Regional growth and absorption speed of EU funds: when time isn’t money

ABSTRACT This paper discusses the effects of European Union (EU) funds on gross domestic product growth by analysing the causal impact of regional absorption speed. The analysis is conducted using a regression discontinuity design approach with heterogeneous treatment on NUTS-2 regions during the period 2000–16. We show that faster absorption of EU funds in the Objective 1 regions, especially the Mediterranean ones, is associated with worse economic outcomes from the Objective 1 treatment. However, this pattern is not observed in the non-treated regions. Regarding policy implications, this study suggests that incentives to increase absorption speeds should be removed in Objective 1 regions.


INTRODUCTION
European Cohesion Policy is designed to foster economic homogeneity across the countries and regions of the European Union (EU) in pursuit of successful market integration. Jacques Delors, President of the European Commission between 1985 and 1995, said in 1989 that the Cohesion Policy is meant 'to give every region an opportunity to benefit from the enormous advantages the single market will bring'. 1 For the last programming period 2014-20, Cohesion Policy constitutes the secondlargest budget line after the EU's Common Agricultural Policy (CAP), accounting for almost a third of the European budget. A special scheme has been designed for NUTS-2 regions characterized by a gross domestic product (GDP) per capita lower than 75% of the per capita European GDP average, making them eligible for Objective 1 treatment. Since the programming period 1989-94, this status has allowed some regions to benefit from markedly increased EU transfers so as to hasten the convergence process.
To make efficient use of these European funds, recipient regions must use these transfers in investment projects that promise to generate additional economic growth. A high regional absorption capacity is therefore necessary to reach these policy goals. The European Commission defines absorption capacity as 'the ability to use the financial resources made available … on the agreed actions and according to the agreed timetable'. 2 Therefore, a high absorption speed for EU funds constitutes a policy target for the European Commission as it is considered to be a signal of the absorption capacity of a recipient region. 3 To accelerate absorption, a portion of the budgetary commitment is automatically decommitted by the Commission if it has not been used or if no payment application has been received by the end of the second year following that of the budgetary commitment (the n + 2 rule). This rule was introduced in 1999 due to a growing concern at the EU level about the poor financial performance of some EU regional development programmes. The programming period 2014-20 was characterized by a softer rule, the decommitment procedure having been extended to three years after the end of the programming period (n + 3 rule). Observing a slowdown in the absorption speed, the Commission has proposed to return to the n + 2 rule for the programming period 2021-27 (Bachtler et al., 2019). Figure 1 indicates the share of EU payments implemented after the end of their corresponding programming period, that is, the late payments, for each NUTS-2 region for the programming periods 1994-99, 2000-06 and 2007-13. It can be seen that the map becomes more reddish across time, indicating that late payments are an increasing phenomenon. During the 2007-13 period, the vast majority of regions had more than 50% of late payments, this share exceeding 75% in most of the English, Belgian and Portuguese regions. It is worth mentioning that only 25% of the observations of this study have a share of late payments lower than 20%, while 30% of observations exceed the 80% threshold. According to Figure 1, it appears that regions with the fastest absorption are mostly located in Sweden, Finland and Greece.
Fast absorption is helpful in the sense that it avoids decommitment of EU payments. Regarding the programming period 2000-06, for instance, substantial amounts were decommitted in the Netherlands (11.1%), Luxembourg (10.8%) and Denmark (6.1%) due to slow absorption (Bachtler & Ferry, 2015). However, one drawback of spending faster might be spending worse: indeed, '[s]ome Member States are critical of n + 2 and argue that it will lead to a recurrence of problems with preparing and managing large, high-value projects, and encourage a less strategic approach to project selection' (Bachtler et al., 2019, p. 39).
The novelty of this paper is that it assesses whether fast absorption of EU funds constitutes a desirable policy outcome of the Cohesion Policy. In other words, should we trust absorption speed as a means for evaluating the absorption capacity of a recipient region? Is it a suitable proxy for absorption capacity?
To study this question, this paper contributes to existing research by exploiting a new source regarding the conditional impact of the Cohesion Policy on regional economic growth: the absorption speed of EU funds in recipient NUTS-2 regions. Our analysis aims to determine whether the delays in EU payments may generate a heterogeneity in the Objective 1 treatment's effect on the  2007-13 (c). Note: [0.25; 0.5] denotes a NUTS-2 region where between 25% and 50% of total EU payments of a given MFF (1994-99, 2000-06 or 2007-13) have been executed after the end of this MFF. The same logic applies for [0; 0.25], [0.5; 0.75] and [0.75; 1]. Source: Author's own elaboration based on data from Lo Piano et al. (2017). EuroGeographics for the administrative boundaries.

512
Benoit Dicharry economic growth of recipient regions. In other words, we intend to determine whether the magnitude of the impact of the EU funds in lagging regions is fully determined by their pace of spending. The estimation methodology of this paper is based on Becker et al. (2013), which exploits the discrete jump in the probability of EU transfer receipt at the 75% threshold to conduct a fuzzy regression discontinuity design (RDD) with heterogeneous local average treatment effect (HLATE). While Becker et al. estimates the impact of the Objective 1 treatment based on regional governance quality and human capital level, we focus on the regional absorption rate of the EU funds. To increase the reliability of our estimates, we consider real EU payments as drawn from the database of Lo Piano et al. (2017), which follows the dates on which expenditures took place on the ground. This contrasts, therefore, with expenditure commitments, which are usually employed in the literature that studies the economic effectiveness of the Cohesion Policy (e.g., Becker et al., 2010Becker et al., , 2013Becker et al., , 2018Pellegrini et al., 2013;Rodriguez-Pose & Garcilazo, 2015).
Our paper shows that faster absorption of EU funds reduces the effectiveness of the Cohesion Policy in Objective 1 regions, that is, the ability of the EU funds to stimulate economic growth. In other words, the faster the EU funds are absorbed in Objective 1 regions, the lower is the impact on economic growth. This result reveals the tension between spending well and spending fast in the European lagging regions, which are generally characterized by a lower absorption capacity.
This illustrates the fact that fast absorption might be the outcome of strategic behaviour by recipient regions or governments to send a signal of good management to the European authorities (Aivazidou et al., 2020;Huliaras & Petropoulos, 2016). A quantile regression analysis suggests that this result is especially valid for the regions with the lowest economic growth performances, the latter being mostly located in Mediterranean Europe. A second result is that slow absorption has a negative impact on economic growth in non-treated regions. As they are wealthier, they receive significantly less EU transfers and are generally characterized by a higher absorption capacity (Becker et al., 2012), which gives little room to conduct the strategic behaviours aimed at increasing absorption rates. Therefore, in non-treated regions, slow absorption would tend to be the outcome of lower management quality (Dudek, 2005;Incaltarau et al., 2020;Milio, 2007;Surubaru, 2017;Tosun, 2014). These results are robust to different specifications, sample compositions and outcome variables.
The policy implications of our interpretation are easily implementable by policymakers: we propose to remove the one-size-fits-all logic of the decommitment rule. We propose instead the introduction of a place-based approach, one that takes into consideration the lower absorption capacity of Objective 1 regions. Therefore, a differentiated decommitment rule between Objective 1 and wealthier regions, or even a suspension of the rule for the Objective 1 regions, could help to mitigate the use of strategies detrimental to the effectiveness of the Cohesion Policy.
The remainder of this paper is organized as follows. Section 2 provides a review of related literature. Section 3 deals with the methodology and data used to conduct our analysis. Section 4 provides the estimation results and the robustness tests, along with the discussion. Section 5 concludes.

RELATED LITERATURE
Within the large literature dealing with the European Cohesion Policy, the local quality of government has unanimously been treated as a promoting factor for the conditional impact of EU funds on regional economic growth resulting from a higher absorption capacity (e.g., Becker et al., 2013;Dall'erba & Fang, 2017;Ederveen et al., 2006;Mendez et al., 2013;Rodriguez-Pose & Garcilazo, 2015). For instance, Dall'erba and Fang (2017) offers a meta-regression analysis of the impact of EU funds on the regional growth of recipient regions based on 323 estimates in 17 econometric studies. Human capital and quality of institutions are identified as 'characteristics of the recipient regions that condition the effectiveness of the funds' (p. 10).
Some recent studies highlight that fast absorption is a signal for high absorption capacity resulting from a sound institutional environment (Dudek, 2005;Incaltarau et al., 2020;Milio, 2007;Surubaru, 2017;Tosun, 2014). Tosun (2014) explores the determinants of absorption pace with regard to the European Regional Development Fund's (ERDF) 2000-06 programming period and finds that member states' government effectiveness is positively associated with the speed of absorption of the ERDF. Meanwhile, Surubaru (2017) associates faster absorption with better institutions and stronger administrative capacity. This comparative study mentions that in the case of Bulgaria, the result of the favourable political and institutional environment has been a higher progression of the absorption speed than in Romania for the period 2007-13. A similar study conducted by Incaltarau et al. (2020) concludes that government effectiveness has a positive effect on national absorption rate.
However, the view that fast absorption results from high absorption capacity is not unanimously shared (Aivazidou et al., 2020; Centre for Industrial Studies (CSIL), 2010; European Court of Auditors (ECA), 2004; Huliaras & Petropoulos, 2016;Polverari et al., 2007). Notably, Huliaras and Petropoulos (2016) provide a case study of Greece for the programming period 2007-13, highlighting the weaknesses in administrative capacity and the bad institutional environment of the authorities in charge of the implementation of the Cohesion Policy. As a result, the observed fast absorption has been more the result of easy-to-spend solutions than of a good use of EU funds resulting from a high absorption capacity. Indeed, '[i]n 2010, one of the top priorities of the newly elected government was not to lose "a single euro" of the National Strategic Reference Framework 2007-2013 money' (p. 8).
Similarly, Aivazidou et al. (2020) conclude that the low performance of EU funds in the Italian regions for the programming period 2007-13 was due to the adoption of strategies aiming at increasing absorption percentages instead of fostering administrative capacity.
Specifically regarding the decommitment rule, while it has been effective in accelerating absorption (Bachtler & Ferry, 2015), it has also led the authorities in charge of the implementation of the Cohesion Policy to focus on the pace of spending rather than the quality of interventions (CSIL, 2010;Polverari et al., 2007). Moreover, the rule has had a detrimental impact on the ability of the Cohesion Policy to adapt to specific regional and national contexts (European Commission, 2011). It could also be mentioned that the decommitment rule has put strong pressure on local administrative resources, as 50% of payments are submitted between September and December (ECA, 2004). To sum up, the faster absorption induced by the n + 2 rule might have been detrimental to the conduct of the Cohesion Policy and its ability to foster regional economic growth. Therefore, our study seeks to provide insights into whether fast absorption has a nurturing or detrimental impact on the ability of Objective 1 treatment to stimulate growth at the regional level.
Regarding the estimation approach, Becker et al. (2010) is the first study to adopt an RDD design to exploit the existence of a threshold in the attribution of the treatment status (set as 75% of the EU per capita GDP in purchasing power parity). An extended use of the RDD is then proposed in Becker et al. (2013), where heterogeneous local effects are estimated. The analysis, based on the HLATE, showed that the degree of absorptive capacity is important for explaining differences in outcomes. This approach has since been followed by numerous studies aiming to reveal the different sources of heterogeneity on the impact of EU funds on regional growth:  provides evidence that the initial distribution of land matters, since rural areas close to city centres are those where the impact of EU funds is the strongest. For example, Percoco (2017) finds that the size of the regional service sector is detrimental to the impact of EU funds on regional growth. Becker et al. (2018) explores the heterogeneity across recipient regions in terms of their exposure to the last European financial and economic crisis and reveals that in spite of a positive impact, the effects of the European transfers are weaker in countries that have been hit harder by the crisis.
The next section presents the methodology and data employed in our analysis.

RDD estimation
In this study, we focus on the potential heterogeneity of treatment effect according to the share of late payments a i,r , which is defined as: where eu i,r−1 late denotes the payments of the last programming period r − 1 made for a region i after the end of the corresponding programming period. We consider the programming periods 1994-99, 2000-06 and 2007-13. 4 eu i,r−1 denotes the total allocation provided to region i for the associated programming period r − 1. To sum up, late payments can be defined as the payments of programming period r −1 made in programming period r. Finally, a i,r is bounded to [0;1].
We recall that the main contribution of this study is to analyse whether a i,r can be considered a suitable proxy for regional absorption capacity by evaluating its impact on the effectiveness of the Objective 1 treatment. To answer this question, we frame the hypothesis that a i,r−1 is associated with the programming period r. More precisely, the share of late payments of period 1994-99 is associated with 2000-06, that of 2000-06 is associated with 2007-13, and that of 2007-13 is associated with 2014-20. The motivations behind this assumption are threefold: (1) operational programmes, or the detailed plans in which the member states set out how money from the EU funds will be spent during the programming period r, are written in the final years of the programming period r − 1; (2) the way the EU funds are managed in the first years of r might be crucially determined by the absorption capacity inherited from the period r − 1; and (3) regarding the empirical strategy, it has the advantage of avoiding the potential endogeneity of the interaction variable.
To conduct the analysis, we adopt a HLATE estimation where the absorption rate may amplify or reduce the treatment effect. We rely on an RDD in line with recent studies (e.g., Becker et al., 2013Becker et al., , 2018Cerqua & Pellegrini, 2018;. RDD is based on the principle that there is an exogenous eligibility rule built on an observable variable, called the forcing variable. In this study, this is the relative GDP per capita of one NUTS-2 region expressed in purchasing power standard (PPS) with respect to the European average. For the programming period 2000-06, the eligibility status is determined on the basis of years 1994-96 (1997-99 for countries that joined the EU in 2004), years 2000-02 for the programming period 2007-13, and years 2007-09 for the programming period 2014-20. 5 The treatment is a binary Objective 1 indicator for a NUTS-2 region i. We recall that Objective 1 status leads to increased transfers aiming at reducing the gap in per capita GDP between non-treated and treated regions.
One key feature here is that the treatment rule is not perfectly respected. Indeed, in reality, and for several reasons, there are some exceptions to the treatment rule. We could mention, for instance, the sparsely populated regions in Austria, Finland and Sweden that are eligible for funds despite being above the relevant threshold of 75%.
Another group comprises the outermost regions of France, Portugal and Spain, where the Canary Islands are above the 75% threshold. Finally, the last exception is the phasing-out status, that is, NUTS-2 regions that were granted Objective 1 transfers in 1994-99, although having a GDP higher than the 75% threshold for the period 2000-06. In essence, due to the imperfect compliance with the eligibility rule, we must implement a fuzzy RDD design. As indicated by Imbens and Lemieux (2008), applying ordinary least squares (OLS) would lead to biased estimates because of the fuzziness of the treatment. Therefore, a two-stage least squares (2SLS) method, where the actual treatment is instrumented by the eligibility rule, should be implemented in order to provide reliable estimates. We rely on Becker et al. (2013) for the entire econometric strategy.
The second stage of the 2SLS under fuzzy RDD, with a HLATE identification where the interaction variable is the share of late EU payments, is given by: where y i,r represents the GDP per capita growth of region i averaged for the programming period r, a 2 is a constant and m i,r is the error term. x i,r is the deviation from the 75% threshold while a i,r and K k k i,r , a set of K control variables, are expressed as the deviation from their sample mean. t denotes the coefficient directly associated with the fitted value of the treatment tˆi ,r . a i,r is associated with coefficients z 1,n and h 1,q when the treatment is switched on (tˆi ,r = 1). z 0,n and m 0,q are the same coefficients when the treatment is switched off.
Regarding the first-stage regression, we use an eligibility rule represented through a binary variable taking the value 1 if the NUTS-2 region has a GDP per capita below 75% of the EU average, and 0 otherwise. A linear probability model is implemented, the first-stage regression being given by: where t i,r represents the instrumented variable that is the treatment status of region i for the programming period r, a 1 is a constant and e i,r is the error term of the firststage estimation. The eligibility rule for treatment in programming period r, r i,r , is determined according to the 75% threshold for region i that is eligible for treatment: r i,r = 1 when the forcing variable is lower than or equal to 75%, r i,r = 0 in the opposite case. x i,r,T is the forcing variable normalized around the 75% threshold. a i,r,T , the interaction variable, normalized around its mean value, is associated with coefficients z 1,n and h 1,q when there is eligibility for the treatment (r i,r = 1). z 0,n and m 0,q are the same coefficients when r i,r = 0, or when a region is not eligible for Objective 1 treatment.
The following subsection describes the data used in the analysis and their descriptive statistics.

Data and descriptive statistics
We collected most of the data from Eurostat Regional Statistics. They have been completed with data from Cambridge Econometrics. The information about Objective 1 status and eligibility and about expenditures comes from the European Commission. We provide all data sources in Table A1 in the supplemental data online. Our sample covers a panel data set of the EU's NUTS-2 regions for the period 2000-16. We do not include Bulgaria, Romania or Croatia for reasons of data availability. The resulting number of NUTS-2 regions is 244. We used the NUTS2-2013 classification employed by the European Commission (2019) which provides the input data used to build the following index. Regarding the time dimension of the dataset, the data employed in the analysis are averaged for the programming periods 2000-06 and 2007-13. Regarding the programming period 2014-20, the latest available year is 2016, so the data correspond to the averages of the period 2014-16. 6 Such a transformation is implemented because the treatment variable is determined for each programming period r.
It should be mentioned that only actual received payments have been considered in this study, and not commitments as in most studies in the literature (e.g., Becker et al., 2010Becker et al., , 2012Becker et al., , p. 2013Pellegrini et al., 2013;Tosun, 2014;Rodriguez-Pose & Garcilazo, 2015;Surubaru, 2017;Cerqua & Pellegrini, 2018;Incaltarau et al., 2020). As Lo Piano et al. (2017) observe, this dataset has the advantage of following the dates in which expenditures took place on the ground. This is not the case for commitments, which: may negatively affect the analytic work subsequently done by the experts to carry out policy assessments or to run counterfactual impact evaluations estimating the effects of the varying intensities of the EU funds on regional growth variables. The misalignment between COM reimbursement cycle and date of the interventions on the ground (beneficiaries' expenditures) may represent a disturbance acting either as a noise or as a bias.
(p. 6) Hence, we consider this modelled annual expenditure as our actual EU funds expenditure variable so as to increase the reliability of our estimates. As control variables we include population density, as the European authorities consider that a low population density is a structural handicap to achieving economic growth. We also use both the share of the manufacturing sector and the share of financial and business services in regional gross added value (GVA). Moreover, we consider the share of the active population and the unemployment rate in order to have a proxy for the size of the labour force, and the share of the active population having Regional growth and absorption speed of EU funds: when time isn't money 515 achieved tertiary education as a proxy for human capital. Finally, to control for the effects of the asymmetric shocks from the Great Recession and the following Euro Crisis, we consider the difference between the yield spreads of the national 10-year government bond (GBYS) of the region and the yield spreads of the national German government bond. The rationale behind this choice of variable is that Germany is legitimately to be considered the benchmark economy thanks to its very favourable market conditions in issuing public debt, especially over the last decade (Debrun et al., 2019). Table 1 displays summary statistics for key variables of interest averaged and pooled over the programming periods 2000-06, 2007-13 and 2014-16. The outcome variable, GDP per capita growth, is calculated as the difference between the logged-GDP per capita and its lagged value. The forcing variable, relative GDP per capita, is then displayed as a deviation from the 75% threshold of the EU average by the time of decision of the European Commission. The interaction variable is expressed in terms of deviation regarding the pooled sample mean value, as are the above-mentioned control variables. Regarding the interaction variable, it appears that the mean is relatively similar between regions lying below and above the 75% threshold, although one subsample is more than twice as large. Identification of a causal effect of Objective 1 treatment on growth by means of RDD requires that there is a discontinuity at the threshold, which is obvious in Figure 2. The jump of the outcome variable at the threshold amounts to about 0.4 percentage points. 7 This result strengthens the usefulness of the RDD in apprehending the impact of the EU funds on regional GDP growth. Second, Figure 3 displays the density distribution of GDP per capita expressed using pooled averaged observations of programming periods 2000-06, 2007-13 and 2014-16. The RDD set-up would not be valid if a spike before the 75% threshold were observed, as it would invalidate the exogeneity of the Objective 1 treatment. This is not suggested by Figure 3, though, since the density peak can be observed around a level of 90%. Figure A1 in the supplemental data online then plots the interaction variable against the forcing variable. There is no indication of a jump at the 75% threshold, which ensures the validity of the RDD estimates. A similar pattern is observed for the control variables used in the analysis in Figure A3 online.
Finally, Figure A2 online illustrates graphically how the probability of Objective 1 treatment relates to region-specific per capita GDP relative to the European average prior to each programming period (forcing variable). While a probability jump is visible at the 75% threshold, the fuzziness of the RDD design is revealed, as some regions having a relative GDP per capita higher than 75% of the European average at the time of the European Commission's decision are treated, and vice versa.

Estimation results
In this subsection, we present the main results from our analysis regarding regional GDP per capita growth and the share of late EU payments. In general, our results support the view that the later the payments are made, that is, the slower the absorption of the EU funds is, the higher is the effectiveness of the Cohesion Policy in Objective 1 regions. Table 2 reports estimates of the local average treatment effect (LATE) of Objective 1 status on regional economic growth. These simple RDD regressions stand for the average effect of the Objective 1 treatment on regional growth.  The LATE is estimated in two different samples: averaged observations of regions having a share of late payments below (column 1) and above (column 2) the sample average. The sample size is restricted in order to increase the reliability of the RDD estimates: we propose a subsample including regions with a relative GDP per capita 25% higher and lower than the European average at the time of decision by the European Commission, that is, between 50% and 100%. Indeed, the RDD approach is based on observations that are close to this threshold since they are likely to be very similar to each other with respect to observed and unobserved characteristics, except for the outcome variable. Therefore, the mean difference in the outcomes can be attributed to the treatment effect. This average treatment effect (ATE) sacrifices external validity by focusing only on observations close to the cut-off point, that is, the 75% level of the average European regional GDP per capita. Finally, we include estimates of panel fixed effects to capture all the unobserved factors related to each NUTS-2 region.
As can be observed, the Objective 1 treatment has a systematic positive and significant effect for regions characterized by a share of late EU payments higher than the sample average. However, the same cannot be said for the fast-spending regions, as the LATE is positive and significant only for the RDD estimate that includes the entire sample, which could be considered a less reliable estimate because of an interregional comparability problem. Otherwise, the Objective 1 treatment does not have any significant effect on regional per capita GDP growth. Consequently, the estimates displayed in Table  2 might reveal a heterogeneous impact of the Objective 1 treatment according to the absorption pace of EU transfers. This legitimates the study of the HLATE of the Objective 1 treatment based on the share of late EU payments.
The estimation results for the heterogeneous effects (HLATE) are displayed in Table 3. To increase the reliability of the RDD estimates as much as possible, we restrict our sample to 12.5% around the eligibility threshold, that is, NUTS-2 regions having a GDP per capita from 62.5% to 87.5% of the European average (columns 1-2). One drawback of this procedure is the sharp reduction of sample size, since the number of observations falls to 219. Columns (3) and (4) include regions with a relative GDP per capita between 50% and 100% of the European average, which allows us to nearly double the sample size to 394 observations. Columns (5) and (6) include the entire sample, where only regional fixed effects are included. It is worth mentioning that some non-linearity is introduced with the squared term of the share of late payments in columns (2), (4) and (6). The analysis shows that weak instruments and endogeneity tests are generally verified. For sake of brevity, we report only second-stage estimates.
The first striking result is that a faster absorption of EU funds reduces the effectiveness of the Cohesion Policy in Objective 1 regions, or the ability of the EU funds to stimulate economic growth. Indeed, in all specifications, the coefficient on the interaction term between the share of late payments and the treatment exhibits a positive sign. The introduction of a quadratic interaction term even reinforces this result. In all specifications, we obtain ∂y i,r ∂a i,r . 0 for Objective 1 regions, which indicates that the net effect of an increase in the share of late payments is beneficial to regional growth. This result validates our supposition that fast absorption might be the outcome of strategic behaviour by recipient regions or governments intended to send a signal of good management to the European authorities (Aivazidou et al., 2020;Huliaras & Petropoulos, 2016). This finding supports our suggestion that there might be a

518
Benoit Dicharry ground to the conflict between spending fast and spending well in lagging regions, as they are generally characterized by a lower absorption capacity (Becker et al., 2013). In other words, less-wealthy local management authorities may encounter more difficulties in spending European subsidies efficiently in a given time period compared to an economically developed region. A second result is that slow absorption has a negative impact on economic growth in regions having a relative GDP per capita higher than 75% of the European average. Indeed, as they do not benefit from the Objective 1 treatment, we find that ∂y i,r /∂a i,r , 0. As these regions are wealthier than the Objective 1 regions, they receive significantly less EU transfers and are generally characterized by a higher absorption capacity (Becker et al., 2012), which leaves little room for them to conduct the strategic behaviours that are aimed at increasing absorption rates. Therefore, in non-treated regions, slow absorption Table 3. Objective 1, late payments and regional gross domestic product (GDP) per capita growthheterogeneous local average treatment effect (HLATE) (instrumental variables (IV) second-stage estimates) and panel fixed effects.
(1) Note: Reported are the results from the two-stage least square estimation of the HLATE with a sample restricted to 12.5% (columns 1 and 2) and 25% (columns 3 and 4) around the 75% threshold of the forcing variable (GDP per capita). The forcing variable is the relative GDP per capita of 1996-98 (1997-99) for years 2000-06, of 2000-02 for years 2007-13, and of 2007-09 for years 2014-16. The two-stage least square (panel IV) estimations using regional fixed effects are reported in columns (5) and (6) using the full sample. The dependent variable presents regional GDP per capita growth. Robust standard errors are reported in parentheses. *p < 0.1, **p < 0.05, ***p < 0.01. Source: Author's own calculations based on data from European Commission, Cambridge Econometrics and Eurostat.
Regional growth and absorption speed of EU funds: when time isn't money 519 would instead tend to be an outcome of lower quality of management (Dudek, 2005;Incaltarau et al., 2020;Milio, 2007;Surubaru, 2017;Tosun, 2014). A third result is that the treatment does not have any robust direct impact on regional economic growth, making its impact purely conditional. Indeed, in all regressions, the magnitude of the impact of the EU funds in lagging regions is fully determined by their pace of spending. Therefore, the Objective 1 treatment does not promote economic growth per se. This finding is in line with a large majority of the literature underlining that the effectiveness of the Cohesion Policy mostly relies on regional governance quality and human capital level (e.g., Becker et al., 2013Becker et al., , 2018Cappelen et al., 2003;Rodriguez-Pose & Garcilazo, 2015).
Regarding control variables, half of them are characterized by insignificant effects. The remaining ones are associated with the expected significant effects: (1) it could be noticed that the proxy for human capital, that is, tertiary education achievement, is associated with a positive and significant impact on per capita GDP growth in most of the specifications; (2) a similar outcome appears for the share of the manufacturing sector in regional gross added value, indicating that the industrial sector is a powerful growth driver (Baumol, 2001); and (3) it is worth mentioning the robust negative significant impact of the GBYS on per capita GDP growth, since this feature reveals that the inclusion of this control variable is relevant to capture the shocks inherited from the Great Recession and the ensuing Euro Crisis.
To add strength to these results, we conduct additional regressions using a different outcome variable, namely the growth of per capita regional investment, since the initial aim of the Cohesion Policy is to stimulate public and private investment and so to foster regional GDP growth. The structure of Table A2 in the supplemental data online is the same as Table 3. The estimation results in Table A2 online are qualitatively similar.
Following the methodology of Becker et al. (2013), we implement non-parametric regressions based on a local linear estimator with bootstrapped estimations (500 times). The optimal bandwidth is selected using the improved Akaike information criterion (AIC) of Hurvich et al. (1998). The non-parametric estimates are derived from a specification with both linear GDP per capita and share of late payments. The variability of the HLATE function according to the share of late payments is displayed in Figure 4. It can be observed that an increase in the share of late payments has a positive effect on the effect of the treatment on regional per capita GDP growth, since the HLATE is an increasing function. It should be noted that the non-parametric HLATE function is steeper. Moreover, while the HLATE estimated with the RDD estimator is always positive, the non-parametric estimated HLATE is negative for all late payments below the sample mean value. Figure A4 in the supplemental data online displays similar estimates where the dependent variable is per capita investment growth. The estimation results are qualitatively similar.
Given the nature of the projects financed by the Objective 1 financial transfers (e.g., transport infrastructure or research projects), our previous estimation results may be affected by spatial autocorrelation. This is confirmed by Moran's I test, which is always below 0.2 yet systematically significant, indicating modest spatial autocorrelation. 8 To tackle this issue, spatial autoregressive fixed-effects estimates are conducted. A weighting contiguity matrix based on the 244 NUTS-2 regions of our sample is created where firstand second-order neighbours have the same weight. The estimation results are reported in Table 4. While remaining robust, it can be noted that the significance Figure 4. Heterogeneous local average treatment effect (HLATE) and regional per capita gross domestic product (GDP) growth for different levels of the share of late EU payments (parametric (a) and non-parametric (b) estimates). Note: The solid line illustrates the point estimates, the dashed lines represent the 95% confidence intervals. The confidence intervals are derived from bootstrapped standard errors with 100 replications. Source: Author's own elaboration. 520 Benoit Dicharry of late payments is reduced to the 10% level where per capita GDP growth is the dependent variable. It can be observed that: (1) an increase in the share of late payments in Objective 1 regions is not detrimental to economic growth; (2) the opposite holds in non-treated regions; and (3) the effect of the Objective 1 treatment is mostly conditional. The next subsection deals with additional regressions, so as to increase the precision of our estimates. First, quantile regressions are implemented in order to investigate whether the treatment effects are homogeneous across per capita GDP growth levels. Moreover, following the conclusions of Becker et al. (2012), we investigate whether the intensity of the European transfers is relevant in determining their capacity to stimulate economic growth in recipient regions.

Additional results
The quantile regressions estimates are displayed in Table  A3 in the supplemental data online. It can be observed that the absorption speed appears to be relevant only in regions exhibiting the lowest economic growth patterns, which are mostly located in Southern Europe. Indeed, 47% of the regions in the lowest 25% quantile, in terms of economic growth, belong to Mediterranean Europe and 5% to the CEE countries. On the contrary, if we consider the upper 25% quantile, where the absorption speed appears to be irrelevant, the Central and Eastern Europe Table 4. Objective 1, late payments and regional gross domestic product (GDP) and investment per capita growthspatial autoregressive (SAR) fixed effects (instrumental variables (IV) second-stage estimates).
(1) Note: Reported are the results from the spatial autoregressive fixed-effects model where the dependent variable is GDP per capita growth (columns 1 and 2) and investment per capita growth (columns 3 and 4). r dependent variable denotes the spatial lag coefficient for the dependent variable, and the same logic applies for r residuals. Their significance legitimates the use of the SAR model. Robust standard errors are reported in parentheses. *p < 0.1, **p < 0.05, ***p < 0.01. Source: Author's own calculations based on data from European Commission, Cambridge Econometrics and Eurostat.
Regional growth and absorption speed of EU funds: when time isn't money 521 (CEE) countries count for 31% of the sample, while this share falls to 30% for the Mediterranean ones. We then conduct a heterogeneity analysis based on the conclusions of Becker et al. (2012). This study states that high EU transfer levels may not be appropriate for regions characterized by a low absorptive capacity. Indeed, 'regions with a transfer intensity of more than 1.3% of GDP could give up EU transfers without experiencing a significant drop in their average annual per-capita income growth rate' (p. 664). To ensure the best comparability of our estimates, we divide our total sample into two equal subsamples: (1) regions with an EU funds intensity below the median (0.15% of per capita GDP); and (2) regions with an EU funds intensity above the median. The rationale behind this approach is to test whether the fast spenders endowed with substantial European transfers, generally associated with a low absorption capacity, make good use of these resources. The estimation results in Table A4 in the supplemental data online indicate that the outcomes of the low-intensity regions are similar to those associated with non-treated regions, and vice versa with high-intensity and Objective 1 regions. Indeed, in regions with lower transfers, the share of late payments is detrimental to economic growth and can be interpreted as a signal of a lower absorption capacity. On the contrary, in regions receiving a lot of EU transfers, the share of late payments is not a suitable signal for regional absorption capacity. The next subsection discusses the estimation results.

General discussion
First, our results indicate that fast absorption in the Objective 1 regions is not a desirable policy outcome, since faster absorption is significantly associated with a lower effectiveness of the Cohesion Policy in terms of stimulation of economic growth. These results corroborate the findings of Huliaras and Petropoulos (2016), who focus on Greece, especially during the 2007-13 period, and reveal that every time a programming period end was approaching, the political authorities targeted easy-to-spend solutions, such as unconditional direct subsidies to small and medium-sized enterprises or the construction of parking facilities to keep authorities satisfied and exhibit that all the European money had been spent on time. Moreover, the conclusions of Huliaras and Petropoulos also corroborate our estimation results, as we have shown that fast absorption is the most detrimental in Objective 1 regions with poor growth performances (see Table A3 in the supplemental data online), where the Greek regions count for 18% of our observations. Regarding the n + 2 rule in particular, our results are in line with the literature pointing out that this rule has resulted in an increased focus on the pace of spending rather than the quality of the investment projects (CSIL, 2010), especially in regions with limited administrative resources (ECA, 2004), such as the Objective 1 regions. While a strand of the literature concludes that there is a positive association between regional administrative capacity and the speed of the implementation of the European Cohesion Policy in Spain (Dudek, 2005), Italy (Milio, 2007), Romania and Bulgaria (Tosun, 2014), we posit that absorption pace is failing as a signal for absorption capacity. Indeed, it does not capture local strategies implemented to hasten absorption at the cost of lower economic effectiveness. We may consider, for instance, the use of retrospective projects, which consist in funding projects which have incurred expenditure, or are completed before the EU co-financing has been formally applied, that is, they are financed retrospectively. As these projects are often selected, initiated or carried out without having been expressly linked to a programme's objectives or to specific legal requirements linked to EU assistance, they exhibit a significant risk of low economic effectiveness (ECA, 2018). Aivazidou et al. (2020) mentions the reduction of regional share of contribution as a strategy to increase absorption rates. This study proposes, then, an alternative measure of absorption, that is, the net absorption rate of total funding based on the initial total commitments (net initial total absorption rate -ITAR), to alleviate the bias of this strategy on absorption rates.
Our results corroborate the idea that there is a tension between spending fast and spending well. The origins of this trade-off have been the subject of some discussion in the literature dealing with the political economy of EU funds (e.g., Charron, 2016;Dellmuth, 2011). This literature underlines the existence of two objectives: on one side, the full and fast absorption of European funds; on the other, achieving regional cohesion by aiding lagging regions. In the context of the implementation of the Cohesion Policy, the European Commission and the member states can be considered as principals, and recipient regions as agents. The policy goal of the European Commission is to maximize the absorption rates of recipient regions to send a signal that the EU funds are fully used, so as to provide incentives to the member states to increase their financial contribution for the next programming period, and this tends therefore to favour regions with a past track record of high absorption rates when it comes to the allocation decision (Dellmuth, 2011). Charron (2016) shows that even member states' central governments do not have an overriding interest in going against the full absorption policy goal of the European Commission. Indeed, they may intend to send a good signal of the use of the EU funds to the European Commission. As a result, the member states push to hasten the absorption of the EU funds in recipient regions, even in the lagging ones. Resorting to retrospective projects or reducing the regional share of contributions are illustrations of strategic behaviours aimed at speeding up absorption.

CONCLUSIONS
This study investigates the effects of EU funds on regional growth in Objective 1 NUTS-2 regions using a panel dataset of 244 regions for the period 2000-16 by employing an RDD with heterogeneous treatment based on the methodology of Becker et al. (2013). We focus on the speed of absorption of EU funds, which for present purposes we have interpreted as the share of real payments allocated for a given programming period implemented after the end of that programming period.
The main result of this study is that faster absorption of EU funds reduces the effectiveness of the Cohesion Policy in Objective 1 regions, or in other words the ability of EU funds to stimulate economic growth. This result corroborates our suggestion that fast absorption might be the outcome of strategic behaviour on the part of recipient regions or governments seeking to increase absorption rates so as to send a signal of good management to the European authorities (Aivazidou et al., 2020;Huliaras & Petropoulos, 2016). This finding confirms that there is a conflict between spending fast and spending well in lagging regions, namely the ones that are generally characterized by a lower absorption capacity (Becker et al., 2013). A more detailed analysis suggests that this result is especially valid for the regions with the lowest economic growth performances, these being mostly located in Mediterranean Europe. A second result is that slow absorption has a negative impact on economic growth in non-treated regions. As they are wealthier, they receive significantly less EU transfers and are generally characterized by a higher absorption capacity (Becker et al., 2012), which gives little room to conduct the kinds of strategic behaviours that are aimed at hastening absorption. Therefore, in non-treated regions, slow absorption would tend rather to be the outcome of lower management quality (Incaltarau et al., 2020;Milio, 2007;Surubaru, 2017;Tosun, 2014). A third result is that the treatment does not have any robust and direct impact on regional economic growth, making its impact purely conditional. Indeed, the magnitude of the impact of the EU funds in lagging regions is strongly determined by their pace of spending. This finding is in line with the large majority of the literature that underlines the conditional effectiveness of the Cohesion Policy (e.g., Becker et al., 2013Becker et al., , 2018Cappelen et al., 2003;Rodriguez-Pose & Garcilazo, 2015).
Regarding policy implications, we believe that the decommitment rule suffers from a major design issue: it is characterized by a one-size-fits-all logic. The early work by Batterbury (2002) had already mentioned the need for a place-based approach ('The Commission needs to adapt better its Structural Fund policies to suit the characteristics of particular regions having diverse cultures and norms'; Batterbury, 2002, p. 15), and this has indeed been applied in several areas of the Cohesion Policy since the Barca (2009) report. Therefore, a differentiated decommitment rule that distinguishes between Objective 1 and wealthier regions, or even a suspension of the rule for the Objective 1 regions, could help to mitigate the use of strategies detrimental to the effectiveness of the Cohesion Policy. This would be especially relevant for the period 2021-27 as the budget allocated to the Cohesion Policy would be reduced globally while being increasingly focused on the lagging regions, a trend likely to be helpful for future programming periods.