We find no evidence that the dams included in the sample are more or less likely to be used for irrigation purposes or to supply water to cities. However the excluded and included dams differ in terms of their height, the size of their reservoir, their capacity, and their average capacity lost due to sedimentation. These results suggest that our inclusion criteria is somewhat biased toward larger dams that can retain more water, but it is uncertain whether this is likely to introduce bias in our analysis in terms of dam performance and its impact on child nutritional status. In their influential paper Duflo and Pande use Indian districts as their unit of analysis and proceed to identify which areas are upstream and downstream from each other. However it is unclear whether one can apply this strategy in Africa. In particular, visual inspection of administrative regions in Africa reveals that the borders of many regions run at least partially along rivers, see for instance the case of the Southern African tip in Figure 1.3. As a consequence many regions contain both the catchment and the command area of a dam. Strobl and Strobl propose an arguably superior spatial breakdown in terms of upstream and downstream relationships that is based on actual river flow data. The U.S. Geological Survey Data Center has developed a geographical database, the HYDRO1K, providing a number of derivative products widely used for hydrological analysis. They use the drainage basin boundaries data from the HYDRO1K which divide the African continent into 7131 6-digit drainage basins with an average area of 4200 km2 . More critical and important for our analysis,plastic flower buckets wholesale the database assigns to each basin a code that allows one to determine whether it is upstream, downstream or not related to another basin.

Figure 1.4 depicts the spatial breakdown of the African continent according to our 6-digit basins. For comparison the figure depicts these jointly with the outline of the country borders. Basins vary greatly in shape and size, with a large number crossing national borders. Figure 1.5 depicts the 6-digit basins and the outline of administrative regions in the Southern African tip. The figure confirms that even at the sub-national level there is little correspondence between administrative regions and 6-digit basins. A similar picture emerges for the Southern Indian region where there is no obvious correspondence between administrative regions and 6-digit basins, see Figure 1.6. The main challenge for estimating the effect of dams on child nutrition is that dams are unlikely to be randomly allocated across regions, leading to a serious endogeneity problem . Moreover with a cross-section of 6-digit river basins, we are unable to control for invariant basin characteristics that influence dam location and are correlated with child nutrition, a strategy that would attenuate the endogeneity problem. In their study of Indian dams, Duflo and Pande use the share of dams in a state prior to their period of analysis interacted with a district’s suitability for dam construction based on the district’s river gradient to construct an estimation of the number of dams in each district. They then use these estimated number of dams as instruments for the actual number of dams in a district. In this paper we implement an instrumental variable strategy developed by Strobl and Strobl who modify Duflo and Pande’s approach along several dimension. Strobl and Strobl use the fact that starting with European colonization, a number of treaties were signed between African states to clarify the management of water resources.

Treaties, especially those signed in the colonial period, focused on the division of water resources or encouraged the construction of dams. For instance Lautze and Giordano note that about three quarters of the treaties cited as a goal the construction of dams for hydropower purposes and/or to expand the area of irrigated land. Strobl and Strobl use the fact that every country on the African continent has territory in at least one treaty basin. In the HYDRO1K data set, treaty basins correspond to 1-digit and 3-digit Pfaffstetter code classification and cover 60 per cent of Africa’s total land area. Starting with European colonization, a number of treaties were signed between African states to clarify the management of water resources. Treaties, especially those signed in the colonial period, focused on the division of water resources or encouraged the construction of dams. For instance Lautze and Giordano note that about three quarters of the treaties cited as a goal the construction of dams for hydro power purposes and/or to expand the area of irrigated land. To construct the relevant geographical delineation of the policies influencing dam construction, Strobl and Strobl use two databases. The first is the International Freshwater Treaties Database, which provide a comprehensive collection of international freshwater related agreements since 1820, including summaries of these as well as coding them according to the year signed and the river basins and countries involved. The second is the database on historical formation of treaty basin organizations in Africa compiled by Bakker. Combining these two databases reveals a total of 98 treaty basin organizations that were formed and which involve 53 countries and 59 river basins since 1884. Figure 1.7 depicts these treaty basins. As emphasized earlier the treaty basins are clearly transnational, cutting generally across several countries. Moreover their size, ranging from 1-digit to 3-digit Pfaffstetter code and their potential extent of coverage is at a substantially larger scale than the individual regions that we use as our unit of analysis, 6-digit Pfaffstetter code. In this paper we use this specific policy context and the approach in Duflo and Pande to develop an instrumental variable strategy to estimate the effect of dams on child nutrition.

As in Duflo and Pande we use the fact that a 6-digit basin’s suitability to dams should influence the number of dams built in the basin relative to other 6-digit basins in the same treaty basin. More specifically we interact a 6-digit basin’s river gradient and the proportion of dams in the treaty basin it falls into as an instrument for the number of dams in the 6-digit basin. As such we only rely on within treaty basin differences in suitability to dams to estimate the effect of dams on child nutrition. Moreover, in the African context, an important distinction needs to be made between perennial and ephemeral rivers, where the former’s flow is continuous, but for the latter water only flows for part of the year. Ephemeral rivers tend to be located in the dry lands of Africa and are much less suitable for dams, see Seely et al.. For instance, to intercept a large volume of water, a dam on an ephemeral river must be large in relation to the average inflows, but such dams are under high risk of failure because of the unpredictability of flash floods. Nevertheless because of the lack of sufficient perennial water sources,black flower buckets many countries rely at least in part on ephemeral rivers for dam location as well. For example, in Namibia only 10 per cent of the population rely on perennial rivers for their livelihood, and only 3 of the 19 major dams of the FAO database are located on those rivers. Treaty basin fixed effects ηb control for time-invariant characteristics that affect child nutrition, which are correlated with the likelihood of dam construction allowing us to only use within river basin and cross sub-basin variations for identification. However even in this situation, there might be unobservable determinants of child nutritional status that are correlated with the incidence of dam construction. In this case OLS estimates of the effect of dams will be biased. For instance if sub-basins where households are relatively richer are more likely to receive dams then the OLS estimate of β1 will be biased upward while the OLS estimate of β2 is likely to be biased downward. As in Duflo and Pande, we use the non-monotonic relationship between river gradient and the incidence of dam construction to implement an instrumental variable strategy. The approach consists in using exogenous variation in geographic features of different river basins to estimate the number of dams in a sub-basin. These estimated number of dams are then used to instrument for the actual number of dams. We construct measures of of a sub-basin geography such as elevation and river gradient using topographic information for multiple cells in each river basin. These information are used to compute the fraction of each sub-basin in different elevation categories and the fraction of a river basin falling into four gradient categories. Lastly to compute river gradient we restrict to cells in a sub-basin through which a river flows and compute the fraction of area in the above four gradient categories. Our panel on dam construction allows us to use all the information available to estimate the number of dams in a given sub-basin located in a river basin at certain points in time. Three sources of variation are used to predict the number of dams in a sub-basin: differences in dam construction across years in Africa, differences in the contribution of each each river basin to the increase in dams built, and differences across sub-basins driven by geographic suitability. First, we show that the river gradient matters for dam location. As a first step we regress the number of dams in 2000 on the fraction of river gradient in each gradient category by type of river, the average gradient in the 6-digit basin, river length by type of river, total area of the basin, and treaty basin fixed effects. We only show the coefficients on our main variables of interest.

The results of this analysis are reported in Table 1.3, columns and , and are consistent with Duflo and Pande’s finding for perennial rivers: moderate gradients in perennial rivers are more likely to be associated with dam construction. We also find that high gradients are less likely to be associated with dam construction. For ephemeral rivers we find that moderate and high gradients are less likely to receive dams. One possible explanation for this is that ephemeral rivers tend to require wider water flow for dam construction and tend to be less steep than perennial rivers. Moreover many of the dams with water supply purpose tend, in our data, to be located on low gradient ephemeral rivers. We also estimated the model in the sample of dams with irrigation as one of the major purpose and find qualitatively similar results. Overall these results provide support for using river gradients calculated separately for perennial and ephemeral rivers as predictors of dam construction. Next we report in columns and of Table 1.3 the estimated coefficient of RGrjks×Dbt from the first step regression in the pooled sample over all years. Column shows the results for all dams while column reports the results for dams with some irrigation purpose only. The results for perennial rivers are overall similar to the cross-sectional results. For ephemeral rivers we find that as the share of dams in the treaty basin increases, additional dams are less likely to be built in 6-digit river basins with very small river gradients . Table 1.4 presents estimates of the effect of dams on the nutritional status of children. Panel A provides Feasible Generalized Least Squares estimates, and panel B Feasible Optimal IV estimates. The coefficient on “own dam” captures the impact of dams built in that 6-digit river basin, while “upstream dam” measures the effect of dams in upstream 6-digit river basins. In this table each row corresponds to a separate regression; row 1 and 3 present estimates where the dependent variable is height-for-age z-score or an indicator equal to one if a child’s height-for-age z-score is below -2 points of standard deviation; while row 2 and 4 present estimates using weight-for-age z-score or an indicator equal to one if a child’s weight-for-age is below -2 points of standard deviation. The models in columns 4 to 6 and 10 to 12 are estimated using a linear probability mode. In columns 2 to 6, the analysis is restricted to dams with irrigation as one of their main purposes, while in columns 7 to 12 we include all dams.