scaling core activities through processes of diversification and complementary innovation.
For many, the policy ambition of Foray and colleagues is woefully under-developed. Indeed, the authors themselves admit as much (Foray et al., 2011) in a paper that did little to dampen the enthusiasm of their critics. Early concerns raised by McCann and Ortega-Argilés (2015) question whether lagging regions have the potential to develop effective Smart Specialisation policy. This theme is echoed by Rodriguez-Pose et al. (2014) who note the absence of effective governance and the institutional supports vital for innovation policy to succeed in peripheral regions of Europe. Moodysson et al. (2015), building on Borrás (2011), argue that S3 fails to connect required innovations in the policy environment, reaching across multiple spatial scales, that successful implementation of regional economic innovation and new path creation will demand. Veugeleurs (2015) is in broad agreement, doubting whether there is sufficient scope for heterogeneity within Smart Specialisation policy at all. Kroll (2015) laments the lack of integration between place-based theoretical development of innovation systems in economic geography and S3, and finds the focus on local techno-economic potentials 'somewhat chaotic'. He is joined by Pugh (2018) and Gianelle et al. (2020a) in questioning whether S3 policy is flexible enough to operate across the heterogeneous institutional environments found in EU regions. Morgan (2015) is more direct, claiming that S3 lacks any empirical foundations. For Nauwelaers et al. (2014), S3 constitutes a rushed application of under-developed theoretical ideas and an inadequate assessment framework. Boschma (2014) recognizes some of the same theoretical failings and attempts to shore-up the policy framework by closer integration with the concept of related variety. Hassink and Gong (2019) summarize many of the criticisms levelled against the S3 programme spurring a spirited defence by Foray (2019).
At its core, S3 is a regional policy framework built around identification of competitive, local specializations and on extending the capabilities existing within those specializations to diversify economies along new, innovative pathways (Hassink & Gong, 2019;McCann & Ortega-Argilés, 2013). In many respects, this policy model sits squarely on top of the economic geography literature that focuses on the generation and maintenance of regional competitive advantage (Maskell & Malmberg, 1999). It recognizes the importance of innovation to long-run growth and imagines innovation as growing out of the capabilities, the sets of tangible and intangible assets, that support competitive clusters of economic activity found in different locations. In this way, regions are not expected to follow a 'one-size-fits-all' model of development (Tödtling & Trippl, 2005), but rather to chart their own course, giving rise to greater differentiation across the EU economic landscape. Operationalization of this framework requires the identification of local/regional capabilities upon which existing competitive industrial specializations have been developed and a model of diversification that is explicitly linked to the evolution of those capabilities.
Identification of the existing regional capabilities is the role of the entrepreneurial discovery process (EDP) of the S3. Local economic actors are seen as having the placespecific knowledge of industrial organization, institutions, innovation systems and markets around which regional forms of competitive advantage have been created (Boschma, 2014;McCann & Ortega-Argilés, 2013). The concept of self-discovery of capabilities is a cornerstone of the development models of Hausmann and Rodrik (2003). They also see state support for resulting policy choices as critical to economic growth, helping to offset market failures in terms of the absence of investment in new growth possibilities and too much diversification that may diffuse productive potentials. Other concerns regarding EDP have focused on the self-interest of local entrepreneurs and on whether they have the ability to detect the 'right stuff' on which self-sustaining growth depends, especially in peripheral regions (Sotarauta, 2018). Hassink and Gong (2019) raise further concerns related to regional 'lock-in' and the inability to make radical structural transformations under the EDP. Boschma and Gianelle (2014) offer the clearest model of diversification linked to Smart Specialisation policy. They view diversification as key to the process of knowledge-based economic transformation, as a process of building new growth paths out of existing national and regional capabilities. At the core of the diversification process is how to leverage the existing capabilities of the region into new ventures. These capabilities, as mentioned early, are the tangible and intangible assets that are most effectively deployed within the region in the creation of value. They may be distinct pools of tacit knowledge, modes of organizing firms and entire industries, the formal and informal institutional supports that undergird local networks of education, learning, knowledge production, its diffusion and effective absorption. One of the most difficult questions that the EDP must face in identifying these capabilities is which of them are actually placebased and thus provide a platform to generate further growth at the local/regional level. The success of diversification is likely to depend upon the strength of existing capabilities within the region and upon the possibilities for deploying those capabilities in new domains of economic growth. In this sense, related diversification is critical for successful regional branching (Frenken & Boschma, 2007;Neffke et al., 2011). Related diversification occurs when the capabilities that support specific kinds of economic activity are successfully redeployed to support other activities. The hope is that over time, sets of regional capabilities will be broadened as each type of economic activity rests, at least to some extent, on slightly different skills, technologies, organizational and institutional structures. Larger, more developed regions have a relatively easy time diversifying their economies as they already possess many sets of related and unrelated capabilities and thus can branch into many different activities. Less developed economies have smaller sets of capabilities and their options for related diversification are therefore limited. At the same time, Grillitsch et al. (2018) argue that regions that focus just on related sectors may hinder long-run growth due to potential lock-in, whilst unrelated diversification may ensure long-term competitive advantages.
It is important to highlight that the current debate around Smart Specialisation, although very intense, has remained mostly speculative, with limited evidencebased analysis. An exception is , who develop a theoretical framework to assess the S3 policy, but not its implementation, using patent data to compute measures of relatedness and complexity. Rigby et al. (2019) take the same analytical framework and use historical data to assess whether regions that followed a technological trajectory coherent with the S3 approach enjoyed improved economic performance. Along this same research path, Balland and Boschma (2020) explore the relevance of interregional linkages, especially in complementary capabilities, on the potential process of S3 diversification. While important, none of these studies directly tackles the assessment of the actual implementation of S3. Only recently have a few studies assessed the adherence of real policy actions to the Smart Specialisation conceptual framework directly: D'Adda et al. The present contribution moves along these same lines, though it is broader in geographical focus, aiming to address the partial lack of evidence-based systematic and comprehensive analysis on S3 operationalization for the whole of Europe. We propose a novel empirical framework to examine the coherence of regional policymakers' choices with the theoretical foundations of S3 and related EC recommendations for defining Smart Specialisation strategies. To illustrate our research question, we start by reconsidering the EU guidelines to the smart strategy definition. In the official S3 platform of the EC it is clearly stated that the regional S3 'should prioritise domains, areas and economic activities where regions or countries have a competitive advantage or have the potential to generate knowledge-driven growth'. To test whether and to what extent the practice of S3 policy is coherent with these recommendations, we articulate our research agenda into two main hypotheses: Hypothesis 1. Regions selected their S3 targets in domains/areas/ activities in which they currently exhibit revealed comparative advantage (RCA).
Hypothesis 2. Regions seek to generate knowledge-driven growth by developing new RCA in S3 target sectors that are related to their existing patterns of RCA.
To examine these hypotheses, we build an empirical representation of the policy choices selected by all EU regions. For each of these regions, we match the S3 sectors targeted to their current pattern of economic specialization and to the potential for local knowledge-driven growth identified by the relatedness of the S3 targets to the existing economic base.
The EC has recently reclassified the S3 domains, often declared in creative or hazy ways, by assigning to each priority the complete set of economic sectors involved in its implementation. This allows us to obtain a complete representation of the economic dimension of S3 domains. Mapping the existing economic structure of regions and their potential growth paths use existing EU data and measures of relatedness between economic sectors derived from a European production space. Testing Hypothesis 1 is possible by comparing the existing economic specialization of regions, identified using RCA indicators, to the S3 policy priorities selected. Testing Hypothesis 2 is more challenging because we have to deal with the potential notion of knowledge-driven growth, likely to be triggered by new specializations. To operationalize this idea, we use the concept of relatedness density proposed by Hidalgo et al. (2007). In the context of our analysis, relatedness density measures the degree to which an S3 sector uses economic capabilities that are readily available within a region. Higher relatedness density implies that a sector is more likely to be successful in activating growth through diversification. Such density measures are derived from the comprehensive description of the production space based on current and prospective co-specializations that we provide for Europe as a whole.

DATA AND METHODS
To investigate our research questions, we need to build three homogeneous blocks of regional data on: S3 selected sectors; current production specialization; and potential related production specialization. In the following, data and methods applied to build these blocks are discussed, while a descriptive analysis is provided in the supplemental data online.
The Smart Specialisation Strategy (S3) As remarked by D'Adda et al. (2019), drilling into the details of regional Smart Specialisation strategies is not an easy task given the absence of a codified system for the classification of targets. As a result, each region has specified its S3 domains in a flexible and creative way so that comparisons across regions and quantitative evaluations are almost impossible. The EC has developed a S3 platform where information on the regional strategies is gathered. 1 In 2018 the EC has enriched the platform classifying each strategy according to standard taxonomies. 2 We will focus on the economic dimension, based on NACE twodigit sectors in both manufacturing and services. This implies that S3 targets can be analysed beyond both the 'pure' technological domain and the manufacturing perimeter by including services. This is crucial given that several regions based their S3 policy on service activities such as tourism, culture, archaeological heritage and health. S3 has been implemented at different territorial levels: national and NUTS-1, -2 and -3 regional levels. 3 From now on we simply refer to 'regions', regardless of the NUTS level. The choice to perform S3 at the national level seems reasonable for small countries, while it is more surprising for large countries such as Hungary and Bulgaria, given that the policy was originally intended as a local strategy. 4 According to the EC guidelines, each region was supposed to build its S3 on a limited number of priorities, namely the economic activities where the region had a competitive advantage or the potential to generate knowledge-driven growth (Foray, 2015). The key idea of the strategy is to concentrate the managerial and financial resources available in the region on a few well-defined priorities to avoid policy dilution following the EDP (Gianelle et al., 2020b).
We might expect a higher number of priorities by richer regions that have wider technological opportunities and are more likely to face greater requests by local stakeholders. At the same time, less developed regions, where private investments are scarce, might prefer a more flexible and inclusive strategy, enlarging the number and the scope of their priorities to exploit all potential investment opportunities (Trippl et al., 2020). The average number of priorities is six, with a high variability among regions. Thus, a first consideration is that the number of priorities pursued by many regions seems higher than we would have expected, although the S3 foundations do not provide very clear guidance on this issue.

Evaluating the implementation of Smart Specialisation policy
According to McCann and Ortega-Argilés (2015) there is no clear rationale to the selections made: regions with similar socio-economic background chose diverse thematic and sectoral priorities in both quantitative and qualitative terms. Looking at the regions according to their geographical location, our results confirm Iacobucci's (2014) prediction and Kroll's (2015) preliminary assessment: Southern regions have a higher number of priorities (on average 7.2) than Central-Northern and Eastern regions (an average less than 6.0). Looking at the economic sectors included, they appear as diverse as the nature of the selected priorities with a high regional heterogeneity.
The great variability in regional strategies in terms of priorities and sectors was partly expected as an obvious consequence of the S3 general strategy, which aims at avoiding all regions following the same direction. However, it is difficult to understand the S3 selection process in many cases. The choices made do not all appear to be clearly linked to a strategy of evidence-based assessment. At the same time, it is important to remark that each region had to allocate a significant part of its ERDF resources (such as firms' financial incentives) to the priorities and sectors indicated in the S3. Therefore, rational behaviour by regional policymakers was to define priorities in a generic fashion so that a larger set of economic sectors was included in their strategies, thus enhancing private sector investments. Overall, there is little sign of proliferation and a widening of specialization domains, but rather diverse approaches to the decision process, apparently independent from the development status and the institutional setting, and more related to the mode of governance (Kroll, 2015).
Finally, for operational purposes, we consider for each region the entire set of S3 targeted sectors, regardless of the original priorities; this constitutes a sort of regional 'unified' S3 strategy, since we have included all the NACE sectors selected at least once in the original priorities. The result is a binary matrix of 169 regions and 82 sectors, where each entry s3 ri takes the value of 1 if sector i is included in the S3 strategy of region r, and 0 otherwise.

Existing regional production specialization
The second block of data we need to test Hypothesis 1 entails the drawing of a comprehensive map of economic specialization across regions in Europe, which can be juxtaposed to the S3 pattern. We provide this representation by computing the RCA index, based on employment in 2016, classified by NACE economic sectors and extracted from Eurostat Structural Business Statistics (SBS). 5 We collect data for 243, mainly NUTS-2, regions. In a similar fashion, Balland and Boschma (2019) have recently explored EU regional specialization by using a different database, derived from the EU Labour Force Survey. Since our aim is to examine for each region the degree of association between S3 target sectors and the existing production specialization, we need to match the regional and sectoral dimension of S3 with the employment data. The intersection between the two datasets contains 166 territorial units 6 and 64 sectors (see Tables A1 and A2 in the supplemental data online).
Finally, we use the employment data to compute the RCA index and a regions/sectors matrix with entries taking the value of 1 if a region has a comparative advantage (RCA > 1) in a given sector, and 0 otherwise.
The European production space and regional relatedness density In the third data block, we provide a measure for the potential to activate growth through diversification by computing the relatedness density of each S3 target sector to each region''s existing economic base. As already mentioned, this measure proxies the availability of local capabilities and it is derived from the empirical representation of the European production space in terms of co-specializations. Thus, we provide an alternative measure of such space, complementing existing spaces based on patents, as in  and Rigby et al. (2019). The importance of relatedness for regional innovation and economic development is emphasized by Boschma (2005) and . The pace and direction of technological dynamics in a region are shaped by the costs and benefits of exploiting new ideas given the existing mix of knowledge and industry. This cost-benefit balance of diversifying from one technology to another is more favourable when two technologies are related. Several studies (Boschma et al., 2015;Boschma & Iammarino, 2009;Kogler et al., 2013;Maggioni et al., 2019;Rigby, 2015;Rigby & Essletzbichler, 1997) have shown that knowledge production within regions accumulates in a path-dependent fashion going from an existing technology to a related one.
Following Hidalgo et al. (2007), we thus proceed by building a European production space using the employment data. First, we compute the proximity matrix for the 64 sectors considered in our analysis. Proximity or relatedness between any two sectors, i and j, is given by the minimum of the pairwise conditional probability of a region being specialized in the production of sector i ( j) given that it is also specialized in the production of sector j (i): The proximity parameter w i,j provides a measure of the strength of co-specialization between sectors i and j and it is computed using all 243 EU regions for which sectoral employment data are available in order to maximize the information on economic co-specialization. 7 The resulting matrix represents the European production space, which is depicted in Figure 1 as a network. For the sake of visualization, we aggregate the 64 sectors into 13 macro-sectors. The graph shows the relevance and centrality of services sectors across most EU regions. Economic sectors that cluster together are more highly related than those which are relatively distant. We interpret the relatedness between sectors as an indication of shared capabilities in terms of production requirements. Thus, if a region has the capabilities to produce output in industry i it is also likely to own the capabilities also needed to produce output in industry j if the industries i and j are related to one another. In general, the 116 Emanuela Marrocu et al.
production space based on employment data allows an assessment of the interactions among sectors in a more comprehensive way with respect to patents. The next step is to calculate the 166 regions by 64 sectors relatedness density matrix V. For sector j in region r the matrix entry is computed as follows: where I r i is an indicator function taking the value of 1 if RCA r i . 1, and 0 otherwise; and f i,j is the sectoral proximity parameter discussed above.
Finally, using matrix V, we compute two measures of the average relatedness density for each region considering: (1) all sectors of the regional economic structure; and (2) the S3-specific sectors including only the sectors selected for the region's S3 policy. The average value computed considering all economic sectors provides the existing pattern of relatedness density of a region''s production structure regardless of the S3 implementation. The correlation between the two regional average relatedness densities is very high (0.98). We return to this finding in the fourth section. Figure 2 maps the density measure considering all economic sectors and the regional value of aggregate relatedness density appears highly differentiated across Europe. Overall, the average relatedness density is equal to 0.35 and ranges from a minimum of 0.14 (Nord-Est in Romania) to a maximum 0.61 (in Hungary). As expected, the higher levels of relatedness density are found for regions identified at the national scale: Hungary, Czech Republic, Slovakia, Slovenia, Croatia and the three Baltic States. The portfolio of specializations at the country level is in general wider than at the regional level, thus it is more likely for a given sector to be surrounded by many related sectors. High values are also found in well-developed regions, such as Lombardia in Italy, Île-de-France and Hessen in Germany, while only few of them are developing regions such as Dolnoslaskie and Malopolskie in Poland. In contrast, the lowest levels of relatedness density are detected in the small and less developed areas of Greece, Romania and Southern Italy because of the weak and sparse production space of these regions, where co-specializations are rare. Interestingly, a low relatedness density is also shown by several French regions (Languedoc-Roussillon, Bretagne, Lorraine, Basse-Normandie), signalling strong territorial specialization of the French production space.
In general, both the overall and the S3-specific relatedness densities largely depend on the size of the economy and on to the specialization pattern of the production structure. Countries and rich regions show the highest aggregate relatedness densities, while less developed and small regions lie in the low part of the ranking. This implies that the initial conditions of each region are highly differentiated: the higher the overall relatedness density exhibited by a region due to its sectoral specialization, the more likely that region will reach a higher S3 relatedness density, given the number of S3 target sectors. Thus, in the fourth section we use the general and S3-specific densities to compute a new measure suitable to assess the relationship between S3 choices and the extent of potential growth.

S3 AND REGIONAL PRODUCTION SPECIALIZATION
The purpose of this section is to address the first research question by examining whether sectors selected in the S3 regional strategies are those in which regions exhibit Evaluating the implementation of Smart Specialisation policy 117 comparative advantage. Using the two binary matrices presented in the previous section, as a preliminary step in our analysis we compute the share of S3 sectors in which each region currently exhibits RCA > 1. We find 48 regions with a share below 33% and only 16 regions with shares higher than 66%. On average, regions have RCA in 43% of the target sectors they have chosen to prioritize for their Smart Specialisation policy, indicating that the degree of coherence between current specializations and S3 sectors is relatively low. We proceed by formally testing the degree of association between the two sectoral distributions: that in which the region already has RCA and that reflecting regional S3 policy choices. The first remarkable result is that the average of the estimated correlation coefficients between these distributions is rather low at 0.13. 8 To test this association, we also computed the Pearson''s chi-squared test. It is worth recalling (Guilford, 1936) that for the case of two binary variables the Pearson correlation coefficient is equal to the mean square contingency coefficient f (with f = x 2 /n , where χ 2 is Pearson's chisquared test and n ¼ 64 is the number of sectors). We find that that in 107 out of 166 regions, the null hypothesis of the test (no association) is not rejected at conventional significance levels. 9 This means that, on average, there is little association between the S3 target sectors and the actual production specialization of most regions.
Looking at the geographical representation of the degree of association (Figure 3), we observe a positive and statistically significant association in some Greek regions. Indeed, four out of the six regions with the highest correlation coefficients belong to Greece. Among the top 10 regions, we find two from Poland and one from Romania, France and Spain. The same differentiated scenario emerges also for the regions with the lowest association. Overall, the regional variability in the correlation coefficients does not seem to exhibit any clear spatial pattern (Moran's I test ¼ 1.19, p ¼ 0.233).
The novel evidence provided so far seems to clearly reject our research Hypothesis 1. On average, while designing their S3 policy, regions have not selected those sectors where they already have comparative advantage. This result stands in stark contrast to the theoretical recommendations and relative guidelines for the favourable implementation of the S3 in Europe. It is clear that S3 policy is built around diversification. With the aggregate nature of the sectors identified in this analysis, it should be clear that there is plenty of scope for diversification across individual industry and product lines within, as well as beyond, each of these sectors. Thus, the relatively low share of priority S3 sectors corresponding with existing regional specializations is rather worrisome.  Emanuela Marrocu et al.

S3 AND REGIONAL PRODUCTION RELATEDNESS
The purpose of this section is to address research question Hypothesis 2 by assessing the relatedness between the sectors selected in the S3 and the existing economic cores of the regions examined. More precisely, we want to measure the degree to which a sector included in a region's S3 policy uses economic and knowledge capabilities that are readily available within that region. For this, we build on the two measures of regional average relatedness density previously discussed: the aggregate density score for all existing sectors in a region and the density score focused only on the S3 target sectors.
It is worth noting that if we compare the average relatedness density scores for the existing sectors in a region and the aggregate S3-specific densities, they are very similar for most regions. Indeed, across regions, the average density score for existing sectors is 0.3453, whereas the S3 density score averages 0.3544. This is a first indication that regions have not targeted their most highly related sectors in their S3 policy, but have tended to replicate the mean underlying features of their current production structure. This could be the result of rational choices made by policymakers, as already anticipated above. Those policymakers may have paid more attention to the requests of local stakeholders and/or to the chances of attracting private investment flows for the economy as a whole, rather than trying to exploit as much as possible the growth potential of their own region's economic base.
In order to measure the extent of the loss in terms of unexploited relatedness density (and ultimately in terms of growth opportunities), we rank for each region in decreasing order all the sectors according to their relatedness densities. We then calculate the average maximum potential relatedness density (from now on max-potential) achievable given the number of sectors the region has selected in its smart strategy. For example, the max-potential relatedness density attainable by a region with, say, 10 S3 sectors is the average of the relatedness densities for its 10 sectors with the highest relatedness density. Finally, we evaluate how close the region has come in its actual S3 choices to this benchmark by calculating the difference between the average S3-specific relatedness density and the max-potential. In order to take into account starting conditions (i.e., having a specialized or diversified production space), we compare regions on the basis of the percentage ratio of the difference computed above with respect to the max-potential.
The percentage ratio is a measure of the 'loss' in terms of economic growth potential that a region may experience related to its S3 sectoral policy choices. If the loss is approximately zero, it means that a region has targeted S3 sectors that maximize relatedness density to its current economic structure. The larger the loss, the more distant is the focal region choice from the 'maximizing' S3 strategy, Evaluating the implementation of Smart Specialisation policy given initial production conditions. This loss measure provides an appropriate indicator to assess our research question Hypothesis 2. Given the value of the max-potential, a small loss implies that the region chose S3 sectors with a better 'fit' to the regional economy and, as a result, it enables higher potential economic growth through diversification across sectors. In other words, the higher a sector's relatedness density to the economic core of a region, the lower the costs and risks for the region of developing that sector. This is because as the relatedness density of a target sector increases within a region, the more likely the pool of capabilities, skills and knowledge required in that sector is already locally available. Figure 4 shows that the loss in S3 relatedness density is highly spatially differentiated, though clear spatial patterns are barely discernible with a Moran's I test statistic of −1.86 (p ¼ 0.062). The region, which comes closest to maximizing its S3 relatedness density, given the available potential in the region, is Mazowieckie in Poland with a loss of just −1.6%, followed by three Italian regions: Veneto, Toscana and Campania, and then Östra Mellansverige in Sweden and Etelä-Suomi in Finland. Interestingly, among the best performers we find regions with a high max-potential such as Veneto (0.53) together with regions where it is quite low, such as Campania (0.29) and the Greek Peloponnisos (0.21). These regions, although characterized by very different production structures and specialization patterns, were all able to choose S3 target sectors with high relatedness density. Similarly, the territorial composition at the low end of the ranking is highly differentiated: the highest loss, −35%, is presented by Bulgaria, followed by the South East of the UK, Hovedstaden in Denmark, Helsinki-Uusimaa in Finland and Podkarpackie in Poland. Again, the loss seems independent from the starting conditions in terms of maxpotential. Helsinki presents a loss of −28% starting from a high max-potential (0.63), while a similar loss (−27%) is found in Sud-Vest Oltenia in Romania, which had a much lower max-potential (0.24).
If we apply the geographical topology to explore the distribution of these results, one interesting outcome arises. Southern regions have chosen relatively densely related sectors among those available, with an average loss of −10.3% and quite similar is the performance of the Central-Northern regions (−11.8%). In contrast, Eastern regions, despite the presence of several territorial units at the country level, are on average the most distant from their max-potential, with a loss of −15.1%. 10 The evidence provided in this section for testing the research Hypothesis 2 shows a highly differentiated behaviour by regional policymakers. Only 17 regions have a loss smaller than 5%. A loss higher than 10% is found for 92 regions, and in 19 regions losses exceed 20%. Thus, it appears that many regions have selected their S3 sectors without considering the available local core of knowledge, and thus with little attention to the risk that unrelated diversification poses for potential future growth. It is worth considering that this might have been the result of political choices based on different grounds or of policymakers lacking relevant and readily usable information on the production features of their own regions and an adequate benchmarking system.

A COMPREHENSIVE EVALUATION FRAMEWORK
Combining the two research hypotheses Our main results indicate that the S3 policy choices of EU regions have not, in general, tended to target sectors in which they have an existing comparative advantage or in which they have significant potential to develop new specializations. We have shown that in their implementation of S3, regions across the EU exhibit considerable heterogeneity and, most importantly, that heterogeneity does not closely reflect the recommendations of Smart Specialisation theory or EC guidelines. The evidence presented raises concerns regarding the likelihood that Smart Specialisation target sectors will stimulate successful growth trajectories that leverage existing or related regional capabilities. This finding does not imply any kind of judgement on the choices made by regional policymakers, or that the policy will necessarily result in ineffective outcomes. However, growth strategies that are relatively unrelated to a region's current and prospective assets are riskier and do seem inconsistent with the bottom-up, evidence-based policy framework at the heart of the Smart Specialisation programme.
In order to highlight the S3 trajectories that emerged from the evidence gathered in testing our two hypotheses, Figure 5 locates EU regions in a two-space that indicates their relative location in terms of the correlation of their S3 targets and the existing sectors in which they exhibit RCA (Hypothesis 1 on the vertical axis) and in terms of the percentage relatedness density loss associated with their S3 policy choices (Hypothesis 2 on the horizontal axis). Note that intercepts of the horizontal and vertical lines are set at medians across the regions (0.123 for Hypothesis 1 and −10.8 for Hypothesis 2). Thus, Figure 5 plots simultaneously how many regions have targeted S3 sectors linked to their existing RCA core and/or to other parts of the regional economy that are high related to that core. Four possible scenarios are identified, varying in terms of their association with the existing economic core of the region and the potential of the region to be able to leverage growth in related activities. See also Figure  6 for the mapping of the European regions in the four quadrants of Figure 5.
In the upper right part of the graph Q1 are regions that have chosen a 'virtuous path' as their targeted S3 sectors are closely linked with their current specialization patterns in terms of sectoral overlap and relatedness. If, for instance, we consider the subset of regions with a correlation higher than 0.24 (significant at the 5% level) for Hypothesis 1 and below the 8% loss for Hypothesis 2, we end up with 15 regions (the shaded area in Figure 5). Among them four Spanish regions (País Vasco, La Rioja, Illes Balears and Andalucía), three Greek regions (Notio Aigaio, Sterea Ellada and Peloponnisos), Wien and Oberösterreich in Austria, and one region in Italy, Portugal, Sweden, Finland, Denmark and Germany. These regions have good chances of developing new comparative advantages because they have selected S3 sectors that are both related to their current specialization and to the core of available knowledge, hence they are less hazardous to develop. The territorial composition of this subset is quite differentiated though it may be remarked In the bottom right quadrant Q2 are regions that we classify as 'out of the beaten path'. The S3 targets of these regions do not overlap closely with the current pattern of RCA, though these targets are relatively highly related to existing specializations in the regions. Interestingly, in this portion of the diagram we find several rich German and Swedish regions together with some innovative regions, such as Lombardia and Emilia-Romagna in Italy, and Cataluña in Spain. At the same time, this quadrant includes developing regions that are trying to diversify their production specialization towards new sectors related to their economic cores. Among these, it is interesting to mention the case of Sicilia, which has explicitly stated its intention to exploit S3 opportunities to radically renew its strategic orientation (Bellini et al., 2021).
In the upper left quadrant Q4, we find regions that have chosen a 'conservative' or safe path, as their S3 strategy is shaped by existing RCA-based specializations though not by high levels of relatedness to new sectors. This scenario might bolster existing strengths, in line with EC recommendations, but it also elevates the risk of getting locked into the current pattern of specialization. This might be the case for the developing regions of Poland, Romania, Greece and Spain included in this quadrant, which risk the perpetuation of a weak equilibrium.
Finally, regions in the lower left quadrant Q3, have chosen a quite different and 'risky' path: they have designed their S3 policy targets with little regard to existing patterns of specialization and, at the same time, away from those sectors that are closely related to existing specializations. These unrelated diversification scenarios depend almost entirely on external capabilities, or on a broad transformation of local capabilities. This strategy is very risky for the regions involved. Regions in this group appear to be quite heterogeneous including Berlin and Lazio (Rome''s region) together with several French (eight), Polish (six) and UK (five) NUTS areas.

Looking for the determinants of regions' S3 choices
Is the selection process that identifies S3 target sectors associated with the institutional, economic or structural characteristics of NUTS regions? To investigate this issue, we compute an S3 policy 'coherence' indicator as a dependent variable and then regress that on a comprehensive set of potential covariates. To build the coherence indicator, the two measures adopted to test Hypotheses 1 and 2 were standardized using the min-max procedure and then averaged. We assign equal weights to the two measures because the EC recommendations (see the statement reported above) places the development of actual (Hypothesis 1) and potential (Hypothesis 2) competitive strategies on the same footing. As S3 is still clearly a 'policy running ahead of theory' (Foray et al., 2011), we cannot proceed by testing theoretical propositions to single out the main determinants of the S3 regional choices. However, given the bottom-up nature of the policy, regional government authorities and local authors played a key role in selecting the priorities. Therefore, we expect the overall coherence of the policy to be positively related to the quality of local institutions. We maintain that high-quality local institutions are less likely to be influenced by external conditions and more likely to follow EU Guidelines. Hence, they are more capable of designing S3 policy to maximize their own region's growth (Capello & Kroll, 2016;D'Adda et al., 2020). It is also possible that weaker regional institutions may be 'captured' by local stakeholders with specific sectoral interests. In this case, the number of S3 targets may proliferate, reflecting pressure from local firms rather than the exploitation of real growth opportunities. In our analysis, we proxy the quality of local institutions by the European Quality of Government Index (EQI), a multidimensional metric resulting from the combinations of three indices: high impartiality, quality of public service delivery and low corruption (Charron et al., 2015). 11 As a possible driver of S3 policy coherence, we also considered the general level of regional economic activity, measured by gross domestic product (GDP) per capita. In this case, the expected association is less straightforward. On the one hand, wealthy regions have more opportunities to diversify since their production structure is wider and more articulated. Moreover, such regions host a larger number of firms, likely to be involved in the strategy and implementation of S3 funding calls. On the other hand, lagging regions with a weak production structure have fewer investment opportunities. These regions may select a relatively large number of S3 targets, simply hoping that at least one of them delivers. In general, we expect a positive association between regional GDP and the S3 policy coherence indicator.
We add two more intangible factors to our model: human capital and technological capital. Human capital is measured by the share of the population aged 25-64 with a university degree (International Standard Classification of Education (ISCED) 5-8). Technological capital is proxied by R&D expenditure per inhabitant or, alternatively, by the number of patent applications to the European Patent Office (EPO) per million inhabitants. Although we expect these intangible assets to complement and reinforce S3 effects once the policy is implemented, we have no clear-cut expectation on the direction of association with the S3 coherence indicator.
We control for the level of agglomeration by including population density in the econometric model. We also add a measure of the structure of the regional economy by including specialization indices for low-technology manufacturing and knowledge-intensive services. 12 Finally, we include two territorial dummies to account for additional economic, institutional, and social features not entirely accounted for by the variables mentioned above. A 'southern' dummy flags the southern regions of Greece, Italy, Spain and Portugal and a 'new' dummy flags the 11 new accession countries in the EU. 13 Before carrying out the regression analysis, we test the coherence indicator, as well as each standardized indicator, for spatial autocorrelation. The Moran's I test computed by using the max-eigenvalue normalized inverse distance matrix returned no significant results.
The main results are reported in Table 1. As for the coherence indicator (columns 1-3), we find evidence of significant positive association with respect to the Quality of Government variable. GDP per capita exhibits a significant positive coefficient only when the model excludes the EQI variable (column 3). The two variables are highly collinear (correlation coefficient ¼ 0.70); thus, when both are included in the model (column 1) a multicollinearity problem arises. No other variables seem to play a significant role as possible drivers of regional S3 choices except for the Southern regions dummy. It is worth noting that human capital and technological capitaleither proxied by R&D or patents per capitado not exhibit significant coefficients even in specifications in which they are included one at a time as the main S3 policy driver (EQI and GDP per capita excluded, controls for productive structure and territorial features included). 14 These results are in line with those in Di Cataldo et al. (2021), who find weak evidence for technological capacity and no evidence at all for human capital as drivers of S3 development axes, economic or scientific domains and policy priorities. This could be due to human capital having no direct effects, rather only indirect ones through the quality of government. As for technological capital, it is reasonable to expect a direct effect on the scientific domains rather than on the selection of productive sectors; as discussed in the previous sections, their set is rather heterogeneous and includes also traditional and low-tech economic activities. Overall, although the empirical literature on regional performance has provided extensive and robust evidence on the role played by intangible assets, such as human and technological capital, in determining economic outcomes, the S3 selection policy does seem to be almost entirely unaffected. 15 Even though we maintain that to assess the role of regional structural characteristics on S3 choices it is more appropriate to consider the composite coherence indicator, Table 1 also reports the results for the Hypothesis 1 (columns 4-6) and Hypothesis 2 (columns 7-9) indicators. Results confirm the key role of the Quality of Government in driving the accordance between S3 choices and the EC recommendations for Hypotheses' 1 and 2 indicators. Per capita GDP exhibits a positive coefficient though not significant at conventional levels. Technological capital turns out to be negatively associated with S3 choices targeting sectors with existing comparative advantage, while we find some evidence suggesting that regions might have leveraged existing knowledge and innovative capabilities to develop potential comparative advantage in related sectors (columns 7-8). Population density is positively associated only with Hypothesis 1, whereas no evidence is found for human capital as a driver of S3 choices. Overall, the estimation results indicate that policy decisions on S3 target sectors are robustly associated only with the quality of local governments. Institutions with high quality are able to lead the regions towards a positive path of potential growth consistent with EC guidelines by adopting more precise and more focused strategies. It appears to be the case that low-quality governments are more prone to fulfil the expectations and requests of local stakeholders. Indeed, the inclusion of a specific sector in the S3 means that private investments in that sector become eligible for EU financial grants. Therefore, the local authorities' choices are likely to be influenced by stakeholder pressures, which might be exerted by pure rent-seekers in the worst cases. Moreover, low-quality institutions might also face more difficulties accessing and processing the necessary information required for the complex and challenging S3 policy agenda.

CONCLUSIONS
As the S3 priorities are currently being implemented with the assigned resources to be spent by 2023 (the n + 3 rule of the 2014-20 EU programmes applies in this case), it will not be possible to evaluate the economic impacts of the Smart Specialisation programme until at least a certain number of years have elapsed since the end of the term above. Nonetheless, it is possible to assess how regions and countries have interpreted the conceptual framework of S3 and how they have moved from theory to practice. Most countries and regions have included S3 in their development policies and devoted a share of available EU resources to their 2014-20 regional operational programmes. The strategy has attracted a lot of attention from policymakers and academics because it represents one of the largest experiments of place-based development policy centred on the selection of local priority sectors. In this paper, we have empirically assessed how much the choices made by regions in selecting S3 sectors are consistent with the EC aim to prioritize economic activities where regions are already specialized or have the potential to generate economic growth through related diversification within and beyond existing specializations.
Our analysis of regional strategies draws from the EC official S3 website, where all regions have disclosed their industrial and technological priorities, and from the employment data in manufacturing and services provided by Eurostat. These data allow us to examine for most EU regions the degree of association between S3 and both current and potential production specializations, in terms of competitive advantage and relatedness.
Results show that S3 practice has taken many different routes with respect to the guiding principles stated in the EC guidelines. Only a handful of regions have chosen Smart Specialisation 'by the book'. Most regions have only partially targeted sectors in which they have an existing competitive advantage or the potential to develop one.
Although our findings do not imply in any way a negative assessment either on policymakers or on the policy itself, it is important to remark that growth strategies unrelated to a regions' current or prospective specializations are much riskier and might entail higher implementation costs. We summarize regional strategies across four different trajectories by matching policy choices with existing and related capabilities within EU regions. These paths are characterized by strengths and weaknesses as much as opportunities and risks. In the future, it will be essential to assess whether the economic performance of regions is linked to the coherence of the S3 trajectory chosen. Unfortunately, the overall effectiveness of S3 policy is going to prove difficult to assess due to the economic impact of the Covid-19 pandemic and the different national responses to it.
Results from a cross-sectional econometric model suggested that individual and composite indicators of regional 'coherence' with S3 policy were positively and significantly related to quality of governance. There was also evidence of a positive relationship between S3 policy coherence and regional GDP. A dummy variable also revealed that southern regions of the EU prioritized S3 targets following EU guidelines more closely. Indicators of regional economic structure and human and technological capital were insignificant in the model. Until the results from the regional S3 choices are revealed, highlighting successful policy prescriptions remains impossible.
All in all, the results presented here should lead to further reflections on S3 policy from both a theoretical and practical perspective for potential future adjustments and improvements. Regional policymakers should have a comprehensive and collective base of information to lead the consultation process towards the best possible strategy. We believe that the novel analysis proposed in this paper should be part of the ex-ante information set available to regions in order to make more conscious and possibly more effective decisions for a truly 'evidence based' strategy. Moreover, the EU authorities should reflect on the opportunity to provide regional policymakers with more detailed and strict guidelines for the next operational programmes 2021-27 and more generally for policies with complex design and implementation criteria, as was the case for the S3.
There are several prospective research lines that might stem from this contribution. As regions are not independent units, as implicitly considered by the S3 programme, they are more or less connected depending on their geographical and technological proximity, and thus have varying capabilities to monopolize internal capabilities and also exploit external possibilities. More work is needed to assess interregional interdependencies, within and between countries. The duplication of S3 policy targets across many regions raises several questions, but also permits interesting research designs given that not all regions chasing the same industrial targets are likely to be equally successful. This calls for coordination efforts at all levels of governmentregions, countries, EUin order to Evaluating the implementation of Smart Specialisation policy maximize the beneficial effects of the integrated regional potentials.
In the near term, the S3 implementation process has already raised a number of policy-related concerns. While the 'bottom-up' initiative of the EDP is to be lauded for generating flexibility across regions in terms of sectoral targets, that same process raises several issues in terms of how effective identification of strategic priorities. First, vested business interests might be pushing rentseeking interests at the expense of choices that might have broader socio-economic impact. What should be the relative role of entrepreneurs in the EDP vis-à-vis local regional policymakers and other 'experts'? Second, how should countries, and even the EU as a whole, deal with the inevitable duplication of target sectors across different regions? Who chooses, and with what criteria, which regions might be favoured in this process? Third, although diversification is to play a critical role in S3 policy, what is much less clear, given the aggregate priorities listed by many EU regions, is how much diversification might occur within existing sectors relative to developing ones? Furthermore, what are the prospects of recombinant knowledge and broader forms of economic development resulting from growing breadth within existing specializations versus developing new sectors, and is there an optimum number of specializations for regions of different scales?

DISCLOSURE STATEMENT
No potential conflict of interest was reported by the authors.

FUNDING
David Rigby acknowledges funding from the Visiting Professor programme, sponsored by the Regione Autonoma della Sardegna (RAS), for his visiting period at the University of Cagliari in February 2020.

NOTES
1. See https://s3platform.jrc.ec.europa.eu/home. For a detailed description of the platform, see Sörvik and Kleibrink (2015). McCann and Ortega-Argilés (2016) provide a first overview of the regions' choices. 2. The economic dimension is classified according to the NACE rev2, the scientific dimension thanks to NABS 2007 and the policy dimension by referring to EU objectives.
3. We have aggregated the S3 defined at the NUTS-3 level for Sweden and Finland to the corresponding NUTS-2. 4. In six countries (Austria, Denmark, Germany, Greece, Poland and Portugal) the S3 carried out at the regional level was complemented with national projects effective for the whole country. We have excluded these national priorities to avoid the overlapping of different decision levels. 5. Four macro-sectors (A, Agriculture; K, Financial and insurance services; O-P, Public administration, education, health; and R-T, Arts, entertainment, recreation) are not available in the SBS, and the corresponding employment levels were retrieved from Eurostat Regional Accounts. 6. We excluded three small countries, Cyprus, Luxembourg and Malta, due to missing data for employment. 7. The use of the entire set of 243 regions allows us to obtain a more accurate measure of the proximity parameters, which, however, does not differ remarkably (correlation coefficient ¼ 0.94) when it is computed using the set of 166 regions involved in the S3 policy. 8. The results are robust with respect to the use of the RCA values rather than their transformation into binary values: the correlation with the S3 matrix is 0.15. 9. Similar results are obtained by estimating the conditional probability of selecting an S3 sector given the current RCA based on logit models. 10. This distribution may explain why Deegan et al. (2021) find that regions in their subsample (with all Southern countries and Romania) are inclined to choose related diversification strategies. 11. Table A3 in the supplemental data online reports detailed information on variable definitions and data sources. 12. We have also included additional indicators for the production structure, for example, the specialization in high-and medium-tech manufacturing, but the results remain unchanged. 13. In a preliminary analysis, we replaced the two territorial dummies with a set of 15 national dummies for the largest countries included in the sample. Although the results are remarkably similar, we opt to report in Table 1 the more parsimonious specification with the two territorial dummies described above. 14. All results are available from the authors upon request. 15. For robustness, we also carried out the regression analysis on the composite coherence indicator obtained as the average of the normalized single indicators; the results, not reported to save space, are very similar.