Detecting economic growth pathways in the EU’s lagging regions

ABSTRACT We analyse growth pathways of European Union NUTS-3 regions from 2003 to 2017. We focus on lagging regions, using a taxonomy based on income level and long-run growth rate that combines the Cohesion Policy classification with that proposed under the ‘Catching Up’ initiative. We find that lagging areas can sometimes be found within larger and more prosperous regions, especially in Western Europe. We analyse the role of industrial structure, innovation and inward foreign direct investments as growth drivers, and find that economic growth is associated with different economic dimensions in different types of regions. The NUTS-3 scale of analysis is helpful to inform the design and implementation of development strategies catering to different opportunities at this smaller geographical scale.


INTRODUCTION
Since its creation, the European Union (EU) has been confronted with sharp regional inequality and an economic growth conundrum. Indeed, despite profound political and financial efforts by EU institutions and governments of the member states to redress regional economic inequalities by stimulating growth through successive rounds of Cohesion Policy funds, slow growth and laggardness in some regions have meant the goal of 'reducing disparities between the various regions and the backwardness of the least-favoured regions' (Single European Act, 1986, Article 130a) still remains elusive (Iammarino et al., 2019;Rodríguez-Pose & Ketterer, 2020). 1 Researchers have contributed to this debate both theoretically and empirically. There is a growing evidence base on the efficacy of EU policy interventions targeting the most problematic and lagging-behind regions, thus providing EU and national policymakers with insights on some of the causesand potential remediesof regional inequality.
Empirical works have traditionally focused on regions at level 2 of the Nomenclature des Unités Territoriales Statistiques (NUTS), given the targeting of Cohesion Policy funds at this geographical level and the better availability of data. However, some more recent contributions have highlighted the existence of high heterogeneity and marked differences within large NUTS-2 regions, thus turning the focus of analysis on smaller NUTS-3 regions to illuminate the dynamics of regional heterogeneity (e.g., Becker et al., 2013) and long-run, rooted growth pathways (e.g., Webber et al., 2018). This increase in geographical resolution of the analysis can help identify regional profiles with a finer detail (Geppert & Stephan, 2008;Postiglione et al., 2020), can help identify why some NUTS-2 regions fail in reaching their full growth potential, and can inform the process of policy design at the local level (e.g., Dijkstra & Poelman, 2011). This idea is not specific to scholarly work, rather it has been emphasized more recently also by European policymakers, who have increased their focus below the NUTS-2 level through instruments such as integrated territorial investments that operate at the NUTS-3 level or similar scales.
In this paper, we provide two main contributions to the research literature and policy debate on laggardness in the EU. First, we propose a novel regional taxonomy based on income levels and long-run growth rates. The taxonomy combines the classification adopted under Cohesion Policy with that proposed in 2015 by the European Commission under the 'Catching Up' initiative to target 'low-income' and 'low-growth' regions, and applies it to the NUTS-3 rather than the NUTS-2 level. The rationale for considering NUTS-3 regions is to uncover local economic growth patterns and specificities that are hidden within large and highly heterogeneous NUTS-2 regions. This categorization allows us to uncover specific economic performance patterns hidden within NUTS-2 regions, and to distinguish between two main types of laggardness -'low income' and 'low growth'in line with the most recent view of the European Commission (i.e., the 'Catching Up' initiative).
Second, we use this taxonomy to investigate empirically growth pathways of NUTS-3 regions across the entire EU territory over the period 2003-17, with a special attention to lagging ones. We provide novel evidence by adding to a still relatively scarce cross-country literature at the NUTS-3 levelwith a majority of works looking at a single country (e.g., Panzera & Postiglione, 2014) or a reduced number of countries (e.g., Crescenzi & Giua, 2020) and, particularly, by exploiting detailed regional heterogeneity to shed light on systematic differences in correlates of growth across different types of regions. We evaluate the role played by three key economic dimensions: industrial structure, innovation, and inward foreign direct investment (FDI). These three factors have been traditionally identified as key economic growth drivers, and can be influenced directly by local policymakers with ad hoc intervention measures and policies.
By analysing correlates of growth across the EU over a long time span , we aim to inform local policymakers about the growth pathways achieved by similar regions over time. We aim to provide some answers for local policymakers about the realistic pathways to economic growth in their own 'type' of region.
It is worth noting clearly that we do not attempt to attribute a causal interpretation to our empirical results, merely to examine how different types of regions across the EU have behaved in recent times, and highlight existing differences in regional growth determinants, pathways and profiles.
The rest of the paper is organized as follows. The next section develops the conceptual framework underlying the empirical analysis. The third section presents the dataset, the regional taxonomy adopted and the empirical modelling. The fourth section presents some stylized facts on regions' economic geography and dynamism. The fifth section presents the empirical results. We then conclude discussing the main findings and drawing some policy implications.
A disappointing empirical finding is that despite income inequality among EU member states has declined, inequalities among regions both across and within countries have dramatically increased (Iammarino et al., 2019). For example, since 2004, when the last big wave of new countries joined the EU, the average gross domestic product (GDP) per capita of the three poorest EU countriesnamely, Bulgaria, Latvia and Romaniahas risen from 15.7% of the EU average in 2004 to 29.3% in 2018. However, most of that convergence has been driven by economic growth in leading regions within each country, usually the capital cities -Sofia achieved a 9.2% average yearly GDP per capita growth rate over the period 2004-17, Riga an 8.3% and Bucharest a 12.5%. 2 Overall, while on average NUTS-2 and NUTS-3 regions have recorded an average yearly GDP per capita growth rate of 2.9% and 2.8%, respectively, over the period 2004-17, there is a much larger variation across NUTS-3 regions than NUTS-2 regions. In particular, the 28 EU capital city-regions achieved a 4.4% average yearly GDP per capita growth rate, while non-capital city urban regions and rural regions achieved a 2.6% and a 3.2% growth rate, respectively (see Table A1 in the supplemental data online). Heterogeneity within NUTS-2 regions becomes more understandable once their large size is recognized. NUTS-3 regions in the EU have a mean population of 373,000 people, which is broadly equivalent to a medium-sized city, while NUTS-2 regions have a mean population of 1.8 million people, which is larger than several EU member states (see Table A2 online).
In line with this evidence, some more recent empirical works have increased their geographical resolution of analysis from NUTS-2 to NUTS-3 level, in order to leverage the heterogeneity within NUTS-2 regions and thus to provide a deeper picture of growth performance across small EU territories. Among prior studies, Becker et al. (2010Becker et al. ( , 2012 analyse the effects of Cohesion Policy on income per capita growth, and highlight large heterogeneity with respect to the eligibility threshold for Objective 1 funding across NUTS-3 regions within the same NUTS-2 region, meaning that high-and low-income NUTS-3 regions coexist within the same larger spatial unit. Gagliardi and Percoco (2017) find that Objective 1 funding has highly heterogeneous effects on GDP per capita growth across urban and rural NUTS-3 regions, while Percoco (2017) highlights heterogeneity in the effects of Cohesion Policy with respect to the development and weight of the services sector across NUTS-3 regions. Similar insights emerge looking at economic growth and convergence. For example, Geppert and Stephan (2008) find heterogeneous processes in income per capita growth related to NUTS-3 regions' urbanization degree, and Butkus et al. (2018) highlight sharp heterogeneity in convergence between urban and rural territories.
This substantial variation in economic performance within NUTS-2 regions makes it challenging to reach reliable conclusions on economic growth at the NUTS-2 level. NUTS-2 regional boundaries hide the coexistence of high and low income, fast and slow growth, and more or less dynamic NUTS-3 regions. This heterogeneity may also help explaining why development strategies and growth policies targeting NUTS-2 regions have partially failed in reducing inequality, promoting sustained growth and pushing convergence across territories, as they might have picked also 'accidental winners' for Cohesion Funds and overlooked local specificities to be leveraged (Gagliardi & Percoco, 2017).
Thus, in a context of large government spending and unevenly distributed results, it becomes important to understand better the dynamics of regional economic growth, and the different growth pathways undertaken by different types of regions, in order to uncover the causes of persistent laggardness and growth differentials in the EU and to provide local policymakers with information about their own regional profile in order to maximize the growth returns of policy interventions. In fact, policies and investments that work in leading and fast-growing regions may not work in lagging and slow-growing ones, and the need for guidance may be strongest among lagging NUTS-3 regions, which cannot find many examples of success or turnaround among their peers.

Dataset
We analyse economic growth at the NUTS-3 level by focusing on three key economic dimensions: industrial structure, innovation and inward FDI. The motivation for this is threefold. First, the literature has identified industrial structure, innovation and inward FDI as key growth-engine factors (e.g., Crescenzi, 2005;Menghinello et al., 2010;Percoco, 2017;Webber et al., 2018). Second, local policymakers can influence these dimensions with measures stimulating a particular industrial sector, promoting firms' innovation capability or attracting foreign companies. Third, from a practical viewpoint, data are available on these three dimensions at the NUTS-3 levelindeed, EU statistical sources provide information on a large set of variables for NUTS-2 regions, but only on a reduced number at the NUTS-3 level.
We collected data to cover the longest possible time period, namely from 2003 to 2017. First, data from Eurostat's Regio database on GDP, population, employment, land area and sectoral gross value added (GVA) for agriculture, industry, construction, market services and nonmarket services. 3 Second, microdata on patents filled under the Patent Co-operation Treaty (PCT) drawn from the REGPAT database (Organization for Economic Co-operation and Development -OECD), that have been aggregated at the NUTS-3 level by priority year and inventor's residence using the fractional count criterion. Third, data on inward 'greenfield' FDI drawn from the fDi Markets database (Financial Times), that collects information on year, destination region and business activity run in the host economy for individual investment projects. 4 The cleaning procedure left a sample of 1321 NUTS-3 regions, which represents the 98.5% of the EU-28 territoryhaving excluded a priori the French Overseas Departments and the Spanish extra-territorial autonomous cities of Ceuta and Melilla. The sample covers entirely all EU member states except Poland, for which data were partially unavailable for 20 out of 73 regions (see Table A4 in the supplemental data online).

Defining lagging regions
Under Cohesion Policy, NUTS-2 regions are classified according to their GDP per capita level as 'more developed' (GDP per capita over 90% of the EU average), 'transition' (GDP per capita between 75% and 90% of the EU average) and 'less developed' (GDP per capita less than 75% of the EU average). An additional classification was developed under the European Commission's 'Catching Up' initiative when it was launched in 2015 to provide technical assistance to support development of a subset of lagging NUTS-2 regions classified as 'low income' (GDP per capita under 50% of the EU average in 2013) and 'low growth' (GDP per capita under 90% of the EU average in 2013, and not converged towards the EU average between 2000 and 2013).
We propose a taxonomy that categorizes all EU regions with special attention to heterogeneity in income level and growth performance amongst 'lagging' NUTS-3 regions. The taxonomy combines the two abovementioned classifications: on the one hand, the Cohesion Policy classification which compartmentalizes regions by income category, and, on the other, the 'Catching Up' classification which differentiates the two main problems constituting laggardness. In principle, either of the two classifications could be transposed from the NUTS-2 to the NUTS-3 level, but some hybrid of the two is necessary to compare both types of lagging regions with non-lagging regions. To combine them into a new hybrid classification, we first classify NUTS-3 regions with respect to income level. We calculate the average yearly GDP per capita For the sake of consistency with the Cohesion Policy's 90% threshold value for 'more developed' regionscorresponding to 'high-income' regions in our taxonomywe set also the threshold value identifying the high-growth performance in the long run at 90%. 5 We then develop our taxonomy based on income level and long-run growth rate criteria starting from the identification of the two types of 'laggardness' considered by the 'Catching Up' initiative, namely regions that are lagging behind due to a 'low-income' problem, and regions that are lagging behind due to a combination of a 'middleand low-income' problem and a 'low-growth' problem. Thus, 'lagging low-income' (LLI) regions are those with an average yearly GDP per capita lower than 50% of the sample average, but a high long-run GDP per capita growth ratei.e., with GDPpc , 90% of the sample averages. These two regional categories represent our targets in terms of income-and growth-related laggardness, respectively, consistently with the 'Catching Up' initiative.
We then categorize the remaining regions not identified as 'lagging behind' into three different and mutually exclusive types in order to exploit further cross-region heterogeneity, and identify comparable regional profiles. First, we classify as 'high income, high growth' (HIHG) those high-income regions that have also recorded a high long-run GDP per capita growth ratei.e., with , 90% of the sample averages. Regions belonging to this profile, although they cannot be considered properly as 'lagging behind', deserve certain attention as their slow-growth dynamics could potentially lead to a sustained reduction in income level in the long run. Finally, we classify as 'middle income, high growth' (MIHG) those middle-income regions that have recorded a high long-run GDP per capita growth ratei.e., with and DGDPpc 2003−2017 r ≥ 90% of the sample averages. Regions belonging to this profile represent dynamic territories that, through sustained economic growth, could potentially improve their relative position in the territorial distribution of income.
Formally, we can define the following categorical variable capturing the different profiles of NUTS-3 regions according to our taxonomy: Therefore, the combined classification allows us to focus our subsequent analysis on the two main types of lagging regions (i.e., LLI and LLG), while providing comparator results for regions with higher income and/or growth performance. Table 1 summarizes our classification, while Figure 1 maps the spatial distribution of NUTS-3 regions according to our taxonomy. Two striking features can be noted about Figure 1. First, there are some broad geographical 'groupings' across the EU: LLI regions are mostly located in post-2004 enlargement countries; LLG regions are mostly in the Southern Mediterranean area; and MIHG regions are often found as capital cities and hinterlands, such as Bucharest, Budapest or Riga. Second, there is a very widespread heterogeneity within NUTS-2 regions across almost all of Western Europe. For example, in Germany, the Brandenburg region around Berlin contains both extremes in our typology: two HIHG areas, Uckermark and Dahme-Spreewald, and two LLG areas, Oberhavel and Märkisch-Oderland. Across Western Germany, most NUTS-2 regions contain at least one lagging NUTS-3 area. In France, almost all NUTS-2 regions exhibit stark inequality between HIHG areas and LLG areas. NUTS-3 regions belonging to all regional profiles can be found in almost all Austrian and Belgian NUTS-2 regions. In Spain and Portugal, some NUTS-2 regions are homogenous, but most combine low-growth and high-growth NUTS-3 regions. In Sweden and Finland, the NUTS-2 regions are a patchwork of high-growth and low-growth NUTS-3 areas. The exceptions to this pattern are found mainly in the newer member states of the EU (e.g., Poland, Romania, Bulgaria and the Baltic States) in which NUTS-2 income and growth profiles are more homogenous, and in Greece and Ireland, which are predominantly lagging and leading, respectively. Table 2 provides details on the distribution of NUTS-3 regions according to our taxonomy. LLI regions represent the 14.5% of the sample, while LLG regions represent 19.6%. HIHG regions represent 16.4%, HILG regions represent 40.4% and MIHG regions represent 9.1%. 6 Moreover, almost all EU member states show a relatively high within-country regional variability in average yearly GDP per capita growth over the period 2003-17 (see Figure A1 in the supplemental data online). In particular, 96.9% of LLI NUTS-3 regions are located in post-2004 enlargement countries, while only 3.1% of LLG NUTS-3 regions can be found in those same countries. In other words, lagging regions are a widespread problem, and the two types of laggardness highlighted by the European Commission in the 'Catching Up' initiative emerge as different and geographically well-defined, thus asking for specific analysis and policy interventions.

Empirical modelling
We analyse economic growth pathways of NUTS-3 regions by explicitly accounting for the heterogeneity related to the different profiles identified by our taxonomy through the following empirical growth equation estimated via a two-way fixed effects (FE) estimator: where the dependent variable for region r = 1, . . . , 1321 at time t = 2003, . . . , 2017 is defined as the log-difference of GDP per capita between t and t − 1. The righthand side of equation (2) includes the vector X c rt−1 of time-varying region-specific explanatory variables, and a series of interaction terms between the categorical variable capturing a region's profile (Taxonomy r ) and each explanatory variable included in the vector X c rt−1 aimed at evaluating the heterogeneous association between economic growth and each individual growth driver across regions of different profile. Specifically, the vector X c rt−1 includes controls for growth-initial GDP per capita (GDPpc rt−1 ) in logarithmic form; short-run population dynamics, captured by a dummy variable taking a value of 1 if a region has recorded a weakly positive change in population between t and t − 1, and 0 otherwise (NonNegative Population Change rt ); and employment density, defined as the logarithm of employment/km 2 , to capture agglomeration-related forces (Employment Density rt−1 ). It also includes the explanatory variables of interest for industrial structure, innovation and inward FDI. We proxy the industrial structure of a region through a set of log-transformed variables defined in terms of sectoral share of regional GVA with respect to agriculture (Share GVA Agriculture rt−1 ), industry (Share GVA Indusry rt−1 ), construction (Share GVA Construction rt−1 ), market services (Share GVA Market Services rt−1 ) and non-market services (Share GVA NonMarket Services rt−1 ). Regions' innovativeness is captured by the log-transformed fractional number of PCT patents per 100,000 inhabitants (Patents rt−1 ). The role of inward FDI is captured by the log-number of investments set up in a region per 100,000 inhabitants (Inward FDI rt−1 ). Finally, the right-hand side of equation (2) includes the terms d r and z t denoting region and year fixed effects, respectively; and the error term 1 rt . 7 We also modify equation (2) by replacing the variable capturing the log-number of inward FDI per 100,000 inhabitants with a categorical variable (Max Inward FDI rt−1 ) capturing whether a region has received FDI in a certain year and, if so, the business activity that has characterized the highest number of investments set up. We classify business activities according to the fDi Markets taxonomy as: headquarter; innovation; production; logistics, distribution, and transportation; and marketing and sales. 8 This exercise aims at evaluating whether regions that have received FDI have registered a 'growth premium' with respect to non-receiving ones, and if this premium is associated with a particular business activity run by the multinational company in the host economy. 9

ECONOMIC GEOGRAPHY AND DYNAMISM OF REGIONS
As previously discussed, Cohesion Policy has been more effective in reducing disparities among countries than regions. However, since the early 2000s, inequality has started to rise again at both country and regional levels after a successful reduction during the 1990s, and it has risen among regions within countries in particular (see Figure A2 in the supplemental data online).
Increasing inequality among regions could possibly be the result of a Cohesion Policy that has targeted also 'accidental winners' (Gagliardi & Percoco, 2017)i.e., NUTS-3 areas that would not be eligible for policy Detecting economic growth pathways in the EU's lagging regions 45 support but did in fact receive it because the eligibility criterion was applied at the NUTS-2 level. Indeed, 'good' and 'bad' performing NUTS-3 regionsdefined in terms of average yearly GDP per capita growth ratecoexist within the same NUTS-2 region (see Figure A3 and Table A11 in the supplemental data online).
Turning to the economic performance of lagging regions (see Table A12 in the supplemental data online), we observe how LLI regionsi.e., the relatively poorest regions in the EU but able to record high economic growthhave grown, on average, about 1.7 times more than MIHG ones over the period 2003-17, and about 1.9 times more than HIHG regions. By contrast, LLG regions have grown about 5.3 times less than LLI regions, and about 1.3 times less than HILG ones. The latter evidence is a special cause for concern, that is, the existence of a group of 259 regionsrepresenting the 19.6% of the samplethat are relatively 'poor'i.e., characterized by middle-and low-income levelsbut are not growing.
The varied performance of NUTS-3 regions is shown particularly in their trajectories after the 2008 Great Recession. Considering such an exogenous shock as a cutting point, we observe how none of LLI regions has recorded a decline in the average yearly GDP per capita between the pre-crisis period 2003-07 and the subsequent period 2008-17 (see Table A13 in the supplemental data online). By contrast, the share of 'declining' regions equals 11.5% in the case of HILG regions and about 8% in the vase of LLG regions. Indeed, LLI regions have, on average, recorded a better GDP per capita trend than regions in the other categories (see Figure A4 in the supplemental data online). Finally, zooming on LLI and LLG regions, we find that some regions have been able to improve their relative position between 2003 and 2017 (see Figures  A5 and A6 online).
These stylized facts highlight, first, how 'rich' and 'poor' territories, as well as more and less dynamic ones, coexist within the EU. Second, the persistent gap in income level for 259 regions, which are middle and low income but are not growing, shows that Cohesion Policy has not yet succeeded in addressing challenges in substantial proportion of lagging regions. In this respect, the NUTS-3 lens on regional economic performance can help identify places that need additional support, and could represent an important step forward for the design, implementation and, consequently, efficacy of Cohesion Policy. Third, there are some successful cases from which we can learn, that is, lagging regions that have achieved a sustained growth in income level over a long time period (2003-17).
Detecting economic growth pathways in the EU's lagging regions

EMPIRICAL RESULTS
In this section we present the results of the two-way FE estimation of equation (2) and its modified version accounting for FDI-related business activities. 10 Table 3 reports the estimated marginal effects of each explanatory variable for each regional profile defined according to our taxonomy. Specifically, marginal effects are obtained as partial derivatives of each explanatory variable evaluated at the different values (i.e., regional types) of the regional taxonomy categorical variable with which it is interacted. This allows us to evaluate whether and to what extent the relationship between economic growth and each individual growth driver varies across regions of different types.
Looking at specification (1), we find that all types of regions are experiencing a convergence process, but also that different growth pathways exist for different types of regions. Looking at the control variables included in the empirical growth equation, we find that short-run population dynamics does not matter for economic growth except for HILG regions, while it emerges as negatively associated with economic growth in MIHG regions. By contrast, employment densityas a proxy for agglomeration forcesis positively associated with economic growth in all but HILG regions. Specifically, we estimate that a 1% increase in employment density is associated with an increase of GDP per capita growth equal to 0.11% in HIHG regions, 0.17% in MIHG regions, 0.05% in LLI regions and 0.09% in LLG regions. Looking at the explanatory variables of interest, we find, first, that a 1% increase in the agriculture share of GVA is associated with an increase in economic growth equal to 0.55% in HILG and to 0.36% in LLG regions; by contrast, we find that it is associated with a decrease of GDP per capita growth equal to 0.75% in HIHG and to 0.15% in LLI regions. Increases in the industry share of GVA are positively associated with economic growth in all types of regions but HIHG ones; specifically, we estimate that a 1% increase in the relative weight of industrial production is associated with increases of GDP per capita growth equal to 0.05% in HILG regions, 0.04% in MIHG regions, 0.03% in LLI regions and 0.06% in LLG regions. A 1% increase in the construction share of GVA is associated with a 0.01% increase in economic growth in HIHG regions, with a 0.05% increase in HILG regions and with a 0.06% increase in LLG regions, while it leads to a 0.03% decrease of GDP per capita growth in LLI regions. We estimate that a 1% increase in the market services share of GVA is associated with a 0.03% increase in economic growth in LLI regions, with a 0.09% increase in MIHG regions and with a 0.1% increase in HILG regions; by contrast, we find a positive but statistically negligible association in the case of both HIHG and LLG regions. Interestingly, economic growth in all but HIHG and MIHG is lowered by increases in the non-market services share of GVA. In particular, we estimate that a 1% increase in the relative contribution of non-market services to total GVA is associated with a decrease of GDP per capita growth equal to 0.05% in HILG regions, 0.03% in LLI regions, and 0.09% in LLG regions. Second, innovation is a growth-enhancing factor in all but HILG and MIHG regions. Third, FDI matters for economic growth especially in LLI regions where a 1% increase in the number of FDI per 100,000 inhabitants is associated with a 0.02% increase of GDP per capita growth; however, inward FDI emerges as an economic growth driver also in HILG and MIHG regions.
Overall, our results highlight differences in the way a region's internal industrial structure and innovation capacity, as well as its availability of foreign-owned capital, are associated with economic growth across different regional profiles. More importantly, they suggest how the issue of 'laggardness' cannot be limited to relatively 'poor' regions such as the 'less developed' ones traditionally targeted under Cohesion Policy, rather laggardness should be evaluated systematically by considering differentials in both income level and growth performance. Indeed, as properly proposed under the 'Catching Up' initiative, there are at least two different profiles of 'lagging behind' regions, namely those that are suffering from a potential 'low-income trap' but exhibit good economic growth potential (i.e., LLI regions), and those that, besides being relatively 'poor', also show little growth capacity (i.e., LLG regions). These two types of regions show differences in growth pathways both with respect to more advanced and dynamic territories, and between each other. Indeed, comparison between the two types of lagging-behind regions suggests how their economic growth is driven by different growth-enhancing factors. On the one hand, and despite to a different extent, economic growth in both types of regions benefits from agglomeration forces, industry-type production activities, and innovation capacity, while is harmed by an enlargement of the nonmarket services sector. On the other hand, economic growth seems to be driven by lower value-added activities such as agriculture and constructionin the less dynamic LLG regions, while it seems to be driven mainly by highvalued market services and inward FDI in the more dynamic LLI regions. A possible explanation for the high relevance of inward FDI for economic growth in LLI regions could be that new capital entering through foreign investments is especially important in places with a scarcity of successful entrepreneurs, but characterized by high economic dynamism and growth potential.
Specification (2) in Table 3 reports the estimated marginal effects obtained through the two-way FE estimation of the modified version of equation (2) considering the FDI categorical variable. We focus our attention on the set of results concerning the different types of activities run by multinational companies, as the results concerning all the other variables are consistent with those of specification (1). First, inward FDI emerges a key growthenhancing factor for LLI regions; indeed, economic growth in this type of region is positively associated with three types of investments: production; logistics, distribution and transportation; and marketing and sales. Despite being relatively low-valued activities, their relevance for economic growth in 'poor' but dynamic regions could be explained through both labour market and value chain mechanisms. On the one hand, these types of activities may contribute to substantial job creation. On the other hand, they may favour the establishment of newor the enlargement of existinglocal activities through inter-firm and inter-industry backward and forward linkages, thus contributing to value added generation. Second, and by contrast, only headquarter-related FDI activities seem to be positively associated with economic growth in LLG regions. A possible explanation could be related to the little dynamism of these territories and their economic structure, such that positive spillovers from inward FDI in production-or service-related activities do not materialize through linkages with local firms, and multinational companies exploit only location-related advantages for overseeing local markets. This finding is potentially of keen interest to LLG regions, since it implies their FDI promotion activities may have the most impact on local growth if targeted towards headquarter-related activities. Finally, economic growth is positively associated with innovation-related FDI activities in HIHG regions, with logistics, distribution and transportation FDI in HILG regions, and with headquarterrelated FDI in MIHG regions. 11 Overall, the results highlight substantial heterogeneity in the economic growth returns of inward FDI across the different regional profiles. This evidence reinforces the idea that regional specificities should be accounted for when designing and implementing policy interventions aimed at promoting growth. In the context of inward FDI, it is clear how local policymakers should not rely on 'imitation strategies', rather invest to stimulate the location of those foreign-owned activities that can effectively contribute to economic growth in their type of region. 12 We now turn to the topic of heterogeneity in economic performance related to regions' rural versus non-rural profile. 13 Prior research generally indicates that convergence among countries has been driven by capital cities in many 'poor' member states, and that there is substantial heterogeneity in economic performance between urban and rural areas (e.g., Butkus et al., 2018;Gagliardi & Percoco, 2017). Table 4 reports the estimated marginal effects obtained through the two-way FE estimation of the modified version of equation (2) disentangling FDI-related business activities, and estimated separately for rural and non-rural regions. Despite our focus on the two types of lagging regions, it can be seen that sharp heterogeneity in growth determinants and pathways encompassing regions' industrial structure, innovation capacity, and inward FDI-related activities characterizes rural and non-rural regions across the regional profiles identified by our taxonomy.
In particular, looking at LLI regions, we find that economic growth is positively associated with agglomeration forces, market services, and innovation only in non-rural regions, while rural regions' economic growth is mainly driven by increases in the industry share of GVA which, however, represents a growth-enhancing factor also for non-rural regions. Interestingly, different inward FDI-related business activities play a different role in rural versus non-rural LLI regions. Economic growth in rural LLI regions seems to benefit from FDI in headquarter, production, and logistics, distribution and transportation activities. By contrast, we find in non-rural LLI regions that inward FDI in production and marketing and sales activities matters for economic growth. Similar insights in terms of high heterogeneity characterize also LLG regions. Economic growth in rural LLG regions is positively associated with agglomeration forces, increases in the agriculture, construction, and market services shares of GVA, and with headquarter-related FDI, while it is negatively associated with production-related FDI. By contrast, economic growth in their non-rural counterparts benefits from increases in the industry and construction shares of GVA, as well as from innovation and agglomeration-related forces. Moreover, for this type of LLG regions, we do not find any statistically significant effect associated with inward FDI. 14

DISCUSSION AND CONCLUSIONS
The EU has been characterized by profound internal inequalities since its creation, and these inequalities have been magnified by the last waves of Eastern countries joining the EU from 2004. Despite European and national policymakers increasing the political and monetary efforts to reduce disparities and promote sustained growth in recent years, inequality has actually been rising across regions, especially within each country. Many NUTS-2 regions in Europe, especially in Western Europe, show a substantial internal heterogeneity in income levels or growth rates. It is provocative and useful to consider if the shortcomings of Cohesion Policy in dealing with subnational inequality could be due to policy design problems, particularly the identification of target regions that are too heterogeneous for unified programming of Cohesion Fundsnamely, NUTS-2 regions.
Drawing on this rationale, we have attempted to contribute to the debate on regional economic growth in the EU by looking at NUTS-3 regions to identify the different pathways that have characterized different types of regions, with a special focus on lagging regions. We have relied on a two-dimensional regional classification based on income level and long-run growth performance, which combines the standard Cohesion Policy taxonomy with that recently proposed under the 'Catching Up' initiative, and we have attempted to provide subnational policymakers with novel empirical evidence useful for planning development interventions that are coherent with their specific regional typology.
We find that 14.5% of NUTS-3 regions are LLI, and 19.6% are LLG. These lagging regions are found in 22 out of 28 member states, including the richest ones. What have we learned in this paper about these types of regions? Table 4. Two-way FE estimates by rural versus non-rural regional profile.
Detecting economic growth pathways in the EU's lagging regions LLI regions are almost all (97%) located in post-2004 enlargement countries. They have grown relatively fastan average of 6.7% per year between 2003 and 2017, compared with 2.6% for the other categories. As such, they have been converging with high-income regions. From a sectoral standpoint, economic growth is associated with growth in industrial production and market services, and with a diminished role for agriculture, construction and non-market services. Innovation and inward FDI are both growth-enhancing factors. Interestingly, economic growth in these regions is not strongly associated with employment density, suggesting that growth is not reliant on agglomeration forces and/or larger cities. Thus, policymakers in LLI regions may wish to support industrial transformation that is associated with economic growth, and to encourage innovation and inward FDI.

REGIONAL STUDIES
LLG regions meanwhile are almost all (97%) located in pre-2004 enlargement countries. Policymakers are increasingly concerned about this category of region in which economic development is somehow 'stuck' in a low-growth mode. From a sectoral standpoint, economic growth is associated with growth in agriculture, industrial production, construction, and with a diminished role for non-market services. Innovation is a growth-enhancing factor, but for FDI only headquarter-related investments are growth-enhancing. Thus, policymakers in LLG regions may wish to focus their efforts on tradable sectors (agriculture and industrial production) rather than locally focused sectors (non-market services), encourage more innovation, and target headquarter-related FDI rather than production-, logistics-or sales-related foreignowned activities.
Overall, we find that different types of regions are characterized by different economic growth pathways. As such, 'one-size-fits-all' approaches to policy design will be ineffective. Differences in growth-enhancing factors emerge not only among different types of regions defined in terms of income level and long-run growth performance, but also between rural and non-rural regions belonging to the same performance group. The two types of lagging regions have no uniform pattern in the role of economic sectors in contributing to growth, but show a positive and similar relationship between innovation and growth. Moreover, they show marked differences on the role that inward FDI plays on economic growth. Economic growth in LLG regions shows a positive correlation with construction, industry and, especially, agriculture, while a negligible relationship with market services. By contrast, economic growth is associated with a move away from agriculture among rural LLI regions, where a key role seems to be played by industry and inward FDI.
How can these results inform policymakers? First, they can help local policymakers to take actions for growth coherent with the different types of regions. Regions designing their growth strategies must look at their own profile, endowments, and realistic opportunities. The analytic results for each type of region show the factors that are typically correlated with growth in that type of region.
Local governments can influence several of the variables considered in this analysis, through targeting industries for support, fostering innovation, and attracting foreign companies. Second, the results should alert NUTS-2 level and national governments to the heterogeneity within NUTS-2 regions, and the need for deliberate actions to link leading NUTS-3 areas with lagging ones. Regional strategies need to be designed differently if better outcomes for lagging NUTS-3 regions are desired. Richer NUTS-3 areas can provide opportunities for poorer ones, but this will require an explicit spatial strategy to link these arease.g., through transport infrastructures, accessibility to neighbouring markets, firms' participation in supply chainssince it does not appear to happen naturally. Third, by combining our results with recent evidence about the correlation of inequality and discontent (e.g., Dijkstra et al., 2020), it can be inferred from our results that European and national policymakers should address inequality at the NUTS-3 level as a source of discontent. Inequality is increasing among NUTS-3 regions, such that cohesion in the EU will require renewed and concerted action at this geographical level.
Our final takeaway is that subnational policymakers should pay attention to the specificities of their regions, as replication strategies could not necessarily work everywhere. Examining the experiences of 'successful' lagging regions can sometimes be helpful, especially if focusing on regions with similar economic profiles.
A18 online. All these exercises fully corroborate our main findings. Finally, we replicated specification (2) in Table 3 by specifying the empirical growth equation according to a spatial Durbin model (SDM) specification to account for endogenous interaction effectsthrough the spatial lag of the dependent variable capturing yearly GDP per capita growthand exogenous interaction effects depending on observable characteristics of neighbouring regionsthrough spatial lags of the explanatory variables. We estimated the SDM specification via maximum likelihood by controlling for spatial and time fixed effects, and used a first-order row-standardized binary spatial weights matrix to construct spatially lagged variables. Table A19 online reports the estimated direct and indirect marginal effects for each regional profile obtained from the SDM specification: the estimated direct marginal effects fully confirm the results obtained from the corresponding a-spatial model; the parameter of the spatially lagged dependent variable is positive and statistically significant; evidence for the estimated indirect marginal effects varies substantially across the different regional profiles identified through our taxonomy. 13. The European Commission classifies regions as 'predominantly urban', 'intermediate' and 'predominantly rural'. We consider two regional typologies due to the limited number of regions resulting in the different categories when also considering our taxonomy: ('predominantly') rural versus non-rural (including 'intermediate' and 'predominantly urban') regions. 14. Table A20 in the supplemental data online reports the results for the whole sample of regionsi.e., obtained without accounting for heterogeneity across regions related to our taxonomyby accounting for their rural versus non-rural profile.