The Spatial Wage Structure in the European Union –What Role for Economic Geography? Holger Breinlich LSE and CEP WORK IN PROGRESS! 1 Introduction Within the last ten years, the “New Economic Geography” has experienced rapid theoretical advances. Empirical research, however, is still in an early stage. The existing literature has investigated the effects of geography on trade flows, the structure of production, and the spatial variation of income levels1. This paper is concerned with the latter aspect and is thus part of a small but growing number of empirical studies that have used the New Economic Geography framework to analyse the impact of distance from markets for the level of wages. The basic idea is that firms in locations further away from consumers will face higher transport costs which reduces their exports receipts, resulting in a lower value added attributable to the factors of production. A first strand of the literature is concerned with the effects of geography at a national level, assuming perfect labour mobility between regions (and thus equalisation of real wages). Hanson (1998) shows for a panel of US counties that the market access of a location (measured as the weighted sum of all regional GDPs, where weights are inversely related to distance from the location in question) has indeed a significant positive impact on local (nominal) wages. Mion (2002) finds similar results for Italy, De Bruyne (2002) for Belgium and Brakman et al. (2000) for Germany2. Redding and Venables (2000) look at the relation between economic geography and income on an international level. In their model, which builds on Krugman and Venables (1995) and Fujita, Krugman and Venables (1999), wage levels are also influenced by the existence of intermediate factors of production: distance from suppliers of these factors increases intermediate input prices and thus further reduces value added. Taking the location of demand and production as given, they derive a structural equation relating wages to market and supplier access, which measure the proximity of a country to markets and intermediate goods suppliers, respectively. Using data on 101 countries, they find that a significant fraction of income differences can be explained by these two variables. The results are robust to the inclusion of a number of other potential determinants. This paper is concerned with the role of economic geography for the spatial structure of regional wage levels in the European Union during the period 1980-1997. Methodologically, it is very much related to the approach used by Redding and Venables, which seems to be more appropriate to the issue in question, given the very low mobility of labour both between and within countries in the European Union. For example, Barro and Sala-i-Martin (1995) estimate the impact of income differences on regional migration for several European countries3. On average, a ten percent increase in local real GDP per capita leads, ceteris paribus, to a yearly population inflow of less than 0.1%. The present analysis contributes to the growing literature on the consequences of economic geography for income levels by applying the theoretical framework developed by Krugman and Venables (1995) and Fujita, Krugman and Venables (1999) to a new kind of data. A priori, it is not clear whether the importance of access to markets and intermediate goods is as significant on a regional level as it is in the international context studied by Redding and Venables. Other mechanisms, such as technological spillovers or density effects on productivity might be more important here. Also, given the panel-character of the data set used, a more detailed investigation of the implications of the underlying model is possible. As 1 See Overman, Redding and Venables (2001) for an overview of the literature. De Bruyne tests for employment density rather than for wages. 3 The countries are Germany, Italy, France, Spain and the United Kingdom; the period of observation is 19501990. For the 1980s, data is available for Germany, Italy and Spain, only, and indicates a slightly higher mobility, which is still only 0.1%. 2 compared to the papers using national data, this study has the advantage of taking detailed account of demand and supply linkages that go beyond national borders 4. This seems to be especially important in the case of the three European economies studied (Italy, Belgium and Germany) which are more economically integrated with their neighbours than the United States are5. The structure of this paper is as follows. Section 2 presents a brief description of the spatial structure of wage levels in the European Union and how it has changed during the period 1980-1997. Section 3 introduces the theoretical framework from which the econometric specification is derived that is used in the subsequent sections. Section 4 estimates a trade equation for trade both within the European Union and between the EU and the rest of the world in order to derive estimates which will be used in section 5.1 to calculate the market and supplier access of EU-regions. In sections 5.2-5.5, regional income levels are regressed on these measures and several robustness checks are undertaken. Section 6 notes alternative explanations for the spatial income structure that is found to exist within the EU and section 7 concludes. As this paper is very much work in progress, ideas and extensions that go beyond the basic framework are so far only briefly sketched. 4 De Bruyne and Brakman et al. include GDP of other EU-countries. However, they use distance/commuting time to capital cities as their transport cost measure which reduces cross-regional variation of this part of market access and underestimates the advantage of border regions. 5 A rough indicator of integration is the share of external trade in GDP (calculated as (imports+exports)/GDP). Figures are 25% for the USA, 49% for Italy, 55% for Germany and 145% for Belgium (source: Penn World Tables 6.1, year:1997). 2 Wage Gradients in Europe 1980-1997 Though differences are smaller than between countries in a worldwide perspective, income levels within the European Union still vary by a substantial amount. For example, in 1997, per capita GDP in the 5% richest Nuts2-regions was more than three times higher than that of the 5% poorest regions6. Also, there seems to be a strong income gradient, i.e. per capita GDP at the geographical periphery is lower than at the centre. Table 1 shows this relation by regressing the log of per capita GDP on distance from Luxembourg, the approximate geographical centre of the EU15. On average, per capita GDP decreases by 4% for every 100km distance from Luxembourg. Also, the gradient seems to have flattened out slightly since 1980, as displayed in figure 1 (the figure shows the distance coefficient of above regression for every year since 1980 for the subsample of the EU127). The existence of a strong but declining centre-periphery gradient is at least not inconsistent with market access being a determinant of income levels. Peripheral countries might be at a disadvantage due to the distance from main European markets. Also, as European integration proceeds and inter-regional transport costs decline, centrality might become less important for income levels, leading to a flattening out of gradients. Whether this cursory impression is correct or not will be the subject of the subsequent analysis. 3 3.1 Theoretical Framework and Econometric Specifications The Model The theoretical framework underlying the empirical analysis in this paper is a reduced version of a standard New Economic Geography model with intermediate goods (see, for example, Fujita, Krugman and Venables (1999), chapter 14). I consider a world with R locations, where the focus is on the manufacturing industry which produces under increasing returns to scale and product differentiation. Manufacturing goods can be used for final consumption and as intermediate inputs. Final demand for goods in location j is derived via utility maximisation of the representative consumer’s CES utility function: R ( 1) / U j ni xij l 1 /( 1) R s.t.[ ni pij xij ] Y j i 1 Where ni is the number of firms in location i and xij is the amount of consumption in location j of a variety produced in i8. Sigma is the elasticity of substitution between varieties and pij the price of location i varieties in j (consistent of the mill price pi and iceberg transportation costs Tij between the two location: pij=piTij). Finally, Yj is income in location j. Solving the optimisation problem we obtain the demand facing a firm in i from location j: 6 Figure for EU15, GDP measured in 1995-ECU. See appendix A for a description of the data. There is no GDP data available at the Nuts2-level for 1980-1988 for Austria, Finland and Sweden. This exclusion also explains the (absolutely) larger coefficients in figure 1 as Finland and Sweden are peripheral highincome countries. Including these three countries for 1988-1998 shows a qualitatively similar picture of flattening wage gradients (from –0.05 to –0.04). 8 In equilibrium, the amount of the different varieties produced in location i and consumed in location j can be shown to be equal (due to symmetric weights of varieties in the utility function and identical production technologies). 7 xij pij R n p 1 nj n n 1 Yj Defining a price index for manufacturing goods Gj and rewriting final consumption yields: R G j [ nn p nj 1 1 / 1 ] xij n 1 Cons pij Yj G j ( 1) Demand for intermediate goods is derived analogously. Assuming a Cobb-Douglas production function with share α of intermediates and a sub-production function for aggregation of intermediates similar to U we get: xij IM pij I j Gj where I is the amount of intermediates used for production in j. Thus total demand facing a firm in i from j is: xij pij I j G j pij Yj G j 1 pij Gj 1 (I jG j Yj ) Summing over all products produced in location i, we obtain the value of total exports from i to j as (“trade equation”): ni pij xij ni pi (1) 1 (Tij )1 G j 1 (E j ) where the last term (Ej) stands for total expenditure in j, summarising final and intermediate demand. Turning to the supply side, firms maximise the following profit function with respect to prices: R i pi xin Gi wi vi ci [ F xi ] n 1 where w is the wage rate, v the price of other factors of production, c the unit input requirement and F fixed costs. Optimal prices turn out to be set at a mark-up σ/(σ-1) over marginal costs. Free entry assures that long-run profits will be zero, implying: R pi x E j G j j 1 1 (Tij )1 Inserting the profit maximising price I obtain: R ß 1 x Gi wi vi ci E j G j (Tij )1 1 j 1 This expression can be transformed to yield the maximum remuneration firms can afford to pay the factors of production other than intermediate goods. Focusing on wages, I derive the so-called “wage equation”: R ß 1 x Gi wi v i c i E j G j (Tij )1 1 j 1 (2) wi A( SAi ) ( 1) 1 ( MAi ) 1 vi ci where R SAi snTin n 1 1 R nn p n n 1 1 1 Tin R and MAi m n Tin n 1 1 R E n Gn n 1 1 Tin 1 The terms s and m stand for supply and market capacity of a location n, respectively. The first reflects the number and price competitiveness of intermediate goods produced in n, the second summarises the market potential of a location. The terms SA and MA, in turn, stand for supplier access and market access and are transport cost weighted sums of the supply and market capacities of all regions. These expressions summarise how well a location is endowed with access to cheap and numerous intermediate products (supplier access) and with access to markets for the products it produces (market access). As explained in the introduction, firms in locations with higher market and supplier access incur less transportation costs and are able to pay higher wages. In this model, the location of firms and demand (n and E) is taken to be exogenously given. This assumption allows the direct derivation of the econometric specifications from the model. In a fully fledged version of this model (such as in Fujita, Krugman and Venables (1999), chapter 14), both variables are endogenised. This leads to the possibility of cumulative causation and multiple equilibria which would make empirical analysis much harder. 3.2 Econometric Specifications The two key relationships that will figure centrally in the rest of the paper are equations (1) and (2). Equation (1), the “trade equation”, can be rewritten using the definitions of market and supply capacities to relate bilateral exports to these measures: ni pij xij si (Tij )1 m j Taking logs, I obtain the econometric specification used in the estimation in section 4: (3) ln(exp ij ) ln( si ) (1 ) ln( Tij ) ln( m j ) ij In practice, market and supplier access will be proxied by the trading partners’ GDP, transport costs by bilateral distances. From the coefficients of this regression, estimates of market and supply capacities of European Regions and countries around the world can be derived (details in section 4). These in turn can be used to calculate market and supplier access of the Nuts2 regions (details in section 5). The resulting access measures will then be used in the second key equation of this paper, the wage equation (2). This equation relates wages to market and supplier access. Again, by taking logs I get the specification used in the estimation in section 5: (4) ln( wi ) 1 ln( SˆAi ) 2 ln( Mˆ Ai ) i 4 Estimation of Trade Equation In this section, I will estimate the trade equation derived in section 3. The underlying idea is to use the information contained in trade flows to get estimates for market and supply capacities and bilateral transport costs. The obvious problem this approach encounters is that there is almost no data on bilateral trade flows at a regional level for the European Union. To circumvent this problem, the assumption needs to be made that interregional trade flows are governed by the same forces as international ones. That is, it is assumed that estimates obtained from a gravity equation on an international level can be used with regional data on GDP and bilateral distances to calculate market and supply capacities. To make these estimates more applicable to regional European data, only trade flows within the EU12 and between the EU12 and the rest of the world (ROW) are included in the estimation9. Furthermore, I allow the estimates of transport costs to vary between within EU and ROW-EU trade, possibly capturing differences in the importance of distance for the two subsamples. The trading partners’ GDPs are used to proxy market and supply capacity. A specification using country dummies would be preferable, especially for supply capacity which summarises number and prices of intermediate goods and might be poorly proxied by GDP10. However, such estimates would not be transferable to the regional level. Data on bilateral manufacturing trade flows and distances is taken from the NBER World Trade Database (Feenstra et al., 1997; Feenstra, 2000), GDP data from the World Development Indicators 2001. The number of countries present in both data sets is 148. The sample had to be reduced further due to missing GDP-data, leaving us with 119 countries11. In order to make observations comparable over time, all data are expressed in 1995 US-dollars12. The data are aggregated into three six-year periods (1980-1985, 1986-1991, 1992-1997) to smooth short-run fluctuations in trade flows and GDP. Finally, it is known from the literature on border effects that capital-to-capital distances between neighbouring countries are usually overmeasured13. In the sample used this could lead to the erroneous conclusion that distance is more important for intra-EU12 trade than for trade with the ROW. To reduce this problem, adjusted distances are calculated as the GDP-weighted sum of bilateral distances between the Nuts2-regions of the two trading partners (for intra-EU12 trade) or as the sum of GDPweighted distances between regions and non-EU12 countries for EU12-ROW trade. That is: dist IJ si s j dist ij where si is share of region i in country I’s GDP. For J being a iI jJ non-EU12 country, this simplifies to: dist IJ si dist iJ . iI The specification estimated is derived from equation (3) in section 3.2 by replacing market and supply capacities with the trading partners’ GDPs. Furthermore, transport costs are split up in a multiplicative and an exponential component (Cmult and Cexp), allowing for a more flexible relationship between distance and transport costs: 9 The estimation was also performed for the EU15 and the ROW, yielding only marginal differences in results. Redding and Venables (2000) use dummies in their estimation. 11 See appendix A for a list of countries included. 12 The trade data are deflated using the US GDP deflator (source: World Development Indicators, 2001). 13 See for example Head and Mayer (2002). 10 ln( X ij ) ln( s i ) (1 ) ln( Tij ) ln( m j ) ln(exp ij ) ln( GDPi ) (1 ) ln( Cmult ij dist ij C exp ij ) ln( GDPj ) ln(exp ij ) (1 ) ln( Cmult ij ) ln( GDPi ) (1 ) (C exp ij ) ln( dist ij ) ln( GDPj ) ln(exp ij ) 1 EU int 2 EUext 1 ln( dist ij ) EU int 2 ln( dist ij ) EUext 1 ln( GDPi ) 2 ln( GDP j ) ij where EUint and EUext are dummy variables that take the value 1 when a trade flow is intraEU12 or EU12-ROW, respectively. This specification is separately estimated for all three sixyear periods. Table 1 presents results for both OLS and Tobit-estimation. The latter takes into account that a small fraction of the trade data is left-censored at zero. As expected, the Tobit estimates show a slightly lower coefficient on distance, though differences are small due to the small number of zero-observations14. The Tobit estimates are used in the calculation of transport costs, and market and supplier access in the following. From the intercept and distance coefficients, an estimate of the transport cost term in the trade and the market and supplier access equations can be calculated for each of the three periods as follows: Tij 1 e1 dist ij 1 e 2 dist ij Tij 1 2 for intra-EU12 trade for trade ROW-EU12 These formulae show that it is actually important to allow for time varying intercepts in the trade equation when comparing transport costs over time as changes in the coefficient of the distance variable only reflect changes in the slope of the distance-transport cost relationship. Results indicate a steady rise of Tij1-σ for both trade within the EU12 and for EU12-ROW trade with an acceleration in the 1990s (note that absolute figures are difficult to interpret as they depend on the choice of units in the trade equation). Using the median adjusted distance between EU12-countries of approximately 1200km, the increase is 22% between the first two periods (80-85 and 86-91) and 66% between the two last periods (86-91 and 92-97). For EU12-ROW trade the figures are 16% and 101%, respectively. Assuming a demand elasticity of σ=6, this corresponds to a decrease of intra-EU12 iceberg transportation costs of 4% and 10% (3% and 13% for EU12-ROW15). Though the gap seemed to be closing slowly in the 1990s, exports to a non-EU12 country that was 1200km away were still 8% more costly in the last period under consideration (1992-97) than exporting to a EU12 country at the same distance. Extensions: - Use regional trade data where available and compare results of regional gravity-equation with values obtained here. Existing data are: o Trade flows between British regions and the rest of the world (from HM Customs & Excise, http://www.uktradeinfo.com/index.cfm?task=aboutreg). o Interregional trade in France (see Combes, Lafourcade, Mayer, 2002). 14 Ceteris paribus, bilateral trade flows have a higher probability of being left-censored (i.e. equal zero) for higher bilateral distances, introducing an upward bias on the distance coefficient. 15 The median distance between EU12-countries and the ROW is 6600km. Using this figure in above calculations, we get a decrease in iceberg transportation costs of 6% and 11%. - Compare estimated distance coefficient with actual data on transportation costs (e.g. from Combes and Lafourcade, 2001). 5 Market Access, Supplier Access and Regional Wages 5.1 Construction and Summary Statistics Market and supplier access are constructed following the formulae in section 3.1, using the results of the trade equation estimation. Thus, market access of a Nuts2-region is the weighted sum of GDPs of all other Nuts2-regions and the rest of the world. Formally, market access of region i in period t is given by: Mˆ Ait GDP ˆ 2 t jt jEU 12 (eˆ1t dist ijt ˆ1t ) GDP ˆ 2 t jt jEU 12 (eˆ 2 t dist ijt ˆ 2 t ) Supplier access is calculated analogously, using the estimated parameter on reporter GDP in the trade equation: SˆAit GDP jEU 12 jt ˆ1t (eˆ1t dist ijt ˆ1t ) GDP jEU 12 jt ˆ1t (eˆ 2 t dist ijt ˆ 2 t ) To calculate the part of market or supplier access a region derives from itself, intraregional transport costs are calculated by replacing own-distances with values obtained from the formulae distii=0.33(area/)1/2 which gives the average distance between two points in a circular region. Due to the fact that I allow for varying transport cost coefficients, both access measures can be split up into an EU12-part and a remaining part (the first and the second sum in above equations). Table 3 provides some information on composition and changes of these variables. Figure 2 graphs market and supplier access as a function of distance from Luxembourg, the approximate geographical centre of the EU15, for the period 1992-97 (using distance from the French region of Lorraine which corresponds a little better to the centre of the EU12 regions used here, does not change results perceptibly). 5.2 Wage equation – baseline specification I now proceed to the estimation of the wage equation. The baseline specification is that derived in section 3.2: ln( wi ) 1 ln( SˆAi ) 2 ln( Mˆ Ai ) i As had to be expected from the construction of both market and supplier access as a function of distance weighted GDPs of surrounding locations, the two measures are highly correlated (correlation coefficient: 0.82). Thus, due to multicollinearity problems the baseline specification is estimated separately for market and supplier access. To avoid the most obvious source of endogeneity, both estimations are done excluding own market access 16. The dependent variable is log of 1995 per capita GDP (in ECU), serving as a proxy for manufacturing wages. Whether this will bias the results certainly requires further investigation. However, note that with labour being relatively immobile across regions (in contrast to capital), workers will probably bear the main burden of low market and supplier 16 Including own market and supplier access introduces a positive correlation with the error term of the wage equation as a shock to own GDP raises both market and supplier access as well as per capita GDP unless GDP changes are driven by population changes only. Indeed, including own access measures raises coefficients and R2 slightly. access, whereas the remuneration of capital will tend to be equalised across regions. Thus, it is to be expected that using GDP data actually introduces a downward bias on the access coefficients. Using GDP also has the advantage that such data is readily available for most of the countries and periods of interest. Still, missing data restricts our analysis to the subsample of the EU12 (excluding Austria, Finland and Sweden) when looking at the full period of 1980-1997. In section 5.4, I will replicate results using GDP data for the full sample of countries for the period 1989-1997 and for a subsample using data on manufacturing wages. Table 4 shows the results of the baseline specification. Both market and supplier access are both statistically and economically highly significant. Doubling market access increases per capita GDP by approximately 70% (95% for supplier access). The access measures explain between 55-60% of income variations in the 171 EU12 Nuts2 regions used in the estimation. The next sections examine the robustness of these results. As results for market and supplier access are quite similar in all the following, I focus on market access only, which is probably more exactly measured given the problems of proxying supply capacity by GDP (see section 4). 5.3 Instrumentation and country fixed effects Another potential source of bias of the estimates comes from omitted variables that influence per capita GDP and are correlated across regions. As a first step, I instrument 1992-1997 market access by its 1980-85 value and regress average 1992-1997 per capita GDP on it. This should eliminate problems arising from intertemporarily uncorrelated variables (e.g. nationwide strikes). The lag length used will also reduce the impact of omitted variables correlated over time and space (e.g. business cycle fluctuations). Still, other important determinants of income levels, such as technological differences or the quality of institutions of a region, are likely to be highly persistent over time. Such differences will probably be smaller within countries than between countries and I try to capture them by additionally including country dummies in the regression. It should be noted that these dummies also capture the part of market access that is common across a country and thus identification relies on intranational variation only. As a further check on the consistency of the estimates, I instrument log market access by log distance from Luxembourg and the United States. The first measure is a good proxy for within-EU access, the second for extra-EU access. In a first stage regression, both explain approximately 80% of the variation in market access. Results of the various specifications are presented in table 5. Column one presents results for the IV-estimation using lagged values of market access, column two adds country-dummies to the regressors and column three shows the results of using above distances as instruments. Using lagged values doesn’t change the baseline results of section 5.2 by much due to the only small changes in the spatial variation of market access over time. Including country dummies raises the explanatory power of the regression enormously and reduces the coefficients on the access measures somewhat. However, both stay statistically and economically significant. Using distances as instruments also confirms the robustness of the results presented here17. 5.4 Full sample and manufacturing wages A potential concern with the results presented so far is that Austria, Finland and Sweden had to be excluded due to data limitations. As the last two countries are peripheral but high income countries, this will bias our results in favour of finding an important role for our access measures. Indeed, as seen in section 5.1, both market and supplier access decrease with 17 Note also that regression the log of per capita GDP directly on the two distance measures lowers the R2 of the regression to 47%, showing the superiority of an access-based approach. distance from Luxembourg. Including the three countries reduces the time period for which data are available to 1988-1997. Secondly, wages have been proxied by per capita GDP. As argued above this should bias our results on the economic importance of market access downwards, if anything. This section investigates this claim. Availability of manufacturing wages is very limited. There is data for a subsample of countries containing average total manufacturing wages for 1995-1997 (in 1995-ECU) for the regions of Austria, Sweden, Finland, Belgium, Spain, France, the Netherlands, Portugal, Italy and Ireland. A further problem arises when trying to calculate wages per employee as data on manufacturing employment is not always available. I thus present two sets of results. The first uses wages per capita where the denominator is total population of a region, yielding 112 observations. The second set uses actual data on manufacturing employment which reduces the number of observations to only 53 (no employment data is available for Austria, Belgium, France, the Netherlands and Portugal). Table 6 presents results for all three estimations. Column one contains estimates for a regression of log of per capita GDP for 1995-1997 (in 1995 ECUs) on market access 19891991 for the 194 EU15 regions. As expected from above considerations, both the explanatory power of the regression and the magnitude of the coefficient on market access is reduced. Columns two and three shows the result for the wage estimations, where again, current market access is instrumented by its 1989-92 value. Consistent with prior reasoning, the coefficient on market access is indeed increased even though central high income countries as Germany and Luxembourg had to be excluded (this may also explain the lower R2 of the regressions). 5.5 First Differences Besides making predictions about wage levels, the theoretical model presented in section 3 also implies that changes in market and supplier access should trigger changes in wages. A cursory look at the development of wage gradients in the European Union proved to be at least not inconsistent with this implication. Indeed, market access of peripheral regions has improved relative to that of central regions over the past two decades. This is mainly due to a decrease in the importance of distance within the European Union (see the changes of the distance coefficient “distint” in table 2). Figure 3 shows this development by plotting market access growth rates against distance from Luxembourg. The aim of this section is to investigate whether the cursory evidence presented so far stands the test of more formal econometric investigation. To this purpose, column one of table 7 presents results of a regression of changes of per capita GDP on changes in market access. For comparison, results from a fixed-effects estimation are displayed in column two. The first differenced regression confirms the above impression that regions with larger market access growth rates have indeed grown faster. However, the estimated coefficients are reduced somewhat as compared to sections 5.3 and 5.4. This might be explained by the fact that market access may influence per capita GDP through other channels, which may be of a more long run nature than the reduction of value added through transport cost (e.g. Redding and Schott (2003) point out the potential effect on human capital accumulation). Also, as the importance of manufacturing industry declines as compared to the service sector, changes in market access may have less impact on changes in per capita GDP, though the level of market access is still important for the level of income. Finally, the decrease in the coefficient might also be due to the reduction of omitted variable bias as time-constant determinates of per capita income are eliminated through first differences or fixed effects. Though the results of the different econometric experiments in this and the previous section have reduced concerns about omitted variable bias, it could still be that regional income increases are driven by changes in other variables which happen to be positively correlated with market access growth and show considerable variation within countries. The next section considers such variables. Also, the force that drives the findings of this section is the decline in the estimates of the distance coefficient within the European Union. As these estimates are derived from a gravity equation using international instead of interregional trade flows, one might be concerned about robustness. Thus, in the following I will also consider time-invariant variables that might drive results of the preceding sections. 6 Inclusion of control variables - Density of employment: Ciccone (2002) shows for five European countries that employment density raises productivity. For a start, I use population density instead of employment density as the former is more readily available. Following Ciccone, I instrument density by area size to control for the endogeneity arising from the fact that high density might be a consequence, not a cause of high productivity. Table 8 presents results for the period 1992-97 of a regression of log of per capita GDP on log of market access, log of population density and country dummies, where I use 1980-85 log market access and log area size as instruments. - Technological spillovers: available data for Nuts2-regions from Eurostat Regio are 1) number of patent applications; 2) R&D-expenditure and R&D-employess by sector (less complete than patent data). - Human and physical capital 7 Conclusion Large differences in income levels remain in the European Union despite ever deeper integration. This paper examines one potential explication why such differences are not bid away by firms taking advantage of cheaper production costs. Distance from markets for final goods and suppliers of intermediate goods may reduce the maximum amount of factor remuneration firms can pay which will disproportionately affect immobile factors such as labour. From a New Economic Geography model, econometric specifications are derived which are used to investigate the relevance of access considerations. Results indicate a significant role for market and supplier access for the level of wages in the European Union. Also, improved access of peripheral regions due to declining transportation costs seems to have had a positive impact on wage levels there, as shown by a first differenced regression. Though the preliminary results suggest an important role for market and supplier access, the analysis has to be improved in several ways before a final verdict can be given. First, the results of the trade equation should be confirmed by both estimations using interregional trade flows and actual data on transportation costs, where available. Secondly, inclusion of a series of control variables will help to reduce remaining doubts about potential biasedness of the results due to omitted variables. Finally, it would be desireable to find a more precise proxy for supplier access than simple GDP data which would allow to make more meaningful statements about the importance of closeness to sources of intermediate inputs. References: - Barro, R.J. and Sala-i-Martin, X. (1995):”Economic Growth”, New York, McGraw-Hill. - Brakman, S., H. Garretsen and Marc Schramm (2000):”The Empirical Relevance of the New Economic Geography: Testing for a Spatial Wage Structure in Germany”, CESifo Working Paper 395, Munich. - De Bruyne (2002), “The Location of Economic Activity: Is There a Spatial Employment Structure in Belgium?”, mimeo, KULeuven. - Feenstra, R., R. Lipsey and H. Bowen (1997), “World Trade Flows, 1970-1992, With Production an Tariff Data”, NBER Working Paper No. 5910. - Feenstra, R. (2000), “World Trade Flows, 1980-97”, University of California, Davis, mimeograph. - Fujita, M., P. Krugman and A.J. Venables (1999):”The spatial economy: cities, regions and international trade”, MIT Press, Cambridge MA. - Hanson, G. (1998):”Market Potential, Concentration”, NBER Working Paper 6429. - Head, K. and T. Mayer (2002):”Illusory Border Effects: Distance Mismeasurement Inflates Estimates of Home Bias in Trade”, CEPII Working Paper 2002-01. - Krugman, P. and A.J. Venables (1995):”Globalization and the Inequality of Nations”, Quarterly Journal of Economics, p.857-880. - Mion, G. (2002): “Spatial Externalities and Empirical Analysis: The case of Italy”, mimeo, Université Catholique de Louvain. - Overman, H.G. , S. Redding A.J. Venables (2001): “The Economic Geography of Trade, Production, and Income: A Survey of Empirics”, CEP Discussion Paper, 508 and CEPR Discussion Paper 2978. - Redding, S. and P. K. Schott (2003):”Distance, Skill Deepening and Development: Will Peripheral Countries ever get rich?”, NBER Working Paper 9447. - Redding, S. and A.J. Venables (2000):”Economic Geography and International Inequality”, CEPR Discussion Paper, 2568. Increasing Returns, and Geographic APPENDIX A: DESCRIPTION OF DATA Regional Data: Data on GDP, population, area, wages and employment for the 211 Nuts2-regions are taken from Eurostat’s Regio database. GDP and wage data are expressed in current ECU and are deflated to 1995-ECUs (the deflators used are calculated by Eurostat as current ECU value series divided by constant 1995 series for each country). The main dependent variable used in the empirical analysis is per capita GDP. Due to data limitations, the following adjustments had to be made: - The French Overseas Territories, the Portugal regions Acores and Madeiras and Eastern Germany (including Berlin) are excluded completely. Data for Austria, Finland and Sweden is only available between 1988-1997. These countries are included in part of the analysis, only. Data on Ireland, Luxembourg and Denmark corresponds to the Nuts0-level (countrylevel). Distances between regions are great circle distances between the main cities of the regions. EU12: Belgium, (West) Germany, Denmark, Spain, France, Greece, Luxembourg, Ireland, Italy, Luxembourg, Portugal, United Kingdom. EU15: EU12 plus Sweden, Finland and Austria. International Data: Country GDP from World Development Indicators. Countries included in the trade equation and in the calculation of market and supplier access: … Manufacturing trade flows in current USD from the NBER World Trade Database (Feenstra et al., 1997; Feenstra, 2000). USD deflator from World Development Indicators 2001. APPENDIX B: ESTIMATION RESULTS Table 1: Per Capita GDP Gradient in the EU15 lcgdp95 dist100 -0.041 (7.26)** Constant 10.023 (255.71)** Observations 194 R-squared 0.36 Notes: Table shows result of regression of log of per capita GDP (in 1995-ECU) of EU15 Nuts2regions on distance from Luxembourg (in 100s of km). t statistics in parentheses based on HuberWhite heteroskedasticity robust standard errors. * significant at 5%; ** significant at 1% Table 2: Trade Equation Estimation 1980-85, 1980-85, 1986-91, 1986-91, 1992-97, 1992-97, OLS Tobit OLS Tobit OLS Tobit lexp lexp lexp Lexp lexp lexp 0.871 0.880 0.897 0.901 0.900 0.900 (46.25)** (42.59)** (54.11)** (51.10)** (62.15)** (56.92)** 1.287 1.304 1.255 1.263 1.230 1.232 (60.44)** (62.92)** (65.57)** (71.44)** (68.69)** (77.81)** -1.050 -1.036 -0.977 -0.970 -0.959 -0.957 (9.12)** (3.18)** (8.71)** (3.44)** (9.29)** (3.68)** -0.794 -0.808 -0.713 -0.720 -0.774 -0.775 (11.26)** (13.09)** (10.64)** (13.38)** (17.03)** (15.65)** -20.046 -20.646 -20.611 -20.899 -20.419 -20.482 (18.21)** (8.53)** (19.92)** (9.96)** (21.68)** (10.63)** -22.403 -22.766 -23.076 -23.250 -22.114 -22.161 (28.85)** (26.26)** (32.51)** (30.35)** (33.87)** (32.32)** Observations 2486 2486 2486 2486 2486 2486 % zeros 2.9% 2.9% 1.5% 1.5% 0.5% 0.5% R-squared 0.91 - 0.92 - 0.93 - lpgdp lrgdp distint distext EUint EUext Notes: lpgdp and lrgdp are partner`s and reporter`s GDP (in logs), distint is ln(dist ij)*EUint, distext is ln(distij)*EUext. Robust t statistics in parentheses. * significant at 5%; ** significant at 1% Table 3: Summary Statistics on Market and Supplier Access Average fraction of MA 1980-1985 1986-1991 1992-1997 85% 82% 80% 61% 61% 59% - 8.1% 11% - -2.1% 5.4% derived from EU12 Average fraction of SA derived from EU12 Average yearly growth rate of MA between periods Average yearly growth rate of SA between periods Table 4: Baseline specification of wage equation Period lMA (1) (2) (3) (4) (5) (6) 1980-85 1986-91 1992-97 1980-85 1986-91 1992-97 lcgdp lcgdp lcgdp lcgdp lcgdp lcgdp 0.670 0.729 0.730 (15.14)** (15.05)** (15.10)** 0.909 0.987 0.970 (16.71)** (16.73)** (16.59)** lSA Constant 18.203 18.719 18.359 16.913 17.790 17.422 (31.29)** (30.70)** (31.82)** (37.57)** (36.03)** (37.12)** Observations 171 171 171 171 171 171 R-squared 0.56 0.55 0.57 0.58 0.58 0.59 Robust t statistics in parentheses * significant at 5%; ** significant at 1% Table 5: Wage regression with instruments and country-dummies lMA (1) (2) (3) lcgdp (1992-97) lcgdp (1992-97) lcgdp (1992-97) 0.724 0.361 0.739 (14.94)** (3.40)** (15.82)** dummybe 14.040 (11.55)** dummyde 14.277 (11.62)** dummydk 14.602 (11.28)** dummygr 13.598 (10.17)** dummyes 13.840 (10.56)** dummyfr 14.080 (11.19)** dummylu 14.655 (11.99)** dummypt 13.552 (10.21)** dummyuk 13.850 (11.00)** dummynl 14.052 (11.51)** dummyit 13.925 (10.88)** dummyie 14.126 (10.93)** Constant 18.286 - (31.62)** 18.466 (33.15)** Observations 171 171 170 R-squared 0.57 0.98 0.57 Robust t statistics in parentheses * significant at 5%; ** significant at 1% Table 6: Results for full EU15-sample and using manufacturing wages (1) (2) (3) lcgdp lcwage95 lemplwage95 0.614 0.853 1.121 (8.91)** (5.03)** (2.87)** 18.737 20.333 27.011 (18.62)** (8.01)** (4.51)** Observations 194 112 53 R-squared 0.37 0.24 0.21 lMA Constant Robust t statistics in parentheses * significant at 5%; ** significant at 1% Table 7: First Differences and Fixed Effects dslMA (1) First (2) Fixed Differences Effects dslcgdp lcgdp 0.195 (24.50)** lMA 0.198 (32.25)** Observations 342 513 R-squared 0.63 0.75 (within) Robust t-statistics in parenthesis * significant at 5%; ** significant at 1% Table 8: Density Effects lcgdp lMA 0.258 (2.26)* ldensity 0.072 (2.03)* dummybe 12.938 (9.98)** dummyde 13.178 (10.06)** dummydk 13.500 (9.84)** dummygr 12.500 (8.86)** dummyes 12.727 (9.18)** dummyfr 13.035 (9.81)** dummylu 13.605 (10.49)** dummypt 12.444 (8.86)** dummyuk 12.718 (9.53)** dummynl 12.939 (9.92)** dummyit 12.818 (9.44)** dummyie 13.089 (9.62)** Observations 171 Robust t statistics in parentheses. Dependent variable is log of per capita GDP (in 1995 ECUs) * significant at 5%; ** significant at 1% Figure 1: Evolving Wage Gradients in the EU12 dcoeff -.054867 -.061006 1980 1997 year Notes: Figure displays distance-coefficient of a regression of log per capita GDP (in 1995-ECU) on distance from Luxembourg (in 100s of km) for Nuts2-regions within EU12 (excluding Austria, Finland, Sweden). The coefficients are all statistically significant at the 1%-level. Figure 2: Market and Supplier Access and Distance from Luxembourg .000011 SA MA .000511 2.2e-06 .000167 0 3044 Distance to Luxembourg (km) 0 3044 Distance to Luxembourg (km) Figure 3: Average Yearly Growth Rates of Market Access and Distance from Luxembourg MAygrowth .120557 .092514 0 3044 Distance from Luxembourg in km