WPS4104 Road Infrastructure in Europe and Central Asia: Does Network Quality Affect Trade? Ben Shepherd and John S. Wilson*, Development Economics Research Group--Trade The World Bank Mail Stop MC3-303 1818 H St. NW Washington D.C. 20433 Abstract: We present a new database of minimum distance road routes connecting 138 cities in 27 countries across Europe and Central Asia. We use it to show that improved road network quality is robustly associated with higher intraregional trade flows. Gravity model simulations suggest that an ambitious but feasible road upgrade could increase trade by 50% over baseline, exceeding the expected gains from tariff reductions or trade facilitation programs of comparable scope. Cross-country spillovers due to overland transit are important: total intraregional trade could be increased by 30% by upgrading roads in just three countries--Albania, Hungary and Romania. Keywords: International Trade; Europe and Central Asia; Road Transport; Trade Facilitation; Gravity Model. JEL Codes: F13; F15; H54. World Bank Policy Research Working Paper 4104, December 2006 The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the view of the World Bank, its Executive Directors, or the countries they represent. Policy Research Working Papers are available online at http://econ.worldbank.org. * The authors are respectively Consultant and Lead Economist (DECRG). This work is part of a broader project on trade facilitation and development supported through a Trust Fund of the U.K. Department for International Development. Sincere thanks to Piet Buys, Uwe Deichmann and David Wheeler (World Bank) for their advice and assistance with the data and analytical framework, and to Yu Cheng Kuo for helping with compiling the road distance database. We are also grateful to Andreas Kopp (OECD), Henry Kerali, Bernard Hoekman, Beata Smarzynska Javorcik, David Cieslikowski, Tsukasa Hattori, Leonardo Iacovone and Souleymane Coulibaly (World Bank) for their comments and suggestions, as well as to seminar participants in Washington, D.C. and an OECD- ECMT conference in Berlin. Ayako Suzuki and Witold Czubala provided very helpful and expert research assistance. Comments to: bshepherd@worldbank.org or jswilson@worldbank.org. - 1 - 1 Introduction In this paper, we show that better roads are strongly associated with larger trade flows within the Eastern Europe and Central Asia region (ECA). Simulation results demonstrate that the benefits for ECA from upgrading its road network may well be greater than from tariff reforms or customs streamlining programs of comparable ambition. This is the case even when the up front costs of road improvement are netted out. We also find evidence of the need for regional coordination of road network upgrades. This is due to large cross-country spillovers stemming from transit effects. Our results indicate that road investments in Albania, Hungary and Romania are likely to have particularly large trade payoffs for the region as a whole. We focus on road transport because of its particular importance in the ECA region (Molnar and Ojala 2003; ADB 2006; Cadot et al. 2006). Part of this importance comes from the fact that eleven of the 27 countries that we study are landlocked. The available empirical evidence suggests that being landlocked adds significantly to the cost of trading internationally (Raballand 2003; Cadot et al. 2006). Exporting firms rely not only on the quality of infrastructure provided by their home governments, but also on that of neighboring countries through which goods must transit. Because of this, the relationship between road quality and trade may not be entirely linear: for example, upgrades in important transit countries or resolution of regional bottlenecks could have impacts well beyond the individual countries concerned. Our results provide strong support for this view. An important aspect of our paper is its comparative outlook. We are primarily interested in identifying the intraregional trade benefits that can come from improved roads. We also want, however, to compare them with what is available from different policy approaches, such as tariff - 2 - reductions or more streamlined customs procedures (largely an issue of "trade facilitation" as reflected in World Trade Organization disciplines). The relative costs and benefits of these options are of direct interest to policymakers. At the same time, there is mounting empirical evidence that in the current environment of historically low tariff levels, "traditional" trade policy accounts for an increasingly small proportion of overall trade costs (Anderson and Van Wincoop 2004). This analysis suggests that the impact of policy reform in the areas of trade- related infrastructure and trade facilitation might be correspondingly greater. Our results support this conclusion, which is not surprising considering the current state of trade policy in Europe and Central Asia (relatively low traditional barriers, but high transport and transaction costs--see Tables 2-4). Moreover, initiatives to reduce trade costs in these ways are more attractive when tariff reforms are dependent upon on a complex, uncertain, and slow multilateral process. Against this background, the empirical literature has produced a number of model-based evaluations of the sensitivity of trade flows to infrastructure and trade facilitation. Studies of general scope such as Bougheas et al. (1999), Wilson et al. (2005) and Francois and Manchin (2006) use the gravity model to show that improvements along those dimensions are associated with increased trade flows. Another strand of the trade facilitation literature has emphasized the importance of time delays (Djankov et al. 2006; Nordås et al. 2006). Meanwhile, the results in Limão and Venables (2001) demonstrate that infrastructure plays an important role in determining transport costs, and thereby impacts trade flows. Nordås and Piermartini (2004) and Cadot et al. (2006) arrive at similar conclusions. In particular, the latter paper shows that the infrastructure of neighboring countries is important for landlocked Central Asian economies, due to transit effects. - 3 - Two recent papers deal more specifically with road infrastructure. The first of them, Coulibaly and Fontagné (2006), focuses on West Africa. The authors find that a composite measure of road quality in the importing and exporting countries has a positive and statistically significant effect on trade. Transit effects are also important, with the authors using a count of the number of borders crossed as a proxy. By contrast, Buys et al. (2006) examine road network quality across the whole of Sub-Saharan Africa (SSA). They use detailed road transport data to construct measures of international distance on an overland basis. They then build up a multi-dimensional measure of road quality, which is aggregated taking proper account of transit effects. Results from their gravity model show that network quality has a significant impact on intra-regional trade, while simulations suggest that the benefits of a road upgrade are very substantial: around $250bn over 15 years. The trade benefits far outweigh the costs, which include an initial investment of the order of $20bn and yearly maintenance of $1bn. Our paper builds on and extends this literature in five main ways. In terms of substance, our analysis incorporates a new ECA road distance database similar to the one used by Buys et al. (2006) in the SSA context (Section 2). It allows us to identify road transit routes in detail, and to construct bilateral road quality indicators based on actual distances traveled in the exporting, importing and transit countries (Section 3). Secondly, our data and gravity model (Sections 4-5) allow us to simulate the trade impacts of road upgrades in different countries, taking account of cross-border spillovers due to overland transit. Thirdly, our model includes policy variables covering applied tariffs and trade facilitation, in addition to road network quality. This enables a comparison of the relative impacts of different reform scenarios (Section 6). On the methodological front, we use the "theoretical" gravity model specification due to Anderson and Van Wincoop (2003, 2004). Fixed effects are used to take account of unobservable country- and - 4 - sector-specific factors, including multilateral resistance. Finally, to ensure that our results are robust to the presence of zero trade flows in the dataset, we use a variation on the Poisson Pseudo-Maximum Likelihood approach of Santos Silva and Tenreyro (2006), in addition to standard OLS. 2 Mapping Road Networks in Europe and Central Asia As defined in this paper, the ECA region covers 27 countries stretching from the Czech Republic in the West to Russia (Siberia) in the East, and from Turkmenistan in the South to the Baltic States and Russia in the North (Table 1 and Figure 1). Anecdotal evidence suggests that the road network in this area is extensive, but of variable quality. This observation is particularly true in those countries where the post-Communist transition has been long and difficult. (See Molnar and Ojala 2003; ADB 2006; and Cadot et al. 2006 for further details.) Computerized maps and spatial analysis software make it possible to develop a detailed picture of road transport routes in the ECA region, as Buys et al. (2006) have done for SSA. We construct minimum-distance routes connecting 138 cities, i.e. all regional cities with a year 2000 population of over 300 000 people. In all, we have 9453 inter-city routes along 2411 individual arcs (Figure 1). For each route, we are able to identify the exact road distance traveled in each of the sample countries. For instance, the minimum distance route from Prague to Moscow includes 128.6km of road travel in the Czech Republic, 723.6km in Poland, 547.2km in Belarus and finally 454.4km in Russia. We will be analyzing international trade data in the remainder of this paper. We need, therefore, to aggregate our intercity road distance data to the country level. To do that, we adopt the convention that the distance between two countries will be treated as the unweighted mean of the - 5 - minimum road distances between all cities for which we have data in those two countries. This is the same approach as in Buys et al. (2006).1 As a check on the reliability of our international distance measure, we compare it with the CEPII distance database (Mayer and Zignago 2006). The most commonly used CEPII measure expresses the distance between two countries as the great circle distance between their respective largest cities. Over the full sample, we find a correlation coefficient equal to 0.93. However, at long distances (greater than 3000km) the relationship between the two series is substantially weaker ( = 0.66). Visual evidence (Figure 2) suggests that the CEPII measure is systematically lower than ours over long bilateral distances. We take this as indicating that while the great circle approximation is reasonable for short overland distances, it loses much of its relevance as those distances grow. A detailed mapping is therefore particularly important for long international routes, such as those we are dealing with in the ECA context. One potential drawback with our approach--which might also explain part of the difference between our measures and CEPII's--stems from the considerable cross-country variation that our network map displays in terms of detail. The minimum population threshold that we have chosen results in 16 out of 27 countries being represented by a single city, while the largest country (Russia) is represented by 63 cities (Table 5). At first glance, it seems plausible that Russia might therefore exert an undue influence on our results. However, even with Russia excluded from the sample, the pattern identified above with respect to the CEPII distance measure still stands (full sample = 0.94; for distances greater than 3000km, = 0.62). Further, we find below that exclusion of Russia from the estimation sample in fact has little bearing on 1Weighting average distances by city population, or using the median rather than the mean, does not change our results. - 6 - the estimated coefficients. We are therefore confident that our population criterion, while necessarily arbitrary, strikes an acceptable balance between detail and tractability. 3 Measuring the Quality of Road Networks The literature cited above generally uses the percent of paved roads in a country as a proxy for road quality (Limão and Venables, 2001; Coulibaly and Fontagné, 2006; and Cadot et al., 2006). Buys et al. (2006), by contrast, also include GDP per capita and a corruption indicator.2 Such an approach is valuable insofar as it highlights the fact that upgrading road quality is not just about bitumen, but also requires maintenance capacity and the ability to control unofficial payments. However, aggregating all three dimensions into a single composite indicator makes it difficult to relate regression coefficients to specific policy actions. In particular, the relative importance of each dimension in relation to the others is determined by the researcher's priors, and not by the data themselves. The desire to map our results directly to policy space motivates our decision to limit our consideration of "quality" to the percent paved roads criterion. In constructing a road quality measure, we compare data from different sources to resolve (usually minor) disagreements amongst them. We have, however, found instances of apparently spurious time-series variation in percent paved roads data. Such instances appear to be related to definitional changes that can significantly alter the apparent percent of paved roads, even though no physical changes have in fact taken place. Table 6 provides our consolidation of the available data for ECA. In arriving at our preferred measures, we have been guided by expert opinion from 2The Buys et al. (2006) indicator Qj combines the percent of paved roads (Pj), per capita GDP (Gj) and the World Bank's Country Policy and Institutional Capacity Index (Cj) such that Qj = Pj Gj C j 1 2 3 . Imposing slightly increasing returns, the alpha parameters are set to 1=0.8, 2=0.2 and 3=0.2. - 7 - the World Bank. As a result we believe that our measures represent a reasonable approximation of the reality on the ground. To limit the risk of spurious time-series variation, we use a single base year only (2003). One important advantage of our road distance mapping approach is that it enables us to produce detailed measures of bilateral road quality that take full account of transit effects. We can consider road quality in third countries, not just the exporter and importer (cf. Nordås and Piermartini 2004). We can also use actual transit distances to weight road quality in each country along the route, rather than using an approximation such as a count of the number of border crossings (cf. Coulibaly and Fontagné 2006). As in Buys et al. (2006), we construct two measures of road quality. The first is a distance weighted average (paved_ave). We construct it using the percent of paved roads in the exporting and importing countries, as well as in all transit countries along the set of minimum distance routes used to calculate international distances as set out above. The paved roads data are weighted in each case by the proportion of the total distance traveled in each country. Our second measure (paved_min) is calculated using the same information, but taking the minimum percent of paved roads observed across the exporting, importing and transit countries. The purpose of this measure is to help identify bottlenecks and assess the potential for cross-country infrastructure spillovers. An incidental benefit of our approach is that our quality measures are unlikely to suffer unduly from endogeneity to bilateral trade flows, since they depend also on policy decisions by countries not involved in a given bilateral relationship. Interestingly, we find strong geographical concentration (65%) of paved_min scores in just three countries: Albania, Hungary and Romania. To the extent that paved_min is found to be a significant determinant of bilateral trade, this finding suggests that improved road quality in - 8 - those three countries is likely to impact a substantial number of third-country trade flows. Given their geography and current income levels, this would perhaps not be a surprising result. But it could be an important one in terms of regional investment initiatives, given that Hungary is an EU Member State and that Romania will accede on 1 January 2007. 4 Model Description and Data As noted at the outset, our primary interest is in assessing the impact of road quality on intraregional trade. We would also like to compare that impact to what can be had through tariff reductions and improved trade facilitation. To do that, we will use a standard tool of empirical international trade, namely the gravity model. Our specification is based on the micro-founded gravity model due to Anderson and Van Wincoop (2003, 2004): log Xij = log Ekj +logYi -logYk + 1-k )logtij - 1-k )log Pj - 1-k )logi +ij ( )k ( ) ( ) k ( ) ( ( ) ( k ( ) ( k ( ) k k (1) Where: Xij = Exports from country i to country j in sector k; Yi = Output of country i in sector k k k; Ekj = Expenditure of country j in sector k; Y k = Aggregate (world) output in sector k; k = Elasticity of substitution in sector k; tij = Trade costs facing exports from country i to country j k ( ) N in sector k; Pj ( ) k 1-k ( ) ( ) N = i 1-k 1-k 1-k k-1 i tij k k ; i k = Pj k-1kj tij k ; i = Country i's output k i=1 j=1 share in sector k; kj = Country j's expenditure share in sector k; ij = Random error term, k satisfying the usual assumptions. As is well known, the most important innovation of this model is its inclusion of the "resistance" or "remoteness" terms Pj and i . Inward resistance Pj captures the fact that j's imports from i k k k - 9 - depend on trade costs across all suppliers. Outward resistance i , by contrast, captures the k dependence of exports from i to j on trade costs across all importers. In applied work with (1), bilateral trade costs tij need to be specified in terms of observables. We k postulate: M tij = dij k (b ) zij k ,m ji m m=1 (2) log tij = log dij + log(bm)zij ( ) k ( ) M k,m m=1 Where: = elasticity of exports with respect to distance; dij = distance between countries i and j; bm = set of m constants; zij = set of observable bilateral determinants of trade costs. Combining (1) and (2) gives our baseline gravity model: log Xij ( )=log(E )+log(Y )-log(Y )+(1- ) ( )+log(b )z M k k k k k,m-... j i k logdij m ij m=1 (3) ...-(1-k )log Pj - 1-k )logi +ij ( ) ( k ( )k k The resistance terms Pj and i are not directly observable. We will account for them using k k fixed effects. In a panel data context, Baldwin and Taglioni (2006) have recently shown that proper specification of the fixed effects version of (3) has given rise to considerable confusion in the applied literature. When estimating over a single year, (3) should, strictly speaking, include fixed effects in the exporter-sector, importer-sector and sector dimensions. In addition, given that the elasticity of substitution k varies across sectors, it is necessary to allow the reduced form coefficients in the trade cost function to do likewise. A strict derivation therefore suggests the following estimable form for (3): - 10 - log Xij =i +j +k + 1 logdistij + 2 log(paved_ave)+ 3 log(paved_min)+... ( ) ( ) K K K k k k k k k k=1 k=1 k=1 (4) ...+ 4 log(1+tariff)+ 5 log(docs)+ 6border+ 7colony+ 8language+ij K K K K K k k k k k k k=1 k=1 k=1 k=1 k=1 Depending on the size of the dataset, it may prove difficult to identify all coefficients in (4) due in particular to lack of variation in the exporter-sector or importer-sector dimensions. We therefore propose a simplification in which trade cost elasticities are assumed to be constant across sectors, and country-sector fixed effects are taken to be subsumed by country fixed effects. The resulting estimation equation (5) uses fewer degrees of freedom than (4), but can be expected to provide a reasonable approximation in cases where cross-sectoral variation is not too strong. log Xij = c +i + j +k + 1 log distij + 2 log(paved _ave)+ 3 log(paved _min)+... ( ) k ( ) ...+ 4 log(1+tariff )+ 5 log(docs)+ 6border+ 7colony+ 8language+ij (5) k Our data and sources are set out in detail in Table 8. For bilateral trade, we use the value of 2003 imports by BEC 1-digit sector, taken from the WITS database. Whenever import data are missing, we use export (mirror) data. Trade cost dummies based on geographical and historical factors (contiguity, colonization and common language) are drawn from the CEPII distance database (Mayer and Zignago 2006). Distance is measured using average intercity road distances obtained by computer mapping, as set out above. Paved_ave and Paved_min refer to our measures of average and minimum road quality. Our tariff variable is drawn from effective applied tariffs as recorded in the WITS-TRAINS database.3 For trade facilitation, we use data 3We use the simple average tariff when aggregating to the BEC 1-digit level. Use of the trade weighted average does not affect our results. - 11 - from the 2006 Doing Business Report (World Bank, 2006) on the number of documents required to export and import (docs), summing across the exporting and importing countries.4 We prefer this measure to the more commonly used indicator of time to export and import (Djankov et al. 2006; Nordås et al. 2006) because it does not suffer from endogeneity to trade flows in the same way: while an unexpectedly large trade flow this year might lead to congestion and thereby increase trading time, the same is not true for the number of documents required by customs authorities.5 Another appealing feature of our measure is that it bears a close relationship to the core interpretation given to the term "trade facilitation" at the WTO level, which emphasizes the streamlining of trade-related administrative procedures and formalities (Wilson 2005). 5 Estimation Results We use (5) as our baseline gravity model for estimation purposes. We first estimate using standard OLS. Next, we present the results of a number of alternative estimation methods drawn from the recent literature.6 In all cases, we gauge robustness by considering three different specifications of (5). Model 1 is our preferred specification, and takes the exact form of (5) above. Model 2 drops paved_min and uses paved_ave only, while Model 3 does the opposite. Finally, we briefly discuss some additional robustness checks. The first three columns of Table 9 contain our OLS estimates. The dependent variable is log(trade), with zero or missing observations simply dropped from the dataset; this is an issue we 4Due to lack of data availability, we use customs formalities in 2005 as a proxy for 2003. Similarly, when TRAINS data are missing for a given year, we take the most recent available data prior to 2003. 5Our argument is indirectly supported by the results of Djankov et al. (2006), who instrument for trading time using two related variables, namely the number of signatures required to export and import. Their model is overidentified and does not reject the relevant restriction, which suggests that the proposed instruments are valid. 6All calculations were performed using Stata 9.2SE. - 12 - return to below. Models 1, 2 and 3 all perform well, with R2 of around 62%. All estimated coefficients carry the expected signs and have economically reasonable magnitudes. Except in the case of paved_ave, coefficient estimates are quite stable across specifications. The distance elasticity is greater than 2 in absolute value in all three models, which is stronger than the central tendency of the gravity literature (around 0.9 according to the meta-analysis of Disdier and Head 2005). We put the difference down to three factors. Firstly, we use overland distances and not the more common great circle measures. Secondly, our data are disaggregated at the sectoral level, whereas many gravity models use total trade. Thirdly, our sample covers just one geographic region, in which it is conceivable that road distance plays a particularly important role, for the reasons set out above. Only distance, paved_min, tariffs and common language are statistically significant at conventional levels. A 1% improvement in paved_min is associated with a 0.6% increase in trade, while a 1% cut in applied tariffs increases trade by 3.5% (evaluated at the approximate sample mean of 8%). Even though the estimated coefficients for paved_ave and docs are not statistically significant, we still regard them as economically significant: a 1% improvement in average road quality is associated with a 0.2% to 0.6% increase in trade, while a similar percentage reduction in the number of export and import documents is associated with a 2.4% to 3.2% increase in trade. Based on OLS estimates, we tentatively conclude that improved road quality, lower tariffs and better trade facilitation are all associated with stronger bilateral trade flows. To be sure that this result holds, we need to deal more carefully with the issue of zero or missing trade flows. In our case, around 1500 observations (nearly one-third of the potential dataset) fall into that category. Unfortunately, we are also missing data for applied tariffs and trade - 13 - facilitation. This means that in terms of our effective sample, i.e. the number of observations for which data are available across all variables, the zero problem in fact only affects 159 observations. Although this is just 6% or so of the effective sample, we still believe it is important to ensure that our results are robust in this sense. Our approach to this problem follows Santos Silva and Tenreyro (2006).7 First, we express (5) in non-linear form (i.e., prior to taking logarithms): X0ij = exp...+ ( ) k c +i + j +k + 1 log distij + 2 log(paved _ ave)+ 3 log(paved _ min)+... 4 log(1+tariff )+5log(docs)+6border+7colony+8language +ij (6) k The notation X 0ij is designed to emphasize the fact that the trade flow data include zeros. The k essential difference between (5) and (6) relates to the error term, which we have relabeled ij in k (6). Equation (5) assumes that the error is additive in a log-linear specification, or alternately that exp ij is multiplicative in the original non-linear specification. On the other hand, equation (6) ( )k more naturally assumes that the error is additive in the original non-linear specification. If (6) represents the "true" model, then the OLS estimator derived by log-linearization will generally be inconsistent (see Santos Silva and Tenreyro 2006 for details). The first order conditions for estimation of (6) by weighted nonlinear least squares are identical to those for maximum likelihood estimation using the Poisson model for count data (Gourieroux et al. 1984; Davidson and MacKinnon 2004). Estimated coefficients from Poisson can still be 7The most common alternative is the Heckman sample selection model (e.g., Francois and Manchin 2006; Helpman et al. 2006). However, reliance only on the non-linearity of the inverse Mills ratio for identification implies a strong distributional assumption that may often be rejected in practice (Davidson and MacKinnon 2004, 488-489). At the same time, overidentification via variable exclusion has tended to rely on unconvincing assumptions. - 14 - interpreted as elasticities, as under log-linearized OLS. Expressing (6) in matrix form as Y = XB + w (dropping subscripts), the first order conditions for Poisson take the form: ((Y -eXB X = 0 ) ) (7) Continuing with the above notation, the weights applied in terms of nonlinear least squares - XB estimation can be seen to be e 2 .8 Although the Poisson estimator can still be consistent under alternative distributional assumptions (i.e., it is a pseudo-maximum likelihood estimator), we also apply a Negative Binomial estimator (NB2 in the terminology of Cameron and Trivedi 2001) that allows for overdispersion in the data and may therefore provide a better fit in this case. First order conditions for maximum likelihood estimation of the NB2 model are again equivalent to weighted nonlinear least squares estimation of (6), though with different weights from Poisson: 1+e (Y -e )X XB = 0 (8) XB The NB2 model puts less weight on large observations than does Poisson, the difference between the two depending on the size of the NB2 overdispersion parameter (estimated from the data). In the limiting case of no overdispersion (i.e., = 0), the NB2 model collapses to Poisson and both estimation methods therefore apply the same weighting system.9 8 The first order conditions for unweighted nonlinear least squares estimation of (6) take the form ((Y -eXB XeXB = 0. ) ) 9Cravino et al. (2006) and Soloaga et al. (2006) also use the Negative Binomial model to estimate gravity models of foreign investment and trade flows respectively. In their Monte Carlo simulations, Santos Silva and Tenreyro (2006) consider the closely related Gamma model, which has first order conditions equivalent to weighted nonlinear least - 15 - Columns 4-6 of Table 9 report estimates using the Poisson model, and columns 7-9 show results for the NB2 model. The overdispersion parameter is estimated in all three NB2 models to be 3.0. A likelihood ratio test of the (unrestricted) NB2 model against the (restricted) Poisson strongly rejects the null for all three specifications (2 = 9.3e10 for models 1 and 2, 2 = 9.5e10 for model 3, prob. = 0.00). We conclude that the NB2 model is to be preferred over Poisson, and our discussion of results will therefore focus on the former. As for the OLS case presented above, we find that coefficient estimates (except for colonization) have the expected signs and economically sensible magnitudes.10 The coefficient on distance has fallen somewhat in absolute value--an effect also noted by Santos Silva and Tenreyro (2006)-- and all policy variables are now significant at the 10% level. The coeffcients on tariffs and export/import documents are respectively a little weaker and stronger than those obtained with OLS. While paved_min enters with almost the same elasticity as under OLS, paved_ave is considerably stronger. In general, coefficient estimates are quite stable across specifications, although paved_ave and docs exhibit some variance according to the presence or absence of other variables. In terms of magnitude, we find that 1% improvements in paved_ave and paved_min are associated with trade increases of 0.8% and 0.6% respectively. By comparison, 1% reductions in tariffs and export/import documents are associated with trade increases of 3.0% and 4.0% respectively (evaluated at the sample mean for tariffs). squares with weights e-XB (Gourieroux et al. 1984). The effect is similar to NB2, in the sense that it downweights large observations compared with Poisson. 10An additional reason for preferring NB2 estimates to Poisson is that the latter suggests that docs has a positive but statistically insignificant impact on trade flows. This is highly counterintuitive, and against the existing evidence using closely related variables (e.g., Djankov et al. 2006). - 16 - In Table 10, we present the results of additional robustness checks using our preferred NB2 model.11 Exclusion of Russia from the sample (column 2) makes no significant difference to our results. Bootstrapping (column 1) results in larger standard errors than using asymptotic results. Our variables paved_ave and docs are no longer statistically significant at the 10% level, but the other variables of interest remain significant. Specifying fixed effects in line with (4), i.e. by importer-sector, exporter-sector and sector, but imposing constant slope coefficients across sectors, results in a lower (and statistically insignificant) coefficient on paved_ave, but does not result in large changes in the other parameters (column 4). Finally, adding country-pair random effects to (5) to account for possible omission of bilateral trade determinants (cf. Carrère 2006) leads to paved_ave and docs entering with unexpected signs, although both estimated coefficients are statistically insignificant (column 3).12 Paved_min remains significant at the 5% level, although its magnitude is reduced to 0.25. Overall, we conclude that our main findings are robust, in particular the importance of road infrastructure bottlenecks as captured by paved_min. Finally, we present results disaggregated by BEC sectors 1-6 (Table 11).13 As expected, coefficient estimates vary considerably across sectors. This is due to two factors. Firstly, it follows from (3) that the reduced form parameters in (4), (5) and (6) will vary to the extent that the elasticity of substitution varies across sectors. Secondly, trade flows in different sectors may themselves be more or less sensitive to particular factors, due to certain product characteristics such as unit value, perishability and bulk. 11An appendix (available on request) provides additional specification and robustness checks. 12 This model converged extremely slowly under BFGS optimization, and required the default tolerances to be relaxed slightly. Combined with the unexpected signs referred to in the text, this suggests that great caution should be exercised in interpreting these results. 13Estimates obtained in this way are equivalent to pooled estimates of (4). We exclude BEC sector 7 (other goods), since the products it groups together are too heterogeneous to allow the drawing of meaningful conclusions. - 17 - In terms of our road quality variables, Table 11 shows that the impact of paved_min is uniformly positive across sectors and is of comparable magnitude to our core elasticity estimate of 0.6 (Table 9 column 7). Paved_ave, on the other hand, displays much greater variation, and is not significant in most cases. Taking results for the two coefficients together, we conclude that trade flows in food, fuel, capital goods and transport equipment are particularly sensitive to road network quality, whereas industrial supplies and consumer goods are less so. Industrial supplies and capital goods are particularly sensitive to tariffs, while improved trade facilitation seems to be relatively important for food, transport equipment and consumer goods. In sum, the above results disclose strong and consistent evidence to the effect that road network quality affects intraregional trade in ECA. We find that both the average and minimum levels of quality across transit countries are important, but our result is clearest in terms of the latter. We interpret this as suggesting that bottleneck effects and, by corollary, cross-country spillovers are important factors in determining intraregional trade. These results withstand numerous robustness checks, as well as estimation by individual sector (subject to cross-sectoral differences in the estimated elasticities of most variables). 6 Policy Simulations For the remainder of this paper, we will concentrate on the NB2 results reported in column 7 of Table 9. Under this specification, 1% improvements in paved_ave and paved_min are associated with trade increases of 0.8% and 0.6% respectively. By comparison, 1% reductions in tariffs and export/import documents are associated with 3.0% and 4.0% trade increases respectively (evaluated at the sample mean for tariffs). - 18 - The above policy elasticities should be considered with caution. The marginal impact of tariff reductions and trade facilitation would appear to be much stronger than for improved roads, since the estimated elasticities are greater in absolute value. Analysis of the standardized coefficients corresponding to the OLS estimates in column 1 tells a similar story: one standard deviation reductions in tariffs and export/import documents are associated with 0.1% and 0.2% increases in trade respectively, while similar innovations in paved_ave and paved_min are associated with 0.01% and 0.07% changes in trade. However, it would be unwise to draw policy conclusions from such an analysis. As is clear from the way the two paved roads variables are constructed, the relationship between bilateral road quality and the percent of paved roads in any given country is a complex one. It depends on overland transit routes, and on the quality of road infrastructure in transit countries. More important than the results from simply shocking either aggregate variable, is to trace through the impact of a change in individual country indices, allowing for transit effects. We deal with these difficulties through simulations (cf. Wilson et al. 2005). We define four counterfactuals in terms of particular changes to national policies. Following this, we use our gravity model elasticities to estimate the resulting change in intraregional trade flows. We are conscious of the limits of this approach, in particular to the extent that it assumes parameter constancy across policy shifts and treats each policy change in isolation from the others. Our simulation results should therefore be taken as indicative of the orders of magnitude involved only. Given the scope of this research, our simulation results do not measure economic welfare, but focus exclusively on projected trade impacts. Nonetheless, comparison of results across simulations is likely to prove a useful tool in assessing different policy options. We therefore identify two initial policy simulations: - 19 - I. Road networks in all ECA countries are upgraded to the regional mean level of quality, namely 74.52% of paved roads; and II. Road networks in Albania, Hungary and Romania only are upgraded as in I. The motivation for these simulations is that raising each country's level of road network quality to the currently prevailing regional average represents an ambitious but feasible scenario. Concretely, this means that under Simulation I, 13 out of 27 ECA countries receive an upgrade, while under Simulation II it is limited to only the three countries identified above as connected with 65% of minimum quality routes in the region. We conduct the simulations as follows. First, we set up the policy shock by recalculating both weighted average and minimum quality measures for all inter-country routes, in exactly the same way as described in Section 3. The only difference is that country scores below the regional average are increased to that level before recalculation. Next, we calculate the resulting percentage changes in paved_ave and paved_min. Using our trade data and our estimated elasticities (0.79 and 0.60 respectively), we then map these policy shocks to changes in bilateral trade values. Finally, we sum estimated bilateral trade impacts across countries to give the estimated overall increase in intraregional trade. Results in Table 12 show that the potential trade gains from an ambitious but feasible program of road upgrades are large in absolute terms. It does not seem unreasonable to consider positive impacts of the order of 50% of baseline trade, or just over US$55 billion based on total intra- regional trade in 2003. These figures are based exclusively on the projected increase in intra- regional trade. They do not include any flow-on effects to extra-regional trade. We therefore consider that our results lie towards the lower bound of expected total trade benefits from a road network upgrade. - 20 - A comparison of results from Simulations I and II also makes clear the crucial role played by just three countries in driving the above estimates. Focusing a road upgrading program of similar ambition on Albania, Hungary and Romania could bring intra-regional trade benefits equal to over 50% of those projected from the region-wide program in Simulation I. That cross-country infrastructure spillovers are important is demonstrated by Figures 3 and 4, which show the distribution of export and import changes by country for both simulations. Given the significant cost reduction likely to result from focusing infrastructure investments on three countries rather than 13--a point to which we return below--the expected return on investment from such a focused program is likely to be impressive from a regional point of view. To provide a comparative context for the above results, we also conduct simulations designed to assess the projected trade impacts of policy changes affecting applied tariffs and trade facilitation (export/import documents): III. Applied tariffs in all ECA countries are cut such that no tariff above the regional mean of 8% ad valorem is applied; and IV. The numbers of documents required to export and import are reduced to their regional means, namely 8 and 12 respectively. We regard these counterfactuals as representing ambitious but feasible reform programs in terms of tariffs and trade facilitation. On a substantive level (if not a formal one), we consider them to be comparable in scope to the road network upgrade analyzed above. Results for both simulations are again reported in Table 12. It is notable that the increases in intra-regional trade associated with region-wide improvements in both traditional and "new" trade policies are considerably lower than for a road upgrade program conducted on a - 21 - comparable scale. Trade flow changes from the tariff scenario are in the region of 6%--nearly an order of magnitude smaller than the trade increases that flow from an infrastructure upgrade.14 The impact of trade facilitation measures is, however, considerably stronger than for a tariff reduction, of the order of 20% of baseline trade. It is still small relative to the gains from a road upgrade, even if it is focused on just three countries. 6.1 The Cost Dimension The policy simulations discussed here focus exclusively on the intraregional trade benefits that could be expected from the different policy options under consideration. However, in order to make a balanced assessment of those options, we also need information on costs. This is particularly true when one of the options--an infrastructure upgrade--has much higher direct costs than do the others.15 Our purpose here is not to provide a detailed cost breakdown of the type that would be required before undertaking a specific road upgrade project. Our analysis has taken place at a higher level of generality, and in particular has not considered the state of individual road arcs. For that reason, our assessment of the costs will focus on producing a general estimate only (cf. the more detailed approach of Buys et al. 2006). 14Since we do not have data on quantitative restrictions and other measures that might restrain trade following a tariff cut, we would argue that this estimate is, if anything, on the high side. 15Both tariff reductions and trade facilitation also involve costs. In a direct sense, they are likely to be quite limited. Indirectly, or in a political economy sense, they may well be substantial from the point of view of individual actors. The political economy of reform affecting infrastructure, tariffs and procedural barriers is an area that would benefit from increased attention in the future, although it is outside the scope of this paper. - 22 - The World Bank's ROad Costs Knowledge System (ROCKS) provides the starting point for our analysis.16 ROCKS is a standardized database of costs associated with various types of road works. It classifies individual projects by country and type of work, and allows the user to obtain cost per km information in a common (real) currency. Most database entries also include extensive additional information as to the tasks performed, as well as geographical conditions that can be expected to affect costs. Since we do not have information on the exact work that would need to be performed on each road arc in order to bring it up to the level of quality assumed in our counterfactuals, we simply assume that all arcs in countries undergoing an upgrade would require "development" or "reconstruction" work in terms of the ROCKS classification. This includes partial and full widening and/or reconstruction work, along with improvements to the road surface. The types of work that we are considering lie towards the high end of the full range of unit costs in ROCKS (excluding those relating to entirely new construction projects). Table 13 provides US$ per km cost data from ROCKS based on the types of work we have identified. We only take account of actual, incurred costs (not estimates), and focus on those from Eastern Europe and the Former USSR; the Western Europe and World cost columns are provided for reference only.17 Based on this data, the range of expected unit costs for the countries under consideration here runs from around $36,000 per km to $666,000 per km, with an average of approximately $269,000 per km. We use these baselines to provide low, average and high cost estimates for the road upgrades implied by our Simulations I and II. We take the length of road to be upgraded in each country as the total length of arcs passing through that 16ROCKS can be downloaded from http://www.worldbank.org/transport/roads/rd_tools/rocks_main.htm. 17We have also eliminated two outlying observations, with unit costs around double the next highest data point. - 23 - country as per our computerized map. In other words, we do not calculate the cost of upgrading the entire road network in each country, but only those parts of it connecting cities with year 2000 population above 300,000 people. Results are presented in Table 14. Since our range of unit costs covers a wide variety of work types, the total cost estimates cover a correspondingly broad range. Focusing on mean unit costs for the region, we can see that a full upgrade (i.e., 13 countries) would involve a total up front cost of the order of US$8 billion. By contrast, focusing on three countries only would reduce that cost very considerably, to just over US$3 billion. Comparing these numbers with Table 12 shows that even once the costs of an upgrade are netted out, the trade benefits to the region from a road upgrade are very substantial: of the order of $45 billion for a region wide program, and $30 billion for a three country program, without allowing for any amortization of the cost of the upgraded road network over its expected lifespan.18 7 Conclusions and Directions for Further Research In this paper, we have built on and extended recent work by Buys et al. (2006) to show that an ambitious but feasible road upgrade program in ECA has great potential to boost intra-regional trade--by as much as 50%. Moreover, it is possible for the region to reap a large proportion of the overall gains by focusing attention on just three countries which are important transit corridors but exhibit significant limitations in terms of infrastructure quality: Albania, Hungary and Romania. Such a concentrated program of road upgrading would come at significantly 18Even using the extreme upper tail of ROCKS unit costs results in net benefits of around $35 billion and $27 billion for the 13 and 3 country programs respectively. - 24 - reduced cost (perhaps 40%) compared with attaining the same level of road quality on a region- wide basis, yet would bring around 60% of the total expected trade benefits. The results we have presented suggest a number of considerations for policy in this area. Firstly, road quality and infrastructure clearly matter for trade in the ECA region. In quantitative terms, our simulation results suggest that a feasible but ambitious scenario of road upgrading is likely to bring greater intraregional trade benefits than comparable actions affecting either tariffs or customs procedures. In any case, the combined impact of upgrading road network quality and improving trade facilitation appears likely to produce gains well in excess of those that could be expected from comparable tariff reductions. This result aligns well with the recent literature on trade facilitation using CGE models, which suggests that the expected gains from such measures may indeed be of greater quantitative significance than those from liberalization of "traditional" trade policy measures (e.g., Francois et al. 2005). It is also consistent with other recent work that has shown the importance of transit country conditions, in particular in the Central Asia region (Cadot et al. 2006). A second important policy implication is that once transit is taken into account, infrastructure projects can have important intraregional spillovers. This dynamic does not generally operate in the same way for traditional trade policy measures, such as tariff cuts. Spillovers therefore need to be taken into account when assessing costs and benefits of various options for trade facilitation and development assistance strategies. They may support an argument for regional coordination and shared funding responsibilities for infrastructure projects (see Schiff and Winters 2002 for a review of the issues involved). In the present case, that suggestion takes on particular importance in light of the fact that Hungary is now a member of the EU, while Romania is soon to be such. Future allocation of EU funding could benefit from taking account - 25 - of potential trade impacts not only on a national level, but also through those countries' regional links. A final policy message to highlight given our results is that the trade benefits from infrastructure upgrades can be obtained by countries acting unilaterally, or through regional cooperation. As is the case for many policy measures under the broad heading of trade facilitation, it is not necessary to wait for multilateral agreement before taking action to bring about greater integration into the trading system. Indeed, national and regional trade facilitation programs sponsored by the World Bank, regional development banks, bilateral donors, and public-private partnerships, for example, could be seen as important ways in which countries and regions can position themselves so as to reap maximum benefit from future rounds of multilateral liberalization. While our results are highly suggestive in policy terms, there nonetheless remain a number of important research questions to be considered in future work. The trade facilitation literature has shown that according to country circumstances, the various modes of transport--road, rail, sea and air--can all be important determinants of trade performance (e.g., Wilson et al. 2005). Future research could usefully focus on the relative benefits and costs of upgrading infrastructure quality for each mode. As there is likely to be considerable variance in results across countries, regions and even sectors, it will be necessary to take a detailed approach to these questions, including through an attempt to account for the interactions amongst the different modes. Our paper has focused exclusively on intra-regional trade. It will be important in future work to consider in addition the impacts of infrastructure upgrades on extra-regional trade. To do this, it will be necessary to compile a detailed dataset that interfaces road and international air or sea routes, taking account of the location of principal sea and air ports. It will also be important to - 26 - take account of possible trade creation or diversion effects. By helping move towards a more complete picture of the benefits of infrastructure upgrades, such an exercise would provide important additional information for policymakers. Finally, the available data have not allowed us to pay detailed attention to the state of upkeep of particular road links. We have had to rely on national aggregates in assessing the extent to which network quality matters for trade. The flipside of this is that our cost estimate does not take account of the detailed work needed as part of a concrete upgrade program. There is thus considerable scope for additional work on specific cost-benefit analyses to be undertaken in this area. - 27 - References Anderson, James E. and Eric Van Wincoop, 2003, "Gravity with Gravitas: A Solution to the Border Puzzle", The American Economic Review, 93(1), 170-192. Anderson, James E. and Eric Van Wincoop, 2004, "Trade Costs", Journal of Economic Literature, 42(3), 691-751. Asian Development Bank (ADB), 2006, Central Asia: Increasing Gains from Trade Through Regional Cooperation in Trade Policy, Transport and Customs Transit, Manila: ADB. Baldwin, Richard E. and Daria Taglioni, 2006, "Gravity for Dummies and Dummies for Gravity Equations", Working Paper No. 12516, NBER. Bougheas, Spiros, Panicos O. Demetriades and Edgar L. W. Morgenroth, 1999, "Infrastructure, Transport Costs and Trade", Journal of International Economics, 47, 169-189. Buys, Piet; Uwe Diechmann and David Wheeler, 2006, "Road Network Upgrading and Overland Trade Expansion in Sub-Saharan Africa", The World Bank, mimeo. Cadot, Olivier, Céline Carrère and Christopher Grigoriou, 2006, "Landlockedness, Infrastructure and Trade in Central Asia", Mimeo. Cameron, A. Colin and Pravin K. Trivedi, 2001, "Essentials of Count Data Regression" in Badi H. Baltagi (ed.), A Companion to Theoretical Econometrics, Malden, Ma.: Blackwell. Carrère, Céline, 2006, "Revisiting the Effects of Regional Trade Agreements on Trade Flows with Proper Specification of the Gravity Model", European Economic Review, 50, 223-247. Coulibaly, Souleymane and Lionel Fontagné, 2006, "South-South Trade: Geography Matters", Journal of African Economies, 15(2), 313-341. - 28 - Cravino, Javier, Daniel Lederman and Marcelo Olarreaga, 2006, "Substitution Between Foreign Capital in China, India, the Rest of the World and Latin America: Much Ado About Nothing?", The World Bank, mimeo. Davidson, Russell and James G. MacKinnon, 2004, Econometric Theory and Methods, New York, N.Y.: Oxford University Press. Disdier, Anne-Celia, and Keith Head, 2005, "The Puzzling Persistence of the Distance Effect on Bilateral Trade", Mimeo. Djankov, Simeon, Caroline Freund and Cong S. Pham, 2006, "Trading on Time", Policy Research Working Paper No. 3909, The World Bank. Francois, Joseph, Hans van Meijl and Frank van Tongeren, 2005, "Trade Liberalization in the Doha Development Round", Economic Policy, April, 349-391. Francois, Joseph and Miriam Manchin, 2006, "Institutional Quality, Infrastructure and the Propensity to Export", Mimeo. Helpman, Elhanan, Marc Melitz and Yona Rubinstein, 2006, "Trading Partners and Trading Volumes", Mimeo. Limão, Nuno and Anthony Venables, 2001, "Infrastructure, Geographical Disadvantage, Transport Costs and Trade", World Bank Economic Review, 15, 451-479. Mayer, Thierry and Soledad Zignago, 2006, "Notes on CEPII's Distance Measures", CEPII, Mimeo. Molnar, Eva and Lauri Ojala, 2003, "Transport and Trade Facilitation Issues in the CIS 7, Kazakhstan and Turkmenistan", Paper prepared for the Lucerne Conference of the CIS-7 Initiative, 20th to 22nd January 2003, Mimeo. - 29 - Nordås, Hildegunn K. and Roberta Piermartini, 2004, "Infrastructure and Trade", Staff Working Paper No. ERSD-2004-04, WTO. Nordås, Hildegunn K., Enrico Pinali and Massimo Geloso Grosso, 2006, "Logistics and Time as a Trade Barrier", Working Paper No. TD/TC/WP(2006)3/FINAL, OECD. Raballand, Gael, 2003, "Determinants of the Negative Impact of Being Landlocked on Trade: An Empirical Investigation Through the Central Asian Case", Comparative Economic Studies, 45, 520-536. Santos Silva, J.M.C. and Silvana Tenreyro, 2006, "The Log of Gravity", Review of Economics and Statistics, 88(4), 641-658. Soloaga, Isidro, John S. Wilson and Alejandro Mejia, 2006, "Moving Forward Faster: Trade Facilitation Reform and Mexican Competitiveness", Policy Research Working Paper No. 3953, The World Bank. Wilson, John S., Catherine L. Mann and Tsunehiro Otsuki, 2005, "Assessing the Benefits of Trade Facilitation: A Global Perspective", The World Economy, 28(6), 841-871. World Bank, 2006, Doing Business in 2006: Creating Jobs, Washington, D.C.: World Bank. World Bank, 2007, Doing Business in 2007: How to Reform, Washington, D.C.: World Bank. - 30 - Tables Table 1: Merchandise exports as a percentage of GDP, 1995-2004. (Source: World Development Indicators.) Country or Region 1995 2000 2004 Albania 41.37 36.60 37.73 Armenia 64.36 61.52 65.71 Azerbaijan 47.17 55.32 83.67 Belarus 74.20 125.40 131.50 Bosnia and Herzegovina 66.11 92.24 90.19 Bulgaria 84.04 89.74 100.88 Croatia 64.55 66.85 71.71 Czech Republic 84.00 109.81 129.11 Estonia 101.28 162.29 130.56 Georgia 20.05 32.09 47.98 Hungary 63.44 128.47 113.41 Kazakhstan 44.45 75.73 80.69 Kyrgyz Republic 56.05 77.32 75.28 Latvia 59.59 65.35 80.73 Lithuania 84.66 81.42 96.81 Macedonia, FYR 65.69 94.90 84.70 Moldova 90.42 96.94 106.36 Poland 38.21 48.39 67.70 Romania 51.27 63.21 76.74 Russian Federation 35.91 57.84 48.11 Serbia and Montenegro 7.82 63.17 65.56 Slovak Republic 89.42 122.22 138.75 Slovenia 88.93 98.85 102.63 Tajikistan 126.67 148.98 110.46 Turkmenistan 132.47 150.45 116.59 Ukraine 59.34 91.26 95.13 Uzbekistan 46.29 40.07 64.15 World 35.30 41.32 44.87 Low income 31.84 33.30 37.79 Lower middle income 36.84 44.96 57.51 Upper middle income 47.86 58.85 67.04 High income 34.25 39.64 41.51 - 31 - Table 2: Delays at export and import (days) and cost to export and import (USD). (Source: World Bank, 2005 & 2006.) 2005 2006 Country or Region Export Time Import Time Export Time Import Time Export Cost Import Cost Albania 37 38 34 34 818 820 Armenia 34 37 34 37 1,600 1,750 Azerbaijan 69 79 69 79 2,275 2,575 Bulgaria 26 24 26 25 1,233 1,201 Bosnia and Herzegovina 32 43 22 25 1,150 1,150 Belarus 33 37 33 36 1,472 1,472 Czech Republic 20 22 20 22 713 833 Estonia 12 14 3 5 640 640 Georgia 54 52 13 15 1,370 1,370 Croatia 35 37 26 18 1,250 1,250 Hungary 23 24 23 24 922 1,137 Kazakhstan 93 87 93 87 2,780 2,880 Kyrgyz Republic NA 127 NA 127 NA 3,032 Lithuania 6 17 6 17 704 782 Latvia 18 21 11 12 965 965 Moldova 33 35 33 35 1,185 1,285 Macedonia, FYR 32 35 32 35 1,070 1,070 Poland 19 26 19 26 2,260 2,260 Romania 27 28 14 14 1,300 1,200 Russia 29 35 39 38 2,237 2,237 Serbia and Montenegro 32 44 11 12 1,240 1,440 Slovak Republic 20 21 20 21 1,015 1,050 Slovenia 20 24 20 24 1,070 1,107 Tajikistan NA NA 72 44 4,300 3,550 Turkmenistan NA NA NA NA NA NA Ukraine 34 46 33 46 1,009 1,025 Uzbekistan NA 139 44 139 2,550 3,970 Europe & Central Asia 31.6 43 29.2 37.1 1,450.20 1,589.30 East Asia & Pacific 25.8 28.6 23.9 25.9 884.8 1,037.10 Latin America & Caribbean 30.3 37 22.2 27.9 1,067.50 1,225.50 Middle East & North Africa 33.6 41.9 27.1 35.4 923.9 1,182.80 32 2005 2006 Country or Region Export Time Import Time Export Time Import Time Export Cost Import Cost OECD 12.6 14 10.5 12.2 811 882.6 South Asia 33.7 46.5 34.4 41.5 1,236.00 1,494.90 Sub-Saharan Africa 48.6 60.5 40 51.5 1,561.10 1,946.90 33 Table 3: Overall Trade Restrictiveness Index for ECA countries (% ad valorem equivalent). Source: Kee et al. (2006). Country or Region OTRI - Tariffs OTRI - Tariffs & NTBs Albania 10.9 11.4 Belarus 9.1 15.9 Czech Republic 4.0 5.0 Estonia 1.1 2.3 Hungary 6.1 11.3 Kazakhstan 5.4 14.0 Kyrgyz Republic 6.9 7.4 Lithuania 2.0 5.0 Latvia 3.0 9.8 Moldova 4.7 7.4 Poland 10.8 15.2 Romania 11.9 15.8 Russia 10.4 22.6 Slovenia 9.8 18.2 Ukraine 9.3 21.6 ECA Average 7.0 12.2 OECD Average 5.6 11.4 a. The OTRI represents the uniform tariff required in each country to achieve an equivalent level of total imports as under current policy settings. Table 4: Market Access Overall Trade Restrictiveness Index for ECA countries (% ad valorem equivalent). Source: Kee et al. (2006). Country or Region MA-OTRI - Tariffs MA-OTRI - Tariffs & NTBs Albania 11.3 16.7 Belarus 9.8 15.4 Czech 6.2 10.7 Estonia 9.3 15.3 Hungary 7.6 13.3 Kazakhstan 5.7 15.3 Kyrgyzstan 11.8 19.2 Latvia 10.8 20.0 Lithuania 14.5 23.0 Moldova 17.1 25.9 Poland 8.2 13.8 Romania 8.8 15.7 Russia 4.3 12.2 Slovenia 8.0 13.9 Ukraine 7.1 15.2 ECA Average 9.4 16.4 OECD Average 7.0 13.1 a. The MA-OTRI represents the uniform tariff required in the rest of the world in order to achieve an equivalent level of total exports from a given country as under current policy settings. 34 Table 5: Breakdown of cities included in the ECA road network. Country No. of Cities > 300 000 ALB 1 ARM 1 AZE 2 BGR 3 BIH 1 BLR 5 CZE 3 EST 1 GEO 1 HRV 1 HUN 1 KAZ 7 KGZ 1 LTU 2 LVA 1 MDA 1 MKD 1 POL 10 ROM 6 RUS 63 SRB 1 SVK 1 SVN 1 TJK 1 TKM 1 UKR 18 UZB 3 35 Table 6: Comparison of percentage paved roads data. (Sources: World Road Statistics, World Development Indicators, CIA World Fact Book online). Country WRS Year WDI Year CIA Year Preferred Albania 39 2002 39 2002 39 2002 39 Armenia 100 2003 97 1998 100 2003 97 Azerbaijan 49 2004 47 2003 47 2003 47 Belarus 87 2004 100 2003 100 2003 87 Bosnia and Herzegovina 52 1999 52 1999 52 2005 52 Bulgaria 99 2004 92 2002 92 2003 92 Croatia NA NA 85 1999 85 2004 85 Czech Republic 100 2003 100 2002 100 2003 100 Estonia 24 2004 23 2003 23 2003 23 Georgia 39 2004 39 2003 39 2003 39 Hungary 44 2003 44 2002 44 2005 44 Kazakhstan 93 2004 96 2003 96 2003 96 Kyrgyz Republic 91 1999 90 2004 91 1999 90 Latvia 100 2004 100 2003 100 2003 100 Lithuania 28 2004 27 2003 89 2003 89 Macedonia, FYR 64 1999 64 1999 64 1999 64 Moldova 86 2004 86 2003 86 2003 86 Poland 70 2003 70 2003 70 2003 70 Romania 30 2004 50 2002 30 2003 30 Russian Federation NA 2001 67 1999 85 2004 85 Serbia and Montenegro 96 2004 62 2002 62 2002 62 Slovak Republic 87 2004 87 2003 87 2003 87 Slovenia 100 2004 100 2003 100 2003 100 Tajikistan NA NA 83 1995 NA 2000 83 Turkmenistan 81 1999 81 1999 81 1999 81 Ukraine 97 2004 97 2003 97 2003 97 Uzbekistan 87 1999 87 1999 87 1999 87 Table 7: Main sources of minimum paved road percentages across 702 international (country-pair) routes. Country No. of Routes Percentage of Total Routes Percentage Paved Rank Albania 130 18.52 24 Hungary 108 15.38 23 Romania 220 31.34 26 36 Table 8: Variables and sources. Variable Description Year Source Borderij Dummy variable equal to 1 if countries i and j share a common land border NA Mayer & Zignago (2006) Colonyij Dummy variable equal to 1 if countries i and j have ever had a colonial link NA Mayer & Zignago (2006) Comlang_Ethnoij Dummy variable equal to 1 if the same language is spoken by at least 9% of the populations of countries i and j NA Mayer & Zignago (2006) Dist_Cepiiij Great circle distance between countries i and j NA Mayer & Zignago (2006) Distance_Meanij or Distij Distance between countries i and j calculated as the mean of road distances between city pairs in those countries NA Own calculations Docsij Sum of number of export documents in origin country and number of import documents in final destination country 2005 World Bank (2006) Average of quality index in origin country i, destination country j and all Paved_Aveij transit countries (based on road routing), weighted by distance traveled 2003 Own calculations in each country as a fraction of total distance between i and j. Paved_Minij Minimum of quality index in origin country i, destination country j and all transit countries (based on road routing) 2003 Own calculations Tariffij 1+Simple average tariff applied by country j to imports from country i 2003 WITS ­ UNCTAD TRAINS Trade_k Merchandise imports in BEC sector k (aggregated from HS-1996) into destination country from origin country, in US dollars 2003 WITS ­ UN Comtrade Trade0_k Trade_k with zeros inserted for missing bilateral trade flows 2003 WITS ­ UN Comtrade and own calculations 37 Table 9: Regression results for equation (5) using BEC 1-digit data (2003). OLS Poisson NB2 Variable Model 1 Model 2 Model 3 Model 1 Model 2 Model 3 Model 1 Model 2 Model 3 -2.08*** -2.19*** -2.06*** ldist -1.32*** -1.36*** -1.40*** -1.74*** -1.86*** -1.68*** 0.15 0.14 0.14 0.18 0.18 0.17 0.17 0.16 0.18 0.18 0.56 lpaved_ave 1.37*** 1.50*** 0.79* 1.21*** 0.44 0.41 0.52 0.52 0.40 0.38 0.56*** 0.59*** lpaved_min 0.20 0.33* 0.60*** 0.74*** 0.21 0.21 0.19 0.19 0.20 0.20 -4.72*** -4.76*** -4.73*** ltariff -6.71*** -6.97*** -6.44*** -4.03*** -3.96*** -4.05*** 0.91 0.91 0.91 2.08 2.08 2.10 0.78 0.79 0.78 -3.06 -2.39 -3.19 ldocs 2.40 2.41 2.15 -4.03* -3.12 -4.79** 2.47 2.48 2.47 3.41 3.26 3.38 2.21 2.30 2.23 0.24 0.24 0.25 border 0.19 0.17 0.20 0.01 0.02 0.04 0.17 0.17 0.17 0.14 0.15 0.16 0.15 0.15 0.15 0.29 0.32 0.28 colony -0.09 -0.09 -0.24 -0.14 -0.09 -0.14 0.31 0.33 0.31 0.23 0.23 0.21 0.23 0.25 0.24 1.01*** 1.09*** 0.99*** comlang_ethno 0.90*** 0.93*** 0.58** 1.09*** 1.17*** 0.98*** 0.33 0.34 0.32 0.22 0.22 0.25 0.25 0.25 0.25 24.60*** 26.59*** 25.47*** 3.34 9.71 9.81 26.48*** 26.72*** 30.91*** _cons 7.54 8.29 7.30 9.68 7.67 9.45 6.75 5.92 6.64 Observations 2440 2440 2440 2559 2559 2559 2559 2559 2559 R2/Pseudo R2 0.62 0.62 0.62 0.78 0.78 0.78 0.03 0.03 0.03 Model F/2 58.99*** 58.19*** 59.30*** 16491.58*** 17830.7*** 17733.07*** 3761.55*** 3574.03*** 3677.03*** a. Dependent variable is log(trade) in columns 1-3, and trade0 in columns 4-9. b. All models include fixed effects by exporter, importer and sector. c. Robust standard errors (adjusted for clustering by country-pair) are in italics. d. Statistical significance at the 10%, 5% and 1% levels is indicated by *, ** and *** respectively. 38 Table 10: Sensitivity analysis of regression results. Variable Bootstrap Russia Country- Country- Excluded Pair REs Sector FEs -1.74*** -1.84*** -1.10*** -2.49*** ldist 0.19 0.19 0.06 0.18 0.79 0.81* -0.19 0.34 lpaved_ave 0.52 0.45 0.22 0.39 0.60** 0.62** 0.25** 0.59*** lpaved_min 0.25 0.27 0.11 0.19 -4.03*** -3.33*** -2.31*** -2.77*** ltariff 0.95 0.84 0.39 0.76 -4.03 -3.93* 0.91 -3.51* ldocs 3.03 2.32 0.97 2 0.01 -0.02 -0.33*** 0 border 0.19 0.17 0.09 0.16 -0.14 -0.18 0.35*** 0.14 colony 0.31 0.57 0.13 0.26 1.09*** 1.15*** 0.54*** 1.08*** comlang_ethno 0.32 0.27 0.12 0.28 36.11*** 30.18*** 0.75 32.91*** _cons 10.23 6.15 2.93 5.36 Observations 2559 2290 2559 2559 a. All models are NB2. Dependent variable is trade0. b. Models in columns 1-2 include fixed effects by exporter, importer and sector. c. The model in column 3 includes fixed effects as in b, and random effects by country-pair. d. The model in column 4 includes fixed effects by exporter-sector, importer-sector and sector. e. Robust standard errors (adjusted for clustering by country-pair) are in italics in columns 2-4. f. Bootstrapped standard errors (500 replications, allowing for clustering by country-pair) are in italics in column 1. 39 Table 11: Regression results for equation (4) by BEC 1-digit sector (2003). Variable Food & Industrial Fuels & Capital Transport Consumer Beverages Supplies Lubricants Goods Equipment Goods ldist -2.50*** -2.23*** -4.92*** -2.01*** -2.81*** -2.45*** 0.23 0.21 0.71 0.21 0.21 0.2 lpaved_ave 1.29** 0.4 -2.09 -0.23 -0.41 -0.38 0.62 0.42 2.03 0.52 0.7 0.37 lpaved_min 0.71** 0.44* 0.93 0.62** 0.70** 0.37 0.32 0.26 0.58 0.24 0.28 0.23 ltariff -2.19** -6.04*** 2.09 -9.61*** -2.83** -1.44 0.99 1.99 8.46 3.39 1.29 1.23 ldocs -5.53* 2.29 5.63 -5.25* -7.82* -8.11*** 3.02 2.3 10.44 3.06 4.11 2.79 border -0.48** 0.04 2.14*** -0.19 -0.74*** -0.40* 0.21 0.18 0.47 0.2 0.2 0.21 colony 0.67** -0.21 -0.35 0.37 0.59* 0.5 0.32 0.26 0.66 0.3 0.3 0.32 comlang_ethno 1.59*** 1.31*** -0.35 1.22*** 0.93*** 1.60*** 0.37 0.27 0.8 0.3 0.29 0.35 40.03*** 20.82*** 28.18 36.48*** 54.98*** 49.15*** _cons 9.78 6.24 37.93 8.02 14.45 7.21 Observations 437 474 279 445 370 457 R2/Pseudo R2 0.04 0.04 0.04 0.05 0.04 0.05 Model F/2 1516.44*** 2397.91*** 2324.02*** 2621.91*** 2726.26*** a. All models are NB2 and include fixed effects by exporter and importer. Dependent variable is trade0. b. Robust standard errors (adjusted for clustering by country-pair) are in italics. c. Statistical significance at the 10%, 5% and 1% levels is indicated by *, ** and *** respectively. 40 Table 12: Simulation results (increase in aggregate intra-regional trade) using estimated coefficients from Table 9 column 7. US$bn % of baseline Simulation I (region wide road upgrade) 56.71 50.4 Simulation II (3 country road upgrade) 34.99 31.07 Simulation III (tariff reduction) 6.19 6.38 Simulation IV (trade facilitation) 19.02 17.56 a. Implied baselines are slightly different across simulations due to rounding and variations in effective sample size. Table 13: Estimated costs (US$ per km) of road reconstruction and development work. Source: ROCKS. Eastern Europe Former USSR Combined Western Europe World Observations 82 8 90 2 205 Average 266686 295560 269253 359172 280691 Median 227031 283737 234153 359172 211445 Minimum 36762 128935 36762 306353 8219 Maximum 666219 464811 666219 411991 2678092 Std Deviation 147025 118359 144373 74698 276780 41 Table 14: Estimated costs (US$ million) of upgrading principal national roads (km). Simulation I Simulation II Country Road Low Mean High Low Mean High Length Cost Cost Cost Cost Cost Cost Albania 375 14 101 250 14 101 250 Armenia 328 0 0 0 0 0 0 Azerbaijan 989 36 266 659 0 0 0 Bosnia and Herzegovina 1880 69 506 1252 0 0 0 Bulgaria 3628 0 0 0 0 0 0 Belarus 5673 0 0 0 0 0 0 Croatia 2679 0 0 0 0 0 0 Czech Republic 3397 0 0 0 0 0 0 Estonia 1059 39 285 706 0 0 0 Georgia 1246 46 335 830 0 0 0 Hungary 4100 151 1104 2732 151 1104 2732 Kazakhstan 13006 0 0 0 0 0 0 Kyrgyzstan 1685 62 454 1122 0 0 0 Latvia 1847 0 0 0 0 0 0 Lithuania 2331 0 0 0 0 0 0 Macedonia 910 0 0 0 0 0 0 Moldova 1075 40 289 716 0 0 0 Poland 12818 0 0 0 0 0 0 Romania 7664 282 2064 5106 282 2064 5106 Russia 41438 0 0 0 0 0 0 Serbia and Montenegro 3834 141 1032 2554 0 0 0 Slovakia 2655 0 0 0 0 0 0 Slovenia 1016 0 0 0 0 0 0 Tajikistan 1713 63 461 1141 0 0 0 Turkmenistan 1310 48 353 873 0 0 0 Ukraine 14071 0 0 0 0 0 0 Uzbekistan 2986 110 804 1989 0 0 0 Total 135713 1100 8055 19930 446 3269 8088 42 Figures Figure 1: Network of major roads in Europe and Central Asia. 43 Figure 2: Scatter plot of mean road distance against great circle distance. 7000 6000 5000 EAN 4000 NCE_MA 3000 STID 2000 1000 0 0 1000 2000 3000 4000 5000 6000 7000 DIST_CEPII 44 Figure 3: Simulation I, percent increase over baseline. 180 160 140 120 asee 100 Incrt Exports Imports encreP 80 60 40 20 0 AlbaniArmeAze a nia ia rus oniaorgiaCroatiHun a garykhstan ubLithuania lic LatviaMoldovnia, FYRPo a ic tan ine p p n R Ge Rom ed eki F Re Ta snia and He rbaijanBulgarrzegovinaBelah epublicEst Kazargyz Re cedo land ania erationtenegro ublSloveniajikisrkme istanUkraUzb stan Mon ak Tu Czec Ky Ma Russiana and Slov rbi Bo Se Country 45 Figure 4: Simulation II, percent increase over baseline. 140 120 100 ase 80 crenIt Exports Imports cen 60 Per 40 20 0 AlbaniaArmeniaerbaijanBulgarrze ia nia n gro c ia ma jikistannistanUkraine istan bek Az govinaBelaruRepublicEstoniaGeorgiaCroatiHungary khstanepublLithuania s a ic land R LatviMoldovnia, FY a a R Po Ro RepubliSlovenTa He ech Kaza yz Federationtenek g cedo rkme Uz Tu Bosnia and an Cz Kyr Ma Russirbia and MoSlova Se Country 46