Policy Research Working Paper 10658 Are Regional Fiscal Multipliers on EU Structural and Investment Fund Spending Large? A Reassessment of the Evidence Federico Fiuratti Desislava Nikolova Steven Pennings Marc Schiffbauer Development Research Group & Macroeconomics, Trade and Investment Global Practice January 2024 Policy Research Working Paper 10658 Abstract The European Commission’s “NextGenerationEU” by estimating regional short-term multipliers using recent COVID-19 recovery package has underscored interest in data on EU fund spending and a leave-one-out predicted the size of regional fiscal multipliers in Europe. While the disbursement schedule instrument. In contrast with much objective of these funds is the long-term transformation of the recent literature, there is little evidence of large rela- toward more sustainable green growth and digitalization tive GDP multipliers at either the national or subnational in EU economies, several recent papers have also focused level in the short term. This is despite a strong response of on their short-term stimulatory effects and have estimated regional investment to EU funds, which often increases large short-term regional multipliers on historical EU struc- euro for euro. The results suggest that expectations should tural and investment fund spending. This has contributed be tempered on using EU structural and investment funds to a view that EU funds can boost growth substantially as a tool for short-term regional fiscal stimulus, and instead not only in the long term, but also in the short term in policy makers may want to focus on the long-term benefits countries receiving large flows, particularly in Central of EU funds, in line with their original purpose. and Eastern Europe. This paper reevaluates the evidence This paper is a product of the Development Research Group, Development Economics and the Macroeconomics, Trade and Investment Global Practice. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at ffiuratti@worldbank.org, dnikolova@worldbank.org, spennings@worldbank.org, and mschiffbauer@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Are Regional Fiscal Multipliers on EU Structural and Investment Fund Spending Large? A Reassessment of the Evidence Federico Fiuratti, Desislava Nikolova, Steven Pennings and Marc Schiffbauer1 1 All World Bank. Corresponding Author: Steven Pennings (spennings@worldbank.org). JEL: E62; F45. Keywords: Fiscal Multiplier, European Union, Monetary Union. We thank several World Bank staff for comments (especially Reena Badiani-Magnusson); Luigi Durand and Raphael Espinoza for sharing links to EU fund data and assistance replicating their results; Guido Damonte for helping us building a neighbors dataset for each NUTS2 region; and helpful comments from European Commission seminar participants and our discussant Philippe Monfort. The views expressed here are those of the authors and not necessarily those of the World Bank, its executive directors or the countries they represent. The authors are grateful for financial support from the Knowledge for Change (KPC) Program, administered by the World Bank, and funded by The Swedish International Development Cooperation Agency (SIDA), Agence Française de Développement (AFD) - French Development Agency, the Government of Japan, and the European Union. An online appendix is available at: https://sites.google.com/site/stevenpennings/ or https://drive.google.com/file/d/1BT-mTW-J26Pt42Jn6FIqgI8F2WhE8Cmu/view?usp=sharing 1. Introduction The 2008 financial crisis, European debt crisis, and COVID-19 pandemic have renewed interest in the size of the fiscal multiplier. With interest rates in Europe and the US having been constrained by the zero lower bound (ZLB), aggregate monetary policy was more constrained as a countercyclical tool than previously, with some of the burden for stabilization policy falling on fiscal policy. The fiscal response to the COVID-19 pandemic was many times larger than the response to previous recessions in the US, Europe, and elsewhere. The European Commission initiated the debt-funded €500-€800 billion ($550-$900 billion) NextGenerationEU (NGEU) COVID-19 recovery plan with the objective to raise long-term economic growth by facilitating the transformation towards more sustainable green growth and digitalization in EU economies. Across its 7-year life, the program averages about 0.5%-0.75% of EU GDP, which European Commission simulations using a large New Keynesian model suggest would boost GDP by 1%-1.5% over 2021-27, much of which is through fiscal short-term stimulus channels (Mahieu et al. 2021).2 The pandemic and financial/debt crises have also illustrated the importance of regional fiscal multipliers—in individual European countries, or US states—as one monetary policy applies to the whole monetary union. The European debt crisis was strongly asymmetric across countries in Europe—with much larger falls in GDP and spending cuts in the South than the North. Blanchard and Leigh (2013) argue that fiscal multipliers in many individual countries were larger than previously thought, partly explaining the larger output falls. In the US, the subprime housing bust hit some regions much harder than others. The most salient aspect of EU funds is not their Europe-wide effects, but their differential effects across European countries and subnational regions because EU spending is heavily focused in the union’s newer and poorer members in Central and Eastern Europe (henceforth CEE). For example, planned NGEU spending for Bulgaria was originally penciled-in at EUR 7.7 billion in current prices, including EUR 6.3 billion or 10.2% of the 2019 GDP in grants under the Recovery and Resilience Facility – the second highest amount in the EU as a share of GDP (after Croatia). 3 The European Commission estimated that as a result, GDP could be up to 4% higher in Bulgaria (Mahieu et al. 2021). The scale of these effects in the short term depends on the size of the fiscal multiplier for EU fund spending. In the US, several papers have found sizable federal cross-state spending multipliers from similar federal stimulus programs of around 1.8 (Chodorow-Reich 2019) (though other research disputes the robustness of some of these findings; see Ramey 2019). In the European context, Coelho (2019) finds large fiscal multipliers on past rounds of structural EU fund spending, of around 1.8 contemporaneously to 0.9 three years later when estimated at the subnational level. 4 More recently, Durand and Espinoza (2021) also find large short- term multipliers on EU funds at the national level, ranging from 1.2 on impact to 1.8 a year later.5 In work 2 https://voxeu.org/article/stylised-quantitative-assessment-next-generation-eu-investment 3 On 30 June 2022, the European Commission recalculated the maximum grant amounts under the RRF for all Member States resulting in a cut for Bulgaria of just over 9 percent and a new total of EUR 5.69 billion (European Parliament 2023). 4 The multipliers that Coelho (2019) reports in the text are 60% of the numbers reported in the tables, to adjust for 40% co-financing. For example, the main contemporaneous multiplier in her Table 3 is listed as equal to 3 but reported as 1.8 in the text. Cumulative multipliers—defined in the literature as the cumulative response of GDP to the cumulative expenditure—are reported in her Table 11 as 1.5, but are reported in her text as 0.9. 5 Spatial CGE models produce mixed short-run multipliers for Cohesion Fund spending, though large positive long-run benefits (in part through spillovers across regions), for example see Crucitti et al. (forthcoming). 2 contemporaneous to ours, Canova and Pappa (2022) find large fiscal multipliers in the short term for the European Regional Development Fund (ERDF) of 1.4 contemporaneously to 0.22 two years later. However, they find the European Social Fund (ESF) has negative multipliers in the short term, though positive in the medium term. 6 Methodology. In this paper, we investigate the size of the spending multiplier on EU structural funds at both the country and subnational (NUTS2) level. We face two common challenges in this literature: (i) measurement and (ii) identification. Measurement of the size of EU fund spending in a year is more complicated than is typically the case in the fiscal multiplier literature because the European Commission records spending when it is disbursed by the Commission to national governments, not necessarily when the spending takes place. To address this challenge, we use recently released adjusted (modeled) estimates of the timing of spending from the European Commission at the NUTS2 subnational level. The second challenge is one of identification as, in general, growth shocks can affect fiscal policies or other omitted variables can affect both fiscal policy and growth. This is less of a problem in the EU fund setting as expenditure is committed in advance at the start of the EC’s seven-year program cycle. 7 However, it could be the case that expenditure is slowed down or sped up in response to a growth shock. To overcome this problem, we instrument the share of spending in each year of the seven-year program cycle. Building on the approach in Kraay (2014) and Durand and Espinoza (2021), we use a leave-one-out instrument where the predicted disbursement schedule is estimated using data on all other regions or countries. This instrument turns out to predict actual expenditure very strongly in most specifications, with first stage F-statistics at the subnational level running from 60-800 (smaller F-stats at the country level). Results. In contrast with much of the recent literature, we find little evidence of large relative GDP multipliers on EU fund spending at either the national or subnational (NUTS2) level in the short term. Estimated multipliers vary across samples, and are occasionally significant at the 5% level, but those occasionally significant estimates are always less than 1. This implies that each euro spent on physical or social investment raises GDP by less than 1 euro in the short term, and possibly not at all. Using our preferred instrumental variable speciation, we find a precisely estimated subnational multiplier of zero, with country-level multipliers also being insignificant (though estimated less precisely). We generally find that multipliers in CEE countries are slightly larger than in the rest of Europe, though differences are usually not significant. 8 The failure to estimate large multipliers is not because of a lack of response of investment to EU funds, mismeasurement issues, or a delayed response of output. Even though EU funds do not lead to a large boost to GDP, they boost investment almost euro-for-euro in most subsamples. We also find even lower GDP 6 An earlier literature evaluated the effect of EU Funds on longer-term growth (Mohl and Hagen 2010, 2011; Becker et al. 2010, 2012, 2013, 2018; Bargain et al. 2013; Sala-i-Martin 1996; Beugelsdijk and Eijffinger 2005; Ederveen et al. 2006). 7 The programming periods we are including in the paper lasted seven years (2000-2006, 2007-2013 and 2014-2020), previous programing periods that we are not using had different durations (1989-1993 and 1994-1999). 8 Motivated by a recent policy debate, we tried to estimate “green” multipliers on environment-related spending (and differentiate them from “brown” multipliers on other spending). Unfortunately, we were unable to separate green EU fund spending from other categories or find an instrument for green spending by itself. Consequentially, we can only produce OLS estimates which should not be reported causally. Nonetheless we do not find any evidence that “green” investment spending is more substantially correlated with growth than “brown” spending (see Online Appendix Table A3.5 for details). 3 multipliers—and a lower response of investment—when using the raw disbursements of EU funds instead of the recently released adjusted estimates of the timing of spending. Estimated cumulative multipliers of up to two years after the initial spending are typically insignificant, suggesting that results are not driven by a delayed response of short-term GDP. We subjected our results to a range of robustness tests to check that they are not caused by outliers, influential individual observations, or omitted variables. First, because growth rates in smaller regions are often very volatile, we winsorize all our growth data—though we also find little robust evidence of large relative multipliers even using non-winsorized data, or when winsorizing at different levels. 9 Second, we carefully check for influential observations by rerunning all regressions dropping countries and years one-by- one, and omitting two country-year observations that are extremely influential. 10 Finally, our default approach includes both time and region fixed effects, as well as the one-year lagged growth rates. We also fail to find large multipliers adding other controls in the literature (including institutional indicators). We also consider some extensions to try to understand the heterogeneity and economic mechanisms behind our multiplier estimates. In contrast with some of the literature, we find little evidence of regional spillovers—spending in a region boosting growth in neighboring regions. Even when we disaggregate our sample beyond CEE vs non-CEE countries—into northern CEE vs southern CEE blocks, for example—we also find little robust evidence of large regional multipliers. The same is true when we split our sample in the time dimension or consider results separately for different EU funds. 11 We also find some suggestive evidence of some negative anticipation effects (one year ahead of EU fund spending) at the subnational level, partially driven by anticipated falls in investment (mostly in non-CEE countries). This could be because some investment is put off until it is eligible for EU funding, though more research is needed to test this hypothesis. Contribution and related literature. Our paper contributes to a recent literature on the short-run growth impacts of EU fund spending in different countries and regions. We differ from the empirical literature mostly in terms of conclusions—we do not find evidence of large short-term multipliers—and we also offer some methodological improvements. Coelho (2019) is the first paper to estimate short-run fiscal multipliers on EU spending, and, like us, she uses NUTS2 subnational data. Our paper seeks to improve on Coelho’s pioneering paper in three main ways: data quality, controlling for region-specific effects, and using an alternative instrumental variable (IV) approach. First, Coelho (2019) did not observe EU Fund expenditures at the NUTS2 level, and instead she interpolated them using country-level disbursement schedules combined with NUTS2 commitments. We have access to more recent data which should improve the accuracy of the estimates. Second, Coelho’s interpolation 9 Winsorizing is a less extreme approach than dropping outliers. Rather than dropping the extreme observations, winsorizing keeps them in the sample but rounds them to a bound. 10 The differences between our country-level results and those of Durand and Espinoza (2021) are mostly due to the selection of influential observations. Other approaches in the literature drop periods with austerity programs or drop observations that Cook’s distance or standardized residuals deem to be influential, often resulting in more dropped observations. Often these dropped observations are non-random; Coelho (2009, p582) finds these “are mostly observations with abnormally sharp output fluctuations and very low transfers changes”. 11 While there is some tentative evidence of an all-Europe subnational multiplier above one for the Cohesion Fund (mostly environmental and transport investment), it is surprisingly insignificant for the CEE sample that receives most of the funds. 4 approach meant that she could not include NUTS2 fixed effects but only country fixed effects. This leaves her results more susceptible to confounding differences in trend growth within countries, a problem we avoid with the inclusion of subnational (NUTS2) fixed effects. Finally, Coelho uses EU fund commitments as her instrument, which tend to predict levels of spending rather than changes.12 In contrast, in our paper changes in our instrument are a strong predictor of changes in spending in most specifications. Durand and Espinoza (2021) use a similar adjusted EU-fund data and IV approach, though they only examine effects at the country level. 13 In contrast with Durand and Espinoza (2021), we produce estimates at the NUTS2 level, which improves the sample size and accuracy of the estimates (though reduces the length of the sample). However, even when we use country-level estimation, our results are still quite different, with those authors finding large (>1) and significant GDP multipliers, whereas we find small (<1) or insignificant multipliers. The different results seem to be mostly explained by the treatment of 2009 (when Europe suffered a severe recession): Durand and Espinoza remove it whereas we control for the average depth of the recession in Europe. When we replicate Durand and Espinoza’s all-Europe results, we get a similarly large IV multiplier of about 1.26 (they get 1.21), but this drops to 0.79 when 2009 is included in the sample (see Section 4.1 for a discussion). While both their approach and ours are reasonable, the fact that a multiplier above one hinges on the modeling assumptions of a single year suggests that that finding should be used with caution in policy discussions. 14 Canova and Pappa (2022) also estimate the effect of EU funds on economic growth at the subnational and national levels, using similar adjusted EU-fund data to ours. There are two main differences between their approach and ours. First, they estimate the effect of two of five EU funds, whereas we examine the combined effect of all spending types. 15 The most important difference is the exclusion of the Cohesion Fund (CF) in their paper, which funds infrastructure and sustainable development in lagging regions, mostly in Central and Eastern Europe. The second difference with Canova and Pappa (2022) is in the estimation methodology and the interpretation of multiplier estimates. They use an “instrumental variable Bayesian local projection approach” that allows for multipliers to vary across regions but does not include time fixed effects. We use a classical panel estimator that is standard in the multiplier literature and does include time fixed effects. Their instrument is the residual of aggregate EU funds regressed on euro area macro variables like aggregate GDP and interest rates, which has some similarities to how shocks are identified in structural vector autoregression models (SVARs). In contrast, we use differences in predicted spending as the instrument. The different approach means that Canova and Pappa (2022) are estimating a combination of relative and aggregate multipliers, whereas we are only estimating a relative multiplier. A relative multiplier is defined as follows: if region A gets an extra 1€ of spending relative to region B, how much will relative GDP increase in region A relative to region B (other things equal)? 16 The relative multiplier does not incorporate the aggregate response to fiscal spending (or any aggregate shocks, including the effects of monetary policy), as they are 12 Coelho (2019) does not report her first-stage estimates, but they are included in her replication file. 13 They also disaggregate the multiplier estimates across different sectors and investment types, which we do not. 14 However, both our papers are in agreement that the effect on investment can be above one at the country level. 15 Canova and Pappa (2022) study the European Regional Development Fund (ERDF) and the European social fund (ESF), but do not consider the Cohesion Fund (CF), the European Agricultural Fund for Rural Development (EAFRD) or the European Maritime and Fisheries Fund (EMFF). Our spending data also includes the Fund for European Aid to the Most Deprived (FEAD), and Youth Employment Initiative (YEI), though these are smaller. 16 Coelho (2019) and Durand and Espinoza (2021) also estimate relative multipliers. 5 captured by the time fixed effects—and so is robust to those shocks. In contrast, Canova and Pappa’s (2022) captures both the relative multiplier and the effect of an increase in aggregate EU fund spending across all regions. As discussed further below in Section 5 (and in Nakamura and Steinsson 2014), some of those aggregate effects may be of interest to policy makers, but at the cost of a lack of robustness to aggregate shocks and the monetary policy response.17 The rest of this paper is organized as follows. Section 2 outlines the empirical methodology, including measurement issues, identification, and data. Section 3 presents the main results. Section 4 presents robustness tests, extensions, and estimates of heterogeneity in various dimensions. Section 5 discusses mechanisms and the interpretation of multipliers, and Section 6 discusses policy implications and conclusions. 2. Data and Methodology 2.1 Main Specification Our primary goal in this paper is to estimate the effect of EU structural fund spending on contemporaneous economic growth at the national or subnational level. Our main specification is given by Equation 1, where the left hand side variable (, − ,−1 )/ ,−1 is the real GDP growth rate at the subnational level i in year t (real is denoted by “r”) (substitute subnational level i for country c for country-level specifications). 18 The right-hand side independent variable is the change in real EU fund spending in region i in year t, as a share of GDP ( , − ,−1 )/,−1 . 1 is our parameter of interest, the impact multiplier of EU funds. Because the right hand (RHS) and left-hand sides (LHS) are both defined as a share of GDP, 1 is interpreted as an impact multiplier: the increase in regional GDP in euros for each extra 1€ of spending. and capture regional and time fixed effects respectively and , are other controls. In our main parsimonious specification, the only control is a lagged dependent variable, i.e., , = (,−1 − ,−2 )/ ,−2 (though we considered other controls as robustness tests, too). , −,−1 � , −,−1 � (1) = 1 + , + + + , ,−1 ,−1 � , −,−1 , −,−1 If using a 2SLS/IV specification, is replaced by its predicted value from Equation 6. ,−1 ,−1 To capture longer run effects, beyond the first year, we also estimate cumulative multipliers, where both the effect on GDP (LHS) and the spending itself (RHS) are cumulative for = 2 and = 3 years. (For example, H=2 is the contemporaneous effect plus one extra year.) Equation 2 with H=1 collapses to the main specification for impact multipliers in Equation 1. ,+ℎ −,−1 � ,+ℎ −,−1 � (2) ∑−1 ℎ=0 = ∑−1 ℎ=0 + , + + + , ,−1 ,−1 17 One further difference is that in our paper, the IV specification estimates the response to predicted spending, whereas Canova and Pappa (2022) estimate the effect of more unexpected spending (at least not predicted by macro variables). 18 Equations 2 and 6 can also substitute country c for subnational level i in a similar way. 6 2.2. Data The main sample consists of 25 countries (EU27 countries minus Ireland and Lithuania) for which we have country level data and NUTS2 level data. 19 Our sample generally runs from 2000-2018 (programming periods 2000-2006, 2007-2013 and 2014-2020), though the sample starts later for some countries and specifications due to missing data. 20 The main data required for Equation 1 are real GDP and real EU payments, both measured in real euros. Real investment data are required for variants of Equation 1 explaining investment rather than GDP. Each real variable is constructed from nominal data taken from Eurostat for GDP and Investment and the European Commission for EU payments. Specifically, nominal GDP data are “Gross domestic product (GDP) at current market prices by NUTS2 regions” (downloaded 19-Oct-2022), 21 EU payments are from “Historic EU payments - regionalised and modelled” (downloaded 09-Sept-2021) 22 and Investment data are “Gross fixed capital formation by NUTS2 regions” (downloaded 19-Oct-2022). 23 In order to convert nominal series denominated in euros ( , ) to a real series ( , ), we take the nominal value of each of the series for NUTS2 region i that belongs to country c(i), at time t and divide it by that € country’s GDP deflator at time t measured in euros ( ( ), ) i.e , = € , 24 . The deflator data are measured (), in euros, and are taken from Eurostat (downloaded 14-Oct-2022). 25 Data Cleaning. Growth data are often highly volatile, and so extreme observations will receive a disproportionate importance in any least-squares algorithm. This is especially true if EU payments are 19 We exclude Ireland and Lithuania as changes in the definition of some NUTS2 regions mean that the NUTS2 regions in the payments data are different NUTS2 regions in the GDP data. The EU27 are the 27 member states of the European Union after Brexit, so it does not consider the UK. We also use a larger country-level sample in table A4.6 for comparison with Durand and Espinoza (2021), which includes EU27 + UK and has data that runs from 1994–2018. 20 EU payment data for Croatia only starts in 2007. 21 https://ec.europa.eu/eurostat/databrowser/product/page/NAMA_10R_2GDP GDP data at the NUTS2 level uses the 2021 NUTS2 definitions (see https://ec.europa.eu/eurostat/web/nuts/history for the changes in NUTS2 regions), but payments data use historical NUTS2 definitions. The two differ as historical NUTS2 regions sometimes had a change in their name or were split into smaller regions. We combined the new definitions of the GDP data to be consistent with the old definitions of the payments data. Some examples of changes include Croatia (HR04=HR02+HR05+HR06), Hungary (HU10=HU11+HU12) and Poland (PL12=PL91+PL92). Also, many names for NUTS2 regions have been changed across programming periods and we also had to adjust for that (in France and Poland). 22 Downloaded from https://cohesiondata.ec.europa.eu/Other/Historic-EU-payments-regionalised-and-modelled/tc55- 7ysv, based on Lo Piano et al (2017). We thank Luigi Durand for referring us to these data. 23 https://ec.europa.eu/eurostat/databrowser/product/page/NAMA_10R_2GFCF. As Investment data use the 2021 NUTS2 definitions, we used the same procedure we have used for GDP data (explained in footnote 19) to make investment data match payments data. 24 We used country-level GDP deflators because NUTS2 level GDP deflators are not available. Note that this means that we are not able to tell whether an increase in GDP in a NUTS2 region is due to an increase in the price level in that region (relative to the national average), or an increase in quantity produced. This is an important caveat for the NUTS2 level results. 25 Note that we apply the same methodology for countries with flexible non-euro exchange rates; which is equivalent to regressing real GDP growth in local currency on the change in real local-currency spending as a share of lagged GDP. This is because the real GDP growth rates in euros reported by Eurostat are the same as the real growth rates in national currency, and euro deflator inflation is the same as national currency deflator inflation adjusted for the change in the exchange rate. https://ec.europa.eu/eurostat/databrowser/product/page/NAMA_10_GDP__custom_3592126. 7 extreme during the same period (high leverage). Consequently, care needs to be taken to ensure the results reflect a general relationship, rather than a few high-influence periods. To improve our robustness to influential observations, we first winsorized GDP growth rates data at the 98% level for country-level data and at 90% for NUTS2-level data. 26 Second, we dropped countries and year observations one-by-one to identify influential observations (where the multiplier estimates change substantially with their exclusion). At the country level, this algorithm identifies Estonia in 2011 (EE11) as influential in OLS regressions, and Croatia in 2016 (HR16) for IV specifications (the first-stage changes). Consequently, we drop both for all country-level specifications. No individual observations are influential by the algorithm in NUTS2-level specifications. Descriptive statistics are presented in Online Appendix Table A2.1 and illustrate the extreme volatility of many of our variables. 27 2.3 Measuring EU Fund Spending Unlike most types of government spending, EU structural funds mostly operate on a reimbursement basis which can disconnect the recorded timing of spending from the time it actually occurred. 28 This is potentially problematic for estimating fiscal multipliers – which require spending and its effects to happen at the same time or within a fixed interval after that—and can lead to biased multiplier estimates. EU structural and investment fund spending proceeds in a roughly three-step process. First, is the planning stage at the start of the 7-year programing period, when the EU spells out the type and size of each program, governments make submissions, and those submissions are approved by the EU. This means that the plan on types and total amount of spending to each country and/or subnational NUTS2 region is usually known well in advance before any actual spending takes place. 29 Second, is the implementation step during the 7-year program which involves (i) some prepayments from the EU to member governments in advance of anticipated work to “provide liquidity” (for example, to provide an advance on a contract), (ii) work on the actual project, and (iii) the firm or government applying for reimbursement from the EU as the work is done. The EU usually does not reimburse 100 percent of the costs (depending on the Fund) as the national government often has to make a copayment. The final step is the closure of the program, which involves the EU paying any remaining balances, conditional on the project being implemented satisfactorily. The closure of the program can take place years after the end of the program period. For example, the 2007–13 program was closed on 31 March 2017 (European Court of Auditors, 2018, p13). 30 26 That is, suppose that 1% of observations , are greater than a cutoff 1 , and 1% are below 1 (which implicitly defines 1 and 1 ). Then the winsorized data replaces the top 1% extreme observations with 1 (rather than dropping them) and replaces the bottom 1% of observations with 1 . We call this 98% winsorized. Relative to dropping observations (an alternative approach), this winsorized data keeps the extreme observations in the sample but reduces their influence. We winsorized at the 90% level for NUTS2 data because it is more volatile than country-level data. 27 Other data we use for other regressions can be found in Online Appendix Tables A2.2 and A2.3. 28 There are also prepayments (payments to provide liquidity for future work), which further disconnects the timing of payments from the timing of activity. 29 More specifically, the Cohesion Fund is allocated nationally, while the structural funds (regional development, social fund) are allocated by NUTS2 region. 30 There are also some delays relative to the planned closing in March 2017. For example, the six NUTS2 regions in Bulgaria (BG31, BG32, BG33, BG34, BG41 and BG42) received payments in 2018 that were for the 2007-2013 programming period. For the programming periods we are using, modeled payments include payments that go up to 2009 (3 years) for 2000-2006 and up to 2015 (2 years) for 2007-2013. While actual payments go up to 2015 (9 years) for 2000-2006 and up to 2018 (5 years) for 2007-2013. 8 The standard data on EU structural fund spending from the European Commission (EC) record when reimbursement payments (or prepayments) are made to national/local governments, rather than when the activities took place on the ground. 31 Reimbursement can take a long (and variable) time, depending on the details of the project and how long it takes for the EC to verify that its requirements have been met, and different projects have different shares of prepayments. We deal with this issue by using EU “modeled” payments provided by the European Commission as a measure of expenditure. These data attempt to make some adjustments to the raw payments to more accurately account for when the real expenditure took place. Most importantly, they reallocate late payments (those taking place 2-3 years after the end of the funding window), to earlier in the programming period (though they also make several other adjustments). The “modeled” payments are based on Monte Carlo simulations of the possible allocations of spending (they do not use any external data on when expenditures took place). While changes in modeled payments are highly correlated with changes in raw payments (Figure 1), they are not identical; a 1ppt GDP increase in raw payments is associated with a 0.6ppts increase in modeled payments. 32 This is important for accurately estimating the size of multipliers. 33 Figure 1 2.4. Identification Identifying the effect of EU spending on output faces two main challenges: omitted variable bias and reverse causality. These are discussed in detail below. 31 “The yearly breakdown of the dataset follows the cycle of the European Commission payments to the Member States and not the date on which real expenditures took place on the ground; accessed September 9, 2021. https://cohesiondata.ec.europa.eu/Other/Historic-EU-payments-regionalised-and-modelled/tc55-7ysv. 32 The raw and modeled payments are more correlated in level terms, as this is driven by the size of commitments (determined in the EU 7 budget) rather than year-to-year variation in the timing; see Online Appendix Figure A1. 33 One concern is that our independent variables may suffer from some residual measurement error, despite the modeling process, potentially biasing the results towards zero. However, we find that modeled payments are a strong predictor of investment, especially public investment, which helps to assuage this concern. 9 Omitted variable bias (OVB) emerges if the same external shock affects both fiscal spending and economic growth. 34 In our setup in equation 1, the set of possible confounding factors is reduced by time fixed effects, , that incorporate all time-series shocks like monetary policy, global growth, commodity prices, and the variation of aggregate EU spending year-by-year. Our specification is in changes (rather than in levels) and with region fixed effects , which means that (i) any deviations in levels are differenced out 35 and (ii) differences in growth rates across regions are incorporated into the region fixed effects. This combination of time and region fixed effects, and estimating in differences, means that potential confounding variables must affect economic growth (rather than levels) and vary across regions and by time, which removes many possible confounding factors. OVB is further reduced by controlling for lagged growth (and so accounting for dynamic growth trends) and estimating equation 1 with 2SLS (see below). Reverse causality is typically the biggest challenge in the estimation of fiscal multipliers, as causality often runs from shocks to economic growth to fiscal spending (, → Δ, in equation 1). In developing countries with limited fiscal space, fiscal policy is typically pro-cyclical (Frankel et al. 2013) resulting in upward-biased fiscal multipliers. Developed countries with countercyclical spending have the opposite problem, multipliers will be downward-biased as stimulus packages are enacted during downturns. Reverse causality is less of a problem for EU fund spending, because of the multi-year budgeting process (though it is still a concern). As mentioned above, the EU structural and investment funds are organized by 7-year programming periods, where the commitments are allocated to regions or countries at the start of the period. This means that the total amount of spending in a region cannot respond to contemporaneous shocks and is “more exogenous” than other types of fiscal spending in the literature. However local economic shocks can still affect the timing of EU fund spending, or whether that spending is taken up at all. On one hand, because most EU spending has a co-payment requirement – which can be smaller (typically for CEE) or larger (non-CEE) – and local governments might find it easier to fund their co- financing component when economic growth is strong and local government coffers are full. 36 On the other hand, governments may be more willing to organize (and co-finance) spending under EU programs to boost their economy in a recession. Because it is difficult to know in which direction or how large these biases might be, we supplement our OLS estimation of equation 1 with a 2SLS specification, which we discuss now. 2.5 Instrument: Leave-One-Out Predicted Disbursement Schedule To address potential concerns that governments might speed up or slow down spending of EU structural funds in response to local economic shocks (reverse causality), we use a leave-one-out predicted disbursement instrument based on Kraay (2014) (and also used by Durand and Espinoza 2021). The instrument focuses on the most endogenous part of the spending in region i in country c(i) – the share , 34 An example is a war that might cause an increase in military spending and increased uncertainty causes a drag on growth. 35 This makes our approach more robust to confounding factors than some of the earlier literature where the specification is in levels. For example, Coelho (2019) reports that her level estimates are robust before 2006, but appear to be confounded after 2006 by the effects of the European debt crisis, which disproportionately affected regions receiving larger EU payments. 36 EU fund programs also usually involve borrowing by beneficiaries to pay for the expenditure, and then seeking reimbursement afterwards, and the access to these loans could be easier in boom times. 10 € € of total 7-year program P spending (∑∈{} , ) that occurs in year ( , ), as in Equation 3 (with everything defined in nominal euros, denoted with a € superscript). For example, if spending was allocated evenly across the 7-year programming period (plus a 2-year extension after its end), , = 1/9. 37 The total spending in € euros in each region ∑∈{} , is fairly exogenous to local conditions as it is set at the start of the programming period (approximately), so it does not need instrumenting. 38 € ∑ € (3) , ≡ , / ∈{} , Once , are defined, we calculate the share of spending in that year based on the disbursement shares in other regions R (excluding i). For example, suppose that all other regions spend their EU funds evenly over the 9 years, whereas region i spends all of it in the first year. Then our instrumented spending in region i in ,€ € the first year 1 of the programming period , would be 1/9 of the total spending of the region ∑∈ , th (rather than all of it). We implement this by running a regression of the shares on time dummies for all regions R other than i (as in equation 4). 39 For country-level regressions, we construct the shares the same way, except that the region for the regression also excludes all subnational regions in the same country ( ), i.e. \{} in Equation 4 is replaced by \{( )}. 40 Excluding other subnational regions in the same country is important as growth shocks in country c might affect EU spending in those other subnational regions, violating the exclusion restriction. Then spending in subnational region i is constructed using a combination of the predicted spending shares �, (equal to the estimated time dummy ̂, ), multiplied by the total program spending in € region i, ∑∈ , (equation 5). (4) , = ∑∈ , + , ∀ ∈ \{} (or ∀ ∈ \{( )} at the country level) ,€ € (5) �, ∑∈ , ≡ , where �, = ̂, 37 The previous 7-year spending period (2007-2013) allowed spending in the two years after the program has finished, so an equal allocation would be , = 1/9; the last programming period (2014-2020) allowed spending until 2023, so an equal annual allocation would be 1/10. 38 More specifically, the amount of funds committed are fixed at the start of a programming period, and total disbursements cannot exceed commitments. However, these can be underutilized. Our data record the total amount spent and not the original commitments, and so it is possible that the underutilization rate could be affected by economic shocks. However, Durand and Espinoza (2021) in their Figure 7 (p44) show that the ratio of disbursements to commitments is very high and almost always exceeds 90%, suggesting this is less of a concern. Moreover, any underspending in the construction of our instrument is spread across all years within the programming period, reducing the impact in any single year. 39 For example, spending shares in the Yugozapaden NUTS2 region of Bulgaria (which contains the capital city of Sofia), are instrumented by spending shares in all other NUTS2 regions in CEE (including other NUTS2 regressions in Bulgaria). 40 For example, spending shares in Bulgaria are instrumented by the spending shares in the NUTS2 regions in all other CEE countries except Bulgaria over the same programming period. 11 ,€ For country-level regressions, we generate country-level predicted payments , as the sum of all NUTS2 ,€ € 41 level payments in the same country c: , ≡ ∑∈ , . , , The final step in creating the instrumental variable is calculating ( , − ,−1 )/,−1 , which involves first € , deflating the nominal predicted spending expressed in euros using a deflator for that country , : , = ,€ € , /, and then expressing the predicted change in spending as a share of local GDP. This yields our first- stage regression in equation 6. For many of our regressions, is quite close to 1 (predicted and actual EU spending track each other one-to-one) and is highly significant (F-stats above 500), but sometimes – particularly for CEE at the country level – is close to zero and insignificant (we discuss this further below). The second-stage regression is as in Equation 1. , , (6) First stage: ( , − ,−1 )/,−1 = ( , − ,−1 )/,−1 + , + + + , A practical issue is the set of regions used to predict these spending shares in the regression in Equation 3. In our default specifications at the country level and at the subnational level, R includes all regions at the same level in the same CEE or non-CEE group (excluding the NUTS2 region i or country c(i)). The main reason for the CEE/non-CEE split is economic (discussed below), but it is also convenient for the instrument to be identical in the split CEE/non-CEE and the pooled samples. The first-stage regressions usually have stronger first-stage regressions at the NUTS2 level than at the country level (and a value of closer to 1). There is important country-level and CEE/non-CEE variation in how the spending shares are allocated over time, which likely reflects the country’s implementing capacity relative to the size of the EU funds it has to spend, as well as cultural and institutional factors. The spending shares for the 2000-06 programming period at the country level are plotted in Figure 2A and 2B. Spending in the CEE group tends to peak later in the programming period around 2006, with very little spending in the first few years. This suggests that CEE countries may have difficulties absorbing the large EU funds they are entitled to. 42 In contrast in most non- CEE countries, spending follows a symmetric inverted-U shape which peaks around the middle of the programming period, likely because of their higher implementing capacity and small share of EU funds. 43 Consequently, it is important to instrument spending shares within each of the CEE and non-CEE groups rather than across Europe as a whole. In the robustness section we also consider calculating shares using four groups, where each of the CEE/non-CEE groups is broken down into a northern group and a southern group. The results are discussed in Section 4.3. 41 As a robustness test, we instead generated the country-level instrument using only country-level data (i.e. not just aggregating the NUTS2 instrument to the country level). More specifically, when we estimate a version of Equation 4, but at the country-level rather than NUTS2 level, results are almost identical to those in Table 1 for Country Level IV regressions (not reported). 42 Bulgaria and Romania joined the EU in 2007. From 2001-2006 they were eligible for CF payments, but only became eligible for EAFRD, ERDF, and ESF payments from 2007. This may explain the boost in payments at that time. 43 Nonetheless, there are exceptions, such as Greece, with spending that peaks later than other non-CEE countries. 12 Figure 2A Figure 2B Notes: Luxembourg, Malta, Cypress, Lithuania, Ireland and the United Kingdom not shown to limit figure size. Country-level spending shares across all programs. 3. Results 3.1 Main Multiplier Estimates (GDP) Our main multiplier estimates are displayed in Table 1, which show that we find little evidence of a large contemporaneous multiplier on EU funds and often the multiplier is close to zero or is insignificant. In most cases we can reject a multiplier above 1. There is some weak evidence that multipliers are larger at the 13 country level than subnational level. While OLS estimates are often statistically larger for CEE than non-CEE countries at the subnational level, for all other specifications they are insignificantly different. 44 Panel A of Table 1 presents OLS estimates of Equation 1, which are subject to the caveat that there may be some endogeneity on the timing of disbursements (though total disbursements are fixed in advance at the beginning of the EU programming period). Panel B1 presents the second-stage IV estimates (Equation 1, but using predicted rather than actual payments), and Panel B2 reports the IV first-stage estimates of . 1 is the impact multiplier (our coefficient of interest), the relative € increase in GDP in year t from a relative 1€ increase in EU funds. Table 1: Main Impact Multipliers for EU Structural and Investement Fund Spending NUTS2 Subnational Regions Country Level 1. All Europe 2. CEE 3. Non-CEE 4. All Europe 5. CEE 6. Non-CEE Panel A: OLS Impact Multiplier 0.15** 0.67*** -0.06 0.80** 1.22 0.20 (β0) (0.08) (0.17) (0.11) (0.29) (0.68) (0.47) Obs 3,930 921 3,009 416 161 255 Controls/FEs: Year FE, State FEs, Lagged GDP Growth Panel B.1: Instrumental Variables (Second stage) Impact Multiplier -0.02 0.13 -0.14 0.67* 4.89 0.26 (β0) (0.09) (0.21) (0.15) (0.34) (3.43) (1.19) Obs 3,930 921 3,009 416 161 255 Controls/FEs: Year FE, State FEs, Lagged GDP Growth Panel B.2: Instrumental Variables (First stage) Coeff (θ) on Leave- 0.90*** 0.90*** 0.80*** 0.82*** 0.35 0.66*** one-out spending (0.03) (0.04) (0.10) (0.09) (0.23) (0.14) F-stat 684.1 546.6 65.28 91.48 2.296 22.85 Notes: Robust Standard Errors shown in parentheses (clustered at the NUTS2 for NUTS2 regressions). OLS and IV (second stage) are regressions on the annual growth rate of real GDP on the scaled "modelled" EU spending. Country level regressions exclude influential observations Croatia (2016) and Estonia (2011) discussed in the text. Dependent variable is winsorized at the 90% level at NUTS2 level and at 98% level at country level. *, ** and *** stars indicate significance at the 10%, 5% and 1% levels (respectively). CEE is Central and Eastern Europe. See Equation 1 (OLS/IV second stage) and Equation 6 (IV first stage). At the subnational level (Table 1, Columns 1-3) most multiplier estimates are insignificant and close to zero, except for a modest and significant multiplier of 0.67 for Central and Eastern European (CEE) countries in the OLS specification (Panel A, Column 2). These estimates are generally “precise zeros” – standard errors range from 0.08 to 0.21 – which means insignificance is driven by the small point estimates and not by lack of precision.45 We can always reject that multipliers at the subnational level are above 1 at the 95% significance level. Interestingly, the first stage at the subnational level is extremely strong for CEE countries, with a θ coefficient in Equation 5 of close to unity, suggesting that on average a 1€ increase in predicted EU spending increases actual spending by close to 1€. The first-stage [weak instrument] F statistics are over 500 for the 44 OLS CEE point estimates are also larger at the country level, but standard errors are too large for the difference to be significant. 45 For example, if the point estimate of the multiplier was 1 (rather than 0-0.6 as currently), then the t-statistics in the subnational sample would range from 4 to 12. 14 CEE and all-Europe specifications and are even above 50 for the non-CEE sample. This is much larger than the 12-28 range for a weak-IV F-statistics in Kraay (2014) and well above Staiger and Stock’s (1997) rule of thumb of 10. Country-level multiplier estimates (Columns 4-6) are generally larger than those at the subnational level, though they are typically less than one, more noisily estimated and significant only for the all-Europe sample. We focus on the all-Europe sample first, as with annual country-level data sample size is a limiting factor; wider standard errors lead to a general lack of significance for the country-level split CEE/non-CEE sample. For the OLS specification, the whole-Europe multiplier is 0.8 and is significant at the 5% level (Panel A, Column 4). The all-Europe IV estimates are marginally smaller at 0.67, but 50% wider standard errors mean significance is now at the 10% level. The first-stage coefficient is smaller than for the NUTS2 subnational � = 0.8), with a first stage F-statistic above 80. As foreshadowed earlier, both CEE regressions, but still large ( and non-CEE multipliers are always insignificant, however the CEE point estimates are larger than those for the non-CEE group. The insignificance of the CEE 2SLS multipliers is because the first stage is weak. 46 Robustness to reported EU spending. In the discussion above, we emphasized that one advantage of our approach was using modeled rather than reported EU spending, which is likely to be closer in time to when the actual spending takes place. In Online Appendix Table A3.1 we re-estimate our results using reported EU spending. At the NUTS2 level for All-Europe and non-CEE, results are fairly similar: small or insignificant multipliers at the 5% level for OLS and IV specifications. However, at the NUTS2 level for CEE countries the raw spending multipliers are sometimes a little higher, though never close to 1. More specifically, the NUTS2 CEE 2SLS multiplier is now 0.53 and significant at 1% (up from 0.13 and insignificant in Table 1), and the OLS CEE multipliers are now 0.47 (down from 0.67 in Table 1). At the country level, estimates using raw spending are similar for raw and modeled spending for IV specifications — both insignificant — but more volatile for the OLS specifications across samples using raw spending. More specifically, OLS CEE multipliers are now positive and significant, but OLS non-CEE multipliers are negative and significant (both were previously insignificant using modeled spending In Table 1). Positive and negative multipliers are difficult to rationalize and might be due to mismeasurement in the raw spending data. 3.2 Effect on Investment One possible explanation for the modest or null effects of EU funds on regional GDP growth is that EU funds do not affect investment spending. 47 This could be because of measurement issues, or perhaps EU funds are used to finance projects that would have been implemented anyway. To test these hypotheses, Table 2 presents the results of a different regression specification, where on the LHS, the growth rate of GDP (from Table 1) is replaced by the change in total investment as a share of GDP (Table 2). The RHS and IV approach stays unchanged. 48 The regression now has an “investment multiplier” specification, namely the € total investment generated by an extra 1€ of EU funds. We focus on total investment (rather than public investment) as public investment is not available at the NUTS2 level. 46 We tried weighting the country-level regressions by the number of NUTS2 regions and we got a similar weak first stage for CEE countries as in column 5 of table 1 (not reported). This suggests that the strong first stage for CEE countries at the NUTS2 level is due to variation in EU funds at NUTS2 level rather than the weighting of the different countries. 47 Most projects they are supposed to fund might be classified as investment, especially public investment. 48 The results are similar, though slightly smaller, if lagged change in investment as a share of GDP is added as a control (rather than the lagged growth rate of GDP as in equation 1). 15 Table 2: Effect of EU funds on Total Investment NUTS2 Subnational Regions Country Level 1. All Europe 2. CEE 3. Non-CEE 4. All Europe 5. CEE 6. Non-CEE Panel A: OLS Impact Investment 0.63*** 0.95*** 0.14 0.97*** 1.22** -0.11 Multiplier (β0) (0.12) (0.17) (0.17) (0.30) (0.39) (0.41) Obs 3,897 921 2,976 416 161 255 Controls/FEs: Year FE, State FEs, Lagged GDP Growth Panel B.1: Instrumental Variables (Second stage) Impact Investment 0.92*** 1.01*** 0.57** 1.17*** 3.68 -0.09 Multiplier (β0) (0.13) (0.24) (0.23) (0.41) (3.19) (1.19) Obs 3,897 921 2,976 416 161 255 Controls/FEs: Year FE, State FEs, Lagged GDP Growth Panel B.2: Instrumental Variables (First stage) Coeff (θ) on Leave- 0.90*** 0.90*** 0.80*** 0.82*** 0.35 0.66*** one-out spending (0.03) (0.04) (0.10) (0.09) (0.23) (0.14) F-stat 683.6 546.6 64.89 91.48 2.296 22.85 Notes: Robust Standard Errors shown in parentheses (clustered at the NUTS2 for NUTS2 regressions). OLS and IV (second stage) are regressions on the annual change of investment as share of GDP on the scaled "modelled" EU Spending. Country level regressions exclude influential observations Croatia (2016) and Estonia (2011) discussed in the text. GDP growth (as a control) is winsorized at the 90% level for NUTS2 regions and at 98% for country level regressions (though investment data is not winsorized). *, ** and *** stars indicate significance at the 10%, 5% and 1% levels (respectively). CEE is Central and Eastern Europe. See Equation 1 (OLS/IV second stage) and Equation 6 (IV first stage). Table 2 shows that EU funds tend to boost investment, with all-Europe and CEE total investment multipliers often being close to 1 (against smaller and insignificant investment multipliers for non-CEE countries). This suggests that we are capturing actual changes in investment driven by EU funds. At the NUTS2 level, we find that an extra 1€ of EU funds spending in CEE countries boosts total investment by around 1€ (OLS and IV) with both being statistically significant at the 1% and we are unable to reject an investment multiplier of 1. At the country level, CEE OLS multipliers are also close to 1, but IV multipliers are imprecisely estimated due to weak instruments. For the all-Europe sample the subnational- and county-level investment multiplier is significant and usually around 1 for both the OLS and 2SLS specifications except for an OLS NUTS2 multiplier which is smaller at around 0.6 (though still significant). For non-CEE countries, investment multipliers are smaller and sometimes insignificant (IV of 0.57 significant at 5% being larger than insignificant 0.14 of OLS), and for OLS are significantly smaller than CEE investment multipliers. This is not surprising as most EU structural fund spending is focused on Central and Eastern Europe. Robustness. We conduct three robustness tests of the investment multiplier results: (i) using raw (rather than modeled) payments, (ii) splitting total investment into public investment and private investment portions (which is only available at the country level) and (iii) using winsorized investment data (our default specification is non-winsorized). First, if we used raw EU payments on the right-hand side (instead of using modeled EU payments), the estimated investment multipliers are substantially smaller in all cases (see Online Appendix Table A3.2). As measurement errors often lead to attenuation bias, this suggests that modeled payments may be more 16 accurate than raw payments. For example, the OLS subnational investment multiplier for the whole of Europe is 0.63 for modeled payments in Table 2, but only about half as large (0.38) when calculated using raw payments (Online Appendix Table A3.2). The corresponding IV multipliers are 0.92 using modeled payments, but 0.72 using raw payments. Multipliers are also smaller at the country level using the raw payment data in the all-Europe sample; they are almost half as large using OLS (0.97 from Table 2, 0.6 from Online Appendix Table A3.2), and IV estimates lose significance using the raw payments. 49 Second, Online Appendix Table A3.3 reports investment multipliers winsorizing investment data at the 90% level for NUTS2-level data and at the 98% for country-level data (analogous to approach for GDP). Results are broadly similar to those in Table 2; investment usually responds robustly to EU spending. For subnational regions (Columns 1-3), investment multipliers with winsorized data are about 0.1-0.2 smaller than those in Table 2 (about one standard error) though are still significant and at least 0.5 for most specifications. Country- level investment multipliers with winsorized data are typically 0.05-0.2 larger than those in Table 2. Finally, Online Appendix Table A3.4 (only at the country level) shows that a 1€ increase in modeled EU fund spending increases public investment by about 0.6€ for all countries, driven by effects in CEE countries. In contrast, effects on private investment are always insignificant. For non-CEE countries both public and private investment multipliers are insignificant and usually close to zero. These estimates reinforce the findings that EU funds are actually spent, usually on public investment, but there is not much evidence of crowding-in of private investment (neither there is of crowding-out private investment, as shown earlier). 3.3 Cumulative Multipliers over Longer Horizons So far, we have found that while EU funds tend to stimulate investment, they have only modest immediate effects on GDP in the first year of the construction/implementation period. One possible explanation is that the one-year horizon considered above might be too short to observe an effect, which might appear over longer time horizons. There are two main arguments for exploring a delayed effect of EU funds on GDP. First, it could be that examining short windows misses some of the benefits of EU spending – newly-employed construction workers on an EU-funded road project might take some time to spend their wages and further boost demand. Indeed, Coelho (2019) and Durand and Espinoza (2021) find some evidence of larger multipliers over longer horizons. 50 Second, there might be residual concerns about some mismeasurement of the timing of spending, and so smoothing multipliers over several years would reduce mismeasurement. 49 Likewise, country-level CEE and non-CEE investment multipliers are always smaller (and more often negative) using raw payments than modeled payments, though for the non-CEE sample both are insignificant. 50 Coelho (2019) reports a medium-run multiplier of 4 after 3 years, but this is not a cumulative multiplier because it measures the cumulative response of output to a one-time spending shock. As spending is also very persistent, Coelho says she finds a cumulative multiplier of 0.9. Durand and Espinoza (2021) estimate cumulative multipliers of 1.8 over a two-year horizon, up from 1.2 contemporaneously. Canova and Pappa (2022) find cumulative multipliers above 1 after 3 years for the ESF program. 17 Table 3: Two-year Cumulative Multipliers for EU Fund Spending NUTS2 Subnational Regions Country Level 1. All Europe 2. CEE 3. Non-CEE 4. All Europe 5. CEE 6. Non-CEE Panel A: OLS Cumulative 2 year 0.14 0.62*** -0.01 0.68 1.13 -0.23 Multiplier (β2) (0.10) (0.20) (0.14) (0.41) (0.80) (0.53) Obs 3,698 866 2,832 389 149 240 Controls/FEs: Year FE, State FEs, Lagged GDP Growth Panel B.1: Instrumental Variables (Second stage) Cumulative 2 year -0.04 0.05 0.02 0.64 4.88* -1.19 Multiplier (β2) (0.13) (0.29) (0.23) (0.39) (2.85) (1.21) Obs 3,698 866 2,832 389 149 240 Controls/FEs: Year FE, State FEs, Lagged GDP Growth Panel B.2: Instrumental Variables (First stage) Coeff (θ) on Leave- 0.92*** 0.92*** 0.89*** 0.84*** 0.53** 0.78*** one-out spending (0.03) (0.03) (0.11) (0.07) (0.26) (0.19) F-stat 840.2 718.2 71.74 168 4.055 17.47 Notes: This table is the same as Table 1, except that instead of the impact mutliplier we estimate cumulative multipliers as in Equation 2. The estimated two-year multipliers are displayed in Table 3 and suggest that cumulative multipliers are generally insignificantly different from zero and point estimates are slightly smaller than the impact multipliers in Table 1. 51 OLS multipliers in Panel A are generally insignificant, except for a subnational CEE sample, which is significant at the 1% level (point estimate of 0.62). 52 IV estimates in Panel B are always insignificant at the 5% level. The subnational multipliers are “precise zeros” where we can reject a multiplier above 1 at the 95% level, though country-level multipliers in the CEE/non-CEE subsamples are imprecisely estimated. Except for the CEE country-level sample, first-stage coefficient estimates are close to 1 and F-stats are very large. 53 Three-year cumulative multipliers are generally either insignificant or can even turn slightly negative (Online Appendix Table A3.6). 51 Two-year cumulative multipliers are the sum of the impact and the “long difference” multipliers which take the two- ,+−1 −,−1 � ,+−1 −,−1 � year difference in both GDP and EU spending: = + , + + + , . The long- ,−1 ,−1 difference estimates, displayed in Online Appendix Table A3.7, are smaller than the contemporaneous multipliers. ,+−1 −,−1 52 An alternative specification is regressing growth over longer periods on the one-year spending shock: = ,−1 �, −,−1 � + , + + + , . Online Appendix Table A3.8 shows that these “forward multipliers” are ,−1 insignificant at the country level, but similar (OLS) or much larger and significant (2SLS) at the sub-national level. However, because EU spending shocks tend to be positively correlated over time, this multiplier specification overstates the economic effect of spending shocks. , , 53 The instrument here is the sum of cumulative change in predicted spending: ∑−1 ℎ=0 ( ,+ℎ − ,−1 )/ ,−1 . Cumulative multipliers also drop influential observations that appear in the forward years: for example, we drop Croatia 2016 in for impact multipliers, Croatia 2015-16 in the two-year multipliers and Croatia 2014-16 in three-year cumulative multipliers as all these observations include the initial Croatia 2016 influential observation. 18 4. Robustness Tests, Anticipation Effects and Heterogeneity 4.1 Robustness Tests This section presents an additional set of results to investigate whether our main results (presented in Table 1) are robust to additional controls or ways of identifying outliers, and also compares to the literature. The result tables are in Online Appendix 4. Alternative treatment of outliers and influential observations. This subsection tests the robustness of our findings to alternative assumptions about winsorizing growth data and excluding influential observations. Online Appendix Table A4.1 Column 1-6 starts this analysis by re-estimating Table 1 including both influential observations (EE2011 and HR2016 in country-level regressions) and not winsorizing any growth data. Even in this case, we find little evidence of large multipliers (>1), with the exception of IV country-level results for CEE countries, which are now very large (2.36) and significant at 5%. This large and significant coefficient is due to the inclusion of one influential observation (specifically Croatia 2016); in Column 8 we remove both influential observations, and the CEE multiplier becomes insignificant (because the first stage is now weak). Including influential observations and not winsorizing also makes multipliers for OLS CEE (subnational) and All Europe (country level) insignificant at the 5% level (they were significant in Table 1). The insignificance of the former is due to the larger standard errors/smaller coefficient with winsorized data, and the insignificance of the latter is due mostly due to the inclusion of influential observation Estonia 2011 (excluding influential observations makes the latter significant in column 7). An alternative approach is to winsorize at a different level. In our default results, we winsorized at the 90% level for the more-volatile NUTS2 data, and at the 98% level for the country-level data. Reversing this in Online Appendix Table A4.2 so NUTS2 data are winsorized at 98%, and country-level data are winsorized at 90%, generates broadly similar results (few multipliers above one). While a more aggressive winsorizing level (lower percentage) often leads to marginally larger estimates with slightly smaller standard errors, results are still similar to our baseline, even winsorizing at the 80% level (Online Appendix Table A4.3). Alternative Controls. Our default specification is parsimonious; it only includes lagged GDP growth as a control (in addition to time and region fixed effects). In this section we show that our results are robust to a range of controls used in the literature. First, we add the change in the institutional index (also used by Durand and Espinoza 2021). 54 Results are almost identical to those in Table 1 (not reported). A second possible control is population growth, as much of the literature seeks to explain per capita GDP growth rather than aggregate GDP growth (as we do in Equation 1). Adding population growth (sourced from Eurostat) has almost no effect on the results; significance at the 5% is the same as in the baseline results, and most estimates are almost identical (not reported). 54 To get the institutional index, we use data from the International Country Risk Guide (ICRG). In particular, we construct an index of institutional quality as an average of subindices of government effectiveness, the rule of law, regulatory quality, and control of corruption, where each subindex is first normalized to be between zero and one. 19 Comparison with Durand and Espinoza (2021). Our finding of a lack of a large multiplier at the country level is somewhat surprising given Durand and Espinoza (2021) found all-Europe multipliers of around 1.2 using similar country-level data and empirical methodology. 55 However, the country-level part of our paper still differs from theirs in several dimensions: their sample is longer than ours (including the 1994-99 programming period, whereas we start in 2002 as this is when NUTS2 GDP sample starts, allowing for lags) and they include some extra countries (like the UK, IE and LT), they have several extra controls (lagged change in institutional quality; lagged change in financial risk), they exclude 2009 (we do not), they drop 7 influential observations based on austerity programs and judgement 56 (we drop 2 observations based on one-by-one exclusion regressions), they drop extreme standardized residuals (we winsorize the data), they use a subsample of 5 CEE countries based on similarity to Slovenia (we consider 10) and our instrument construction varies slightly. To try to understand which of these factors drive the different results we replicated their country-level all- Europe results, and then changed modeling assumptions one-by-one from the replication. The first column of Online Appendix Table A4.7 presents their all-Europe results: a contemporaneous GDP country-level of 1.2. Applying their sample (which does not include payments from FEAD and YEI funds), outliers, controls and a similar construction of the instrument, we find a similar multiplier of 1.26, also significant at the 1% level. We are also able to replicate their investment multiplier of 1.5. The remaining columns of Online Appendix Table A4.7 make changes that move towards our sample and specification. The most important difference seems to be that Durand and Espinoza exclude 2009, whereas we do not. Specifically, in column 3 we include 2009 (by removing their control variable of spending X a 2009 dummy), the all-Europe 2SLS multiplier falls to 0.8 and significance falls to the 5% level. This is within one standard error of our all-Europe 2SLS estimate of 0.67, which is borderline insignificant at the 5% level. It is debatable whether 2009 should be included or excluded in the sample — Europe clearly experienced a large growth shock that period, but our specification already includes time fixed effects, which remove bulk the effect of the recession (the 2009 time fixed effect has a value of -7% and is significant at the 1% level). 57 Other changes appear to be less important and generate estimates similar to those in Durand and Espinoza (2021). In Table A4.7 column 4 we start with our replication of Durand and Espinoza but then start the sample in 2000, which yields large all-Europe multipliers of around 1.3. Likewise removing their outliers and/or their influential observations (Columns 5 and 6) continue to generate large 2SLS multipliers of 1.3. 58 4.2 Spillovers from Spending in Neighboring Countries One of the possible reasons for a small multiplier estimated above could be that the higher demand in a region from EU spending is leaked to surrounding regions: for example, a firm constructing an EU-funded road in a region might employ workers or buy supplies from a nearby region. Indeed, Coelho (2019) finds a 55 They find an all-Europe OLS country-level multiplier of 0.9, which is fairly similar to our default multiplier of 0.87 for the same sample (both significant), so we do not discuss this further in the text. 56 Cyprus (2012, 2013, 2014), Portugal (2011), Spain (2011), Ireland (2011, 2015, 2016), Finland (2015, 2016). 57 Instead, it is the cross-sectional correlation between the change in EU spending and the growth that is driving the lower multiplier estimate. 58 We were also able to replicate a significant CEE 2SLS multiplier of 1.9 using their specification. However, the size of that multiplier is entirely dependent on including Croatia in 2016. Without including it, the multiplier becomes insignificant and the first stage becomes weak. 20 strong effect of import-weighted expenditure on local growth, Canova and Pappa (2022) find larger multipliers when estimating using national spending and Crucitti et al. (forthcoming) find many of the long- term benefits of cohesion fund spending operate through spillovers. Motivated by these concerns, this subsection tests for the spillover effects. Unlike Coelho (2019) and Canova and Pappa (2022) we focus on spillover from neighboring NUTS2 regions (rather than import-weighted spending, or national spending respectively), as this is where the spillovers are likely to be the largest. Our specification involves adding an extra term 1 � , − ,−1 �/,−1 to the regression in Equation 1, where , is the total real spending in all NUTS2 regions neighboring region i. The first step in the analysis is to identify the set of NUTS2 regions that share a border with region i. , is € € calculating by summing the nominal modeled payments in euros (, ), for NUTS2 region ∈ ; , = € , , ∑∈ , . For 2SLS specifications, we instrument �, − ,−1 �/,−1 with �, − ,−1 �/,−1 , where , ,€ , is predicted spending in neighboring regions, calculated by adding up predicted spending , of ,€ ,€ , neighbors from equation 5 (i.e. , = ∑∈ , ). Real versions , and , are calculated dividing by € nominal neighbors spending by the country-level deflator in the region of interest ( (), ). The first stage , , regressions for � , − ,−1 �/,−1 and �, − ,−1 �/,−1 now include both � , − ,−1 �/,−1 and , , � , − ,−1 �/,−1 , with the coefficients on the predicted own (neighbors) spending being close to one in their own (neighbors) regression, and first stage F stats are large (bottom panel of Table 4). Table 4: NUTS2 Multipliers with Spillovers from Neighbors’ Spending 1. All Europe 2. CEE 3. Non-CEE Own Region Neighbors' Own Region Neighbors' Own Region Neighbors' Spending (β1) Spending (φ1) Spending (β1) Spending (φ1) Spending (β1) Spending (φ1) Panel A: OLS Multiplier 0.14 0.00 0.53*** 0.04* -0.01 -0.04** (β1 or φ1) (0.09) (0.01) (0.19) (0.02) (0.11) (0.02) Obs 3,930 921 3,009 Controls/FEs: Year FE, State FEs, Lagged GDP Growth Panel B.1: Instrumental Variables (Second stage) Multiplier -0.05 0.01 0.06 0.01 -0.08 -0.03 (β1 or φ1) (0.13) (0.02) (0.27) (0.02) (0.18) (0.03) Obs 3,930 921 3,009 Controls/FEs: Year FE, State FEs, Lagged GDP Growth Panel B.2: Instrumental Variables (First stage) Regression Own Region Neighbors' Own Region Neighbors' Own Region Neighbors' explaining: Spending Spending Spending Spending Spending Spending Coeff on own 0.88*** -0.08 0.94*** 0.20 0.79*** -0.22 spending (0.07) (0.17) (0.05) (0.22) (0.12) (0.15) Coeff on neighbor 0.00 0.97*** -0.01 0.91*** 0.00 1.06*** spending (0.01) (0.04) (0.01) (0.06) (0.01) (0.07) F-stat 78.09 238.9 22.28 Notes: Robust Standard Errors shown in parentheses (clustered at the NUTS2). OLS and IV (second stage) are regressions on the annual growth rate of real GDP on the scaled "modelled" EU Spending and neighbor "modelled" EU spending. Dependent variable is winsorized at the 90% level at NUTS2 level. *, ** and *** stars indicate significance at the 10%, 5% and 1% levels (respectively). CEE is Central and Eastern Europe. 21 Table 4 shows little evidence of short-run spillover effects from the EU spending in neighboring NUTS2 regions. Specifically, the estimated multipliers on own-region spending are not larger or more significant than those in Columns 1-3 of Table 1 and are sometimes slightly smaller or less significant. The multiplier on neighbors’ spending is either close to zero, insignificant, or both. 59 4.3 Anticipation Effects EU funds are different from many other fiscal policies, in that they are mostly anticipated. That is, the amount of spending for EU funds in different countries (and sometimes NUTS2 regions) is allocated at the start of the 7-year EU programming period rather than year-to-year and quarter-to-quarter. In contrast, typical fiscal stimulus packages are implemented rapidly during a recession and so may be more of a surprise. Anticipation effects can mean a smaller multiplier when policies are implemented, because firms start to adjust in advance (by either raising prices or increasing capacity and employment). In the EU fund literature, Coelho (2019) tests for anticipation effects, but does not find any. We test for anticipation effects in a similar way as Coelho (2019)—by regressing GDP growth at t-1 on EU fund spending at t, as in Equation 7 (naturally we also remove lagged GDP growth from the specification). ,−1 −,−2 � , −,−1 � (7) = + + + , ,−2 ,−1 In contrast to Coelho (2019), we find some tentative evidence of negative anticipation effects (Online Appendix Table A5.1); that is, GDP growth is usually negative in advance of an increase in EU fund spending. More specifically, we find that for OLS specifications, there is a negative anticipation effect of around -0.36 at the NUTS2 level, and -0.95 at the country level for non-CEE countries (for OLS estimates, others are insignificant). While the NUTS2-level estimates are robust to using a 2SLS specification (and even more negative), the country-level multipliers for the 2SLS specification are insignificant. Anticipation effects of this size and sign are surprising theoretically but can be partially explained by the effect of anticipated EU spending effect on investment. Results in Online Appendix Table A5.2 regress investment growth in year t-1 on the change in EU funds (share of investment) in year t. Results suggest a reduction in investment at the NUTS2 level for non-CEE countries almost euro-for-euro, which is similar in both OLS and 2SLS specifications. In other words, it seems that for some non-CEE subnational regions, investment falls in anticipation of future EU fund spending. This could be because firms and governments delay projects until they get EU funding. In other words, although the strong contemporaneous investment multiplier in Table 2 means there is little conventional crowding-out, there could be a subtler dynamic crowding out in some regions, where investment is delayed until EU funding is approved. Nonetheless, falling investment does not explain all the negative anticipation effects on GDP; for example, anticipation effects of EU fund spending on investment is always insignificant for CEE countries, through there is some evidence of negative anticipation effects for GDP in the same sample. Negative anticipation effects not due to investment might be explained if firms started increasing prices (and so reducing output) in anticipation of higher demand, but the effects are not expected to be so large. More common in the literature are positive anticipation effects, as firms start to hire and invest in advance of higher demand, though in the EU funds case this might happen on announcement (at the start of the programming period), rather than a 59 The results are robust to dropping years, dropping countries, and dropping extreme neighbor observations. 22 year before spending. Either way, the estimated anticipation effects presented here represent an interesting area for further investigation. 4.4 Multiplier Heterogeneity Results by EU fund. Several other papers in the literature have found heterogeneous multipliers across different EU funds. Canova and Pappa (2022) investigate the effects of two funds, the European Regional Development Fund (ERDF) and the European Social Fund (ESF). The ERDF focuses on several priority areas, such as innovation, digitalization, support for SMEs, or the “green transition” whereas the ESF mostly invests in employment-focused education and training. Canova and Pappa (2022) find that ERDF has a positive impact multiplier of 1.4, falling to 0.2 after 3 years (cumulatively). In contrast they argue the ESF has a negative short-run multiplier, but a positive and large medium-run multiplier. Coelho (2019) estimates a large multiplier on funds allocated to “Objective 1” regions (regions with low per capita income, to encourage convergence), but an insignificant multiplier for those allocated to “Objective 2” regions (regions with structural changes). In this section, we disaggregate our main results in Table 1 by the type of fund, with the results presented in Online Appendix Table A6.1. 60 Perhaps due to the smaller sample size, we never get significant multipliers in the country-level sample, so we focus our discussion here only on the NUTS2 sample. The largest economic effects are from the CF, which for the whole of Europe has a significant multiplier of 0.86 (at 1%) for the OLS specification and of 1.21 (at 1%) for the IV specification (with a strong first-stage coefficient). However, CEE multipliers for the CF are insignificant—which is perhaps surprising given spending is focused on CEE regions—and makes us not to put too much weight on the all-Europe CF results. For the ERDF, we get an OLS multiplier of 0.39 for all Europe NUTS2 sample (significant at 1%) and of 1.22 for CEE countries (significant at 1%), but these are not robust in the 2SLS specification. All multipliers for EAFRD and ESF are insignificant at the 5% level. Results by programming period/split sample. The economic situation in Europe has changed substantially over our 2000-18 sample period (or starting in 2002, accounting for lags). Most important has been the European debt crisis and its aftermath, but also the expansion of the European Union to the east. The sample also includes three different programming periods of EU funds: 2000-06; 2007-13 and 2014-20 (though our sample finishes in 2018). In this subsection we test whether multipliers vary over time. Specifically, we either simply split our sample in the middle (allow to vary across the 2000–09 and 2010–18 periods); or split out sample across the three sample periods (while keeping other controls and fixed effects the same): , − ,−1 � ,−,−1 � = ∑ + , + + + , for p={2000–09, 2010–18} or p={2000-06; 2007-13; 2014-20} ,−1 ,−1 We find little robust evidence of large multipliers in a particular subsample. Online Appendix Table A6.2 reports the results for the mid-sample split. The occasionally significant OLS multipliers in Table 1 seem to be driven by the 2010-18 subsample. However, differences in estimated multipliers are never statistically significant at the 5% level across subsample. The very high 2SLS multiplier that is significant—country-level 60 Our variable of interest is now the change in modeled payments for fund f in region i (or country c) over lagged GDP growth and our instrument is calculated as in equation 5, but with spending for fund f (instead of for all programs). We also use the change in modeled payments for all the other funds except for fund f over lagged GDP growth as a control. 23 CEE for the 2001-09 period (borderline significant at the 5% level)—is not robust to excluding influential observations. 61 Online Appendix Table A6.3 instead splits samples by programming period. 62 2SLS multipliers are almost always insignificantly different from zero and are always insignificantly different across the three programming periods – perhaps due to the shorter sample. 63 Greater regional disaggregation. In this section we disaggregate beyond the CEE/non-CEE classification in Table 1 into North and Southern Regions. There are two reasons for this. First, the economic effects (structure of economies, efficiency of spending etc.), might be different in Northern and Southern Europe. The differences between Northern and Southern Europe became stark following the European debt crisis around 2010 and continue to drive the political discussion in Europe. 64 Second, it could be that the size and dynamics of EU spending differs in the Northern and Southern parts of Europe, for example: Spain, Portugal, Italy and Greece should not be grouped with richer Norther countries based on how they spend EU funds. Indeed, Figure 2 shows some differences between the spending patterns of Bulgaria and Romania and other CEE countries. Naturally, the disadvantage of greater disaggregation is smaller sample sizes, especially at the country level. Online Appendix Table A6.4 shows that even with further disaggregation there is little robust evidence of large multipliers on EU spending. At the subnational level, OLS multipliers are all less than one or insignificant, as are IV multipliers. At the country level, there are occasionally some larger multipliers, such as 2.65 for Southern CEE using OLS, but they are not robust to the IV specification. Greater disaggregation also sheds some light on what has been driving the results in the main text. For example, Table 1 found positive and significant OLS subnational multipliers for CEE countries (albeit one less than 1), which is repeated in Online Appendix Table A6.4 (column 2, Panel A). The problems of weak instruments for CEE countries at the country level in Table 1 are in part due to a negative first stage for northern CEE countries (Column 3 Panel D). 5. Discussion: Mechanisms and Interpretation 5.1 Possible Mechanisms EU fund spending aims to boost sustainable long-term growth by supporting the transformation towards more sustainable green and digital EU economies. In the short term, it is also sometimes thought to stimulate economic activity through the standard Keynesian aggregate demand channels which, in theory, provides policy makers with a tool to boost GDP also in the short term, for instance, to support the recovery from the COVID-19 crisis. That is, an increase in government spending in a subnational region or country increases demand for locally produced goods and services. As prices and wages are sticky in the short term, firms respond to higher demand by increasing output (rather than increasing prices); for example they could hire 61 Specifically, if either Poland (2009) or Estonia (2016) or both are excluded from the sample, the estimated coefficient becomes insignificant. 62 To run that specification, we use the same data as Table 1. This means that we are getting the coefficient of the payments received those specific years, and so this means that payments received during the years of a certain programming period can belong to previous programming periods. 63 While the 2001-06 country-level IV multiplier is large and significant for the All-Europe sample, this is not robust: the multiplier becomes insignificant when dropping either Latvia or Estonia. 64 For a discussion on the structural differences within the eurozone, see, for example, Bagus (2012). 24 more workers, increase the utilization rate of those workers (through higher hours or working employees harder) or utilize existing capital more effectively. Higher local incomes could then encourage local consumption – if households are credit constrained or higher incomes are expected to be persistent – which then further increases demand and output. But we do not find evidence of large Keynesian multipliers. So why might the Keynesian mechanism be weak? The first possible reason is that, unlike most of the fiscal policy literature, EU fund spending is anticipated. EU fund allocations are announced in advance for the 7-year programming period with commitments by country and/or subnational region (though there is still some uncertainty about whether individual projects might be approved until closer to implementation). This means that firms might be able to start adjusting pricing and capacity in advance, potentially reducing the multiplier. In an extension in Section 4.3, we analyzed the effects of EU spending the year before it was recorded and found a negative anticipation effect on GDP. This might be consistent with some price adjustment, but less consistent with increases in output (adding employment or capacity) in advance of EU-funded projects. A second possible explanation is some form of fungibility or crowding out – the EU funds projects that would have happened anyway. In Table 2 we showed that for most cases – especially in CEE countries – investment does actually increase contemporaneously in response to EU funds, reducing the relevance of this explanation. However, we also find some evidence in favor of a dynamic crowding out: in regions of non-CEE countries investment falls in anticipation of EU spending. In these cases, it could be that the availability of EU funding has the unintended effect of simply delaying projects that would have happened anyway. We only have tentative evidence for this mechanism, but it is an interesting topic for future research. The third possible reason is the presence of demand leakages—a substantial fraction of demand might accrue outside the region with the increase in the EU spending. This is the largest concern for NUTS2 subnational regions, as they are the smallest hence the most open on average. However, the analysis in Section 4.2 finds little evidence of sizable spillovers to neighboring regions—where these leakages are expected to be the largest. Nonetheless, some of the demand might accrue to regions further away, or even to countries outside the EU. The final possible reason for an absence of large multipliers is that prices or wages are not that sticky and so demand leaks to changes in prices rather than output. 65 While there is a large literature estimating price stickiness in Europe, which mostly finds that prices are indeed sticky (see Alvarez et al. 2005 for a survey), most of this applies to Western Europe, rather than CEE countries where the largest EU funds are disbursed. In addition, there is substantial new evidence in favor of the longstanding notion that wage stickiness is asymmetric – wages are only downwardly sticky, but upwardly flexible (Schmitt-Grohe and Uribe 2016). Barnichon et al. (2022) argue that for this reason, multipliers would be larger for spending cuts or during downturns, but multipliers would be smaller for spending increases or during normal times. As the EU fund spending has predominantly taken place during normal times and is an increase rather than decrease in spending, this could explain an absence of large multipliers, as the spending may have added to wage growth rather than output. Indeed Coelho (2019) finds that wage income increased substantially in response to EU funds, even though employment did not. 65 This would mostly explain country-level results, as the NUTS2-level growth rates are only deflated using country-level deflators (rather than by NUTS2-level deflators, that would capture local price inflation, but are not available). 25 However, there are several factors that might explain why empirical papers should estimate larger multipliers (even if multipliers are actually small). The most important argument is that EU Fund spending usually requires local co-financing, which means that the actual total expenditure is larger than the measured EU funds’ spending. The size of the co-financing is usually smaller for lagging regions (especially in Central and Eastern Europe), but co-financing is higher for richer regions in Western Europe (though this also depends on the specific EU fund and other details). The implication is that the size of multipliers might be overestimated, due to the understatement of total expenditure (actual multipliers should be less than 100% of the estimated size, especially in the non-CEE subsample). 66 The second factor is that EU funds are financed by joint taxation in all parts of the EU, rather than taxation at the national or subnational level. In a New Keynesian model, this would tend to increase the size of fiscal multipliers (though in a neoclassical model it would reduce them, as the negative wealth effects on labor supply drive a positive multiplier). Pennings (2022) shows that the size of the boost to the multiplier in a New Keynesian model due to external financing depends on the persistence of the stimulus. Most types of fiscal stimulus are relatively transitory, and so the boost to the fiscal multiplier is small. However, EU funds are much more persistent, and so the gains from external financing could be larger. 5.2 Interpretation: Relative versus Absolute Multipliers The multipliers identified in this paper are open-economy relative multipliers (as in Coelho 2019, Durand and Espinoza 2021, Nakamura and Steinsson 2014, and Pennings 2021), which are different from the closed- economy multipliers often discussed in the literature. The open-economy relative multiplier captures the change in GDP of region A vs region B in comparison to size of the relative fiscal shocks in the two regions. That is, only the difference between the change in spending in region A and B appears on the RHS of the estimated equation, and only the difference in growth appears on the LHS. This has the advantage that it makes multiplier estimates more robust, because other aggregate factors like monetary policy or foreign shocks are differenced out.67 The disadvantage of the open-economy relative multiplier is that it does not capture the complete output effect in which most policy makers may be interested in; that is, whether fiscal stimulus increases growth in an absolute sense. Specifically, a large open-economy relative multiplier is consistent with fiscal stimulus that reduces growth in all regions (for example, due to higher taxes), but just reduces growth by less in regions that received more fiscal spending. For this reason, one has to interpret the policy implications of fiscal multipliers with some care. 66 For example, if there was a 15% local copay, then the estimated multipliers – as they are usually defined – would be roughly 15% larger than those estimated in Table 1, because the measured change in spending is only the EU component, which understates the total increase in spending. 67 For example, consider two fiscal shocks, one under an accommodative monetary regime (like at the zero lower bound), and another under a less accommodative regime (a Taylor rule). The open-economy relative multiplier will be the same multiplier under both regimes, whereas the closed-economy aggregate multiplier will be much larger under more accommodative monetary policy (Nakamura and Steinsson 2014). 26 6. Conclusions and Policy Implications Motivated by the post-pandemic expansion in European Union spending—the NextGenerationEU (NGEU) recovery package—this paper estimates the effect of historical EU structural and investment fund spending on regional GDP growth in Europe. In order to assuage concerns about the endogeneity of EU funds, we instrument the time allocation of spending by spending in other regions (a leave-one-out instrument). In contrast with much of the recent literature, we find little evidence of large relative short-term multipliers on EU fund spending at either the national or subnational (NUTS2) level. Estimated multipliers vary across samples and are occasionally significant, but even significant estimates are less than one. Using our preferred instrumental variable specification, we find a precisely estimated subnational multiplier of zero, with country- level multipliers also being insignificant (though estimated less precisely). This is despite a strong contemporaneous response of investment to EU funds, which often increases euro-for-euro. The implication of our results is that policy makers should have more modest expectations of the short-run stimulatory effects of EU structural and investment funds on regional economic growth in regions receiving large inflows—and instead could focus on the fund’s other longer-run benefits, consistent with the original purpose of these funds. While there are other papers in the literature that find large stimulatory effects of EU funds on regions, our results suggest that the relative multipliers are generally less than one or close to zero in the short term. If our multipliers were to be applied to NGEU spending, they would suggest that the boost to short-term growth in regions with relatively high NGEU spending would only be modestly larger than in those regions with low NGEU spending (other things equal). While policy makers should never put too much weight on a single study, our results do encourage some tempering of growth expectations for policy makers hoping that EU spending could substantially boost short-run growth in Central and Eastern European countries receiving large inflows. Having said that, it is important to emphasize that our paper makes an argument only about the relative increase in growth in regions receiving more or less EU funding and does not make a statement about whether the NGEU program increases European short-run growth in the aggregate (or the value of the program in general). Most importantly, EU funds – if used efficiently – are expected to increase the long-run productive capacity of CEE countries or may help to achieve other important social welfare objectives. 27 References Álvarez, L., E. Dhyne, M Hoeberichts, C. Kwapil, H. Le Bihan, P. Lünnemann, F. Martins, R. Sabbatini, H. Stahl, P. Vermeulen and J. Vilmunen (2005) “Sticky Pries in the Euro Area: a Summary of new micro evidence,” ECB Working paper 563. Barnichon, R. D. Debortoli and C. Matthes (2022) “Understanding the Size of the Government Spending Multiplier: It’s All in the Sign” The Review of Economic Studies, 89(1): 87–117 Bagus, Phillipp (2012) The Tragedy of the Euro. Ludwig von Mises Institute; 2nd edition. Bargain, Olivier, Mathias Dolls, Clemens Fuest, Dirk Neumann, Andreas Peichl, Nico Pestel, and Sebastian Siegloch (2013) “Fiscal union in Europe? Redistributive and stabilizing effects of a European tax- benefit system and fiscal equalization mechanism.” Economic Policy, 28(75): 375–422. Becker, Sascha O., Peter H. Egger, and Maximilian von Ehrlich. (2010) “Going NUTS: The effect of EU Structural Funds on regional performance.” Journal of Public Economics 94(9–10): 578–590. Becker, Sascha O., Peter H. Egger, and Maximilian von Ehrlich (2012) “Too much of a good thing? On the growth effects of the EU’s regional policy.” European Economic Review, 56(4): 648–668. Becker, Sascha O., Peter H. Egger, and Maximilian von Ehrlich (2013) “Absorptive capacity and the Growth and investment effects of regional transfers: A regression discontinuity design with heterogeneous treatment effects.” American Economic Journal: Economic Policy, 5(4): 29–77. Becker, Sascha O., Peter H. Egger, and Maximilian von Ehrlich (2018) “Effects of EU regional policy: 1989–2013.” Regional Science and Urban Economics, 69: 143–152. Blanchard, O., and D. Leigh. (2013). "Growth Forecast Errors and Fiscal Multipliers." American Economic Review Papers and Proceedings, 103 (3): 117-20. Beugelsdijk, M., and S. Eijffinger (2005). “The effectiveness of structural policy in the European union: An empirical analysis for the EU-15 in 1995–2001.” Journal of Common Market Studies, 43(1): 37–51. Canova F. and E. Pappa (2022). “The macroeconomic effects of EU Structural funds” Working Paper (November 2022) Chodorow-Reich, G. (2019). "Geographic Cross-Sectional Fiscal Spending Multipliers: What Have We Learned?" American Economic Journal: Economic Policy, 11 (2): 1-34. Coelho M. (2019) “Fiscal Stimulus in a Monetary Union: Evidence from Eurozone Regions”, IMF Economic Review 67: 573–617 Crucittia F., N. Lazarou, P. Monfort, and S. Salotti (forthcoming) “Where does the EU cohesion policy produce its benefits? A model analysis of the international spillovers generated by the policy” Economic Systems 28 Durand L. and R Espinoza (2021) “The Fiscal Multiplier of European Structural Investment Funds: Aggregate and Sectoral Effects with an Application to Slovenia” IMF Working Paper WP/21/118 Ederveen, Sjef, Henri L.F. Groot, and Richard Nahuis (2006) “Fertile soil for structural funds? A panel data analysis of the conditional effectiveness of European cohesion policy.” Kyklos, 59(1): 17–42. European Court of Auditors (2018), “Commission’s and Member States’ actions in the last years of the 2007-2013 programmes tackled low absorption but had insufficient focus on results” https://www.eca.europa.eu/Lists/ECADocuments/SR18_17/SR_ABSORPTION_EN.pdf European Parliament (2023) “Next Generation EU (NGEU) delivery – How are the Member States doing? Bulgaria's National Recovery and Resilience Plan Latest state of play” European Parliament Research Service” https://www.europarl.europa.eu/RegData/etudes/BRIE/2022/733662/EPRS_BRI(2022)733662_EN.pdf Frankel J, Végh C, Vuletin G. (2013) “On Graduation from Fiscal Procyclicality.” Journal of Development Economics 100 (1) :32-47 Kraay, A. (2014) "Government Spending Multipliers in Developing Countries: Evidence from Lending by Official Creditors." American Economic Journal: Macroeconomics 6(4): 170–208. Lo Piano, Samuele, Rosaria Chifari and Andrea Saltelli (2017) “Regionalisation of ESIF payments 1989-2015” European Commission, Contract CCI 2016CE16BAT081 Mahieu G., P. Pfeiffer, J. Varga, and J.in 't Veld, (2021) “A stylised quantitative assessment of Next Generation EU investment”, VOXEU 4 August 2021 link Mohl, P. and T. Hagen (2010) “Do EU structural funds promote regional growth? New evidence from various panel data approaches.” Regional Science and Urban Economics, 40(5): 353–365. Mohl, P. and T. Hagen (2011) “Do EU structural funds promote regional employment? Evidence from dynamic panel data models.” Working Paper Series, European Central Bank 1403, ECB Nakamura, E., and J. Steinsson (2014) "Fiscal Stimulus in a Monetary Union: Evidence from US Regions." American Economic Review, 104 (3): 753-92. Pennings, S. (2021) "Cross-Region Transfer Multipliers in a Monetary Union: Evidence from Social Security and Stimulus Payments." American Economic Review, 111 (5): 1689-1719. Pennings, S. (2022) “Locally financed and outside financed regional fiscal multipliers”, Economics Letters, Volume 213 (April 2022) Ramey, V. (2019) "Ten Years after the Financial Crisis: What Have We Learned from the Renaissance in Fiscal Research?" Journal of Economic Perspectives, 33 (2): 89-114. Sala-i-Martin, X. (1996) “Regional cohesion: Evidence and theories of regional growth and convergence.” European Economic Review, 40(6): 1325–1352. 29 Schmitt-Grohe S. and M Uribe (2016) “Downward Nominal Wage Rigidity, Currency Pegs, and Involuntary Unemployment,” Journal of Political Economy 124: 1466-1514. Staiger, Douglas, and James Stock, “Instrumental Variables Regression with Weak Instruments,” Econometrica, 65 (1997), 557-586. 30