Food Prices, Road Infrastructure, and Border Effects in Central and Eastern Africa

Transcription

1 Work in progress Food Prices, Road Infrastructure, and Border Effects in Central and Eastern Africa Paul Brenton Alberto Portugal-Perez Julie Régolo February 2013 Abstract Better integrated regional markets for agricultural goods would contribute to food security on Africa. We estimate the effect of distance and border-crossing impediments in Central and Eastern Africa for three food staples: maize, rice, and sorghum. We use high-frequency consumer price data for 152 towns in 13 African countries and a rich dataset on the road infrastructure separating the towns that includes details of the road quality for each segment. We estimate the impact of road infrastructure, border impediments and exchange rate fluctuations between countries on deviations from the Law of One Price, and find that substantial distance and border effects in explaining price differences between towns. Yet, according to our estimates, cross border trade costs impede regional market integration more than costs related to distance. These results suggest that priority should be given to reducing transaction costs and trade restrictions between African countries to improve market integration and to contribute to food security. JEL classification: Keywords: border effects, Africa, road infrastructure World Bank, pbrenton@worldbank.org; World Bank, aportugalperez@worldbank.org; University of Geneva, Julie.Regolo@unige.ch; 1

2 I. Introduction There are enormous opportunities for cross-border trade in agricultural products from food surplus to food deficit areas and as a result of differing seasons and production patterns. For example, Southern Malawi is not well endowed with agricultural potential and is a persistent food deficit area. Nearby Northern Mozambique is a productive area for growing maize, the main staple of the region, but it is distant from the main area of national consumption in the south of the country. Differences in weather patterns entail low correlations in production between countries and regional production is far less variable than production at the country level. Hence, regional trade integration can have a substantial impact by better linking farmers to consumers across borders and in ameliorating the effects of periodic national food shortages and increasing global food prices. Nevertheless, as summarized in World Bank (2012), while there is enormous opportunity for trade across borders between food surplus and food deficit areas, such potential is undermined by barriers to trade, which include regulatory constraints that raise the price and limit access to key inputs such as seeds and fertilizers, high transport costs reflecting limited competition among providers of transport and logistics services, the costs of getting goods across borders and opaque and unpredictable trade policies, including export bans. These policy-related barriers compound the high costs of physically moving food and people within and between countries in Africa due to poor infrastructure. The Africa Infrastructure Country Diagnostic study (World Bank, 2010) found a deficit across all the key core infrastructure, transport, telecommunications and energy. Clearly, there is a need to scale up the levels of investment in trade related infrastructure. Though road infrastructure along the major international trade corridors is increasingly in fair to good condition, the same is not the case on the intra-regional links which require much greater attention in discussions about filling the infrastructure gap in Africa. An important step towards the implementation of policy to improve food security is to evaluate by how much trade barriers at the border and transport costs impede the integration of agricultural markets. The depth of market integration in a region is typically assessed by the volume of intra-regional trade. This is difficult in Africa because of the substantial volumes of informal trade in food that are not captured in official statistics. A recent report (USAID (2013) suggests officials statistics in West Africa may underestimate flows of trade in food by as much as 80%. This paper uses an alternative procedure to assess the extent of regional market integration for basic food staples. It measures market integration in Africa by examining food prices in different markets and the extent to which they move together. In well-integrated and well-functioning markets, price differences should be arbitraged away through intraregional trade. If prices of goods are equal across locations, once converted into the same currency, the law of one price (LOP) holds. A less restrictive version of the LOP, its relative version, admits that even if prices are different due to structural differences, highly integrated markets should have similar price variations, through a customer arbitrage across markets. In other words, the change in relative prices should be zero under the relative LOP. Yet, high transport costs, border barriers, and imperfect markets may 2

3 prevent arbitrage and this would result in large price dispersion and higher price variations among locations; which lead to potentially high welfare costs. In this paper, we examine movements in relative prices of 3 key agricultural commodities, maize, rice and sorghum for 152 towns located in 13 countries of the region 1. We estimate the extent to which road distances and traveling time, national border frictions, and nominal exchange rate variations have an impact on deviation on relative prices (LOP absolute version) and on the monthly change in relative prices (what we call the relative version of the LOP). 2 The use of specific commodity prices leads to precise estimates of border coefficients. Indeed, recent research shows that using aggregate prices indicators instead of disaggregated good prices could lead to larger and less precise estimates of border affects (e.g. Broda and Weinstein (2008) and Engel et al. (2005) for the US-Canada border; Grafe et al. (2008) for five central Asian countries, Aker et al. (2010) for the Niger-Nigerian border, and Versailles (2012) for five East African countries. To assess the role of distance and travel times on market integration, we use information from a rich database on road infrastructure between African towns that includes information of the quality of roads. Border impediments are evaluated by estimating a model where the explained variable is a measure of the relative price of a good between two locations (or ratio of prices in a common currency) and the explanatory variables are a border-dummy for town-pairs located in different countries, a measure of the distance or travel time separating the towns, and other controls, as in Engel and Rogers (1996), Parsley and Weil (2001), Broda and Weinstein (2008), and others. We introduce interactions of month-product and town-product fixed effects to control for unobserved variations along these dimensions. For the LOP, the coefficients on the border dummies evaluate the magnitude of relative price divergence between countries with respect to relative prices within countries. A recent paper by Gorodnichenko and Tesar (2009) recently points out that estimates of this standard model fail to take into account the so-called country heterogeneity effect: if relative price variability across locations within the same country differs systematically country-by-country, the border effect measured by a regression comparing within-country and cross-country price dispersion will be confounded with the difference between these countries' internal price distributions. We take into account this critique and compare the distribution of relative prices among countries. In addition, we provide lower and an upper bound estimates for estimated border effects. Our estimates suggest that road distances and travelling time between towns as well as border frictions largely impede market integration in the region. On average, distance increases price differences between countries by 17% in the region and, once controlled for distance, price differences are estimated to be 11.3% larger between than within countries. Second, we provide comparable estimates for border effects for the 25 pairs of contiguous countries in the sample. Yet, we find that distributions of relative prices could sharply differ within countries in the 1 Burundi (BDI), Djibouti (DJI), Democratic Republic of Congo (DRC), Ethiopia (ETH), Kenya (KEN), Mozambique (MOZ), Malawi (MWI), Sudan (SDN), Somalia (SOM), Rwanda (RWA), Tanzania (TZA), Uganda (UGA) and Zambia (ZMB). 2 Deviations from the Law of One Price, estimated by taking the absolute log of relative prices in level in the LHS, give the long run border effect, as argued by Versailles (2012), while deviations from the Relative Law of one price, estimated by taking the absolute first difference of relative prices, count for the short term border effect. 3

4 region. In particular, domestic relative prices in DRC and in Somalia are around 40% greater than the average relative prices of town-pairs within Djibouti, the country with the best domestic integration. According to Gorodnichenko and Tesar s critique, for these country pairs with sharp differences of domestic price distributions, border effects are highly volatile according to the benchmark used for their estimation. We then provide lower and an upper bound estimates for border coefficients. We find that for most country-pairs, large border effects estimated using the absolute LOP are associated with large border effects estimated using the relative LOP and provide the distance equivalent of these border effects. According to the results, the most integrated country pair in the sample is Tanzania and Zambia while one of the least integrated pair is Somalia and Ethiopia. Finally, we discuss the potential value of this analysis in monitoring the impact of regional integration and, in particular, at identifying borders that can be subject to potential policy reforms aimed at facilitating cross-border trade in Africa. The rest of the paper is organized as follows. Section 2 develops the empirical model and discusses the methodology of the identification of border effects. Section 3 presents the data and describes the sample. Section 4 presents and discusses the results for the impact of transport costs and border impediments on deviation from the LOP and the RLOP. Section 5 provides our main conclusions. 2. Empirical Model Deviations from the LOP and Market integration. The Law of One Price (LOP) states that, once converted to a common currency, commodity prices should be equal across locations, when markets are fully integrated. In a less restrictive version the relative version of the law of one price (RLOP) prices may be different in two towns, but price movements should be the same in these towns. 3 If the LOP holds, the RLOP also holds automatically. In segmented markets, where arbitrage is not possible, the price level could vary across locations for a variety of reasons. One is that production costs may be different across locations. For example, the price of non-tradable inputs or services could differ across locations according to the resource endowments and/or productivity of workers (see Gorinchas et al., 2012). Another reason is that preferences and/or market size may vary across locations and lead to difference in prices. If markets are perfectly integrated and final goods perfectly homogenous, economic agents will arbitrage until equalization of goods prices, i.e. until the LOP hold. However, trade costs or other market imperfections may impede arbitrage so that markets are imperfectly integrated. In that case, and if structural price differences between locations are high enough with respect to transport costs to motivate arbitrage, the price differences between locations will be equal to 3 The analogy is drawn from the absolute and relative version of the purchasing power parity that focuses on price indices. 4

5 trade costs (Gopinath et al, 2011): pikt p jkt T, where p ikt and p jkt are the price of product k at time t respectively in towns i and j and T are the trade costs associated to trade between these towns (which include transport costs and transaction costs). According to this expression, the greater the trade costs, the greater are the price differences between locations. The RLOP focuses on the speed to which price shocks are spread and allows price levels to be different across locations. Therefore, structural differences (such as resource endowment, market size, differences in productivity) do not necessarily prevent the RLOP to hold, and deviations from the RLOP should be directly related to trade costs. Trade costs between locations can be decomposed into (i) transport costs that are variable and positively related to distance and (ii) all fixed trade barriers between locations. Denote T w=τ.distance+η the trade costs between towns within the same country (with τ >0 the transport costs associated to distance and η the transaction costs/fixed trade barriers between towns). For cross-country trade, T B=τ.distance+η+B with B the border effect or the transaction costs and trade barriers associated with crossing the border. It follows that the so-called border effect is the difference between transaction costs (and degree of integration) associated to trade between countries T B and those associated to trade within countries. p p.distance B, if towns i and j are in different countries and ikt ikt jkt p p.distance, if towns i and j are in the same country. jkt A similar reasoning holds for the variations of relative prices. Therefore, to assess the transaction costs associated with crossing the border and transport costs behind deviations from the absolute Law of One Price, we estimate the following specification: P P 0 1 RoadDist Border, ln / ln (I) ikt jkt ij IJ ij ik jk kt i jkt IJ and for deviations from the relative version of the LOP: P P 0 1 RoadDist Border u, ln ikt / jkt ln ij IJ ij ik jk kt i jkt (II) IJ P ikt In (I) and (II), is the price of commodity k in town i and month t expressed in dollars. As P ikt / P jkt is the ratio of prices labeled in dollars, it is also a measure of the real exchange rate for commodity k between towns i and j in month t. In (II), ln Pikt / Pjkt is the first difference of the log of relative prices of good k between towns i and j for months t and t-1, or the monthly variation of the log of the real exchange rate of good k. RoadDist ij is the road length between towns i and j. Its coefficient is expected to be positive as price differences are deemed to be 5

6 bigger between more distant cities. different countries, and zero if not. 4 The variables ik and jk Border ij is a dummy that equals one if towns are located in are town-product dummy interactions and control for town-product kt specific characteristics affecting price differentials such as the potential non-perfect homogeneity of products across cities (differences of quality, color, shape) or differences of input of services; are month-product dummy interactions and control for specific monthproduct variations. In model II measuring deviations of relative version of the LOP (henceforth referred as RLOP), town-product dummies control also for idiosyncratic measurement error or seasonality in some towns that increase price variations on average Interpreting coefficient estimates As a dummy designating town-pairs located in the same country (e.g. RWA-RWA, DRC-DRC, MWI-MWI ) are not included in models I and II, the coefficients of border dummies measure by how much a transition from the average within-country town pair (RWA-RWA or DRC-DRC or BDI-BDI ) to a between country town pair (RWA-DRC or DRC-BDI or RWA-BDI ) raises the average relative prices (model I) or its variation (model II). In the general version of the model with a single border-dummy for the region (as opposed to dummies specific to each border), the border coefficient, γ, measures this transition on average in the region. Results for this specification are discussed in section 4.1. When dummies for each country-pair border (I-J) are introduced, the coefficient of each dummy, γ I-J, reflects the relative price between countries I and J with respect to the average within-country relative prices in the sample. The results from this specification are shown in section 4.2. Measuring the relative price between countries with respect to a common benchmark (the average within country relative price in the region) has the advantage of estimating relative price levels (or variations) at borders which are directly comparable across country-pairs. However, one has to be careful with the interpretation of border coefficients, in particular if countries have different price distributions 5. First, border coefficients may not reflect the direct effect of crossing the border on relative prices (relative prices variations). For example, the border effect estimated between Tanzania and Kenya should reflect the magnitude of relative prices between Tanzania and Kenya with respect 4 The border dummy can be specified as (i) a single dummy measuring the average effects of borders in the sample (γi-j=γ) or (ii) border dummies specific to country-pairs (γi-j for all (I-J) {DRC-BDI, KEN-ETH, MWI-MOZ, RWA- BDI, RWA-DRC, SDN-ETH, SDN-KEN, SOM-DJI, SOM-ETH, SOM-KEN, TZA-BDI, TZA-DRC, TZA-KEN, TZA- MOZ, TZA-MWI, TZA-RWA, UGA-DRC, UGA-KEN, UGA-RWA, UGA-SDN, UGA-TZA, ZMB-DRC, ZMB-MOZ, ZMB- MWI, ZMB-TZA}). We first evaluate the average border effects in the region by introducing a single border dummy that takes the value of 1 if towns are located in the same country, 0 otherwise. We later, replace the dummy variable with 25 border dummies, each one defined for each pair of contiguous countries in the sample. 5 The distribution of prices can vary across countries with the degree of integration of their domestic labor market and good markets. 6

7 to relative prices within Tanzania and/or within Kenya. If relative prices within Tanzania differ substantially from the average relative price within the countries in the sample, the border coefficients of model I may not capture the effect on prices of crossing the border from Tanzania to Kenya. Second, estimates may reflect heterogeneity in relative prices across countries as pointed out by Gorodnichenko and Tesar (2009). When the distribution of relative prices (or relative price variations) in a country is different from the distribution of relative prices in the country at the other side of the border, the border coefficients may reflect what Gorodnichenko and Tesar (2009) call the country heterogeneity effect. For example, in the context of the US- Canada border, these authors found that the US-Canada border effect is much bigger when estimated from the Canadian perspective (i.e. when including a dummy for US city pairs and not the dummy for Canadian city pairs) than when estimated from the US perspective (i.e. when including a dummy for Canadian town-pairs and removing the US city-pairs dummy). This difference is explained by a relatively lower variation of Canadian prices than US prices. Therefore, when within-country price distributions are very different, it is only possible to estimate a lower and an upper bound for the border effect by omitting alternatively the dummy of one of the countries, which becomes the benchmark for the border coefficient. To deal with these potential issues, we examine the level of intra-national market integration of countries in the region and compare their relative prices distributions in section 4.3. Further, in section 4.4, we estimate the magnitude of the border from the perspective of each bordering country (benchmark country) by introducing in regressions a dummy for town-pairs located in the same country (e.g. RWA-RWA, BDI-BDI ) and omitting the dummy for the benchmark country. This provides lower bounds and upper bounds for each border coefficient estimate of models (I) and (II). In addition, border coefficients measure the degree of integration of markets between countries relative to the degree of integration of domestic markets. The less integrated domestic markets within African countries are, the lower the border coefficients are, ceteris paribus. We also provide estimates of the degree of integration of domestic markets in the countries of our sample. Finally, border coefficients can capture four confounded factors behind cross-countries price differences. First, they can reflect transaction costs related to crossing the border, such as trade policy regulations (bans, tariffs, licenses, non-tariffs measures), customs efficiency or currency exchange costs. Second, they can capture idiosyncratic barriers between countries, which may impede trade partnership, such as language or ethnicity differences 6. Third, the markup across countries can be different when the degree of market competition varies. Fourth, border coefficients can also capture the different degree of price stickiness in markets on both sides of the border. Indeed, if the price of a good in local currency does not change when the nominal exchange rate changes (i.e. prices are sticky), cross-border relative prices would fluctuate with 6 As an example, Aker et al found that at the border between Niger and Nigeria, price differences may vary from 8% to 24% regarding if the people are from the same ethnic group or not. 7

8 the nominal exchange rate and these variations will be captured by the border dummies. Note however, that price stickiness may be related to border barriers, as a seller can resist cutting the price of their good when markets are imperfectly integrated. To isolate the last effect, we introduce a measure of monthly nominal exchange rate variation on the right hand side of the regression ( ln ER ). ijt We restrict the sample to market-pairs that are less than 1000 kilometers apart, the average distance between markets in our sample 7 to diminish the effect of country-level unobserved heterogeneities, by assuming that they are weaker for market pairs located close to the border The distance-equivalent of border frictions To provide a measure of the size of the border, we follow the literature and compute the distance-equivalent of the estimated border effects. Our measure, Dist, was used by Engel and Rogers (1996) and provides the distance between towns in the same country that would increase price dispersion by the same magnitude than that due by the estimated border coefficient 8 : Dist IJ exp IJ / 1 (III) This measure of the distance-equivalent of the border does not change whether distance is measured in miles, or kilometers, or any other unit Data Detailed consumer price data for 3 commodities (Maize, Sorghum, and Rice) was compiled from Fewsnet and FAO for 152 towns in 13 countries in East and Central Africa on a monthly basis. Because of data availability, we restrict our analysis to the common period May 2008-Oct We match the commodity price data with a detailed dataset on the road network in these countries compiled from GIS. The rich road network data not only includes the length of the fastest or shortest road separating any pair of towns in the sample of countries, but also includes detailed measures on the road quality for each segment, which is used to obtain the average number of hours required to travel between any two towns. Figure 1 shows the geographical dispersion of the towns in the sample as well as paved and unpaved roads. Here: Figure 1 7 In their main analysis Gopinath et al (2009) restrict the sample to markets within 500 km of the U.S.-Canada border, although they find very similar results for bandwidths of 100 km or 350 km. 8 Dist. solves the equation: IJ 1lnDist IJ. 9 A second measure, used by Parsley and Weil (2001), compute the extra-distance that has to be added to the average distance between two towns to generate as much price difference as the border coefficient 9. This measure has the advantage to give a distance equivalent in the unit in which the distance is measured in the sample. Yet, its dependence on the size of the average inter-town distance in the sample makes it very sensitive to small variations in the estimated coefficients. In addition, the measure is much larger than Dist and looks less credible. Consequently, we prefer the distance equivalent as computed by Engel and Rogers (1996). 10 Data for DRC is only available over this period which constraints the sample. As data for other countries is available over a larger period (January 06-August 2011), we check the robustness of results on the larger period by excluding DRC. 8

9 In the raw data, prices of commodities are sometimes available in different units (e.g. kilograms, tons, etc) and for different varieties (e.g. maize or white maize). In order to keep the largest possible number of observations, we transform the price units to dollars per kilogram using the monthly average exchange rate and assimilate a variety (e.g. white maize) to the main commodity (e.g. maize) when the latter is not available. In appendix A, we compare the relative prices of our expanded sample with a more restricted sample where we exclude goods that are labeled differently than the main good (e.g. white maize) and find that the distribution of relative prices are similar and not very likely to affect results (see Figure A.1). Table 1 provides a summary description of the number of cities and the price data availability for each of the three commodities studied. Depending on the country, commodity prices may be available for one to three commodities. The number of cities in the data varies also greatly between countries. Most of the countries have price data for at least two commodities and 6 towns; the greater data availability being for Rwanda with 30 towns and 3 commodities. Table 1 here: Table 2 reports summary statistics for the two dependent variables in our regressions: the absolute value of the log of relative prices and the absolute value of the first difference of the log of relative prices. Summary statistics are broken down into: (i) within-country price differentials where both towns are located in the same country, and (ii) between-country price differentials where each town is in a different neighboring country. A zero value of the log of relative prices implies that the LOP holds. Likewise, a zero value of the log of first differences in relative prices implies that monthly price variation in both locations is the same, and thus, the relative version of LOP (RLOP) holds (even if the LOP does not necessarily hold). Column (1) reports an average measure of the distance between towns (in kilometers) and column (2) reports our measure of average travelling time (in hours) between towns in each subsample. Comparing the average within-country price differentials with between-country price differentials gives a first insight on the potential border effects between countries. Withincountry prices are 30% (=exp(0.26)-1) higher on average than the LOP benchmark, and withincountry relative prices variations are on average 19% (=exp(0.17)-1) higher than the RLOP benchmark. Both statistics are higher in the between-country sample with relative prices being 49% (=exp(0.41)-1) higher on average than the LOP benchmark, and relative-prices monthly variation being 20% (=exp(0.18)-1) higher on average than the RLOP price benchmark. Table 2 here For example, price differences between a town in Malawi and a town in Mozambique are on average 28 % (=exp(0.25)-1), a figure greater than within-countries price differences (of respectively 19%(=exp(0.17)-1) in Malawi and 22% (=exp(0.20)-1) in Mozambique). Yet, the econometric analysis below attempts to disentangle the contribution of inter-town distances and the contribution of border barriers on price differences. Across commodities, relative prices in levels and variations are on average larger for Maize than for Sorghum and Rice. 9

10 East and Central African countries seem to have heterogeneous distributions of within-country relative prices, as countries with greater price dispersion tend to have larger distance between towns in our sample. For example, DRC and Somalia are the countries with the largest average price differences, whereas Djibouti and Rwanda are on the other side of the spectrum with the smallest price differences. This is in line with an effect of transport costs on price differences Econometric Analysis. The competing effects of transport costs, barriers associated with crossing the border and nominal exchange rate variations on price levels and price variations between locations are assessed in the next section in four steps. First, we estimate the average effect of these three factors in the region and present the global distance equivalents of the average impact of borders. Second, border effects are estimated for each pair of contiguous countries. Third, we study the differences of price integration within the countries of the sample. Fourth, we check the robustness of the estimations, provide an interval for the border effects and discuss distanceequivalents for each border Border, Distance and Price rigidity in the region. Table 3(a) reports estimates from variations of model (I), where the dependent variable is the absolute value of the log of relative prices. Specifications (1) and (2) estimate the average border effect in the East and Central African regions while controlling respectively for distance and travelling time between cities. Specifications (3) and (4) include the absolute value of monthly depreciation of the exchange rate and column (5) and (6) show results using a restricted sample of town-pairs that are located less than 1000 km of distance of each other. All regressions include interaction terms of town-product dummies and product-year dummies. All coefficient signs are as expected. Our estimated coefficients on road length and border dummy are similar in terms of magnitude to those obtained by Aker et al (2012) with the same specification between Niger and Nigeria. According to estimates of specification (1), a distance of 1212 km (the average inter-town distance in the sample) accounts for an increase in prices of about 17.5%. 12 Border frictions have a substantial effect on price differentials. After controlling for inter-town distance, price differentials for towns separated by a border increase by 11.3% (=exp(0.107)-1) on average in the region. To provide a magnitude of the average effect on prices of crossing a border in East and Central Africa, we compute the distance-equivalent measures, Dist, defined by equations (III). 11 Only price data of Sorghum are available for Sudan. Sorghum price differences between locations are relatively lower on average in the sample than those of Maize and Rice. This and the low number of observations could explain why, despite its relatively large size, Sudan within price differences are relatively low on average. 12 Computed as: exp[0.0227*ln(1212)]-1. 10

11 Accordingly, crossing a border in the region corresponds, on average, to a distance-equivalent of 110km. There are multiple reasons that could explain low market integration between these African countries. Nevertheless, the distance equivalent estimated here is lower than the one found by Engel and Rogers (1996) between Canada and US (they found Dist equal to ); and are in the same range than those reported by Broda and Weinstein (2008), who use microdata between US and Canada. Even if border estimates are difficult to compare given the heterogeneity of products or level of price aggregation between studies, is it possible that the border effects between African countries are as low as those found between developed countries? Border coefficients measure the degree of market integration across different countries compared to the degree of integration of markets located within the same countries. The less integrated markets within African countries, the lower the estimated border coefficients are, ceteris paribus. Therefore, low values of border coefficients could either reflect low barriers at the border or a substantial lack of integration within countries. On the one hand, weighbridges and policy controls increase transport delays and therefore constitute substantial barrier to trade within African countries. Moreover, ethnicity and language can be different within countries and can preclude full market integration. On the other hand, the lack of control at the border and the corruption of customs may facilitate trade of primary commodities between countries; which is reflected in the large amount of informal trade estimated by Stryker (2012) based on FEWSNET data. For example, according to their results, MT of Maize have been informally traded between Tanzania and Kenya in In this regard, it is less surprising that estimated average border coefficients between African countries can be low. Table 3 here In specification (2), we replace the inter-town kilometer distance variable with the measure of inter-town travel time in hours. While estimates do not vary significantly, the effect of travel time on relative prices seems to be lower than the effect of distance in previous estimates. The small difference may be due to the fact that inter-town travel variable takes into account the quality of all road segments and is therefore richer than the distance variable. Analogous to distance-equivalent measures computed for specification (I), we compute the time-equivalent measure of borders in a similar way to equation (III) and report this figure at the bottom table 3. We find that crossing a border has the same effect on relative prices as travelling 518 hours between towns of a same country. Specifications (3) and (4) include exchange rate variations. The estimated coefficient on this variable is positive and significant, and its addition reduces the estimated border coefficient. Consistent with the literature, the strong impact of this coefficient reveals that nominal price rigidities explain substantially the deviations from the Law of One Price (Crucini et al, 2009). In the absence of nominal price rigidities, the exchange rate coefficient should be non-significant. However price stickiness may be dependent upon market segmentation as producers will adjust more quickly their nominal price if markets are integrated. In that sense, the effect of variations of the nominal exchange rate on relative prices between countries is a consequence of border barriers and therefore is part of the border effect. 11

12 To reduce heterogeneity between town-pairs and the possibility of spurious correlation when considering towns that are far away from each other, specifications (5) and (6) restrict the sample to town-pairs that are no more than 1000 kilometers apart, which is roughly the median inter-town distance in the sample. As expected, the coefficient on the border dummy becomes smaller. Fewer differences in language, of ethnicity and more generally of culture between towns geographically close to each other might explain this finding. Transport costs play a greater role in this subsample where the distance equivalent of the border effect is 14 km (column (5)). We turn the analysis to the deviations from the relative version of law of one price around model II. Analogous to specifications in table 3a, table 3b reports estimates of similar variations of model II, where the explained variable is the first difference of relative prices (in logs) taken in absolute value. Across specifications, the coefficients on distance and on the border dummy are positive and significant. More precisely, in column (1), the average distance in the sample (1 212km) increases the monthly variation of relative prices by 3.5% (=exp(0.0049*ln(1212))-1), on average. Similarly, the monthly variation of relative prices for towns separated by a border increases by 1.3% (=exp(0.0134)-1) on average, after controlling for distance and market-product fixed effects. According to these results, barriers at the border would appear to play a relatively greater role in explaining deviations from the absolute law of one price than from the relative law of one price. Coefficients on nominal exchange rate variations in columns (3) and (4) are not significant. As in previous estimates, the coefficient of the border dummy becomes smaller when the sample is restricted to town-pairs located at less than 1000 km (columns 5 and 6). Finally, deviations from the absolute and the relative law of one price are influenced by intertown distance and are, on average in the region, greater between countries separated by a border than within countries. In next section, we attempt to estimate the border effects specific to each pair of contiguous countries Specific bilateral border effects. Table 4 reports estimates of specifications that include border dummies for each pair of contiguous countries (25 country-pairs in the sample), instead of a single border dummy. Specifications (1) to (4) account for deviations from the Law of One Price and specifications (5) to (8) for deviations from the Relative Law of One Price. Introducing 25 border dummies allows for heterogeneity of border effects across country-pairs. Most border coefficients are positive and significant. The coefficients on distance (in column (1), (3), (5) and (7)), travel time (in column (2), (4), (6) and (8)) and exchange rate (in column (3), (4), (7) and (8)) are almost unaffected by the decomposition of the border coefficients between each country-pair. Table 4, here 12

13 In columns (1) to (4), the positive sign of most border dummy coefficients reveals that relative prices for pairs of towns located in different countries tend to be larger than the average withincountry relative prices in the region. By contrast, the border dummies for Kenya-Uganda, Uganda-Sudan, Tanzania-Burundi, and Tanzania-Zambia are not significant, and thus not significantly different from the average within relative prices in the region. Thus, relative prices between these country-pairs are the lowest of the region and reflect greater integration between them for Maize, Sorghum and Rice. The larger border coefficients found are between Somalia and Ethiopia and Zambia and DRC, for which price differences are respectively 47% (=exp(0.388)-1) and 36% (=exp(0.306)-1) greater than the average within the country of the region. The other border coefficients are in the range [0.042 ; 0.19]. For example, price differences between Mozambique and Malawi are 4% (=exp(0.0425)-1) greater than the regional average within countries. As a member of both regional agreements, EAC and the SADC, and as a costal country, Tanzania tends to be relatively more integrated with all its neighbors. The inclusion of the exchange rate variation in column (3) and (4) decreases slightly all border coefficients. This reveals that price stickiness and exchange rate variations influence price differences between all contiguous countries in the sample. Figure 3 compares the border coefficients in the baseline regression focusing on the LOP (column 1) and in the regression focusing on the RLOP (column 5). Overall, there is a positive correlation between coefficients as greater relative prices between countries are associated with greater variations of relative prices. This correlation supports the argument that both approaches (LOP and RLOP) provide similar measures of market integration. Country-pairs Zambia-DRC and Somalia-Ethiopia seem to be the exception, as estimated border coefficients show large differences in prices, but similar movement of relative prices. Country pairs Uganda- DRC and Zambia-Mozambique have not only one of the largest relative prices in the region but also substantial monthly prices movements 13. On the contrary, Tanzania and Zambia are relatively well integrated as price differences and relative price variations are low compared to other country pairs. Figure 3 here 4.3. Comparison of market integration within countries. To examine in more details the heterogeneity of relative prices within countries, figures 2a and 2b show the distributions of within-country relative prices and relative price variations after controlling for distance. More precisely, we take the residuals that come from the estimation of the following specification: Q ln RoadDist ijkt 0 1 ij ik jk kt i, jkt ijkt ikt jkt with Q ln P / P for Figure 2(a) and Q ln P / P for Figure 2(b). ijkt ikt jkt 13 Note that for these country-pairs the border effect are computed only on Maize given the data availability 13

14 Figure 2 here. The distributions of relative prices after controlling for distance are heterogeneous (Figure 2(a)), with relative prices being more dispersed in DRC and Somalia. By contrast, the distributions of monthly variations of relative prices are more homogenous across countries (Figure 2(b)). To compare more precisely the market integration within countries, we took a subsample of town pairs located in the same countries and estimate specifications similar to model I and II but where the border dummies are replaced by country dummies (D C-C): Q 1 2 lnroad _ Dist D φ φ φ ijkt ij C C C ik jk kt ijkt C D C-C is a dummy equal to one if the towns i and j are located in country C, and equal to zero otherwise. In table 5, column (3) show the coefficients obtained on the country dummies for the P test of the LOP, i.e. when Qij kt ln ikt P, and column (6) for the test of the RLOP, i.e. when ikt Q. The omitted country dummy in both regressions is DJI-DJI, i.e. the ijkt P ln ikt Pikt coefficients on country dummies reflect the integration of these countries with respect to Djibouti. In the table, countries are ranked according to the level of market integration, from the most to the least integrated. Table 5: here Here again, the results suggest that the level of integration sharply differ between countries. The most integrated countries are Sudan and Djibouti whereas the least integrated are Somalia, DRC and Burundi. This indicates that for these three latter countries, estimated border effects could be very different with respect to the benchmark used. For example DRC is relatively less integrated than the average country in the sample. It follows that the border coefficients previously estimated for DRC are likely to overestimate the barrier that the borders represent for this country. Worse, the border coefficients may capture the fact that price differences are different between DRC and the other countries (Gorodnichenko and Tesar, 2009) Border effects and Robustness. Previous estimations of border coefficients provide estimations for between-country market integration disregarding the integration within the countries on either side of the border. To provide accurate estimates of the border barriers to trade, we follow Gorodnichenko et al (2009) and control for within countries heterogeneity by introducing dummies for towns-pairs located in the same country 14. As a country dummy has to be omitted in each regression to avoid 14 The introduction of country dummies in equation (I) and (II), in addition to border dummies and town dummies generate a multi-collinearity problem which leads to inaccurate estimations of the border coefficients. This is the case, even if one country dummy is excluded (see Gorodnichenko and Tesar (2009)). To avoid this collinearity problem, we follow Gorodnichenko and Tesar (2009) and constrain the sum of the town dummies to be equal to zero for each country. 14

15 multicolinearity, serving as benchmark, we estimate separately 13 regressions, one for each country in our sample. This provides two border coefficients for each country-pair (one for each contiguous country) : the lower bound and the the upper bound estimates of the border coefficient. Table 6 (a) and (b) report estimates for the deviations from the LOP and the RLOP, respectively. In both tables, Column (1) reports estimates of the baseline regression with border coefficients for each country-pair (models I and II, respectively). Column (2) and (4) show estimates of the lower and upper bounds for estimated border effects respectively and columns (3) and (5) report the benchmark country for each bound. Columns (6), (7) and (8) show the distance equivalents of each coefficient. Table 6: here. When within-country price differentials are similar, the lower bound is close to the upper bound, as in the case for example for Kenya-Ethiopia. In contrast, large differences between lower and upper bounds reveal heterogeneity of within-country price differences. In this case, border coefficients reflect both heterogeneity in price variability between the countries and the costs associated with crossing the border. Ethiopia and Kenya have similar price distributions. Therefore, in this case, the border coefficient between these two countries could be interpreted as trade frictions and the border effect is accurately estimated. The first row of column (4) reports coefficients estimated by running model I with all within-country dummies except the Ethiopia-Ethiopia dummy. According to this estimate, crossing the border for Ethiopia increases price differences by 18% (exp(0.166)-1). The lower bound in column (2) provides the border estimate from the Kenya perspective, which is similar (of 16.6%). Both bounds are also similar to the previous estimates of the border between Ethiopia and Kenya (of 17.3%). The distance equivalent of crossing the border between Kenya and Ethiopia is about 704km(551 and 901km respectively from the Ethiopia and the Kenya perspective). Compared to other country pairs, the Kenya-Ethiopia border effect is substantial. The border coefficient between Tanzania and Mozambique measures accurately border frictions. The estimated distance equivalent for this country pair is relatively low, i.e. around 9 km. Country pairs for which the lower and upper bounds are substantially different often involved either DRC or Somalia, the most different countries in terms of within price distribution in the sample, as shown in the Figure 2. In that case, the border effects between these countries and their neighbors are likely to be confounded with the price distributions differences of these countries. Despite the apparent similarity of within country price distributions, the border effects coming from deviation from the RLOP are less accurately estimated as the interval given by the estimations of the bounds may be wide. Finally, while having a large sample of countries is worthy and provides comparable estimations of border effects over the region, the large amount of heterogeneous data could add noise to the estimations and the coefficients on distance representing an average effect. To check the 15

16 robustness of our results, we re-estimate the border coefficients and the lower and upper bounds by running a separate regression for each country-pair sub-sample (i.e. 25 sub-samples). Table A3 in Appendix 3 reports these estimates for deviations from the LOP (column (1) to (3)) and the RLOP (column (4) to (6)) and show that border coefficients are fairly similar and confirm our previous estimations 15. Figure 4 illustrates the border effects obtained with local regressions and compares it with within-country relative prices. It shows that for all country pairs, the intercountry relative prices are greater than within countries, and that the border effect may be viewed differently from each side of the border. Figure 4: Here To sum up, we find a substantial effect of border frictions and transport costs on average in the region on relative prices levels and variations. The magnitude of border effects is heterogeneous between country pairs, Tanzania-Zambia being the most integrated country-pair and Somalia and Kenya one of the least (with an estimated distance equivalent of 0km and km respectively). The market integration within countries differs also substantially in the sample. In particular, relative prices are high between towns within DRC, Somalia and Burundi (of respectively 49.6%, 43% and 13.6% above the average relative prices of towns within Djibouti, the country with the best integration between its towns). Finally, all cross-country relative prices are correlated with exchange rate variations, which suggest price stickiness within countries. 5. Conclusions This paper seeks to contribute to the literature on the importance of policy and infrastructure constraints affecting the capacity of countries to trade food across borders in Africa. As summarized in World Bank (2012), while there is enormous opportunity for trade across borders between food surplus and food deficit areas, such potential is undermined by barriers to trade along the whole of the value chain. We evaluate the relative impact of high transportation costs and border impediments on market integration between 25 pairs of contiguous countries in the East and Central African region and for 3 key agricultural commodities, maize, rice and sorghum. To this, we assess the extent to which road infrastructure and international borders explain deviations of absolute and relative Law of One Price (LOP) between 152 towns of the region. Results suggest that, having controlled for exchange rate variation, transport distance or travelling time between towns and border frictions largely impede market integration. On average in the region, distance increases price differences between countries by 17.3% and, once controlled for distance, price differences are estimated to be 11.3% larger between than within countries. For most country pair, strong 15 Further robustness checks have been applied and are shown in the annex. Border coefficients have been estimated in sub-sample with several restrictions according to distance between cities (taking only the town-pairs which are not further than 500, 1000 or 2000 km from each other) and taking a larger period of time for some country-pairs. Overall the estimated coefficients are very similar. 16

17 deviations from the LOP are associated with strong deviations from the relative LOP, the most integrated country pair being Tanzania-Zambia while one of the least is Somalia-Ethiopia. We find also that the market integration within countries differs substantially in the sample. In particular, relative prices are high between towns within DRC and Somalia, of respectively 49.6% and 43% above the average relative prices of towns within Djibouti, the country with the best integration between its towns. In terms of policy, the results suggest that market integration is relatively low both between and within some African countries. Priority could be focused on removing policy constraints that impact both the movement of goods within countries and on trade across borders, such as lack of competition in the provision of transport and logistics services. In many African countries, transport cartels are still dominant and the cost of transportation is very high compared to other regions. Measures which improve the availability of information on prices both within and across countries could also have an important impact on reducing differences in the relative price of food in different markets. Nevertheless, measures to reduce transactions costs at the border, particularly in fragile states could have a large payoff in terms of facilitating trade in food products across markets. Finally, we raise the issue of whether the approach adopted here could be useful in identifying the impact of specific measures to reduce policy constraints to trade. If procedures at the border are improved we should expect to see a decline in relative prices and/or relative price variation between markets on either side of the border. Having monthly data would enable potential impacts to be closely linked to the time that the policy change was introduced. 17

18 References Aker, Jenny C., Klein, Michael W., O'Connell, Stephen A. and Yang, Muzhe, 2010 Borders, Ethnicity and Trade (May 2012). NBER Working Paper Buys, Piet, Uwe Deichmann, and David Wheeler "Road Network Upgrading and Overland Trade Expansion in Sub-Saharan Africa" Journal of African Economies 19(3): Buys, Piet, Uwe Deichmann, and David Wheeler "Road Network Upgrading and Overland Trade Expansion in Sub-Saharan Africa" Journal of African Economies 19(3): Ceglowski, Janet, 2003, The Law of One Price: Intranational Evidence for Canada, Canadian Journal of Economics, vol. 36, no. 2, pp Chen & William D. Nordhaus, "The Value of Luminosity Data as a Proxy for Economic Statistics," NBER Working Papers 16317, National Bureau of Economic Research. Chen, B, Intra-national versus international trade in the European Union: why do national borders matter? Journal of International Economics, 63: Engel, C., Rogers, J., How wide is the border? American Economic Review 86, December. Engel, C., Rogers, J., 2001 Deviations from Purchasing Power Parity: Causes and Welfare Costs. Journal of International Economics, 55: Engel, Charles & John Rogers, "Relative price volatility: what role does the border play?," International Finance Discussion Papers 623, Board of Governors of the Federal Reserve System. Engel, Charles; John H. Rogers; and Shing-Yi Wang, 2005, Revisiting the Border: An Assessment of the Law of One Price Using Very Disaggregated Consumer Price Data, in Exchange Rates, Capital Flows and Policy, Rebecca Driver, Peter Sinclair and Christoph Thoenissen, eds. (London: Routledge), Gopinath, Gita, Pierre-Olivier Gourinchas, Chang-Tai Hsieh & Nicholas Li, "Estimating the Border Effect: Some New Evidence," NBER Working Papers Gorodnichenko, Yuriy and Linda Tesar. Border Eff3ect or Country Effect? Seattle May Not Be so Far from Vancouver After All. American Economic Journal: Macroeconomics, 1(1):219{241, Grafe, Clement, Martin Raiser, and Toshiaki Sakatsume Beyond Borders - Reconsidering Regional Trade in Central Asia. Journal of Comparative Economics, 36:

19 Henderson, J. Vernon, Adam Storeygard, and David N. Weil Measuring Economic Growth from Outer Space. National Bureau of Economic Research Working Paper Michalopoulos, Stelios and Elias Papaioannou (2010), Divide and Rule or the Rule of the Divided? Evidence from Africa, CEPR Discussion Paper Morshed, A.M., What can we learn from a large border effect in developing countries? Journal of Development Economics 72, Morshed, A.M., 2007 Is There Really A `Border' Effect? Journal of International Money and Finance, 26(7): Nunn, Nathan (2008), The Long Term Effects of Africa s Slave Trades, Quarterly Journal of Economics, 123(1): Odedoku. M.O. Fulfilment of Purchasing Power Parities in Africa: The Differential Role of CFA and non-cfa membership. Journal of African Economies, 9(2):213{234, Parsley, David C., and Shang-Jin Wei Explaining the Border Effect: The Role of Exchange Rate Variability, Shipping Costs, and Geography. Journal of International Economics, 55(1): Stryker D., Study of Policy Options for Increasing Tanzanian Exports of Maize and Rice in East Africa While Improving Its Food Security to the Year 2025, Final Report, NAFAKA/AIRD. USAID (2013) Assessment of the Volumes and Value of Regionally Trade Staple Commodities, USAID West Africa Trade Hub, Paper presented at the Food Across Borders Conference, Accra January. Versailles, Bruno Market Integration and Border Effects in Eastern Africa. CSAE Working Paper WPS/

20 Price differences Figures and Tables for Food Prices, Road Infrastructure, and Border Effects in Central and Eastern Africa Figure 1: Towns and Road quality Figure 2. Distribution of price variables within countries (a) Distance-adjusted within countries relative prices BDIBDI DRCDRC KENKEN MWIMWI SDNSDN TZATZA ZMBZMB DJIDJI ETHETH MOZMOZ RWARWA SOMSOM UGAUGA (b) Distance-adjusted within countries relative price variations. 20