Place-Based Policy in the European
Union
Quantitative Analysis of policy programs by the European Regional Development Fund
Master’s Thesis
Supervised by prof. dr. J.H. (Harry) Garretsen & prof. dr. S. (Steven) Brakman
Abstract
In this work I evaluate the efficiency of place-based policy funded by the Euro-pean Regional Development Fund. By targeting spatial units instead of inhabitants directly, place-based policies allow policymakers to easily address market failures and public good provision. This freedom comes at the cost of potential arbitrage opportu-nities and market distortions. As a result, economists have been sceptical regarding the efficiency of place-based policy. I make use of regional EU data to evaluate one of the largest place-based policy programs in existence. Starting with a difference-in-difference model with a fixed counterfactual, I evaluate the effect of the policy on regional GDP and employment. However, this methodology is shown to depend on assumptions I can not make. To resolve this matter, I construct a double robust es-timator using a flexible counterfactual based on propensity scoring. The estimated regional GDP and employment effects of the policy largely depend on the used method-ology. Where a simple model using strong assumptions finds that treated regions do not catch up in terms of GDP and employment, the double robust estimator sides with the policy makers and produces evidence of positive income effects.
1
Introduction
Currently, European regional policy can not be seen separately from the European Union. The discussion regarding the efficiency of European regional policy, however, predates the EU and goes back to the times of the European Coal and Steel Community and the subsequent European Economic Community (EEC). At the time, regional policies were mainly a national affair. Huge sums where spend on rural infrastructure by European nations, in an attempt to alleviate rural poverty and regional imbalances (Baldwin et al., 2009). The EEC developed some small regional policy programs at the time, but with total spending amounting to less than 5% of the EEC budget, these programs were almost negligible.
The first appearance of the large regional policies set out by European nations in the fifties and sixties of the 20th century sparked an interesting discussion regarding their effi-ciency. The pioneering work of Chenery (1962) evaluated one of Europe’s poorest regions, the Italian Mezzogiorno. The Mezzogiorno region in Southern Italy was the recipient of a large inflow of capital from the Italian national government, in an attempt to help the region catch up with the richer north. Despite the volume of these investments, the accomplishments of the regional development policy were disappointing. Even though the Mezzogiorno grew economically, the region’s economy diverged further from the even faster growing northern economy. Since then, many studies have tried to assess the effi-ciency of regional policy. As the European Union grew larger and its influence in regional policy increased, this discussion became ever more important. Nevertheless, the current consensus regarding the effectiveness of regional policy is unfavorable, to say the least. The question remains whether this consensus is the result of a shortage of sophisticated policy analyses or wasteful regional policies. After all, it seems unlikely for the EU to continuously spend a large share of its budget on an ineffective policy.
questions that lie at the foundation of research regarding so-called place-based policies. Place-based policies are policy programs that allocate resources to a specific geo-graphical area to alleviate poverty and stimulate development. These resources come in the form of subsidies, tax incentives, infrastructure investments and more. The policy programs funded by the ERDF are a prime example of place-based policies. Next to the ERDF, examples of recent place-based policy programs are the Chinese special economic zones and the United States’ Empowerment Zone program. Place-based policies go back even further than the outset of the first European regional policies. Perhaps one of the most ambitious place-based policies, the United States’ Tennessee Valley Authority, first commenced in 1933 while annual funding continued till 1960. Even though these programs have played a role in policy debates and academics in prior decades, economist only fairly recently reopened their interest in the field. Now that evidence of the effects of more recent place-based policy programs like the ERDF is flowing in, the path for sophisticated empirical policy analysis is opening up.
From a political viewpoint, place-based policy programs are often initiated with either efficiency or equity in mind. The EU currently states that their regional policy is designed to “strengthen economic and social cohesion in the European Union by correcting imbal-ances between its regions” (European Commission, 2019a). This statement makes clear that equity between regions is a strong driver for European policymakers. Economists, however, criticize these programs fearing for large distortions in economic behavior and the creation of arbitrage opportunities. We could argue that they are a sign of waste and an unfortunate cost of the political process (Kline, 2010). After all, in a general equilibrium environment, subsidizing some regions will come at the expense of others, while generat-ing many distortions along the way. An overview of evaluations of EU place-based policy program effectiveness performed by EU memberstates shows that even the recipients of policy funding have doubts regarding the effectiveness (European Commission, 2019c). This debate between the policymaker on the one hand and the economist on the other is ongoing and fuels a lively field of both theoretical and empirical contributions regarding the effectiveness of place-based policies.
as-sess the effectiveness of the United States Empowerment Zone program, with the most complete analyses (Busso et al., 2013) finding only modest program efficiency. Contrary, Wang (2013) shows that the Chinese Special Economic Zones beneficially influence munic-ipalities through agglomeration economies and wage effects. Even though data describing the effects of the ERDF is trickling in, the program’s effectiveness has yet to be analyzed academically.
In this work, I aim to advance the field of place-based policy research by utilizing the opportunities of this new case. I will make use of recent EU data to discuss the effec-tiveness of European place-based regional policies by evaluating the impact of the ERDF. The main question I aim to answer is how place-based policy in the EU affects the tar-geted regions in regards to stock, growth and per-capita variables related to employment and income. By creating a quasi-experimental framework, in which control and treatment groups are selected based on ERDF eligibility criteria and the member states’ order of EU accession, the effect of the ERDF can be evaluated in isolation. This framework will be tested using a difference-in-difference estimation procedure. Additionally, I use propensity scoring and decomposition techniques to produce treatment effect estimators with flexi-ble counterfactuals. Whereas program evaluations performed by EU memberstates focus on specific sectoral outcomes in small jurisdictions, I take a broad view to evaluate the program’s economic impact on all targeted regions. This perspective, combined with the quasi-experimental setup and sophisticated treatment effect estimators is unique to this work. The results regarding the effectiveness of these place-based policies are instrumen-tal in helping this ongoing discussion forward. Was Chenery (1962) right all along, or are contemporaneous research techniques sufficient to side with the European policymakers who envision a reduction in social and economic disparities between regions?
The remainder of this work consists of the following chapters: Section 2 discusses regional policy in more detail. I will start by setting out the theory behind regional policies, followed by an overview of empirical results and methodologies. Section 3 will be used to give a detailed description of the ERDF. This description will be compared to the theory discussed in section 2 to form a hypothesis regarding ERDF efficiency. In section 4 I will set out the experimental framework, describe the data and potential issues, and set out a trajectory to resolve the issues and construct an appropriate model. Section 5 discusses the starting point of this trajectory, difference-in-difference estimation. I will discuss this model and present the results as well as the drawbacks of this method. In section 6 I introduce more sophisticated identification techniques to tackle some of the problems discussed in section 4 and 5. New results will be presented. Section 7 discusses the results, as well as the trajectory and the remaining issues. Section 8 concludes.
2
Regional Policy
2.1 Theoretical Foundation
When designing policy programs, policymakers have a choice between different tar-geting strategies. The two most straightforward strategies are to target people directly, or indirectly through the characteristics of the place in which they live. In the context of regional development policy aiming to alleviate regional poverty, a policymaker can implement the first strategy by targeting the share of the population that has an income below a certain threshold. Since by definition, poverty is concentrated in poor regions, this strategy will result in more funds moving towards the poor regions. These people-based policies have the benefit that the funds will always move towards the poorest inhabitants, and eligibility can only be altered by changing earnings. As a result, arbitrage oppor-tunities do not arise. A drawback, however, is that people-based policies are difficult to distribute towards firms, industries and the local provider of public goods and services like infrastructure and schools.
and firm competitiveness, and combat existing market failures. However, this additional freedom comes at the cost of creating arbitrage opportunities and potential new market distortions. The theoretical foundation of these disadvantages and potential rational in favor of place-based policy will be discussed below.
The field of place-based policy received substantial theoretical attention over the last decade. One of the earliest papers to thoroughly evaluate different types of place-based policies using a spatial equilibrium model is Glaeser & Gottlieb (2008). The basic spatial equilibrium model the authors construct suggests that place-based policies will induce indirect effects that counter the redistribution the policy aims to achieve. A regional policy program subsidizing wages, for example, will be met with higher local (housing) prices in equilibrium. The only real effect of the policy will be the transfer of population towards economically less productive areas. The real beneficiaries of the policy will be property owners who can capitalize on increasing house prices.
However, the authors show that a case for place-based policy exists when non-linear positive externalities like agglomeration economies and human capital spillovers are present. Agglomeration economies and the resulting agglomeration effects are among the oldest theories in economics, going back as far as Marshall (1890). Agglomeration economists argue that the productivity of an area is a function of the number of workers available in that area. Hence, a densely populated agglomeration is expected to have higher productivity and concentration of economic activity. This effect is caused by, for example, knowledge spillovers and economies of scale. Glaeser & Gottlieb (2008) argue that in the presence of agglomeration economies, resources should be pushed towards regions with productivity that is both higher and more elastic with respect to agglomeration forming. In a similar framework, Kline (2010) shows that place-based policies targeting employment are most effective when relative demand and supply are inelastic. This counter-intuitive result suggests that place-based policy aiming to create jobs, is most successful in welfare terms when the least amount of jobs are created but local prices change. A further result by Kline (2010) is that in a theoretical framework incorporating agglomeration effects and heterogeneity, a large temporary intervention can have long-lasting equilibrium ef-fects. This so-called “big push” argument can be found in many theoretical models and is driven by the forces of agglomeration. The presence of strong agglomeration effects results in the existence of multiple spatial equilibria. A large temporary intervention should, in theory, be sufficient to push a region into a different, preferable equilibrium state.
difficult to identify in real markets. Additionally, Duranton (2011) shows that the absence or presence of imperfect worker mobility, imperfect information about wages and cost of living, imperfect property rights or a co-ordination failure have dramatic effects on optimal regional policy. The regional presence of these and other market failures is difficult to confirm empirically. Duranton (2011) argues that due to this ambiguity in market failure detection, it is unclear what exactly place-based policies should target, and how they should be implemented. It is argued that policymaker incompetence combined with the complexity of these policies can ultimately result in harmful outcomes. Instead, Duranton (2011) suggests that policymakers invest in easier to identify regional issues in, for example, the realm of infrastructure and transportation. Infrastructure improvements benefit the natural development of regional clusters which benefit from agglomeration effects. They do so by reducing the costs of transportation, reallocation, and commuting.
The main theoretical driver of place-based policy success, agglomeration effects, shows that an efficiency focus is instrumental in designing welfare-enhancing place-based policy (Glaeser & Gottlieb, 2008). Efficiency, however, is not the only aim of policymakers. Equity motives play a large role in politics and consequently are a large driver of regional policy. These equity motives played a role in the Mezzogiorno and many other regional development policies of the past decades. Equity motives might lead to governments accepting the creation of market distortions and inefficiencies, as long as the transfer of resources leads to a sufficient reduction in regional inequality. Hence, it is important to understand the policy’s aim and policy maker’s prioritization of equity relative to efficiency when evaluating program success.
2.2 Empirical Contributions
Now that the theoretical foundation of regional policy has been discussed, I will evaluate contemporary results and the current stance on place-based policy effectiveness in an em-pirical setting. Place-based policies have been analyzed in the context of many different fields within economic research. Researchers in the field of urban economics, for example, evaluate place-based policies targeting certain neighborhoods or zones within an urban area. On the other hand, development economics analyses place-based policies target-ing underdeveloped regions, often larger and more rural than the policies evaluated by urban economic scholars. Additionally, place-based policies play a role in the economics of housing, spatial economics, and spatial econometrics. This multidisciplinary aspect of place-based policies provides the opportunity to evaluate the topic from different angles.
agglomera-tion effects they discuss in their theoretical framework cannot be shown empirically. This result is striking, since without the presence of non-linear agglomeration effects, place-based policies have no way to produce economic gains. In addition, by evaluating the allied bombing of Japanese cities, Davis & Weinstein (2002) show that even the largest exogenous shocks do not alter long-run spatial equilibria, which counters the rhetoric that multiple spatial equilibria exist which we discussed earlier. In absence of these multiple equilibria, place-based policies will not be able to induce a “big push”. The disparities between the proposed theoretical solutions and empirical findings show that the already weak case for place-based policy becomes even weaker in the real world. I will discuss the existing research on different place-based policies, which makes use of a range of method-ologies, to further illustrate this point.
The empirical analysis of place-based policies mainly targets policy programs initiated in the United States and China. The federal urban Empowerment Zone (EZ) is one of the largest based policies in the United States and a popular program to assess place-based policy efficiency. Neumark & Kolko (2010) focus specifically on the Californian Enterprise Zone program. The novelty in the method used by Neumark & Kolko (2010) is how they describe zone boundaries and business locations using maps and geocoded observations. This method guarantees a correct assessment of treatment, even when policy boundaries do not follow census tracts or zip codes. In doing so, they find that the EZ does not increase employment, nor does it shift employment towards the targeted low-wage sectors.
Contradicting results are presented by Busso et al. (2013), who assess the efficiency of the general federal urban Empowerment Zone. By limiting their data to the first round of the policy program, the authors create the possibility to construct a control group consist-ing of future applicants to the EZ policy. This way the authors are better able to isolate the effects of the EZ from other factors. They find that the EZ program increased local wages and employment without the corresponding price and population increases spatial equilibria predict, thereby contradicting the results of Neumark & Kolko (2010). Wang (2013) takes a long-run perspective in evaluating the Chinese Special Economic Zones (SEZ). The SEZ differ from the United States’ EZ in the sense that the SEZ specifically targets foreign investors, whereas the EZ does not. Wang (2013) finds that the SEZ suc-ceed in their aim to increase per capita foreign direct investment, without crowding out domestic investment. Income implications, however, are not derived.
(2018) assess the efficiency of another German place-based policy, this time during the cold war period. The German Zonenrandgebiet, consisting of municipalities close to the iron curtain, received large scale transfers. Using a discontinuity design, they find increased economic density in the Zonenrandgebiet. However, the authors are rather pessimistic regarding the net effect of the policy, as the local increases in economic density are the result of reallocation instead of new economic activity.
Mayer et al. (2015) evaluate the French Enterprise Zones using spatial and time differencing. They find that the Enterprise Zones do affect the firms’ localization decision. However, the positive regional employment effect is offset by a negative effect on wages. As a result, the authors can not conclude on potential positive welfare effects. Givord et al. (2018) take a look at the same French Enterprise Zones. By taking a long-run perspective, the authors find a similar positive effect on employment and business location in the targeted regions during the first years of the policy. However, increased business discontinuation offsets this result in later years.
Many of the above-discussed papers are limited in their analysis by the fairly recent first initiation of the evaluated policies. As a result, none of the analyses can give a through long-run perspective. The exception of this observation and perhaps one of the most significant contributions to this field is the work by Kline & Moretti (2014a). Kline & Moretti (2014a) set out to evaluate the Tennessee Valley Authority (TVA), one of the largest and oldest place-based policies. Due to its size and ambitions, it is a prime example of a “big push” strategy as advocated by Kline (2010). Additionally, the TVA funding started in the 1930s and peaked between 1950 and 1955. As a result, the real long-run consequences and permanent effects can be assessed. By setting up a quasi-experiment, in which the control group consists of proposed additional authorities which were never implemented, the authors find long-run effects on manufacturing employment. These effects continued to intensify after the transfers were scaled down. It is argued that this effect is unique to the manufacturing sector due to the large role of agglomeration effects in this sector. Due to the higher manufacturing wages, the average income in the recipient counties increased during and after the TVA transfers. Aggregate income effects, however, are limited as other regions lose out.
co-determines the steady-state regional output levels. By increasing this steady-state output level, a regional policy can theoretically speed up convergence, as convergence increases the further a region is below its steady-state.
This section has shown that even though there exists a theoretical case for place-based regional policy, the empirical foundation of this result is worrisome. Nevertheless, policymakers do not fair the introduction of new place-based regional policies. Is this a sign of ignorance, or are economist simply not identifying the policy outcomes correctly? Identification of policy outcomes relies on the existence of a naturally occurring experi-mental setting or appropriate identification strategies. This should be the main focus of any research attempting to conclude on policy effectiveness. Therefore, I will extensively describe the setting and identification strategies I apply to research European regional policy
3
The European Regional Development Fund
Making use of the methodological advancements discussed above, I will be conducting an empirical analysis using data regarding the European Regional Development Fund (ERDF). According to the European Commission, the ERDF is a fund that ’aims to strengthen economic and social cohesion in the European Union by correcting imbalances between its regions’. The ERDF tries to correct these imbalances by channeling resources towards underdeveloped regions, making the ERDF one of the largest place-based policy programs in history.
The ERDF has had a long history of providing Europe’s less favored regions with financial aid. It was first created in 1975, shortly after the accession of Ireland, to redis-tribute money to the poorest regions European Commission (2019d). However, the limited budget prohibited the ERDF from any significant policy implementations. It took until the late 1980s, when new members Greece, Spain, and Portugal used their voting power, for the budgetary share to rise. This trend repeated itself consistently. Poorer regions joined the EU, which lead to budgetary changes in favor of regional development policy (Baldwin et al., 2009). Since the 2000s, roughly a third of the EU budget is spent on the less-favored regions.
and doubling efforts by increasing the budget to 168 billion euros. The 2000 to 2006 programming round followed the Lisbon Strategy. As a result, new policies shifted their focus on growth, employment, and innovation. Additionally, this round focused on the admission of ten new, mostly poor, member states. The ERDF budget increased towards 213 billion euros, with an additional 22 billion euros to support the new member states. The budget grew further for the 2007-2013 programming round, towards 347 billion eu-ros. Next to the accession of Romania, Bulgaria, and Croatia, this round introduced two key investment areas. Environmental infrastructure and climate change, as well as re-search and innovation. Furthermore, the emphasis on communication and transparency increased. The current programming round 2014-2020 aims to simplify funding rules and focuses on smart, sustainable and inclusive growth.
The main focus of this analysis is the 2000-2006 round of ERDF funded place-based policy. I have made this decision based upon the opportunity this period offers for a quasi-experimental setting. Additionally, standardized regional data are available from the start of this round until multiple years after the program round completion. This round of funding consists of different types of projects. The main project category, receiving over 100 billion euros of the total regional policy budget, is Objective 1 (SWECO, 2008). Objective 1, funded by the ERDF, targets the development of the least-favored regions of the EU. To be eligible for the Objective 1 program, the region’s per capita GDP needs to be below 75% of the EU average. Next to Objective 1, the ERDF provides funds to Objective 2, which targets reasonably developed regions facing difficulties in specific aspects of their economy or labor market. Furthermore, the first ERDF round funds some smaller programs focusing on interregional cooperation, sustainability, and innovation. As reflected by the budgetary share, Objective 1 is the main priority and the most likely to result in measurable effects. Consequently, I will target the ERDF Objective 1 program in the analysis below.
focuses on relatively safe investments in infrastructure. The remaining 37% is invested in the productive environment (33%), human resources (2%) and miscellaneous (2%). The share of investment in the productive environment is largest for richer countries, where infrastructure is often already highly developed.
These numbers show that the ERDF beliefs that infrastructure is key to fulfill its aim to “narrow the gap between the development levels of the various regions”. Taking note from the work discussed in section 2 (both theoretical (Duranton, 2011) and empirical (Kline & Moretti, 2014a)), I believe this could be the right call. Hence, I hypothesize that the ERDF Objective 1 program should lead to some, potentially small, income gains in the targeted regions. I expect the effect of the investment in the productive environment on employment to be marginal, as labor markets are complex and market failures are difficult to correctly identify and target. Since convergence between regions is the equity driven ERDF’s main objective, policy success can not be achieved without evidence of a decreasing income gap.
4
Experimental Design and Data
To produce meaningful results concerning the effectiveness of a place-based policy, a solid experimental framework is instrumental. Taking inspiration from Kline & Moretti (2014a), I aim to isolate the effect of Objective 1 on the recipient regions from other factors by selecting a control group based on the program eligibility requirements. In an ideal world, this would mean that we select some comparable EU regions with a GDP below 75% of the community average that did somehow not receive Objective 1 resources, as the control group. However, both this ideal world and proposed control group do not exist. Only two NUTS3 regions that are eligible based on their GDP did not receive funding because the larger NUTS2 region of which they are a part was not eligible. Fortunately, as of the first of January 2007, Romania and Bulgaria joined the EU. In 2013, the EU grew even further as Croatia joined. Large parts of these new member states would have been eligible for Objective 1 funding if they were members at the time. Furthermore, they received substantial resources in the second and third round of the ERDF, confirming their program eligibility. As a result, these regions form a suitable control group for the Objective 1 treatment during the first ERDF round.
as it provides the best opportunity for modeling and understanding regional differences. Furthermore, it assures a sufficient number of observations for the statistical analysis to be conducted. Objective 1 participation is established based upon the GDP statistics at the NUTS 2 level. However, as NUTS is a hierarchical system, all NUTS 3 regions that lie within a selected NUTS 2 region benefit from the Objective 1 funding. The exact list of treated NUTS 3 regions based on spending and financial commitment data is taken from the European Commission (2019b) and SWECO (2008).
NUTS3 regions are defined by several characteristics and thresholds. First of all, the population of each region is bounded by a minimum of 150 thousand and a maximum of 800 thousand. As a result, urban or densely populated areas will consist of geographically small NUTS3 regions, whereas the opposite holds for rural and sparsely populated areas. Secondly, the NUTS classification system favors local administrative divisions. Hence, most NUTS regions will follow territorial borders and consist of one or more adminis-trative units of the respective member state. Thirdly, the NUTS classification is prone to amendments. These amendments follow changes of the member states’ administrative structure and territorial borders. Amendments are normally processed once every three years at the most, except for the 2014 amendment caused by the substantial reorganization of the Portuguese administrative system.
The described treatment regions are depicted in figure 1. The figure, taken from the SWECO (2008) report, shows how the ERDF Objective 1 budget is used across the treated regions. The figure shows the size of the commitment and the composition across three categories for all treated NUTS 3 regions. The data are aggregated to the NUTS 2 level for the east German regions to allow for better visualization. The three categories depicted are infrastructure in green, the productive environment in blue, and human resources in red. As discussed in section 3, infrastructure investment is large with almost two-thirds of the funds being committed. However, there are some exceptions to this rule. Most notably in Scandinavia, the UK, and the France-Belgium border region, where a large share of funds is committed to improving the productive environment.
Using an overview of the Operational Programs adopted by the European Commis-sion, all NUTS 3 regions that receive ERDF funding in the first round between 2000 and 2006 can be described and categorized based on the type and amount of received fund-ing. These investments range from physical infrastructure to enhancing competitiveness, producing clean energy and even digital development. A complete panel dataset including tens of variables on Agriculture, Demographics, Welfare, Employment, Property rights, Tourism, Transport and more can be constructed on NUTS 3 level from 2000 onwards using the Eurostat database. For sake of consistency, all data are collected for the NUTS-2016 classification following the latest amendments.
working with international data. Using only one data source lowers the risk of inconsis-tency and improves the reliability of results. This dataset provides the opportunity to construct models with extensive controls. Numerous dependent variables like GDP, un-employment, and health can be analyzed in stock, per-capita and growth-rate variations. This allows isolating the effects of population mobility. I apply logarithmic transforma-tions to variables in levels and create appropriate ratios where necessary.
The resulting dataset consists of variables measured in 1298 regions between 2000 and 2010. This is less than the previously discussed total of 1348 NUTS 3 regions. This unfortunate discrepancy is due to data availability. The main cause is missing NUTS 3 GDP data for Belgium before 2003. As a result, no pre-treatment observations can be collected. Since no Belgium regions are part of the control or treatment group, this is of no real concern. I do not collect observations after 2010 since the effects of the second round of ERDF can be expected to interfere with the results. Of the 1298 included regions, 431 received funding from the first round of the ERDF. 91 regions received ERDF funding from the second or third round onwards, and are in the control group. The remaining 776 regions will play a role when I start evaluating the more flexible treatment and control definitions in section 6.
Using this dataset and the quasi-experimental framework discussed above, I aim to identify the effect of the ERDF objective 1 program on regional GDP and employment. As discussed in section 2.2, the correct identification and isolation of policy effects is the main empirical challenge in this field. Therefore, I will be implementing numerous techniques taken from the latest contributions in this field and that of policy evaluation in general. I will first discuss the basic difference-in-difference model, as it played a significant role in this field so far. This model is used to estimate the effect of the policy on regional GDP and employment. I perform sensitivity analyses to show how cross-sectional dependence interferes with the estimation results. Serious issues arise when the assignment of the ERDF treatment is found to be correlated with the output variables the model aims to explain. Section 6 attempts to resolve these issues by relaxing the fixed treatment and control conditions of the quasi-experimental framework. I will discuss multiple methods to do so, taken from the field of place-based policy and treatment analysis in general. The first method I use is the simple Oaxaca-Blinder decomposition, as proposed by Kline & Moretti (2014a). Next, I turn to Wooldridge (2010) and use statistically appropriate propensity scoring methods.
5
Difference-in-difference
regarding the assumptions needed for this method to produce unbiased results. First, the model will be estimated separately for a set of output variables. The results are presented in Table 1, where each column corresponds to an estimated model. Next, I perform a sensitivity analysis in which I show the influence of cross-sectional dependence on the difference-in-difference estimator. These results are presented in Table 2, where I show the difference-in-difference estimator for six models in six outcome variables. I end this section by discussing the necessary assumptions and show that difference-in-difference estimation is not an appropriate method to perform this specific analysis.
5.1 Model and results
Difference-in-difference estimation is the workhorse model of policy analysis. Hence, it will be the point of departure to assess the incidence of the ERDF. The basic dynamic panel data difference-in-difference estimation equation is given in (1).
Yit= αi+ β1Ti+ β2DDit+ δYit−1+ γXit+ ct+ it (1)
Yitis the output variable under investigation. To fully understand the effect of the ERDF
on targeted regions, I will evaluate six different output variables. I will estimate the effect of the ERDF Objective 1 program on regional GDP, GDP per capita, and the GDP per capita growth rate to evaluate income effects. Total regional employment, the regional employment rate, and the employment growth rate will be evaluated to understand the employment effects. I evaluate regional aggregates, per capita numbers, and growth rates to understand the role of migration and the ERDF’s effect on convergence. These variables correspond with the output measurements used by the European Commission to determine which regions are lagging behind and need additional support. Hence, improvements in these measurements caused by the place-based policy program signal policy success.
Ti is the treatment dummy, which is 1 for every region in the treatment group. DDit
is the difference-in-difference variable, which is equal to 1 for treatment regions when the policy is physically implemented. Since converting available funding into tangible out-comes that benefit the local economy takes time, this implementation date is later than the start of the funding round in 2000. The absorption of policy funds is a time-consuming process and different for each regional program. I follow SWECO (2008) and argue that significant absorption of the funds commenced from 2004 onwards. Hence, from 2004 on-wards the policy is treated as being physically implemented. Since absorption from the next round of funding takes a similar amount of time, the policy effects can be assessed in relative isolation up to 2010. Xit is a vector of numerous control variables, selected
employment. The model is estimated using time and region fixed effects and standard errors are clustered at the regional level. The estimation results are presented in Table 1. Every column represents a separate regression of the respective outcome variable on the treatment dummy, the difference-in-difference variable, and a set of control variables.
Table 1: Difference in difference estimates of treatment effect on selected outcome variables
(1) (2) (3) (4) (5) (6)
GDP GDP GDP p.c. Employment Employment Employment
per capita growth rate rate growth rate
Treatment 0.131∗∗∗ 0.128∗∗∗ 0.091∗∗∗ 0.001 -0.001 0.015∗∗∗ dummy (0.009) (0.009) (0.009) (0.004) (0.002) (0.005) Diff. in diff. -0.085∗∗∗ -0.082∗∗∗ -0.023∗∗∗ 0.021∗∗∗ 0.010∗∗∗ 0.007 (0.009) (0.010) (0.009) (0.004) (0.002) (0.005) Lagged output 0.743∗∗∗ 0.760∗∗∗ -0.094∗∗∗ 0.619∗∗∗ 0.641∗∗∗ -0.122∗∗∗ variable (0.013) (0.012) (0.021) (0.019) (0.020) (0.024) Employment 0.319∗∗∗ 0.277∗∗∗ 0.253∗∗∗ (0.027) (0.028) (0.031) Employment 0.458∗∗∗ 0.494∗∗∗ 1.204∗∗∗ 0.283∗∗∗ 0.105∗∗∗ 0.241∗∗∗ share of manu. (0.101) (0.106) (0.136) (0.067) (0.026) (0.077) Employment -0.511∗∗∗ -0.490∗∗∗ -0.079 0.593∗∗∗ 0.255∗∗∗ 0.358∗∗∗ share of AFF (0.071) (0.073) (0.052) (0.023) (0.056) (0.059) Population -0.204∗∗∗ -0.365∗∗∗ -0.111∗ -0.013 -0.030 0.026 density (0.051) (0.064) (0.060) (0.051) (0.026) (0.068) Population 3.423∗ 2.125 -6.916∗∗∗ 1.553 0.504 -2.258∗ change (1.754) (1.839) (1.998) (1.031) (0.433) (1.264) GDP 0.300∗∗∗ 0.002 -0.089 (0.060) (0.030) (0.062) GDP -0.199∗∗∗ 0.034 0.172∗∗∗ per capita (0.058) (0.029) (0.060) # Obs 4723 4723 4281 4723 4717 4281 R squared 0.893 0.897 0.181 0.667 0.630 0.127
Estimated using fixed effects and including a time trend and constant Clustered and robust standard errors in parentheses
∗
p < 0.10,∗∗ p < 0.05,∗∗∗p < 0.01
convergence theory (Barro et al., 1991).
The implications of these results for policy effectiveness are far from ideal. Receiving financial support from the ERDF Objective 1 program is associated with negative income effects in terms of aggregate, per capita, and growth levels. This model predicts that treatment is responsible for a reduction in GDP and GDP per capita of roughly 8 percent compared to the control regions. This claim does not take into account possible other policy programs, events, or unobserved variables that intervene with the Objective 1 program. Hence, it should be noted that the previously stated 8 percent is by no means an isolated effect of the Objective 1 program. The most notable interference is due to one of the fundamentals of this quasi-experimental framework, namely the accession to the EU of part of the control regions in 2007. Even though EU funding and policy is not expected to have an immediate impact on these new member-states, the mere notion of joining the EU is likely enough for a significant economic boost. This simultaneous effect interferes with the policy program and hence compromises the identification of the treatment effect. It is a prime example of the fundamental issues related to difference-in-difference estimation, which are often overlooked.
The effect of the policy on GDP and GDP per capita is almost identical. This suggests that the outcomes are not affected by large scale migration, which could potentially dilute the welfare effects. Other noteworthy results are the significant effect of employment composition on GDP and GDP per capita. A high share of workers in manufacturing is associated with high GDP, whereas the opposite holds for the share of workers in Agriculture, fishing, and forestry. Population density is found to inversely relate to income, suggesting that cities and urban regions have more difficulty in maintaining a reasonable welfare standard. The effect of the ERDF Objective 1 program on employment is found to oppose the income implications. Even though the growth rate of the employment rate is unaffected, treatment significantly increases both the total employment and the employment rate in the treated regions. The size of the estimated treatment effect on total employment is considerably larger than the effect on the employment rate. This result suggests that the policy potentially induced some form of labor migration.
might be considered the policy evaluation workhorse, without perfect conditions it does not perform adequately.
5.2 Robustness and issues related to identification
The estimation results presented in Table 1 are highly significant. However, this signifi-cance could potentially be caused by model misspecification or incorrect identification of the treatment itself. These potential issues will be evaluated in this section. First, I per-form a sensitivity analysis to evaluate potential model issues and validate the robustness of the results. Second, I test for the necessary independence of treatment on pre-treatment outcome variables. This independence is instrumental for correct identification of treat-ment effects using a difference-in-difference model.
One of the issues that arise in the specification given in equation (1), is cross-sectional dependence. Both regional GDP and the regional employment rate tend to move hand in hand with national levels as well as unobserved national characteristics. As a result, the er-ror component may be cross-sectionally correlated. This suspicion is confirmed by running a CD-test for cross-sectional dependence as constructed by Pesaran (2004, 2015). The test confirms that both GDP and employment are cross-sectionally dependent. The presence of cross-sectional correlation in the error term is the consequence of not adequately spec-ifying the model to account for this co-dependence. Had the cross-sectional component been specified correctly, the resulting disturbance would be uncorrelated across observa-tions (Sarafidis & Wansbeek, 2012). Due to the presence of cross-sectional correlation, the estimation results in Table 1 are potentially biased and inconsistent. Unfortunately, resolving this issue is not straightforward. Cross-sectional dependence can be modeled in different ways. Examples are the use of spatial weight matrices to describe the spatial interaction between regions (LeSage & Pace, 2009), common factors to allow regions to heterogeneously move along with national averages (Pesaran, 2006), or a combination of both (Vega & Elhorst, 2016; Bailey et al., 2016). The usage of spatial weight matrices requires the dataset to be free of missing observations, as the spatial relationship between regions can not be defined otherwise. The method of common factors requires the addi-tion of N variables to allow a heterogeneous response. Hence, this method is only feasible with small N , large T datasets. Unfortunately, the current data fit neither of the above requirements.
es-Table 2: Difference-in-difference estimates of treatment effects using selected models
(1) (2) (3) (4) (5) (6)
Base CD CD errors All All regions All regions CD
model errors and trends regions CD errors errors and trends
GDP -0.085∗∗∗ -0.085∗ -0.054 -0.008∗∗∗ -0.008 -0.003 (0.009) (0.042) (0.046) (0.003) (0.016) (0.007) GDP -0.082∗∗∗ -0.082∗ -0.052 -0.009∗∗∗ -0.009 0.001 per capita (0.010) (0.043) (0.046) (0.003) (0.016) (0.008) GDP p.c. -0.023∗∗∗ -0.023 -0.141∗∗ -0.017∗∗∗ -0.017 -0.027 growth rate (0.009) (0.040) (0.062) (0.004) (0.019) (0.021) Employment 0.021∗∗∗ 0.021 -0.002 0.005∗∗ 0.005 0.011∗∗ (0.004) (0.014) (0.017) (0.002) (0.007) (0.004) Employment 0.010∗∗∗ 0.010∗ -0.003 0.002∗ 0.002 0.004∗∗∗ rate (0.002) (0.006) (0.007) (0.001) (0.003) (0.001) Employment 0.007 0.007 -0.009 -0.006∗∗∗ -0.006 0.007 growth rate (0.005) (0.018) (0.012) (0.002) (0.006) (0.006) # Obs ≈ 4500 ≈ 4500 ≈ 4500 ≈ 11000 ≈ 11000 ≈ 11000
Estimated using fixed effects and including a time trend Clustered and robust standard errors in parentheses
∗
p < 0.10,∗∗ p < 0.05,∗∗∗ p < 0.01
timation results with standard errors clustered at the national level in column 2 of Table 2. Additionally, to control for correlated trends within countries I include country-specific time trends in column 3. Column 1 summarizes Table 1 and presents the results of the base difference-in-difference model. Table 2 is constructed to present the estimated treatment effect using a unique regression in every cell. The row determines the outcome variable under investigation, whereas the column gives the used estimation method.
the treatment itself should be randomly assigned. Randomization of treatment, however, is often unattainable in policy research. Ideally, the perfect counterfactual would be observable, meaning that every region could be observed in two states at the same time. One state in which it received the treatment, and one state in which it did not. This observation of a perfect counterfactual is impossible to obtain in the real world. In the second-best case, treatment is assigned randomly and the control group serves as the counter-factual. However, in the case of place-based policy targeting disadvantageous regions, the treatment assignment is not random. Regions are being targeted because of certain characteristics. Even though the quasi-experimental framework discussed and used here tries to produce an appropriate counter-factual, it likely is not.
Table 3: The dependence of treatment on base year GDP and employment
(1) (2) (3) (4)
Counterfactual Counterfactual All regions All regions
GDP 0.312∗∗∗ -0.221 (0.095) (0.141) Employment -0.296∗∗∗ 0.235 (0.089) (0.160) GDP per capita 0.304∗∗∗ -0.209 (0.093) (0.148) Employment rate -0.813∗∗ -0.408 (0.308) (0.624) # Obs 439 439 1105 1105 R squared 0.617 0.616 0.152 0.185
Base year estimation including a constant and covariates describing demographics and other regional characteristics.
Standard errors in parentheses, clustered at country level.
∗
p < 0.10,∗∗ p < 0.05,∗∗∗p < 0.01
The assumption that assignment of treatment is independent on pre-treatment values of the outcome variable Yi0, is necessary for the difference-in-difference estimator to be
Figure 2: The average employment rate in the treatment and control regions
finding by plotting the population-weighted average GDP per capita and the employment rate for the control group, treatment group, and all other EU regions. Contrary to the results in column 1 and 2, column 3 and 4 of Table 3 show that when evaluating the complete dataset, consisting of the treatment regions and all other regions, GDP and employment do not significantly affect the treatment eligibility. It should be noted that this result holds only when estimation adequately controls for region-specific characteristics by including a set of covariates.
The conditional independence of GDP and employment on being assigned treatment when evaluating the complete dataset suggest that it is beneficial to drop the counter-factual altogether. Hence, I evaluate the effectiveness of the place-based policy using all non-treatment regions as the counterfactual. These results are presented in columns 4, 5, and 6 of Table 2. First of all, this estimation procedure leads to a reduction in the size of the income and employment effect compared to the base model. Whereas the effect of treatment on GDP remains negative and highly significant, the employment effect is re-duced. Correcting the specification to control for cross-sectional dependence produces the final results in column 5 and 6. In line with the estimation in column 3, the effect of treat-ment on GDP diminishes completely. However, a significant positive effect of treattreat-ment on employment levels and the employment rate remains.
Figure 3: The average GDP per capita in the treatment and control regions
counterfactuals. Therefore, constructing an adequate counterfactual using the appropriate statistical tools is essential to reveal the true treatment effects. I find that correcting for cross-sectional dependence as well as improving upon the counterfactual diminishes the initial results. Only the positive effect of the place-based policy under evaluation on local employment is robust to these corrections. The next step is to further improve upon the identification of control and treatment regions and evaluate the true effectiveness of the ERDF Objective 1 program.
6
Methods assuming ignorability of treatment
Now that I have shown that the ERDF treatment dummy is correlated with the output variables, an additional assumption needs to be made to correctly identify treatment ef-fects. This assumption, ignorability of treatment, was first introduced by Rosenbaum & Rubin (1983) and states that conditional on the set of control variables Xi, the treatment
indicator and the output variables are independent. When this assumption holds, it is possible to estimate the effect of treatment on the treated regions even when treatment is not randomly assigned (Wooldridge, 2010). As shown in Table 3, conditional on the co-variates Xi, treatment is independent of the output variables GDP and employment when
evaluating the entire sample. In other words, the ignorability of treatment assumption holds when the counterfactual consists of all non-treated regions. As a result, I can apply different methods which assume ignorability of treatment to estimate the average treat-ment effect conditional on Xi. I start by constructing an Oaxaca-Blinder decomposition,
flexible counterfactual, it leaves room for improvements. I try to find these improvements by following Wooldridge (2010) and constructing propensity score weighting estimates.
6.1 Oaxaca-Blinder decomposition
Following Kline & Moretti (2014a), I aim to improve upon the basic difference-in-difference estimation by implementing an Oaxaca (1973) Blinder (1973) decomposition as con-structed in Jann (2008). This decomposition divides the mean difference in an outcome variable between two groups into a part that is explained by group characteristics and covariates and an unexplained part that is unaccounted for. This unexplained part en-compasses, amongst other unobservables, the effect of the treatment under investigation. Since the Oaxaca-Blinder decomposition can not decompose observations over time, I fol-low Kline & Moretti (2014a) and construct a cross-sectional dataset consisting of base year observations of the covariates, as well as the difference between post and pre-treatment values of the outcome variables GDP and employment. More precisely, I compute the increase of GDP and employment over the treatment period as Y = YiT − Yit, where
t = 2000 and T = 2010. I use a logit model to determine the probability of a region re-ceiving treatment conditional on the pre-treatment covariates. Following Kline & Moretti (2014a), I drop all non-treated regions with a probability of treatment lower than 0.25. These regions are the least informative regarding policy effectiveness. Dropping them in-creases the comparability of the ERDF Objective 1 and control regions. The remaining sample consists of 369 treated regions and 332 control regions.
The first step of the decomposition is to fit basic regression models to the pooled sample, including a group dummy for the control and treatment regions.
Yi= α + βXit+ γ + i. (2)
Y is the constructed increase in the dependent variable over the treatment period. Xit is
the set of pre-treatment covariates and γ is the group dummy. The resulting estimates for β, β∗, are used to predict a counterfactual mean for the treated regions. As a result, the difference between the treatment group A and the control group B can be decomposed in
YA− YB= E + U, (3)
where
E = [ ¯XA− ¯XB]0β∗, U = [ ¯XA0 (βA− β∗) + ¯XB0 (β ∗− β
B)]. (4)
predictors, weighted by the pooled regression coefficients. U is the remainder of the mean difference that is not explained by the pre-treatment covariates. Decomposed sample variances are estimated following Jann (2008) and clustered at the country level. The result of the Oaxaca-Blinder decomposition on GDP increase, GDP per capita increase, employment increase, and employment rate increase over the 2000-2010 period is presented in Table 4.
Table 4: Oaxaca-Blinder decomposition of regional differences
(1) (2) (3) (4)
GDP GDP Employment Employment
per capita rate
Difference 0.058 0.057 0.010 0.007 (0.114) (0.136) (0.038) (0.010) Explained 0.024 0.033 -0.004 0.002 (0.107) (0.130) (0.031) (0.007) Unexplained 0.035 0.024 0.014 0.006 (0.035) (0.031) (0.021) (0.008) # Obs 701 701 701 701
Standard errors in parentheses.
Standard errors clustered at country level.
∗
p < 0.10,∗∗ p < 0.05,∗∗∗p < 0.01
decomposition regression, which allows for the construction of a double robust estimator.
6.2 Propensity score weighting
In the previous section, I discussed the assumption of ignorability of treatment, and how this assumption is sufficient for the identification of conditional treatment effects. To identify unconditional treatment effects, an additional assumption has to be made. This assumption, the overlap assumption, states that for all settings of covariates in the assumed population, there is a change of seeing units in both the control and treatment groups (Wooldridge, 2010). The simultaneous acceptance of the ignorability of treatment and the overlap assumption is called strong ignorability (Rosenbaum & Rubin, 1983). The overlap assumption can be tested using the propensity score, the probability of treatment as a function of the covariates. The overlap assumption rules out the case where the propensity score is one or zero. Hence, for strong ignorability and the resulting identification of unconditional treatment effects to hold, the overlap assumption needs to be checked and validated. I estimate the propensity score using a logit function regressing treatment on a set of economic and demographic covariates. Using the results, I follow Kline & Moretti (2014a) and drop all regions with a propensity score below 0.25 from the sample to assure the overlap assumption holds. Additionally, this assures the remaining regions are informative about the policy program effectiveness.
The propensity scores will be used to estimate treatment effects in two ways. First, I will use the propensity scores directly in a method called inverse propensity score weighting (IPW) to compute the treatment effect. Second, I use the propensity score to perform a regression adjustment called the augmented inverse propensity score regression (AIPW). I use the same dataset, consisting of pre-treatment covariates and treatment period increases of GDP and employment, as used for the Oaxaca-Blinder decomposition.
The IPW estimation equation for the average treatment effect (ATE) and the average treatment effect on the treated (ATT) as constructed by Wooldridge (2010) are given by equation 5 and 6. The ATE estimator is
ate = N−1 N X i=1 [wi− ˆp(xi)]yi ˆ p(xi)[1 − ˆp(xi)] , (5)
where N is the number of observations, wi is the treatment dummy for region i, ˆp(xi) is
the estimate for the propensity score of region i given the vector of covariates xi, and yi
is the output variable. The ATT estimator is given by
where the unconditional probability of treatment is simply given by ˆp, the fraction of regions that receive treatment. This unconditional probability of treatment suffices for the ATT estimator due to the weaker assumptions that are needed to correctly identify the ATT (Wooldridge, 2010). I bootstrap both the propensity score and the estimation equations to obtain proper standard errors. The results are presented in column 1 and 2 of Table 5.
Column 1 presents the effect of the ERDF Objective 1 policy on regional income and employment. I find a significant positive effect on aggregate GDP, combined with a moderately significant effect on employment and the employment rate. Column 2 presents the policy effects on the treated regions alone. It can be seen that the effect on GDP does not translate from the general treatment effect to the treatment region-specific effect. The effect on employment, however, is robust for both estimators. These results do not convince me of the beneficial effects of the ERDF Objective 1 program. They hint back at the basic difference-in-difference specification, suggesting that the policy program is not capable of countering the unfavorable outset of the treatment regions. However, just as the basic difference-in-difference model, this first propensity score method has some drawbacks I need to discuss and tackle before jumping to conclusions.
6.3 Augmented propensity score weighting regression
The results of the IPW estimation are highly dependent on the correctness of the spec-ification of the parametric propensity score model. The logit model assumed earlier is the go-to approach, but it is not unchallenged. Additionally to biased estimates, mis-specification will lead to false standard error calculations and potentially wrong inference (Wooldridge, 2010). As a result, the estimates presented in column 1 and 2 of Table 5 are not robust to different specifications of the propensity score model.
Additional robustness to misspecification can be obtained by combining the IPW es-timator with regression adjustments (Robins et al., 1995; Robins & Rotnitzky, 1995). The resulting estimator is called double robust since it requires only the conditional mean or the propensity score model to be specified correctly. This method is called the augmented inverse propensity score weighted regression (AIPW). The first step is to estimate the propensity score ˆp(xi). Next, this propensity score is used to construct the weight
vi = wi ˆ p(xi) + 1 − wi 1 − ˆp(xi) . (7)
Using vi to weight observations, I regress a constant, the treatment dummy wi, and the
interactions between the treatment dummy and the vector of covariates xi on yi to obtain
ˆ
Table 5: Treatment effects of the ERDF Objective 1 program
(1) (2) (3) (4)
IPW - ATE IPW - ATT AIPW - ATE AIPW - ATT
GDP 0.030∗∗ 0.021 0.036∗∗∗ 0.038∗∗∗ (0.014) (0.025) (0.012) (0.012) GDP per capita 0.018 0.004 0.026∗∗ 0.028∗∗ (0.014) (0.027) (0.011) (0.012) Employment 0.015∗ 0.019∗∗ 0.015∗ 0.013 (0.008) (0.009) (0.008) (0.009) Employment rate 0.005∗ 0.005 0.006∗ 0.005 (0.003) (0.004) (0.003) (0.003) # Obs 701 701 701 701 # Bootstrap reps 500 500 500 500
Bootstrapped standard errors in parentheses.
∗
p < 0.10,∗∗ p < 0.05,∗∗∗ p < 0.01
predict the outcome ˆy1 and ˆy0 for a treated and untreated sample respectively. The
resulting double robust ATE estimator is then given by
ate = N−1
N
X
i=1
[(α1+ xiβˆ1) − (α0+ xiβˆ0)], (8)
the mean difference between the predicted outcomes for treated and untreated regions using the probability of treatment as weights to determine the coefficients. The ATT is estimated using the same equation, only averaging over the treated sample. Standard errors are computed using the bootstrapping technique over both the propensity score and the estimation equation. The results are presented in column 3 and 4 of Table 5.
7
Discussion
Even though AIPW produces interesting estimates for the effect of the place-based policy on European regions, I would argue that the most interesting result of this work is the displayed sensitivity of treatment effects to estimation methods and the counterfactual selection. It shows that the model and estimation decisions the researcher makes are decisive and should be adequately discussed. Hence I will describe my decisions and the robustness of the results to potential alternative decisions in detail in this section.
Starting with the difference-in-difference estimation, two choices have to be made. First of all, the control and treatment group have to be constructed. Where treatment is a given, the selection of a control group is largely based on the researchers’ intuition. An alternative control group can be constructed based on, for example, a region’s GDP per capita. I construct two alternative control groups. One contains all regions that are not in the treatment group with a GDP per capita below 80% of the EU average. For the second control, group I increase this threshold to 85%. Estimating the model using these alternative control groups produces results that are similar in sign, but less significant than those presented in Table 1 and 2.
Second, I make some alternations to the treatment date. In the previous estimates, I assumed the policy to be in effect from 2004 onwards, due to the relatively slow absorption of funds as shown by SWECO (2008). Alternatively, I ignore this claim and set treatment to be in effect from 2002 or 2003 onwards. This results in a reduction of the negative income effect and the complete disappearance of the positive employment effects in the base difference-in-difference specification. Increasing the absorption time even further and setting the treatment commencement date to 2005 or 2006 does not change the results. All in all, the sign of the difference-in-difference estimator is robust to alternative control group selections and treatment dates. The significant levels, however, can be reduced considerably by making different decisions.
this alternative decision. This does not mean the results in Table 5 are wrong. Ignoring a part of the policy program is expected to reduce the significance of the results.
Generally speaking, measuring policy effectiveness is a challenging task. Making ap-propriate decisions regarding timing and control and the treatment group selections is just one of the challenges in this field. Perhaps the largest challenge is the isolation of treat-ment effects. The reality is that in the real world, countless observable and unobservable factors interfere with the policy program. As a result, all estimates for the effect of the policy I present here, are not isolated from other unobserved policy programs, events, and factors. They should be interpreted as the effect of the ERDF Objective 1 program on European regions, allowing for the potential response of other policy programs and events that occur over the same period. An interesting observation, for example, is how almost all targeted regions are located in southern or eastern Europe. These regions have specific political systems that differ from those in northern and western Europe. How these politi-cal differences potentially influenced the allocation and success of regional policy programs could be an interesting path for future research.
So far I discussed the dependence of the results on the methodology and the selection of a counterfactual. A third factor to potentially influence the estimated policy effec-tiveness is the composition of the policy program itself. Previously, I followed Duranton (2011) and argued that the ERDF’s decision to assign the majority of available funds to-wards regional infrastructure is the right call and should lead to efficient outcomes. This is a strong assumption to make, which completely ignores one of the potential causes of policy effectiveness. Unfortunately, it is outside of the scope of this work to challenge and verify this assumption by evaluating regions based on the local composition of the policy program. Hence, I will leave it as a suggestion for future research. The results presented in this work are specific for a policy mainly targeting the development of infrastructure and are not tested for robustness to changes in the policy decomposition.
8
Conclusion
References
Ahlfeldt, G. M., Maennig, W., & Richter, F. J. (2016). Urban renewal after the berlin wall: A place-based policy evaluation. Journal of Economic Geography, 17 (1), 129–156.
Bailey, N., Holly, S., & Pesaran, M. H. (2016). A two-stage approach to spatio-temporal analysis with strong and weak cross-sectional dependence. Journal of Applied Econo-metrics, 31 (1), 249–280.
Baldwin, R. E., Wyplosz, C., & Wyplosz, C. (2009). The economics of european integration (Vol. 3). McGraw-Hill Education Berkshire.
Barro, R. J., Sala-i Martin, X., Blanchard, O. J., & Hall, R. E. (1991). Convergence across states and regions. Brookings papers on economic activity, 107–182.
Blinder, A. S. (1973). Wage discrimination: reduced form and structural estimates. Journal of Human resources, 436–455.
Busso, M., Gregory, J., & Kline, P. (2013). Assessing the incidence and efficiency of a prominent place based policy. American Economic Review , 103 (2), 897–947.
Chenery, H. B. (1962). Development policies for southern italy. The quarterly journal of economics, 515–547.
Davis, D. R., & Weinstein, D. E. (2002). Bones, bombs, and break points: the geography of economic activity. American Economic Review , 92 (5), 1269–1289.
Duranton, G. (2011). California dreamin’: The feeble case for cluster policies. Review of Economic Analysis, 3 , 3.
Ehrlich, M. v., & Seidel, T. (2018). The persistent effects of place-based policy: Evidence from the west-german zonenrandgebiet. American Economic Journal: Economic Policy, 10 (4), 344–74.
European Commission. (2019a). European regional development fund. https://ec .europa.eu/regional policy/en/funding/erdf/. (Online; accessed 30-July-2019) European Commission. (2019b). The eu’s main investment policy. https://ec.europa
.eu/regional policy/en/policy/what/investment-policy/. (Online; accessed 14-June-2019)
European Commission. (2019d). History of the policy. https://ec.europa.eu/regional policy/en/policy/what/history/. (Online; accessed 14-June-2019)
Givord, P., Quantin, S., & Trevien, C. (2018). A long-term evaluation of the first gener-ation of french urban enterprise zones. Journal of Urban Economics, 105 , 149–161.
Glaeser, E. L., & Gottlieb, J. D. (2008). The economics of place-making policies. Brookings Papers on Economic Activity, 2008 (1), 155–253.
Jann, B. (2008). The blinder–oaxaca decomposition for linear regression models. The Stata Journal , 8 (4), 453–479.
Kline, P. (2010). Place based policies, heterogeneity, and agglomeration. American Economic Review , 100 (2), 383–87.
Kline, P. (2011). Oaxaca-blinder as a reweighting estimator. American Economic Review , 101 (3), 532–37.
Kline, P., & Moretti, E. (2013). Place based policies with unemployment. American Economic Review , 103 (3), 238–43.
Kline, P., & Moretti, E. (2014a). Local economic development, agglomeration economies, and the big push: 100 years of evidence from the tennessee valley authority. The Quarterly Journal of Economics, 129 (1), 275–331.
Kline, P., & Moretti, E. (2014b). People, places, and public policy: Some simple welfare economics of local economic development programs. Annu. Rev. Econ., 6 (1), 629–662.
LeSage, J., & Pace, R. K. (2009). Introduction to spatial econometrics. Chapman and Hall/CRC.
Marshall, A. (1890). Principles of economics. Macmillan, London.
Mayer, T., Mayneris, F., & Py, L. (2015). The impact of urban enterprise zones on establishment location decisions and labor market outcomes: evidence from france. Journal of Economic Geography, 17 (4), 709–752.
Neumark, D., & Kolko, J. (2010). Do enterprise zones create jobs? evidence from califor-nia’s enterprise zone program. Journal of Urban Economics, 68 (1), 1–19.
Oaxaca, R. (1973). Male-female wage differentials in urban labor markets. International economic review , 693–709.
Pesaran, M. H. (2006). Estimation and inference in large heterogeneous panels with a multifactor error structure. Econometrica, 74 (4), 967–1012.
Pesaran, M. H. (2015). Testing weak cross-sectional dependence in large panels. Econo-metric Reviews, 34 (6-10), 1089–1117.
Robins, J. M., & Rotnitzky, A. (1995). Semiparametric efficiency in multivariate regression models with missing data. Journal of the American Statistical Association, 90 (429), 122–129.
Robins, J. M., Rotnitzky, A., & Zhao, L. P. (1995). Analysis of semiparametric regression models for repeated outcomes in the presence of missing data. Journal of the american statistical association, 90 (429), 106–121.
Rosenbaum, P. R., & Rubin, D. B. (1983). The central role of the propensity score in observational studies for causal effects. Biometrika, 70 (1), 41–55.
Sarafidis, V., & Wansbeek, T. (2012). Cross-sectional dependence in panel data analysis. Econometric Reviews, 31 (5), 483–531.
SWECO. (2008). Final report - erdf and cf regional expenditure contract no 2007.ce.16.0.at.036y. https://ec.europa.eu/regional policy/sources/docgener/ evaluation/pdf/expost2006/expenditure final.pdf. (Online; accessed 14-June-2019)
Vega, S. H., & Elhorst, J. P. (2016). A regional unemployment model simultaneously ac-counting for serial dynamics, spatial dependence and common factors. Regional Science and Urban Economics, 60 , 85–95.
Wang, J. (2013). The economic impact of special economic zones: Evidence from chinese municipalities. Journal of development economics, 101 , 133–147.
