614 quarterly journal of economics duflo and pande...
TRANSCRIPT
productivity are also likely to make more of these investments,implying a spurious positive relationship between poverty reduc-tion and agricultural growth and dam building. Indeed, Mer-rouche [2004] finds larger poverty reductions for states that builtmore dams, but these findings cannot be given a causalinterpretation.
Our identification strategy, therefore, relies on within-statedifferences in dam construction, specifically differences acrossdistricts in a state. We can, therefore, examine spillover effectsfrom dams in neighboring districts but not state-wide economiceffects of dam construction, such as their effect on prices deter-
FIGURE IDistribution of Dams across Indian Districts, 1970
Data on the number of completed dams per district is from the World Registryof Dams published by the International Commission on Large Dams (ICOLD). SeeAppendix for further details.
614 QUARTERLY JOURNAL OF ECONOMICS
mined at the state level. We discuss the possible direction of suchstate-wide effects as we interpret our results.
Consider the following regression:
(1) yist � �1 � �2Dist � �3DistU � �4Zit � �5Zit
U � �i � �st � ist,
where Dist denotes the number of dams in the district and DistU the
number of dams located upstream from district i. �i is a districtfixed effect, �st is a state-year interaction effect and ist a dis-trict-year specific error term. Zit and Zit
U are a set of time varyingcontrol variables for the district and for upstream districts (thelist of relevant control variables is discussed below).
FIGURE IIThe Distribution of Dams across Indian Districts, 1999
Data on the number of completed dams per district is from the World Registryof Dams published by the International Commission on Large Dams (ICOLD). SeeAppendix for further details.
615DAMS
while most of western India, which has seen the maximum damconstruction, 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 anddistrict river gradient. Our regressions include state fixed effects,district elevation, river length, and district gradient as controls.The omitted river gradient category is the proportion of river in
FIGURE IVAverage River Gradient (in Percentage), by District
The average river gradient is computed using a digital elevation map whichprovides information on surface elevation. We have elevation data for multiplecells per Indian district. To compute river gradient we restrict attention to cellsthrough which a river passes. River gradient is a measure of the steepness of theground surface in the vicinity of district rivers and is defined as the tangent of thesurface. See Appendix for further details.
618 QUARTERLY JOURNAL OF ECONOMICS
the flat gradient category (0–1.5 percent). Our results confirm theimportance of engineering considerations: A gentle river gradient(1.5–3 percent) increases the number of dams, while a steepgradient reduces it. However, a very steep river gradient (morethan 6 percent) increases dam construction. The last effect isattributable to some multipurpose 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 predominantlyhydroelectric multipurpose dams. Excluding the fraction of area in very steepgradient from the instrument set provides qualitatively identical results.
TABLE IIGEOGRAPHY AND DAM CONSTRUCTION
Number of dams
Cross-section(1999) Poverty sample Production sample
Notinteracted
Interacted with predicted number ofdams in the state
(1) (2) (3)
Fraction river gradient1.5–3%
0.278(0.122)
0.153(0.040)
0.176(0.094)
Fraction river gradient3–6%
�0.210(0.127)
�0.191(0.065)
�0.219(0.128)
Fraction river gradientabove 6%
0.014(0.033)
0.075(0.031)
0.097(0.043)
F-test for river gradient 1.764 6.372 7.68[p-value] [0.15] [0.000] [0.053]Geography controls Yes Yes YesState*year and river
gradient*yearinteractions No Yes Yes
Fixed effects State District DistrictN 374 1855 7743
Geography controls are river length (in kilometers), district area (in sq. kms), three elevation variables(fraction district area with elevation 250–500 m, 500–1,000 m, and above 1,000 m, respectively), three districtgradient variables (fraction district area with gradient 1.5–3%, 3–6%, and above 6%, respectively). Columns (2)and (3) regressions also include a full set of state*year interactions and river gradient*year interactions. Columns(2) and (3) regressions include a full set of state*year interactions and river gradient*year interactions. For theseregressions the geography controls and the river gradient measures are all interacted with predicted dams. Thecolumn (2) regression is estimated by OLS with standard errors clustered by NSS region*year and the column (3)regression by GLS with standard errors clustered by district. Regression coefficients are multiplied by 100.Standard errors are in parentheses. In columns (1) and (2) the sample includes 374 districts as defined by 1981census. The poverty sample includes the years 1973, 1983, 1987, 1993, and 1999. The production sample includesannual data for 271 Indian districts (using the 1961 Indian census definition) for the years 1971–1999. Missingdistrict-year observations account for actual sample size.
619DAMS
TA
BL
EII
ID
AM
SA
ND
AG
RIC
UL
TU
RE
Are
aIn
puts
Agr
icu
ltu
ral
prod
uct
ion
Gro
ssir
riga
ted
area
Gro
sscu
ltiv
ated
area
Fer
tili
zer
use
Pro
duct
ion
Yie
ldP
rodu
ctio
n
Lev
elL
ogL
evel
Log
All
crop
sW
ater
-in
ten
sive
crop
sN
on–w
ater
-in
ten
sive
crop
s
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
Par
tA
.FG
LS
Dam
sO
wn
dist
rict
14.5
28(1
3.30
0)0.
131
(0.1
56)
114.
493
(47.
838)
0.09
4(0
.059
)0.
231
(0.3
42)
0.18
4(0
.334
)0.
152
(0.1
96)
0.06
3(0
.334
)0.
640
(0.5
85)
Ups
trea
m17
.830
(12.
639)
0.19
8(0
.162
)77
.641
(48.
233)
0.02
8(0
.054
)0.
256
(0.3
39)
0.53
0(0
.155
)0.
227
(0.1
41)
0.56
9(0
.243
)0.
801
(0.3
07)
Par
tB
.Fea
sibl
eO
ptim
alIV
Dam
sO
wn
dist
rict
232.
092
(235
.847
)0.
728
(1.0
02)
325.
358
(263
.509
)0.
875
(0.5
90)
0.56
3(1
.244
)0.
085
(0.6
99)
�0.
033
(0.4
51)
0.36
6(0
.782
)�
0.10
5(1
.349
)U
pstr
eam
49.7
54(2
2.33
9)0.
328
(0.1
54)
58.6
02(3
5.67
4)0.
088
(0.0
62)
0.16
9(0
.175
)0.
341
(0.1
18)
0.19
3(0
.097
)0.
470
(0.1
54)
0.18
1(0
.307
)N
4,53
64,
536
4,52
24,
522
4,52
17,
078
7,07
77,
143
6,78
6F
irst
stag
eF
-sta
tist
ic(o
wn
dist
rict
)8.
488.
488.
518.
518.
59.
229.
229.
039.
14
Reg
ress
ion
sin
clu
dedi
stri
ctfi
xed
effe
cts,
stat
e*ye
arin
tera
ctio
ns,
inte
ract
ion
ofth
en
um
ber
ofpr
edic
ted
dam
sin
the
stat
ew
ith
dist
rict
grad
ien
t,ki
lom
eter
sof
rive
r,di
stri
ctar
ea,
and
dist
rict
elev
atio
nan
dri
ver
grad
ien
t*ye
arin
tera
ctio
ns
(see
not
esto
Tab
leII
for
afu
llde
scri
ptio
nof
geog
raph
yva
riab
les)
.Th
eyal
soin
clu
dein
tera
ctio
nof
the
nu
mbe
rof
pred
icte
dda
ms
inth
est
ate
wit
h(a
vera
ge)g
radi
ent,
kilo
met
ers
ofri
ver,
area
and
elev
atio
nin
ups
trea
mdi
stri
cts,
rive
rgr
adie
nt
inu
pstr
eam
dist
rict
s*ye
arin
tera
ctio
n,a
nd
anin
dica
tor
for
wh
eth
erth
edi
stri
cth
asan
yu
pstr
eam
dist
rict
s.R
egre
ssio
nco
effi
cien
tsar
em
ult
ipli
edby
100.
Sta
nda
rder
rors
,clu
ster
edby
dist
rict
,are
repo
rted
inpa
ren
thes
es.P
rodu
ctio
nan
dyi
eld
vari
able
sar
ein
logs
.W
eu
seth
em
onet
ized
valu
eof
prod
uct
ion
for
six
maj
orcr
ops
(des
crib
edin
not
esto
Tab
leI)
.Y
ield
isde
fin
edas
crop
prod
uct
ion
per
un
itof
lan
d(R
s.pe
rh
ecta
re).
Non
wat
er-i
nte
nsi
vecr
ops
are
sorg
hu
m(j
owar
),pe
arlm
ille
t(b
ajra
),an
dm
aize
,an
dw
ater
-in
ten
sive
crop
sar
ew
hea
t,ri
ce,a
nd
suga
rcan
e.T
he
sam
ple
incl
ude
san
nu
alda
tafo
r27
1di
stri
cts
in13
stat
es(d
efin
edby
1961
cen
sus
bou
nda
ries
).A
rea
and
fert
iliz
erda
taco
ver
1971
–198
7an
dpr
odu
ctio
nan
dyi
eld
data
1971
–199
9.D
evia
tion
sin
sam
ple
size
are
due
tom
issi
ng
data
.T
he
last
row
prov
ides
the
F-s
tati
stic
sfr
omth
ere
gres
sion
ofth
en
um
ber
ofda
ms
inth
edi
stri
cton
the
pred
icte
dn
um
ber
ofda
ms
inth
eow
ndi
stri
ct.
624 QUARTERLY JOURNAL OF ECONOMICS
TA
BL
EV
ID
AM
SA
ND
RA
INF
AL
LS
HO
CK
S
Agr
icu
ltu
ral
prod
uct
ion
Hea
dcou
nt
rati
oP
over
tyga
p
(1)
(2)
(3)
(4)
(5)
(6)
Rai
nsh
ock
0.06
5(0
.030
)0.
008
(0.0
44)
�0.
041
(0.0
19)
�0.
036
(0.0
26)
�0.
012
(0.0
07)
�0.
0001
(0.0
10)
Dam
s�
0.01
1(1
.227
)0.
109
(1.2
28)
0.76
5(0
.324
)0.
713
(0.3
20)
0.29
4(0
.111
)0.
255
(0.1
08)
Dam
s*ra
insh
ock
0.89
8(0
.364
)�
0.24
3(0
.191
)�
0.20
3(0
.088
)U
pstr
eam
dam
s0.
722
(0.1
97)
0.73
4(0
.195
)�
0.14
9(0
.067
)�
0.13
0(0
.068
)�
0.03
8(0
.019
)�
0.03
0(0
.019
)U
pstr
eam
dam
s*ra
insh
ock
�0.
184
(0.0
92)
0.10
9(0
.059
)0.
034
(0.0
22)
N7,
078
7,07
81,
799
1,79
91,
799
1,79
9
All
colu
mn
sre
port
2SL
Sre
gres
sion
s.R
egre
ssio
ns
incl
ude
dist
rict
fixe
def
fect
s,st
ate*
year
inte
ract
ion
s,in
tera
ctio
nof
the
nu
mbe
rof
pred
icte
dda
ms
inth
est
ate
wit
hdi
stri
ctgr
adie
nt,
kilo
met
ers
ofri
ver,
dist
rict
area
,an
ddi
stri
ctel
evat
ion
and
rive
rgr
adie
nt*
year
inte
ract
ion
s(s
een
otes
toT
able
IIfo
ra
full
desc
ript
ion
ofge
ogra
phy
vari
able
s).
Th
eyal
soin
clu
dein
tera
ctio
nof
the
nu
mbe
rof
pred
icte
dda
ms
inth
est
ate
wit
h(a
vera
ge)
grad
ien
t,ki
lom
eter
sof
rive
r,ar
eaan
del
evat
ion
inu
pstr
eam
dist
rict
s,ri
ver
grad
ien
tin
ups
trea
mdi
stri
cts*
year
inte
ract
ion
s,an
dan
indi
cato
rfo
rw
het
her
the
dist
rict
has
any
ups
trea
mdi
stri
cts.
All
coef
fici
ents
are
mu
ltip
lied
by10
0.S
tan
dard
erro
rsre
port
edin
pare
nth
eses
.Th
ese
are
clu
ster
edby
dist
rict
inco
lum
ns
(1)–
(2),
and
byN
SS
regi
on*y
ear
inco
lum
ns
(3)–
(6).
Rai
nsh
ock
isth
efr
acti
onal
ann
ual
devi
atio
nof
dist
rict
rain
fall
from
its
his
tori
cm
ean
(defi
ned
over
1971
–199
9).
Agr
icu
ltu
ralp
rodu
ctio
nis
inlo
gs.W
eh
ave
ann
ual
agri
cult
ura
ldat
afo
r27
1di
stri
cts
and
the
year
s19
71–1
999,
and
pove
rty
data
for
374
dist
rict
san
d19
73,1
983,
1987
,199
3,an
d19
99.
Mis
sin
gdi
stri
ct*y
ear
obse
rvat
ion
sac
cou
nt
for
actu
alsa
mpl
esi
ze.
630 QUARTERLY JOURNAL OF ECONOMICS
TA
BL
EV
III
DA
MS
AN
DR
UR
AL
WE
LF
AR
E
Hea
dcou
nt
rati
o
Per
-cap
ita
expe
ndi
ture
Ori
gin
alA
ssu
me
poor
in-m
igra
nts
Ass
um
eri
chin
-mig
ran
tsP
over
tyga
pG
ini
coef
fici
ent
Agr
icu
ltu
ral
wag
egr
owth
Mal
aria
inci
den
ce
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Par
tA
.O
LS
/F
GL
SD
ams
Ow
ndi
stri
ct�
0.28
9(0
.115
)0.
273
(0.0
84)
0.40
7(0
.083
)0.
174
(0.0
81)
0.08
1(0
.030
)0.
014
(0.0
27)
0.01
8(0
.043
)0.
60(0
.89)
Ups
trea
m0.
093
(0.0
57)
�0.
083
(0.0
39)
�0.
079
(0.0
38)
�0.
082
(0.0
38)
�0.
027
(0.0
13)
0.00
7(0
.013
)0.
057
(0.0
20)
0.20
(0.6
5)
Par
tB
.2S
LS
/F
easi
ble
Opt
imal
IVD
ams
Ow
ndi
stri
ct�
0.45
7(0
.467
)0.
772
(0.3
24)
0.87
9(0
.314
)0.
651
(0.3
15)
0.29
7(0
.112
)0.
104
(0.1
39)
�0.
045
(0.2
42)
�0.
40(2
.02)
Ups
trea
m0.
142
(0.0
84)
�0.
154
(0.0
68)
�0.
149
(0.0
66)
�0.
150
(0.0
66)
�0.
039
(0.0
20)
0.00
0(0
.020
)0.
069
(0.0
31)
0.07
(0.5
2)N
1,79
91,
799
1,79
91,
799
1,79
91,
794
4,27
27,
001
Fir
stst
age
F-s
tati
stic
(ow
ndi
stri
ct)
7.71
7.71
7.71
7.71
7.71
7.71
8.71
10.9
5
Reg
ress
ion
sin
clu
dedi
stri
ctfi
xed
effe
cts,
stat
e*ye
arin
tera
ctio
ns,
rive
rgr
adie
nt*
year
inte
ract
ion
s,an
din
tera
ctio
nof
nu
mbe
rda
ms
inth
est
ate
wit
h(i
)di
stri
ctgr
adie
nt
vari
able
s,(i
i)ki
lom
eter
sof
rive
ran
ddi
stri
ctar
ea,a
nd
(iii
)ele
vati
onva
riab
les
and
rive
rgr
adie
nt*
year
inte
ract
ion
s(s
eeT
able
IIfo
ra
full
desc
ript
ion
ofth
ege
ogra
phy
vari
able
s).T
hey
also
incl
ude
inte
ract
ion
ofth
en
um
ber
ofda
ms
inth
est
ate
wit
h(a
vera
ge)g
radi
ent,
kilo
met
ers
ofri
ver,
area
and
elev
atio
nin
ups
trea
mdi
stri
cts,
rive
rgr
adie
nt
inu
pstr
eam
dist
rict
s*ye
arin
tera
ctio
ns,
and
anin
dica
tor
for
wh
eth
erth
edi
stri
cth
asan
yu
pstr
eam
dist
rict
s.R
egre
ssio
nco
effi
cien
tsar
em
ult
ipli
edby
100.
Col
um
ns
(1)–
(6)
regr
essi
ons
repo
rtO
LS
(Par
tA
)an
dIV
(Par
tB
)es
tim
ates
.Sta
nda
rder
rors
,clu
ster
edby
1,97
3N
SS
regi
on*y
ear,
are
inpa
ren
thes
es.T
he
un
itof
anal
ysis
isth
eIn
dian
dist
rict
asde
fin
edby
1981
cen
sus,
we
hav
eda
tafo
r37
4di
stri
cts
and
year
s19
73,1
983,
1987
,199
3,an
d19
99.C
olu
mn
(7)r
egre
ssio
nco
ver
1971
–198
7fo
r27
1di
stri
cts
(as
defi
ned
by19
61ce
nsu
s).C
olu
mn
(8)r
egre
ssio
nco
vers
1975
–199
5fo
r41
3di
stri
cts
(defi
ned
by19
81ce
nsu
s).I
nbo
thco
lum
ns
we
repo
rtF
GL
S(P
art
A)
and
Fea
sibl
eO
ptim
alIV
(Par
tB
)es
tim
ates
wit
hst
anda
rder
rors
clu
ster
edby
dist
rict
.Per
-cap
ita
expe
ndi
ture
isin
loga
rith
ms.
Agr
icu
ltu
ral
wag
egr
owth
isde
fin
edas
log
(wag
e(t)
�w
age(
t�
1)).
Col
um
n(2
)u
ses
hea
dco
un
tra
tio
figu
res
asco
mpu
ted
from
NS
Sda
ta.C
olu
mn
s(3
)an
d(4
)ad
just
this
usi
ng
the
Tab
leV
in-m
igra
tion
esti
mat
es.C
olu
mn
(3)a
ssu
mes
in-m
igra
nts
are
poor
,an
dco
lum
n(4
)in
-mig
ran
tsar
eri
ch.M
alar
iain
cide
nce
ism
easu
red
aslo
gan
nu
alP
aras
ite
Inci
den
ce(A
PI)
�lo
g(N
o.of
bloo
dsm
ears
fou
nd
posi
tive
for
mal
aria
/tot
alpo
pula
tion
un
der
surv
eill
ance
).
634 QUARTERLY JOURNAL OF ECONOMICS
organize themselves and obtain compensation in nonlandlorddistricts, then the poverty impact of dam construction should bemuted in these districts. To examine this, we construct and use anonlandlord dummy that equals one: The tax revenue system ofthe district was not landlord-based prior to independence.
In Table IX, we report regressions, which include the sepa-rate interactions of the dam’s variable with the 1971 districttribal population share and the nonlandlord district dummy asadditional regressors (our instrument set is as before, plus theinteraction with the tribal and landlord variables). Our sample isrestricted to districts that were under direct British rule.
Column (1) shows that the impact of dams on agriculturalproduction did not differ across landlord and nonlandlord dis-tricts. Nor did it vary with the tribal population share. Thissuggests that technological rather than institutional factors de-
TABLE IXINSTITUTIONS AND DAMS: 2SLS ESTIMATES
Agriculturalproduction
Headcountratio Poverty gap
(1) (2) (3)
Dams �0.439(2.129)
1.072(0.539)
0.332(0.178)
Dams*nonlandlorddummy
�0.125(1.067)
�0.639(0.309)
�0.193(0.102)
Dams*tribalpopulation share
1.354(2.637)
0.711(0.790)
0.087(0.276)
Upstream dams 1.015(0.708)
�0.393(0.293)
�0.131(0.102)
Upstream dams*Nonlandlorddummy
0.160(0.705)
0.196(0.264)
0.061(0.085)
Upstream dams*tribalpopulation share
�0.609(0.701)
�0.057(0.191)
�0.015(0.060)
N 4,090 914 914
The regressions include district fixed effects, state*year interactions, river gradient*year interactions,and interaction of (i) district gradient variables, (ii) kilometers of river and district area, and (iii) elevationvariables (see Table II for a full description of the geography variables) with number of dams in the state (weuse the predicted number in column (1) and actual number in columns (2) and (3)). They also includeinteraction of the number of dams in the state with (average) gradient, kilometers of river, area and elevationin upstream districts, an indicator for whether the district has any upstream districts and river gradient inupstream districts*year interactions. In column (1) we use predicted number of dams in a state. Standarderrors in column (1) are clustered by district and in columns (2) and (3) by NSS region*year. The sample isrestricted to the 151 districts under British direct rule. The nonlandlord dummy is a district-level dummywhich equals one if the majority of land in the district was under nonlandlord arrangements for land taxationpurposes (from Banerjee and Iyer [2005]). The tribal population share is the fraction of district populationthat belongs to a tribal group, as measured in 1971. All coefficients are multiplied by 100.
638 QUARTERLY JOURNAL OF ECONOMICS