Assessing the impact of special economic zones on regional growth through a comparison among EU countries

ABSTRACT The establishment of incentive zones represents a cluster policy approach that a country can implement to boost economic growth. Being influenced by regional characteristics, the success of specific incentive types (special economic zones, free-trade zones, free zones) may vary considerably. By using a new composite indicator, differently designed according to each type of development plan, this study analyses 51 European incentive zones so as to identify best practices. Moreover, propensity score matching allows the net impact of the industrial policy to be assessed. The findings show significant positive results achieved by the various industrial policy instruments with differing levels of success.


INTRODUCTION
Since the mid-1990s there has been extensive academic research into industrial policy instruments. The establishment of different kinds of economic incentive zones (IZs) represents one of the main policy instruments used to attract investments, thereby boosting industrial production, economic growth and even exports (Davies & Mazhikeyev, 2019). Such instruments, used in both industrialized and developing countries, aim to support specific sectors or promote the development of peripheral areas (Pasha, 2019). They are designed to generate circular and cumulative growth processes and to initiate or expand export-oriented manufacturing, promoting structural change (Foreign Investment Advisory Service (FIAS), 2008; Asian Development Bank (ADB), 2015). According to Narula and Zhan (2019), IZs are defined as delimited areas with location-specific advantages differing from those available in the rest of the country or countries concerned.
The generic term 'incentive zone' (IZ) covers various types of policy incentives that may coexist in the same country Narula & Zhan, 2019), such as special economic zones (SEZs), free zones (FZs), industrial parks, free-trade zones (FTZs), free ports, foreign trade zones and export processing zones (ADB, 2015;Pasha, 2019;Bost, 2019). However, whatever the term used, the core concept remains unchanged.
At present, according to the United Nations Conference on Trade and Development (UNCTAD) (2019), there are nearly 5400 active IZs in 147 countries worldwide, and the trend is increasing. Most of the world's IZs are located in developing countries, with the majority in Asia (4046), followed by Latin America (486) and Africa (237). About 25% of such zones may be termed under-utilized and 22% heavily under-utilized (UNCTAD, 2019).
While in many countries IZs have met with enormous success, achieving their set objectives, in others few spillover effects have been generated (Jarosiński & Maśloch, 2016). Subsequently, the global picture is somewhat heterogeneous in both developing and industrialized countries.
Although research worldwide has identified the best practices as well as shortcomings in creating an IZ, with regard to the European Union (EU), where at least 23 member states play host to at least one such zone, comprising a total of 97 zones, research is both scant and less specific. Since EU zones are regulated by the Union Customs Code (UCC) and must comply with the rules on state aid, IZs, as part of the EU's wider policy on economic development, must be established in regions that benefit from regional assistance (Ambroziak, 2016). This regulation should have led to similar benefits on an EU-wide basis. However, the possibility of differentiating each benefit within each IZ, as well as their magnitude, has generated enormous disparities in the impacts achieved (Jarosiński & Maśloch, 2016;Nazarczuk & Umiński, 2017). Given the complexity of the regulatory mechanism and the rate at which IZs have spread, the need to assess the performance of such areas, in terms of spillovers on the surrounding areas, appears fundamental.
Recent studies have progressively documented both the successes and failures of special zones in the EU, based on country-specific cases, yet have been unable to suggest implications for policymakers . Given the lack of an overall vision of this policy approach, in-depth, long-term investigation would appear to be required to assess the impacts and potential benefits of IZs.
Our research sets out to provide a methodology to both evaluate the policy implemented by each country involved and compare their activities among the European special zones. By using a combined approach, the long-term effects of IZs on the surrounding regional economies were analysed, distinguished by type of development plan (SEZs, FTZs, FZs).
According to the theory of change (Rogers, 2014), we first identify the desired long-term goals and then work back from these to identify all the conditions (outcomes) that must be in place (and how they relate to one another causally) for achieving the goals. In so doing, the building of composite indicators (CIs) should allow us to analyse the impacts of such investments on the local economy. These indexes provide a 'measure' that summarizes information on current movement, trends and processes able to support the policymaker to make decisions that could influence future results. Despite their limitations, according to the Organisation for Economic Co-operation and Development (OECD) (2017) CIs are now widely recognized as a fundamental part of all governance models and can be seen as tools to evaluate policies aimed at attracting investments and increasing economic growth. Subsequently, counterfactual analysis allows us to evaluate the net impact of the industrial policy as the gap differential compared with other European regions without IZs, and to highlight potential spillover effects in the surrounding regions as well.
In relation to the existing literature, we seek to add empirical evidence on the economic effects of European economic zones. By considering the whole sample of IZs in the EU, we fill the void of comparable cross-country data measuring IZ outcomes and provide interesting insights for policymakers. Moreover, following Frick and Rodríguez-Pose (2019), we identify the drivers of IZ dynamism across countries from several comparative perspectives: by considering all the IZs in the EU; by providing a comparison with the control groups; by providing clearer policy implications; by identifying a support index that can be replicated for analysing other case studies (The World Bank Group, 2008;Zeng, 2016;UNCTAD, 2019).
The paper is structured as follows. The next section reviews the existing literature on the impacts of SEZs.
The third section provides a description of the data, the methodology implemented and the empirical results. Finally, the main conclusions are drawn.

BRIEF LITERATURE REVIEW
The economic impacts that the establishment of an IZ could have on a country's economy have been extensively explored by scholars and institutions (The World Bank Group, 2008). However, a lack of cohesion emerges in many studies: most encompass only individual aspects of impact evaluation, remaining less explicit about the overall impact of IZs as effective industrial policy instruments. Our research seeks to fill this gap, appraising this industrial policy from two viewpoints: the first concerns the direct impact of the policy, according to the administrators' stated goals for each zone, while the second entails evaluation of the achieved objectives of the zone in order to assess the spillover effect of the policy in the surrounding areas. In order to define a coherent framework for policy assessment, it helps, in looking at the economic literature, to capture the main determinants, hitherto studied separately. In so doing, we grouped the main determinants expected by policymakers into six main macro areas.
. Research and development (R&D): technological innovation becomes an important tool of the long-term IZ strategy since the overriding aim is to increase firm productivity and competitiveness (OECD, 2017). According to Zeng (2016), such positive outcomes are strictly correlated with both well-focused R&D expenditure and the strengthening of university-industry linkages. . Production structure: since IZs are an important instrument of industrial policy, one of the necessary factors for achieving growth is economic diversification and structural change (UNCTAD, 2019). Lall (2000) suggests that IZs in their role as catalysts could promote technology transfer and upgrading from industrialized to developing economies, with a subsequent increase in skills levels, and in research and design activities. . Infrastructural endowments: infrastructural support is one of the main objectives of IZ policy, especially in developing countries (UNCTAD, 2019). A proper level of infrastructure will certainly lead to a greater positive impact on the region involved. Thus, connections to key infrastructure nodes and links such as ports, railways and highways allow IZs to become catalysts for economic change on a nationwide basis. Such infrastructures should be created before attracting investors to avoid the risk of the IZ becoming an enclave (Aggarwal, 2010). On the other hand, it may be the implementation of an IZ itself that provides off-site infrastructures (Pakdeenurit et al., 2017). Therefore, national programmes should provide companies that invest in special areas with a favourable endowment of infrastructures, such as reliable energy and water supply, and efficient road, rail and telecommunication networks (Zeng, 2016).
. Environment: the spread of IZs has led to several challenges to environmental regulators (Richardson, 2004). Several studies have highlighted the link between manufacturing activities and greenhouse gas emissions (Davies et al., 2018). At the same time, the literature includes many country-specific studies showing high environmental standards for IZs, therefore indicating that the results obtained cannot necessarily be generalized (Madani, 1999). Furthermore, given the severe environmental challenges at both local and global levels, there is widespread evidence of new types of zones, that is, so-called eco-industrial zones, applying appropriate environmental standards (Sosnovskikh, 2017). . Contextual factors: with respect to contextual factors, the role of the state is considered a major determinant with its regulatory and institutional frameworkweak governance structures or the coexistence of too many institutions can lead to the failure of IZs (Cheng, 2019;UNCTAD, 2019). Furthermore, IZ growth spillovers may also depend on the institutional environment of the host country . IZs, in turn, can have a positive impact on government budgets, whose main gains come from both taxation on stakeholders' personal income and from corporate income tax (Lion, 2014). . Economic background: the creation of special areas can positively influence the level of exports, the performance of local companies as well as employment, especially in relation to other countries where there are no IZs (Johansson & Nilsson, 1997). Among the main benefits expected from IZs, the increase in foreign currency earnings is one of the direct consequences of higher export levels. Recently, Zeng (2016) concluded that foreign companies can provide local companies in SEZs with high levels of knowledge and professional skills from which, in the absence of such areas, they would not be able to benefit. Furthermore, local businesses, in turn, influence the entire economic apparatus of the host country: therefore, national governments use special areas as a way of attracting investments to those sectors where they have no obvious comparative advantage or as a means of increasing the added value of their production and export activities.
The impact of IZs on employment is a widely debated topic in the literature since it concerns several issues: gender equality, wages, workers' rights and working conditions in developing countries (Cling et al., 2005;Aggarwal, 2007).

DATA AND METHODOLOGY
In order to evaluate the impact of IZs on the regional economy, we implemented the following two-step methodology: . We constructed CIs by means of principal component analysis (PCA) to evaluate the benefits achieved in each region on implementing the industrial policy.
These CIs summarize a complex set of information (impacts) in a single useful index. In this research, we chose to build two different indexes, both with a view to analyse the different zone types: the first was constructed by selecting for each IZ (SEZ, FZ or FTZ) some variables according to the main objectives stated for that specific special area; and the second index allows us to consider the impact obtained through a broader vision which consists in evaluating spillover effects, such as those described by the economic literature. In so doing, this analysis has the potential to offer preliminary indications to policymakers concerning the strengths and weaknesses of their policies. Moreover, by using panel model analysis, it is possible to test the goodness of CIs and compare them with disaggregated indicators. Finally, as an additional robustness check, a forecasting exercise is implemented to test the strength of the hypothesized effect of IZs on regional gross domestic product (GRDP) growth. . We then performed counterfactual analysis to verify the capacity of public policy to change the conditions of a given target population in a desired direction by making comparisons with other regions that do not benefit from IZs.

Sample
Data on active zones were collected from several sources to avoid potential issues arising from the lack of a rigorous and detailed register of the number and location (Bost, 2019). In all, the EU has 92 special zones: according to Farole and Akinci (2011) Table B2 in Appendix B in the supplemental data online). The sample is clustered according to the type of development plan (Bost, 2019; European Commission, 2019) by building three subsamples: . SEZs are special areas within the customs territory of the Community that offer a vast of incentives including infrastructure, tax and customs exemptions and simpler administrative procedures. . FZs are areas with a fixed perimeter with entry and exit points that are subject to customs supervisions, aimed at boosting the activities with foreign markets by providing differential regulation for overseas companies, enabling them to operate with economic incentives and free-trade conditions. . FTZs are fenced-in, duty-free areas or warehouses, located along the main transportation axes such as seaports and major airports, along the development corridors, or in border regions, that provide logistics facilities aimed at trade facilitation in globalization.

Data
The construction of each CI requires detection of the main variables able to describe the phenomenon in question. The first composite index, namely CI_Stated (CI_S), is built according to the main objectives as stated by the administration during its implementation (see Table B2 in Appendix B in the supplemental data online). Starting from in-depth research for each IZ, a set of main goals was defined for each cluster as follows: . SEZs: R&D, gross fixed capital formation, firms, employment, motorways, railways, foreign direct investment (FDI) and GRDP. . FTZs: firms, employment, marine transport, air freight transport, exports and GRDP. . FZs: gross fixed capital formation, firms, employment, marine transport, exports, FDI and GRDP.
In order to build the second composite index, CI_Total (CI_T), we take into account the main determinants of successful IZs, as defined by the above literature review. Thus, the model considers the six previously defined macro areas in order to select several variables for each topic, chosen from those most widely recognized and for which data are available ( Table 1). The variables used refer to data provided by the Statistical Office of the European Union (Eurostat) over the period 2006-18.
. Research and development reflects the improvements in innovative activities, which is one of the objectives of implementing an SEZ (Nazarczuk & Umiński, 2018). Thus, trade mark applications, R&D expenditure, and R&D personnel and researchers were included in order to assess whether they are effective at producing innovation (OECD, 2017;Nazarczuk & Umiński, 2018). . Productive structure variables, such as gross fixed capital formation and local units, are likely to attract investments and subsequently accelerate regional development through the implementation of SEZs (Maslikhina, 2016;. . Infrastructural endowments play a prominent role since better infrastructure has a huge impact on regions. Thus, we include air freight transport, maritime freight transport, motorways and railway lines (Sosnovskikh, 2017;. . Environmental variables are used by several studies. We use air pollution as a proxy for environmental factors in order to assess the environmental impacts for selected SEZs (Jauch, 2002). . With respect to contextual factors, the role of the state is considered a major determinant with its regulatory and institutional framework: weak governance structures or the coexistence of too many institutions can lead to IZ failure (Cheng, 2019;UNCTAD, 2019 IZs, in turn, can have a positive impact on government budgets, whose main gains come from both taxation on stakeholders' personal income and from corporate income tax (Lion, 2014). Accordingly, we introduce the European Quality of Government Index (EQI) to measure institutional quality at regional level and thus assess its role in influencing IZs growth . The index in question is the only measure of institutional quality available at regional level in the EU. Institutional quality is defined as a multidimensional concept consisting of high impartiality and quality of public service delivery, along with low corruption. . Economic background includes exports, FDI and employment. According to the literature, the increase in foreign currency earnings is one of the main benefits expected from IZs and is directly related to the level of exports. Furthermore, the implementation of IZs may lead to an increase in annual FDI (Pradhan & Zohair, 2015;Nazarczuk & Umiński, 2018). Employment is widely recognized by the literature to increase the competitiveness of IZs in attracting investors, which is largely based on the productivity of their workforce (Aggarwal, 2007;Caraway, 2007) (Table 1).
Having defined the variables, one of the main problems in comparing different regions is their initial value. Undoubtedly, the latter is linked to the economy of the country where the IZ is established, and this could lead to a biased evaluation when we compare different IZs from different countries. To offset this problem and to allow a comparison among them, to build the sample we consider the deviation of each indicator from the average value of the European sub-economic area.
In order to compare their impact among the zones, setting the local economic development policy j ¼ 3 (the three sub-economic areas) and the regions of each subeconomic area i ¼ 1, 2, 3, … , n, the indicators x i,j (labelled in Table 1) are constructed as a ratio of region i with the average value of the same indicator for all the n -1 regions belonging to that sub-area, as described in equation (1): We can thus mitigate the problem of the initial conditions and compare results of similar industrial policies in more homogeneous areas. Subsequently, for measuring the evolution of the phenomenon, comparisons were made between the intensities of the CIs at different times (i.e., calculating the variations in intensity from one period to another). The periods analysed are as follows: 2006-10, 2011-14 and 2015-18. For each sub-period, the methodological approach was applied by creating different CIs; the 2006-18 value is the average obtained from the findings of the three sub-periods.

The CI
In recent decades, in order to respond to the ever-increasing need for systematic information on complex realities, attempts have augmented to create synthetic and complex indicators, in many areas of knowledge, capable of integrating a large amount of information in formats that are easily understood. CIs are useful tools for policymaking in conveying information on the performances of countries and regions in fields such as environment, economy, society or technological development (OECD, 2008). For instance, Bandura (2006) cites more than 160 CIs, and this number has been growing year after year. In general, the synthesis has the advantage of avoiding the presentation and interpretation of a large number of elementary indicators in order to perform simpler and faster analyses, especially in comparative terms. A CI is, therefore, the mathematical combination of individual indicators that represent different dimensions of a concept whose description is the objective of the analysis (Saisana & Tarantola, 2002). According to this approach, indicators are not substitutable (ignoring an indicator is omitting part of the construct). Hence, the measurement model could not be influenced by the correlations between indicators. In fact, internal consistency is of marginal importance because two uncorrelated indicators can both serve as meaningful indicators of the same construct (Maggino, 2017).
The ongoing debate and criticism on the statistical validity of CIs is expressed by an extensive literature, which is worth investigating briefly, to strengthen, or not, the rationale of the empirical applications.
Many stages are involved in the construction process of a composite index, and criticisms could simultaneously grow with reference to each of them.
What a CI is able to measure largely depends on the choices made in this process. This requires a robustness analysis of the CI defined, to verify the dependence on the choices made. Robustness verification allows dispelling controversies surrounding CIs and their construction. Therefore, this study should evaluate the robustness and sensitivity of the vulnerability CIs aggregated by the conventional additive and multiplicative utility functions from the vulnerability assessment components (Choi, 2019). Among others, the literature on the topic has stressed the relatively high subjectivity present in many phases of the construction process (Booysen, 2002) and the lack of a transparent methodological procedure and reasonable justification, which might leave room for outcome manipulation ( Assessing the impact of special economic zones on regional growth through a comparison among EU countries Overall, criticisms can be summarized in two different groups, referring to: (1) the methodological approach underlying CI construction; and subsequently, (2) its empirical application.
Common problems associated with methodology deal with (Bräutigam & Xiaoyang, 2019;Greco et al., 2019): the definition and adoption of assumptions, in order to make choices concerning the model for estimating the measurement error in the data; the mechanism for including or excluding indicators in the index; the process of transforming the data used in the indicator (choice of procedures for the treatment of missing data, for the standardization and normalization of data); the method for the attribution of weights; and finally, the method of aggregation.
Therefore, the main consequences of these criticisms might be linked to the subsequent application into empirical analyses (McKenna & Heaney, 2020). Indeed, according to the OECD (2008), if poorly constructed or misinterpreted, CI might: (1) send misleading policy messages; (2) invite simplistic policy conclusions; and (3) be misused, for example, to support the desired policy, especially if the construction process is not transparent and/or lacks sound statistical or conceptual principles. Moreover, improper selection of indicators and weights could be subject to political dispute and disguise severe failings in some dimensions, increasing the difficulty of identifying proper remedial action if the construction process is not transparent. This would result in inappropriate policies if dimensions of performance that are difficult to measure are ignored (Kuc-Czarnecka et al., 2020). Keeping all these criticisms in mind enhances both transparency and soundness of the process. Our research would refer to the OECD's (2008, p. 15) guidelines in CI construction.
Nevertheless, we would like to stress that CIs should never be comprehended as a goal per se, given the presence of large limitations that are related to their construction, but to help in evaluating the phenomenon in order to give an interpretation inclusive of several related aspects.
With the aim of constructing composite indices, the PCA was used to evaluate the degree of correlation between the different parameters of the seven selected macro-economic indicators. This method is applied when there is the need to clearly define the relationships between different basic indicators that could be highly correlated. Hence, the broad set of information can be compressed into a single useful support index by: . reducing confusion generated by the different pieces of information gathered (Callens & Tyteca, 1999); . explaining the information more simply and concisely (Jollands et al., 2004); . formalizing the aggregation process of the starting parameters.
This analytical approach is based on factorial analysis in order to highlight the correlations between the main macro areas of the indicators considered. The aim was to summarize the information contained in the relationships among the parameters, thus simplifying it. The methodology applied to construct the CI started from the choice and analysis of initial parameters. In so doing, we used 13 of the 15 selected variables since the pillars 'Environmental' and 'Contextual factor' gather only one indicator each. Moreover, we applied factorial analysis to identify common factors and then calculated the weights. Four different factorial analyses were implemented (one per macro area) so as to obtain a single factor, common to each macro area. Figure 1 shows, starting from the basic variables, the different levels of aggregation we used to construct the composite index.
Through this procedure, the weights to be associated to each variable were obtained as the variance of those components, together explaining more than 50% of cumulative variance. Given the composition of our sample, we consider this value a sufficient threshold for the construction of our composite index. Thus, the coefficient of the derived component of the factors was used as a weight and each variable was then multiplied by its weight.
Following the above procedure, four new variables were acquired. Together with EQI and air pollution, the variables were used to perform another PCA. We calculated the weights to attribute to the basic indicators and finally defined the CI through the aggregation of basic indicators. 1 The conditions of applicability of the PCA were then evaluated by carrying out the most common and most indicative preliminary tests. For the basic index I1, the calculation of the Kaiser-Meyer-Olkin index was 0.672, while Bartlett's spherical test resulted in a high level of significance (0.000) and a very high Chi-square value (1261.87). Similar results to those of Bartlett's spherical test were obtained for the other basic indices too. The result of such tests confirms the possibility of proceeding to a subsequent analysis of the factors.
In order to facilitate data interpretation, we adopted a normalization process. Among the main procedures, we chose min-max normalization, since we were able to obtain scores between 0 and 1, comparable with others: This procedure allowed us to rank IZs according to the normalized score. The same procedure was then carried out considering the two CIs: stated objectives and total objectives. The explanatory power of CIs is reported in Appendix A in the supplemental data online.

Interpretation of the CIs
The results for the first step of the analysis are presented in Tables 1-3. The disaggregated values of each indicator, reported as the variation among the three periods considered, are presented for each IZ in Tables B6-B8 in Appendix B in the supplemental data online. As emerges clearly from Table 2 (see also Table B6 online), the analysis of the FZs shows there are large differentials among the best performers (the top three in the ranking, belonging to France, Spanish and Estonia) and the other FZs established in Croatia and Lithuania.
The FZ of Le Verdon in Aquitaine represents a best practice since it obtains the greatest effects from the industrial policy both regarding the stated target (A) and product spillovers (B) (achieving a score of 1 for both indexes). Matching high values for both indices show achievement of the goals set by the administration during policy planning. Moreover, the success of this French FZ is confirmed by the constant and increasing positive variation between the periods analysed. Besides substantially expanding the transport infrastructure endowment, the policy of attracting FDI represents a winning strategy. The Le Verdon terminal is specialized in transporting containers and heavy packages and allows the stopover of cruises. Its best practice is based on no customs duties, excise taxes or VAT being applicable to the goods declared as transit and stored within the FZ for three years (see Table B6 in Appendix B in the supplemental data online).
Different findings are obtained from the other FZs. Good results concerning the stated objectives may also be confirmed from Estonia (Eesti), in which three FZs are located: the northern port of Paldiski, Muuga Arbour and the port of Sillamae (2006-18 CD_S score ¼ 0.76). The region presents a high GRDP growth rate and good transport endowment. The incentive scheme involves no customs duties, excise taxes or value added tax (VAT) applicable on goods declared as transit and stored within the FZ for three years.
The Spanish regions (Canarias and Cantabria) achieve good levels of FDI (respectively, 2006-18 CD_S scores ¼ 0.80 and 0.72). The port of Santa Cruz de Tenerife in Canarias offers full exemption from tariffs and direct and indirect taxation on imported goods. Streamlined procedures ensure that businesses encounter a bare minimum of red tape. Additionally, plant operators installed in the FTZ may also benefit from other fiscal instruments in the Canary Islands such as a lower corporate income tax rate of only 4%. Businesses need only generate five full-    Assessing the impact of special economic zones on regional growth through a comparison among EU countries time jobs in order to benefit from the exceptionally advantageous fiscal regime. In Cantabria, all the advantages are applicable to the companies established in them. This regime is aimed at stimulating exports, by allowing imports of goods from a third country, with total exemption from import duties and/or trade policy measures, provided that the products obtained in their transformation (where the imported goods are used as intermediary goods) are then re-exported outside the customs territory of the Community. No good performance is identifiable from the activity of the other regions. Table 3 presents the findings for the FTZs in Germany and Spain. Unlike the FZs, the FTZs situated in Germany show good results in relation to the stated objectives. Freihafen Hamburg (2006-18 CD_S score ¼ 1) has undergone an industrialization process based on both the localization of a large number of firms and corresponding capitalization due to FDI, which has increased steadily during the whole period in question (see Table B7 in Appendix B in the supplemental data online). Export values are high: under the incentive scheme, goods within the EU can be stored, processed and traded free of customs duties. This FTZ has become one of the busiest ports in the word for container traffic. Despite this great success the total CI shows low values of development for the surrounding area (2006-18 CI_T score ¼ 0.35).
The same characteristics are encountered by other FTZs in other German regions, albeit with less impressive results: Freihafen Deggendorf in Niederbayern involves ships, railways and road transport (2006-18 CD_S score Finally, Table 4 shows the performances of SEZs in 16 Polish regions. The Katowice SEZ in Slaskie records a 2006-18 CD_S score of 1, higher than the rest of the SEZs, and a 2006-18 CI_T score of 0.92. In presenting consistent data both regarding the stated and total objectives, Slaskie confirms the complete success of the industrial policy in question. The presence of high transport infrastructure endowment, good quality of institutions, good levels of exports, FDI and GRDP make Slaskie one of Europe's best special areas (see Table B8 in Appendix B in the supplemental data online). It was recognized as the best FZ in Europe in 2015, 2016, 2017 and 2019. In 2019, the Katowice SEZ was recognized as the second best special zone in the world. Like most of Poland's SEZs, this incentive scheme involves exemption from income tax, tax relief on real estate, haulage vehicles and custom duties, tax incentives for hiring new staff, and related to investment procedures overall.
For three other regions in Poland the 2006-18 CD_S score lies between 0.64 and 0.56 (Dolnoslaskie, Malopolskie and Lodzkie). The first region covers three zones, namely Kamienna Gora special economic zone for small entrepreneurship (SEZSE), Walbrzych SEZ and Legnica SEZ, which specialize in the automotive industry, besides related industries such as metal and plastic processing, as well as logistics and warehouse services for the above sector. The location of Krakow SEZ in Malopolskie, which allows it to be perfectly connected with the rest of Poland and Europe, has made global companies decide to invest here. Malopolskie received, among other titles, that of the European Region of Entrepreneurship and took first place in the international competition RegioStars 2016.
The Łódź SEZ in Lodzkie has built a nationwide innovation centre and created a local start-up support system. Thanks to this, every current and new investor in the SEZ has direct access to modern technologies; the Łódź SEZ provides support in the form of corporate income tax (CIT) or personal income tax (PIT) exemptions for individuals conducting business there.
All SEZs in Poland benefit from income tax exemption, tax relief on real estate, haulage vehicles and custom duties, tax incentives for hiring new staff, and related to investment procedures and exemption from property tax. The tax exemption has a maximum that varies according to the voivodship. By considering the target over time, there is a puzzling picture among the special areas, probably characterized by the economic impact of the two crises (2008 financial crisis and 2010-11 European sovereign debt crisis) as well as by the resilience of the considered country.
Looking at the dynamic of each single IZ, the FTZs show increasing values over time for the regions belonging to Germany, while for the Spanish regions constant values between the first and second periods persist, while they definitely grow in the third period. These results could be explained by the reaction of these countries to the economic crises, how they reacted during the central years and, starting from 2015, how they started having a recovery.
The situation for free trade is not as clear as in the previous IZ. Except for regions that have been labelled as best practices (Eesti in Estonia; Canarias in Spain; Aquitaine in France) and show increasing values compared with the targets considered throughout the period, the other regions offset the previous growth trend during the crises.
Finally, the situation for the SEZ in which only the Polish regions are presented is once more different. In fact, they show values mostly increasing in the medium term, which corresponds to the years of crisis. This situation could be explained by the quite strong Polish economic growth, which has suffered little from the public debt crisis, because although it has been a member of the EU since 1 May 2004, Poland currently continues to use its national currency. Poland has a high-quality education system and sound public finances: debt below 50% of GDP and a deficit of around 1%, which, with those growth rates, is practically zero. This allows for a mass of public investments which, added to European funds, reach 4-5% of GDP every year. And it explains the great steps forward that Poland is making. Moreover, even when it was a Communist country, Poland already had a widespread and diversified industrial base which, with its entry into the EU, was strengthened thanks to the enormous contribution of investments from abroad, not only from the EU, which affected up to 10% of GDP and, above all, they have brought new technologies and renewed the existing industrial base.

Statistical robustness check of the CIs
We have shown above the different phases in the construction of the CI that required choices, such as the selection of indicators, the choice of the method for relativizing the data, the imputation of missing values and the attribution of weights. What the CI will be able to measure largely depends on these choices. This requires an analysis of the robustness of the CI defined in order to verify the level of dependence on the choices made.
The verification of the robustness is required whenever the model adopted (in our case, the construction model of a CI) is uncertain or when there is a possibility of obtaining an incorrect result even in the presence of correct application. The verification of the robustness of the built CI allows the dispelling some of the controversies surrounding the CIs and their construction.
This literature has debated the relatively high subjectivity present in many phases of the construction process of CIs, in particular, the phases that require the definition and adoption of assumptions in order to make choices concerning the model for estimating the measurement error in the data. The mechanism for including or excluding indicators in the index; the process of transforming the data used in the indicator (choice of procedures for the treatment of missing data, for the standardization and normalization of data); the method for the attribution of weights; and finally, the method of aggregation. Assessing the impact of special economic zones on regional growth through a comparison among EU countries There is no doubt that the assumptions adopted and the decisions taken in these phases can heavily influence both the information that is included in the construction of the indicator and, consequently, the information transmitted by the constructed indicator. The subjectivity required calls into question both the robustness of the indicator and the message transmitted by the indicator and requires a particular and delicate analysis.
In order to evaluate the robustness of the CI, increase and improve its transparency, it is possible to apply one of the two different analysis methodologies: . Uncertainty analysis, which analyses how uncertainty in the input factors spreads and is transmitted in the structure of the CI and affects the values it produces. In other words, it measures how much a given CI depends on the information that composes it. . Sensitivity analysis, which divides the total variance of the output produced by evaluating the individual contribution of all potential sources of uncertainty. In theory, all potential sources of uncertainty should be considered.
The two approaches are almost always treated in separate contexts and in the context of the construction of CIs (Jamison & Sandbu, 2001). The use of the two analysis approaches after the development of a CI proved to be more useful and powerful Tarantola et al., 2002).
The robustness of the results of the CI is evaluated by verifying and comparing the possible different performances that would have been obtained by taking different decisions at different points of the analysis.
From a logical point of view, the procedure is equivalent to the definition for each unit of different scenarios in the production of the CI value . Each of the scenarios corresponds to the value that the CI would have had if different choices had been made at different moments in the construction of the indicator. Subsequently, the goal is to identify how sensitive the result (or model) is to changes in the various parameters that define it. A small change reveals low sensitivity of the parameter. This approach is part of the broader sector of what-if analysis of a model ('What would happen to the result if we inserted a particular change to a parameter?').
Judgments are an important part of constructing CIs, for example, on the selection of indicators, data normalization, weighting and aggregation methods, etc. The robustness of the CIs depends on the underlying assumptions and methods chosen. By using a sensitivity analysis, the robustness of the CI_S and CI_T (for each IZ) can be evaluated and transparency can be improved.
The sensitivity analysis results are generally presented in terms of the sensitivity measure for any input source of uncertainty. We implement a simulation-based sensitivity analysis to explore how robust this result is to potential confounding from omitted variables (Franks et  The analysis estimates three quantities to help researchers assess the internal validity of their results. The first quantity, RVq ¼ 1, represents the amount of residual variance in both the treatment and the outcome that an omitted variable would need to explain in order to change the sign of the correlation between the variables used. The second quantity, RVq ¼ 1, α ¼ 0.05, represents the amount of residual variance in both the treatment and the outcome that an omitted variable would need to explain in order to nullify our point estimates at the conventional level of statistical significance. The third quantity, the partial R 2 of the treatment with the outcome, measures the percentage of variation of the outcome explained by the treatment, after taking into account the part explained by the remaining covariates. This quantity matters for sensitivity analysis because the amount of outcome variation explained by the treatment determines how strong unobserved confounders would need to be to explain away the observed effect. Tables 5-7 present the results of the sensitive analysis of the two indices and the IZ, respectively. The CD_S of the SEZ shows a partial R 2 of the treatment with the outcome (R 2 Y D|X ) and is 17.72%. This means that if confounders explained 100% of the residual variance of the outcome, they would need to explain at least 17.72% of the residual variance of the treatment to fully account for the observed estimated effect. The robustness value for the point estimate describes the minimum strength of association (measured in terms of partial R 2 ) that unobserved confounders would need to have, with both the treatment and the outcome, to bring the effect estimate down to exactly zero. The robustness value for the point estimate (RV q=1 ) is 36.87%. This means that unobserved confounders would need to explain at least 36.87% of the residual variance both of the treatment and of the outcome to bring your estimated effect to 0. Finally, the robustness value for the t-value (RV q=1, a=0.05 ) describes the minimum strength of association (measured in terms of partial R 2 ) that brings the point estimate not to exactly zero but rather within a range where it is no longer 'statistically different' from zero. The robustness value of the t-value is 27.74%, and means that unobserved confounders would need to explain at least 27.74% of the residual variance of both the treatment and the outcome for the null hypothesis that the true treatment effect is equal to zero to not be rejected at the significance level of 0.05.
The CI_T of the SEZ shows a partial R 2 of the treatment with the outcome and is 12.17%; the robustness value for the point estimate is 30.94%. Hence, unobserved confounders would need to explain at least 30.94% of the residual variance for both the treatment and the outcome to bring the estimated effect to zero; and the robustness value for the t-value is 20.86%. This means that unobserved confounders would need to explain at least 20.86% of the residual variance of both the treatment and the outcome for the null hypothesis that the true treatment effect is equal to zero to not be rejected at the significance level of 0.05.
We also show the regression results augmented with sensitivity statistics in Tables 6 and 7 for FTZ and FZ, and observe that they follow similar paths along with the results obtained for the SEZ. The robustness value for the point estimate around 35%, the partial R 2 of the treatment with the outcome around 16% and the robustness value for the t-value around 23%. For all three IZs, the estimated effects are positive and statistically significant. Recall that these reasonably large values such as these indicate only that large confounders would be required to alter our conclusions and they say nothing about whether such confounders exist. However, we believe that these quantities summarize what we need to know in order to safely rule out confounders that are considered to be problematic. Researchers can then argue about whether they fall within plausible bounds on the maximum explanatory power that unobserved confounders could have in a given application. Where investigators are unable to offer strong arguments limiting the absolute strength of confounding, it can be productive to consider relative statements, for instance, by arguing that unobserved confounders are likely not multiple times stronger than a certain observed covariate. In our application, this is indeed the case. One could argue that, given the composite and complex nature of the indices, it is hard to imagine that unobserved confounding could explain much more of the residual variance of targeting than what is explained by the observed variable.

Counterfactual analysis
In order to provide a comprehensive assessment of industrial policy effectiveness in the regions selected, a comparison with non-IZ regions is of fundamental importance. To meet this objective, we applied counterfactual impact evaluation.
Where applicable, this type of evaluation makes it possible to understand whether the policy in question has actually produced the desired effects and, if so, to quantify them, isolating the changes resulting from the policy, compared with the overall changes observed. This approach allows us to gauge the net impact, that is, without the effects due to other factors external to the policy (changes in the socio-economic context, social and/or economic dynamics, etc.).
Thus, three different control groups belonging to the seven countries selected were identified: . Group 1: includes the other regions in which there are no SEZs belonging to the seven member states considered. The sample consists of 74 NUTS-2 (DE ¼ 31; Assessing the impact of special economic zones on regional growth through a comparison among EU countries . Group 2: starting from the sample of group 1, group 2 was constructed by considering only the remaining regions in which there are no SEZs that border SEZ regions. The sample consists of 23 NUTS-2 (DE ¼ 6; EE ¼ 1 ¼ ES ¼ 8; FR ¼ 3; HR ¼ 2; LT ¼ 2; PL ¼ 1). . Group 3: starting from the sample of group 1, group 3 was constructed by considering the other regions with the same low levels of GRDP growth as the SEZ regions. In order to build such a sample, we performed quartile analysis and chose the regions belonging to the first and second quartiles with the lowest values; the sample consists of 28 NUTS-2 (DE ¼ 6; ES ¼ 9; FR ¼ 13).
Depending on the characteristics of the policy, the appropriate evaluation strategy is thus defined. Hence, among different models proposed in the literature, propensity score matching (PSM) is a model that can implement a variety of matching methods to adjust for pre-treatment observable differences between treated and untreated groups. Since there has been little research at European level, with few studies mainly based on firm samples (IZ versus non-IZ firms) and focusing on specific case studies (Huang et al., 2017;Nazarczuk & Umiński, 2017), the initial problem was to define a proper sample for the analysis (Ambroziak & Hartwell, 2018). The first step in PSM analysis is to estimate the propensity score. Normally, a logit or probit function is used for this purpose, given that treatment is typically dichotomous (e.g., let T i be a binary variable which takes the value of 1 for region i having access to the treatment (SEZs), and 0 for non-treated regions). The effect of SEZs on the regional growth rate is estimated using the PSM procedure. 3 We present below the results pertaining to estimation of propensity scores, average treatment effect on the treated (ATT) and bootstrap analyses.
The first step in implementing counterfactual analysis using PSM is to determine the conditional probability of regional special zones improving regional GRDP growth. We perform this estimate by using a logistic regression model. The model considered all the observable covariates 4 that affect regional participation in SEZ and non-SEZ regions and for which observational data were available. These factors are responsible for regional differential participation in improved GRDP growth. Since we are interested in computing the propensity scores, which will be used later in the matching process, we will not go into the details of why and how each of the covariates affected regional participation in the intervention. Nevertheless, we indicate that as we proceed with our analysis, these before-matching differences are no longer significant in the aftermath of matching, which is an indication that the PSM was successful in experimentally creating a comparable group of control individuals whose outcomes can be compared with those of the treated ( Table 8).
Estimation of the ATT is performed using nearest neighbour matching (with five-nearest neighbour matching algorithms) ( Table 6). The estimates are robust to this type of matching considered and support the balancing test. Based on the findings, we note the existence of a statistically significant difference between treated (n ¼ 320) and control (n ¼ 760) in group 1; treated (n ¼ 330) and control (n ¼ 240) in group 2; and treated (n ¼ 124) and control (n ¼ 228) in group 3 for the whole sample.
In addition to the mean values of the outcome variable, Table 6 contains mean differences between treated, control groups (columns 3 and 4), and bootstrap standard errors (with 500 replications) on the mean difference (column 8). Overall, we found convergence of results between the two matching algorithms (column 9). The outcomes in this section are based on the results obtained using the five-nearest neighbour matching algorithm as this resulted in a higher level of statistical significance. Accordingly, the results show a statistically significant gain in regional GRDP growth as a result of improving IZ policies in the study areas. More specifically, we found that the three groups that implemented specific industrial policies (see SEZ areas) achieved, on average, a higher increase in GRDP growth compared with those which did not benefit from such policies. The difference between treated and control groups is greater in group 1 and less pronounced in groups 2 and 3. The greater differences shown in group 1 could be explained through the heterogeneity of the regions involved in this group. The difference is less pronounced in the second group, that is, in the areas bordering SEZ regions. This could confirm, indirectly, the existence of some spillover effects as widely recognized in the literature (Jarosiński & Maśloch, 2016;. Overall, this is a significant result, indicating that the adoption of special economic areas in these regions resulted in improving the wealth of the regions concerned. The implementation of special economic areas triggered the process of regionalism through the new industrial clusters that are able to promote economic cooperation between different institutional actors and increase national competitiveness on the global scene, through a multi-centric network in which the public sector joins forces with private initiatives (the well-known top-down and bottom-up approach). In line with these results, the authorities should consider the differences between IZ-based and non-IZ-based regions (which mean, respectively, enterprises and local governments which benefit from large incentives and those which do not) and be able to reduce the gaps in the regional growth rates.

DISCUSSION
In order to attract investment, foster economic growth and job creation, many countries have experienced the creation of 'zones of advantage' for the establishment of new businesses. Worldwide there are currently about 5400 IZs, that is, areas which enjoy not only tax benefits but also financial, infrastructural and logistic support, besides being subject to special legislation and procedures that may differ from other parts of the same country.
As we showed in the previous section, the ability to establish a successful IZ depends on long-term industrial policy objectives, aligning policy instruments such as IZs with the corresponding stage of industrialization of the country, and on the degree of complementarity of IZs with other industrial policy instruments, as occurs in Aquitaine (France) and Slaskie (Poland). In the latter regions there is strategic importance in building productive capacity for exports, enhancing manufacturing capabilities as a leading long-term development goal. This capability is confirmed by a continuous development of the areas in various stages ('early', 'middle' and 'late') of industrialization. Large differentials among the IZs depend on the lack of these requirements. Moreover, the magnitude of the impacts of this industrial policy can be better understood with regard to the fiscal regulation of special areas.
The advantages conferred, namely tax benefits, financial, infrastructural and logistical support measures, as well as bureaucratic simplification, come under the customs framework of the EU. At a detailed level, generalized fiscal support supplied by each country can be adapted to each special area on the basis of the strengths and weaknesses identified, as happens in Spain. Indeed, each Spanish zone has different levels of tax relief and other related benefits, according to its specific needs. However, this policy instrument remains a solid support for the development of areas with a lag in economic growth. Consistent with the findings, authorities should consider the differences between the IZs regions and those not based on the IZ and try to reduce the gaps in their regional growth rates.

CONCLUSIONS
Despite the widespread interest in IZ implementation as a useful industrial policy tool to support regional economic growth, current studies do not provide complete understanding of the impacts of such zones for several reasons. Indeed, each type of zone (SEZ, FZ, FTZ) presents specific features which make the impact on the area different (Chen, 2019). There are few detailed accounts of IZs numbers and locations (Bost, 2019). The economic literature tends to follow a country-wide case approach . Finally, studies providing the necessary counterfactual analyses remain scarce (UNCTAD, 2019). Policymakers have thus been provided with little or no solid support.
In order to tackle such issues, we supplied methodology to evaluate the impacts of implementing 51 IZs upon 35 European regional economies. Several problems were addressed, comprising the definition of the sample, the period to be investigated and the indicators to analyse. The choice of European IZs, activated from 1995 to 2005, allowed us to assess the impacts (drawn from the main evidence provided by the literature) of the zones in question 12 years after being set up (until 2018). By both building two CI indexes and using a counterfactual approach, this research yielded insights into the impact of the industrial policy implemented through the creation of IZs besides comparing European regions with IZs and those without.
Our models revealed that the composite index is a good instrument for assessing the main impact of IZs. Through joint evaluation of the CIs it was possible to highlight the best results achieved mainly by the FZ of Le Verdon in Aquitaine and the Katowice SEZ in Slaskie region (Poland) (recognized as the best FZ in Europe in 2015, 2016, 2017 and 2019). These regions are the only two cases that present the highest value for both CIs besides showing constant values over time.
By considering the reached target over time, the findings show also that FTZs and FZs need more time to benefit from their implemented policies (third period) while less time is necessary for the SEZs to reach their targets (second period).
By comparing the findings among the various types of IZs, it emerges that FZs show large differentials among the best performers (the top three in the ranking, belonging to France, Spanish and Estonia) and the other FZs set up in Croatia and Lithuania. Good findings of the FTZs emerged only in Germany, albeit limited to the established objectives (the Spanish regions did not perform well). Finally, the SEZs, except for that mentioned above, achieved the least success. With respect to the counterfactual analyses, the findings show that even if, as expected, growth is constant in each region analysed, the level