The long and winding road to find the impact of EU funds on regional growth: IV and spatial analyses

ABSTRACT We contribute to the analysis of the impact of European Union funds on European regional development. We find that the European funds have a significantly positive effect on regional economic growth in the European Union. This result is obtained both with ordinary least squares (OLS), and with two-stage least squares (2SLS) using the presence of environmentally protected areas as an instrument. Furthermore, we find that interregional spillovers are important: a significant part of the favourable effect seems to take place in nearby regions rather than in the recipient region.


INTRODUCTION
One of the key principles of European integration has always been solidarity.This has been manifested, most prominently, in the funding set aside to support regional development within the European Union (EU): Cohesion Policy, the primary tool of regional policy in the EU's toolbox, accounts for approximately one-third of its budget. 1  Its importance, furthermore, has been increasing over time: from less than 10% of the European Economic Community (EEC) budget in the 1970s and the early 1980s to 32.5% in the 2014- 20 EU budget. 2 This reflects the changing priorities of the EU budget in line with the EU's expansions.Initially, the EEC comprised six founding members who were at similar levels of economic development, so that there was little perceived need for regional redistribution.Rather, on the backdrop of food shortages during the Second World War, the bulk of EU spending was dedicated to supporting agriculture so as to ensure the security of food supply.However, the accession of Greece in 1981, followed by both Spain and Portugal in 1986, resulted in considerable income differentials among the member states and their regions.The Southern European countries were, furthermore, concerned about greater competitive pressure following their entry to the Single Market: the increased allocation of funds to regional policy served to alleviate their fears that they would be adversely affected.These considerations were further strengthened after German Reunification in 1990 and the Eastern enlargements in 2004, 2007 and 2013.Correspondingly, the relative importance of EU funds has increased and the Cohesion budget has almost caught up with the funding earmarked for the Common Agricultural Policy (CAP).
The evidence for the growth-boosting effect of EU regional policy, however, has been rather mixed so far.Some studies do find that Cohesion Policy has had a positive impact on economic growth.Other analyses, however, yield an insignificant or even a negative effect.Broad overview studies by Dall'erba and Fang (2017) and Marzinotto (2012) observe a general lack of consensus in the literature over the sign of the effect of regional policy on growth.Given the considerable amounts of money that the EU spends on Cohesion Policy, it is disappointing that clear and overwhelming evidence of a positive effect remains elusive.
Nevertheless, the absence of evidence of a positive effect does not necessarily imply that no such effect exists.The positive effect could elude the researchers for a number of reasons.First, the analyses can be plagued by measurement errors in data on growth or Cohesion Policy transfers.Second, the estimated impact of EU funds on regional growth is likely to be endogenous either because of omitted variables (such as structural issues and other factors jointly affecting eligibility for regional aid and economic growth) or because of reverse causality (Cohesion payments being shaped by regional economic development).Third, the effects of an investment in any particular region can spill over to other regions.Not accounting for such spillovers could diminish the estimated regional effect of Cohesion Policy.We discuss these issues in detail in the following section.
In our paper, we join the fellow travellers seeking to correctly identify the effect of Cohesion Policy on regional economic growth.We introduce a number of innovations.First, we use an updated and more extensive dataset, with Cohesion transfers and growth measured in annual frequency (most previous studies were limited to data measured for the so-called Programming periods comprising six to seven years).Second, we seek to account for the potential endogeneity of Cohesion Policy by using a novel and previously unused instrumental variable (IV): the presence of environmentally protected areas (under the Natura 2000 programme) as an instrument.Third, we estimate a spatial Durbin model (SDM) so as to capture the potential spatial spillover effects of Cohesion Policy.
Our results suggest that environmental protection, proxied by the fraction of the region's area under protection and the number of protected sites in the region, is positively correlated with the amount of funding that the region receives from Cohesion Policy.The Cohesion Policy funding, in turn, translates into higher regional economic growth.The results of our spatial analysis, furthermore, suggest that Cohesion Policy indeed has important cross-regional spillover effects on growth.
How much bang does the EU get for its Cohesion Policy buck?In the next section, we briefly discuss the existing literature on the impact of Cohesion Policy on regional development.The data we use are described in section 3.In section 4, we explain the construction of our instrument and present our analysis of endogeneity-robust effect of Cohesion Policy on regional economic growth.We then present the result of our spatial analysis in section 5. We summarise our findings and put them into the broader context in the last section.

THE ROAD SO FAR
The EU spends around one-third of its budget on Cohesion Policyregional aid to less-developed regions.This makes Cohesion the second largest expense item, just slightly lower than the CAP.Unsurprisingly, questions have been raised about the merits of dedicating such a large share of EU funds to regional aid.
In principle, Cohesion Policy should stimulate regional development.It prioritises lagging-behind regions, often with structural problems, where it targets areas with significant growth-promoting potential: infrastructure improvements, education, innovation, job creation, environmental sustainability and climate change, social inclusion, and ageing (Barca, 2009).Nevertheless, generally, all member states receive transfers from Cohesion Policy, not only the less affluent ones.Therefore, Cohesion Policy should be seen as relatively uncontroversial and broadly beneficial.Are these optimistic expectations borne out by the evidence?
The existing literature on the effect of EU funds on economic growth paints a rather mixed picture.Some studies find a positive impact on economic growth (Beugelsdijk & Eijffinger, 2005;Bradley et al., 2004;Bradley & Untiedt, 2007;Cappelen et al., 2003;Cerqua & Pellegrini, 2018;Maynou et al., 2014;Radvansky et al., 2015;Rodríguez-Pose & Fratesi, 2004;Venables & Gasiorek, 1999).Others conclude that the effect is either insignificant or even negative (Boldrin et al., 2001;Dall'erba & Le Gallo, 2008;Eggert et al., 2007;Fagerberg & Verspagen, 1996).The meta-analysis by Dall'erba and Fang (2017) reviews the quantitative evidence in 17 studies which together yield 323 estimates of growth elasticities.The average estimate is close to zero at 0.174 and the range of estimates is high: from −7.6 to 6.3.Marzinotto (2012), in another survey, also finds the literature inconclusive, and points out that the effects obtained in empirical studies fall short of those predicted by macroeconomic simulations.
The broad range of findings in the previous literature can be attributed to differences in methodological approaches, or it can reflect heterogeneity in the way such transfers affect regional economies.As for the former, Dall'erba and Fang (2017) observe that the later the research was conducted, the higher was the estimated impact of EU funds.This could suggest that there is a learning effect on the side of the researchers whose methodological approaches improve over time, or that the member states learn over time how to use European subsidies better and more efficiently.
It is reasonable to expect that the member states allocate the resources received from EU funds first to the projects with the highest rate of (social) return, followed by those with a lower return, etc. Member states with more generous allocations of funds should then display lower average rates of return.The effect of EU funds need not be monotonous, however.For example, funding may need to reach a minimum critical mass to have a measurable positive impact on regional growth (Becker et al., 2012).In line with this reasoning, Becker et al. (2012) estimate a non-linear relationship between Cohesion transfers and regional growth and find that it peaks when the EU funds account for just above 1% of the region's gross domestic product (GDP).Fiaschi et al. (2018), in a later study, obtain a higher optimal level of around 3%.Dicharry (2020) puts forward a different argument: the return to regional policy can depend on the speed of implementation.The reason for this is the EU rules that stipulate that regions lose their Cohesion Policy allocations if the funds are not used within a specified period of time.This may induce them to apply for funding to finance projects that are simple and easy to implement.Such safe projects, however, may be lacking in terms of originality and innovative capacity.Dicharry (2020) finds that while the overall impact of Cohesion Policy on regional development is positive, it indeed declines with the speed of absorption.He suggests that inefficiently fast absorption of EU funds may be a factor behind the comparatively poor growth performance of Southern European countries, in particular.
Several studies argue that the effect of EU funds on regional growth is conditional on other outcomes.The impact of regional policy has been found to be reinforced by sound (national or local) institutions (Becker et al., 2013;Casula, 2021;Rodríguez-Pose & Garcilazo, 2015;Rodríguez-Pose & Ketterer, 2020), quality of human capital (Becker et al., 2013), decentralisation of decision-making from the national government to the local level (Bähr, 2008;Védrine, 2020), quality of local government (Crescenzi et al., 2017;De Matteis et al., 2021), and by the effective oversight of how the EU funds are used (Wostner & Šlander, 2009).Hence, the absorption capacity of the recipient regions can crucially depend on the presence (or absence) of specific conditioning factors in the recipient region.
Other contributions find evidence of heterogeneity across different sectors of the economy, or across EU funding objectives.Mohl et al. (2008) and Fiaschi et al. (2018) argue that the positive effect of Cohesion Policy is primarily limited to the Convergence Objective (which was previously known as Objective 1 and targets regions whose GDP per capita is less than 75% of the EU average).Percoco (2017) observes that EU funds disproportionately flow into the service sector, where their effect is higher when this sector is still relatively underdeveloped.Rodríguez-Pose and Fratesi (2004), in turn, conclude that investments in human capital and education have a positive impact on economic growth, in contrast to spending on infrastructure or business support.Such differences can matter because member states have different priorities for the use of European fundsthe more developed countries tend to invest more in innovation and education while the less developed ones prioritise infrastructure (Berkowitz et al., 2020).Moreover, the less developed countries receive mainly Convergence Objective funding whereas if the more developed member states receive funding, it is under other objectives.Such differences could therefore help explain why the estimated effects funds vary.
Another potentially important explanation for the weak and mixed evidence on the effect of European funds on growth is endogeneity.This can even stem from the institutional set-up of Cohesion Policy, with less developed regions being given more funding.Consider two regions, both initially at just under 75% of the EU average GDP per capita, and thus eligible for Convergence Objective funding.If one region is stricken by structural problems (such as being dominated by a declining industry), it will tend to report a lower rate of growth.As a result, this region will remain eligible for Cohesion Policy funding longerand it may even receive additional funding to help it overcome its structural issues.The other region, in contrast, will quickly lose the eligibility for Convergence Objective funding.Being free of structural issues, it may continue growing faster than the first region even after crossing the Convergence Objective threshold.This would result in a negative correlation between fiscal transfers and growth, which would be spurious as it would be driven by an omitted factorthe presence of underlying structural problems in the first regionrather than by a causal negative effect of regional policy.Dall'erba and Fang (2017) observe that most studies on the effectiveness of Cohesion Policy tend to ignore the challenge posed by endogeneity .This is largely owing to the difficulty with finding suitable instruments for Cohesion Policy.One exception to this is Dall'erba and Le Gallo (2008), who use geographical distance and travel time to Brussels as IV.However, as these instruments are constant over time, they can be used only in a crosssection analysis.In their analysis, Dall'erba and Le Gallo conclude that the process of convergence among European regions is ongoing, but the EU funds do not play much role in it.
Recently, a number of papers sought to tackle endogeneity by means of quasi-experimental methods.Becker et al. (2010Becker et al. ( , 2018)), Pellegrini et al. (2013), Gagliardi and Percoco (2017), Ferrara et al. (2017), Percoco (2017) and Cerqua and Pellegrini (2018) take advantage of the fact that the bulk of EU funds is distributed in line with the Convergence Objective (recall that under this Objective, a large share of Cohesion funds is set aside for regions whose GDP per capita is below 75% of the EU average).This makes it possible to apply the regression discontinuity design (RDD): arguably, regions that are just below and just above this threshold should be very similar to each other in every respect other than their eligibility to receive Cohesion transfers.Giua (2017) and Crescenzi and Giua (2020) follow a similar approach, but consider municipalities and areas on either side of the geographical borders between Convergence Objective regions and adjacent regions that do not qualify for funding under this objective.The studies using RDD find that Cohesion Policy payments have a positive endogeneity-robust impact on growth and, in some instances, also on employment.Again, there is evidence of heterogeneity across regions and countries.Crescenzi and Giua (2020) find especially robust growth effects for German regions, compared with less compelling evidence for Italy and Spain.Gagliardi and Percoco (2017), in turn, conclude that the benefits are especially pronounced in rural regions close to major urban agglomerations.
Other quasi-experimental studies rely on finding (or creating) a suitable control group of regions that are similar in every aspect other than being beneficiaries of regional policy funding.Barone et al. (2016) and Di Cataldo (2017) apply the synthetic control method to construct so-called synthetic clones for the regions that have lost the Convergence Objective status (in Italy and the UK, respectively).The clones are constructed as weighted averages of other regions, selected to ensure that the original region and its clone are as similar as possible in the The long and winding road to find the impact of EU funds on regional growth: IV and spatial analyses 585 pre-treatment period.This allows them to compare the actual performance of the region with its hypothetical performance had it remained eligible for Cohesion funding.Both studies conclude that the loss of the Convergence Objective eligibility is associated with a deceleration of growth.Finally, Bachtrögler and Hammer (2018) and Bachtrögler et al. (2020) use propensity score matching at the firm level to compare the performance of beneficiaries of Cohesion Policy funding with otherwise similar firms without such subsidies.They find that European funds help firms grow bigger; however, the recipient firms do not seem to become more productive as a result of this funding.Yet another explanation for weak or insignificant results could be ignoring regional spillovers of the Cohesion Policy (Berkowitz et al., 2020;Hagen & Mohl, 2011).EU funds are disbursed for projects located in specific regions.Most studies therefore seek to identify an impact of such expenditure on the recipient region.The EU, however, is a closely integrated market where goods, services, labour and capital readily cross regional or national boundaries.EU-financed spending, therefore, can also boost output in a broader geographical area if the investments are realised by firms or with deliveries from nearby regions.Similarly, the benefits of these investments, when completed, can accrue to other regions too.However, the results of this part of the literature, are once again mixed.On the one hand, Fiaschi et al. (2018) propose a growth model that stipulates that EU funds can affect regional growth both directly (in the recipient region) and indirectly, by generating spatial externalities.They then present empirical evidence that such spatial spillovers were indeed important in the EU-12 countries.Similarly, Bachtrögler-Unger et al. (2022), using satellite-captured light emissions, find a positive impact of Cohesion Policy both in the recipient regions as well as in the neighbouring regions.On the other hand, Le Gallo et al. (2011) conclude that the Cohesion transfers do not contribute to economic growth in a spatial-spillover model estimated for the EU as a whole, although they do appear to boost growth in peripheral regions of the EU (the UK, Greece and Southern Italy).De Dominicis (2014) similarly fails to obtain a significant impact for the EU as a whole but finds a positive effect for the less developed regions of the EU-15.Dall'erba and Le Gallo (2008) likewise find no overall effect of EU funds on economic convergence, but they identify convergence from core to peripheral regions.Breidenbach et al. (2019) present even more disappointing finding based on a model incorporating spatial spillovers: their results suggest that the EU funds may have a negative effect on regional growth.
With our analysis, we build on and extend the existing literature by contributing further to the quest to determine the nature and sign of the effect of Cohesion Policy on growth.We use a more extensive data set than most previous studies, including both old and new member states.
Our study is also one of very few based on annual data.We proceed along two avenues.First, we consider the impact of Cohesion Policy funding on recipient regions in ordinary least squares (OLS) and two-stage least squares (2SLS) framework.For the latter, we use the presence of environmentally protected areas in each region to construct IV for Cohesion funding.An advantage of this approach is that it allows us to account for the intensity of regional aid receivedrather than using a dummy variable to measure qualifying (or not) for a specific objective.Furthermore, it allows us to include funding under all objectives rather than focus only on one objective of Cohesion Policy spending (such as the Convergence Objective).Second, we estimate a spatial model in order to allow for spatial spillovers.We apply these two approaches one at a time, to demonstrate how the results differ from the baseline OLS estimates.

DATA
Given the nature of our analysis, we have to combine data from a number of different sources.First, we use the annual receipts of EU funds at the level of NUTS-2 regions as reported by the European Commission. 3 Until recently, only the total payments over whole programming periods were available at the NUTS-2 level.To the best of our knowledge, ours is one of the first papers to use these newly available annual data for an economic analysis of the Cohesion Policy impact. 4Only data covering the Cohesion Policy payments during the last three completed programming periods, 1994-99, 2000-06 and 2007-13, have been made available in annual frequency so far, which determines our choice of time period. 5Funds allocated to a region in any given year have to be spent during that year and/or the next two (occasionally three) years.Our data record the annual payments (rather than commitments) of Cohesion Policy funds.Because of this, some of the funds committed during the last two to three years of the 2007-13 programming period were only paid out during 2014-16.We keep the information for 2014 (when most of the spending is likely to be funds allocated during the preceding programming period), but drop 2015 and 2016 from our analysis. 6 Finally, we use only the total amounts of EU funds received by each region, without breaking it down further by categories of spending. 7 We complement the Cohesion Policy data with regional economic statistics, provided by the Cambridge Econometrics European Regional Database 2016: regional output per person, population growth and investment in physical capital. 8We also use the Worldwide Governance Indicators compiled by Kaufmann et al. (2011).These cover six areas: Voice and accountability; Political stability and lack of violence; Government effectiveness; Regulatory quality; Rule of law; and Control of corruption.We collapse them into a single composite indicator by means of principal component analysis.The governance indicators are available from 1996 to 2016 at the country level only.Furthermore, they are available in yearly frequency only since 2002.Therefore, we impute the missing years using a Monte Carlo simulation of the regression of the composite index on a polynomial of time.
We thus have data from 1997 until 2014 at the NUTS-2 level for 272 regions altogether (in their 2013 definitions).An overview of the main variables (including those related to our instruments discussed in the next section) is provided in Table 1, while Table 2 displays the descriptive statistics.Table 3, in turn, reports the correlations between the various variables.The regional distribution of the Cohesion Policy spending is depicted in Figure 1.It shows averages over three-year windows at the beginning of the period considered , in the middle and just after the Eastward EU enlargement (2004-06), and at the end .The maps clearly show that the geographical focus of Cohesion spending has shifted over time.Traditionally, Cohesion Policy mainly benefited regions at the periphery of the EU-15: Southern Europe, East Germany, the northern UK and Ireland.After the Eastern enlargements in 2004 and 2007, the bulk of funding was redirected towards the new member states in Eastern and South Eastern Europe.9

Constructing the instrument
To deal with endogeneity of regional aid, we require instruments that are correlated with the transfers that regions receive under Cohesion Policy, but uncorrelated with the error term in the growth equation.We propose to use the presence of environmentally protected areas (designated as such under the EU's Natura 2000 programme) in all EU NUTS-2 regions.In this subsection, we explain how the Natura 2000 programme works, what are the practical implications of bestowing a protected status on an area, and why we believe that environmental protection can serve as an instrument for Cohesion Policy payments.
The Natura 2000 programme was established in the EU in 1992 to grant protected status to endangered species and their habitats.The protection granted by the Natura 2000 programme is enshrined in two EU Directives: the Birds Directive and the Habitats Directive. 10At present, the Natura 2000 programme comprises over 27,000 terrestrial and marine conservation areas, covering some 18% of the EU's land area and 8% of its sea surface.This makes Natura 2000 the largest network of protected sites in the world.
For an area to receive protected status, member states first identify sites that they deem eligible for protection under the Natura 2000 rules. 11These are then proposed for inclusion in the Natura 2000 network.The merits of the member states' proposals are evaluated by the European Commission with input from the European Environment Agency.The member states' proposals and the Commission's decisions are to be based strictly on objective scientific criteria: the two Directives explicitly list the species and habitat types which are considered to be of European importance and eligible for protection.If approved, the sites receive protection under the Natura 2000 programme.
The fact that the species and habitats eligible for protection are determined centrally by the EU makes it less likely that regions receive conservation status based on national or local considerations, including economic motives.Such a feedback effect from economic conditions to decisions on conservation status could invalidate our instruments.Another issue would arise if some regions were better at applying for Natura 2000 recognition than others.For example, this could be the case if the administrative capacity or the quality of bureaucracy were The long and winding road to find the impact of EU funds on regional growth: IV and spatial analyses 587 correlated with the success of applications for Natura 2000 status.Such local-level factors could also be correlated with the local economic performance. 12However, such local determinants are likely to be highly persistent over time so that they should be picked up by the regional fixed effects (see the Methodology subsection).Moreover, once a site receives the Commission's endorsement, it generally keeps the protected status whereas both the regions' receipts from Cohesion Policy and local factors can change over time.After a proposed site has been approved, the member state designates it as either a Special Protection Area (SPA) or a Site of Community Importance (SCI). 13The member states then have an obligation to maintain the sites, and even restore them where required, to ensure the long-term survival of the protected species or habitats.Note that candidates for EU membership are expected to submit similar proposals during the accession negotiations so that new member states can already have several protected sites by the time they join the EU.
Given the bureaucratic nature of the process, the Natura 2000 network has been only growing gradually.This is documented in Table 4, which reports the increase in the size of the Natura 2000 network for the average NUTS-2 region during the period considered, 1997-2004.The first column reports the change in the area under protection (as share of the overall area), while the second column shows the number of sites added to the network.The average NUTS-2 region granted new protection to some 15% of its area during the full period, and this amounts to 65 new protected sites per region on average.Looking at the individual years, the area under protection grew the most in 2004 (by 3.0%) and 1998 (by 1.6%), while the number of protected sites increased the fastest in 2004 (by 13 new sites per region on average) and 2000 (11 sites per region added).Figure 2 shows how the size of the network expanded in individual regions.It depicts the proportion of the area of Natura sites to the total area of each region over the same three-year windows as those in Figure 1 : 1997-99, 2004-06 and 2012-14.The size of the Natura 2000 network has increased markedly in the Mediterranean countries (especially Spain, Greece and France) and the new member states (most notably Bulgaria, Romania, Slovakia and Poland).By 2014, the Natura 2000 network was represented in the vast majority of EU regions. 14 It is important to note that most Natura 2000 sites are not free from human presence or activity.Some 50% of the protected area is forested (but only around 13% is classified as wild), with most of the remaining area (around 40%) being agricultural ecosystems. 15Some sites are of urban or industrial character: examples include ports, (disused) mines and roofs of old buildings.The protected sites also vary considerably in size: from 1 m 2 (an abandoned mine in Slovakia which serves as a nesting site for several protected species of bats) to vast national parks and marine conservation areas.
To construct the instrument, we use the information on the location and size of Natura 2000 areas. 16Based on the geospatial data from the European Environment Agency and Eurostat, 17 we match each site (excluding marine areas) with NUTS-2 regions based on the coordinates of the centroids of the Natura 2000 sites.In some cases, the centroid of a protected site appears to lie in a different region because of the specific shape of the site and the location of administrative borders of regions.For that reason, we web-scraped the publicly available application forms of the Natura sites and corrected the mistakes that arose in this way.Lastly, we aggregate the individual sites in each NUTS-2 regions, yielding two summary regional figures: the proportion of each region's area that is protected by the Natura 2000 programme, and the number of environmentally protected sites per regions.These two variables constitute our instruments.Since they are different expression of the same underlying information, we use them one at a time rather than jointly.
The regions with a high density of Natura 2000 sites tend to receive more EU funds: the correlation between the proportion of region's area that is protected by Natura 2000 and the ratio of Cohesion transfers to GDP is 0.33: positive although not very high. 19On the other hand, the correlation between the area under protection and the growth of output per capita is close to zero (and in fact moderately negative) at −0.10 (Table 3).While the exclusion restriction cannot be formally tested, this suggests that our instrument is correlated with the endogenous variable but not with the dependent variable.
We can see several potential reasons why there should be a correlation between environmental protection and Cohesion Policy spending.First, the presence of environmentally protected sites limits the nature and scope of industrial activity and infrastructure building in the region.Because of their protected status, human habitation and economic activity within protected sites have to be environmentally sustainable and are subject to restrictions.This means that some activities may not be allowed at all, such as industrial or agricultural production resulting in excessive pollution, noise, or habitat destruction.Other activities may be possible subject to specific constraints.For example, a new road through a protected area may have to be built according to more demanding specifications (such as adopting noise-reducing measures  The long and winding road to find the impact of EU funds on regional growth: IV and spatial analyses 589

REGIONAL STUDIES
or having additional features to allow safe and unhindered access to a nesting site of a protected species) or it may have to be planned differently (such as going around the nesting site instead of taking the shortest possible route).
Hence, environmental conservation may preclude some investment and activities, while others are likely to be more costly on account of the additional requirements.
The net effect on the volume of investment can be either positive or negative.Second, the regions hosting protected sites are likely to have better access to Cohesion Policy funding.The Natura 2000 network has some dedicated direct funding but this is limited in nature: Gantioler et al. (2014) estimate that the annual funds available to finance Natura 2000 sites are just under €6 million.The primary responsibility for financing the conservation activities in Natura 2000 sites thus lies with the member states.Nevertheless, the member states have various co-financing options from European sources.Specifically, the European Regional Development Fund (ERDF), Cohesion Fund (CF), European Agricultural Fund for Rural Development (EAFRD) and also the CAP explicitly offer funding to protect and restore biodiversity and to promote ecosystems (although the bulk of their funding is dedicated to purposes unrelated to environmental conservation). 20 A third possibility is that receiving EU funds affects a region's eligibility for the Natura 2000 recognition.However, with the species and habitats deserving protection being decided centrally, the success of an application for protected status should be primarily determined by the pre-existing environmental conditions and biodiversity. 21 The use of EU funds could affect these but mainly adversely: for example, EU funds could be used to build infrastructure which destroys a habitat that would have been eligible for protection.Such a pattern, however, would imply a negative correlation between Cohesion spending and the presence of Natura 2000 sites, which is the opposite of what we observe.
The preceding discussion suggests that the relationship between Cohesion Policy spending and the extent of environmental protection could be either positive or negative.The positive correlation that we observe suggests that the net effect is positive.

Methodology
To assess the impact of Cohesion Policy on regional growth, we estimate a standard Solow-Swan growth model (Islam, 1995;Mankiw et al., 1992) augmented to include Cohesion Policy transfers as well as institutional quality.The model is estimated with annual data at the regional level (NUTS-2, except for institutions, which we only observe at the national level, as discussed above).Specifically, we estimate: where the dependent variable is the log-difference of per capita output of region j located in country i and observed at time t.The first three terms of equation ( 1) are the standard elements of the Solow model: the lagged output per capita y ijt−1 , measured in the preceding year, the ratio of gross fixed capital formation to GDP, s ijt , and a term containing population growth, n ijt , technological progress, g ijt , and depreciation, δ ijt .Since we do not observe g ijt and δ ijt , we follow the practice common in the literature and replace their sum with a constant term equal to 0.06. 22 To this, we add the weighted average of the World Governance Indicators, wgipca it , with the weights determined by principal component analysis.
Next comes our variable of interest: the ratio of the EU funds to GDP (efpayr ijt ).Since some regions receive no funds in some years, we add 1 to this ratio before taking logs.In the baseline specification the Cohesion Policy can only impact economic growth in the same way across all region types.However, we also estimate an alternative specification where we account for the fact that the less developed regions receive priority access to funding under the Convergence Objective.Therefore, we include a dummy for qualifying for the Convergence Objective, obj1 ijt .Approximately half of EU funds are set aside for Convergence Objective regions, making these regions the primary beneficiaries of Cohesion Policy. 23Previous contributions (e.g., Becker et al., 2012;Fiaschi et al., 2018) find heterogenous effects of regional transfers depending on the objective under which the funds are disbursed.This can be because of the types of investment financed under different objectives, or because some regions display limited absorption capacity as the amount of funding they receive exceeds the growth-maximising level (Becker et al., 2012).Finally, we also include fixed effects for regions, µ j , and time, τ t . 24

Results
We estimate equation (1) first by OLS, and then by 2SLS, with the proportion of the region's area taken up by Natura 2000 sites and the number of Natura 2000 sites per region as instruments.Table 5 reports the OLS results.In the first column, we report the standard Solow model (Islam, 1995;Mankiw et al., 1992) as a benchmark.The coefficients of all variables are strongly significant and have the expected signs: negative for the lagged output per capita (consistent with real convergence, whereby less developed regions grow faster), positive for the investment rate (higher investment/savings rate implies a higher steady state and, in turn, faster growth rate when the region is below its steady-state level), and negative for the term comprising population growth (the dependent variable is the growth rate of output per capita, so that higher population growth implies that incremental increases in output need to be shared among a larger number of individuals).
In the next five columns, we explore the effect of adding the variable of interest: transfers under the EU Cohesion Policy.In column (2), we add the ratio of EU funds to GDP: it has a positive and strongly significant effect on regional economic growth.In column (3), we add (country-level) institutional quality, which is also strongly significant and positive: regions in countries with good institutions grow more dynamically.Importantly, adding the quality of institutional environment does not diminish the positive effect of EU funds; in fact, the estimated impact is somewhat strengthened .Adding an interaction between institutional quality and EU funds (column 4) changes little: the coefficients of EU funds and institutions remain approximately the same as before, while the interaction term is insignificant: the effect of regional policy on growth does not appear conditional on good institutions. 25 In column (5), we add a dummy distinguishing regions that qualify for EU funding under the Convergence Objective: as discussed above, the less developed regions are prioritised as recipients of Cohesion Policy transfers.Previous studies (Fiaschi et al., 2018;Hagen & Mohl, 2011) conclude that these regions tend to benefit more from EU funds than other regions.Our results suggest that the less developed regions indeed grow faster; the effect of Cohesion Policy transfers is diminished somewhat but remains strongly significant.Finally, in column (6), we allow the effect of EU funds to differ in less developed regions and in the rest of the sample.The results are a bit peculiar: the EU funds retain their significantly positive effect on regional growth in the broad sample, and the less developed regions tend to report significantly higher growth overall as signified by the significant and positive Objective 1 dummy.However, the positive effect of EU funding is cancelled out in the less developed regions by the negative and significant interaction term between the EU funds and the Objective 1 dummy.Hence, less developed regions grow on average faster than other regions, but this growth premium does not appear to be correlated with the amount of regional aid that they receive.
As we argue above, the Cohesion Policy transfers may be endogenous in regional development.If this is the case, the effects estimated in Table 5 are likely to be biased.Therefore, Table 6 reports the 2SLS estimates of the effect of EU funds on regional growth, obtained with the presence of Natura 2000 sites as instruments.In column (1), the EU funds to GDP ratio is instrumented by the proportion of the area covered by Natura 2000 sites.Column (2) reports the associated first-stage regression.As expected, the instrument is strongly and positively correlated with the ratio of Cohesion Policy transfers to regional GDP.The EU funds, in turn, have a robust and positive effect on growth.The coefficient estimate is approximately four times larger than that obtained with OLS, suggesting that endogeneity biases the estimated effects downwards. 26The results change little when we replace area proportion with the number of protected sites (columns 3-4): the effect of EU funds strengthens a little, and the instrument is again strongly correlated with the Cohesion Policy spending rate.The next four columns replicate there regressions while adding the Convergence Objective dummy.Adding it reduces the estimated impact of Cohesion Policy funding on growth but the effect remains positive and statistically significant (at 10% when using the area proportion as instrument in columns 5-6, and at 1% in columns 7-8 when using the count of protected sites): this mirrors the pattern observed with The long and winding road to find the impact of EU funds on regional growth: IV and spatial analyses 591 REGIONAL STUDIES  OLS.Likewise, the Convergence Objective dummy is again positive and statistically significant. 27 Our results thus suggest that the effect of Cohesion Policy on regional development in the EU is positive, and that studying this effect by OLS, without accounting for endogeneity, may result in the estimates being biased downwards. 28The high F-statistics in columns (2), ( 4), ( 6) and ( 8) confirm that both instruments are strong: the first-stage F-statistic is always well above the rule-ofthumb threshold of 10.

Methodology
So far, we considered the impact of Cohesion Policy funding on the recipient regions only.The EU is a single free market with free trade in goods and services and unhindered mobility of labour and capital.Therefore, the effects of regional policy are unlikely to be confined only to the region receiving the funding.Rather, we can expect the EU funds to translate into an increase in aggregate demand in the recipient region as well as in other (especially nearby) regions.Such cross-regional spillovers can be captured by estimating a spatial model, which allows for linkages among regions based on a chosen spatial weight matrix.The choice of weights reflects the assumptions made in the model.Since some NUTS-2 regions are located on islands, we opt for using a spatial weight matrix based on distance rather than contiguity.Specifically, we use the squared inverse of d ij , the great circle distance (km) between the centroids of regions i and j. 29 As Kopczewska et al. (2017) note, the squared inverse distance matrix captures well the global links between all units and local clusters, as the strength of the relationship between regions declines exponentially with distance.
Since faraway regions are unlikely to exert much influence (and given that we use the inverse square distance, their weight would approach zero), we only consider spillovers from regions located within a threshold distance from the recipient region.Specifically, we use quartiles (k) of the overall distribution of the great circle distances among all region pairs.The quartile distance values, D (k), are: {D(1) ¼ 660; D(2) ¼ 1090; D(3) ¼ 1594}. 30 Some EU countries have overseas territories far away from the country's mainland.Including these regions would inflate the cut-off distances for the quartiles.Moreover, the impact of the nearest EU regions on such remote territories might not be economically plausible.For this reason, we exclude all EU overseas territories from the spatial analysis. 31 Spatial econometrics offers a variety of ways of capturing spatial spillovers.These include the spillovers in the dependent variable (spatial autoregressive model), independent variable and error terms (spatial error model).Since the spatial Durbin model (SDM) nests also several other models, it is recommended to start from the SDM specification and conduct common factor tests to discriminate between the SDM and other models (Beer & Riedl, 2012).
We thus estimate a SDM: where Δy is a 1 × nt vector of the log differences of output per capita, X is a m × nt matrix collecting the explanatory variables, W is the spatial weight matrix of size n × n, and ρ is the spatial lag, with n, m and t denoting the number of observations, number of explanatory variables and years, respectively. 32The Kronecker product of the identity matrix I T with the dimensions t × t accommodates W for use in a panel regression.Additionally, θ denotes the coefficients for the spatially lagged explanatory variables in X.The full spatial model is a spatial version of the model in equation ( 1).
The spatial weight matrix W has the following form: where w ij denotes an element of the spatial weight matrix W in row i and column j.That is, W ij equals the inverse of distance squared, as long as the distance is below a critical threshold, zero otherwise.Note that in contrast to OLS or 2SLS, the partial derivatives of the SDM are more complex to interpret.A change in a variable affects the region itself as well as other regions, giving rise to direct and indirect effects.Let S be defined as: where x k and β k represent the kth column of the X and β matrices, respectively.The average direct effect is then given as ADE = 1 nt tr(S), the average total effect is ATE = 1 nt i ′ Si and the average indirect affect is then In other words, ADE is the average of the sum of all diagonal elements of S, AIE is the average of the sum of all off-diagonal elements of S, and ATE is the average of the sum of all elements of S, with i being a summation vector of ones.

Results
Table 7 reports the results of the SDM estimation of the effect of EU funds on regional growth, which allow for interregional spillover in the effects of Cohesion Policy.The long and winding road to find the impact of EU funds on regional growth: IV and spatial analyses 593 Overman , 2012;Beer & Riedl, 2012).The coefficient of spatial dependence, ρ, is significantly positive, implying positive spillovers in GDP per capita growth rates.The direct effect of the EU funds approaches zero and is insignificant.In contrast, the indirect effect is positive and (weakly) significant.The total effect is also positive (but again only weakly significant): EU funds received by a region help the surrounding regions more than the recipient region itself.The same can be said about the investment rate: the direct effect is relatively small, but the spatial spillovers are positive and statistically significant, resulting in a statistically significant and positive total effect. 33 How to make sense of the fact that the indirect effect seem more important?We believe that this result suggests that the benefits from Cohesion spending accrue to the wider local area and not only to the recipient region: in fact, most of the effect accrues to the nearby regions.For example, the investments financed by regional policy could be carried out by firms located in other regions, or the increased spending could increase the imports of goods (e.g., building materials, machinery,

REGIONAL STUDIES
etc.) from other regions.This should not be surprising: the EU is an integrated market for trade in goods and services.There is no reason to expect the work financed by EU transfers to be carried out only by firms from the recipient region.Moreover, our estimates only capture the short-term effect of executing the investment financed by Cohesion Policy.The medium-and long-term effects of using the resulting improvement or facilities, however, may be more locally based.Finally, the weak significance of the spatial effects suggest that more work may be required to fully determine the nature and importance of interregional spillovers.The long and winding road to find the impact of EU funds on regional growth: IV and spatial analyses 595 REGIONAL STUDIES

CONCLUSIONS
In this paper, we report on the results of our journey to identify the effect of the EU Cohesion Policy on regional growth in the EU.We argue that possible reasons for the weak and mixed findings in the previous literature could be attributed to measurement errors, endogeneity of EU funds in regional economic performance, and/or presence of interregional spillovers in the effect of Cohesion Policy on regional growth.
To address the first of these potential shortcomings, we use an updated data set covering both old and new member states and with information on Cohesion Policy receipts in annual frequency.In contrast, most of the previous studies were limited to data covering only whole programming periods.Our data span three programming periods, from 1997 to 2014.Estimating a stylised panel version of the Solow growth model with OLS, we find that EU funds are strongly and positively correlated with regional growth.
Next, we address the potential endogeneity of Cohesion Policy by employing a novel and previously unused IV.Specifically, we use the presence of environmentally protected areas under the EU Natura 2000 programme as instruments for Cohesion Policy transfers.The presence of such areas is strongly and positively correlated with the amount of EU funds that NUTS-2 regions receive.We show that the payments from European funds significantly boost regional economic growth, both when applying OLS and 2SLS.The IV estimates, however, are larger than those obtained with OLS: this suggests that the OLS estimates may be biased downwards.This may help explain why many previous studies failed to yield significant findings.Our results, furthermore, confirm that there is substantial heterogeneity in the Cohesion impact on economic development, as reported in the literature: not all regions are affected equally.Further research may be needed to satisfactorily identify the sources of this heterogeneity.
Finally, we account for the likely presence of interregional spillovers in regional policy impact by estimating a spatial model of growth.The results suggest that interregional spillovers in the growth effects of the EU funds may be important (although our results are at the margins of statistical significance).In fact, the impact of Cohesion Policy seems to take place as much (or even more) in the nearby regions as in the recipient region.Our findings of positive spillover effects on nearby regions is in line with the previous results of Fiaschi et al. (2018) and Bachtrögler-Unger et al. (2022).
Our results thus confirm that accounting for endogeneity and spillover effects is important when assessing the impact of Cohesion Policy on regional development.Therefore, the preponderance of weak and mixed findings in the previous literature can be attributed to the fact that most of the earlier studies were estimated by simple OLS.As such, they are likely to have underestimated the Cohesion impact on regional growth and economic development.
Our findings, furthermore, give support to the recent effort to mitigate the consequences of the COVID-19 pandemic by increased redistribution through the Next Generation EU (NGEU) instrument.The NGEU envisages an additional spending package of €750 billion. 34 Bańkowski et al. (2022) estimate that this programme will boost output in the EU by up to 1.5 percentage point.The results of our analysis suggest that such favourable predictions may indeed be justified.
The results of our analysis are especially reassuring in light of the mixed findings on the economic effects of government spending (Ramey, 2011).A potential explanation of the disappointing findings of that literature is Ricardian equivalence an observation that the growth-stimulating effect of debtfinanced government spending is mitigated by the expectation of a future tax liability required to repay the ensuing debt.As Cohesion spending is financed from European rather than from national sources, the future tax liability of the recipient regions is negligible, so European funds should indeed have a better the growth-boosting potential.
The previous literature on the economic effectiveness of regional aid has been mixed and generally disappointing.The largest item of EU spending, the CAP, likewise largely fulfils a redistributive function, with little regard for improving economic efficiency.On that backdrop, our finding of a growth-stimulating effect of the Cohesion Policy spending is indeed a positive result.

NOTES
1. We use the terms 'Cohesion Policy', 'EU funds' and 'regional policy' interchangeably throughout the paper because they are broadly similar (though not entirely identical).When referring to Cohesion Policy, we refer not only to the EU's Cohesion Fund (CF) but also to other related funding instruments such as the European Regional Development Fund (ERDF), the European Social Fund Plus (ESF+), the European Agricultural Fund for Rural Development (EAFRD) and the European Maritime, Fisheries and Aquaculture Fund (EMFAF). 2. The EEC was the original name of what later came to be known as the European Community (EC) and finally the European Union (EU). 3. See https://cohesiondata.ec.europa.eu/EU-Level/Historic-EU-payments-regionalised-and-modelled/tc55-7ysv.The dataset contains information from the following funds: ERDF, European Social Fund (ESF), CF, EAFRD, European Maritime and Fisheries Fund (Roemisch, 2017).4. Most previous analyses used programming-period averages, or used these averages to intrapolate and estimate annual figures.An example of an innovative approach to constructing annual regional data is Roemisch (2017), who used annual national project-level data to create estimates of regional-level annual data, since project names often contained information about the recipient region.Among the more recent studies that use annual data similar to ours are Rodríguez-Pose and Garcilazo (2015) and Di Cataldo and Monastiriotis (2020).5. Cerqua and Pellegrini (2018) also use the new yearly data.However, they restrict their analysis to the regions of the EU-15 and1994-2006 period. 6.In principle, the same issue applies to the end of each programming period.However, for the 1994-99 and 2000-06 budgets, the end-of-budget spending overlaps with spending allocated during the first two to three years of the next period, which also appears in our data.It is only the beginning-of-budget spending in the 2014-20 programming period that our data set misses. 7.Only information on spending by the various EU funds (ERDF, CF, ESF and EAFRD) is available in annual frequency.A breakdown by sectors is available for whole programming periods only.8.The Cohesion Policy spending could be correlated with investment in physical capital.However, the correlation coefficient reported in Table 3 is relatively low at 0.1669.9.The new member states were ineligible for Cohesion Policy before joining the EU.We therefore assign them zero values of Cohesion Policy funds in the pre-accession years.Given that the regions in the new member states objectively had no receipts from Cohesion Policy in the pre-membership period, this solution seems appropriate.Assigning them zero EU funds values rather than treating the pre-membership years as missing observations increases the variation in the data, thus helping improve 22. Mankiw et al. (1992) use 0.05.Using 0.05 would result in the loss of two observations, Thessaly in Greece in 2000 and Nord-Est in Romania in 2012, which recorded a negative population growth rate of 5%.23.See https://ec.europa.eu/regional_policy/en/policy/what/investment-policy/.We are grateful to an anonymous referee for suggesting including this dummy as a regressor.24.The use of a fixed effects model is supported by the result of a Hausman test: the test statistic (for the baseline Solow model augmented to include the ratio of the EU funds to GDP) is 734.11(p ¼ 0.00).25.This contrasts with the previous findings of Becker et al. (2013), Rodríguez-Pose and Garcilazo (2015) and others.A possible reason for our findings deviating from the previous literature is that we use country-level rather than regional information on institutional quality.26.As we argue in the Introduction, this kind of bias could be driven by the fact that slowly growing regions receive more funds because they remain eligible for Cohesion Policy funding for longer.In contrast, fast-growing regions quickly lose eligibility for transfers.27.As an alternative, we split the sample into two subsamples according to the Convergence Objective eligibility.These results (available from the authors upon request) show a strong and statistically significant effect of EU spending in the less-developed regions and a slightly weaker but still positive effect in the rest of the regions; as before, the effect appears more robust when using the number of protected sights as the instrument.28.It is possible that the EU funds affect regional development with a delay rather than contemporaneously.This could be due either to delays in disbursing the funds or because the investments themselves take time to build and bear fruit.Therefore, we also re-estimated the previous results -OLS and 2SLS alikewith all the EU funds lagged by one year, and again with all regressors lagged by a year.These regressions results (available from the authors upon request) are very similar to those reported with contemporaneous regressions.29.We also conduct regressions with a spatial weight matrix based on k-nearest neighbours, with k ¼ {5; 10; 15; 25; 50}.These results (available from the authors upon request) indicate that our model is robust to the choice of the spatial weight matrix.30.We also use a spatial weight matrix based on a cut-off value of the quartiles of great circle distances for each region specifically, which results in an equal number of non-zero values for each region.The difference in the results is negligible.31.The following remote NUTS-2 regions were dropped: Madeira, Azores, Canary Islands, Ceuta, Melilla, Guadeloupe, Martinique, French Guiana and Réunion.For example, the French overseas territory of Réunion (in the Indian Ocean) is more likely to be economically influenced by mainland France rather than the nearest EU region, which is Cyprus.32.We are mindful of the identification problems when using IV in spatial models, as highlighted by Gibbons and Overman (2012).In particular, the methodology for dealing with endogenous variables in a panel spatial model is not well-established.Therefore, rather than present results that might be questionable and potentially weak, we prefer to implement one contribution at a time.In the preceding section, the objective was to compare OLS and 2SLS estimates.Now we compare estimates obtained with OLS with those based on a spatial model.The 2SLS analysis has established that the OLS estimates may be biased downwards.Therefore, the SDM results probably also underestimate the true effects of Cohesion Policy in a spatial framework.33.We conducted additional robustness checks by replacing the distance-based weight matrix with considering nclosest neighbours (with n ¼ 5, 10, 15, 25 and 50) and omitting the interaction term between the EU funds and institutional quality.These results (see the supplemental data online) suggest that the effect of the Cohesion spending hovers at the limits of statistical significance.The results of the spatial model, therefore, need to be taken with caution.34.See https://ec.europa.eu/info/strategy/recoveryplan-europe_en

Figure 1 .
Figure 1.Ratio of approximate yearly structural funds to regional gross domestic product (GDP), 1997-99, 2004-06 and 2012-14.Sources: Authors based on data from Cambridge Econometrics Database and DG Regio.

Figure 2 .
Figure 2. Ratio of the area of Natura 2000 sites to a region's total area, 1997-99, 2004-06 and 2012-14.Sources: Authors based on data from the European Environment Agency and Eurostat.

Table 1 .
Overview of variables.
Note: Values are the average annual changes in the area under environmental protection, and in the number of protected sites, at the NUTS-2 level.The figures for 1997-2014 are the overall changes during this period, not the average changes per year.

Table 5 .
European funds and regional growth: ordinary least squares (OLS).