Stockholm Doctoral Course Program in Economics
Development Economics I — Lecture 8
Infrastructure
Masayuki Kudamatsu IIES, Stockholm University
Big question in this lecture
Does
infrastructure
promote
It’s NOT easy to empirically identify the impact of infrastructure
• Endogenous placement of infrastructure
• Infrastructure may be a response to rising economic opportunities
(reverse causality)
• Govt. may target poor areas to improve their economic conditions
Evaluation of infrastructure: emerging field in development
• Dams (Duflo & Pande 2007 QJE) • Mobile phones (Jensen 2007 QJE)
• Excellent example of when DID works best
• Electricity (Dinkelman 2008) • Careful analysis on heterogenous
treatment effect
1. Dams
• Duflo and Pande (2007, QJE) • Using geography as instrument to
1-1 Research questions
• What’s the impact of irrigation dams on agricultural production and rural poverty?
• What’s the distributional
1-2 Data
• Annual agricultural production, 1971-1999, for 271 districts
• Poverty data in 1973, 83, 87, 93, 99 for 374 districts
Data (cont.)
Ratio to district area of river area with gradient more than 6%, 3-6%, 1.5-3%
• Data source: GTOPO30 (elevation at 30-arc second grid space) & Digital Chart of World (river drainage network)
• Identify GTOPO30 cells where rivers flow
• RGrki: fraction of river areas with
gradient falling in category k
• k: 2 for 1.5 to 3%; 3 for 3-6%; 4 for above 6%
• River flowing at some gradient: ideal for irrigation dams
• Very steep river: ideal for power generation dam
• D¯st: # of dams in India in year t
(Figure III) multiplied by fraction of dams in state s in 1970
• Why not the actual # of dams in state
• νi: district FE
• µst: state-year FE (different trends
across states)
• Mi: area, elevation, overall gradient,
river length
• Why (Mi ∗D¯st) included?
• lt: year dummies
1st stage results (Table II)
• Dictricts w/ more river gradient 1.5-3% or above 6%: more dams built
Empirical strategy: 2nd stage
yist = γi + ηst + δDist +δUDistU
+ZistδZ + ZUistδZU + εist
w/ Dˆist,DˆistU ,Zist,ZUist as instruments
• yist: outcome variable
• γi: district FE
• ηst: state-year FE
• Zist: vector of Mi ∗ D¯st,RGrki ∗lt
• ZUist: vector of Mi ∗ D¯st,RGrki ∗ lt for
upstream districts
• Dˆist,: fitted value for Dist
• DˆU
ist: the sum of fitted values for
Empirical strategy: Method
• Feasible optimal IV with S.E. robust to arbitrary covariance of the
residual w/i state (see ft. 15 for how to implement this)
• Why?
• Autocorrelation at state level
• Feasible GLS: more efficient than OLS with S.E. clustered
⇒ Small effect more likely to be detected (Power of test ր)
Digression: IV estimates &
heterogenous treatment effect
• IV estimates: treatment effect for compliers (“Local Average
Treatment Effect”)
cf. Angrist and Imbens (1994), Imbens (2007)
1-4 Results 1: Impact on
agriculture (Table III)
1 additional dam in upstream ⇒ • Irrigated areas ր by 0.33%
• Production/Yield of 6 major crops ր by 0.34/0.19%
• Production of water-intensive crops ր by 0.47%
Results 2: Interaction w/ rainfall
shocks (Table VI)
Rainfall shocks (deviation from 1971-99 mean) on agricultural production:
• Mitigated if dams built upstream • Amplified if dams built in own
districts
Results 3: Impact on rural
welfare (Table VIII)
Head count ratio:
• 0.77% pt ր by 1 more dam in own district
• 1.5% pt ց by 1 more dam upstream • No impact on district-level
population or in-migration (Table VII)
Results 4 (Table IX)
Impact of dams on poverty in own districts: mitigated if tax collection in colonial days done by farmers, not by landlords
cf. Banerjee and Iyer (2005): non-landlord districts ⇒ • public goods ր
• agricultural productivity ր
1-5 Taking Stock
• Use geography interacted with nation-wide trends & inter-state
variation in infrastructure-building to credibly estimate the impact of
infrastructure
2. Railroads
• Dolandson (2008)
2-1 Research questions
• Did the expansion of railroads in colonial India promote agricultural development?
2-2 Data
• Sample: 239 districts in colonial India
• Annual panel, 1861-1930 • Outputs & retail prices of 17
principal crops
• Bilateral trade flows for 85 commodities, 1880-1920
• Daily rainfall from 3614 stations, 1891-1930
2-3 Background
Transportation means in colonial India • Bullocks on roads (<20-30km/day) • River (65km/day downstream,
2-4 Model (Eaton-Kortum 2002)
• D districts, each denoted by d or o
• K commodities, each w/ a continuum of varieties
• Unit mass of identical agents in each district
• Each owns Ld units of land,
immobile & supplied inelastically • Land: only factor of production • Land rental rate rd
Model: Preference
Model: Preference (cont.)
• CES over varieties (j) of each k ⇒ Indirect utility per acre, Wd, is given
by (cf. equation (9) on p.15):
Model: Production
zdk(j): amount of variety of j of
commodity k produced by 1 unit of land in district d
• Follows type-II extreme value distribution
Fdk(z) = e−Akdz−
θ
k
• Akd: how likely productivity is high • θk: how variable productivity is
Model: Commodity market
• Many competitive firms w/i district ⇒ Each firm makes zero profit
Model: Trade
• To export 1 unit of k from district o
to d, Todk ≥ 1 units must be
produced in o (iceberg trade cost). • Todk ≤Tomk Tmdk
• Took = 1 (normalization)
• Railroads reduceTodk
⇒ Import price of k from o:
Model: Trade (cont.)
• Agents: indifferent about where each k(j) is made
⇒ They pay the cheapest pkod(j), denoted by pdk(j)
⇒ Its distribution is given by
Gkd(p) = 1 − e−
"
%D
o=1Ako(roTodk )−θk
$
Model: Trade (cont.)
commodity k varieties, denoted bypdk
Eaton-Kortum’s result no. 1
• Prob. for district d to import k(j)
from o: (see fn. 16 of Eaton-Kortum)
πodk = A
k
o(roTodk )−θk
%D
o=1Ako(roTodk )−θk
• πodk is also the fraction of varieties of
Eaton-Kortum’s result no. 2
• Price of a variety that district d
imports from o: distributed by Gdk(p)
• See ft. 17 of Eaton-Kortum
⇒ District d’s expenditure for imports from o: same across o for each k
⇒ πodk = Xodk /Xdk where
• Xodk : Trade flow from o to d for commodity k
Model: land market
• Land: inelastically supplied • If Ako UP or Todk DOWN
⇒ πodk UP & demand for land in o UP ⇒ Rental price ro should go up
• Land rental prices rd’s solve the
following system of equations
roLo =
!
k
!
d
2-5 Taking Model to Data
6 empirical steps to estimate the impact of railroads:
1. Trade costs 2. Trade flows
3. Market integration 4. Mean income
5. Income volatility
Prediction 6
• Indirect utility per acre for agents in
d, Wd, is given by • Trade costs & other districts’
Prediction 6 can be used for identifying the mechanism of the railroad impact on welfare
d as control (reduced-form
estimation)
dd as additional
regressor
• Extent to which coeff. on RAILdt
Testing Prediction 6
• Wd: real agricultural income per
acre
• Observed from each commodity’s yield per acre and price & land areas
• µk: k’s consumption share
Testing Prediction 6 (cont.)
• We need to estimate unobserved
Akd & πddk as functions of exogenous variables
• We also need to estimate θk
Step 1
• Estimate the trade cost Todk in the model
Step 1: Prediction 1
• Remember average price of commodity k in d is
pdk = λk1"
produced only in one district
Step 1: Specification
lnpdo = βoto +βdto +φoodt
+δ lnTC(Rt)odt + εo
odt
• Commodity o: salt produced only in a particular district
Step 1: Specification (cont.)
lnpdo = βoto +βdto +φoodt
+δ lnTC(Rt)odt + εo
odt
Prediction 1 tells us: • βoto = lnpoot
• δ lnTC(Rt)odt: time-variant
component of trade cost btw. o & d
Step 1: Measuring
TC
(
R
t)
odt LCR(Rt,α)odt: lowest-cost routedistance in railway-equiv. km
• α = (αroad, αriver, αcoast): trade cost
per km relative to railroad
• Existing transportation network + Rt
⇒ shortest-distance btw. o & d for each α
• α: estimated by NLS together with δ
Step 1: Results (Table 2)
• Railroads did reduce trade cost per km (αˆ > 1)
• More than reported relative freight rates (α = (4.5,3.0,2.25)) suggest
• Over & above linear trends
• Important as LCR(Rt,α)odt ↓ over time
• Elasticity of trade cost to distance in rail-equiv. km: 0.247
Step 2
• Check whether railroads increased trade flows
Step 2: Specification
lnXodtk = βotk + βdtk + βodk + φkodt
−θkδˆlnLCR(Rt,αˆ)odt + εk
odt
Prediction 2 suggests: • βotk = lnAko − θk lnro
• βdtk = ln%D
o=1Ako(roTodk )−θk +lnXdk
• Other terms: −θk lnTodk
Step 2: Specification (cont.)
• Estimate for each k to obtain θˆk’s
• S.E.: bootstrapped
• See Deaton (1997) for references on bootstrap
• Then we obtain lnAˆk
o = ˆβotk + ˆθk lnrot
where rot is measured by nominal
Step 2: Results (Table 3)
• −θkδˆ: significantly negative on
average (column 2)
⇒ Shorter railway-equiv. distance increased trade flow
Step 2b
Extract exogenous component in lnAˆk o:
• RAINotk : total rainfall between
sowing and harvest dates for k in o
• κ: 0.441 (se: 0.082)ˆ
Step 3: Prediction 3
• Check if railroads integrate markets • Remember
pdk = λk1"
D
!
o=1
Ako(roTodk )−θk
$−θ1
k
1. pdk depends less on Akd if Tod ↓
Step 3: Specification
Step 3: Results (Table 4)
• χ1ˆ = −0.402∗∗∗
• χˆ2 = +0.375∗∗: railroad link reduces the dependence of price on own district rainfall
• χˆ3 = −0.021: w/o railroad link, neighboring districts’ rainfall does not affect price
• χˆ4 = −0.082∗∗∗
Step 3b: Model evaluation
• If model correct, predicted commodity price pˆdtk should be close to observed pdtk
• Solving the model & plugging
estimated parameters & observed exogenous variables to obtain pˆdtk
Step 3b: Model evaluation (cont.)
• Then estimate
lnpkdt = βdk + βtk +βdt +ωlnpˆdtk + εkdt
Step 4: Prediction 4
• Solve system of equations (6) to obtain rd for the case D = 3,K = 1
• Then conduct comparative statics on Wd w.r.t. Tod
• Wd ↑ if Tod ↓: arrival of railroads
increase welfare
• Wd ↓ if Too′ ↓: railroads in other
Step 4: Specification
ln(Wo) = βo + βt + γRAILot
+ψ( 1 #No)
!
d∈No
RAILdt + εot
• Prediction 4: γ > 0, ψ < 0 • Estimated by OLS, assuming
Step 4: OLS Results (Table 5)
• Arrival of railroad
⇒ Real agri GDP per acre ↑ by 18.2%
• Railroads in No
⇒ Real agri GDP per acre ↓
• Treatment externality: ignoring this yields understimation (column 1)
Step 4: validity checks
1. Placebo tests: estimate impact of proposed but never built railroads ⇒ No effect (Table 6)
2. IV estimation: 1876-78 rainfall
deviation from long-run mean as IV ⇒ IV estimate: similar magnitude to
OLS (Table 7)
3. Bounds check: estimate
coefficients separately for each category of railroads
Step 5: Prediction 5
1. Akd UP ⇒ Wd UP
2. Todk DOWN ⇒ Wd less responsive
to Akd
Step 5: Results (Table 9)
• Indeed ψ2ˆ > 0,ψ3ˆ < 0
Step 6: Results (Table 10)
Once the openness term is included as a regressor,
• Railroad coefficients become insignificant and close to zero. • ψ2ˆ is close to 1, ηˆ is close to -1! ⇒ Model explains a very large portion
Future research in the literature
(my own view)
• Impact on industrial development • Distributional consequences of
infrastructure construction • Non-economic impact of
infrastructure
• Public service delivery should be affected by infrastructure, too.
References for the lecture on infrastructure
Banerjee, Abhijit V., and Lakshmi Iyer. 2005. “History, Institutions and Economic
Performance: The Legacy of Colonial Land Tenure Systems in India.” American Economic Review 95(4): 1190-1213. !
Deaton, Angus. 1997. The analysis of household surveys. World Bank Publications. !
Dinkelman, Taryn. 2008. “The effects of rural electrification on employment: New evidence from South Africa.”
Donaldson, Dave. 2008. “Railroads of the Raj: Estimating the Impact of Transportation Infrastructure.”
Duflo, Esther, and Rohini Pande. 2007. “Dams.” Quarterly Journal of Economics 122(2): 601-646. !
Eaton, Jonathan, and Samuel Kortum. 2002. “Technology, Geography, and Trade.”
Econometrica 70(5): 1741-1779. !
Hansen, Christian B. 2007. “Generalized Least Squares Inference in Panel and Multilevel Models with Serial Correlation and Fixed Effects.” Journal of Econometrics 140(2): 670-694.
!
Imbens, Guido W. 2007. “Instrumental Variables with Treatment Effect Heterogeneity: Local Average Treatment Effects.” Available at: http://www.nber.org/minicourse3.html.
Imbens, Guido W., and Joshua D. Angrist. 1994. “Identification and Estimation of Local Average Treatment Effects.” Econometrica 62(2): 467-475. !