Dams Duflo

Dams Author(s): Esther Duflo and Rohini Pande Source: The Quarterly Journal of Economics, Vol. 122, No.
2 (May, 2007), pp. 601-646 Published by: Oxford University Press Stable URL: http://www.jstor.org/stable/25098854 . Accessed: 12/06/2013 07:22
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp
.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.
Oxford University Press is collaborating with JSTOR to digitize, preserve and extend access to The Quarterly Journal of Economics.
http://www.jstor.org
This content downloaded from 83.145.66.202 on Wed, 12 Jun 2013 07:22:00 AM All use subject to JSTOR Terms and Conditions
DAMS*
Esther
This tion dams studies paper in India. Our
Duflo
and Rohini Pande
affects gradient to rainfall shocks and vulnerability from a dam, agricultural increases, production in shows an insignificant increase In contrast, agricultural declines. production Rural increases. the dam is located but its volatility the district where poverty in the district where the dam is in downstream declines districts but increases that neither markets built, suggesting distributional adverse impacts of dam nor state institutions have alleviated the construction.
and distributional effects of large irriga the productivity variable estimates instrumental exploit the fact that river a district's In districts for dams. located downstream suitability
I. Introduction
"If you are to suffer, you should suffer in the interest of the country." to those displaced Prime Minister Dam, Nehru, by Hirakud speaking Indian 1948.
to undertake major in dam construc World Bank investments 19 percent of the world's tion. By the year 2000, dams generated electricity supply and irrigated over 30 percent of the 271 million hectares these dams also dis However, irrigated worldwide. over 40 million altered and placed people, cropping patterns, increased and of arable land salination significantly waterlogging on Dams 2000a]. The distribution of the costs [World Commission and benefits of large dams across population groups, and, in extent the to which rural the have poor particular, benefited, are issues that remain widely debated. Dams provide a particularly good opportunity to study the the distributional and productiv between potential disjunction of a public policy. The of dam ity implications technology
* Pande for financial thanks NSF We support under grant SES-0417634. the editor Lawrence three anonymous thank, without Katz, implicating, referees, Marianne Michael Bertrand, Greenstone, Angus Deaton, Greg Fischer, Michael Dominic T.N. Srinivasan, Chris Udry, and, especially, Kremer, Leggett, Abhijit for comments and Chris Hansen and suggestions. We also thank James Banerjee Callie Scott, and Hyungi Woo for excellent research assis Fenske, Niki Klonaris, at Columbia for GIS help, and Abhijit Banerjee, Shawn tance, CIESIN University Siddharth Petia Topalova Cole, Seema Sharma, and, especially, Jayachandran, for sharing their data. ? 2007 by the President and Fellows ofHarvard Technology. The Quarterly Journal ofEconomics, May 2007 College and the Massachusetts Institute of
over 45,000 Worldwide, large dams have been built, and nearly half the world's rivers are obstructed by a large dam. The belief that large dams, by increasing irrigation and hydroelectric and reduce poverty has led ity production, can cause development countries such as the and international developing agencies
601
602
QUARTERLY JOURNAL OF ECONOMICS
is often restricted, especially in rain-scarce years. This upstream increases the vulnerability of upstream agricultural production to on Dams 2000b; Shiva 2002]. rainfall shocks [World Commission What is striking in the policy debates surrounding dam con struction is the absence of systematic empirical evidence on how the average large dam affects welfare, especially of the rural poor. This paper aims to provide such evidence. We focus on India, is the world's third most which, with over 4,000 large dams,
than is typically assumed. of large dam construction emphasize the role of Proponents on rainfall and enabling in reducing dependency large dams In contrast, oppo irrigation, providing water and hydropower. nents argue that while these benefits may be enjoyed by down stream populations, benefit only from the upstream populations from increased economic construction activity and potentially of compensation, activity around the reservoir. In the absence they suffer potentially large losses; flooding reduces agricultural and forest land, and increased salinity and waterlogging reduces the productivity of land in the vicinity of the reservoir [McCully 2001; Singh 2002]. Further, to fill the dam reservoir, water use
from a dam construction implies that those who live downstream stand to benefit, while those in the vicinity of and upstream from a dam stand to lose. From an econometric viewpoint, this implies that we can isolate the impact of dams on the two populations, this suggests compensating losers and from a policy perspective, in such a of compensation is relatively easy. The inadequacy simple case would suggest that the distributional comparatively to remedy of public policies are perhaps harder consequences
prolific dam builder (after China and the United States). Large dam construction is the predominant form of public investment in 1950 and 1993, India was the single irrigation in India. Between Bank of World largest beneficiary lending for irrigation.1 An important justification for such investment, both by the Indian government and theWorld Bank, is agricultural growth and rural [World Bank 2002; Dhawan 1993]. poverty alleviation A comparison of outcomes in regions with and without dams is unlikely to provide a causal estimate of the impact of dams since regions with relatively more dams are likely to differ along other dimensions, such as potential agricultural productivity. To
1. India received roughly 26 percent of Bank loans, and irrigation made up 7 1951 and 1997, Indian [World Bank 2002]. Between percent of total Bank lending in major and medium 33 public investment irrigation projects was approximately
billionUS dollars [Thakkar 2000].
DAMS 603 in the address this problem, we exploit the fact, well documented dam engineering literature, that the gradient at which a river flows affects the ease of dam construction. In particular, there is a nonmonotonic in an area between river gradient relationship and its suitability for dam construction. Low (but nonzero) river gradient areas are most suitable for irrigation dams while very dams. steep river gradient areas are suitable for hydroelectric or where the river is either flat somewhat steep gradient Regions are the least likely to receive dams. is the administrative Our unit of analysis unit below an in dam construction Indian state, a district. We exploit variation induced by differences in river gradient across districts within Indian states
to obtain instrumental estimates. Our variable time control for differential effects, regressions state-specific and for other effects of river gradient, time-varying national district geographic features. outcomes. We Our first set of results relate to agricultural use annual data find and district-level that dam agricultural construction leads to a significant increase in irrigated area and in districts located downstream from the agricultural production dam. Dams also provide insurance rainfall shocks in against downstream districts. In contrast, dams have a noisy and typi in the dis cally insignificant impact on agricultural production tricts where they are built, and the vulnerability of agricultural in these to rainfall shocks significantly increases production districts. Our second set of results relate to rural poverty. Using dis trict-level poverty data at five points in time we find that dams significantly increase rural poverty in districts where they are located. In contrast, poverty declines in districts downstream from the dam, but relative to the increase in the dam's district, this decline is small. Our poverty results are consistent with the and volatility. It is worth production findings of agricultural noting, however, that with only five years of poverty data, it is difficult to disentangle the poverty impact of dam construction time trends in poverty outcomes from differential in districts to dams construction relative to other districts in the suitable
same states.2
together, our results suggest a failure of state-level we cannot redistributive institutions. While identify all the Taken
2. Inclusion of such trends makes own district poverty very large. the standard error of the estimate of the
604
reasons for inadequate of those who lose out, we compensation are able to show that the poverty impact of dam construction is in districts with a history of relatively more extrac accentuated land tenure sys tive institutions (as captured by their historical see and Our tems, Iyer [2005]). Banerjee finding lends support to the view that institutions play an important role in ensuring the distribution of productive gains between winners and losers. in In Section II, we describe the dam construction process on case dam construction review the literature India, large study in India, and use a simple production function framework to identify the expected effects of dams. In Section III, we describe the empirical and conclude strategy. We provide in Section V. empirical results in Section IV
II. Background up over 90 Irrigation dams, the focus of our study, make we In this the dams. first describe of India's section, percent large the Indian irrigation dam construction process and, in particular, in determining dam placement. We then use a role of geography simple agricultural production function framework to outline the main costs and benefits of irrigation dams. Finally, we discuss how Indian district-level data may be used to estimate some of these costs and benefits. II.A. Dam Both Construction in India
in India play an the federal and state governments in The role construction. dam Indian Com Planning important mission (a federal body) sets each state's five-year water storage sur and irrigation targets. Given these targets and topological of each veys of potential dam sites, the irrigation departments state proposes dam projects. Next, a federal committee examines the economic visibility of these projects. The Planning Commis sion selects the final projects on the basis of investment priorities and sectoral planning policies. Construction of a dam and the remain association canal network the state's responsibility, though federal or international funding may be available. The typical Indian irrigation dam is an earth dam: Water is impounded in a "reservoir" behind an artificial wall built across a
river valley. Artificial canals channel water from the reservoir to downstream regions for irrigation. The area upstream from which water and silt flow into the reservoir and the area submerged by
DAMS 605 the reservoir form the catchment area of the dam, while the area covered by the canal network is the command area. downstream The state government agency and farmers jointly manage the irrigation system associated with a dam. The agency deter to release to each branch outlet, and mines how much water farmers served by that outlet decide how to share it. Farmers working on land covered by the dam's canal network are eligible
for dam irrigation. In contrast, to ensure that the reservoir is agency often restricts water withdrawal filled, the government from the reservoir (for instance, by cancelling water upstream in [Shiva years of limited rainfall sites), especially pumping 2002]. The
viability and cost of dam construction at a location de pend crucially on its geography. The ideal dam site is in a narrow river valley where the river flows at some gradient. Since current satellite imagery resolution for India is too coarse for us to iden tify the width of a river valley, we exploit the relevance of river gradient for dam construction. reasons as The dam engineering literature provides multiple a some to why river flowing at is preferred for dam gradient construction. The first reason relates to reservoir construction. to Golz [1977, p. 7], "to obtain economical According storage a reservoir site should be wide in comparison to the dam capacity site and should be on a stream having a low or gentle gradient to obtain a long reservoir in proportion to the height of the dam." The second reason is that water flow from dams to the irrigated area is typically via gravity. Hence, according to Cech [2003, pp. "Dams for irrigation projects are generally constructed 150-152], at an elevation high enough to deliver irrigation water to crop land entirely by gravity." That said, the river gradient should not be too steep because fast flowing water would erode the canal. The U.S. Department of Agriculture states [1971, pp. 320-322] that dam canals "should be designed to develop velocities which are non-erosive for the soil materials through which the canal or
lateral passes." Water for velocity, and therefore the potential erosion, increases with river gradient. As a result, for irrigation dams, the recommended practice is to target dam sites where the river gradient in the direction of irrigation is neither too steep nor completely flat. In contrast, higher river gradient reduces the cost of produc [Warnick 1984]. To quote Cech ing hydroelectricity [2003, pp. "Dams for hydroelectric power generation are located 150-152], at a site where the difference in elevations between the surface of
606 the new
QUARTERLY JOURNAL OF ECONOMICS reservoir to power and the outlet
to the downstream river is turbines." adequate electrical-generating To summarize, engineering considerations suggest that river a on the likelihood of have nonmonotonic effect should gradient dam construction. A river flowing through a flat area is less likely to see dam construction while areas where the river flows at a moderate should see more dam construction. Finally, gradient while areas with a steep gradient are less suited for irrigation, areas with a very steep river gradient favor hydroelectric dam
construction.
II.B.
Benefits
and Costs
of Irrigation
Dams
in India Between 1951 and 2000, food grain production with two-thirds of this increase coming from nearly quadrupled, [Thakkar 2000]. While dams account for 38 per irrigated areas cent of India's irrigated area, estimates of what fraction of the can be attributed to dams vary from 10 increase in production on Dams 2000b] to over 50 percent [World Commission percent 2000]. [Gopalakrishnan To clarify the potential role of dams in affecting agricultural production we describe a simple agricultural production function and McKinsey [1999]. framework, which is based on Evenson We assume that agricultural output is a function of labor inputs L, land surface K, land quality A, inputs such as fertilizer, seeds, and electricity /, climate r (rainfall and temperature), farmer's ability u, and a productivity shock e. We denote the access to an irrigation sys for land without function production tem (via pump or canal) as = y FiiL&AJSyUye) and the production system as function for land with access to an irrigation
y=F2(L,K,A,I,r,u,e).
Evenson and McKinsey these production [1999] estimate functions using Indian data. They find that irrigation reduces the the effect of rainfall shocks volatility of production by mitigating and temperature. Further, irrigation and agricultural inputs, such as fertilizer, electricity, and seeds for high yielding variety (HYV) crops are complements. These findings and other studies, such as Singh [2002], suggest that irrigation enhances productiv and the cultivation ofmore ity by increasing multi-cropping itable water-intensive cash crops, especially sugarcane. prof
DAMS 607 Assume that access to irrigation has a fixed cost. This is the cost of accessing ground water in a region with no dams and the cost of accessing canal irrigation in a dam's command area. If a farmer can obtain the optimal set of inputs, then she will invest in irrigation if its cost is below the long-run difference between the value function with and without irrigation. Her decision pro cess follows a threshold rule: she switches if the productivity shock exceeds some threshold in a given period. to other forms of water harvesting, such as ground Relative cost water and small dykes, dams reduce the fixed of accessing 2001; Dha irrigation in the command area [Biswas and Tortajada wan 1989]. Availability of dam irrigation will not affect the irriga tion choices of farmers who have paid the sunk cost of accessing those farmers who would have ground water irrigation. However, invested in ground water irrigation in the future will instead opt for dam irrigation. Finally, some of the farmers who would not have chosen ground water irrigation will invest in the cheaper dam irri gation. Dams, therefore, increase irrigated area (though by less than the demand the area actually irrigated by the dams). Consequently, of for labor, fertilizer, and seeds will increase, and the dependence on decrease. rainfall will agricultural production The impact of a dam is different in its catchment area. First, a significant fraction of land in the catchment area is submerged on during dam construction. For instance, theWorld Commission
4.5 Dams that dam construction [2000b] estimates submerged million hectares of Indian forest land between 1980 and 2000. is usually accompanied Land submergence by large-scale popu lation displacement. of what Estimates fraction of the Indian population has been so displaced vary between 16 and 40 million on Dams that [2000b] estimates people. The World Commission a the average Indian dam displaced while World 31,340 persons Bank study in the mid-1990s estimated a lower number of 13,000 people per dam [Cernea 1996]. Case study evidence suggests that tribe populations scheduled have historically disadvantaged borne the brunt of displacement.3 Second, water seepage from the canal and the reservoir in creases waterlogging and soil salinity and makes land less pro ductive and Hildyard Indian Water [Goldsmith 1984]. The Resources Ministry estimated that roughly one-tenth of the area
3. Official figures for 34 large dams show that scheduled of India's 47 percent constituted up 8 percent population, on Dams [World Commission 2000b].
tribes, which make of those displaced
608
or salinity/ irrigated by dams suffered from either water-logging on Commission Dams 1991 [World 2000b].4 While alkalinity by around the canal and the reservoir both waterlogging happens command remote also affects the therefore (and area), sensing in the studies suggest that these problems are more pronounced area Khan and show that Sato [2000] (for instance, drainage are to the canals rela network, irrigation drainage compared tively unaffected by this problem). This implies that irrigation in the catchment area is less profitable, but fertilizer use may in crease since poorer soil requires more nutrients. to Finally, unaffected land in the catchment area upstream the reservoir is unlikely to benefit from dam irrigation, as lift [Thakkar 2000]. In fact, irrigation is rarely practiced for dams which control the flow of water government (through agencies, the opening of gates and sluices and the control ofwater pumping water distribution sites), typically maximize through the canal network and, to achieve maximum water storage in the reservoir, often restrict water use upstream from a dam. Such restrictions are particularly in drought years when rainfall is in prevalent sufficient to fill the reservoir (see, for instance Shiva [2002] and Tehri Report of a dam in a [1997]). As a result, the presence district may exacerbate water shortage and, therefore, the vari ance of agricultural in areas upstream to the dam. production, land Taken cultivated and together, irrigated area potentially and production are likely to decline in the catchment area. Our model of agricultural leaves out some bene production fits of dam
construction. Dams may prevent floods and droughts the flow of water downstream. These effects can by regulating There is, however, a trade-off be extend very far downstream. tween using dams for flood control (which requires emptying the reservoir) and for irrigation (which requires filling the reservoir). Dams may also be used for electricity generation. Finally, dam reservoirs often provide a source of fishing and are sometimes as tourism sites.5 developed On the cost side, an often-cited consequence of dam construc tion is adverse health effects for those living near the reservoir. A
4. Some argue that seepage may lead to benefits in the longer run because it allows water [Dhawan 1993]. recharge 5. However, in India, there is very limited development of reservoirs for Indian studies of reservoir fisheries conclude tourism, and most that, while fish can be bred in the reservoir, in reductions in the fish production large neighboring of the river (caused reaches in river flow) imply an by dam-induced changes on fisheries overall negative effect of dam construction [Jackson and Marmulla 1994].
DAMS 609 reservoir provides hence, for diseases a natural for vector breeding habitat and, such as malaria, filariasis, and river blindness
[Sharma 1991]. In this paper, we evaluate the impact of dams on agricultural and rural welfare (as measured by poverty rates). production and public policy, affect how Several factors, notably migration the impact of dams on land productivity translates into individual
welfare outcomes.
ment resolutions. Numerous studies suggest compensation rights of the landless, and those without formal land titles are typically not recognized is usually [Thukral 1992]. Further, compensation to replace lost land by its equivalent insufficient for the displaced in quality and extent elsewhere [Dreze, et al. 1997]. A related policy intervention would be to charge the down stream population for water usage. However, water charges in India remain so low that they often do not even cover the opera costs of the canal system [Jones 1995].6 In tion and maintenance Prasad and Rao [1991] show that tax collection costs often fact, exceed the amount collected. The inability or unwillingness to
to expect labor and capital to migrate away It is reasonable from adversely affected parts of the catchment region towards the command area. Such migration would increase land prices in the command area and reduce the beneficial impact of dam construc and poverty in the command area. It would also tion on wages a smaller poverty increase in the catchment area than is imply several predicted by the direct physical impact of dams. However, recent papers and [2006] (including Jayachandran Topalova [2004]) show that factor market imperfections significantly in in turn, implies that, in hibit such migration. Limited migration, will cause wages line with their productivity to dams effects, increase and poverty to fall in the command area. The converse will be true in the catchment area. Even if labor is immobile, public policy can reduce the impact of dam construction on welfare outcomes. India's rehabilitation on the Land Acqui for is based dam-displaced policy populations to sition Act of 1894. This Act empowers the Indian government acquire land for public purpose in return for cash compensation. and compensation is the responsibility Resettlement of the rele vant project authorities and is based on project-specific govern
6. In 1980, the annual recurrent cost per hectare for access to irrigation was was closer to Rs. 300 roughly Rs. 50 (1 U.S. dollar), while that of dam maintenance (or 6 U.S. dollars).
610
fee for water has three consequences: charge an appropriate to switch to farmers in the command area have been encouraged water intensive crops, thereby, according to some, seriously lim the canal systems are poorly iting the gains inwater availability; their which reduces effectiveness [Jones 1995]; and maintained, of any benefits is limited. redistribution II.C. Districts as the Unit ofAnalysis
We will estimate the effect of dams on economic outcomes at for the district level, which is the lowest level of disaggregation and agricultural which household consumption production data
are available.
of 8,000 square kilometers. extent of the catch The absence of data on the geographic ment and command areas of large Indian dams prevents us from identifying the fraction of district area covered by the catchment and command areas of each dam. Available catchment and com in that district which the dam is mand area maps the suggest or most all of its catchment contains located usually physically
area, even for
A district is the administrative unit immediately under the to a county in the United Indian state (somewhat analogous States) and forms the natural unit for the planning and imple of state policies. In 1991, India had 466 districts, with mentation a district, on average, having a population of 1.5 million and an area
in the command area of a dam is usually via gravity through artificially constructed canals. The command area lies downstream from the reservoir, with the canal network direction along the main canal and extending in the downstream often covering parts ofmultiple districts.8 The estimated effect of a dam in the district where it is built combines the effects in catchment, and unaffected command, The effect of a dam on agricul areas, and is a priori ambiguous. in the neighboring tural production downstream district should be unambiguously positive (since it only has part of the command Water distribution
large
dams.7
are Chari et al. 7. Some and Vidhya [1994] and Chari [1995], examples et al. [2002], and Vidhya et al. [2002]. Chakraborti 8. See again Chari and Vidhya et al. Chakraborti [1994, 1995], [2002], and et al. [2002] for specific examples. there are instances where the Vidhya However, area covers districts command In personal that are not downstream. correspon the Bank's senior water and (until recently dence, John Briscoe professional on water area of one of The command issues) offered one example: spokesperson India's dam, covers part of the district lateral to largest dams, the Bhakra Nangal that in which it is built.
DAMS 611 the command area effect (both be area) but will underestimate cause some (or all) of the downstream district will not be in the command area, and because part of the command area will not be to expect that the dam's own in that district). But it is reasonable downstream district and the neighboring districts will be the most affected by the dams. In Table IV, we examine the effect of dams in all neighboring districts and find that, on average, dams only affect production and rural welfare in the district where they are built and in neighboring downstream districts. Our estimates are not quantitative of the impact of the dam in its estimates command and catchment areas, but rather its impact on the most units. relevant corresponding administrative aside from the fact consistent data are available that Finally, at the district level, a district-level analysis presents one impor are relevant markets tant advantage: and social units Districts within which people might relocate were displaced or their land became tion across district lines in response III. Empirical (for example, because they less productive), but migra to shocks is rare.
Strategy
Our analysis exploits detailed Indian district panel data on district geography, dam placement, and poverty and agricultural outcomes. We have annual agricultural production data for 271 districts for 1971-1999 and poverty data for 374 districts for five years: 1973, 1983, 1987, 1993, and 1999. Table I provides descrip tive statistics, and the Appendix describes the data sources and
variable construction.
I and II, which depict district-wise construction. Figures dam construction in 1970 and 1999, respectively, show that dam construction was concentrated in western India, with relatively in north and northeastern little dam construction India. Figure III graphs overall dam construction in India. Dam construction was between the mid-1970s and late-1980s but slowed rapid down considerably in the 1990s. OLS of how the number of dams in a regression estimates or welfare, outcomes are unlikely to be state affect agricultural, consistent. Richer and fast growing states can build relatively more dams. States that anticipate larger increases in agricultural dam
1971 and 1999 the number of large dams in India from 882 to 3,364, and the average number of dams in quadrupled a district increased from 2.39 to 8.66 (46 percent of the districts had no dams in 1999). There was significant regional variation in Between
TABLE I
Descriptive Beginning A. Geography Fraction district with Fraction Fraction Fraction district with district with district with river gradient river gradient river gradient river gradient 0-1.5% 1.5-3% 3-6% above 6% 0.75 0.08 0.06 0.11 GIS GIS GIS C2 GIS 2 Statistics period ^ End
period So
(0.30) (0.07) (0.08) (0.22) ?> g gj 8.66 (16.58)7 (30.57)G 5.79 3.80 P. (0.27) 0.22 0.46 ^ (0.14) 0.27 0.04 (0.04) 0.47 1.31? (1.85)<3 1.44 1.71 (0.31) $ (324) 169 138 (176)
B.
Dams Number ofdams in district

Number of dams upstream to district
2.39 (4.55) (7.77) (0.20) (0.16) (0.05) (2.10) (0.35) 570 (277)
tons) 52.26 115.27 15,931
13.85 3.63
C. Welfare Per capita expenditure (logRupees) Headcount ratio Poverty gap Malaria incidence Agriculturalwage (logRupees)
D. Agriculture
Gross cultivated area (in '000hectares) Gross irrigatedarea (in '000hectares)

Total production (Rupees (Rupees per '000
03
(151)
Yield
per hectare)
(12,749)
0.079 Rainfall shock
(110,130)
-0.231
81,336
(30)
(62)
(0.25)
(0.23)
TABLE I
(continued) Beginning E. Demographics period End
period S
Rural population (in '000 persons) Rural in-migrants(in '000 persons) Fraction villages in district with anywater facility Fraction villages in district with power
Fraction villages in district with tarmac road
and Public
Goods
1,295 (867) (239) 0.982 (0.03) 0.24 (0.19) 0.10 (0.48) (0.20)
0.31
(0.26)
1,887 (1,254) 366 486 (309) 0.997 (0.01) 0.81 (0.26)

0.47
(0.22)
Non-landlord districts (proportion)
0.61
Tribal population share in district in 1971
malaria 1975 and 1995 Beginning and end periods are (i) dams: 1971 and 1999; (ii)welfaremeasures: 1973 and 1999 (for wages: 1971 and 1987; (v)demographicand public goods: 1971 and 1991.All monetaryvalues are in 1973Rupees (at current exc inparentheses. Total production includesproductionofsixmajor crops: wheat, rice,sorghum (jowar), sugarcane,pearlmillet below the line. the as a is distance below the line of the line average living poverty Poverty gap expressed proportion poverty Malaria incidenceis the logofaverage parasite incidence in thedistrict. Rainfall shock is t being zero distance below the line). mean (over 1971-1999). Non-landlord district is a district-level was un which equals one if dummy majority area in thedistrict sources and variable constructionare in theAppendix.
614

Q N*D?s
k Figure Distribution of Dams across Data of Dams on the number I
Indian Districts, 1970 dams per district is from the World Registry of completed on Large Dams See Commission (ICOLD). published by the International for further details. Appendix
productivity are also likely to make more of these investments, implying a spurious positive relationship between poverty reduc tion and agricultural Indeed, Mer growth and dam building. rouche [2004] finds larger poverty reductions for states that built a causal more cannot be given but these findings dams, interpretation. Our identification strategy, therefore, relies on within-state in dam construction, differences specifically differences across districts in a state. We can, therefore, examine spillover effects from dams in neighboring districts but not state-wide economic effects of dam construction, such as their effect on prices deter
DAMS 615
NoDams |~~1
^Jp|{H^^J/s^,\^J? j
More
than 34 Dams
k Figure The Distribution Data of Dams on the number II
across of Dams 1999 Indian Districts, of completed dams per district is from the World Registry on Large Dams Commission See (ICOLD). published by the International for further details. Appendix
mined at the state level. We discuss the possible state-wide effects as we interpret our results. Consider the following regression:
direction
of such
= + m*+ <>*, (1) ylst Pi + PiA* + Pa^S + P4^ + PsZ? + v,-
where Dist denotes the number of dams in the district and D%t the number of dams located upstream from district i. vt is a district fixed effect, \xstis a state-year interaction effect and u>ist a dis trict-year specific error term. Zit and Zft are a set of time varying control variables for the district list of relevant control variables and for upstream districts is discussed below). (the
616
3,600 -

-i
j********
3,000
/T
2,400
1,800- /
1,200^++++****^ jf
600
0 -I-1-1-1-1-1-.-.-1-1-1 1950 1955 1960 1965 1970 Figure Total Dams Constructed in India, 1975 III ICOLD Dam Register for India 1980 1985 1990 1995 2000
District fixed effects allow us to control for time-invariant in characteristics that affect the likelihood of dam construction the district and state-year interactions for annual shocks, which are common across districts in a state. We only exploit cross district variation in dam construction in a state for identification. Within from regressions using annual states, the residuals In this setting, data are strongly autocorrelated. agricultural least squares rather than OLS, regres estimating generalized sions can potentially increase efficiency. We compute feasible estimates GLS (FGLS) using the method proposed by Hansen [2006]. First, we obtain the time series process for the regression residual (after correcting for biases due to the small sample and inclusion of fixed effects). An AR(2) process best reflects the data, to construct the except for wages.9 We use these parameters in the time series pro FGLS weighting matrix. Misspecification cess can cause the conventional FGLS standard errors to overre [Bertrand et al. 2004]. Therefore, we ject the null hypothesis report standard errors, which are robust to arbitrary covariance of the FGLS residual within the state [Wooldridge 2003; Hansen 2006].10
9. For wages, the first difference over time?wage growth?is approximately i.i.d. on average. 10. If the time series process is correctly specified, then FGLS and these errors are asymptotically the time series "clustered" standard equal. Even when is misspecified, is consistent FGLS and implies efficiency gains relative to process
DAMS617
to be consistent requires that estimates and GLS For OLS across districts within a in construction dam variation the annual shocks. How with other district-specific state be uncorrelated for violated be this ever, instance, agricultur if, assumption may districts witness relatively greater dam ally more productive construction. To address this problem, we develop an instrumen tal variable strategy based on the geography of dam construction. III.A. Instrumental Variables
Strategy
between river gradient and The nonmonotonic relationship in Section the likelihood of dam construction II) (as described forms the basis of our identification strategy. of district To implement this strategy, we construct measures from a digital elevation map, which provides topo geography cells in each Indian district. We graphic information formultiple use information on surface elevation at each of the cells within a district to compute the fraction of district area in four different elevation categories: 0-250 meters, 250-500 meters, 500-1,000 meters, and above 1,000 meters. District gradient characterizes as the of the ground surface and is measured the steepness area of the We the fraction of district surface. compute tangent
falling into four gradient categories: flat (0-1.5 percent), moder ate (1.5-3 percent), steep (3-6 percent), and very steep (above 6 percent). Finally, to compute river gradient we restrict attention to cells in a district through which a river flows and compute the fraction of area in the above four gradient categories. overall gradient, In our analysis, we control for elevation, and river length in a district and use differences in river gradient to predict the annual distribution of dams built in a state across II and IV illustrate our identification strategy. districts. Figures we see In Figure II, that, despite the presence of one of the world's
the Indo-Gangetic basin, almost no dam largest river basins, construction has occurred in north India. Figure IV depicts the north India has very flat rivers, average river gradient?central
OLS better approximates the (since a common AR(2) (rather than an i.i.d. process) of the robust variance-covariance data generating matrix insures process). Use errors. standard consistent Simulations [2006] confirm that reported by Hansen errors do not significantly worsen power when the data is, in fact, robust standard are with an arbi OLS and FGLS results but FGLS AR(2). similar, qualitatively intervals, trary covariance matrix has tighter confidence implying efficiency gains errors robust to time series from estimating FGLS. OLS estimates with standard are available autocorrelation from the authors.
618

River Gradient Avtragr
Figure
IV
River Gradient (in Percentage), by District Average a elevation map which is computed river gradient The average digital using on surface elevation. We have elevation data for multiple information provides to cells river gradient we restrict attention cells per Indian district. To compute of the is a measure River gradient of the steepness through which a river passes. ground surface in the vicinity of district rivers and is defined as the tangent of the for further details. surface. See Appendix
while most
The
dam of western India, which has seen the maximum construction, has rivers with moderate gradient. In column (1) of Table II, we formally examine the relation ship between the number of dams built in a district by 1999 and include state fixed effects, district river gradient. Our regressions district elevation, river length, and district gradient as controls. omitted river gradient category is the proportion
of river in
DAMS 619 TABLE

Geography and Dam
II
Construction Number of dams
Cross-section (1999) Not interacted Poverty Interacted sample with dams Production sample of
number predicted in the state
(1) (2) (3)

Fraction Fraction Fraction river gradient river gradient river gradient 0.278 1.5-3%(0.122) -0.210 0.153 (0.040) -0.191 -0.219 0.176 (0.094)
3-6% (0.127)
0.014
(0.065)
0.075
(0.128)
0.097
above
F-test
6%
(0.033)
1.764
(0.031)
6.372
(0.043)
7.68
for river gradient
[p-value] Geography controls

State*year and river gradient*year Fixed
[0.15] Yes No
[0.000] Yes Yes

District District
[0.053] Yes Yes 7743
interactions
effects
State
374
1855
Geography controlsare river length (in kilometers),districtarea (in sq. kms), three elevation variables districtarea with elevation 250-500 m, 500-1,000 m, and above 1,000m, respectively), threedistrict (fraction districtarea with gradient 1.5-3%, 3-6%, and above 6%, respectively). Columns (2) gradientvariables (fraction and rivergradient*yearinteractions. and (3) regressions also includea fullset ofstate*yearinteractions Columns and rivergradient*yearinteractions. For these (2) and (3) regressionsincludea fullset ofstate*yearinteractions measures are all and the rivergradient interacted with predicteddams. The regressionsthegeographycontrols column (2) regressionis estimatedbyOLS with standarderrorsclusteredbyNSS region*year and thecolumn (3) are multiplied by 100. regressionby GLS with standard errors clustered by district.Regression coefficients Standard errorsare inparentheses. In columns (1) and (2) the sample includes374 districtsas denned by 1981 census.The povertysample includestheyears 1973,1983,1987,1993, and 1999.The productionsample includes annual data for271 Indian districts(using the 1961 Indian census definition) fortheyears 1971-1999.Missing observationsaccount foractual sample size. district-year
the flat gradient category (0-1.5 percent). Our results confirm the importance of engineering considerations: A gentle river gradient a steep increases the number of dams, while (1.5-3 percent) a it. reduces river (more very steep However, gradient gradient than 6 percent) increases dam construction. The last effect is to some multipurpose attributable dams in our sample that pro vide both irrigation and hydroelectricity.11 Our panel regressions build on these findings. To predict the
11. We exclude purely hydroelectric dams but cannot identify predominantly dams. Excluding the fraction of area in very steep hydroelectric multipurpose from the instrument set provides identical results. gradient qualitatively
620
number of dams in a district, we exploit three sources of variation in dam construction: differences in dam construction across years in the contribution in India, differences of each state to this across differences districts within a state increase, and, finally, of districts. that are driven by the geographic suitability of the form We estimate panel regressions 4 (2) Disl = a, + 2
k=2
a2k(RGrki
D~st) + a3(Mt
4
D~st)
+ 2
k=2
ot4k(RGrki
lt) + vt +
+ \Lst
<j>ist,
where Dist is the number of dams in district i of state s at time t. for time-invariant district Vj, the district fixed effect, accounts set of which and the affect dam may characteristics, iLst, building, accounts for of state-level the state-year interactions, impact
macro shocks.
RGrki denotes the river gradient variables. These enter the regression interacted with predicted dam incidence in the state, total dam construction in Dst. This is constructed by multiplying India with the fraction of total dams in the state in 1970. Use of that the predicted, rather than actual, dam incidence ensures measure is exogenous with respect to the number of dams in the district.12 The interaction of the RGrki variables with year dum mies time-varying effects of river gra (lt) accounts for national dient on the outcomes of interest. Mt, the vector of district-specific time-invariant control variables, includes district elevation and overall gradient measures, river length, and district area. coefficient estimates for the and column the for (3) poverty agri cultural production sample (twenty-nine years of data). The re sults are similar to the cross-sectional in column (1), regression the that of the interaction of in the very district except proportion and the number is of dams steep gradient category predicted F-statistics The that demonstrate instruments the significant. have sufficient power. Let Zist denote the vector of right-hand side variables in (2), * for the interactions define except RGrki Dst. Similarly, ZYst as Column (2) of Table II provides sample (five years of data),
12. The use of actual [Duflo and Pande 2005].
number
of dams
built
in the state yields
similar
results
DAMS 621 includes the same control variables the vector, which We estimate district.13 upstream = + (3) yist &! + o2D^ + 8aD^ + Zistb4 + Z^o5 + v* + ^ To generate instruments for Dist and D^, of dams we use for the
coto.
parameters For
from (2) to predict
the number
per district Dist.
upstream districts, the predicted number of dams, Dfst, is the sum districts (it equals of predicted values from (2) for all upstream zero if the district has no upstream district). We estimate (3) using Di8t9 DY8t9Zist, and ZY8t as instruments. is The first-stage equation (4) Aist = + Vi + Vst + + CJ)3D^+ Zist$4 + ZYst<$>5 Cj>!"h <\>2Dist <?ist,
where A^ represents Dist or DYst. information efficiently: by us Our procedure uses available the to districts between district geo all relationship predict ing than just those number of dams and the features (rather graphic we avoid averaging these features when that are upstream), there are several upstream districts.14 In the regression using annual (3) by data, we estimate of FGLS for IV). As with IV (the equivalent feasible optimal errors that are robust to arbitrary FGLS, we report standard of the residual within a state.15 covariance in the Our instrumental variable strategy exploits variation in the district (or in the interaction of the river gradient variable in the state district) with predicted dam construction upstream
sum river length and district area across all upstream 13. We districts and for the other variables (which are proportions). average Controlling separately district provides similar estimates. these variables for each upstream to a 14. If every district has a single upstream district, then this is identical of river gradient 2SLS procedure where the interaction variables with predicted dam incidence in the state are instruments. of dependent 15. If y is the outcome, X is the matrix variable, Z is the matrix of instruments, and tl is the estimated variance-covariance matrix of the regular IV residual each block is of size (for T years, this is a block diagonal matrix where T * T and is identical), IV estimate is equal to the optimal (X,Z(ZfCl-1Z)-1Z'X)-1(X'Z)(Z'Cl-1Z)-1ZfCl-1y, and its estimated standard error is given by
(X'Z(Z'n-1z)-1Z'Z)-HX'Z)(Z'A-1z)-1w(Z'a-1z)-^'Z(Z'n-1z)-1Z'X)-1. For N districts indexed is a row vector from 1 toN, W, in turn, equals Djl * tth row of the matrix thej Cl~1,2Z, = x UjUj, with Uj 2?= x and eJt is the residual
JtZjt, z% of the Optimal
IV regression.
622
the evolution of economic outcomes absent dam construction, across districts located in the same state but with different river differed across states gradients would not have systematically with more dams in 1970 and states with fewer dams in 1970. In were and Maharashtra the 1970, Gujarat, Pradesh, Madhya are and Maharashtra three states with the most dams. Gujarat income among India's richer states, though their agricultural growth is not the highest in our sample. Still, one may worry that our instrument is picking up nondam-related in differences across districts with different river gradient growth patterns in richer and poorer states. characteristics To address this concern, we report additional specifications. First, Figure III suggests that, at least for outcomes forwhich we have annual data, the time pattern of evolution of outcomes due to dam construction can potentially be distinguished from a lin ear trend that differs across regions with different river gradient. We, therefore, examine whether our results are robust to includ control the interactions of a linear trend with ing as an additional the share of dams built by the state in 1970 and the river gradient variables river (and the similar interactions with the upstream we our are check that results robust variables). Second, gradient to including as additional controls the initial tribal population share and initial poverty in the district, each interacted with the (and the corresponding predicted dam share in the state Dst interactions with upstream tribal share and poverty). This ad dresses the concern that regions favorable to dams may have also differed in terms of their initial poverty or tribal population shares, and that these initial conditions, not dam construction, determined and poverty changes. subsequent agricultural In reality, dams vary significantly, both in their physical characteristics (this includes dam height and canal network) and
16. Controlling for district gradient, we would not expect nondam irrigation to be sensitive to river gradient is indepen (ground water irrigation, for instance, on dent of river gradient, but may depend In line with this, our district gradient). or not we include the interaction results are not unaffected of the river by whether gradient with year dummies.
(that is, the interaction of the 1970 share of dams in a state with in India). The inclusion of the interaction of dam construction controls for any differential river gradient with year dummies In particular, trends across regions with different river gradients. we control for the fact that the ease of nondam irrigation in different river with regions gradients may vary.16 our analysis is that The identifying assumption underlying
DAMS 623 of surrounding in their location (for instance, the productivity land). Our IV estimates capture the "local average agricultural treatment effect" of dams, or their effect in districts where dams were built because the districts had favorable river gradients and our which would, otherwise, not have received dams. Likewise, were dams that the effect of variable dam captures upstream the river gradient in the for a district because located upstream was In favorable. other district words, what we do not upstream in of the dams is effect specific districts for, say, placed capture our reasons. This that estimates are close to may imply political a "best case scenario," the economic since we are measuring impact of technologically this may differ from the the explicitly examining could shed some light on for this.17 that dams. We recognize appropriate of the dam. While impact "average" dams impact of, say, different-sized our not have instruments do suf this,
ficient power
IV. Results IV.A. Agriculture We start by examining how large dams have affected irri gated and cultivated area and agricultural production, both in the district where they are constructed and downstream. We have Part A of annual data for 271 districts for the years 1971-1999. III provides FGLS and Part B Table estimates (1), equation The "own feasible optimal IV estimates district" (3). equation coefficient captures the impact of dams built in that district while coefficient captures the impact of dams built in the "upstream" neighboring upstream districts (and we often refer to this as the In the last row, we report the F effect of dams). downstream statistic for the first-stage regression for the "own district" dams "dams predicted in own district"). The instru (using the variable ments appear sufficiently strong to avoid bias caused by weak
instruments.18
A first observation, and one to which we return below, is that errors on the estimated "own district" coefficient the standard coefficient. always exceed those of the "dams upstream"
17. Understanding the extent and impact of such heterogeneity is a very avenue for further work. Duflo and Pande estimates [2005] report OLS promising forwhether the effect of dams varies with dam height and find the poverty impact is the most pronounced for very large dams (more than 30 meters). 18. The first stage F-statistic in the upstream for dams districts (available from the authors) is larger.
TABLE III
Dams and Agriculture
eg
Area Inputs Fertilizer use
Gross
irrigated area
Gross cultivated area
Productio
Level
Log
Level -
Log
All
(2)
Dams Own district Upstream
(1) (3)
(4)
Part A FGLS ^
(5)
(6)
14.528
0.131
114.493
0.094
0.231
0.184 0.530
(13.300)
17.830
(0.156)
0.198
(47.838)
77.641
(0.059)
0.028
(0.342)
0.256
(0.334
(12.639)
(0.162)
(48.233)
(0.054)
(0.339)
(0.155
Part B. Feasible Optimal IV **$
Dams
Own district
232.092 (235.847) 49.754 (22.339) 8.48
0.728 (1.002) 0.328 (0.154) 4,536 4,536 8.48
325.358 (263.509) 58.602 (35.674) 4,522 8.51
0.875 (0.590) 0.088 (0.062) 4,522 8.51
0.563 (1.244) 0.169 (0.175) 4,521 8.5
0.085 (0.699) 0.341 (0.118) 7,078 9.22
Upstream N First stage .F-statistic (own district)
of the number ofpredicteddams in the sta state*year interactions,interaction Regressions include districtfixed effects, and districtelevation and rivergradient*year interactions(see notes toTable II fora fulldescriptionofgeographyvariables river gradient inupstre dams in the statewith (average) gradient,kilometers of river,area and elevation inupstream districts, a aremultiplied by 100. Standard errors,clusteredby district, the districthas any upstream districts.Regression coefficients are in logs. We use themonetized value of production for six major crops (described in notes to Table I). Yield is defin Nonwater-intensivecropsare sorghum (jowar),pearlmillet (bajra),andmaize, andwater-intensive cropsarewheat, rice, and s in 13 states (defined data cover 1971-1987 and productionand yield data 197 by 1961 census boundaries). Area and fertilizer The last row provides theF-statistics fromthe regressionof the number ofdams in the districton the predictednumber of
DAMS 625 for In columns (l)-(4) we report level and log estimates indicate a sig irrigated and cultivated area.19 The IV estimates in downstream in gross irrigated area dis nificant increase tricts?an additional dam increases irrigated area in the down stream district by roughly 0.33 percent or 497 hectares (this from the FGLS esti is statistically estimate indistinguishable mate). The effect on irrigation in own district is also positive. The is similar to the downstream effect in the FGLS point estimate IV and in the the larger specification. However, specification standard errors are also very large (below we suggest associated possible explanations). In columns (6) and (7) we consider the production and yield of the six major and (in logs). Both the FGLS crops in districts IV that estimates dam construction suggest significantly Optimal increases agricultural (0.34 percent) and yield (0.19 production in the downstream districts. In contrast, own district percent) estimates are smaller and insignificant. Case study evidence (and our model) suggests that dam irri causes to farmers substitute towards water-intensive gation crops. In columns (8) and (9) of Table III, we see that dams had an crop production, but insignificant impact on non-water-intensive crops in significantly increased the production ofwater-intensive downstream districts. We find similar sized, but much noisier in own district.20 estimates, for water-intensive crop production III provides a very consistent pattern for how dams Table in downstream affect agricultural districts: They sig production nificantly increase irrigated area and agricultural production, of water-intensive crops. In contrast, the effect on especially in own district is typically insignificant. production To the best of our knowledge, there are no other systematic estimates of the impact of dams on agricultural outcomes to
we can
which
two checks plausible. First, we
on whether can use
compare
our
estimates.
However,
we
can
the magnitude direct estimates
of estimated of the physical
provide
effects area
is
irri
a dam irrigates a certain acreage. 19. Dams may affect area in levels because If dam size and, hence, the associated in irrigated area is related to land increase then log acreage may be affected. In this case, we would productivity, expect dams to have a proportional, not level, effect on irrigated area. Since irrigated area and the average production typically have a log-log relationship, proportional impact of dams on area outcomes is also easier to compare with the average proportional impact on production. as water-intensive 20. We define sugarcane, rice, and wheat crops and sor as nonwater-intensive and maize crops. ghum, pearl millet,
626
in 1985. This estimate suggests that 36 irrigated 8,759 hectares area of the percent irrigated by dams would not have been irri otherwise. gated Second, ifwe assume that the sole effect of dams on produc tion downstream is through irrigation, then our instrument for a is also an instrument for area irrigated.21 dam upstream having This procedure gives us an elasticity of production with respect to dam-induced irrigation of 0.61 (standard error of 0.21, estimate to existing estimates, this is not reported to save space). Relative a plausible in the lower elasticity, though range.22 In Tables IV and V, we check that our produc Robustness. tion results are robust to alternative specifications.23 In Part A of Table IV, we show that the production effects of the average dam do not extend to neighboring districts that are not downstream. This suggests that our focus on the impact of dams on the districts where they are located and downstream In Part B, we examine whether changes in districts is reasonable. include agricultural production precede dam construction. We in own district and upstream dams that are currently under and will be completed in the next five years as construction
for alterna irrigation, in part, substitutes gated by dams. Dam tive forms of irrigation. This suggests our estimates of the addi tional area irrigated due to dams should be lower than the actual area irrigated by dams. Our point estimates suggest that dams increased area irrigated in each downstream district by 497 hect in own district. There are, on average, ares, and 2,321 hectares from a dam. This suggests 1.75 districts downstream that an additional 3,191 hectares are irrigated by a dam. This number can be compared to estimates of the average area irrigated by a estimated that a dam dam. The Indian Planning Commission
in the districts where 21. Dams may directly affect production they are built as an instrument we and therefore are inappropriate for area irrigated. However, would to mainly in downstream affect agricultural dis expect dams production tricts by increasing to existing esti irrigated area. Similar irrigation elasticity includes adjustments made mates, we identify a total derivative, which by farm ers in response to the availability of irrigation. 22. FAO reviews the irrigation literature for Asia and [1996], for instance, of crop yields with respect to production in the range of 1-4. report elasticities India's Planning Commission that "the yields on irrigated [1997, p. 474] assessed areas are generally two times higher than those for rain-fed areas." Crop-specific estimates increase in yield due to irrigation of percentage for India are very similar to our estimate et al. 2005], 73 percent for (75 percent for wheat [Mandal winter maize and for sugarcane [Mishra et al. 2001], and 63 percent [Ramesh 1998]). Mahadevaswamy 23. For brevity, we only report results for agricultural for Results production. are robust to these alternative outcomes other agricultural specifications.
DAMS 627 TABLE

The Reach of Dams:
IV
2SLS Estimates Headcount ratio
Agricultural production
Poverty gap
(1) (2) (3)

Dams Own district Part A. Neighboring 0.345 (1.148) 0.665 Upstream Downstream Neighboring but not upstream/ (0.220) -0.107 (0.260) -0.175 downstream (0.203) N 7,078 Part B. Time Effects Dams Own district Own district, completed next 5 in years 0.109 (1.266) 0.891 (3.376) 0.187 Upstream (0.323) -1.127 5 years (0.683) N 4,992 2.004 0.888 0.417 (0.444) 0.963 (1.213) (0.433) -0.156 -0.049 (0.154) 0.061 -0.025 (0.372) 1,443 (0.058) (0.137) 1,443 (0.194) districts 0.223 0.594 (0.306) -0.170 -0.045 (0.078) 0.073 0.022 (0.060) 0.013 -0.008 (0.072) 1,799 (0.101) (0.023) (0.021) (0.026) 1,799
Upstream,
completed
in next
Regressions include districtfixed effects,state*year interactions,interactionof predicted dams in the statewith districtgradient,kilometers ofriver, districtarea, and district elevation and river gradient*year interactions (see Table II fora description of the geographyvariables). They also includeriver gradient in upstream districts*yearinteractionsand interactionofpredicted dams in the statewith (average) gradient, kilometers of whether thedistricthas any river, area and elevation inupstream districts,and an indicatorfor upstream districts.Part B regressions also include the interactionsof the geographyvariables listed above with the number ofpredicted dams in the state at time t + 5. In Part B, dams completed innext 5 years are thenumber ofdams completed in the districtover thenext 5 years, and dams upstream completed innext 5 years is the number of dams completed in upstream districts in the next 5 years. Column (1) regression includes annual data for271 districts for 1971-1994, and columns (2) and (3) regressions include the years aremultiplied by 100. Standard errorsare inparentheses. 1973,1983,1987, and 1993.Regression coefficients They are clustered by district in column (1) and by 1973NSS region*year in columns (2) and (3).
additional regressors. We find no evidence that agricultural pro duction increases in the five years prior to dam completion. In Table V, we first include an array of additional controls. In column (1), we control for the interaction of initial poverty in the district (and in the district upstream) with predicted dam con struction in the state, and in column (2), we similarly control for initial tribal population. Our production estimates are unaltered. In column (3), we show that our results are robust to controlling for a linear time trend interacted with the state's share of total
TABLE V
The Impact of Dams:
Robustness
eg
Check
oo
Agricultural production Additional control Linear trend
Headcount ratio Additional Dams per 100 sq. km Initial poverty
Tribal Initial
poverty
population
(3)
(1) (2)
(4)
IFGLS fg
(5)
Part A. OLS Dams Own 0.168 district (0.310) 0.498 (0.137) 0.207 0.144 (0.308) 0.470 (0.154) Part B. 2SLSI Dams Own 0.082 district (0.769) 0.379 (0.115) 7,078 N -0.162 0.219 (1.729) 0.363 (0.163) 7,078
0.168 (0.310) 0.498 (0.137) Feasible Optimal IV ^
0.279 (0.074) -0.077 (0.039)
(0.299) Upstream 0.488 (0.140)
0.082 (0.769) 0.379 (0.115) 7,078
0.748 (0.308) -0.168 (0.068) 1,689
(0.676) Upstream 0.369 (0.115) 7,078
interaction of thenumber ofpredicteddams in the st All regressions includedistrictfixed effects, state*year interactions, and districtelevation and rivergradient*year interactions(see notes toTable II fora fulldescriptionofgeographyvariables area and elevation inupstream districts,rivergradient inupstre dams in the statewith (average) gradient,kilometersofriver, are as follows:(i) in columns (1) and (4) the separate interact the districthas any upstream districts.The additional controls of the headcount ratio in 1973 in own districtand inupstream districts,(ii) in columns (2) and (5) the separate interactions own in district and in and share (iii) in columns (3) and (6) 1970 state dam share*linear tim upstream districts, population are in and 100. columns in columns Coefficients districts. coefficients (4) an (5)-(7) (l)-(3) upstream multiplied by Regression FGLS (PartA) and Feasible Optimal IV (PartB) estimateswith standard errorsclustere parentheses. Columns (1H3) report (Part B) estimateswith standard errors clusteredbyNSS region*year.
DAMS629
dams in 1970. Finally, in column (4), we show that our results are the number of dams by district area. also robust to normalizing effects of Variance of Production. Why are the estimated in the dam's own district than in the dams so much noisier downstream district? One reason is that the own district effect combines the (presumably negative) catchment area effect with a in the effect part of the command area that (presumably positive) falls within the district. The combined effect is likely to vary across districts, depending, for instance, on dam size and its
location within the district. Such variability, in turn, would imply noisier estimates of the average effect. Another possible reason (which does not exclude the above) is that dams affect the variance of agricultural production differen districts. An important cause tially across own and downstream in crop production across Indian districts is of annual variations rainfall shocks. Dams are likely to reduce the sensitivity to rain fall shocks downstream, but to increase it in the upstream areas, due to restriction on water usage. or amplify In Table VI, we examine whether dams mediate the effect of a rain shock on agricultural production. We measure rain shock as the fractional deviation of annual rainfall from the district's historical average.24 In column (1), we see that, control ling for dam presence, a positive rain shock enhances agricultural
the mean in the average district. If risk production decreases with income, then this increase is likely to be particularly harmful to the poor. creasing aversion Other
re (2) shows that having a dam upstream production. Column duces the adverse effect of a negative rain shock: The coefficients on the upstream dam's rain shock interaction variable and the rain shock variable have opposite signs, and both coefficients are significant. In contrast, dams amplify the effect of a rain shock in own district. The coefficients on own district dam's rain shock interaction variable and the rain shock variable have the same the and interaction with dams is significant. sign, Our finding suggests that dams increase the variance of own in without in district, agricultural production significantly
with the predictions of a simple Inputs. Consistent we find that dams increase the agricultural production function, of water-intensive in downstream crops areas; two of production
24. We in modelling follow the literature is the main in India. problem more rainfall as beneficial, since
drought
TABLE VI
Dams and Rainfall Shocks
Agricultural production
Headcount ra
(1)
Rainshock
Dams
(2)
0.008 (0.044)
0.109
(3)
-0.041 (0.019)
0.765
0.065 (0.030)
-0.011
Dams*rainshock Upstream dams Upstream dams*rainshock JV
(1.227) 0.722 (0.197)
(1.228) 0.898
(0.324) -0.149 (0.067)
0.734 (0.195) -0.184 7,078
(0.364)
7,078
(0.092)
1,799
All columns report2SLS regressions. fixedeffects, interaction ofthenu Regressions includedistrict state*year interactions, kilometers of river,district area, and district elevation and river gradient*year interactions (see notes to Table II fora f interactionofthe number ofpredicteddams in the statewith (average) gradient, kilometersof river,area and elevation inup aremultiplied by 100. Stand and an indicatorfor All coefficients whether the districthas any upstream districts. interactions, district in columns (l)-(2), and byNSS region*year in columns (3)-(6). Rainshock is the fractionalannual deviation ofdistri is in We have annual for 271 districts the data and years 1971-1999, and povertyd Agricultural production logs. agricultural Missing district*yearobservations account foractual sample size.
DAMS 631 TABLE VII

Dams and Public Good Provision and Demographics: 2SLS Estimates
Water facility Power Tarmac road Population In-migrants
(1)
Dams Own districts 0.090
(2)
0.034
(3)
-0.070
(4) (5)
0.082 -0.293
(0.058)
Upstream -0.014
(0.347)
0.157
(0.240)
0.059
(0.574)
0.057
(0.797)
0.160
(0.011) 847
(0.093) 848
(0.065) 848
(0.059) 947
(0.104) 947
Regressions include district fixedeffects, state*year interactions,river gradient*year interactions,and interaction ofnumber dams in the statewith (i) districtgradientvariables, (ii)kilometers ofriverand district area, and (iii) elevation variables and rivergradient*year interactions (see Table II fora full description of the geography variables). They also include interactionof the number of dams in the statewith (average) gradient, kilometers of river, area and elevation in upstream districts, river gradient in upstream whether the districthas any upstream districts.Regression and an indicatorfor districts*yearinteractions, aremultiplied by 100, and robust standard errorsare inparentheses. The dependent variables in coefficients are with the specifiedpublic good facility. columns (l)-(3) the fractionofvillages in the district Water facility excludes dams. The dependent variables in columns (4)-(5) are in logsand referto the rural population. The unit of observation is the Indian district, and we have data for 1971, 1981 and 1991. Columns (1M3) regressionsuse data for302 districts,as definedby 1971 census, and columns (4)-(5) regressionsdata for339 districtsas definedby the 1961 census. Differences in sample size are accounted forbymissing district*year observations.
Columns that evidence (1) and (2) provide some, albeit weak, water nondam-related infrastructure decreased (the number of canals shows an insignificant increase) and that electrification increased in downstream of infrastruc districts. Other measures ture are unchanged dam construction column (see (3)) and by no in of other (or out) suggest government crowding inputs.
saw increased use of high yielding these crops (rice and wheat) (5) in Table III also shows that dams led to variety seeds. Column an insignificant increase in fertilizer use. The agricultural production function also suggests that dam forms of water irrigation should reduce the use of alternative infrastructure and increase electricity and fertilizer use (because canal irrigation requires electricity). In Table VII, we examine census the impact on other inputs using decennial data.25
25. Limited data availability for census outcomes less precise esti implies mates. We, instrument?the interaction of our therefore, use a more powerful river gradient in the state (the results with variables with actual dam incidence are qualitatively dam incidence the time Since predicted similar, but noisier). errors. For series is very short, we provide robust, rather than clustered, standard the same reason, we also use actual dam incidence for demographic outcomes and when the impact of institutions (Table VII) (Table IX). examining
632 IV.B. Rural
QUARTERLY JOURNAL OF ECONOMICS Welfare
we
a cost-benefit analysis our estimates of to undertake Using about the con dam construction requires additional assumptions costs of dams. We also need to make struction and operation benefits and costs, such as the unmeasured about assumptions value of insurance downstream, possible effects in other neigh are to detect (although, on average, too small that districts boring do not observe any such effects), and agents' optimizing
behavior.
the overall impact of dams, A different approach to assessing and one we take, is to examine how large dams have affected the is the main occupation of a welfare of the rural poor. Agriculture of rural and India's the poor, five-year plan documents majority the government's (which identify, on a five-year basis, policy an in aim of investment that show public explicit priorities)
is a panel of districts, rather than of individu (since our dataset estimated effect the of dam construction on poverty would be als, if dam construction induces either the relatively rich or biased to individuals poor migrate across district boundaries). relatively Columns (4) and (5) of Table VII report insignificant effects of dam construction on district census rural population outcomes, and in-migrants. A potential explanation both overall population inhibit labor mo is that imperfect credit and insurance markets bility by the poor in response to regional economic shocks. Our and case study evidence that findings are in line with anecdotal that suggests prefer to remain near their displaced populations [Thukral 1992]. Other studies also find very original habitats in response to relatively limited migration large district-level economic shocks.26 These findings also imply that it is reasonable to use a district-level panel on poverty outcomes to examine how dam construction affects rural welfare. That said, we compute
productivity, reduce insta irrigation is to increase agricultural the welfare of the rural in and enhance crop production, bility poor. For instance, the opening chapter of India's fifth five-year in view are removal of plan document begins, "The objectives of and achievement self-reliance." poverty We begin by examining the implications of dam construction in and This is an interesting question for district demographics. of itself and is also relevant for interpreting our poverty results
26.
See
for example
Topalova
[2004]
for the response
to globalization.
DAMS 633 bounds on how migration may have affected the estimated impact of dams on poverty [Manski 1990]. Our welfare results are in Table VIII; Part A provides OLS estimates and Part B 2SLS estimates.27 Column (1) shows that dams lead to an insignificant decline in mean per capita expen diture in the district where they are located and a marginally
in per capita expenditure in downstream significant increase districts. In column (2), we examine the most basic rural poverty the fraction of indicator: the head count ratio. This measures level below the official pov rural households with a consumption with a significant poverty in erty line. Each dam is associated crease of 0.77 percent in its own district. In contrast, poverty in downstream districts. Columns decreases (3) and (4) bound our
alternative about mi poverty results by making assumptions use VII the Table estimates for (we gration point in-migrants even though they are insignificant). The poverty results remain
robust.28
Since there are, on average, 1.75 districts downstream of each dam, the poverty reduction in downstream districts is insuf for the poverty increase in the dam's own ficient to compensate district. Another way of computing the overall poverty effect is to start with the observation that between 1973 and 1999 the aver age district had five dams built in own district and ten dams Our point estimate upstream. implies that this led to a 2.35 ? in count increase the head ratio (5*0.77 10*0.15). Over percent this time period, the head count ratio reduced by 23 percent, Indian district suggesting that poverty reduction in the average as as a have suffered much 10 due to dam setback may percent construction. An important caveat is that we have not accounted
27. The relatively short and very spaced time series implies that autocorre to be a problem. lation is unlikely In 1973 and 1983 poverty data is only available by NSS "region," an entity larger than the district, but below the state. To account we cluster standard errors by NSS for this aggregation Ifwe instead region*year. errors increase. However, cluster them by NSS results region, then our standard that are reported as significant at the 5 percent level remain significant at the 10 level or less (results available from the authors). percent 28. A simple back-of-the-envelope that the magnitude calculation of suggests our estimates is plausible. One of the channels dams increase through which on Dams World Commission that each [2000b] estimates poverty is displacement. dam displaced If 46 percent of them were already poor (this is the 31,340 persons. poverty rate in 1973), then 17,000 nonpoor Meanwhile, people were displaced. our 0.77 percent district population, estimate that given the average suggests an additional each dam made caveats 14,530 persons poor. Of course, apply: on are displaced estimates how many is debated; not all those displaced people will become poor, while dams may cause some nondisplaced to become individuals their land less productive), dams made and some nondis poor (if, for instance, placed poor people will escape poverty due to dams.
TABLE VIII
Dams Headcount Assume Per-capita expenditure Original and Rural ratio Assume
Welfare
05
poor
rich
Gini
in-migrants
in-migrants
Poverty ga
(3) Dams
Own district Upstream -0.289 0.273
(1) (2)
(4) 2]
(5)
PartA. OLS IFGLS

0.407 -0.079
(0.115)
0.093
(0.084)
-0.083
(0.083)
0.174 -0.082
0.081 -0.027 IV
&jj
(0.081) (0.038)
Optimal 0.651 0.297 g
(0.030)
(0.057)
(0.039)
(0.038)
(0.013)
Dams
Own district Upstream -0.457
Part B. 2SLSI Feasible 0.772 0.879
(0.467)
0.142
(0.324)
-0.154
-0.149
(0.314)
N
First stage F-statistic (own district) pj =^===================^
(0.084) 1,799
7.71
(0.068) 1,799
7.71 Cq
-0.150
(0.315)
(0.112)
-0.039
(0.066) 1,799
(0.066) 1,799 1,799

7.71 7.71
(0.020)
7.71 7.71
rivergradient*yearinteractions, and interaction ofnu Regressions includedistrict fixedeffects, state*year interactions, (ii)kilometers of riverand districtarea, and (iii) elevation variables and rivergradient*yearinteractions(see Table II fora fu ofthenumber ofdams in the statewith (average) gradient, interaction kilometersofriver,area and elevation inupstream dist and an indicatorfor whether the districthas any upstream districts. are multiplied by 100. Column Regression coefficients estimates. Standard errors,clusteredby 1,973NSS region*year, are in parentheses.The unit ofanalysis is the Indian distri and years 1973,1983,1987,1993, and 1999. Column (7) regressioncover 1971-1987 for271 districts (as definedby 1961 censu (defined A) and Feasible by 1981 census). In both columnswe reportFGLS (Part Optimal IV (PartB) estimateswith standa logarithms. 1)). Column (2) uses head count ratiofiguresas Agriculturalwage growth is definedas log (waged) waged theTable V in-migration estimates. Column (3) assumes in-migrants are poor,and column (4) in-migrants are rich. Mala using = (API) malaria/total population under surveillance). log (No. ofblood smears foundpositive for
DAMS635
for either
area.
state-wide,
or small,
effects that diffused
over a large
driver of poverty. in column dams affect (8), we examine whether Finally, health outcomes, as measured incidence be by annual malaria tween 1976 and 1995. We find no evidence that dams increased the district-level of waterborne here ma prevalence diseases, laria. This suggests that negative health effects did not drive the observed increase in poverty.
Column (6) examines within-district inequality as measured Gini Dam the coefficient. did not significantly construction by dis affect inequality, either in its own district or downstream tricts. This suggests the main unequalizing effect of dams was across, not within, districts. In column (7), we find that dams lead to significant in male increases wage agricultural growth in downstream districts. This is consistent with our poverty find are considered an important wages ings?male agricultural
an alternative measure Column of rural pov (5) examines measures This the the erty, poverty gap. depth of poverty? to how is much income needed specifically, bring the poor to a to In line with our line. level the equal poverty consumption earlier results, we find that dams significantly increase the pov erty gap in their own district, while reducing it downstream.
under construction. This is consistent with the fact that popula tion displacement in the vicinity of the reservoir occurs during the construction period. In columns (5)-(7) of Table V, we introduce additional con trols to check for differential trends across districts with varying initial conditions. Our poverty results are robust to controlling for the separate interaction of the predicted number of dams in the state with initial poverty and tribal population share (columns (5) and (6)). In column (7) we include the triple interaction of a linear
In Part A, Table IV, we show that, similar to the agricultural production, poverty effects of dam construction do not extend to neighboring districts that are not downstream. In Part B, we include dams in own district and upstream that will in the next five years as additional be completed If regressors. dam construction activity itself lowered poverty, then comparing poverty outcomes before and after dam construction may lead us to overestimate their negative effect. Dam construction activity does not reduce poverty; in fact, poverty increases while dams are Robustness.
636
time trend, the initial dam incidence in the state (as a fraction of in the in 1970), and the river gradient total dam construction is in this specification district. As with agricultural production, that the trend in economic tended to control for the possibility outcomes in a district which ismore suitable for dam construction differed from that in other districts in the same state and that this also differed in the state that built more dams. We find that, district hardly change, the while the results for the downstream standard error of the own district estimates more than doubles (to drops (to 0.25). The confidence 0.765), and the point estimates interval includes both the previous estimate and zero.29 A likely reason for this is the limited time series for poverty data; with only five years of data, the time pattern of dam construction from a linear trend, and across India is difficult to disentangle left. That said, this spec there is very little identifying variation ification provides a relevant caveat to the robustness of the own district poverty results. to displacement and the In addition Poverty and Rainfall. loss of productivity of the land around the reservoir, the increased in the district where a dam is built may of production variance to am and closed markets interact with low levels of migration increase shocks and 2006]. poverty [Jayachandran plify negative Given limited access to insurance against risk in rural India and and limited insurance [Morduch 1995; Rosenzweig options a is this and 1993; 1993], poten Wolpin Rosenzweig Binswanger can through which dam construction tially important channel of Table VI we examine increase poverty. In columns (3)-(6) whether dams amplify or reduce the poverty effect of rainfall shocks. In columns (3) and (5) we observe that, controlling for dams, positive rainfall shocks reduce both the head count ratio and the poverty gap. In columns (4) and (6), we examine whether dams alter the impact of rainfall shocks on poverty outcomes. Dams weakly reduce the impact of rainfall shocks on poverty in downstream districts, but significantly amplify it in the district true for the where the dam is constructed. This is particularly our accounts for which of poverty gap, poverty; finding is depth as it suggests dams have worsened out particularly worrisome comes for households whose welfare was the lowest to begin with.
29. We have count ratio and conclusions.
also estimated this regression for the bounded measures the poverty gap in Table VIII), with (columns (3)-(5)
of head similar
DAMS637
ness The inability or unwilling Institutions, Politics, and Dams. of those who benefit from dams to compensate groups of losers, or of the state to require them to do so when both groups are clearly identifiable ex-ante, suggests poorly functioning insti
examine whether the welfare consequences of dam construction with the of vary presence historically disadvantaged population groups and institutional quality. in India face significant socioeconomic Tribal populations case and that these disadvantages, study evidence suggests have faced dam-induced groups significant population displace ment. We, therefore, use the 1971 tribal population share in a district as our measure of the socioeconomic faced disadvantage by the district population. To measure institutional quality, we use district-level data on historic land tenure arrangements. During the colonial period, the British instituted different land revenue collection systems across districts. In some districts, an intermediary (usually a landlord) was given property rights for land and tax collection
tutions of redistribution. The absence of a statutory rehabilitation law, or even a national policy for rehabilitation, implies that state governments face no legal imperative and project authorities to undertake on Dams World rehabilitation Commission [2000b] de planning. scribes the rehabilitation policies of Indian states as "knee-jerk on reactions to the manifestations of disaffection of populations for public purposes." land which is acquired Such arbitrariness suggests that actual compensation may depend on the ability of affected population groups to organize themselves and lobby gov ernment and project authorities. To explore this possibility, we
In other districts, farmers were individually or responsibilities. collectively responsible for tax collection. In "landlord" districts, a class of landed gentry who had conflictual relationships with the and Iyer [2005] show that the ability peasants emerged. Banerjee of the population to organize themselves and obtain public goods exhibited marked differences across regions with different histor ical land tenure legacies. In landlord districts, class relations remain tense, rendering collective action more difficult. These
districts continue to have lower public good provision, agricul tural productivity, and higher infant mortality.30 If the popula tion groups affected by dam construction are better able to
30. Banerjee and Iyer [2005] determined whether characteristics, document that British not politics, a district was a landlord district. district
638
QUARTERLY JOURNAL OF ECONOMICS TABLE

Institutions and Dams: Agricultural production
IX
2SLS Estimates Headcount ratio Poverty gap
(1)
Dams-0.439
(2) (3)
1.072 0.332
(2.129)
Dams*nonlandlord Dams*tribal population Upstream Upstream share dams dams* -0.125
(0.539)
-0.639 -0.193
(0.178) (0.102)
0.087
(1.067) dummy
(2.637) 1.015
(0.309)
(0.790) (0.276) -0.393 -0.131
1.354 0.711
(0.708)
0.160 0.196
(0.293) (0.264)
-0.057 -0.015 (0.191)
(0.102)
0.061
Nonlandlord
dummy dams*tribal share
(0.705)
-0.609
(0.085)
Upstream population
AT
(0.701)
4,090
914
(0.060)
914
The regressions include districtfixed effects,state*year interactions,river gradient*year interactions, and interactionof (i) districtgradient variables, (ii) kilometers of river and district area, and (iii) elevation variables (see Table II fora fulldescriptionof thegeographyvariables) with number ofdams in the state (we use the predicted number in column (1) and actual number in columns (2) and (3)). They also include interaction of thenumber ofdams in the statewith (average) gradient,kilometers ofriver,area and elevation in upstream districts,an indicatorfor whether the districthas any upstream districts and rivergradient in upstream districts*year interactions.In column (1) we use predicted number of dams in a state. Standard errors in column (1) are clustered by districtand in columns (2) and (3) byNSS region*year.The sample is restrictedto the 151 districtsunder British direct rule. The nonlandlord dummy is a district-leveldummy was under nonlandlord arrangements forland taxation which equals one ifthemajority of land in the district purposes (from Banerjee and Iyer [2005]). The tribal population share is the fractionof districtpopulation are multiplied by 100. that belongs to a tribal group, as measured in 1971.All coefficients
and obtain compensation themselves in nonlandlord organize then the of dam construction should be districts, poverty impact muted in these districts. To examine this, we construct and use a nonlandlord dummy that equals one: The tax revenue system of the district was not landlord-based prior to independence. In Table IX, we report regressions, which include the sepa interactions of the dam's variable with the 1971 district tribal population share and the nonlandlord district dummy as additional (our instrument set is as before, plus the regressors interaction with the tribal and landlord variables). Our sample is restricted to districts that were under direct British rule. Column (1) shows that the impact of dams on agricultural did not differ across landlord and nonlandlord dis production tricts. Nor did it vary with the tribal population share. This rather than institutional factors de suggests that technological rate
DAMS 639 the productivity effect of a dam. Columns (2) and (3) poverty outcomes. The poverty impact of dams is inde share of the district. However, pendent of the tribal population historic land tenure arrangements significantly affect the poverty implications of dam construction. The effect of dams on poverty in own district is halved in nonlandlord districts, and we cannot that dams do not increase poverty in non reject the hypothesis landlord districts. We conjecture that in nonlandlord districts, the is more effective in organizing itself to demand com population pensation from the state. It is also possible that the absence of the landed gentry gives the displaced more political power in non termine consider landlord districts.31 These findings point to the relevance of the institutional framework within which public policies, such as dam construc tion, are executed and suggests that "weak institutions" or social conse conflict may help explain the claim that the distributional have been particularly adverse
quences
of dam construction countries. developing
in
V. Conclusion on infrastructure In 2000, public spending in developing 9 percent of government countries averaged spending, or 1.4 the magnitude of such spending, and a percent of GDP. Despite belief that infrastructure is integral to development, widespread on how investment in physical evidence infrastructure affects and individual wellbeing remains limited [World productivity Bank 1994]. In this paper, we have examined in the these questions context of large dam construction in India. We have argued that of large dams must address the fact that any credible evaluation dam placement is likely to be affected by regional wealth and the expected returns from dam construction in a region. This problem of endogenous in the evaluation arises of any large placement infrastructure project [Gramlich 1994]. While cross-country evi dence finds that productive enhances government spending are most to studies unable for unob control growth, convincingly served heterogeneity for et al. [1994] and (see, instance, Canning
31. At the state level, we found evidence in the fraction of that an increase reduced dam construction. This effect was, however, left-wing legislators only at the 10 percent level. A district's tribal population share and the significant other not did affect state-level dam support enjoyed political by parties construction.
640
and Ramirez Esfahani for infrastructure allows returns to infrastructure Large dams in India In contrast, those living
in the district where the dam is of living standards worsening for the poverty outcomes built; though limited data availability the poverty impact of dam limits our ability towholly disentangle time trends in poverty, which construction from district-specific are correlated with geographic suitability for dams.32 In Spring These findings have important policy implications. in grants and $270 million 2005, the World Bank announced for the Nam Theun 2 dam in Laos. The New York guarantees Times (June 5, 2005) quotes a senior World Bank official who justifies the return to dam lending as driven by the need to in a "practical" way since, support infrastructure development "You're never ever going to do one of these in which every single is good for me'" is going to say, This [Fountain 2005]. in statement is this the belief that projects with an Implicit as itwill be pos average positive return should be undertaken, sible to compensate the losers. We, however, find an unequal distribution of the costs and benefits associated with large dam in India and provide suggestive evidence that an construction reason areas for this is institutional In where quality. important the institutional structure favors the politically and economically advantaged, large dam construction is associated with a greater increase in poverty. Whether institutional quality could also ex plain the extent of dam construction across India is left for future person
research.
[2003]). Exploiting suitability geographic us to address this concern and estimate investment. have benefited downstream populations. in the vicinity of the dam fail to enjoy any agricultural productivity gains and suffer from increased volatil production. Our poverty results also suggest a ity of agricultural
Appendix
A. Dams
on dams is from the World Registry Data maintained Commission by the International
of Large Dams, on Large Dams
32. We are unable to examine effects of dam (state-level) market equilibrium the most with increased construction; likely such effect is a price effect associated If the poor are net sellers of agricultural [Deaton 1989], then production. products a decrease in food prices may have further accentuated to what poverty, relative we estimate.
DAMS 641 (ICOLD). A large dam is defined as a dam having a height of 15 meters from the foundation, or, if the height is between 5 and 15 meters, having a reservoir capacity ofmore than 3 million cubic meters. The registry lists all large dams in India, completed or under construction, together with the nearest city to the dam and date of completion. We use city information to assign dams to districts in the year of completion. The nine states in our sample Pun without dams are Arunachal Pradesh, Mizoram, Nagaland, Dadra Daman and and Diu, Delhi, Nagar Haveli, jab, Sikkim, and Pondicherry. and Delhi have dams in neighboring Punjab
upstream states.
B. Geography District overall gradient, area, river kilometers, elevation, and river gradient are collated from two GIS files, which provide information for India: GTOPOSO (elevation data, topographical at http://edcdaac.usgs.gov/gtopo30/gtopo30.html), and available at http://ortelius. "dnnet" (river drainage network data, available The files were processed by CIESIN, maproom.psu.edu/dcw/). Earth Institute Columbia University. GIS data exists formultiple cells in every district. District gradient and elevation was com puted as percent of district land area in different elevation/gra dient categories (summed across the cells in the district). For we river gradient, used the same process but restricted attention to cells through which the river flowed. We identified neighboring districts and, within them, upstream and downstream districts from district census maps. Data
C. Agriculture
crop-wise production, and male agricultural wages. We used the and McKinsey to update the primary sources used by Evenson district year observa production and yield data to 1999. Missing tions lead to variations in sample size. Kerala and Assam are the states. Also absent, but less impor major excluded agricultural are the minor states and Union Territories tant agriculturally, in northeastern Pradesh India, and the northern states ofHimachal and Jammu-Kashmir. We use the average 1960-1965 crop prices to obtain monetary and yield values. All monetary production
We use the Evenson India Agriculture and McKinsey and Climate dataset which (available at http://chd.ucla.edu/dev-data), covers 271 Indian districts across 13 Indian states, defined by 1961 boundaries This dataset provides and the years 1971-1987. information on gross area irrigated and cultivated, fertilizer use,
642
are deflated by the state-specific Consumer Price Index variables for agricultural laborers in Ozler and Ravallion [1996], base year 1973-1974.
D.
Rural We
Welfare Data
with the usual 30-day recall period) for household expenditures on most goods in the 1999-2000 round is believed to have led to an overestimate of the expenditures based on the 30-day recall across surveys Topalova [2004] period. To achieve comparability follows Deaton
data collected by the Indian National (NSS). Survey Sample a district randomly within These surveys sample households (sample size is roughly 75,000 rural and 45,000 urban house holds).33 For 1973, NSS regional averages were obtained from Jain et al. [1988]. For all other years, Topalova [2004] computed district-wise statistics using the poverty lines proposed by Deaton [2003a, b].34 The introduction of a new 7-day recall period (along
use poverty estimates for 374 districts across 23 Indian 1981 boundaries. We have poverty estimates states, defined by and 1999-2000; for 1973, 1983-1984, 1987-1988, 1993-1994, household these are derived from all-India survey expenditure
and imputes, for 1999, the correct district per on a capita expenditure distribution from household expenditures subset of goods forwhich the new recall period questions were not introduced. The poverty, inequality, and mean per capita expen were derived from this distribution. diture measures District are identifiers available from 1987 onwards (in hard copy for 1993). For 1973 and 1983, we have NSS (a region estimates region is a group of neighboring districts forwhich the sample is of sufficiently large for the NSS to deem the data "representative" across censuses the region). We use the district matching and region to district matching provided inMurthi et al. [2001] and in Indian censuses to match regions to districts and account for changes.
district boundary
33. The NSS as it considers does not report district averages, organization the district sample size inadequate for reliable district poverty estimates. This we a does not affect us, since of districts and do report results for larger number not make district. any inference about a particular 34. Poverty lines were unavailable for the smaller states and union territo ries ofArunachal and Diu, Jammu and Kashmir, Pradesh, Goa, Daman Manipur, and Nicobar Mizoram, Sikkim, Tripura, Andaman Meghalaya, Nagaland, Islands, are Dadra and Haveli. Most Chandigarh, Pondicherry, Lakshwadweep, Nagar, no or we lack other data for them. For excluded dams because already they have we use those included, the neighboring states' poverty line.
DAMS 643 E. Population, Public Goods, and Landlord Data
and public goods data are from the District-level population Decennial 1971, 1981, and 1991. Public goods data exist Census, for 302 districts, defined by 1971 census boundaries. We aggre gate village data (also known as village directory data) to com pute the fraction of villages in the district with a particular public and Somanathan [2005]). Popula good (obtained from Banerjee tion data are available for 339 districts defined by 1961 census Indian District Database, boundaries http://www. (Maryland These data en bsos.umd.edu/socy/vanneman/districts/index.html). ter the regressions in logs. Finally, district-level data on colonial land tenure systems is from Banerjee and Iyer [2005] and is avail able for 151 districts that were under direct British rule.
F. Rainfall use the rainfall dataset, Terrestrial Air Temperature and and Annual Time Series (1950-1999), Precipitation: Monthly Version and Kanji Maturate 1.02, constructed by Cord J.Willows at the Center for Climatic Research, University of Delaware. The rainfall measure for a latitude-longitude node combines data from 20 nearby weather stations using an interpolation algorithm based on the spherical version of Shepard's distance-weighting method. We define a rainfall shock as the fractional deviation of over the district's rainfall from the district mean (computed We 1971-1999). G. Malaria Annual district-level malaria data for India is collected by the National Malaria Eradication It conducts a Program (NMEP). smears were fever surveillance in India. Blood collected campaign for a sample of fever cases in every Indian district. Our measure (API), where
incidence in the log of the annual parasite incidence API is defined as (number smears positive for P. under surveillance. District-wise annual faliciparum)/populdLtion data on API was collated from the following NMEP publications: in India, NMEP and Control of Malaria (i) Epidemiology 1996, and (ii) Malaria and its Control in India, Volumes 1-3, NMEP 1986.
Massachusetts Research, Harvard and Institute Center of Technology, National Bureau for Economic Policy Research Center for Economic Policy of Economic
of malaria
University
and
Research
644
QUARTERLY JOURNAL OF ECONOMICS References
Perfor and Economic and Lakshmi Iyer, "History, Institutions Abhijit, Banerjee, mance: in India," American Land Tenure The Legacy of Colonial Systems Economic 1190-1213. Review, XCV(4) (2005), of Public "The Political and Rohini Somanathan, Economy Banerjee, Abhijit, MIT from India," Mimeograph, Goods: Some Evidence (2005). and Sendhil "How Much Esther Mullainathan, Bertrand, Marianne, Duflo, Should We Trust Differences-in-Difference Journal Estimates," of Quarterly CXIX(l) Economics, (2004), 249-275. A Global and Large Dams: Asit, and Cecilia Biswas, Tortajada, "Development International Journal XVII(l) of Water Resources Development, Perspective," (2001), 9-21. Marianne and Roberto "Infrastructure and Perotti, David, Fay, Canning, in International in Growth Rates: Market Globalization Growth," Differences and L. Paganetto, eds. (New York: St. and Economic Areas, M. Baldassari Martin's Press, Inc., 1994). Cech, Thomas V., Principles of Water Resources: History, Development, Manage ment and Policy and Sons, 2003). (New York: John Wiley The Bankwide "Resettlement and Development: Review of Cernea, Michael M., World Bank Resettlement, 1986-1993," involving Projects Involuntary "Performance and A.V. Suresh Babu, A.K., V.V. Rao, M. Shanker, Chakraborti, of an Irrigation Project using Satellite Remote and Evaluation GIS Sensing India (2002). GPS," Map of Nagarjuna "Satellite S.T., and R. Vidhya, J., Harikishen, Inventory Chari, (1994). Area," Technical report, NRSA Sagar Command in "Satellite Monitoring of Salandi Command Area Chari, S.T., and R. Vidhya, Technical (1995). Orissa," report, NRSA in Thailand: A Non-Para "Rice Prices and Income Distribution Deaton, Angus, The Economic XCIX 1-37. metric Analysis," Journal, (1989), Indian Estimates for Economic and Political -, 1999-2000," Poverty "Adjusted 322-326. (2003a) Weekly, in India, Economic and Political "Prices and Poverty -, 1987-2000," Weekly, Claims and Counter Claims Common (New Delhi: Dhawan, B.D., The Big Dams: 1989). wealth, Indian Water Resource and -, Issues, Development for Irrigation: Critiques Reviews 1993). (New Delhi: Commonwealth, and Sanjay and the Nation: Dis Dreze, Jean, Meera Samson, Singh, The Dam in theNarmada and Resettlement (New Delhi: Oxford Uni placement Valley 1997). versity Press, and Rohini Pande, Duflo, Esther, "Dams," NBER 11711, 2005. Working Paper, and Maria Infrastructure and Economic Hadi, Esfahani, Ramirez, "Institutions, LXX Economics, Growth," Journal (2003), 443-477. of Development in the Green Climate Interactions Evenson, R., and James McKinsey, "Technology Economic Growth Center, Yale (1999). Revolution," and Food Security Food Summit, (Rome: World 1996). FAO, Agriculture "Unloved but Not Unbuilt," New York Times (June 5, 2005). Fountain, Henry, and N. Hildyard, The Social and Environmental Goldsmith, Edward, Effects of (United Kingdom: 1984). Large Dams Centre, Ecological Wadebridge (New York: Van Nostrand Golz, Alfred R. (ed), Handbook of Dam Engineering Reinhold 1977). Company, on Report of World "ICOLD Comment of Commission E.R., Gopalakrishnan, Dams," (2000), www.dams.org/report/reaction/incoldindia.htm. A Review Edward Investment: Journal M., "Infrastructure Gramlich, of Essay," Economic XXXII(3) Literature, (1994), 1176-1196. "Generalized Lease in Panel and Multilevel Inference Hansen, Christian, Square with Serial Correlation Models and Fixed Effects," Mimeograph/Chicago GSB on River Fisher and Gerd Marmulla, "The Influence of Dams Jackson, Donald, Fish and Fisheries: and Conflict ies," in Dams, Opportunities, Challenges Resolution. Gerd Marmulla, ed. (FAO, Fisheries 1994). Department, and S.D. Tendulkar, "Dimensions of Rural Poverty: An Jain, L.R., K. Sundaram,
(1996).
(2003b) 362-368.
(2006).
DAMS 645
Inter-Regional 2395-2408. Profile," Economic and Political Weekly, XXIII (1988),
in De Seema, "Selling Labor Low: Equilibrium Jayachandran, Wage Volatility Journal CXIV(3) (2006), 538-575. of Political Economy, veloping Countries," The World Bank and Irrigation (World Bank OED Jones, William, Study 14908, 1995). and Yohei Sato, "Land Degradation due to Hydro-salinity Khan, Nasir Mahmood, Univer in Semi-arid Regions using GIS and Remote Sensing," Mimeograph, (2000). sity of Tokyo K.K. and M. Mohanty, Mandal, K.G., K.M. Hati, A.K. Misra, Bandyopadhyay, Effects on Growth and Water Yield Relationship of and Nutrient "Irrigation and Wheat (Triticum aestivum L.) in Central India," Journal of Agronomy CXCI (2005), 416-425. Crop Science, on Treatment Charles Bounds F., "Nonparametric Effects," American Manski, Economic LXXX(2) Review, (1990), 319-323. Rivers: The Ecology and Politics (Lon Patrick, Silenced of Large Dams McCully, don: Zed Books, 2001). on Poverty and Inequality: "The Impact of Dam Construction Ouarda, Merrouche, Evidence from Indian States, Institute 1960-1992," Mimeograph, European Florence (2004). T.R. Rathore, and U.S. "Water-use of Irrigated H.S., Savita, Mishra, Efficiency Winter Maize under Cool Weather Conditions of India," Irrigation Science, XXI (2001), 27-33. and Consumption "Income Smoothing Journal Morduch, Jonathan, Smoothing," IX (1995), 103-114. of Economic Perspectives, P.V. Srinivasan, and S.V. Subramanian, Indian Census Murthi, Mamta, "Linking with National Economic and Political (2001), Sample Survey," Weekly 783-793. on Poverty "A Database B. Datt, and Martin and Ozler, Gaurav, Ravallion, in India," Mimeograph, Growth World Bank (1996). Ninth Five Year Plan of India, 1997). (Government Commission, Planning in India," International Journal Prasad, K., and P.K. Rao, "On Irrigation Pricing VII (1991) 274-280. of Water Resources Development, on "Effect of Formative Phase Ramesh, P., and M. Mahadevaswamy, Drought of Shoots, Shoot Mortality, Yield and Different Classes Cane Attributes, of Four Sugarcane Journal and Crop Sci Cultivars," Quality of Agronomy ence, LXXIV (1998), 138-147. Mark R., and Hans "Wealth, Weather Risk, and the Rosenzweig, Binswanger, of Agricultural and Profitability Economic Jour Investments," Composition nal, CIII (1993), 56-78. Mark I. Wolpin, "Credit Market Con R., and Kenneth Constraints, Rosenzweig, and the Accumulation of Durable in Production Assets sumption Smoothing, Low-Income in Bullocks Countries: Investments in India," Journal of Politi cal Economy, CI (1993), 223-244. "Health Aspects ofWater Resources in Environmental Sharma, V.P., Projects," Central of Irrigation Board of Water Resources Impact Assessment Projects, and Power, eds. (New Delhi: Discovery 1991). Publications, Water Wars: Pollution and Profit (Pluto Press, Privatization, Shiva, Vandana, 2002). theWaters: The Political in India Singh, Satyajit, Taming Economy of Large Dams (New Delhi: Oxford University Press, 2002). on Rehabilitation and Environmental TehriReport, "Report of Expert Committee Tehri Hydro-Electric II" (1997). Aspects: Project: Volume of Irrigation in India," World "Assessment of Commission Thakkar, Himanshu, Dams (2000). Rivers Rivers Thukral, Enakshi, Displaced of Sorrow, Big Dams, People: of (New Delhi: 1992). Change Sage, "Factor and of Trade Effects Liberalization: Petia, Topalova, Immobility Regional on Poverty and Inequality Evidence from Indian Districts," Mimeograph, MIT US of Agriculture, Soil Conservation Service, National Department Handbook of Dams, (World Commission 1971). and K. Venugopal, "GIS based Diagnostic R., M. Karmegam, Vidhya,
(2004).
Engineering Analysis of
646
at of Bhadra Area Performance Assessment Command Irrigation System India (2002). Level," Map Disaggregated Cliffs: Prentice-Hall, 1984). C.C., Hydropower Warnick, Engineering (Englewood in Applied Methods Ameri Econometrics," Jeffrey, "Cluster-Sample Wooldridge, can Economic 133-138. Review, XCIII (2003), 1994: Infrastructure World Bank, World Development for Development Report D.C.: World Bank, 1994). (Washington forWater Resources "India: World Bank Assistance -, (2002). Management," on Dams, and Development: A New Dams Framework World Commission for (London: EarthScan 2000a). Publications, Decision-Making on Dams, World Commission India's (London: India, Large Dams: Experience of Dams, World Commission 2000b).

Dams Duflo

Încărcat de

Informații document

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

Dams Duflo

Încărcat de

Drepturi de autor:

Formate disponibile

Dams Author(s): Esther Duflo and Rohini Pande Source: The Quarterly Journal of Economics, Vol. 122, No.

and Rohini Pande

QUARTERLY JOURNAL OF ECONOMICS

billionUS dollars [Thakkar 2000].

QUARTERLY JOURNAL OF ECONOMICS

606 the new

QUARTERLY JOURNAL OF ECONOMICS reservoir to power and the outlet

tribes, which make of those displaced

QUARTERLY JOURNAL OF ECONOMICS

QUARTERLY JOURNAL OF ECONOMICS

Dams Number ofdams in district

Gross cultivated area (in '000hectares) Gross irrigatedarea (in '000hectares)

1,887 (1,254) 366 486 (309) 0.997 (0.01) 0.81 (0.26)

Non-landlord districts (proportion)

Tribal population share in district in 1971

QUARTERLY JOURNAL OF ECONOMICS

k Figure Distribution of Dams across Data of Dams on the number I

k Figure The Distribution Data of Dams on the number II

= + m*+ <>*, (1) ylst Pi + PiA* + Pa^S + P4^ + PsZ? + v,-

QUARTERLY JOURNAL OF ECONOMICS

QUARTERLY JOURNAL OF ECONOMICS

DAMS 619 TABLE

number predicted in the state

(1) (2) (3)

for river gradient

[p-value] Geography controls

[0.000] Yes Yes

[0.053] Yes Yes 7743

QUARTERLY JOURNAL OF ECONOMICS

12. The use of actual [Duflo and Pande 2005].

in the state yields

from (2) to predict

per district Dist.

JtZjt, z% of the Optimal

QUARTERLY JOURNAL OF ECONOMICS

Area Inputs Fertilizer use

Gross cultivated area

Part B. Feasible Optimal IV **$

232.092 (235.847) 49.754 (22.339) 8.48

0.728 (1.002) 0.328 (0.154) 4,536 4,536 8.48

325.358 (263.509) 58.602 (35.674) 4,522 8.51

0.875 (0.590) 0.088 (0.062) 4,522 8.51

0.563 (1.244) 0.169 (0.175) 4,521 8.5

0.085 (0.699) 0.341 (0.118) 7,078 9.22

Upstream N First stage .F-statistic (own district)

two checks plausible. First, we

on whether can use

the magnitude direct estimates

of estimated of the physical

QUARTERLY JOURNAL OF ECONOMICS

DAMS 627 TABLE

(1) (2) (3)

Agricultural production Additional control Linear trend

Headcount ratio Additional Dams per 100 sq. km Initial poverty

0.168 (0.310) 0.498 (0.137) Feasible Optimal IV ^

0.279 (0.074) -0.077 (0.039)

(0.299) Upstream 0.488 (0.140)

0.082 (0.769) 0.379 (0.115) 7,078

0.748 (0.308) -0.168 (0.068) 1,689

(0.676) Upstream 0.369 (0.115) 7,078

Dams*rainshock Upstream dams Upstream dams*rainshock JV

(1.227) 0.722 (0.197)

(0.324) -0.149 (0.067)

0.734 (0.195) -0.184 7,078

= + m+ <>, (1) ylst Pi + PiA* + Pa^S + P4^ + PsZ? + v,-

Damsrainshock Upstream dams Upstream damsrainshock JV