While I find that doubling agricultural productivity under either existing high or counterfactual low trade costs does lower food prices, the effects on farmer incomes are dramatically different in the two cases, with net agricultural revenues actually falling by 71.4% under existing high trade costs and increasing by 12.4% under counterfactual low trade costs. These results underscore the importance of implementing policies to lower trade costs and improve market access in tandem with technology adoption initiatives. This chapter is most closely related to a recent literature on trade costs along intranational spatial transportation networks that has expanded rapidly since the seminal work of Donaldson 2012. Atkin and Donaldson 2015 estimate the distance-dependent component of intra-national trade costs within two sub-Saharan African countries using price and origin data for specific, narrowly-defined manufactured goods. Sotelo 2015 uses a richer dataset from Peru to explore how intra-national trade costs lower agricultural productivity by preventing agricultural producers in particular locations from specializing in the crops in which they have a comparative advantage, a mechanism which is less important in the African context for the range of crops that I consider. This chapter goes beyond the existing literature in several important ways, including covering a larger network of African markets and using a dynamic monthly model with storage, which I show is important for identifying when trade occurs so as to avoid underestimating trade costs and welfare effects. In my case, using a static annual model underestimates trade costs by 23% and welfare effects by 33%. The balance of this chapter proceeds as follows. In Section 2, I describe the context and data. In Section 3, I present my model. In Section 4, I detail my estimation strategy, present my estimates for the model parameters including trade costs,dutch buckets and examine the goodness of fit of my estimated model. In Section 5, I present the results of my counterfactual analysis and robustness checks. Section 6 concludes.Due to data limitations and comparability issues, I restrict my attention in this chapter to the consumption and production of the six major staple cereal grains: maize, millet, rice, sorghum, teff, and wheat2 .

Table 1.1 shows the relative share of cereal grains and other categories of agricultural goods in production value, caloric intake, and gross value of international trade in sub-Saharan Africa. Although they make up only 17.2% of the total value of agricultural production in sub-Saharan Africa, cereal grains are by far the most important source of calories in African diets. Tubers like cassava and yams are another important source of staple carbohydrates, but their perishability and low value-to weight ratios severely constrain their trade and storage. Cash crops like cocoa and tea make up the largest share of the value of African countries’ international agricultural trade, but they differ from cereals in that their production is often localized near ports or in certain geographic niches and is nearly all exported to the world market. Grain trade in sub-Saharan Africa can be roughly classified into two types: farm-to market and market-to-market trade. Although farm-to-market trade may involve much higher trade costs than market-to-market trade due to extremely poor rural infrastructure, I will not be able to capture farm-to-market trade and trade costs in a continent-level model due to data limitations and will focus exclusively on market-to-market trade instead. One important difference between the two types of trade is the level of competition — while farm-to-market trade may be conducted by relatively few traders with significant market power, market-to-market trade at the level considered here tends to be highly competitive with many traders, low firm concentration ratios, homogeneous products, and few barriers to entry . I will therefore assume that traders are competitive price-takers. Grain is bought and sold in thousands of open-air markets across sub-Saharan Africa. I seek to identify and include in my model the larger, regionally important hub markets that collect grain from surrounding smaller markets for trade with other hub markets. I do so in three steps. First, I include the 178 towns and cities in my 42 countries of interest which have a population of at least 100,000 people and are at least 200 kilometers apart . Second, I add smaller towns that are located at important road junctions or ports. Third, I add additional major towns in countries which still have high population-to-market ratios after my first two steps.

Together these steps produce a list of 263 markets . In order to be able to include a particular market in my model, I must have grain price data for it. Using my “ideal” list of 263 markets, I conducted an exhaustive search for monthly grain price series from these markets and obtained price series for 230 of them. I then used maps of road networks and navigable waterways to identify the pairs of these markets between which direct trade is feasible. A map of my final network of 230 markets with the 413 direct links between them is shown in figure 1.3. A complete list of markets and further details on the market selection process are contained in the appendix. The median market town has a population of 207,000, and the median transport distance between directly linked markets is 337 kilometers. Among the 230 markets, I identify 30 major ports that trade with the world market and include direct links between them and the most important world market for each crop . I also treat Johannesburg, South Africa like the world market for maize in my model due to its special circumstances3 . The monthly grain price series for the 230 markets cover a 10-year period from May 2003 to April 2013. The price series include series for the 6 cereal grains most produced and consumed in sub-Saharan Africa – maize , sorghum , millet , rice , wheat , and teff 4 . In each market, only a subset of these major grains are sold – 54 markets have price series for 1 grain, 111 have series for 2 grains, 24 have series for 3 grains, and 41 have series for 4 grains. Maize is by far the most common grain with price series from 180 of the 230 markets, followed by rice , sorghum , millet , wheat , and teff .Of the 512 total price series, 42% were obtained from the World Food Programme’s VAM unit and 25% from FAO’s GIEWS project, which both maintain online databases of staple food price series collected by themselves or by national government agencies. The remaining 33% were obtained directly from national ministries of agriculture and statistical offices or through USAID’s FEWS NET project, non-governmental organizations,grow bucket and other researchers. Each original source typically employs teams of surveyors who observe and record prices at multiple points of sale in each location on a weekly or monthly basis and then relay them to analytical teams in the capital city who compile and publish monthly and annual reports. The price series are not all complete — the average series has 72 observations worth of data.

The original price series are in local currency. Of the 512 series, 76% are identified as retail price series for quantities ranging from 0.5 to 3.5 kg, while the remaining 24% are identified as wholesale price series for quantities ranging from 50 to 100 kg. I convert all price series to USD/kg using monthly exchange rates and conduct a statistical analysis of 37 series for which I have both retail and wholesale prices that fails to reject a hypothesis of equality between retail and wholesale prices. This is consistent with interviews of market participants which suggest that separate retail and wholesale markets typically do not exist and that prices per kilogram often do not vary with quantity sold. Details on this statistical test as well as the grain types and data sources by market are contained in the appendix. Across all time periods and all markets, average prices are $0.41/kg for maize, $0.44/kg for wheat, $0.45/kg for both millet and sorghum, $0.58/kg for teff, and $0.84/kg for rice. Regressions with market fixed effects comparing price levels within particular markets show maize significantly cheaper than sorghum, which is significantly cheaper than millet and wheat, which are significantly cheaper than teff, which is significantly cheaper than rice . My next step is to acquire production and population data for sub-Saharan Africa and assign each of the 230 markets a monthly production and population to match its monthly prices. I start by obtaining annual national totals for production of all cereal grains from FAO and annual national totals for population from the UN Population Division. To allocate the production data by month, I use agricultural calendar data from FAO to divide the continent into three zones: a Northern Hemisphere zone with a single annual grain harvest in October , an Equatorial zone with a larger grain harvest in July and a smaller grain harvest in December , and a Southern Hemisphere zone with a single annual grain harvest in May 5 . Allocating the national-level data by market is more challenging. I first obtain GIS grid cell level data for population and production of each crop for the year 2000 at the 5 arc-minute level from the GAEZ project of FAO and IIASA and the Harvest Choice project of IFPRI and the University of Minnesota6 and use it to derive the percentage of national population and production of each crop belonging to each grid cell.

Under the assumption that these percentages stay constant during my study period, I combine them with my monthly national production and population data to get monthly production and population series at the grid cell level. The final step is to assign grid cells to particular markets. I do this by constructing market catchment areas following the methodology of Pozzi and Robinson 2008. The underlying assumption of this methodology is that if producers and consumers in a given grid cell have to choose one of the markets in the network at which to sell and buy their grain they will choose the market to which they can travel in the least time. To identify which of the 230 markets is the closest in terms of travel time for each of the 292,000 grid cells, I combine information from the following GIS datasets: the roads layer from the World Food Programme’s SDI-T database7 , the FAO Land and Water Division’s Rivers of Africa and Inland Water Bodies in Africa datasets, the USGS-EROS Global 30-Arc Second Elevation dataset, the European Commission Joint Research Centre’s Global Land Cover 2000 dataset, and the US Department of State’s Large Scale International Boundaries and Simplified Shoreline datasets. Following Pozzi and Robinson 2008, I assign different average travel speeds to different categories of road and different average walking speeds to different land cover classes, and I then adjust these speeds based on the degree of slope of the terrain. I assign inland water bodies and rivers with Strahler number of at least 4 a travel speed of zero 8 . I also assign a travel speed of zero to international borders so as to keep market catchment areas within countries to match my national production and population data. Combining all of this information, I assign each pixel a travel cost in minutes and then use a least-cost path algorithm to identify the minimum travel time from each grid cell to any market in the network. I then assign each grid cell to the market catchment area of its nearest market in terms of travel time. Figure 1.4 shows maps of estimated grid-cell level travel time to the nearest market in the network and the resulting market catchment areas. Once each grid cell has been assigned to a market catchment area, it is straight-forward to add up the production and population data for all of the grid cells in a given market catchment area and assign the total production and population to that market and its price series. Although my 512 price series do not include a price for every grain in every market, 86.3% of total cereal grain production in my countries of interest is covered by a price series in its associated market.