Measuring the Impact of High-Speed Rail on Economic Performance: Evidence for the Madrid-Barcelona corridor

Transcription

1 Measuring the Impact of High-Speed Rail on Economic Performance: Evidence for the Madrid-Barcelona corridor Daniel J. Graham, Ruben Brage-Ardao, Patricia C. Melo Centre for Transport Studies Department of Civil and Environmental Engineering Imperial College London London SW7 2AZ ABSTRACT During the period between 2000 and 2010 Spain carried out the largest high-speed rail construction programme in Europe. As a result, in 2011 the Spanish HSR network became the largest in Europe, exceeding France and Germany. By 2020, it is planned that 90% of the country population will live within a 50km radius to the nearest high-speed rail station. Investments in high-speed rail projects are frequently justified on the basis of alleged positive effects on regional and national economic growth. Nevertheless, many of these benefits are unclear and depend on a case to case evaluation. The debate about the real economic impact of high-speed in Spain has not fully arisen until the current economic recession. In this research we measure the economic impact of the Madrid-Barcelona corridor. We have carried out an ex-post analysis to assess how the high-speed rail corridor has influenced economic output at the regional level. Our results show that the Madrid-Barcelona corridor has not produced any significant positive effects on the output growth of the Spanish provinces with access to the high-speed corridor. Keywords: Difference-in-differences, high-speed rail, Spain

2 1. INTRODUCTION 2 Investments in major transport infrastructure projects are frequently justified, or even planned, on the basis of presumed positive transformational effects on the spatial economy. The extent to which such effects actually materialise has been the subject of a large academic literature spanning several decades (Munnell, 1992, Gramlich, 1994, Rietveld, 1994, Boarnet, 1997, Button, 1998, Banister and Berechman, 2000). In the policy domain, this issue now features prominently in contemporary debates over the benefits of high-speed rail investment. In the UK, for instance, supporters of a proposed highspeed rail link from London to the North of England have argued that economic growth and job creation will be stimulated along the route of alignment and that the investment will help rebalance the national economy by reducing wealth inequalities (e.g. Anselin, 1988). Similarly, debate over the merits of the California high-speed line have revolved around arguments emphasising the potential for regional economic growth, benefits from agglomeration economies, and other positive spatialeconomic effects (e.g. Angrist and Krueger, 2001). High-speed rail (HSR) investments are worthy of attention, not only because they are topical, but also because they offer a valuable opportunity to study the economic impacts that may arise via improvements in the long-distance connectivity of an economy. These are typically large investments, both in terms of physical capacity and financial capital, and it is therefore worth considering whether the investment has delivered positive impacts to the economic performance of the local economies receiving the investment. There are several reasons why investment in high-speed rail can make a positive contribution to economic performance. First, high-speed rail can link distant markets and foster agglomeration economies, hence stimulating competition and increasing productivity (e.g. Graham, 2007). This enlargement in market scope can produce a wide range of benefits to firms: (i) improved access to a larger pool of skilled workers, made possible through improvements in the efficiency of commuting flows; (ii) increased communication and sharing of ideas between firms and workers (knowledge spillovers); (iii) better access to customers and suppliers, including specialized professional services (input-output linkages) (e.g. Duranton and Puga, 2004, Rosenthal and Strange, 2004). During the period between 2000 and 2010 Spain has carried out the largest high-speed rail construction programme in Europe. During this decade four new corridors opened between 2005 and 2010, in

3 3 addition to the Madrid-Seville HSR line which had opened in As a result, in 2011 the Spanish HSR network became the largest in Europe and the highest rate of HSR km per capita in the world (UIC, ). Investments in high-speed rail in Spain have been justified by alleged positive effects on regional and national economic growth. However, many of these benefits remain unclear and have yet to be confirmed. In this research we have carried out an ex-post analysis to quantify how the Madrid-Barcelona (or north-east) high-speed rail corridor has influenced economic output of the Spanish provinces with access to the high-speed rail network. We develop a causal analysis based on a difference-indifferences (DID) method, a causal inference technique commonly used to assess the impact of a given innovation or policy change (called treatment) on a given set of outcomes. To implement the DID method data are needed for two groups - treated and untreated - and two periods - before and after the treatment. The treated group consists of the individuals who are exposed to the specific treatment (here a high-speed rail connection), while the untreated group refers to the individuals that do not receive the treatment. The DID estimator consists of taking the difference in the average outcome in the treated group before and after the treatment minus the difference in the average outcome in the untreated group before and after the treatment. In order to avoid the violation of the parallel trend assumption between the treated and the untreated groups we also use data for Portuguese regions, as these should not have been affected by the Madrid-Barcelona high-speed corridor. Our results indicate that the Madrid-Barcelona high-speed rail corridor has not produced any positive transformational effects on the economic growth of the provinces receiving the high-speed corridor. Differences in the economic performance between these regions and those that did not receive a highspeed rail investment have not been altered by the investment, which suggests that the alleged positive effects proclaimed by the Spanish government cannot be confirmed empirically for the Madrid- Barcelona high-speed rail corridor. The structure of the paper is as follows. Section 2 provides a description of high-speed rail in Spain. Section 3 explains the methodology used based on DID models, while Section 4 describes the data used to estimate the DID models. Section 5 presents and discusses the main results, and Section 6 draws together the main conclusions.

4 4 2. BACKGROUND TO HSR IN SPAIN High-speed rail can be an ambiguous concept as there are several definitions of HSR for Europe, the USA and Asia. According to the European definition (Arellano, 2003), HSR comprises new lines designed for speeds above 250 km/h or upgraded lines for speeds above 200 km/h. In our case, we focus in the new lines which in the Spanish case are clearly distinguishable because they have an international rail gauge instead of the Iberian gauge. In Spain, HSR started with the New Rail Access to Andalucia plan (Nuevo Acceso Ferroviario a Andalucia) (Garcia, 2002) approved in Despite it was firstly conceived as a line upgrade, it was decided to build a new line with high speed trains and international gauge. This first HSR line opened in 1992 linking Madrid with Seville (472 km) in less than 3 hours. Following the decision to build a HSR line linking Madrid and Seville, more studies were conducted for a new HSR line linking Madrid with Barcelona. However, the mid-nineties economic crisis pushed the new HSR line to Barcelona into the background until In 2000 the investment in HSR increased and the first stretches of the line opened in 2003 despite not at high speed. In 2006 the line started operating at high speed from Madrid up to Lleida and, in February 2008, it reached Barcelona linking Madrid and Barcelona (659 km) in 2 hours and 38 minutes. The Madrid-Barcelona HSR corridor was partially funded (27%) by European funds such as Cohesion Fund and the Trans-European Transport Networks Fund. Following the public investment increase in HSR from 2000, the Spanish Government approved a new Infrastructure and Transport Strategic Plan (Plan Estrategico de Infraestructuras y Transporte) (2004) in 2004 which established that by 2020 Spain would have 7200 km of HSR corridors and 90% of the population would live within 50 km to the nearest HSR station. Figure 1 shows the map of HSR corridors in Spain. In 2009 there were three main corridors: the south corridor between Madrid and Seville / Malaga; the north-west corridor between Madrid and Valladolid; and the north-east corridor between Madrid and Barcelona. In this paper we focus on the north-east corridor linking Madrid to Barcelona.

5 5 FIGURE 1 HSR network in Source: Wikipedia. 3. METHODOLOGY In this paper we use a difference-in-differences (DID) estimator to estimate the casual effects of investment in high-speed rail. We divide the Spanish economy into 47 peninsular provinces and treat access to HSR as a binary treatment. Provinces receiving / not receiving a high-speed rail connection are called treated / untreated and provide the basis for the DID analysis. This section provides a description of the difference-in-differences method. A binary treatment, which is administered to unit,, has two potential outcomes: is the response under control and is the response under treatment. We observe the outcome

6 6 but we cannot observe the joint density since the two outcomes never occur together. Instead, we seek to estimate an average treatment effect (ATE) by comparing responses across treated and untreated units. Clearly, there may be differences between the treated and untreated units which affect potential outcomes and which will therefore render simple treatment effect measures based on comparisons of means between groups subject to bias. Furthermore, for unbiased ATE estimation we need to adjust for temporal trends in the outcome variable and for the effects of extraneous events that are unrelated to the treatment. The DID estimator addresses such sources of bias by using information for both treated and control groups in both pre and post treatment periods. In the basic DID set-up we model the outcomes,, for units in binary time periods, with representing the pre-treatment period and t = 1 the post-treatment period as (1) where is the treatment indicator variable such that if unit has been exposed to the treatment prior to period and otherwise, is a time specific component, is a unit specific component, and is a potentially autoregressive error with mean zero in each time period. The effect of the treatment is captured by the parameter, but is not identified from (1) without further restrictions. The key identifying assumption for the basic DID model requires independence of the treatment assignment and the error, for. Adding in (1) gives (2) with and we can write, and similarly,. Defining and allows to rewrite (3) as

7 7 (3) The identifying assumption of the DID model means that we can obtain an unbiased estimator It follows that (4) As Imbens and Wooldridge (Imbens and Wooldridge, 2009) point out the double-differencing of the DID estimator removes two potential sources of bias. Firstly, it eliminates biases in second period comparisons between the treated and controlled groups that could arise from time invariant characteristics. Secondly, it corrects for time varying biases in comparisons over time for the treated group that could be attributable to time trends unrelated to the treatment. The identifying restriction on the basic DID model can be modified to allow for differences between units in covariates predetermined at that are thought to be associated with the dynamics of the outcome variables. By adjusting for differences in covariates with distributions that may differ between the treated and control groups we can accommodate heterogeneity in outcome dynamics between these two groups. With the vector of covariates included, the DID model is (5) The impact of the treatment across the population can be generalised by allowing for inter-actions between and. Two additional points are worth mentioning. First, in relation to serial correlation in the DID model, recent papers have shown that the existence of correlation within groups or over time periods can adversely affect the performance of the DID estimator. For instance, Bertrand et al. (Bertrand et al., 2004) investigate the problem of inconsistent standard errors arising from serially correlated outcomes. They show that the uncorrected standard errors from the DID estimator substantially understate the

8 8 variance of the estimates, thus greatly increasing the potential for Type I error. Second, the DID approach relies on the strong identifying assumption that the average outcomes for the treated and control groups would have followed parallel paths over time in the absence of the treatment: Clearly, the assumption will also hold conditional on covariates (i.e., and in fact it may be more reasonable to covariates to satisfy a parallel trend. 4. DATA In order to be able to measure the impact of the high-speed corridor linking Madrid to Barcelona on the economic performance of the provinces that received the transport infrastructure, we constructed a regional database comprising both economic and HSR-related data. The database consists of a panel of the Spanish peninsular provinces (47 provinces) and 28 Portuguese NUTS3 regions for the period As we explain in this section, we decide to extend the database with data for Portuguese regions because they can be used in the DID analysis as genuine untreated regions. The identification of untreated Spanish regions is more difficult because of potential interference effects between provinces that received a HSR connection and those that did not receive a HSR connection. As described in Section 3, the effect of the HSR corridor is represented through the creation of a binary variable,, which divides the data into two spatial groups: the treated ( and untreated provinces. Similarly, we also create a binary time variable,, to distinguish between the pre and post treatment periods. The outcome variable analysed is the provincial total Gross Value Added (GVA) per capita. 1 The data are obtained from the Regional Account Series published by the Spanish National Institute of Statistics (INE). Total GVA values are turned into constant terms applying the national GDP deflator from the BD-Macro dataset published by the Spanish Ministry of Exchequer (Diaz Ballesteros and Garcia Garcia, 2010). Differences in the outcome variable across provinces may occur because of regional characteristics, which may be associated with the outcome variable itself (i.e. GVA per capita). To account for this 1 GVA is equivalent to GDP without including taxes and subsidies on production.

9 source of heterogeneity we include other covariates ( ) in our DID models in order to distinguish the true treatment effect. 9 We include variables for the share of industrial jobs in the economy (shrindo) and the share of services sector jobs (shrservs). In this case, industry has a wide scope in order to match with the Portuguese data definition. Industry jobs refer to the aggregation of manufacturing, construction and energy sector jobs according to the Spanish series. Data regarding the industrial and services sector job share are obtained from Instituto Valenciano de Investigaciones Económicas (Arellano and Bond, 1991). Similar data were collected for the Portuguese regions from the National Institute of Statistics (INE). Table 1 provides basic descriptive statistics for the outcome variable (GVApc; in Euros and 1995 constant terms) and the variables describing the economic structure of the Spanish and Portuguese regions (shrindo, shrservs). TABLE 1 Summary statistics. Variable Obs. Mean Median SD Min Max GVApc Shrindo Shrservs Legend: SD=standard deviation. 5. RESULTS AND DISCUSSION 5.1 Establishing the control group for the DID analysis of the Madrid-Barcelona HSR corridor In order to assess the effect of the HSR corridor linking Madrid with Barcelona we need to ensure that the parallel trend assumption between the treated and the untreated (control) groups holds during the period studied. This is, in the absence of the treatment the outcome variable would follow the same trend over time in the treated and the control groups.

10 10 We start by creating four potential control groups. Three correspond to the Spanish provinces without any HSR station: north (10 provinces), east (11 provinces) and west (13 provinces). 2 The fourth group consists of the 28 Portuguese NUTS3 regions. The Spanish groups are more or less delimited by the Y-shaped Spanish HSR network in Other provinces which have not been treated by the north-east HSR corridor nor been added to the control group are excluded from the analysis. These essentially refer to provinces in the south and north-west HSR corridor. These provinces have also received a treatment (i.e. another high-speed corridor), and hence cannot be used in this study. The north group represents the most industrialized provinces of Spain with an average share of manufacturing jobs of 35%. Both the west and east groups have lower industrial job shares, 27% and 30% respectively, and they have not been merged in one group to ensure we have a balanced number of provinces in each potential control group. Figure 2 shows the provinces included in this analysis: dark grey (treated); light grey (potential control); the remaining provinces in black were excluded from this analysis for the reasons discussed earlier. 2 The Spanish control groups are formed by the following provinces: North (Alava, Burgos, Cantabria, Guipuzcoa, Huesca, Navarra, Palencia, Rioja, Soria and Vizcaya), West (Asturias, Avila, Badajoz, Caceres, Cadiz, Coruna, Huelva, Leon, Lugo, Ourense, Pontevedra, Salamanca and Zamora) and East (Albacete, Alicante, Almeria, Castellon, Cuenca, Girona, Granada, Jaen, Murcia, Teruel and Valencia).

11 11 FIGURE 2 Control and treated provinces for the north-east corridor. We test whether any of the potential control groups satisfies the parallel trend assumption for the period, which corresponds to the pre-treatment period of the analysis, by means of a regression analysis. If the assumption holds, the outcome variable must follow the same trend over time in the treated and each of the potential control groups during the pre-treatment period. We run the regression of GVA per capita as the outcome variable in province i at time t as shown in Equation 6. We run this regression on the year variable and the three interaction terms between year and the three potential control groups north, east and west (yrn, yre and yrw respectively) where the base case is the group of treated provinces. Next we estimate the same equation adding two covariates to take into account the economic structure of the provinces. The covariates are the share of jobs in the industrial, energy and construction sector, shrindo, and the share of jobs in the services sector, shrservs. (6)

12 12 We use ordinary least squares (OLS) and random-effects (RE) estimators and conclude that the random-effects model is preferred as we strongly reject the null hypothesis of zero variance across units by means of a Breusch-Pagan test. Due to the properties of the panel and the risk of serial correlation, we use use two different random-effects estimators. First, a feasible generalized least squares (FGLS) for panel data is used which allows for first order serial correlation within panels and heteroskedasticity in the disturbance term structure. Secondly, generalised least squares (GLS) randomeffects estimator is used which only allows for serial correlation of the disturbance term. When the coefficient of the interaction variables in Equation 6 (, ) is significant we conclude that the respective control group has a significant slope divergence from the base group (treated province) and, consequently, the parallel trend assumption does not hold. According to the results shown in Table 2, only the north group can be effectively treated as a control because there is no statistically significant difference between its slope and that of the treated group. TABLE 2 Test of Spanish groups as potential controls. Spanish potential controls Variables (1) (2) (3) (4) Constant *** *** *** *** yr *** *** *** *** yre *** *** *** *** yrw *** *** *** *** yrn shrindo *** shrservs ** Observations R 2 Within R 2 - Between R 2 - Overall Estimator FGLS GLS FGLS GLS ***, **, * indicate significance at 1%, 5% and 10%, respectively. We also test whether the Portuguese regions can be a control for the DID model following the same model as for the Spanish potential controls. We apply the regression shown in Equation 7 where port refers to a binary variable for Portugal and yearport an interaction variable between the year and port variables. The base case is the group of treated provinces by the north-east HSR corridor. The results

13 13 shown in Table 3 suggest that Portuguese NUTS3 regions can also be used as a control for the northeast HSR corridor DID as the interaction coefficient, yrport, is statistically insignificant with both econometric techniques (FGLS and GLS). In models (7) and (8) we include shrindo and shrservs as covariates and the interaction term remains statistically insignificant. (7) TABLE 3 Test of Portuguese regions as potential controls. Portugal potential control group Variables (5) (6) (7) (8) Constant *** *** *** *** yr *** *** *** *** port *** *** yrport *** *** shrindo *** *** shrservs *** *** Observations R 2 - Within R 2 - Between R 2 - Overall Estimator FGLS GLS FGLS GLS ***, **, * indicate significance at 1%, 5% and 10%, respectively. We also test if the north control group and the Portuguese control group follow the same parallel trend for the period. Thus, we run the same model as in Equation 7 but in this case the base case is the control group of northern Spanish provinces and not the treated provinces of the north-east corridor. As in previous models, we also test the model with covariates to control for the differences in the structure of the economy such as the shrindo and shrservs variables explained above. The results in Table 4 show that Portugal diverges in relation to the group of northern Spanish provinces. Consequently we conclude that the parallel trend assumption does not hold between this two potential control groups.

14 14 TABLE 4 Divergence test of Portuguese regions with Spanish northern group. Portugal and northern Spanish control group Variables (9) (10) (11) (12) Constant *** *** *** *** yr *** *** *** *** port ** * * yrport *** ** * shrindo *** *** shrservs *** *** Observations R 2 - Within R 2 - Between R 2 - Overall Estimator FGLS GLS FGLS GLS ***, **, * indicate significance at 1%, 5% and 10%, respectively. Due to this divergence, we split the group of northern provinces into northern provinces that are contiguous to the north-east HSR treated provinces and, secondly, a group of non contiguous northern provinces. Next we test whether the group of Portuguese NUTS3 units diverge, firstly, with relation to the group of contiguous, and secondly to the group of non contiguous northern provinces. The results in Table 5 show that the Portuguese regions and the contiguous northern provinces do not diverge as the interaction term as the interaction variable yrport is statistically insignificant. However, in Table 6 we test the divergence of Portuguese NUTS3 regions with respect to non contiguous northern provinces. In this case, the insteraction variable yrport is signficant and shows that the group of Portuguese NUTS3 regions follow a divergent trend in comparison with the non contiguous northern provinces.

15 15 TABLE 5 Divergence test of Portuguese regions with contiguous northern provinces. Portugal and contiguous northern provinces Variables Constant *** *** *** *** yr *** *** *** *** port yrport shrindo *** *** shrservs *** *** Observations R 2 - Within R 2 - Between R 2 - Overall Estimator FGLS GLS FGLS GLS ***, **, * indicate significance at 1%, 5% and 10%, respectively. TABLE 6 Divergence test of Portuguese regions with non contiguous northern provinces. Portugal and non contiguous northern provinces Variables Constant *** *** *** *** yr *** *** *** *** port *** *** ** *** yrport *** *** ** *** shrindo *** *** shrservs *** *** Observations R 2 - Within R 2 - Between R 2 - Overall Estimator FGLS GLS FGLS GLS ***, **, * indicate significance at 1%, 5% and 10%, respectively. It is difficult to understand why the paprralel trend assumption holds for the Portuguese regions and the Spanish northern provinces contiguous to the provinces receinving th ehigh-speed rail corridor, but does not hold for the group of all northern provinces. Therefore, in Section 5.2 we present the results obtained from the DID models using the following control groups: (i) Portuguese regions (Table 7); (ii)

16 Portuguese regions and Spanish northern provinces contiguous tho the north-east high-speed corridor (Table 8); and (iii) Portuguese regions and the Spanish northern provinces (Table 9) Results of the DID analysis of the Madrid-Barcelona HSR corridor The results from the DID model (Equation 8) are reported in tables 7 to 9. Equation 8 below is also extended with the two covariates describing the economic structure of the treated and untreated regions (i.e. shrindo and shrservs). (8) We estimate the results with three different control groups. First, with estimated the DID model using the group of Portuguese regions as untreated regions - models (1) to (4), shown in Table 7. Second, we use the group of Portuguese NUTS3 regions and Spanish contiguous provinces as a group of untreated regions - models (5) to (8), in Table 8. Finally, we used the group formed by the Portuguese NUTS3 regions and the whole group of Spanish northern provinces (contiguous and non contiguous) as untreated regions - models (9) to (12) in Table 9. All the models have been estimated twice using the FGLS (odd numbered models) estimator and the GLS (even numbered models) estimator. Table 7 shows the results for Equation 8 with the group of Portuguese regions as the only control group. Models (1) and (2) suggest that there is a statistically significant change in the average GVA per capita after 2006 as the coefficient for post06 is positive and highly significant. However, the value of the coefficient is moderate, and , suggesting that the average GVA per capita for the years after 2005 is, on average, 1.8% and 2.9% higher than the average GVA per capita for the period, holding other factors fixed. The coefficient for the treated variable is also positive and significant but much larger than for post06. According to model (1) and (2), the average GVA per capita of the treated provinces by the north-east HSR corridor is around 70% higher than the average GVA per capita of the Portuguese NUTS3 regions. This is reasonable as the treated provinces comprise provinces such as Madrid, Barcelona and Lleida with high GVA per capita in comparison to other Iberian regions.

17 17 The coefficient of interest is that corresponding to the interaction variable between post06 and treated. As it has been explained above, post06*treated is the DID estimator which corresponds with the effect of the treatment once the time effect and the particularities of the treated provinces have been taken into account by means of the post06 and treated variables respectively. Both models (1) and (2) present a non significant estimate of the interaction term coefficient, suggesting that the effect of the HSR corridor on GVA per capita has been negligible. In models (3) and (4) we estimate the same equation adding the covariates shrindo and shrservs, as explained above. The results are consistent with models (1) and (2). The coefficients of the covariates are statistically significant at a 1% confidence level and have positive coefficients. In the case of the services sector, if the share of employed in the services sector increases by 1 percentage point, the average GVA per capita increases on average 1.3%, holding other factors fixed. The addition of covariates to take into account the economic structure of the regions causes a sharp decrease in the treated and constant estimates of models (2) and (4) in comparison with models (1) and (2). However, the interaction term between post06 and treated, post06*treated, remains statistically insignificant. Therefore, we can conclude that the HSR treatment seems to have had no significant effect on the economic growth of the regions receiving the high-speed rail corridor. TABLE 7 Results with the group of Portuguese NUTS3 regions as control. Portugal as control group Variables (1) (2) (3) (4) Constant *** *** *** *** post ** *** *** *** treated *** *** *** *** post06*treated shrindo *** *** shrservs *** *** Observations R 2 - Within R 2 - Between R 2 - Overall Estimator FGLS GLS FGLS GLS ***, **, * indicate significance at 1%, 5% and 10%, respectively.

18 18 Table 8 shows the results for the same equation 8 using the Portuguese regions and contiguous northern provinces as the untreated group. The estimates are similar, particularly for the interaction variable which shows no statistically significant effects of the treatment. The constant and the post06 variable have similar values to the previous estimates while those of the treated variable are more moderate. This may be caused by the addition of contiguous provinces in the control group as they have similar GVA per capita values with the north-east treated provinces. Therefore, the difference between the control and the treated in terms of GVA per capita values is reduced. TABLE 8 Results with the group of Portuguese NUTS3 regions and contiguous northern provinces as control. Portugal and contiguous provinces as control Variables (5) (6) (7) (8) Constant *** *** *** *** post ** *** * *** treated *** *** *** *** post06*treated shrindo *** *** shrservs *** *** Observations R 2 - Within R 2 - Between R 2 - Overall Estimator FGLS GLS FGLS GLS ***, **, * indicate significance at 1%, 5% and 10%, respectively. Table 9 presents the estimates for Equation 8 with a control group formed by the Portuguese regions and the group of Spanish northern provinces. Overall, the estimates are very similar to the other models and they confirm the null impact of the north-east HSR corridor on the economic growth of the treated provinces. In this case, estimates for the treated variable are lower as the difference between average GVA per capita of treated provinces and average GVA per capita of control provinces is smaller. This is caused by the addition of Spanish northern provinces, which have a high level of income per capita.

19 19 TABLE 9 Results with the group of Portuguese regions and Spanish northern provinces as control. Portugal and northern Spanish control group Variables (9) (10) (11) (12) Constant *** *** *** *** post ** *** ** *** treated *** *** *** *** post06*treated shrindo *** *** shrservs *** *** Observations R 2 - Within R 2 - Between R 2 - Overall Estimator FGLS GLS FGLS GLS ***, **, * indicate significance at 1%, 5% and 10%, respectively. 6. CONCLUSIONS During the period between 2000 and 2010 Spain carried out the largest investment in high-speed rail in Europe, and is currently the country with the largest high-speed network in Europe and the highest rate of HSR km per capita in the world. Investments in high-speed rail projects have frequently been justified on the basis of supposed positive effects on the economic growth of the regions receiving the investment. Nevertheless, many of these benefits are unclear and depend on a case to case evaluation. This paper conducted empirical work based on difference-in-differences (DID) methods and panel data for Spanish and Portuguese regions to quantify the impact of the Madrid-Barcelona high-speed rail corridor on the economic growth of the Spanish provinces with access to the high-speed rail network. The results of the analyses indicate that the alleged claims of a positive economic impact on the economic performance of the regions receiving the transport investment have not taken place. In the case of the Madrid-Barcelona high-speed rail corridor, our results show that there are no significant differences in the pattern of regional economic growth before and after the high-speed rail corridor between the treated and untreated provinces.

20 REFERENCES Plan Estrategico de Infraestructuras y Transporte. Secretaria de Estado de Infrastructuras y Planificacion, Ministerio de Fomento. ANGRIST, J. D. & KRUEGER, A. B Instrumental Variables and the Search for Identification: From Supply and Demand to Natural Experiments Papers 455, Princeton, Department of Economics - Industrial Relations Sections. Journal of Economic Perspectives, 15, ANSELIN, L Spatial Econometrics:Methods and Models, Kluwer Academic Publishers. ARELLANO, M Panel Data Econometrics, Oxford University Press. ARELLANO, M. & BOND, S. R Some Tests of Specification for Panel Data. The Review of Economic Studies, 58, 277. BANISTER, D. & BERECHMAN, J Transport Investment and Economic Development, London, Routledge. BERTRAND, M., DUFLO, E. & MULLAINATHAN, S How much should we trust differencein-differences estimates? Quarterly Journal of Economics 119, BOARNET, M. G Highways and Economic Productivity: Interpreting Recent Evidence. Journal of Planning Literature, 11, BUTTON, K. J Infrastructure Investment, Endogenous Growth and Economic Convergence. The Annals of Regional Science, 32, DIAZ BALLESTEROS, A. & GARCIA GARCIA, E Base de Datos Macroeconómicos de España-BDMACRO Últimafecha de actualización: octubre de ed.: Ministerio de Economía y Hacienda. DURANTON, G. & PUGA, D Micro-Foundations of Urban Agglomeration Economies. In: HENDERSON, J. V. & THISSE, J. F. (eds.) Handbook of Regional and Urban Economics. Amsterdam: Elsevier. GARCIA, J. S La red ferroviaria andaluza: infraestructura y modelo territorial. Cuadernos Geograficos, 32, GRAHAM, D. J Agglomeration, productivity and transport investment. Journal of Transport Economics and Policy, 41, GRAMLICH, E. M Infrastructure Investment: A Review Essay. Journal of Economic Literature, 32, IMBENS, G. W. & WOOLDRIDGE, J. M Recent developments in the econometrics of program evaluation. Journal of Economic Literature, 47, INE Anuário Estatístico. Instituto Nacional de Estatistica. (several years). INE Regional Account Series. Instituto Nacional de Estadistica (several years). MUNNELL, A. H Policy Watch: Infrastructure Investment and Economic Growth. The Journal of Economic Perspectives, 6, RIETVELD, P Spatial Economic Impacts of Transport Infrastructure Supply. Transportation Research Part A, 28, ROSENTHAL, S. S. & STRANGE, W. C Evidence on the Nature and Sources of Agglomeration Economies. In: HENDERSON, J. V. & THISSE, J. F. (eds.) Handbook of Urban and Regional Economics, Vol.4. Amsterdam: Elsevier. UIC High speed lines in the world. International Union of Railways (UIC). Accessed online on:10th March 2012.