2000 palm oil land - Crawford School of Public Policy

Palm oil and poverty in Indonesia
Ryan B. Edwards
Arndt-Corden Department of Economics
Crawford School of Public Policy
ADEW, June 2015
ryan.edwards@anu.edu.au
@ryanbedwards
Question
Oil palm
land use
Poverty
Counterfactual example: oil palm vs. all over land use
Does district d, which has converted land into oil palm
plantations, currently have lower poverty than it would
have if it had not converted the land?
2
Approach
Nationwide district-level balanced panel (DAPOER)
– No RCT or policy discontinuity for natural experiment
– Exploit spatial and temporal variation
Longitudinal features of the data allow me to control for:
– Time-invariant district-specific confounders
– Common nation-wide and regional shocks and policies
– Different initial conditions and time trends
Causal effect identified if no time and district-varying OVB
1. Theory and empirical exercises suggest unlikely
2. Consonant results from instrumental variable estimation
Caveat: data not perfect and omitted variables cannot be ruled out
3
District poverty rate, 2002-2010
Share of population below expenditure-based poverty line
Also look at poverty gap index, no. of poor (i.e., drop denominator)
• Most poor people live on Java; rates highest periphery
• 10 million people lifted from poverty in the decade under study
4
Official (declared) palm oil land / total district area
Tree Crop Statistics; reflects policy settings (actual not available)
• Land key policy issue; central to expansion (92%)
• 2001-2009, 5  17 mill. ha (8.7%), mostly “smallholder” expansion
• Reduced-form; heterogeneity contained with
• Share focuses on changing composition (c.f., other uses)
5
Basic estimating equation
District and period specific error term
Heteroskedasticity robust
District-clustered for serial correlation
Lag District oil palm land / total district area land
SUSENAS measured mid-year; need time to effect
Not logged to retain the ‘control group’
Log poverty rate in district d in year t
Natural logs for skewedness
Semi-elasticity interpretation
6
Fixed effects and estimation
District fixed effects (d=~330)
Time invariant district specific factors
Climate, geography, history, institutions
Island-year fixed effects (i=5; t=9)
Time and district varying factors shared across island
Regional and national growth, shocks, cycles, policies
Mean-differenced FE estimator
Allows level effects
Time-to-effect temporal dimension
Long differences
Remove time-invariant bias
Island, initial pov.,palm, controls
7
Identification
No time and district varying omitted variables correlated with
changes in both poverty and palm oil land influential enough to
systemically shift poverty trends within island groupings
(i.e., conditional independence)
1.
2.
3.
4.
Identifying variation fundamentally driven by bureaucracy
Selection across & within arguably mostly captured by FEs
Placebo tests using future values
Coefficients stable to DAPOER district-specific controls and
adding province-specific time trends (very-rich control vector)
5. Unobservable hetero can’t be ruled out in observational studies
8
Instrumental variables
Panel LIML IV: 2000 district palm oil land*year
• Focus on demonstration effects:
– path-dependence from initial-conditions
– pre-existing knowledge, materials, supply chains
• Main drawback: LATE omits “new” palm oil districts
Long-difference LIML IV
• 2000 palm oil land as above
• Rainfall Better in humid low land tropics
• Slope Less mountainous better for plantation
Significant < 0.1% in first-stage
All three violate exclusion restriction in theory
9
No relationship between IVs
and poverty reduction in
“non-treated” districts.
Cannot rule out a relationship
in treated districts.
Bonus: if an instrument is
valid, over-identification tests
are no longer useless
(informative on lag land)
10
Main result: Effect of palm oil land on district poverty
Annual
Sample
Estimator
FE
LIML
Two yearly
Four yearly
FE
FE
LIML
LIML
Long difference
FE
LIML
Impact on poverty (coefficient, semi-elasticity)
0
-0.01
-0.02
-0.03
-0.04
-0.05
-0.06
-0.07
11
Heterogeneity, by sub-sector
State-owned land
Sample
Annual
Two yearly
Four yearly
Baseline district and year FE model (no IV)
Long-diff not available;
Private owned (company) land
Annual
Two yearly
Four yearly
Smallholder land
Annual
Two yearly
Four yearly
Impact on poverty (coefficient, semi-elasticity)
0
-0.005
-0.01
-0.015
-0.02
12
Answer to question
Districts which have increased oil palm
land have experienced greater
reductions in the poverty rate and the
depth of poverty
• Effect increases from 1-4 years
• Evident a decade later
• Relatively consistent across regions
and sectors
• No evidence of local short-run
economic spillovers (not presented)
13
Main limitations
• Cannot generalise beyond average within-district effect
• Long-shot analysis: mechanisms, spill-overs, GE effects unclear
• Endogeneity still a problem; declared vs. actual
Ongoing improvements to the paper
• Strengthen identification strategy (OVB bounds; controls)
• Reconcile with SUSENAS (distributional analysis)
• Look into mechanisms further (labour demand, logging
income, role of productivity, spill-overs and other sectors)
• More detailed regional analysis
14