WPS8509 Policy Research Working Paper 8509 Transport Costs, Comparative Advantage, and Agricultural Development Evidence from Jamuna Bridge in Bangladesh Brian Blankespoor M. Shahe Emran Forhad Shilpi Lu Xu Development Economics Development Research Group July 2018 Policy Research Working Paper 8509 Abstract This paper studies the effects of a large reduction in transport in a difference-in-difference design. The analysis finds that costs on agricultural development in a developing country, the construction of Jamuna bridge led to increased adoption with a focus on the interactions among the comparative of technology (fertilizer, irrigation, greenness, and cropping advantage and transport costs of a location, and the trans- intensity) and reallocation of land from low-value and non- port intensity and value of a commodity. The paper extends perishable rice to high-value crops, pulses, and vegetables. the von Thunen model of land allocation to incorporate The evidence indicates spatial nonlinearity in the effects costly technology adoption and comparative advantage on cropping intensity and the reallocation of land in areas based on land productivity. The theoretical analysis predicts with comparative advantage in vegetable production. For spatial non-linearity in cropland allocation. A reduction in cropping intensity, the magnitude of the effect is large in transport costs leads to adoption of productivity-enhancing the intermediate distance (130–150 kilometers) from the inputs in the newly-connected region, and an increase in bridge. In areas with relatively higher vegetable productivity, the share of land devoted to the high-value transport-in- land allocated to rice declined, and land was reallocated tensive crop. The strongest effect is felt in areas that are from high-yielding variety rice to vegetables in the inter- not too near or too far from the center and have a higher mediate distance (110–150 kilometers). This improved land productivity in transport intensive crop. The empirical productive efficiency by aligning the cropping pattern more context of the analysis is the Jamuna bridge in Bangladesh, closely with comparative advantage. The bridge thus led which opened in 1998 and reduced the transport costs from to agricultural development through technology adoption, the poor hinterland in the northwest to the capital city higher cropping intensity, and reducing the spatial mis- (Dhaka) by more than 50 percent. Using sub-district level match between land suitability and crop choice. panel data, the paper implements doubly robust estimators This paper is a product of the Development Research Group, Development Economics. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/research. The authors may be contacted at fshilpi@worldbank.org. 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 views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Transport Costs, Comparative Advantage, and Agricultural Development: Evidence from Jamuna Bridge in Bangladesh Brian Blankespoor, World Bank M. Shahe Emran, IPD, Columbia University1 Forhad Shilpi, World Bank Lu Xu, World Bank Keywords: Land Reallocation, Technology Adoption, Cropping Intensity, Agriculture, Bridge JEL Classi…cation: R40; O18; O13; O16 1 Emails for correspondence: shahe.emran@gmail.com, fshilpi@worldbank.org. 1 Introduction The implications of segmented and imperfect rural markets for resource allocation and technological change in agriculture have occupied a prominent position in both theoretical and empirical literature in development economics from the 1970s onward (see, among others, Bardhan (1984), Braverman et al. (1993), Basu (1997)). High trade costs arising from the absence of transport and communications infrastructure are among the most important factors behind spatial segmentation of markets and may result in isolated village economies that are e¤ectively cut-o¤ from the urban growth centers. This paper uses a quasi-experimental study of a major bridge construction in Bangladesh, the Jamuna bridge, to analyze the e¤ects of a large reduction in trade costs on the spatial pattern of agricultural specialization and technology adoption in a lagging region of a developing economy. As noted by Donaldson (2015), there are two important advantages in focusing on agriculture when estimating the e¤ects of trade costs: (i) the main factor of production, land, is immobile, (ii) reliable measures of natural productivity of land are available (GAEZ, FAO).2 This Jamuna bridge crossing a major river (Jamuna) in Bangladesh opened to tra¢ c in 1998. The bridge connected the northwest region of Bangladesh to the growth centers in the east including the capital city Dhaka and the port city Chittagong. As a result of the bridge, freight costs were reduced by more than 50 percent and travel time from areas in the North-west to Dhaka city by 3-4 hours. Such a large reduction in transport costs pro- vides an excellent opportunity to examine the e¤ects on spatial organization of agricultural activities which may not be detectable with marginal infrastructure interventions such as improvements in existing roads, construction of rural feeder roads, or small bridges over canals in a village. The theoretical and empirical analysis of this paper focuses on two major issues. First, 2 Duranton and Venables (2018) note that the reallocation of resources in response to a decline in trade costs will be according to absolute advantage in the manufacturing and services sectors where the main inputs (labor and capital) are spatially mobile across regions of a country. This implies that to understand the role played by comparative advantage derived from initial endowment, agriculture is the most suitable sector. 1 the interaction between transport costs and natural land productivity in determining the spatial pattern of land allocation across crops and technologies (modern vs. traditional) is analyzed. We extend the standard von Thunen model of spatial specialization of crops to incorporate land productivity heterogeneity and costly technology adoption. An important prediction from the extended von Thunen model is that the pattern of crop choices and technology adoption across villages may exhibit nonlinearity with respect to distance from the urban center (spatial non-monotonicity). The positive e¤ect of the bridge on the share of land devoted to high-value transport intensive crops (e.g. vegetables) is felt most prominently in areas that are not too near or far from markets and also have higher relative land productivity in that crop. Second, a major caveat emphasized in the recent literature is that the estimated ef- fects of transport infrastructure in the target region may primarily be due to reallocation (reorganization), without any signi…cant e¢ ciency and growth e¤ects (a spatial zero-sum reallocation) (Redding and Turner (2014), Donaldson (2015)). Our analysis focuses on two factors central to e¢ ciency and growth in agriculture: comparative advantage based on crop suitability of land in a village, and technology adoption through investment in irri- gation and fertilizer. Reallocation improves productivity when land is allocated to crops according to comparative advantage rather than transport cost di¤erences. The interde- pendence between technology choice and crop choice can render some of the widely-used measures of reallocation and productivity change misleading. For example, a reduction in the share of land allocated to the modern variety of rice (HYV) in response to a reduction in trade costs may not imply adverse e¤ects on technology adoption.3 A third source of productivity improvements, largely neglected in the recent literature on the e¤ects of trade costs, is multiple cropping, made possible through irrigation in dry seasons. The changes in cropping intensity may be especially important in land-scarce countries where the tra- ditional extensive margin of agriculture in the standard von Thunen model cannot adjust 3 The share of High Yielding Variety (HYV) of rice in total cropped land is used by many as a measure of technological change in agriculture in Asian countries where rice is the major crop. Our theoretical model shows that the expansion of area under high-value transport intensive crops such as vegetables can come at the expense of less transport intensive crop produced under modern technology such as HYV rice. 2 to a reduction in transport costs. For empirical analysis, we use a subdistrict (upazila) level panel data set, and develop a di¤erence-in-di¤erence strategy where the comparison areas come from a region which were supposed to be connected to the growth centers in the center (Dhaka city) by the proposed, but not yet constructed, Padma bridge.4 The identi…cation is grounded on the following observation: the fact that Jamuna bridge was built in 1998, while the proposed Padma bridge is yet to be built, re‡ects idiosyncratic political factors (birth places of presidents and prime ministers) and thus can be treated as quasi-experimental. We take two additional steps to address potential biases in the DID estimates for the Jamuna treatment areas. First, we include upazila and year …xed e¤ects in all of the regressions. Second, we implement doubly robust estimators that combine two alternative reweighting schemes with regression adjustments as suggested by Kline (2011), Busso et al. (2013) and Moretti and Kline (2014). The empirical analysis uses four di¤erent indicators of technology adoption: proportion of land using chemical fertilizer, proportion of households owning irrigation equipment, cropping intensity, and green-ness depicted by Normalized Di¤erences Vegetation Index (NDVI) during dry months. The analysis of cropland allocation focuses on four crops covering a range of transport costs and prices: High Yielding Variety (HYV) of rice, total rice crop, pulses, and vegetables. Rice is the main crop in Bangladesh; approximately 75 percent of land is allocated to rice (BBS, 2014). Rice (and paddy) is not perishable and can be transported from remote areas, but high-value vegetables are perishable and need quick transport to the urban market. Pulses are also high-value crops, but similar to rice in terms of transport intensity. Since we have data on whether the land in a village is more suitable for rice or vegetables, a comparison of these two crops allows us to analyze the trade-o¤ between comparative advantage and trade costs.5 4 Most of the studies on the e¤ects of trade costs in the context of developing countries we are aware of rely on household level data. As pointed out by Donaldson (2015), among others, estimating the e¤ects at such a disaggregate spatial level is subject to potentially serious biases from spillover (the SUTVA assumption is violated). We focus on a much larger spatial unit, upazila. There were 490 upazilas in Bangladesh in 1991, and most of the upazilas had population between 150,000 and 350,000. 5 Unfortunately, the crop-speci…c land productivity data are not available for pulses. Land productiv- 3 The empirical evidence shows that, on average, subdistricts in the region connected by Jamuna bridge use chemical fertilizer in more land, have higher irrigation equipment ownership and higher cropping intensity, and show greater green-ness (NDVI), especially in the dry months. The results for the cropland allocation indicate a decline in the share of rice land, particularly HYV rice, and an increase in the share of pulses and a modest increase in the share of vegetables in the treatment region compared with the comparison region. The average e¤ects, however, conceal interesting spatial nonlinearity in many cases, driven, in part, by land productivity heterogeneity. While the e¤ects on fertilizer use decline monotonically with an increase in the distance from the bridge, the e¤ects on cropping intensity display a non-linear (concave) spatial pattern. The areas that are 130 km-150 km away from the bridge experience the highest increase in cropping intensity compared with the areas near to or farther away from the bridge. The pattern of reallocation of cropland in areas with relatively higher vegetables productivity: land moved away from rice, particularly from HYV rice, to vegetables in the intermediate distance (110-150km) and into rice particularly in HYV rice in areas farther than 150km. This result suggests that construction of the bridge allowed the cropping pattern in areas located in intermediate distance from bridge to align more closely to their natural advantages. This reallocation is associated with productivity gain even if we ignore the technology adoption since it allows vegetables to be grown in land better suited for vegetable production. We contribute to the literature on the e¤ects of better market access on agriculture in two ways.6 First, we provide evidence on how a large reduction in trade costs improves e¢ - ity for crop production is determined by million years of interactions of natural forces such as rainfall, temperature, wind, river, volcanic and glacial activities along with other terrain characteristics. Data on indicators of land productivity combining all these di¤erent factors have also become available recently. 6 Among recent papers, Jacoby (2000) and Shrestha (2016) …nd positive impact of better access to markets on agricultural land value in Nepal. Several studies also …nd higher propensities for households to use modern inputs (fertilizer, irrigation, high yielding variety of seeds) and sell in the markets (Shamdasani (2016) for India, and Shrestha (2016) for Nepal, Ali et al (2016) for African countries, Kyeyamwa et al. (2008) for Uganda, Omamo (1998) for Kenya) and agricultural yields (Ali et al. (2016)), Dorosh et al. (2012) for Sub-Saharan Africa. The positive impacts of better access to market is con…rmed in the case of developed countries as well (see Donaldson and Hornbeck(2016), and Atack and Margo (2011), Haines and Margo (2006), Chandra and Thompson (2010)). See also Costinot and Donaldson (2016) and Costinot, Donalson and Smith (2016) for broader impacts of trade costs on agriculture. 4 ciency in resource allocation in a poor agricultural region by reducing the spatial mismatch between land productivity and crop choice because of heterogeneity in transport intensity and unit value of di¤erent crops. A lower trade cost allows better matching of crops with suitable land, consistent with comparative advantage. Although there is substantial evi- dence in the existing literature that better market access due to lower transport costs leads to crop diversi…cation, especially in favor of the noncereal crops, it is not clear how to interpret this …nding without evidence on the role played by land productivity heterogene- ity.7 If land in the treatment areas is less suitable for non-cereal production than that in rest of the country, then increased diversi…cation into non-cereal crops may not improve s income in the treatment over-all productive e¢ ciency even though it increases a farmer’ areas. Second, we provide evidence on spatial nonlinearity where the areas in the treatment areas located in the intermediate distance from the bridge experience the strongest e¤ects on cropping intensity and reallocation of land with comparative advantage in vegetables. This spatial nonlinearity has two important implications: (i) the standard practice of using areas close to a bridge (or other transport infrastructure) as the treatment catchment is likely to underestimate the e¤ects of bridge construction on reallocation of land, and (ii) large transport infrastructures such as a bridge may result in spatial inequality within the treatment region, even though the average e¤ect is positive. The rest of the paper is organized as follows. The next section sets up an extended von Thunen model of cropland allocation and technology production and derives testable predictions about the e¤ects of a reduction in the cost of crossing the river. Section (3) discusses the background of the Jamuna bridge. We develop the empirical strategy in the next section, and discuss the data sources and construction of the variables in section (5). Sections (6 and 7) are devoted to the empirical results. The paper ends with a summary of the …ndings and their implications for the literature. 7 Shamdasani (2016) provides evidence that a better access to markets increases the land devoted to non-cereal crops in India, and Emran and Shilpi (2012) …nd that market access leads to crop diversi…cation in Nepal. 5 2 Theoretical Model The Basic Set-Up We extend the canonical von Thunen model of crop land allocation to incorporate tech- nology adoption and land productivity heterogeneity. The possibility of investment in technology can introduce non-linearity in the cropping pattern with respect to the distance from urban markets. The standard von Thunen model focuses on the transport cost varia- tion across areas, but assumes away heterogeneity in land productivity. This assumption, however, excludes any interaction of transport cost with natural advantage in determining the cropping pattern. Heterogeneity in land productivity is introduced in the model to allow for natural (and comparative) advantage to vary across areas. The productivity and transport cost heterogeneity help to explain deviation of the actual cropping pattern from inherent natural advantage when transport cost is especially high. This simple model pro- vides a ‡exible framework to investigate the impacts of a large transport investment such as a major bridge on technology adoption and cropping pattern. Geography We consider the geography where all possible locations are ordered along a line between interval [H1 ; K1 ] (please see Figure 1). The line is divided into three segments by the presence of two rivers. The …rst river (RVH ) is located between H0 and CH ; and the second (RVK ) between CK and K0 : As shown in Figure 1, the presence of rivers de…nes three regions: H = [H1 ; H0 ]; C = [CH ; CK ]; and K = [K0 ; K1 ] where C is the central region and the other two are underdeveloped lagging regions. There are continuum of locations in each of the regions. Each location in region H (K ) is indexed by h (k ); where h (k ) also depicts the distance from riverbank CH (CK ): In the absence of bridges, each river is crossed by using ferry. Two rivers are identical in width and water ‡ow resulting in identical cost of ferry. The crossing of the river using ferry involves a product speci…c cost (FHi = i FH = FKi ) where FH (= FK ) is the travel time to cross the river and i is the marginal cost of crossing the river for product i: To avoid confusion, the notational conventions are: the subscript 6 i is the crop index (i = X; Y ) and subscripts h and k are the location index in regions H and K respectively, and superscripts T and M denote traditional and modern technology respectively. Each location is endowed with one unit of land. Regions H and K are identical to each other with one exception that they are located on the opposite sides of the region C: Region C is a central region consisting mostly of urban population and constitutes the primary market for agricultural goods. Following the standard von Thunen model, we assume that crop prices are determined in the urban markets in the central region C , and are exogenous for farmers in the peripheral regions H and K . Since regions H and K are identical, we can characterize the spatial equilibrium in this economy by focusing on region H . The e¤ect of bridge can be posed as changes in equilibrium outcomes in H in response to building a bridge relative to that in K which remains cut-o¤ without a bridge. Production Technology and the Crop System Each region can produce two crops: X and Y . Both crops can be produced using a traditional and a modern technology. While under traditional technology (T ), each unit of land in an area h can produce Aih unit of output of crop i = X; Y and h 2 H . Farmers can invest in an indivisible irrigation equipment per unit of land (Z ) and without loss of generality, we set Z = 1: The irrigation equipment facilitates the adoption of modern technology (M ) that increases land productivity multiplicatively by i > 1; i = X; Y: To purchase the irrigation equipment, farmers in location h need to pay PZh . Rivers and Transport Costs Let Pi be the price of crop i in the urban central region C where i = X; Y . We assume that X is perishable and transport intensive but also high-value (PX > PY ). Shipping crop i within the region is subject to iceberg cost (1 > i > 0) such that a unit of output shipped from distance h becomes (1 i h) at the riverbank. Crossing the river involves ferry cost and thus unit price of i at any location h is equal to Pih = Pi (1 i Fh i h) = Pi dih ) where Fh is distance of the river in terms of hours of ferry travel and i is product speci…c unit cost of ferry crossing: Irrigation equipment is imported from the central region where 7 its price is …xed at PZ : Crossing the river and traveling inside region H adds to cost of acquiring an irrigation equipment, so its price at h is equal to PZh = PZ (1 + z Fh + z h) = PZ dzh where z is cost of shipping the equipment from riverbank to location h and z is the unit cost of river crossing: Denoting revenue of crop i by rih = Pi Aih ; i = X; Y the bid rent Rih of crop at a distance h 2 [0; H1 ] and under di¤erent production technologies can be written as follows: M Rih = i rih dih PZ dZh under modern technology T Rih = rih dih under traditional technology where superscripts M and T refer to modern and traditional technologies respectively, i = X; Y : Without land productivity heterogeneity across locations, the revenue of crop i is ri = Pi Ai : The slope of the bid rent curve for i under traditional technology is determined by its price, transport cost and land productivity. The slope is steeper and intercept is larger if productivity (Ai ) and price (Pi ) are higher. Thus the bid rent curve for a high- value crop such as vegetables exhibits steeper slope when compared to a low-value crop such as rice. Also, the higher price and transport cost of irrigation equipment imply a steeper slope. Pro…t maximization by the farmers involves two decisions: whether to adopt modern technology and which crop to produce. The optimal decision can be described as: T M T M Rh = M axfRXh ; RXh ; RY h ; RY h ; 0g where Rh is the equilibrium land rent at the location h. Each location produces the crop with the technology that provides the highest land rent, and the equilibrium land rent thus encompasses the upper envelope of all bid rent functions. Given the assumption that crop j j X is more transport intensive, the slope of RXh is steeper than that of RY h for j = M; T: j j Farmers in region H will not produce X if RX 0 RY 0 at the riverbank where h = 0. To j j rule out this trivial case, we assume that RX 0 > RY 0 : The extensive margin of cultivation 8 can be de…ned as H E = minfH ; H1 g where H is determined by setting RY H = 0; since Y is less transport intensive crop. Equilibrium Allocation of Land and Technology Adoption without Land Heterogeneity As a benchmark, proposition 1 below summarizes the equilibrium spatial con…guration of technology adoption and cropping pattern in the absence of land productivity heterogeneity across locations implying that Aih = Ai , and the revenue rih = ri = Pi Ai . This helps us to see how technology adoption alone can introduce non-linear pattern of crop land allocation with respect to the distance to markets. We relax this assumption later. Before describing equilibrium con…guration of technology adoption and cropping pattern, we introduce some notations to help the exposition. Let ij denote crop i produced using technology j where i = X; Y and j = M; T . Let hjm j jm in be the distance from riverbank such that Ri (hin ) = m jm Rn (hin ); i; n = X; Y and j; m = M; T:Thus, hM T M T XY de…nes the intersection of RX and RY and so on. Proposition 1: Under the assumptions that land productivity in each location varies j across crops but is the same for a given crop across locations ( Aih = Ai ) and that RX 0 > j RY 0 , j = M; T , the spatial equilibrium con…guration of technology adoption and crop land allocation depends on the cost of irrigation equipment and the transport costs of crops and irrigation equipment: ^Z ); (i) If the price of irrigation equipment is high and above a threshold ( PZ > P then the farmers do not adopt modern technology and crop X T is produced in all locations ^ T T ] and crop Y T in relatively remote locations h 2 ^ T = [0; h closer to the bridge h 2 H X XY ^ T T ; H E ]; ^ T = [h H Y XY (ii) If PZ is lower than a threshold, PZ < PZ ; then all of the farmers in region H 9 produce both crops using modern technology, crop X M is produced in all locations in the M interval HX = [0; hM M M MM E XY ] located closer to the bridge and crop Y in h 2 HY = (hXY ; H ] located farther from the bridge ; (iii) When the price of the irrigation equipment falls into an intermediate range de…ned ^Z , the pattern of technology adoption and allocation of land to crops with by PZ < PZ < P respect to distance from the riverbank ( h ) depend on the relative transport costs. Farmers in locations h 2 HiM = [0; hM i ] use modern technology in producing crop i. Depending on the relative lengths of hM i ; i = X; Y; determined by the di¤erential transport costs, three subregions can be de…ned in terms of land use. Crop X will be produced using modern technology in the subregion closest to the riverbank and crop Y using traditional technology in the subregion farthest from the riverbank. In the intermediate subregion, either crop Y will be produced using modern technology or crop X using traditional technology or both. Proof: The cost of irrigation equipment is the lowest at the riverbank (h = 0) and increases at the rate of z with an increase in distance from the riverbank (h). Noting M that, at the riverbank (i.e., location h = 0), RX M ^ 0 > RY 0 , PZ in proposition 1(i) can be M determined by setting RX ^ T ^ 0 ( PZ ) = RX 0 : Intuitively, PZ is the price at which the bid rents for crop X at the riverbank are equated across traditional and modern technology. With ^Z ; technology adoption is not feasible in any location h 2 H , and thus both crops PZ > P T T are produced with the traditional technology. Because RXh is steeper than RY h ; areas ^T = H closer to the riverbank h 2 H ^ T T ] are planted with X , and areas farther ^ T = [0; h X X XY ^ T T ) = RT (h ^ T T is determined by setting RT (h away with crop Y, where h ^ T T ): XY X XY Y XY In proposition 1(ii), threshold of irrigation cost PZ is determined by equating the bid rents for crop Y at the boundary of extensive margin H E with and without adoption of M T technology, i.e., RY H E ( PZ ) = RY H E :The intuition for allocation of land is similar to that for proposition 1(i) where hM M M MM M MM XY is determined by equating RX (hXY ) = RY (hXY ): ^Z ; RM > RT at h = 0 and RM < RT at h = H E ; 8 i = X; Y: Farmers For PZ < PZ < P ih ih ih ih i such that Rih = Rih producing crop i will use modern technology up to the distance hM M T for h 5 hM M T M i , and Rih < Rih for h > hi : The border of the zone of modern technology for 10 each crop i (hM M T i ) is determined by equating Rih and Rih : ri0 di0 ( i 1) PZ dZ 0 hM i = for i = X; Y (1) ( i 1) i ri0 + z PZ where di0 = (1 i Fh ) and dZ 0 = (1 + z Fh ):Under the assumptions that crop X is M M more transport intensive and also of higher-value, i.e., x > y and RX 0 > RY 0 , the slope of bid rent curve for X ( x rx0 x + z PZ ) is greater than that for Y ( y ry0 y + z PZ ) when both are produced using the modern technology. The larger is the transport cost ( x ); the greater is the possibility that hM M T T X < hY : Similarly, RX is steeper than RY . The slopes and intercepts of these four bid rent functions determine the equilibrium cropping pattern. In the appendix, we describe the possible outcomes that may result from di¤erent values of transport cost parameters along with prices and land productivity di¤erences of the two crops. The regularity that emerges from these outcomes is that transport intensive X is produced using modern technology near the riverbank and less transport intensive crop Y is produced under traditional technology in the subregion farthest from the riverbank. In the intermediate sub-region, either X is produced under traditional technology, or Y using modern technology or both. When both are produced, their relative location within the sub-region is determined by the underlying slope and intercept parameters. Technology Adoption and Cropping Pattern in Bangladesh It is clear from proposition 1 that many di¤erent outcomes and spatial con…gurations of technology adoption and cropping pattern are possible depending on the magnitudes of productivity parameters, transport costs, product prices and the cost of technology investment. Before describing the possible impact of constructing a bridge over the river, we highlight some distinctive features of land use in Bangladesh that help to narrow down these possibilities. First, population density in Bangladesh is exceptionally high even in rural areas (800/sq km) and all available agricultural land has been under cultivation for 11 many decades.8 To account for the land constraint in agriculture, we relax the standard von Thunen assumption that opportunity cost of land is zero at the extensive margin by normalizing transport cost of Y to zero ( y = 0). This assumption implies that H E = H1 : Second, the HYV rice is a more water and thus irrigation intensive crop than vegetables. We assume that irrigation boosts productivity of Y (rice) more than that of X (vegetables) ( y > x ), but because of higher value of vegetables, the bid rent at the riverbank is higher j j T for vegetables, and we have RX 0 > RY 0 , j = M; T . RX curve (line) is assumed to be M ‡atter than RY : ( x r x0 < z PZ ), partly because of indivisibility of irrigation equipment (PZ ): While viewed as a tax, this assumption implies that the transport tax on irrigation M is higher than that on crop X . Note that the slope of bid rent curve RY h is z PZ whereas T for RXh , it is x r x0 : As shown in the appendix A, several di¤erent cropping patterns may result depending on the slopes and intercepts of the bid rent functions. We focus on the equilibrium where both crops are produced under both technologies.9 This equilibrium land allocation is illustrated in Figure 2a. The equilibrium shows inter- esting and non-linear spatial pattern. The area near the riverbank (closest to the urban M markets in C ) are planted with the transport intensive crop X (h 2 HX = [0; h1 = hM M XY ]) M followed by a subregion that produces Y (h 2 HY = (h1 = hM M MT XY ; h2 = hY X ]) , and both crops are produced using the modern technology. Farther away, land use reverts back …rst T to X (h 2 HX = (h2 = hM T TT T TT E Y X ; h3 = hXY ]) and then to Y (h 2 HY = (h3 = hXY ; H ]) , both produced under the traditional technology. It is illustrative to consider the cropping pattern that would have resulted from a traditional von Thunen set up without productiv- ity heterogeneity and technology adoption. The equilibrium outcome would be to produce X in the interval (0; hT T TT E XY ) and Y in (hXY ; H ]: The possibility of technology adoption in- troduces non-linearity in cropping pattern with respect to distance from market (bridge). This non-linearity is often taken as an evidence of reverting back to subsistence (Fafchamps and Shilpi (2003)). The modi…ed von Thunen model presented here provides an alternative 8 According to the 2008 agricultural census, arable land per person is only about 0.0482 hectare. 9 Both crops are produced using both technologies if ( x rx0 > z PZ ): But the cropping pattern in this case is di¤erent from what is shown in Figure 2a. In this case, X M is produced in the interval nearest to the riverbank followed by X T ; then Y M and Y T : 12 explanation for this non-linearity which arises because of higher transport cost of indivisi- ble irrigation equipment relative to that of perishable high-value crops. Before introducing land productivity heterogeneity, we consider the possible e¤ects of the bridge on technology adoption and cropping pattern in the benchmark model without productivity heterogeneity. The Impact of the Bridge on Technology Adoption and Cropping Pattern Suppose a bridge is constructed over river RVH , but no bridge is built over RVK : A reduc- tion in the cost of crossing the river (FH ) increases prices of both crops received by the farmers and reduces the price of irrigation equipment paid by the farmers. Proposition 2 summarizes the predictions regarding the impacts of bridge on technology adoption and cropping pattern if bridge led to a reduction in cost of river crossing. Proposition 2: A decrease in the ferry cost (FH ) leads to the following results: (i) extends the zones within which farmers adopt modern technology, (ii) increases the extensive margin of cultivation if H E < H1 , (iii) increases land allocated to crop X if H E = H1 and x = y and where x and y are unit ferry/river-crossing costs for X and Y respectively; and (iv ) its impact on cropping pattern in the intermediate subregion is ambiguous. The larger is the decrease in ferry cost, the greater is the extension of zones of modern technology and extensive margins. Proof : Proposition 2(i) follows directly from equation 1. A reduction in FH increases hM i by increasing the price received by farmers for their crop and by decreasing the price they need to pay for the irrigation equipment. Proposition 2(ii) follows from the fact that at the edge of the extensive margin, Y is produced either using modern or traditional j j technologies. At H E < H1 ; RY = 0; j = M; T: As a lower ferry cost increases RY , it @H E follows that @Fh < 0: @hjm For propositions 2(iii) and 2(iv ), we show in the appendix that in @Fh < 0; i; n = X; Y and j; m = M; T; if x = y: A lower ferry cost shifts all of the bid rent curves upward and 13 thus pushes all intervals of crop specialization towards the farthest border of region H (H1 ). This unambiguously increases land under X near the riverbank if H E = H1 : The impacts in the intermediate zone depends on the initial con…guration of cropping pattern which, as shown in proposition 1, in turn is determined by the cost of technology adoption and intercepts and slopes of bid rent functions. In the aggregate, the share of land allocated to X increases as bridge pushes all the circles of crop specialization toward the farthest areas and because the extensive margin of land cannot be increased. (2.2) Implications of Land Productivity Heterogeneity The model so far assumed land productivity of each crop to be homogeneous across areas. To illustrate how heterogeneity in land productivity across areas can a¤ect technology adoption and cropping pattern, we focus on a simple case where land productivity of Y is homogeneous across areas but that of X varies with distance in the following manner: Axh = (1 + h) Ax0 (2) where can be positive or negative. A positive indicates increasing land productivity with an increase in the distance from the riverbank and vice versa. The bid rent function for X becomes nonlinear when land productivity changes with respect to the distance from the M riverbank. As we show in the appendix, the bid rent function RXh is concave (convex) if > 0 ( < 0): For < 0; the bid rent for crop X produced using either technology declines with the distance on account of a decrease in land productivity in addition to transport cost. In other words, the farmers located farther away from the riverbank face double disadvantages due to the higher transportation costs and a lower land productivity. The pattern of technology adoption and land allocation described in proposition 1 would hold however with band/intervals for crop X becoming shorter. Heterogeneity in land productivity with respect to the distance to the riverbank either accentuates or o¤sets the impacts of transport costs on technology adoption and land allocation described in proposition (1). 14 For > 0, land productivity increases with distance raising bid rents above what it would have been with = 0 . The productivity increase can o¤set the decrease in bid rent due to higher transport cost depending on the magnitude of . But the bid rent h i 1 z PZ M curves are now concave. For 1 Fh x+ rX 0 ; bid rent curve RXh is downward x X sloping but lie above the straight line bid rent curve for = 0 described in proposition h i 1 z PZ T 1 (see Figures 2a and 2b). For 1 x Fh < 1 Fh x + rX 0 , bid rent curve RXh is x x X concave but downward sloping. The pattern of technology adoption and land allocation described in proposition 1 still holds, but the intervals for crop X produced under modern and traditional technologies both expand.10 With a large enough ; it may become feasible to adopt modern technology in the production of X in the intermediate sub-region. The basic insights derived from the parametric land productivity function carry over to the case where land productivity is not distributed monotonically over space according to a formula as in equation (2). With random distribution of land productivity parameter over geographic space, the probability of technology adoption and the amount of land allocated to a crop will increase with an increase in land productivity in the intermediate subregion. M M By assumption, RX 0 > RY 0 at the riverbank (h = 0) implying that PX X AX 0 > PX AY 0 PY Y AY 0 . However, this condition may hold even if AX 0 < AY 0 as long as > : Y PY X AX 0 Thus X M is produced near the riverbank because of its high value even though the land there may not be the most suitable for its production. On the other hand, at much far- ther distance from the riverbank, the high transport cost of X may more than o¤set any advantage from a higher land suitability, resulting in the land being used in less transport intensive crop Y . Proposition 3 below summarizes the key insights when land productivity of a crop can vary across areas. Proposition 3: A Moderate land productivity heterogeneity may not a¤ect the technol- ogy adoption and land allocation pattern in the nearest and the farthest sub-regions from the central market while its e¤ects are felt more prominently in the subregion located at the h i 10 1 z PZ When is large enough > 1 x + , production of X may become feasible even if x Fh X rX 0 M M RX 0 < RY 0 at the riverbank. 15 intermediate distance. In the intermediate sub-region, the higher is the land productivity of a crop relative to that of other crops, the higher is the possibility that it is produced in that location. Land Productivity Heterogeneity in Bangladesh The impacts of the bridge depend on the distribution of land productivity with respect to the distance to the bridge. In Figure 4a, we plot the non-parametric graph of subdistricts top-ranked for vegetables relative to subdistricts top-ranked for rice production with respect to the distance to the bridge site. The relative productivity of vegetables (X ) is lower in the subregions located nearest and farthest from the bridge site and higher in the intermediate sub-region. For simplicity, we divide region H into three sub-regions V1 ; V2 and V3 such that V1 is located at the riverbank and consists of all areas in distance interval [0; h1 ), and V2 in the interior and covers all areas in distance interval [h1 ; h2 ): Subregion V3 is located even farther away at distance h2 from the riverbank and covers all locations in distance interval [h2 ; H1 ]. To reproduce the relative productivity of X , we normalize land productivity for Y to unity in each location AY h = 1. We assume that land productivity for vegetables X is equal to AX in V1 and V3 but higher in V2 (AX 2 > AX ): To highlight the source of mismatch between natural advantage and the actual cropping pattern, we assume that (AX 2 > AY = 1 > AX ): In Figure 2a, the borders of the three subregions are identi…ed and J the bid rent curves for X (labeled Rx 2 ; J = M; T ) are shown in brown color. As shown in Figure 2a, actual land use pattern does not overlap well with natural advan- tage re‡ected in land productivity. This mismatch arises partly because of transportation costs for irrigation equipment and partly because of higher value of transport intensive perishable product (X ). Without transport cost of equipment, all land in h 2 [hM M XY ; H1 ] should be planted with Y . On the other hand, if there were no cost of transporting X , then all land in region H should be planted with high-value crop X , resulting in a mismatch of natural advantage and actual cropping pattern in V1 and V3 . 16 Land Productivity Heterogeneity and the E¤ects of the Bridge The impacts of bridge on technology adoption and cropping pattern vary with land pro- ductivity. Proposition 4: A reduction in river crossing cost increases the probability of technol- ogy adoption and land use in a crop that is transport intensive and has relatively better land productivity and this e¤ect is most prominent in the intermediate sub-region. The expan- sion of land under transport intensive crop ( X ) may come at the expense of less transport intensive crop ( Y ) produced under modern technology. M M To see the intuition behind this, we start with initial equilibrium where RY > RX 2 > M M RX at h = h1 ; where RX 2 is the bid rent function at land productivity AX 2 : The minimum reduction in Fh that is required to switch land from crop Y to crop X is then Fh = M 0 (h ) R M 0 (h ) RY 1 1 X2 : The higher is AX 2 ; the lower is the reduction in ferry cost needed to induce x rx0 y ry 0 a change in cropping pattern. Note also that this expansion of crop X produced under modern technology in V2 comes at the cost of a decline in land to crop Y produced under modern technology (Figure 3). Similarly, large enough decrease in Fh can make technology adoption feasible for Y in V3 ; shrinking land allocated to both X and Y produced using traditional technology. As a result of bridge, land allocated to modern variety increases at the expense of traditional variety for each crop, the e¤ects of the bridge on total land allocated to each crop at the regional level may not change. 3 Costs of Crossing the River and the Jamuna Bridge Two major rivers in Asia –the Ganges (locally known as Padma) and Brahmaputra (locally known as Jamuna) run through Bangladesh slicing the coutnry into three separate regions (see map 1). These two rivers e¤ectively cut-o¤ the north-west and southern regions of the country from the growth centers in the middle and east of the country. The 4.8 kilometer long Jamuna bridge connected the north-west region to the main growth center (capital s population city Dhaka). The north-west region accounted for 24.5 percent of the country’ 17 of 105 million in 1991. The bridge led to a signi…cant reduction of travel time and transport costs. Prior to the bridge, river crossing by ferries took more than 3 hours. During heavy tra¢ c periods, the average waiting time at the ferry ran as high as 36 hours (World Bank (1994).11 After the opening of the bridge in June 1998, travel time to cross the river decreased to less than an hour (including waiting time). The bridge cut the average travel time by 4 hours during the normal tra¢ c time, and reduced the freight costs by a half. Travel time by truck between city of Bogra in the north-west region and the capital city Dhaka was reduced from 20 hours to 6 hours.12 The bridge thus led to a very substantial reduction in transport time and costs. We utilize this substantial reduction in transport cost to estimate the e¤ects of trade costs on the spatial pattern agricultural development. To identify the e¤ects of the bridge, we exploit the fact that the southern part of the country is also separated from the growth centers in the capital city Dhaka and port city Chittagong by Padma river. While bridges were proposed to be built on both Padma and Jamuna rivers to connect the southern and north-western regions of the country respec- tively, the bridge over Jamuna river received priority due to exognous factors. Seveteen years of the two decades between 1977 and 1999, Bangladesh was governed by leaders (Zi- aur Rahman, Hossain M. Ershad and Khaleda Zia) who hailed from the north-west region. The construction of the bridge required large investment for which donor funding was nec- essary. The fact that the north-west region su¤ered disproportionate fatality during the 1974 famine made it easier to secure donor funding for Jamuna bridge …rst. The construc- tion of the proposed Padma bridge started only in December, 2015 under the current prime minister whose ancestral home is located in the sourthern region. We use the sub-districts (upazilas) in the southern region as comparisons for the treatment areas in the north-west. 11 The estimate is for 1993. 12 It took much longer for trucks to cross the river by ferry because buses carrying people had priority in getting access to the ferry boats. 18 4 Empirical Strategy To estimate the e¤ects of the Jamuna bridge, we compare the subdistricts in the treatment area with the subdistricts in the appropriately de…ned comparison area with similar pre- bridge characteristics. We use the following …xed e¤ect di¤erence-in-di¤erence (FE-DID) speci…cation: Yijt Yijt 1 = + (T Y r) + 1 Zijt0 + 2 Zijt + T + Y r + "ijt (3) where Yijt is the outcome variable j in subdistrict i and period t. T is a dummy which takes a value of unity if a subdistrict is located in the service area of Jamuna bridge and zero if it is located in the comparison area. Y r is a dummy that takes the value of unity if the year is after 1998 and zero otherwise. Zijt0 is a matrix of pre-bridge characteristics and Zijt is a matrix of contemporaneous and exogenous characteristics (e.g. rainfall). We implement the location …xed e¤ects by …rst di¤erencing of the dependent variable which wipes out the location speci…c and time-invariant factors, whereas captures the common shocks. In this formulation, the estimate of is the treatment e¤ect of the bridge. The vector of pre-bridge covariates includes log of population density in 1991, an index of suitability of land for crop production, dummies for whether the land quality in a subdistrict is top-ranked for rice or vegetables. Since our focus is on agricultural development, the variation in rainfall across subdistricts may in‡uence the estimates of treatment e¤ects. To guard against this possibility, we include contemporaneous rainfall as an additional comparison. To correct for possible spatial correlations, all regressions cluster standard divisions’in local term).13 errors at the regional level (‘ In addition to the …xed e¤ect DID (FE-DID) estimates using OLS for equation (3), we undertake two weighting schemes using the pre-bridge characteristics to improve the com- parability of treatment and comparison areas. The …rst approach uses propensity scores 13 The country is divided into 7 regions/divisions, each of treatment and control areas comprises of two divisions. 19 from a logit model of the probability of being included in the treatment area using the pre-bridge characteristics. The predicted probabilities are used to de…ne weight for each observation (subdistrict) in the comparison subset. The logit regression include pre-bridge characteristics such as log (population in 1991), the ranking of upazilas in terms of suit- ability of land for vegetables production and for rice production, and the distance to bridge (the Jamuna bridge for the treatment and the proposed Padma bridge for comparison) as controls. For vegetation index, distance to the capital city Dhaka is also included in the controls. Note that the DID regressions directly control for the pre-bridge characteristics, and thus the approach is similar to the doubly-robust estimators proposed by Robins et al. (1994) and Wooldridge (2007). We call this approach LWRA (logit weighted and re- gression adjusted) estimator. The second estimator developed by Kline (2011) and Moretti and Kline (2014) uses weights generated from the Oaxaca-Blinder approach as suggested by Kline (2011). The variables used for the Oaxaca-Blinder weights are the same as the ones used in computing the logit probability weights. The Oaxaca-Blinder estimates of the e¤ects of bridge are also doubly robust, as discussed by Kline (2011). 5 Data To estimate the e¤ects of Jamuna bridge on the pattern of agricultural specialization and technology adoption, we rely on subdistrict (upazila) level panel data. Several data sources, including agricultural and population censuses and di¤erent GIS databases, are utilized to create the dependent and explanatory variables in our analysis. The agricultural censuses are available for two years (1998 and 2008). Agricultural specialization is measured by the share of total cropped land allocated to rice, pulse and vegetables.14 Rice is the staple crop and less perishable whereas vegetables are high-value but perishable and transport intensive. Pulse is also high-value, but less transport intensive, similar to rice. Cropping intensity depicts multiple use of land for crop production and thus captures agricultural in- 14 Total cropped land is equal to total agricultural land in use multiplied by cropping intensity where cropping intensity measures the number of times same piece of land is used in cultivation. 20 tensi…cation, especially through irrigation during the dry season. Though agricultural land is approximately …xed in Bangladesh, multiple use of the same land as re‡ected in higher cropping intensity can in practice extend the availability of land similar to an expansion of the extensive margin in the standard von Thunen model. From the census data, two indi- cators of technology adoption are considered: the share of land where fertilizer is applied and the average ownership of shallow tube-wells, the main equipment used in irrigation, in an area. The data for crop land allocation and technology adoption are drawn from two agricultural censuses (1998 and 2008). The data for 2008 come from the sample survey con- ducted as a part of the 2008 agricultural census. For 1998, the data set consists of about 30 percent of the unit records from agricultural census. To make data comparable, we de‡ate all of the variables by total cropped land in the relevant upazila, with the exception of ir- rigation equipment. Irrigation equipment is measured by proportion of households owning a shallow tube-well in the upazila. Shallow tube-well is the most common equipment used for irrigation in rural Bangladesh. We supplement the census data by using remote sensing data on normalized di¤erence vegetation index (NDVI) which depicts green-ness of an area/pixel. Using satellite data on strong plant re‡ectance, The normalized di¤erence vegetation index (NDVI) is de…ned using sattellite data on strong plant re‡ectance (see appendix B for more detail). To minimize the gaps in the early satellite data, we restrict our analysis to the period covering 1996-2014 and de…ne quarterly averages from bi-weekly data.15 The …rst quarter corresponds to the driest months in the year whereas third quarter covers the monsoon time. While NDVI data have been used to examine changes in forest covers, its use in detecting changes in agricultural practices in the context of Bangladesh is aided by couple of factors. The forest cover is very limited in the country, concentrated mainly in three areas: Sundarban in the south, Hill tract districts in Chittagong and the tea gardens in Sylhet division. The rest of the land outside of urban settlements are utilized in agriculture. The land constraint for agriculture is evident in the average farm size which is less than an acre. For the 15 The NDVI data are available for a su¢ ciently long period of time (bi-weekly data from mid 1980s to 2014 but not for every year before 1996). 21 empirical analysis, we restrict our sample to the areas not covered by forest/tea gardens. Second, the leaf canopy on cultivated land changes depending on the utilization of land as well as irrigation, particularly in the dry months (…rst and last quarters). Thus changes in NDVI can capture changes in technology adoption and agricultural intensi…cation. In the empirical analysis, we consider annual average vegetation index along with its average during two relatively dry seasons: …rst and fourth quarters of the year. To create a consistent upazila level panel from the censuses and the remote sensing data, digital maps are used to identify the borders of upazilas overtime. The upazilas in the panel are de…ned using 1990 upazila boundaries. For upazilas that did not change boundary overtime, the matching exercise is simple. For upazilas that changed boundaries, s upazilas. Total number of we use area weights to link the newly created upazilas to 1990’ upazilas in our data is 122 in the treatment region (Jamuna bridge service area) and 105 in the comparison region (Padma hinterland). Among other variables, population data are drawn from census. The original data on rainfall are from the Climate Research Unit (CRU) of the University of East Anglia. The CRU reports estimated monthly rainfall for most of the world at the half degree resolution from 1902 to 2014. The CRU method combines weather station data with other relevant information to arrive at the estimates. To estimate the sub-district (upazila/thana) level rainfall from the CRU data, we use area weighted averages. The crow-‡y distance between the geographical center of a subdistrict to the Dhaka city is estimated using GIS software. Data on agro-ecological zones are drawn from the Bangladesh Water Board database which s broader GAEZ database.16 The advantage of was prepared as background work for FAO’ this data set is that in addition to providing information on agro-ecological zones, it also ranks land in terms its suitability to production of certain crops. Ranking is provided in a scale of 1 to 5 with 1 being best. This ranking is available for rice and vegetables but not for pulses. 16 These detailed data sets were put together by researchers and scientists at Bangladesh Agricultural Research Council in collaboration with FAO researchers under a project by the Water Board and formed the basis for Global Agro-Ecological Zone data on Bangladesh compiled by FAO. 22 6 Evidence on the Plausibility of the Research Design (6.1) Comparability of Treatment and Comparison Areas The treatment sample consists of 122 upazilas, located in the North West (henceforth NW) region that was connected by the Jamuna bridge to the central region where the capital city Dhaka is located. The upazlias in the south that remained cut-o¤ from Dhaka city due to the delay in constructing a bridge over Padma river serves as our comparison. After dropping 4 upazilas that constitute the protected natural forest in Sundarban, our comparison sample consists of 105 upazilas. To see whether the south provides a good counterfactual region for the treatment region (NW), we provide summary statistics during the pre-bridge period in Table 1. Column 1 reports the means for the treatment areas in the NW and columns 2-4 report unweighted and weighted means for the comparison areas in the south, and the last three columns provide the respective p-values of a test of the null hypothesis that the di¤erence between the treatment and comparison upazilas is zero. As explained in the econometric strategy section above, the weights are derived from Logit and Oaxaca-Blinder regressions. The top panel in Table 1 reports the evidence on land productivity measured by the average rank of land in terms of its suitability in crop production. This suitability index can be taken as a measure of natural advantage of land. A higher average for the rank indicates less suitability and less land productivity for the crop in question. The evidence suggests an absence of statistically or numerically signi…cant di¤erences in land productiv- ity between the treatment and the comparison regions (the smallest p-value=0.17). The second panel reports the means of a number of pre-bridge characteristics of treatment and comparison areas, and the two regions appear quite similar in terms of total population and its density, and rainfall and its variability. In terms of the level of NDVI, the comparison areas are on average greener, and the di¤erence between the comparison and treatment areas are statistically signi…cant in the driest months during the …rst quarter of the year (p-value=0.03 for the unweighted means di¤erence). In the case of annual change in NDVI, the di¤erence in means is numerically small and is statistically signi…cant only in the …rst 23 quarter of the year. There are some statistically signi…cant di¤erences in the cropping pattern: both the land under high yielding variety (HYV) of rice and vegetables are larger in treatment areas, whereas that under pulses is higher in the comparison areas. However, there is no signi…cant di¤erence in the proportion of land under chemical fertilizer and of household owning irrigation equipment. When considered along with the evidence of no signi…cant di¤erence in land productivity discussed above, this evidence on productivity enhancing inputs suggests strongly that the treatment and comparison areas were similar in the pre-bridge period in terms of agricultural potential and technological development. For most variables, the di¤erences in the weighted averages are smaller than in the unweighted averages, with the exception of some of NDVI variables. (6.2) Doubly Robust Approach: Evidence from Placebo Tests during the Pre-treatment Period The evidence from Table 1 shows that the treatment and comparison areas are bal- anced in terms of some variables, while they di¤er signi…cantly for other variables such as land allocated to high-yielding variety of rice and vegetables. For some of the variables, these di¤erences are not smoothed out by weighting (logit or Oaxaca-Blinder). The recent literature suggests that a doubly robust approach that combines weighting with regression adjustments is likely to be better at achieving pre-treatment balance and providing credible estimates of treatment e¤ects. To see if our treatment and comparison subdistricts are well balanced in terms of pre-bridge characteristics when we use the doubly robust approach, we estimate the e¤ects of a placebo bridge on our dependent variables using the pre-bridge data. We estimate the e¤ects of the placebo treatment on changes in vegetation indices for dry seasons and annual average during pre-bridge period. These false experiments test whether the outcome variables are statistically di¤erent between treatment and compari- son areas once we implement both weighting and regression adjustments. Because tests are done with data prior to the opening of the bridge, these falsi…cation tests should be able to indicate if the doubly robust approach is successful in dealing with any selection bias between the treatment and comparison subdistricts. 24 Table 2 reports the results from these doubly robust placebo regressions. Columns 1 and 2 in Table 2 report the di¤erences between the treatment and comparison subdistricts and p-values when logit weighting is buttressed with direct regression adjustments using the same set of pre-bridge characteristics, and columns 3 and 4 report the results for the Oaxaca-Blinder (OB) regressions. The vector of controls include the log of population den- sity in 1991, suitability of land for crop production, log of average and standard deviation of rainfall in 1991, and whether an upazila is top-ranked for rice and vegetables produc- tion. In contrast to the evidence in Table 1, the estimates in Table 2 indicate the absence of statistically signi…cant di¤erences between treatment and comparison regions for all of the variables. Overall, we …nd no signi…cant di¤erences in the levels of outcome variables between the treatment and comparison areas during the pre-bridge periods. For the out- comes such as vegetation indices for which we have multiple years of observations before the opening of the bridge, we …nd no signi…cant di¤erences in trends either. We interpret this evidence as supportive of the research design based on …xed e¤ect DID and doubly robust estimators. 7 Evidence on the E¤ects of Jamuna Bridge on Agri- cultural Development (7.1) The Average E¤ects of Jamuna Bridge on Technology Adoption, Land Use Intensity and Cropping Pattern The estimated e¤ects of Jamuna bridge on treatment areas in NW compared with the comparison areas in south are reported in Table 3. The FE-DID-OLS estimates conditioned on a small set of pre-bridge characteristics described above are reported in column 1. Columns 2 and 3 report the estimates from logit and OB weighted regressions, using the same set of controls for direct regression adjustments, respectively. A comparison of the estimates across columns indicate some di¤erences among the three sets of estimates, but those di¤erences are numerically small. For most of the regressions, the magnitudes of the 25 estimates are smaller in OB weighted regressions with a few exceptions (e.g. the share of rice land). The weighted estimates have smaller standard errors as well. For the discussion below, we focus on the OB weighted estimates. The upper panel in Table 3 reports the estimates for technology adoption using six indicators. The estimates suggest positive and statistically signi…cant impacts of Jamuna bridge on all six indicators of technology adoption. While cropping intensity and fertilizer use increased in both the treatment and comparison areas during the post-bridge period, the estimates imply an additional 3 percent increase in the cropping intensity, and a 7 percent increase in the share of land using chemical fertilizer in the treatment upazilas compared with the comparison upazilas. The implied additional increase of ownership of irrigation equipment is much larger (0.157) which compares favorably with its level in the pre-treatment period (0.11). The impressive increase in fertilizer use and irrigation adoption is re‡ected in the changes in vegetation index NDVI. The estimates suggest that the treatment areas have become greener compared with the comparison areas after the opening of the bridge. The increase in NDVI is much larger during the dry seasons (…rst and fourth quarters of the year) relative to the average for the year consistent with a substantial increase in irrigation. The results for cropping pattern are more complex. There is a signi…cant decline in the share of HYV rice as well as in the share of total rice, but an increase in the share of land allocated to pulse. The increase in the share of vegetable while statistically signi…cant is modest numerically. The decrease in the share of land devoted to the low-value and low transport-intensive crop rice is consistent with the canonical von Thunen model where a reduction in the transport cost increases the share of land going into transport intensive high-value vegetable crops. However, a considerable decline in the share of HYV rice on the other hand appears puzzling in the light of robust positive response found in technology adoption after the opening of the bridge. The modi…ed von Thunen model presented in section 2 shows that land productivity heterogeneity can lead to such an outcome (Figure 3) when upward shifts in the bid rent curve due to a reduction in costs of transportation 26 for vegetables are larger than that for rice. (7.2) Heterogeneous Land Productivity and the E¤ects of the Bridge A central focus of this study is to understand whether a large reduction in trade costs lead to a better matching of land productivity and crop choices according to comparative advantage. To see if the reduction in transportation cost due to bridge opening helped cropping pattern to align more closely with the natural land productivity, as predicted by the modi…ed von Thunen model in section 2 above, we explore the heterogeneity in the e¤ects of Jamuna bridge with respect to land productivity. The theoretical analysis, however, suggests that the inherent land productivity matters much less if an area is too close to the markets in the central region or too far away. As a …rst step to examining this heterogeneity, we de…ne relative land productivity for transport intensive high-value crop vegetables. Using the ranking of land in terms of its suitability for production of di¤erent crops developed by the agronomists, we de…ne the relative productivity as the ratio of the rankings of rice and vegetables. Recall that land productivity for a crop is ranked in the scale of 1 to 5, with 1 being the best. The relative productivity variable as de…ned (rank of HYV rice/rank of vegetables) indicates how good the land in a subdistrict is for vegetable production relative to the high-yielding variety of rice production. Figures 4a and 4b plot the non-parametric graphs of this relative productivity indicator and of the actual share of land allocated to vegetables in the NW during the pre-treatment year (1998) against the distance from bridge location respectively. The average vegetable productivity relative to HYV rice is low at the riverbank and remains nearly ‡at for the distance up to 100 km from the bridge location, and it rises with distance, reaching its peak at around 200 km from the bridge. In contrast, the share of land devoted to vegetables in 1998 increases with distance up to 100 km from the bridge location and starts falling after 110 km. That the peak of vegetables land share is reached half way to its peak of land productivity is indicative of very high transport costs during the pre-bridge period. According to the modi…ed von Thunen model, the large reduction in transport cost due to the opening of the bridge should help expand the share of land to vegetables in the 27 areas farther from the bridge (beyond 110 km ), and particularly in those areas with higher V vegetable productivity. To test this formally, we de…ne a dummy Di that takes the value of unity if relative vegetable productivity of a subdistrict is greater than unity and zero otherwise. This dummy represents the subdistricts which have better productivity ranking for vegetables compared with HYV rice. We then de…ne a set of distance dummies using DF ar ’because it takes the value of unity di¤erent distance cut-o¤s. The dummy is labeled ‘ if a subdistrict is located farther than the cut-o¤. For instance, for a distance cut-o¤ of 110 F ar km, the dummy is called ‘ D110 ’is unity if a subdistrict is farther than 110 km away from the bridge location and zero otherwise. The average distance from bridge in our sample is 110 km. Each panel in Table 4 reports the results for a distance cut-o¤. The …rst row in each panel reports coe¢ cient on the treatment-year interaction, (T Y r) in equation (3) above, and the second row the coe¢ cient on the quadruple interaction term treatment- year-productivity-distance T Y r DV DF ar . The focus is on the coe¢ cient on the quadruple interaction term which shows how cropping pattern and technology adoption in areas that have relatively higher vegetable productivity and are not close to the bridge location responded to bridge opening in a …xed e¤ect DID model. On the other hand, the coe¢ cient of treatment-year (T Y r) variable indicates the response in areas which are within the distance cut-o¤ from the bridge and have comparative advantage in HYV rice production. In the appendix Table A.1, we also report the coe¢ cients on triple interaction dummies which are insigni…cant either numerically or statistically or both anr are omitted for the sake of brevity. (7.2.A) Spatial Heterogeneity in Technology Adoption and Cropping Inten- sity The results display some interesting patterns for technology adoption and cropping in- tensity (please see …rst three columns on Table 4). For cropping intensity, the estimated coe¢ cient of treatment-year dummy (T Y r) is 0.27 for areas within 110 km of the bridge, and it increases to 0.53 when the distance cut-o¤ is extended to 130 km, and declines to 28 0.43 and 0.35 when the distance cut-o¤s are extended to 150 km and 170 km respectively. The results suggest non-linear e¤ects of bridge on cropping intensity in the Jamuna treat- ment region in the NW. Relative to the comparison areas in the South, cropping intensity has increased everywhere in NW, but increased the most in the interval of 130 km-150 km. This non-linear pattern is observed for the coe¢ cients on the quadruple interaction term T Y r DV DF ar as well though none of the coe¢ cients are estimated with sta- tistical precision. The estimated coe¢ cients on T Y r DV DF ar are twice as large in magnitude as the coe¢ cients on treatment-year dummy (T Y r). For fertilizer and irrigation, the estimated coe¢ cients of treatment-year (T Y r) dummy become smaller in magnitude with an increase in the distance cut-o¤ for the far dummy. This implies somewhat larger impacts near the bridge than farther away. For irrigation, the statistical precision of the estimates also su¤ers with an increase in the distance cut-o¤. None of the coe¢ cients on the quadruple interaction term T Y r DV DF ar is esti- mated with precision except for fertilizer in the areas father than 170 km from the bridge where it has a positive coe¢ cient. The coe¢ cients of treatment-year dummy (T Y r) in the NDVI regressions are indistinguishable across di¤erent distance cut-o¤s. The quadru- ple interaction term T Y r DV DF ar has statistically signi…cant coe¢ cients for the distance cut-o¤ of 170 km but the coe¢ cient is rather small in magnitude. The results show that impact on fertilizer use is higher near the bridge whereas on cropping intensity in intermediate distance of 130-150 km. Even with lower increase in fertilizer use in these areas, total land under modern technology may have gone up due to higher cropping inten- sity. It is reassuring to note that the impacts on NDVI which subsumes intensity of both land and input use did not vary with respect to distance from the bridge. (7.2.B) Spatial Heterogeneity in Land Reallocation across Crops The results shown in Table 4 suggest a signi…cant reduction in the share of land allocated to HYV rice, and an increase in that to pulses, with no signi…cant change in either vegetables or total rice in the areas that are within 110 km of the bridge. However, the areas that enjoy relatively higher vegetable productivity, but are located farther than 110 km saw a 29 signi…cant increase in the share of land to vegetables: the estimate implies a 25 percent increase over its pre-bridge level. The evidence suggests that the increase in share of land used for vegetables remained limited up to 150 km of the bridge. Overall, the non-linear pattern of e¤ects on the share of land to vegetables is consistent with that of cropping intensity described above. In contrast to vegetables, the results suggest diversi…cation of cropland away from rice for all distance cuto¤s and into pulses. But the areas good for vegetable production and farther away from 150 km, the share of land to rice increases, and the increase is much larger for HYV rice. The increase in the share of rice in the areas farther than 170 km is associated with a small decline in the share of pulse. The results indicate a nonlinear pattern in reallocation of cropland in areas with rela- tively higher vegetables productivity: land moved away from rice, particularly HYV rice, to vegetables in the intermediate distance (110 km-150 km), and it reverses in the areas farther than 150km, land moves into rice, particularly HYV rice. This reallocation is associated with productivity gain even if we ignore the technology adoption, since it allows vegetables to be grown in land better suited for vegetables production in the intermediate distance from the bridge. The areas in the intermediate distance also have much higher cropping intensity. All areas regardless of distance experienced increased technology adoption in terms of cropping intensity, fertilizer use, irrigation ownership and greenness particularly in dry seasons. The evidence also indicates that the pattern in green-ness aligns well with use of fertilizer and irrigation and cropping intensity, yet it is unable to detect change in cropping pattern. However, it is technology adoption and increase in cropping intensity made feasible by the bridge that in the end allows actual cropping pattern to align more closely to natural land productivity. (7.3) Discussion The empirical results discussed in section (7) above provide evidence of positive e¤ects of the Jamuna bridge on technology adoption, agricultural intensi…cation and the share of land allocated to higher value crops (pulses and vegetables). The evidence suggests that large reduction in trade costs following the opening of the Jamuna bridge led to agricultural 30 development in the newly connected Jamuna hinterland through both technology adoption, and better matching of land to crops according to comparative advantage. It also con…rms spatial heterogeneity in the e¤ects of the bridge on cropping pattern as predicted by the extended von Thunen model in section 2 above. However, there are a few potential issues regarding the empirical estimates that may s mind. First, one might wonder whether the empirical estimates of the come to a reader’ treatment e¤ects of the bridge and the substantive conclusions are likely to be signi…cantly a¤ected by inter-regional labor mobility. The issue of spatial reallocation (reorganization e¤ect in the terminology of Redding and Turner (2014) e¤ect is of …rst order importance when population density, labor allocation, and wages (and income) are the focus of an analysis, as is the case in many recent studies. Our focus is on allocation of an immobile resource, land among di¤erent crops where the e¤ects of labor mobility is not likely to be of …rst order consequence (see the discussion in Donaldson and Hornbeck (2016)). A related concern is the price e¤ects of the bridge. The theoretical model assumes prices to be determined in the center, and are not subject to change in response to bridge. While the small country assumption is a plausible one in the context of agricultural products such as rice and pulses in Bangladesh, vegetables prices may be more responsive to local supply conditions. Since both treatment and comparison regions trade with the center, a reduction in vegetables price in the center due to an increase in the supply from the treatment region would a¤ect farmers in both treatment and comparison regions, and are not likely to a¤ect the conclusions in the DID-FE estimation in a signi…cant way. An additional concern is that opening of markets may expose farmers to higher price volatility and encourage them to diversify (Allen and Arkokalis (2017)). Since the price of vegetables tend to be more volatile than that of rice or pulse (BBS (2014)), it can not explain the increase in the share of vegetables in cropland in response to the bridge. Neither spatial displacement nor price volatility can explain the heterogeneous e¤ects of the bridge discussed earlier within the treatment region. 31 8 Conclusions This paper utilizes a quasi-natural experiment to study the e¤ects of a large reduction in transport cost (more than 50 percent) due to the construction of a bridge on agricultural specialization and technology adoption, with a focus on spatial heterogeneity. We extend the classical von Thunen model of land allocation to incorporate costly technology adoption and land productivity heterogeneity. Technology adoption introduces non-linearity in crop land allocation with respect to the distance to the urban market. Land productivity hetero- geneity along with technology adoption produces deviation of observed cropping patterns from e¢ cient pattern based on comparative advantage due to land productivity. The areas closer to the bridge devote more land to transport intensive high-value crop (vegetables) even if the land productivity for vegetables is relatively lower, whereas in the areas farther away, transport costs outweight the land productivity advantage. The model predicts that the positive e¤ects of the bridge on the share of land devoted to a high-value transport intensive crop is felt most prominently in areas that are not too near or far from markets and also have higher relative land productivity in that crop. The empirical analysis is based on a subdistrict level panel data set and exploits a di¤erence-in-di¤erence framework motivated by idiosyncratic political factors; the compar- ison region comes from the hinterland of the proposed but yet to be built Padma bridge which remains cut-o¤ from the growth centers in the capital city Dhaka and port city Chit- tagong. The central …ndings are as follows. The Jamuna bridge contributed to agricultural development in the treatment areas in the poor Nortwest region through technology adop- tion, and better matching of crops according to land productivity, thus reducing the spatial mismatch between comparative advantage and the actual cropping pattern in an upazila. The results indicate non-linear spatial patterns in the e¤ects, consistent with the predic- tions from the extended von Thunen model. For cropping intensity, the largest e¤ects are observed in the areas in the intermediate distance (130 km-150 km) from the bridge. The reallocation of cropland in areas with relatively higher vegetables productivity show interesting spatial nonlinearity: land moved away from rice, particularly from HYV rice, to 32 vegetables in the intermediate distance (110-150km), and the pattern changes after 150km where more land is allocated to HYV rice in response to the bridge. References Allen and Atkin (2016), “Volatility and Gains from Trade,”working paper. Ali, R., Barra, A. F., Berg, C. N., Damania, R., Nash, J. (2016), "Agricultural Tech- nology Choice and Transport," American Journal of Agricultural Economics, 2016. Atack, J. and R. Margo (2011) “The Impact of Access to Rail Transportation on Agri- cultural Improvement: The American Midwest as a Test Case, 1850-1860,” Journal of Transport and Land Use, 4(2), 5-18. Basu, K (1997). Analytical Development Economics, MIT Press. Bardhan (1984), Land, labor and Rural Poverty, Columbia University Press. Berg, Claudia N, Uwe Deichmann, Yishen Liu and Harris Selod (2016): “Transport Policies and Development”, The Journal of Development Studies. Braverman, A, K, Ho¤ and J. Stiglitz (1993), The Economics of Rural Organization, Oxford University Press. Busso, M, Jesse Gregory, and Patrick Kline (2013), “Assessing the Incidence and Ef- …ciency of a Prominent Place Based Policy," American Economic Review, vol. 103(2), 897-947. Chandra, A. and E. Thompson (2000) “Does Public Infrastructure a¤ect Economic Activity? Evidence from the Rural Interstate Highway System” Regional Science and Urban Economics, 30, 457-90. Costinot, A., and D. Donaldson, (2016). How Large Are the Gains from Economic Integration? Theory and Evidence from U.S. Agriculture, 1880-1997", working paper Costinot, A., D. Donaldson and C. Smith, (2016). Evolving Comparative Advantage and the Impact of Climate Change in Agricultural Markets: Evidence from 1.7 Million Fields Around the World” Journal of Political Economy, 2016, vol. 124, issue 1, pp. 205- 248 Deichmann, U., F. Shilpi, and R. Vakis. 2009. “Urban Proximity, Agricultural Potential 33 and Rural Non-farm Employment: Evidence from Bangladesh,”World Development, 37(3), p. 645-660. Donaldson D. (2018). "Railroads of the Raj: estimating the impact of transportation infrastructure." American Economic Review. 108(4-5), 899-934. Donaldson, D. and R. Hornbeck (2016), "Railroads and American Economic Growth: Market Access’Approach," Quarterly Journal of Economics, 131(2):799-858 A‘ Donaldson, D, (2015) “The Gains from Market Integration”, Annual Review of Eco- nomics, 7:619-647 Dorosh, P., H.G. Wang, L. You, and E. Schmidt (2012) “Road connectivity, population, and crop production in Sub-Saharan Africa,”Agricultural Economics, 43(1), 89-103. Duranton, G, and A, Venables (2018), Place-Based Policies for Development, World Bank Policy Research Working Paper 8410. Emran, M. S., and Hou, Z. (2013), “Access to markets and rural poverty: Evidence from household consumption in China”. Review of Economics and Statistics, 95(2), 682–697. Emran, M. S., and Forhad Shilpi (2012), “The extent of the market and stages of agricultural specialization," Canadian Journal of Economics, vol. 45(3), pages 1125-1153. Emran, M. Shahe and Shilpi. Forhad, (2002). “Marketing externalities and market development,”Policy Research Working Paper Series 2839, The World Bank. Fafchamps, M, and Forhad Shilpi, (2003). “The spatial division of labour in Nepal," Journal of Development Studies, 39(6), pages. 23–66. Haines, M.R. and Margo, R.A., 2006. Railroads and local economic development: The United States in the 1850s (No. w12381). National Bureau of Economic Research. Jacoby, H.G. (2000) “Access to markets and the bene…ts of rural roads,”The Economic Journal, 110(465), 713-737. Kline, P, and Enrico Moretti (2014), “Local Economic Development, Agglomeration Economies, and the Big Push: 100 Years of Evidence from the Tennessee Valley Authority," The Quarterly Journal of Economics, vol. 129(1), pages 275-331. Kline, P, (2011), “Oaxaca-Blinder as a Reweighting Estimator," American Economic 34 Review, American Economic Association, vol. 101(3), pages 532-537, May. Kyeyamwa, H., S. Speelman, G.V. Huylenbroeck, J. Opuda-Asibo, W. Verbeke (2008) “Rising o¤take from cattle grazed on natural rangelands in Sub-Saharan Africa: a trans- action cost economic approach,”Agricultural Economics, 39, 63-72. Omamo, S. W. (1998). "Transport costs and smallholder cropping choices: An appli- cation to Siaya District, Kenya." American Journal of Agricultural Economics, 80(1). Redding SJ, and Turner MA. (2014). Transportation costs and the spatial organization of economic activity. NBER Work. Pap. 20235. Robins, J.M., Rotnitzky, A., Zhao, L.P. (1994), Estimation of Regression-Coe¢ cients When Some Regressors are Not Always Observed, Journal of the American Statistical Association, 89, 846–866. Shamdasani (2016), "Rural Road Infrastructure and Agricultural Production: Evidence from India," working paper Shrestha, P (2016), "Access to the North-South Roads and Farm Pro…ts in Rural Nepal," forthcoming in Economic Development and Cultural Change. Wooldridge, Je¤rey M., (2007). “Inverse probability weighted estimation for general missing data problems," Journal of Econometrics, Elsevier, vol. 141(2), pages 1281-1301, December. World Bank (1994), Sta¤ Appraisal Report: Bangladesh: Jamuna Bridge Project. South Asia region 35 Figure 1: Geography of the country with two rivers and three regions Figure 2a: Technology Adoption and cropping pattern with homogenous land productivity Note: Blue depicts bid rent curve for X and green for Y. Lighter shade for crop produced under traditional technology and darker for crop under modern technology. Arrowed lines of respective color show land under different crops. Figure 2b: Technology Adoption and cropping pattern with heterogenous land productivity Note: Blue depicts bid rent curve for X and green for Y. Lighter shade for crop produced under traditional technology and darker for crop under modern technology. Arrowed lines of respective color show land under different crops. Land productivity of X increases with distance from bridge location, while it is constant for Y. Figure 3: Cropping pattern before and after bridge in village V2 Note: Brown depicts bid rent curve for X and green for Y. Lighter shade for crop produced under traditional technology and darker for crop under modern technology. Dashed lines depict before bridge and solid after bridge bid-rents and land allocation. Arrowed lines show land under different crops. Figure 4a: Land productivity for Vegetables relative to HYV rice and Distance from Bridge Location Vegetables' Land Productivity Relative to Rice 2.5 2 1.5 1 .5 0 0 100 200 300 Distance from Bridge Location bandwidth = .8 Figure 4b: Vegetables’ share of cropland in per-bridge period (1996) and Distance from Bridge Location Vegetables' Land Share .2 .15 .1 .05 0 0 100 200 300 Distance from Bridge Location bandwidth = .8 Table 1: Pre-Bridge Sample Means in Treatment and Comparison Areas North-West South (Padma Bridge) P-value of Null Hypothesis of no (Jamuna difference between North-West and Bridge) South Un- Logit OB Un- Logit OB weighted weighted weighted weighted Weighted weighted Average Rank in terms of Suitability of Land for all crops 2.91 3.31 2.86 2.94 0.31 0.87 0.95 Suitability of Land for Rice 2.07 2.25 2.09 2.09 0.28 0.87 0.88 Proportion of Upazilas top ranked for rice 0.86 0.88 0.95 0.95 0.82 0.21 0.17 Suitability of Land for Vegetables 2.65 3.19 2.63 2.75 0.35 0.96 0.85 Proportion of Upazilas top ranked for vegetables 0.92 0.67 0.86 0.82 0.23 0.62 0.48 Population in 1991 210041 203068 214034 212908 0.27 0.61 0.70 Population Density in 1991 767 774 754 757 0.82 0.54 0.60 Average Rainfall 52 53 48 50 0.86 0.68 0.80 Standard Deviation of Rainfall 56 48 43 44 0.47 0.26 0.32 Cropping Pattern Share of HYV rice in total cropped land 0.464 0.189 0.259 0.244 0.01 0.03 0.02 Share of rice in total cropped land 0.686 0.675 0.626 0.643 0.85 0.17 0.35 Share of vegetables in total cropped land 0.040 0.022 0.024 0.024 0.00 0.00 0.00 Share of pulses in total cropped land 0.032 0.105 0.099 0.105 0.04 0.04 0.04 Agricultural Technology Cropping Intensity 1.778 1.740 1.748 1.739 0.61 0.63 0.56 Share of land under chemical fertilizer 0.536 0.419 0.466 0.419 0.04 0.13 0.11 Prop. of households with Shallow tube-well 0.111 0.051 0.082 0.074 0.072 0.273 0.198 Change in Normalized Vegetation Index(NDVI) Annual Average -0.013 0.000 0.003 0.003 0.026 0.011 0.009 Average in First quarter (January- March) -0.014 0.017 0.008 0.011 0.032 0.049 0.047 Average in Fourth Quarter(October-December) -0.059 -0.022 -0.013 -0.013 0.031 0.010 0.008 Note: The unit of observation is sub-district (upazila). Data on NDVI from satellite data and crop suitability from Bangladesh Agricultural Research Council, and everything else from agricultural and population censuses. Logit weights are inverse probability weights based on logit regression of treatment status on pre-bridge characteristics. Oaxaca-Blinder (OB) weights are estimated using a procedure suggested by Kline (2011). Both logit and OB regressions used the same set of pre-bridge controls. Table 2: Treatment and Comparison Areas during Pre-bridge period: Effects of a Placebo Bridge DID-FE with Regression Adjustments Logit Weight OB weight Coefficient P-value Coefficient P-value N Agricultural Technology adoption (1998) Cropping Intensity -0.038 0.615 -0.065 0.502 229 Share of land under chemical fertilizer 0.020 0.330 0.014 0.519 229 Prop. of households with Shallow tube- wells 0.013 0.395 0.015 0.309 229 Difference in NDVI (1993-1998) Annual Average -0.003 0.606 -0.004 0.485 401 Average in First quarter (January-March) -0.007 0.374 -0.011 0.155 365 Average in Fourth Quarter(October- December) -0.020 0.167 -0.022 0.112 397 Agricultural Cropping pattern (1998) Share of HYV rice in total cropped land 0.022 0.921 0.035 0.875 229 Share of rice in total cropped land -0.045 0.820 -0.037 0.856 229 Share of vegetables in total cropped land -0.002 0.892 -0.001 0.944 229 Share of pulses in total cropped land 0.003 0.923 0.001 0.981 229 Note: The results for each outcome are reported in a row. The odd numbered column provides the difference-in-difference estimate of coefficient of treatment dummy and adjacent even numbered column its robust standard errors. Column 1 provides the simple OLS results for the full sample, columns 3 and 5 inverse probability weighted and Oaxaca-Blinder weighted estimates. Controls in each regression includes log (population in 1991), log (crow-fly distance to bridge location), log (average rainfall in 1998), log (standard deviation of rainfall in 1998), suitability of land for crop production, and dummies indicating top ranking of land for its suitability for rice and vegetables production. For NDVI regressions, rainfall variables are for 1995-1998. Standard errors are clustered at regional (division) level. Legend: *** p<0.01, ** p<0.05, * p<0.1 Table 3: Jamuna Bridge and technology adoption and cropping Pattern in agriculture: DID-FE with regression adjustments DID-FE with Regression Adjustments Logit OB Un-weighted Weighted Weighted N Agricultural Technology adoption Cropping Intensity 0.059** 0.045 0.047 211 (0.020) (0.023) (0.023) Share of land under chemical fertilizer 0.038* 0.037*** 0.036*** 211 (0.017) (0.006) (0.006) Prop. of households with Shallow tube-wells 0.184 0.143* 0.155* 202 (0.090) (0.058) (0.060) Normalized Difference Vegetation Index (NDVI): Difference Annual Average 0.019*** 0.019*** 0.019*** 2961 (0.002) (0.003) (0.003) Average in First quarter (January-March) 0.038** 0.026* 0.030* 2904 (0.013) (0.011) (0.013) Average in Fourth Quarter(October- December) 0.042** 0.047** 0.048** 2975 (0.014) (0.011) (0.011) Agricultural Cropping pattern Share of HYV rice in total cropped land -0.107** -0.114*** -0.113*** 212 (0.024) (0.015) (0.016) Share of rice in total cropped land -0.021 -0.033** -0.030** 208 (0.021) (0.008) (0.008) Share of vegetables in total cropped land 0.003 0.004** 0.004*** 213 (0.002) (0.001) (0.001) Share of pulses in total cropped land 0.029*** 0.029*** 0.028*** 197 (0.004) (0.003) (0.003) Note: Each labeled row reports results for the labeled dependent variable and its respective standard errors are in parenthesis in the next row (un-labeled). Column 1 provides the simple DID-FE results, columns 2 and 3 inverse probability weighted and Oxaca- Blinder weighted estimates respectively. Controls in each regression includes log (population in 1991), log (crow-fly distance to bridge location), log (average rainfall in 1998), log (standard deviation of rainfall in 1998), suitability of land for crop production, and dummies indicating top ranking of land for its suitability for rice and vegetables production. For NDVI regressions, rainfall variables are for 1995-1998. Standard errors are clustered at regional (division) level. Legend: *** p<0.01, ** p<0.05, * p<0.1. ` Table 4: Heterogeneity of impacts with respect to distance from the bridge: Results from OB weighted DID-FE with regression adjustments Cropping Fertilizer Shallow Normalized Vegetation Index Share of land under 1st Intensity use (prop. Tubewell Average Quarter 4th Quarter HYV Rice All rice Vegetables Pulses of land) Ownership Far:>110km Panel A Treatment 0.027* 0.058*** 0.173* 0.014* 0.029* 0.037*** -0.101*** -0.019 0.000 0.030*** (0.011) (0.010) (0.067) (0.006) (0.011) (0.005) (0.017) (0.010) (0.003) (0.002) Treat*RVeg*Far110 0.087 -0.000 0.171 -0.001 -0.005 -0.008 0.010 -0.019 0.011*** 0.002 (0.091) (0.024) (0.098) (0.003) (0.003) (0.006) (0.020) (0.016) (0.002) (0.004) Far>130km Panel B Treatment 0.054** 0.046*** 0.128 0.015** 0.030* 0.040*** -0.117*** -0.031** 0.004** 0.029*** (0.019) (0.004) (0.064) (0.005) (0.012) (0.007) (0.022) (0.010) (0.001) (0.004) Treat*RVeg*Far130 0.123 0.003 0.238 -0.000 -0.004 -0.003 0.019 -0.007 0.013** 0.005 (0.087) (0.019) (0.167) (0.004) (0.005) (0.005) (0.037) (0.012) (0.003) (0.006) Far>150km Panel C Treatment 0.043** 0.048*** 0.109 0.016** 0.030* 0.041*** -0.112*** -0.025* 0.003** 0.027*** (0.009) (0.005) (0.065) (0.004) (0.012) (0.007) (0.016) (0.010) (0.001) (0.003) Treat*RVeg*Far150 0.114* 0.004 0.129 -0.001 -0.000 -0.007 0.051* 0.023*** 0.009 -0.003 (0.041) (0.007) (0.132) (0.003) (0.003) (0.003) (0.018) (0.003) (0.004) (0.002) Far>170km Panel D Treatment 0.035* 0.044*** 0.099 0.016** 0.031* 0.042*** -0.111*** -0.023** 0.002 0.026*** (0.016) (0.007) (0.072) (0.004) (0.013) (0.007) (0.014) (0.007) (0.001) (0.003) - Treat*RVeg*Far170 0.054 0.008** 0.041 0.003** 0.002** -0.007** 0.061*** 0.018** 0.004 -0.005** (0.038) (0.003) (0.050) (0.001) (0.001) (0.002) (0.004) (0.006) (0.002) (0.001) Note: “FarK”: is a dummy that takes the value of unity if a subdistrict is located farther than the distance cut-off K (e.g. K=110km in panel A) and zero otherwise. Rveg: is a dummy that takes the value of unity if suitability of vegetables production is greater than that for High Yielding Variety (HYV) rice and zero otherwise. Treat is unity if subdistrict is located in North-West region that is treatment region of Jamuna bridge. Each labeled column reports results for the labeled dependent variable and its respective standard errors are below in parenthesis. Each regression uses the same set of controls are reported in Table 3. Standard errors are clustered at regional (division) level. Legend: *** p<0.01, ** p<0.05, * p<0.1. Appendix A: Theoretical Model (proofs) A1. Proof of propostion 1(iii): We have 4 bid-rent functions and can have 12 di¤erent outcomes where one bid rent curve intersects another. For proposition 1(iii), 10 di¤erent unique outcomes: For transport cost of X less than a threshold ( x < ^x such that hM M X (^ x ) = hY ); crop X is produced in h 2 HX = [0; hj X ] either using modern technology (j = M ) or a combination modern and traditional technologies (j = T ) whereas crop Y is produced using traditional technology in subregion farther away from hj j j j T j X . hX is determined by equating RX (hX ) = RY (hX ); j = M; T: For transport cost of X above a threshold ( x > ^x ); two broader cases each with 3 alternative outcomes are possible: T M M Case (a): RX is ‡atter than RY : ( x rx0 < z PZ + y y ry 0 ) : Suppose hi is such that Ri (h = T hi ) = R i (h = hi ): Three outcomes can be identi…ed if hX < hY : M M (1) If RY (h = hX ) < RX M (h = hX ); then crop X M is produced in h 2 HXa MT 1 = [0; hXX ] using T MT TT modern technology and in h 2 HXa1 = (hXX ; hXY ] using traditional technology, and crop Y in h 2 T TT E HY a1 = (hXY ; H ] using traditional technolgy M T (2) If RY (h = hY ) > RX M (h = hY ), crop X M is produced in h 2 HXa MM 2 = [0; hXY ] , and crop Y M MM MT T MT E in h 2 HY a2 = (hXY ; hY Y ] using modern technology and in h 2 HY a2 = (hY Y ; H ] using traditional technolgy; M M M T (3) If RY (h = hX ) > RX (h = hX ) and RY (h = hY ) < RX (h = hY ); then crop X M is produced by M MM T MT TT farmers in h 2 HXa3 = [0; hXY ] using modern technology and in h 2 HXa3 = (hY X ; hXY ] using traditional M MM MT T TT E technology, and crop Y in h 2 HY a3 = (hXY ; hY X ] using modern technology and h 2 HY a3 = (hXY ; H ] using traditonal technology. M M Two more outcomes if hX > hY ; so RY (h = hY ) < RX (h = hY ) M (4) Then either crop X M is produced in h 2 HXa MT 1 = [0; hXX ] using modern technology and in h 2 T MT TT T TT E HXa1 = (hXX ; hXY ] using traditional technology, and crop Y in h 2 HY a1 = (hXY ; H ] using traditional technolgy M (5) or crop X M is produced in h 2 HXa MT T 1 = [0; hXY ] using modern technology and rop Y in h 2 HY a1 = (hM T E XY ; H ] using traditional technolgy T M Case (b): RX is steeper than RY :( x rx0 z PZ + y y ry 0 ): Three cases as well for hX < hY : M M (1) Produce X in h 2 HXb1 = [0; hM M XY ];and Y M M in h 2 HY MM MT b1 = (hXY ; hY Y ]and Y T in h 2 T MT E HY b1 = (hY Y ; H ]; M (2) Produce X M in h 2 HXb MT 2 = [0; hXX ];and X T T in h 2 HXb MT TT 2 = (hXX ; hXY ]and Y T in h 2 T TT E HY b2 = (hXY ; H ]; M (3) Produce X M in h 2 HXb MT T T MT TM 3 = [0; hXX ]; X in h 2 HXb3 = (hXX ; hXY ] and Y M M in h 2 HY b3 = 1 (hT M MT XY ; hY Y ]and Y T T in h 2 HY MT E b3 = (hY Y ; H ] M M Two more outcomes if hX > hY ; so RY (h = hY ) < RX (h = hY ) M (4) Then either crop X M is produced in h 2 HXa MT 1 = [0; hXX ] using modern technology and in h 2 T MT TT T TT E HXa1 = (hXX ; hXY ] using traditional technology, and crop Y in h 2 HY a1 = (hXY ; H ] using traditional technolgy M (5) or crop X M is produced in h 2 HXa MT T 1 = [0; hXY ] using modern technology and rop Y in h 2 HY a1 = (hM T E XY ; H ] using traditional technolgy A2. Land productivity heterogeneity and curvature of bid rent function Consider bid rent function for X M : M RXh = X rX 0 (1 + h) (1 x Fh x h) PZ (1 + z Fh + z h) M @RXh = X rX 0 [(1 x Fh ) x (1 + 2 h)] z PZ @h M @ 2 RXh = 2 x X rX 0 < 0 if >0 @h2 2 Appendix B (Online): Normalized Difference Vegetation Index (NDVI) Live green plants absorb solar radiation in the photosynthetically active radiation (PAR) spectral region, which they use as a source of energy in the process of photosynthesis. Leaf cells have also evolved to re-emit solar radiation in the near-infrared spectral region (which carries approximately half of the total incoming solar energy), because the photon energy at wavelengths longer than about 700 nanometers is not large enough to synthesize organic molecules. Live green plants appear relatively dark in the PAR and relatively bright in the near-infrared. By contrast, clouds and snow tend to be rather bright in the red (as well as other visible wavelengths) and quite dark in the near-infrared. Using satellite data on strong plant reflectance, the normalized difference vegetation index (NDVI) is defined as: NDVI=((NIR-red)/(NIR+red)) where red and NIR stand for the spectral reflectance measurements acquired in the red (visible) and near-infrared regions, respectively. These spectral reflectances are themselves ratios of the reflected over the incoming radiation in each spectral band individually, hence they take on values between 0.0 and 1.0. By design, the NDVI varies between -1.0 and +1.0. The NDVI data are available for a sufficiently long period of time (bi-weekly data from mid 1980s to 2014 but not for every year before 1996). Figure A.1: Location of Jamuna and Proposed Padma Bridges and Treatment and Comparison areas Table A.1: Heterogeneity of impacts with respect to distance from the bridge: Results from OB weighted DID-FE with regression adjustments Cropping Fertilizer Shallow Normalized Vegetation Index Share of land under Intensity use (prop. Tubewell Average 1st Quarter 4th Quarter HYV Rice All rice Vegetables Pulses of land) Ownership Far:>110km Panel A Treatment 0.027* 0.058*** 0.173* 0.014* 0.029* 0.037*** -0.101*** -0.019 0.000 0.030*** (0.011) (0.010) (0.067) (0.006) (0.011) (0.005) (0.017) (0.010) (0.003) (0.002) Treat*RVeg*Far110 0.087 -0.000 0.171 -0.001 -0.005 -0.008 0.010 -0.019 0.011*** 0.002 (0.091) (0.024) (0.098) (0.003) (0.003) (0.006) (0.020) (0.016) (0.002) (0.004) Treat*Far110 0.038 -0.016 -0.121 0.009** 0.003 0.016** -0.031 -0.014 0.004 -0.008 (0.037) (0.030) (0.093) (0.003) (0.005) (0.005) (0.028) (0.014) (0.008) (0.005) Treat*Rveg -0.039 -0.032 0.015 0.003 0.001 0.012 -0.003 0.008 -0.004 0.003 (0.101) (0.023) (0.030) (0.002) (0.001) (0.006) (0.009) (0.019) (0.003) (0.002) Far>130km Panel B Treatment 0.054** 0.046*** 0.128 0.015** 0.030* 0.040*** -0.117*** -0.031** 0.004** 0.029*** (0.019) (0.004) (0.064) (0.005) (0.012) (0.007) (0.022) (0.010) (0.001) (0.004) Treat*RVeg*Far130 0.123 0.003 0.238 -0.000 -0.004 -0.003 0.019 -0.007 0.013** 0.005 (0.087) (0.019) (0.167) (0.004) (0.005) (0.005) (0.037) (0.012) (0.003) (0.006) Treat*Far130 -0.041 -0.012 -0.042 0.009* 0.002 0.014** 0.006 0.012 -0.003 -0.012 (0.063) (0.045) (0.055) (0.003) (0.007) (0.004) (0.022) (0.016) (0.008) (0.010) Treat*Rveg -0.054 -0.027 -0.004 0.002 -0.000 0.008* -0.002 -0.001 -0.004 0.001 (0.100) (0.014) (0.034) (0.001) (0.001) (0.003) (0.019) (0.014) (0.003) (0.002) Continued next page. Table A.1: Heterogeneity of impacts with respect to distance from the bridge: Results from OB weighted DID-FE with regression Adjustments (continued from earlier page) Cropping Fertilizer Shallow Normalized Vegetation Index Share of land under 1st 4th Intensity use (prop. Tubewell Average Quarter Quarter HYV Rice All rice Vegetables Pulses of land) Ownership Far>150km Panel C Treatment 0.043** 0.048*** 0.109 0.016** 0.030* 0.041*** -0.112*** -0.025* 0.003** 0.027*** (0.009) (0.005) (0.065) (0.004) (0.012) (0.007) (0.016) (0.010) (0.001) (0.003) Treat*RVeg*Far150 0.114* 0.004 0.129 -0.001 -0.000 -0.007 0.051* 0.023*** 0.009 -0.003 (0.041) (0.007) (0.132) (0.003) (0.003) (0.003) (0.018) (0.003) (0.004) (0.002) - Treat*Far150 -0.081 -0.070* 0.085 0.008** 0.003 0.017*** 0.045 -0.056 0.002 0.023*** (0.058) (0.029) (0.154) (0.002) (0.004) (0.003) (0.085) (0.030) (0.015) (0.003) Treat*Rveg -0.053 -0.030** 0.035 0.002 -0.002*** 0.008* -0.016 -0.011 -0.001 0.005 (0.079) (0.008) (0.059) (0.001) (0.001) (0.003) (0.007) (0.008) (0.004) (0.003) Far>170km Panel D Treatment 0.035* 0.044*** 0.099 0.016** 0.031* 0.042*** -0.111*** -0.023** 0.002 0.026*** (0.016) (0.007) (0.072) (0.004) (0.013) (0.007) (0.014) (0.007) (0.001) (0.003) Treat*RVeg*Far170 0.054 0.008** 0.041 -0.003** 0.002** -0.007** 0.061*** 0.018** 0.004 -0.005** (0.038) (0.003) (0.050) (0.001) (0.001) (0.002) (0.004) (0.006) (0.002) (0.001) Treat*Far170 0.080 0.009 0.186 0.012*** -0.002 0.019*** -0.021 -0.061* 0.009 -0.001 (0.069) (0.010) (0.171) (0.002) (0.003) (0.004) (0.011) (0.023) (0.005) (0.005) Treat*Rveg -0.011 -0.032** 0.090 0.003 -0.002** 0.009** -0.016* -0.008 0.001 0.006 (0.056) (0.011) (0.066) (0.002) (0.001) (0.003) (0.006) (0.009) (0.004) (0.003) Note: “FarK”: is a dummy that takes the value of unity if a subdistrict is located farther than the distance cut-off K (e.g. K=110km in panel A) and zero otherwise. Rveg: is a dummy that takes the value of unity if suitability of vegetables production is greater than that for High Yielding Variety (HYV) rice and zero otherwise. Treat is unity if subdistrict is located in North-West region that is treatment region of Jamuna bridge. Each labeled column reports results for the labeled dependent variable and its respective standard errors are below in parenthesis. Each regression uses the same set of controls are reported in Table 3. Standard errors are clustered at regional (division) level. Legend: *** p<0.01, ** p<0.05, * p<0.1.