Natural Resource Sectors and Human Development ... · Natural Resource Sectors and Human Development: International and Indonesian Evidence Ryan B. Edwards a thesis submitted for

Natural Resource Sectorsand Human Development:

International and Indonesian Evidence

Ryan B. Edwards

a thesis submitted for the degree ofDoctor of Philosophy at the

Australian National University

April 2016

i

© Copyright by Ryan Barclay Edwards 2016All Rights Reserved

Declaration

This thesis is my own work.

A version of Chapter 2 is published in World Development:

Edwards, R. B. (2016), “Mining away the Preston curve”, World Development,78, February, pp. 22–36.

Ryan B. EdwardsJanuary 2016

i

Acknowledgements

I first and foremost thank Paul Burke, Chair of my PhD supervisory panel. Paul

has been an amazing supervisor and provided exceptional guidance throughout

my PhD journey. I would also like to thank my other two panel members, Budy

Resosudarmo and Robert Sparrow, who were always happy to discuss ideas and

provide useful feedback.

I was fortunate to complete my PhD in the Arndt-Corden Department of

Economics (ACDE) at the Australian National University (ANU), with its unique

focus on the economies of the Asia-Pacific region. Prema–chandra Athukorala,

Sommarat Chantarat, Max Corden, Sarah Dong, Hal Hill, Raghbendra Jha, Heeok

Kyung, Blane Lewis, Chris Manning, Ross McLeod, Kate Mclinton, Nurkemala

Muliani, Arianto Patunru, Daniel Suryadarma, Peter Warr, Ben Wilson, and

Sandra Zec have been fantastic and helpful colleagues.

My month of fieldwork supported by Budy and the Indonesia Project greatly

enhanced my PhD experience, allowed me to develop enduring friendships, and

laid the foundations for my ongoing research program. I thank Thomas Barano,

Ernawati Apriani, Yudi Agusrin, Margaretha Nurrunisa, and WWF-Indonesia for

kindly hosting me in Jakarta and accompanying me around Sumatra, and Daniel

Suryadarma at CIFOR/ICRAF, Matthew Wai-Poi at the World Bank, Kiki Verico

at the University of Indonesia, and Bank Indonesia for being such generous hosts.

Bill Wallace, Asep Suryahadi, Indira Hapsari, Meine van Noordwijk, Ernest

Bethe, Triyanto Fitriyardi, Dhanie Nugroho, and Ari Perdana also provided

useful discussions and feedback.

ii

iii

ACDE PhD students provided a stimulating environment throughout my PhD

journey, particularly Lwin Lwin Aung, Ariun-Erdene Bayarjargal, Rohan Best,

Kimlong Chheng, Omer Majeed, Matthew McKay, Huy Nguyen, Manoj Pandey,

Rajan Panta, Umbu Raya, Marcel Schroder, Yessi Vadila, Samuel Weldeegzie, and

Agung Widodo.

ACDE is part of the Crawford School of Public Policy at the ANU. The Crawford

School was an ideal place for PhD study in many respects, allowing me to

develop my teaching skills, providing stimulating public policy events and

seminars, and providing a collegial environment of PhD students and academic

staff. I thank Fitrian Ardiansyah, Adriyanto, Shiro Armstrong, Bob Breunig,

Alrick Campbell, Bruce Chapman, Leo Dobes, Matt Dornan, Mark Fabian, Ippei

Fujiwara, Yusaku Horiuchi, Wee Koh, Stephen Howes, Llewelyn Hughes, Frank

Jotzo, Tom Kompas, Ida Kubiszewski, Belinda Lawton, Luke Meehan, David

Stern, Julia Talbot-Jones, Ariane Utomo, Peter Whiteford, and Terence Wood for

their friendship and support. I am also grateful to our excellent post-graduate

coursework students for keeping me on my toes when teaching and for their

usually constructive and positive feedback.

Kay Dancey and the CartoGIS unit at the ANU College of Asia and the Pacific

kindly assisted with maps and GIS training.

Indira Hapsari, Robert Sparrow, Kay Dancey, Arianto Patrunu, Agung Widodo,

Yessi Vadila, Budy Resosudarmo, Blane Lewis, Susmita Dasgupta, and the

Australian Data Archive kindly shared data.

iv

Useful feedback on the three chapters of this thesis was received from three

examiners, five anonymous reviewers, one editor, Fabrizio Carmignani, Andrea

la Nauze, Richard Denniss, and from many conference and seminar participants

at the Australian National University, the Australasian Development Economics

Workshop at Monash University, the Centre for International Forestry Research,

the World Bank, Bank Indonesia, the University of Indonesia, the Australian

Conference of Economists at Queensland University of Technology, the World

Congress of Environmental and Resource Economics in Istanbul, and the

Australian Agricultural Resource and Environmental Economics Society.

I gratefully acknowledge financial support for activities undertaken during

the course of my PhD study from the Indonesia Project, ACDE, ANU, Monash

University, and the Australian Government.

Last but certainly not least, I thank my family—Colette, Raymond and Karina—

and my wife Jessica for the motivation and for seeing me through.

Abstract

This thesis collects three papers on natural resource sector-led development.

The first paper examines the long-term health and education impacts of mining

dependence. Exploiting between-country variation in a large international

sample, causal effects are identified through instrumental variable estimation.

Results show that countries with economies more oriented toward mining on

average display poorer health and education outcomes than countries of similar

per capita income. Income from sectors other than mining tends to deliver

better health and education outcomes. Key channels explaining the lower social

productivity of mining sector activity include its impacts on non-mining sectors

and institutions. Similar patterns are observed across Indonesian districts,

suggesting this is not only a country-level phenomenon.

The second paper examines the poverty impacts of the world’s largest modern

plantation sector expansion, Indonesian oil palm in the 2000s. The paper

combines administrative data on local oil palm acreage at the district level

with survey-based estimates of poverty, using an estimation approach in

long-differences. Identification is achieved through an instrumental variable

strategy exploiting detailed geospatial data on crop-specific agro-climatic

suitability. The key finding is that increasing the oil palm share of land in a

district by ten percentage points contributes to around a forty percent reduction

in its poverty rate. Of the more than 10 million Indonesians lifted from poverty

over the 2000s, my most conservative estimate suggests that at least 1.3 million

of these people have risen out of poverty due to growth in the oil palm sector.

Similar effects are observed for different regions of Indonesia, for industrial and

v

vi

smallholder plantations, and at the province level. Oil palm expansion tends to

be followed by a small but sustained boost to the value of agricultural output,

manufacturing output, and total district output.

The final paper presents three quantitative case studies on the local economic

and welfare impacts of rapid natural resource sector expansion in Indonesia. The

paper focuses on three districts that have experienced notably large production

booms for Indonesia’s three largest primary exports: palm oil (Indragiri Hilir, in

Riau), coal (Tapin, in South Kalimantan), and natural gas (in Manokwari, West

Papua). Counterfactuals are constructed for each case study district through

synthetic control modelling. Results suggest that all three resource booms boosted

total economic output and altered the structure of the local economy. Oil palm

expansion in Riau raised agricultural, industry, and services output, while coal

mining in South Kalimantan reduced agricultural and services output. Oil palm

and coal mining booms both appear to have delivered strong poverty reduction.

The Tangguh natural gas project in West Papua delivered a massive increase in

local economic and industry output, but I find no evidence of any discernible

impacts on household welfare and poverty. The three case studies show that

natural resource sectors can make important contributions to poverty alleviation.

Relative to their size, sectors with more concentrated rents tend to provide less

broad-based benefits than diffuse resource sectors using labour more intensively.

Contents

Declaration i

Acknowledgements ii

Abstract v

List of Figures ix

List of Tables 1

1 Introduction 11.1 Natural resource sectors . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Curse or blessing? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31.3 A focus on Indonesia . . . . . . . . . . . . . . . . . . . . . . . . . . . 91.4 Thesis purpose and approach . . . . . . . . . . . . . . . . . . . . . . 111.5 Key research questions and results . . . . . . . . . . . . . . . . . . . 121.6 Methodological contributions . . . . . . . . . . . . . . . . . . . . . . 151.7 Organisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2 Mining away the Preston curve 182.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192.2 Linking mining to health and education . . . . . . . . . . . . . . . . 222.3 Empirical approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

2.3.1 Instrumental variable strategy . . . . . . . . . . . . . . . . . 272.3.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

2.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362.4.1 Health and education effects of mining sector growth . . . . 362.4.2 Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 402.4.3 Health and education elasticities of income, by sector . . . . 412.4.4 Heterogeneity . . . . . . . . . . . . . . . . . . . . . . . . . . . 442.4.5 Potential channels . . . . . . . . . . . . . . . . . . . . . . . . 45

2.5 Within-country evidence from Indonesia . . . . . . . . . . . . . . . 482.6 Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . 522.7 Chapter 2 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

vii

Contents viii

3 Is plantation agriculture good for the poor? 683.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 693.2 Indonesia’s oil palm expansion . . . . . . . . . . . . . . . . . . . . . 72

3.2.1 Linking oil palms to poverty . . . . . . . . . . . . . . . . . . 743.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

3.3.1 Oil palm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 753.3.2 Poverty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 773.3.3 Pemekaran . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

3.4 Empirical approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . 803.4.1 Instrumental variable strategy . . . . . . . . . . . . . . . . . 81

3.5 Main results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 853.6 Short-run and dynamic impacts . . . . . . . . . . . . . . . . . . . . . 923.7 A migration story? . . . . . . . . . . . . . . . . . . . . . . . . . . . . 963.8 Heterogeneity and wider impacts . . . . . . . . . . . . . . . . . . . . 102

3.8.1 Heterogeneity by region . . . . . . . . . . . . . . . . . . . . . 1023.8.2 Heterogeneity by sector . . . . . . . . . . . . . . . . . . . . . 1043.8.3 Wider impacts . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

3.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1093.10 Chapter 3 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

4 Local impacts of resource booms 1214.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1224.2 Synthetic control approach . . . . . . . . . . . . . . . . . . . . . . . . 125

4.2.1 Estimation and inference . . . . . . . . . . . . . . . . . . . . 1264.2.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1294.2.3 Identifying appropriate case studies . . . . . . . . . . . . . . 1324.2.4 Constructing each synthetic control . . . . . . . . . . . . . . 138

4.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1414.3.1 Oil palm expansion in Sumatra . . . . . . . . . . . . . . . . . 1414.3.2 Coal mining in Kalimantan . . . . . . . . . . . . . . . . . . . 1494.3.3 Natural Gas Extraction in West Papua . . . . . . . . . . . . . 152

4.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1564.5 Chapter 4 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . 157

5 Conclusion 163

References 166

List of Figures

2.1 Preston Curve, 2005 . . . . . . . . . . . . . . . . . . . . . . . . . . . 202.2 Educational Attainment, Income, and Mining, 2005 . . . . . . . . . 212.3 Conceptual Framework . . . . . . . . . . . . . . . . . . . . . . . . . 232.4 Contribution of Mining to Value-Added, 2005 . . . . . . . . . . . . 312.5 Per Capita Fossil Fuels, 1971 . . . . . . . . . . . . . . . . . . . . . . 322.6 Health and Education Elasticities of Income, By Sector . . . . . . 432.7 Mining Share of District Regional Gross Domestic Product in

Indonesia, 2009 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 502.8 Global Infant Mortality, 2005 . . . . . . . . . . . . . . . . . . . . . 552.9 Global Years of Educational Attainment, 2005 . . . . . . . . . . . 56

3.1 Oil Palm Land as a Share of District Area, 2009 . . . . . . . . . . . 763.2 District Poverty Rates, 2010 . . . . . . . . . . . . . . . . . . . . . . . 783.3 Attainable Palm Oil Yield Across Indonesia . . . . . . . . . . . . . 823.4 Poverty Impacts of the 2001–2010 Oil Palm Expansion . . . . . . . . 873.5 Regional Heterogeneity . . . . . . . . . . . . . . . . . . . . . . . . . 1023.6 Sector Heterogeneity . . . . . . . . . . . . . . . . . . . . . . . . . . 106

4.1 Case Study Districts—Locations . . . . . . . . . . . . . . . . . . . . 1344.2 Case Study Districts—Treatments . . . . . . . . . . . . . . . . . . . 1364.3 Impacts of Oil Palm Expansion in Sumatra . . . . . . . . . . . . . . 1424.4 Consumption in Indragiri Hilir . . . . . . . . . . . . . . . . . . . . . 1474.5 Kernel Density Estimate–Consumption Distribution in Indragiri

Hilir . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1484.6 Impacts of Coal Mining in Kalimantan . . . . . . . . . . . . . . . . 1504.7 Impacts of Gas Extraction in West Papua . . . . . . . . . . . . . . . 1544.8 Appendix–Impacts of Oil Palm Expansion in Sumatra, non-zero Y

axis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1604.9 Appendix–Impacts of Coal Mining in Kalimantan, non-zero Y axis 1614.10 Appendix–Impacts of Gas Extraction in West Papua, non-zero Y axis162

ix

List of Tables

2.1 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352.2 Health and Education Effects of Mining Dependence . . . . . . . 372.3 Sensitivity Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392.4 National-level Mechanisms . . . . . . . . . . . . . . . . . . . . . . . 462.5 Sub-national Evidence from Indonesian Districts . . . . . . . . . . 512.6 Main Results Instrumenting with Mineral, Oil, Gas, and Coal

Reserves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 572.7 Main Results Instrumenting with only Oil and Gas Reserves . . . 582.8 Main Results using Alternative Time Periods . . . . . . . . . . . . . 592.9 Main Results using Alternative Measures of Mining . . . . . . . . 602.10 Main Results without Resource-rich Mini-states and Other

Outliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 612.11 Health and Education Elasticities of Income from Different

Sectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 622.12 Regional Sub-samples . . . . . . . . . . . . . . . . . . . . . . . . . . . 632.13 Sub-sample Estimation by Institution Type . . . . . . . . . . . . . . . 642.14 Local Average Partial Effects, By Commodity . . . . . . . . . . . . 652.15 National-level Mechanisms—OLS and IV . . . . . . . . . . . . . . 662.16 Main Results, By Gender . . . . . . . . . . . . . . . . . . . . . . . . . 67

3.1 Poverty Impacts of the 2001–2010 Oil Palm Expansion . . . . . . . . 863.2 Impacts of the 2001–2010 Oil Palm Expansion on the Poverty Gap . 893.3 Additional Covariates and Robustness . . . . . . . . . . . . . . . . 913.4 Short-run and Dynamic Poverty Impacts . . . . . . . . . . . . . . . 943.5 Short-run and Dynamic Impacts on Poverty Depth . . . . . . . . . 973.6 Province level Results . . . . . . . . . . . . . . . . . . . . . . . . . . 1003.7 Population, Poor People, and Production . . . . . . . . . . . . . . 1013.8 Regional Heterogeneity . . . . . . . . . . . . . . . . . . . . . . . . . 1033.9 Heterogeneity by Plantation Type . . . . . . . . . . . . . . . . . . . 1053.10 Effects of Oil Palm Expansion on Sectoral and Total District GDP1083.11 Estimated Contribution to Poverty Reduction . . . . . . . . . . . . 1093.12 Panel Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . 112

x

LIST OF TABLES 1

3.13 Main Results—Linear-linear Functional Form . . . . . . . . . . . 1133.14 Main Results—Production Instead of Land . . . . . . . . . . . . . 1143.15 Determinants of Changing Oil Palm Land Shares . . . . . . . . . . 1153.16 Robustness—Panel Fixed Effects . . . . . . . . . . . . . . . . . . . . 1163.17 Robustness–Alternative Samples . . . . . . . . . . . . . . . . . . . . 1173.18 Heterogeneity–By Region . . . . . . . . . . . . . . . . . . . . . . . . 1183.19 Heterogeneity–By Sector . . . . . . . . . . . . . . . . . . . . . . . . 1193.20 Heterogeneity–By Land Quality . . . . . . . . . . . . . . . . . . . . 120

4.1 Case Study Districts— Descriptive Statistics . . . . . . . . . . . . . 1354.2 Impacts of Oil Palm Expansion in Sumatra . . . . . . . . . . . . . . 1414.3 Predictor Balance: Indragiri Hilir . . . . . . . . . . . . . . . . . . 1444.4 Synthetic Indragiri Hilir Weights . . . . . . . . . . . . . . . . . . . 1454.5 Impacts of Coal Mining in Kalimantan . . . . . . . . . . . . . . . . 1494.6 Impacts of Gas Extraction in West Papua . . . . . . . . . . . . . . . 153

Chapter 1

Introduction

1.1 Natural resource sectors

Natural resource sectors have been key drivers of the unprecedented economic

development of the 20th century, with the consumption of fossil fuel and forestry

products powering the industrial revolution, the enlightenment, and the global

human progress that followed.1 Natural resources are assets (i.e., raw materials)

occurring in nature that can be used for economic production or consumption;

natural resource sectors are industries extracting and processing them. Today

natural resource sectors remain important in the global economic landscape. The

total value of natural resource sector exports from developing countries in 2010

was 9.8 and 9 times larger than foreign aid and remittances (Jedwab, 2013; World

Bank, 2014).

In this thesis I focus on two critical natural resource sectors: mining, broadly

defined to include coal and mineral mining and oil and gas extraction2; and palm

oil. Palm oil is a unique export-oriented sector similar to mining in many respects.

Both sectors are characterised by economic enclaves and negative environmental

externalities. Unlike many other cash crops, fresh fruit from oil palms must be

processed shortly after harvest, requiring processing facilities and large up-front

infrastructure investments (e.g., transport) similar to mining. These up-front

1See, e.g., Ayres and Warr (2005; 2009), Cleveland et al (1984), Deaton (2013), Hall et al (2003),Mayumi (1991), Kander et al (2013), Kander and Stern (2014), and Wrigley (2010).

2Although the oil and gas industries have several different characteristics to coal and minerals,these sectors have many similarities and major resource sector companies tend to diversify acrosscommodities for scale and scope efficiencies.

1

§1.1 Natural resource sectors 2

investment requirements ensure the palm oil sector grows principally through

plantations, which are better able to recover large up-front investments and deal

with long gestational periods than family farms or smallholders (Hayami, 2010).

Smallholder oil palm typically emerges once this prerequisite infrastructure is in

place, much like artisanal mining around large mining facilities.

Natural resource sectors can have several unique characteristics requiring

special attention from policy-makers. Natural resource endowments (including

forestry resources) tend to be the property of the nation state, and resource

rents are often legally intended to be distributed amongst countries’ current

and future populations. A government allowing a nation’s natural resources

to be extracted without benefiting the national economy and its people is

unlikely to be a popular one, so natural resource sectors are often of interest to

politicians. Natural resource sectors are characterised by negative environmental

and public health externalities, giant infrastructure investments, and long project

life cycles requiring local and macro-level economic and political stability.

Compared to other sectors, there is a greater role for policy-makers in weighing

up the relative costs and benefits of large natural resource projects requiring

government approval (e.g., through concessions, licensing, permits, special

regulations) and often, for more political than economic reasons, obtaining

direct government support (e.g., through state-owned enterprises, favourable

financing arrangements, tax exemptions, and direct subsidies). Understanding

the contribution of natural resource sectors to human development is important

for informing development strategies and policy settings.

§1.2 Curse or blessing? 3

1.2 Curse or blessing?

Despite their important role in driving economic development in the 20th

century, the role of natural resource sectors for improving well-being today is less

clear. Many resource-rich countries have experienced sustained natural resources

sector-driven economic growth but little progress in improving development

outcomes. Nigeria’s oil revenues increased almost tenfold from 1965–2000, but

real income stagnated, and poverty and inequality increased (van der Ploeg,

2011). Papua New Guinea had an average GDP per capita growth rate of 8%

over the second half of the 2000s yet a virtually unchanged poverty rate of

nearly 40% (World Bank, 2014). Natural resource sector-led economic growth

swept Africa in the 2000s, but agricultural yields across the continent remain low,

manufacturing and services sectors remain relatively small and unproductive,

and poverty reduction has been disappointing (Caselli, 2005; Evenson and Gollin,

2003; Restuccia et al, 2008; Vollrath et al, 2015; and World Bank, 2013). Comparing

these recent experiences with those of resource-poor countries like South Korea,

Singapore, and Taiwan (or the manufacturing-led growth of China), natural

resource sector-driven economic growth appears to correspond to countries

developing distinctly differently.3 Much is known about how natural resource

sectors affect macroeconomic aggregates and political institutions, but many

questions about how natural resource sectors affect broader well-being remain

largely unanswered.

3For greater tractability in this introductory discussion, I focus on mining sectors, as naturalresources have been typically conceptualised in the literature. Palm oil is only produced in ahandful of developing countries and has some similar characteristics to mining (c.f. most countrieshave mining sectors).


The idea that natural resources can “curse” countries that own them has

been widely held. Early studies by Auty and Mikesell (1998), Gylfason,

Herbertsson, and Zoega (1999), and Sachs and Warner (1995; 2001) documenting

negative associations between resource abundance and economic growth led to

a burgeoning field of empirical research contesting the existence of the “resource

curse” (Torvik (2009), van der Ploeg (2011), and Wick and Bulte (2009) provide

reviews). This research focused on the relationship between natural resources

and aggregate economic activity, with institutional quality widely believed to

determine whether natural resources become a curse or a blessing (Collier and

Goderis, 2012; Mehlum et al, 2006; Torvik, 2002, 2009). As highlighted by van

der Ploeg (2011), cross-country evidence has been sensitive to sample period,

countries, variable definitions, omitted variables, measurement error, and other

factors.4 van der Ploeg (2011) concludes “the road forward might be exploiting

variation within a country where variables that might confound the relationship

between resources and macroeconomic outcomes do not vary and the danger of

spurious correlation is minimised”.

Within-country studies on the economic impacts of natural resources are now

common, and more recent national-level studies with improved identification

strategies tend to disprove the idea of an aggregate GDP resource curse.5 Smith

(2015) shows that countries that have become resource-rich since 1950 now have

significantly higher per capita incomes that they would have without the resource

discoveries. Cotet and Tsui (2013) show the same exclusively for oil. For diffuse

natural resources, Nunn and Qian (2011) show that areas more suitable for

agricultural production are significantly more developed today.4For example, studies following the Sachs and Warner (2001) approach arrive at different

results instrumenting for resource abundance (Brunnschweiler and Bulte, 2008), includingcountry fixed effects (Manzano and Rigobon, 2001), and using different measures of resourceintensity (Lederman and Maloney, 2007).

5Examples of within-country studies on the effects of natural resources and a boomingresource sector include Aragon and Rud (2013), Black et al (2005), Caselli and Michaels (2013),Domenech (2008), Dube and Vargas (2013), Fleming and Measham, (2014), James and James (2012)and James and Aadland (2011), Michaels (2010), and Papyrakis and Gerlagh (2007). Cust andPoelhekke (2015) provide a review.


Policy recommendations for managing natural resource wealth typically seek

to address the mechanisms through which a possible resource curse could

act. For example, improved governance and transparency policies are designed

to deal with the direct negative impacts of resource wealth on institutional

quality (e.g., the Extractive Industries Transparency Initiative). Different

macroeconomic policies are often used to manage potential Dutch disease

effects and macroeconomic volatility associated with commodity prices (e.g.,

sterilisation and sovereign wealth funds).6 Such policies have been in place in

several resource-rich countries since the early 2000s (Arezki et al, 2011), yet the

poor development performance of countries with natural resource sector-driven

economic growth seems unchanged.

I follow a different line of enquiry. If we are concerned with overall living

standards, we must look beyond economic aggregates and examine how natural

resource sectors affect well-being across society. Positive aggregate income effects

arising from resource wealth are unlikely to be spread evenly across society, as

natural resource sectors are closely linked to extractive institutions (Acemoglu

et al, 2001). Standard empirical relationships between higher income levels and

improved health (Summers and Pritchett, 1994), education (Barro and Lee, 1993;

2010; Taubman, 1989), and poverty outcomes (Dollar et al, 2016) may be different

in the context of income from natural resource sectors. In this thesis I examine the

links between natural resource sectors and human development, focusing on (a)

human capital, which I consider as as the combination of health and education

outcomes and an important channel for economic and social mobility in all

economies; (b) spillovers to non-resource sectors, which tend to generate greater

employment opportunities and productivity benefits due to higher labour and

skill intensities (discussed further below); and (c) poverty, the first and foremost

Sustainable Development Goal.

6van der Ploeg and Poelhekke (2009), and van der Ploeg (2011) discuss the importance ofvolatility in resource-dependent economies in detail. The Dutch disease is explained below.


The experiences of natural resource exporting countries in the post-war period

support the argument that “natural capital appears to crowd out human capital,

thereby slowing down the pace of economic development” (Gylfason, 2001).

While human capital has typically been seen as a channel for natural resources

to stunt economic growth (Gylfason and Zoega, 2006), primary commodities

abundance and dependence have been shown to lower broad measures of social

development (Bulte et al, 2005; Carmignani, 2013; Carmignani and Avom, 2010).

To understand how natural resource sector-led economic growth affects long-term

social development trajectories, it is instructive to revisit some of the earliest work

on effects of a booming primary sector on the rest of the economy: the “Gregory

thesis”, or as it is more commonly known, “Dutch disease” theory (Corden, 1984;

Corden and Neary, 1982; Gregory, 1976).

The Dutch disease occurs when growth in export-oriented natural resource

sectors reduce activity levels and employment in other tradable sectors

(e.g., manufacturing and highly-skilled services) by raising factor prices and

reducing competitiveness relative to foreign-produced substitutes (Corden, 1984).

Non-tradable services, usually low-skilled, tend to see an increase in activity

levels and employment. Lagging sectors in a Dutch disease affected economy

are those where profitability and investment incentives are trapped between

rising domestic costs and output prices set in world markets: some parts of

agriculture, skilled tradable services, and most of manufacturing. Dutch disease

theory initially found limited empirical support due to the sheer dominance of

the net income effects of natural resource booms and the diverse experiences

of resource rich countries (Auty, 2001). Today, natural resource sector-driven

growth continues to deliver higher per capita incomes and other signs of economic

development (e.g., urbanisation) without the industrialisation-related structural

change that historically accompanied broad development progress (Vollrath et al,

2015), just as the Dutch disease predicts.


But do these missing sectors matter? Export-oriented manufacturing

and skilled tradable services are characterised by technological dynamism,

information and productivity spillovers, and other agglomeration effects that

drive economic growth by making capital and other factors of production more

productive (Ellison et al, 2010; Greenstone et al, 2010; Hanson, 2012; Rodrik,

2015).7 Natural resource sectors and non-tradable services, by contrast, are not

skill-intensive, tend to experience less productivity growth, and deliver fewer

productivity spillovers to other industries.8

A sustained resource boom could thus reduce the growth of skilled jobs, lower

returns to existing human capital, and reduce incentives to invest in new human

capital, as predicted by Gylfason (2001) and highlighted in Grilches’ (1969) early

work on the complementarity of physical and human capital. In the short term,

there is likely to be downward pressure on enrolment, retention, and graduation

rates (particularly at higher levels) due to rising demand for labour in low-skilled

industries and the rising opportunity cost of attaining further education. In the

long term, the natural resource sector-driven economy will be poorly positioned

to transition into innovation- and skills-driven economic growth (Eichengreen,

Park, and Shin, 2013). Given the important role of education as a pathway to

economic mobility, structural change in the labour market shapes the distribution

of income and opportunities.9 In developing countries, formal opportunities

vary by sector and structural change redistributing employment across sectors

alters the probability of informal employment. Unskilled non-tradable services

sectors that often boom with natural resource sectors tend to have high levels of

informality and act as a sink for the underemployed, the unemployed, and the

7In explaining why resource booms could lower economic growth (i.e., the “resource curse”),it was often assumed that it was due to stagnating manufacturing (see, e.g., Matsuyama (1992)and Torvik (2001)), known for being an “escalator” for economic growth (Rodrik, 2015).

8Lederman and Maloney (2007), Corden and Neary (1982), Krugman (1987), Matsuyama(1992), and van Wijnbergen (1984) show why this may be the case. Michaels (2010) and Blacket al (2005) provide some recent evidence to the contrary.

9Cassing and Warr (1985) discuss the distributional impacts in a standard Dutch disease modelin more detail.


poor. A focus on sectors and structural change is thus critical to understand

the human development and distributional consequences of natural resource

sector-driven economic growth.

Distributional aspects of natural resource sectors, particularly mining, have

also been neglected by the literature (Ross, 2007), particularly relating to poverty.

In fact there is a strong negative relationship between a country’s dependence

on resource rents and the amount of data we have about its inequality and

poverty levels (Ross, 2007). Bhattacharyya and Williamson (2013) study Australia,

finding that the richest benefit disproportionately from resource booms, but not

agricultural booms, and note that “ the empirical literature on the inequality and

resource boom connection is relatively thin.” Gylfason and Zoega (2003) and

Goderis and Malone (2011) similarly find that resource booms increase inequality.

Carmignani and Avom (2010) and Carmignani (2013) find that natural resource

abundance and dependence increases income inequality and that this is the key

mechanism through which natural resources can hinder social development.

Turning to poverty, Smith and Wills (2015) exploit detailed global geo-spatial

data and find that oil booms around the world have not benefited the rural

poor. Bhattacharya and Resosudarmo (2015) find that province-level mining

sector growth accelerations actually increase poverty in Indonesia, and Aragon

and Rud (2015) find that gold mining in Ghana significantly worsens poverty in

surrounding areas. In a review paper, Gamu et al (2015) conclude that extractive

industries appear to make limited contributions to poverty reduction. Given the

strong environmental and climate change externalities associated with natural

resource sectors, particularly fossil fuels and palm oil, the development challenges

summarised in this subsection become even more pertinent when considered in

the context of the recently agreed Sustainable Development Goals and the Paris

climate change agreement.

§1.3 A focus on Indonesia 9

1.3 A focus on Indonesia

With a population of 253 million, Indonesia is the world’s third most populous

developing country after India and China. Indonesia’s per capita income was

$3,492 USD in 2014. This places Indonesia on the cusp of being classified as an

upper-middle income country, yet 28 million Indonesians (11.4%) lived below the

national poverty line in 2014 and a further 68 million remain vulnerable to poverty

(World Bank, 2014). Indonesia is the world’s largest natural resource-dependent

economy, with coal, natural gas, and palm oil its three largest exports and

responsible for most of the country’s recent economic growth (Garnaut, 2015).

Today Indonesia is the world’s largest exporter of thermal coal and palm oil, two

sectors at the centre of contemporary sustainable development policy debates.

Natural resource-driven economic growth from the 1970s to the early 2000s

supported Indonesia’s rapid industrialisation and provided broad-based benefits

and poverty reduction (Hill, 1996). Indonesia was often regarded as one of the

few developing countries to have not suffered from the “resource curse” for its

post-war development progress (Temple, 2003). The same cannot be said of

the last two decades. Like some African countries, the Indonesian economy

was buoyed through the global economic crisis of 2007–2008 by the commodity

boom of the 2000s (Burke and Resosudarmo, 2012; Hill et al., 2008). Demand

for Indonesia’s industrial crops and mined commodities—driven by economic

expansion in China and other large Asian economies—helped to sustain an

average annual per capita GDP growth of almost 5 percent over the decade, lifting

Indonesia’s income per capita from 15% of the world average in 2001 to 20% in

2011. The share of Indonesia’s main resource exports (oil, natural gas, coal, copper

and palm oil) in total merchandise exports rose from 34 percent to 46 percent,

driven almost entirely by coal and palm oil.10 The real exchange rate rose by

almost 4 percent per year during the 2000s and growth all of Indonesia’s key10The share of palm oil in merchandise exports rose from 2 percent to 9 percent, and coal from

3 to 14 percent.

§1.3 A focus on Indonesia 10

skill-intensive manufacturing industries slowed (Coxhead and Li, 2008; Coxhead,

2014), contributing to more than half of Indonesia’s GDP growth from 1990–1996

but less than a third from 2000. Following the persistent drop in investment after

the Asian financial crisis, capital stock shifted from mostly manufacturing before

2000 to mostly construction, particularly non-residential. The growth of capital

per worker (excluding construction) was very low, even negative, throughout the

2000s (van der Eng, 2009).

The boom of the 2000s was the first for which a labour-intensive agricultural

product like palm oil played a central role. Indonesia exhibits all the symptoms of

a modern Dutch disease-affected economy, continuing to grow and urbanise while

export-oriented manufacturing and skilled services sectors stagnate. Real wages

across all sectors have been flat since the early 2000s (Coxhead, 2014). Indonesia

performs poorly by most international health and education comparisons, with an

already under-educated labour force showing no discernible catch-up relative to

slower-growing economies (Newhouse and Suryadarma, 2011; Suryadarma and

Jones, 2013; World Bank, 2014). The pace of poverty reduction has slowed and

inequality continues to rise, but there remains little evidence on the role that the

resource boom of the 2000s has played in shaping these outcomes (Yusuf, 2013;

World Bank, 2014). In this thesis, I empirically examine how the mining and palm

oil sectors affect some of these development outcomes. This evidence could assist

countries like Indonesia to better align development strategies across different

economic, social, and environmental policy objectives. The environmental

impacts of natural resource sectors have been well documented and are beyond

the scope of this thesis.

§1.4 Thesis purpose and approach 11

1.4 Thesis purpose and approach

This thesis presents three self-contained empirical research papers on

how two critical natural resource sectors—mining and palm oil—affect

human development outcomes. Human development involves expanding

the opportunities and freedoms offered to all people, and is defined as the ability

to live a long and healthy life, attain knowledge, and have a decent standard of

living.

I focus on health, education, non-resource sectors, and poverty, and apply a

range of econometric techniques to identify the impacts of natural resource sectors

at the (a) country level, using a large international sample; (b) within-country

level, using a national sample of Indonesian districts; and (c) local level, through

three district-specific case studies in Indonesia.

Research methods vary by context and appropriateness for the research

questions at hand. The three research papers use observational data and

exploit quasi-experimental conditions for causal inference. Key methods include

(a) between estimators exploiting variation across units, for long-run effects;

(b) long-differences exploiting differential rates of changes across units, for

medium-run effects; (c) panel data estimators exploiting within unit variation, for

short-run effects; (d) instrumental variable estimators to deal with unobserved

heterogeneity and measurement error; and (e) synthetic control modelling, to

make causal inferences on single case study units.

§1.5 Key research questions and results 12

1.5 Key research questions and results

Chapter 2—Mining away the Preston curve

The first research paper asks the question: how does mining dependence affect

long-term health and educational development? I estimate the long-term national

health and education impacts of having a larger mining sector, instrumenting the

relative size of the mining sector with the natural geological variation in countries’

fossil fuel endowments to obtain causal effects. By comparing countries with

different structural compositions, I examine the “social productivities” of different

types of economic activity, focusing on fossil fuel extraction.

The findings suggest that countries with larger mining shares tend to have

poorer health and education outcomes than countries with similar per capita

incomes, geographic characteristics, and institutional quality. Doubling the

mining share of an economy corresponds to the infant death rate being twenty

percent higher, life expectancy being five percent lower, total years of education

being twenty percent lower, and seventy percent more people having no formal

education. Divergences from the Preston curve—the concave relationship

between cross-country income and life expectancy (Preston, 1975)—are thus

partly explained by the size of the mining sector. I test the generality of my

results by estimating an analogous model using a large cross-section of Indonesian

districts. Similar patterns are observed. I also provide evidence on key causal

mechanisms, finding that mining dependence is associated with lower levels of

non-mining income, lower health investment, and weaker democratic institutions.

The findings of this chapter provide support for a growing body of evidence

linking mining to poorer average living standards, particularly vis-a-vis other

types of income.


Chapter 3—Is plantation agriculture good for the poor?

Chapter 3 turns to the world’s largest modern plantation-based agricultural

expansion. I ask whether Indonesian districts that converted more land into oil

palm plantations over the 2000s now have lower poverty as a result. I combine

administrative information on local oil palm acreage at the district level with

survey-based estimates of district poverty, and relate decadal changes in oil

palm plantation area to changes in district poverty rates to compare the poverty

elasticity of oil palm land against alternative uses for land (e.g., rice and forestry).

Causal effects are identified by instrumenting the size of each districts’ oil palm

expansion with its relative agro-climatic suitability for the crop.

The key finding is that districts with larger oil palm expansion have

achieved more poverty reduction than otherwise-similar districts without oil

palm expansion. The magnitude of the estimated poverty reduction from

increasing the district share of oil palm land by ten percentage points from

my preferred estimator is around 40% of the initial poverty rate. Poverty

gaps significantly narrow, suggesting not only those near the poverty line are

being lifted up. I assess short-term effects and dynamics using standard panel

estimators with distributed lags and I find no evidence of any major effect

heterogeneity when I disaggregate by large plantations and smallholders. Similar

effects are also observed across Indonesia’s major palm oil producing regions and

at the province level. I find evidence of minor spillovers to other local economic

activities. Oil palm expansion tends to be followed by a small but sustained

boost to the value of agricultural, manufacturing, and total district output. While

the links between agriculture and poverty have been widely studied, plantation

agriculture has received relatively little attention. To my knowledge, I provide

the first estimates of the poverty elasticities plantation-based agricultural growth.

That oil palm growth has been pro-poor in Indonesia is consistent with existing


findings on agricultural-based growth.11 This new evidence will be able to inform

an ongoing policy debate on the future of palm oil across the developing world.

Chapter 4—Local impacts of resource booms

The final research paper asks: how does rapid natural resource sector

expansion affect a local district economy and its residents’ welfare? To answer this

question, I present three quantitative case studies. I focus on Indonesia’s three

largest export commodities at the heart of modern climate change debates and

exploit some of the largest and most sudden increases in district-level production:

palm oil in Indragiri Hilir, Riau; coal mining in Tapin, South Kalimantan; and a

giant natural gas project in the Bintuni Bay of West Papua. The three sectors have

all been argued to be economic enclaves, but have starkly different characteristics.

I use a relatively new empirical method—synthetic control modeling—to

construct a “synthetic” comparison district for each resource boom district.

This allows me to compare the booming districts’ key economic and human

development indicators—per capita economic output, its components, average

household expenditures, and poverty rates—with reasonable counterfactuals.

The findings suggest that all three sectors transform the local economy and

reduce poverty. Palm oil expansion in Indragiri Hilir delivered a small boost

to all sectors of the economy, and strong poverty reduction. Coal mining in

Tapin reduced agricultural and services sector output, but also delivered strong

poverty reduction. The Tangguh natural gas project in West Papua delivered

a massive increase in local economic output, but a more muted response to

average household welfare and poverty and a contraction in the agricultural

sector. Together the three case studies highlight how more diffuse natural

resource sectors tend to deliver more broad-based benefits for the local economy

and its residents. For natural resource sectors with highly concentrated rents (e.g.,

11See Dercon (2009) and Dercon and Gollin (2014) for reviews.

§1.6 Methodological contributions 15

natural gas), resource sector growth alone does not appear to deliver economic

development in other sectors or commensurate improvements to local residents’

welfare. The findings from these three case studies contribute to a nascent but

rapidly growing empirical literature on the local economic and poverty impacts

of booming natural resource sectors.12

1.6 Methodological contributions

A focus on sectors

Throughout the thesis, my analytical focus is on the impacts of natural resource

sectors. The literature’s traditional focus has been on resource endowments

exogenously determined by nature, and natural resource exports largely driven by

external demand. A focus on natural resource sectors has greater policy relevance,

as sector size is a function of policy choices and subject to a range of national

and subnational policy instruments. In each chapter I introduce new measures

of natural resource sector size and activity. In Chapter 2, I introduce a new

measure for country-level resource dependence: mining value-added and mining

value-added as a share of GDP. In Chapter 3 I measure district-level palm oil

sector intensity with the share of total district area used for palm oil plantations.

Focusing on land use change allows me to compare the development impacts of

more oil palm land against all alternative uses of land—the opportunity cost. My

three district-level case studies in Chapter 4 define local resource booms as events

where the district economy experiences a sharp and sudden increase in resource

sector output. Examining time series for resource sector output allows me to

identify episodes of resource expansion appropriately classified as dichotomous

treatments for event study.

12See Cust and Poelhekke (2015) and Gamu et al (2015) for reviews.

§1.6 Methodological contributions 16

Empirical strategies

The primary contributions of this thesis are the new empirical results

described above, but I also introduce several methodological innovations. The

empirical method used in Chapter 2 builds on the instrumental variable strategies

of van der Ploeg and Poelhekke (2010) and Carmignani (2013), who instrument

natural resource share of exports with estimated natural resource reserves. I

extend this approach to focus on the mining sectors and explicitly address

potential measurement error in observed historical resource endowments by

controlling for historical exploration effort. My novel proxy for exploration effort

is the total number of exploratory oil and gas wildcats drilled in each country over

the 20th century.

In Chapter 3 I introduce a new instrumental variable strategy to study causal

effects of agricultural sector growth. By exploiting detailed geo-spatial data on

agro-climatic suitability for oil palm for every field in Indonesia, I can identify the

local average partial effect of oil palm expansion arising from purely exogenous

agro-climatic conditions shaping the incentives to develop oil palm plantations.

Data are taken from the Global Agro-ecological Zones database of the Food and

Agricultural Organisation of the United Nations (Fischer et al, 2002). Each pixel

is matched to Indonesian district boundaries. By also controlling for potential

yields of other crops that could share agro-climatic suitability characteristics with

oil palm, I ensure the identifying variation relates only to oil palm and not other

types of agriculture.

An additional empirical innovation in Chapter 3 is use of plausibly exogenous

identifying variation for my panel fixed effects estimates. I exploit the fact that

district heads must apply to the central government for approval to convert

land into oil palm plantations. This generates uncertain (i.e., subject to some

degree of randomness) variation in the timing of approvals within districts

and the outcomes of the decisions across districts. This identification strategy

§1.7 Organisation 17

builds on Burgess et al (2012), who similarly argue that the timing of district

splits through Indonesia’s recent decentralisation is exogenous in a panel data

setting. Convoluted oil palm and forestry regulations across different levels of

government arguably strengthen the case for exogenous timing relative to the

more formalised arrangements for setting up new administrative units (Fitriani

et al, 2005; Resosudarmo, 2005).

My unique application of a relatively new method—the synthetic control

method—in Chapter 4 should also be of demonstrative utility. Sub-national

panel data are becoming increasingly available and policies and decision-making

are commonly devolved to sub-national administrative units. I show how the

synthetic control method can be applied to analyse the impacts of major policies

and economic shocks to single administrative units in the Indonesian context. This

is just the second sub-national application of the synthetic control method to study

the effects of resource shocks and the first in a developing country context.

1.7 Organisation

The thesis has five chapters. Chapters 2–4 present the core research. Chapter

2 examines the long-term health and education impacts of economic dependence

on the mining sector, internationally across countries and across districts in

Indonesia. Chapter 3 turns to diffuse natural resources, studying the poverty

impacts of rapid growth in the palm oil sector in Indonesia. Chapter 4 presents

three district-level case studies of the local economic and welfare impacts of

growth in the palm oil, coal, and natural gas sectors in Indonesia. Chapter 5 briefly

discusses the implications of my findings and outlines some possible directions

for future research.

Chapter 2

Mining away the Preston curve

Abstract

I estimate the long-term national health and education impacts of having a

larger mining share in the economy. By instrumenting the relative size of

the mining sector with the natural geological variation in countries’ fossil fuel

endowments, I provide evidence suggestive of a causal relationship. The

findings suggest that countries with larger mining shares tend to have poorer

health and education outcomes than countries with similar per capita incomes,

geographic characteristics, and institutional quality. Doubling the mining

share of an economy corresponds to, on average, the infant death rate being

twenty percent higher, life expectancy being five percent lower, total years of

education being twenty percent lower, and seventy percent more people having no

formal education. Divergences from the Preston curve—the concave relationship

between cross-country income and life expectancy that has long been of interest

to economists, demographers, and epidemiologists—are thus partly explained by

the size of the mining sector. Within-country evidence from Indonesia paints

a similar picture. My results provide support for a growing body of evidence

linking mining to poorer average living standards, particularly vis-a-vis other

types of income. I also estimate the effects of national mining dependence on

non-mining income, health and education investment, and institutions.

18

§2.1 Introduction 19

2.1 Introduction

Resource-rich Equatorial Guinea had a gross national income of $14, 320 per

capita in 2013, yet more than three-quarters of its population lives below the

poverty line, and life expectancy at birth is 20 years less than other high-income,

non-OECD countries. Africa grew fast on the back of the global commodity boom

in the 2000s, but progress in reducing poverty has been disappointing (World

Bank, 2013). The fate of much of the world’s poor is tied to mining, with at least

half of the world’s known oil, natural gas, and mineral reserves in non OECD,

non-OPEC countries. Resource-driven economies are home to around 70 percent

of the world’s extreme poor.13

The economic and institutional effects of natural resources and a booming

resource sector have been well studied, but evidence on how extractive industries

relate to social outcomes remains thin (see van der Ploeg (2011) and Wick and

Bulte (2009) for reviews). Human capital is typically seen as a channel for

natural resources to stunt economic growth (Gylfason and Zoega, 2006), although

primary commodities can also directly impede social development (Carmignani,

2013). To my knowledge, an international study is yet to focus on mining sector

output nor examine its effects on national health and education outcomes.

In this chapter I compare countries with different structural compositions

to look at the “social productivities” of different types of economic activity,

focusing on fossil fuel extraction. I exploit geological variation in countries’

fossil fuel endowments to identify the long-term effects of mining on health and

education. I find that countries with more mining tend to have poorer health and

education outcomes than countries with similar per capita incomes, geographic

characteristics, and institutional quality. My estimates suggest that doubling the

mining share of an economy corresponds to the infant death rate being twenty

percent higher, life expectancy being five percent lower, total years of education13’Resource-driven economies’ are categorised according to the typology in McKinsey Global

Institute (2014), using World Bank (2014) data.


being twenty percent lower, and seventy percent more people having no formal

education. Within-country evidence from Indonesian districts reveals similar

patterns. Just some types of economic growth (e.g., agriculture) are better at

reducing poverty (Christiaensen et al, 2012), non-mining income is on average

better for health and education than income from the mining sector.

Figure 2.1: Preston Curve, 2005

The findings help to understand divergences from the Preston curve, the

concave relationship between cross-country income and life expectancy that has

long been of interest to economists, demographers, and epidemiologists (Deaton,

2013; Preston, 1975). In Figure 2.1, I plot life expectancy at birth against per capita

income, with countries weighted by contribution of mining to value added. I do

the same for years of education in Figure 2.2. Countries with larger mining sectors

tend to have poorer health and education outcomes than expected at their income

level.


Figure 2.2: Educational Attainment, Income, and Mining, 2005

The chapter proceeds as follows. In Section 2.2, I provide a conceptual

framework linking mining to health and education. Section 2.3 explains the

instrumental variable (IV) strategy used in my main estimates. Section 2.4

presents the national-level results, compares the health and education elasticities

of mining income with income from other sectors, and explores potential

mechanisms. Section 2.5 presents similar evidence from Indonesian districts.

Section 2.6 concludes.

§2.2 Linking mining to health and education 22

2.2 Linking mining to health and education

Why would the mining sector affect a country’s infant mortality rate or life

expectancy? The size of the mining sector can be linked to national health and

education outcomes through three main channels: income and Dutch disease

effects; investment in health and education, by individuals and governments; and

various institutional channels (Figure 2.3, dotted lines and clear boxes).

A larger share of mining in the economy could benefit health and education by

boosting income. While a substantial body of research argues natural resources

can hinder long-term economic growth, evidence remains mixed and recent

studies demonstrate clear long-term positive income effects from the discovery of

large resource wealth stocks (Mideksa, 2013; Smith, 2015). Such discoveries can

lead to significant health improvements (Cotet and Tsui, 2013). Net income effects

arising from a larger mining sector depend on the size of the mining expansion,

its effects on other sectors of the economy, and the distribution of mining and

non-mining income. Dutch disease occurs when resource exports generate large

balance of payments surpluses, appreciating the real exchange rate and increasing

relative prices for non-tradable inputs. Coupling these price and exchange rate

effects with higher demand from a mining boom, other trade-exposed sectors tend

to be less competitive and are often permanently displaced (Corden, 1984). In

extreme cases, booming mining sectors can have similar effects on non-resource

sectors as large tariffs (Gregory, 1976). Because manufacturing and other tradable

sectors tend to more intensively use human and physical capital, booming sector

dynamics often lead to less capital in the economy (Mikesell, 1997). Positive health

and education impacts of a mining-driven income boost are also likely to be offset

by the unequal distribution of new income, as countries with greater primary

commodity dependence tend to have higher inequality, which in turn affects social

development trajectories (Carmignani and Avom, 2010).


Figu

re2.

3:Co

ncep

tual

Fram

ewor

k


The second main channel is human capital investment, which tends to be lower

in mining-dependent countries due to lower expected returns to skills, education,

and knowledge (Blanco and Grier, 2012). At the micro level, a booming mining

sector alters the incentives for human capital development. Trade-exposed

modern sectors are typically more labour and human capital intensive, with

higher wage premiums for educated workers and greater innovation. Conversely,

primary commodity sectors tend to use less skilled labour and have fewer linkages

to other sectors of the economy, effectively taxing human capital if they divert

people and resources away from higher skilled activities (Matsuyama, 1992). For

example, oil resources tend to orient university students towards specialisations

providing better access to resource rents (Ebeke et al., 2015).14 With poorer

micro-level incentives for investment in skills and education, private health

investment is unlikely to respond much differently. Long-term positive spillovers

from natural resources typically hinge on resource revenues strengthening

governments’ fiscal positions, enabling increased investment in health and

education and “spreading the benefits” (Arezki et al, 2011). Volatility—argued by

the van der Ploeg and Poelhekke, (2009) to be the “quintessence of any resource

curse”—-makes this difficult, as short-term political and economic horizons in

volatile economies provide little incentive to prioritise long-term health and

education investments. Rather, commodity price uncertainty often corresponds

to erratic and restrictive public spending and even long-run neglect (Gylfason

and Zoega, 2006; Mikesell, 1997). Acemoglu et al (2013) find the health spending

elasticity of resource-related income is well below one and there remains scant

evidence of resource revenues being converted into effective public investment

(Caselli and Michaels, 2013).

Extractive industries go hand in hand with the extractive institutions

historically associated with poorer health conditions (Acemoglu et al., 2001). The

challenges of managing natural resources tend to be less severe in the presence of14Note that some of these sectors can be highly skilled, or example geology and engineering.


good institutions (Boschini et al., 2013; Farhadi et al., 2015; Mehlum et al., 2006),

but poorer institutions also directly harm health and educational development.

For example, a high level of corruption and poor government effectiveness

can lead to systemic health and education system failure. Resource-related

conflict can not only undermine health and education service delivery and the

incentives governing human capital accumulation, but cause sudden depletions

of human capital stock and long-term health and cognitive impacts (de Soyza

and Gizelis, 2013; Lei and Michaels, 2014; Williams, 2011). Political institutions

in mining-focused economies provide little incentive for broad-based human

capital investment. Increasing the average level of education and facilitating

the growth of an urban middle class undermines elites’ control of rents: a

dynamic underwritten by weaker democratic accountability in resource exporters

and reflected in their generally poorer health outcomes (Besley and Kudamatsu,

2006; Ross, 2001; Sokoloff and Engerman, 2000; Tsui, 2011). Lastly, rent

holders often use their political power to promote sub-optimal policies, resisting

industrialisation (and urbanisation) and reinforcing any Dutch disease and

investment effects (Auty, 1997).

A potential mechanism receiving attention in emergent work on the local

impacts of mining also bears a mention: pollution (see Cust and Poelhekke

(2015) for a review). Pollutants emitted from mining operations are some

of the most toxic, associated with premature births, lower birth weights and

weight-for-height ratios, stunting, anaemia, increased respiratory illness, malaria,

and less intelligence (Factor-Litvak et al, 1999; Iyengar and Nair, 2000; 2014; Saha

et al., 2011). Aragon and Rud (2015) study the impacts of 12 gold mines in

Ghana on local agricultural production, finding large decreases in productivity

and greater poverty likely to stymie human capital development (e.g., through

nutrition). While pollution mostly affects exposed communities (i.e., unlikely

to drive any national-level effects), cumulative impacts on cognitive and other

§2.3 Empirical approach 26

long-term development outcomes could be significant; it would be naive to

rule out their potential role in shaping national social development outcomes,

particularly in smaller economies.

In this chapter I focus on the reduced-form effects of mining on health

and education (solid lines, shaded boxes in Figure 2.3), holding income and

institutions constant (dotted lines, clear boxes). I focus on the mining share of

income and estimate partial elasticities, to compare the long-term health and

education effects of mining income with income from other sectors. I then extend

this approach to empirically test some potential channels outlined in this section.

2.3 Empirical approach

I relate the mining share of the economy to national health and education

outcomes with the equation:

ln(yc) = αln(Mc) + βX ′c + uc (2.1)

where yc is the general health conditions or educational attainment in country c;

M is the percentage contribution of mining to value-added; and X ′ includes per

capita gross domestic product (GDP), the absolute distance from the equator, an

index of institutional quality, and the total number of wildcats drilled in the 20th

century (exploratory drilling, known as wildcat drilling because of wild cats seen

in remote areas explored in the early 20th century). Standard errors are adjusted

for arbitrary heteroskedasticity. Logs account for skewness inM , y, and per capita

GDP, and provide convenient partial elasticity interpretations: when a country’s

mining share of total value-added increases by one percent, health and education

indicators are expected to rise or fall by α percent in the long run, holding all else

constant.


National health and education indicators are some of the slower moving

cross-country variables and exhibit strong serial correlation. The timing of

potential effects is difficult to correctly specify due to the long and non-uniform

lags in the system governing human capital accumulation, but doing so is critical

to detect any average effects and understand long-run magnitudes. My preferred

estimator is the classic “between” approach, ignoring short-run dynamics (unlike

static fixed-effects models) and providing long-run effects (Baltagi and Griffin,

1984; Burke and Nishitateno, 2015; Stern, 2010). By exploiting the long-run

equilibrium differences between countries in a large international cross-section in

2005, I obtain a natural long-run interpretation, retain the cross-country variation

of interest, and make few ad-hoc timing assumptions.

2.3.1 Instrumental variable strategy

Associative relationships between mining and health and education outcomes

cannot be taken as conclusive evidence more mining causes poorer health

and education outcomes. Countries with less human capital might tend

towards the primary sectors, “selecting” into mining (Alexeev and Conrad, 2009;

Brunnschweiler and Bulte, 2008). Estimation with ordinary least squares (OLS)

could lead to biased and inconsistent estimates, e.g., due to measurement error,

reverse causality, or correlation with other unobservable factors. For example,

economic activity in the mining sector may be affected by human capital stock and

capabilities, suggesting potential reverse causality through the supply of mining

engineers, exploration abilities, mining technology, cost competitiveness, and

capabilities in other sectors: factors mostly unobservable and difficult to control

for. National-level mining is affected by domestic policy settings and the decisions

of people and firms, so strong exogenous variation is needed to identify causal

effects.


My identification strategy exploits the fact that countries must be endowed

with natural resources before they can have a mining sector. I instrument the

contribution of mining to value-added in 2005 with national per capita fossil fuel

reserves (i.e., coal, oil, and gas) deeply lagged to 1971 to allow sufficient time to

evolve into mining sector activity.15

Instrument relevance and strength

Instrumenting mining share with deeply lagged per capita fossil fuel reserves

has important implications for the interpretation of my estimates. I estimate

the local average partial effect of a country having a larger mining share due to

greater initial fossil fuel reserves, i.e., relating more to dependence on coal mining

and oil and gas extraction than mineral extraction. First-stage coefficients are

positive and statistically significant at the one percent level, with around half

of the variation in 2005 mining shares explained by 1971 per capita fossil fuel

reserves. The combination of per capita coal, oil, and gas reserves provides the

broadest commodity coverage and strongest identification.16

A weak IV problem can be present even with highly significant first-stage

coefficients, so I provide weak IV diagnostics with all results. I report the

Kleibergen-Paap (2006) rk Wald F statistic with all IV estimates with the highest

Stock-Yogo (2005) critical value of 16.38 for one endogenous regressor, a single

instrument, and 10% maximal IV size. An excluded-F statistic greater than

10 is the more common benchmark for sufficient instrument strength (Staiger

and Stock, 1997). With a large international sample and a single strong IV, I

obtain consistent long-run causal estimates with the smallest bias if the exclusion

restriction is satisfied.15Using 1971 reserves as an instrument for the mining sector in a between-country setting is

the most appropriate use of this instrument, as it does not provide any temporal variation. As mydependent variable is slow-moving and my IV from 1971, I am also not overly dependent on anymoment in time. Estimates from alternative time periods and using country averages are similarand provided in the Chapter Appendix.

16By contrast, national mineral reserves in the 1970s are large in scale but only weakly correlatedwith mining dependence: an empirically irrelevant potential instrument.


Instrument validity

For the exclusion restriction to be satisfied, proven reserves in 1971 can have

no relationship with the dependent variables except through the mining sector,

holding other controls constant. If the geological allocation of fossil fuel reserves is

a product of nature and luck—as argued by Carmignani (2013) and van der Ploeg

and Poelhekke (2010)—then it is orthogonal to the unobservable and unmalleable

factors affecting my outcomes and the exclusion restriction is satisfied. Causal

pathways other than mining would be ruled out by design: the primary channel

for sub-soil reserves in 1971 to affect socioeconomic outcomes since 1971 is

through the current, past, and future size of the mining sector, with future size

likely to be a function of current size and endowments.17

That reserves are proven however gives them a non-random component.

While countries cannot decide and enact policies to create fossil fuel endowments,

they can enact policies that may increase the likelihood unknown resource

endowments will become known. If reserves are random but measured with error

(i.e., depending on exploration effort, income, institutions, and other historical

factors) then the exclusion restriction is valid conditional on including the factors

explaining any systematic measurement error in the reserves. By controlling

for per capita income, institutions, geography, and exploration effort, I seek

to isolate the natural geological component of measured reserves. The critical

identification assumption is that my covariates control for all relevant omitted

variables and systematic error: not an implausible assumption, but impossible

to prove in observational studies with external instruments. I present two

additional pieces of evidence suggesting the probability of my main identification

17Bazzi and Clemens (2013) show how many popular IVs can only be valid in one applicationdue to possible violation of the exclusion restriction, and highlight the common confusion betweenexogeneity and orthogonality in applied work (e.g., rainfall, price, geographical, and laggedinstruments). The exclusion restriction for a given instrument can only be not rejected afterconsidering all previous uses, then providing a convincing argument that: (a) past instrumentedexplanatory variables are part of the same causal chain; or (b) past usage is invalid. van der Ploegand Poelhekke (2010) and Carmignani (2013) use Norman (2009) reserves to instrument naturalresource exports, but mining comes before exporting in the same causal chain.


assumption being violated is likely to be low. Firstly, I adopt the common

heuristic that coefficient stability to additional control variables can be informative

about potential omitted variable bias (Altonji et al., 2005; Bellows and Miguel,

2009) and conduct sensitivity analysis using a wide range of variables plausibly

correlated with potential measurement error in fossil fuel reserves. Secondly, as

unobservable country-specific factors cannot be ruled out, I present analogous

within-country evidence from Indonesia.

2.3.2 Data

My health dependent variables are the mortality rate for infants aged one year

and below (per thousand births) and life expectancy at birth from the World

Bank (2014). Education variables are the average years of education attained

and the percentage of the population with no formal schooling for the total

population and youth from Barro and Lee (2010). The poorest performers in

terms of infant health conditions and years of schooling tend to be equatorial

low-income countries with poorer institutions (see maps in Chapter Appendix).

I estimate impacts on different indicators separately to allow for heterogeneity

across indicators and a more precise interpretation (c.f., interpreting effects on

composite indexes).

My explanatory variable of interest is the contribution of mining to

value-added, taken from the United Nations (UN) Environmental Indicators and

available for 1995–2008. Mining is defined following International Standard

Classification 0509 and includes the extraction of coal, lignite, metal ores, crude

petroleum and natural gas, as well as mining support services (UN, 2013). Mining

value-added is more useful than previous proxies for natural resources—namely,

exports (Sachs and Warner, 2001), estimated natural capital and resource rents

(Brunnschweiler and Bulte, 2008), and commodity prices (Collier and Goderis,

2012)—because it explicitly captures economic activity in the mining sector and


Figu

re2.

4:Co

ntri

buti

onof

Min

ing

toVa

lue-

Add

ed,2

005


Figu

re2.

5:Pe

rC

apita

Foss

ilFu

els,

1971


is affected by policy choices. To my knowledge, the UN data are only publicly

available national-level data for mining sector output across countries and not

yet used in any resource curse studies. Figure 2.4 illustrates the contribution of

mining to value-added around the world in 2005, with noticeable variation across

continents and income groups.

My national per capita fossil fuel reserves IV is taken from Norman (2009),

and generally regarded as some of the more reliable and exogenous of the

different natural resource stock measures currently available (van der Ploeg and

Poelhekke, 2010; Carmignani, 2013). Norman (2009) constructs her dataset by

adding extraction of oil, coal, natural gas, and minerals from 1970—2001 to proven

reserves in 2002. I convert Norman’s original measures into per capita terms

and combine the oil, coal, and natural gas components to obtain per capita fossil

fuel reserves in 1971, in 1971 prices, presented in Figure 2.5. Heavily-endowed

countries are from all continents and income groups, and not necessarily the

countries with the largest mining shares (c.f., Figure 2.4). The conversion of large

fossil fuel reserves to large mining shares is an outcome of choices.

Per capita GDP controls for a country’s income and level of development,

and is taken from the Penn World Tables in purchasing power parity, 2005

constant prices, and in per capita terms to scale for population size (Heston

et al., 2012). Latitude (i.e., the absolute distance from the equator) controls

for countries’ location, remoteness, the effects of tropical diseases, and regional

trade effects (Sala-i-Martin et al., 2004).18 My preferred proxy for institutional

quality is the World Bank’s government effectiveness index, strongly correlated

to traditional measures of institutions, but also picking up service delivery

capabilities (Kaufman et al., 2013). Lastly, I include the total number of wells

drilled in areas where no oil production exists (wildcats), summed over the

20th century (Cotet and Tsui, 2013). Wildcat drilling is a useful proxy for18Regional fixed effects yield similar results, but latitude achieves the same purpose more

parsimoniously and allows for a stronger first-stage in the IV estimates. Results with regionalfixed effects are provided in Column 1 of Table 2.3.


country-specific mining exploration effort, investment, history, and technological

capabilities.

My parsimonious set of controls explain substantial international variation in

health and education outcomes, and has implications for interpreting the mining

coefficients. The estimated partial effect of mining due to higher initial fossil

fuel endowments is from economies of comparable levels of GDP per capita,

location, institutional quality, and exploration effort, excluding effects through

these channels: I compare whether countries at a given income level have better

health and education outcomes because of the structure of their economies.

Summary statistics for mining share, per capita fossil fuel reserves, and key

dependent variables in 2005 are presented in Table 2.1. Column 1 includes the full

sample and the remaining columns split the sample by countries whose mining

share in GDP is larger or smaller than thirty percent. Health and education

outcomes tend to be worse in high-mining countries (Column 2).


Tabl

e2.

1:Su

mm

ary

Stat

istic

s

Varia

ble

All

coun

trie

sH

igh-

min

ing

Low

-min

ing

Col

umn

12

3

Con

trib

utio

nof

min

ing

tova

lue-

adde

d(p

erce

nt)

Mea

n0.

090.

520.

04

Std.

dev.

0.16

0.14

0.06

Cou

ntrie

s18

819

163

Mor

talit

yra

te,i

nfan

t(no

.per

’000

)M

ean

33.5

638

.61

32.4

4

Std.

dev.

30.8

432

.57

30.4

4

Cou

ntrie

s19

335

158

Ave

rage

year

sofs

choo

ling

atta

ined

Mea

n7.

87.

267.

87

Std.

dev.

2.74

1.98

2.83

Cou

ntrie

s13

917

122

Perc

enta

geof

popu

latio

nw

ithno

scho

olin

gM

ean

16.8

21.6

716

.12

Std.

dev.

18.8

216

.48

19.0

8

Cou

ntrie

s13

917

122

Valu

eof

perc

apita

foss

ilfu

elre

serv

esin

1971

USD

)M

ean

20.8

112.

391.

78

Std.

dev.

119.

5227

3.95

5.37

Cou

ntrie

s15

727

130

Hig

hm

inin

gre

fers

toco

untr

iesw

itha

cont

ribut

ion

ofm

inin

gto

valu

e-ad

ded

ofov

er30

perc

ent.

Low

min

ing

refe

rsto

coun

trie

sund

er30

perc

ent.

Sam

ple

isal

lava

ilabl

eco

untr

iesi

n20

05

§2.4 Results 36

2.4 Results

2.4.1 Health and education effects of mining sector growth

My main results are presented in Table 2.2. Panel A presents results for

children and youth and Panel B for the total population. OLS and IV estimates are

presented for all outcomes. Coefficients for mining are statistically significant in

all IV estimates, and the signs and magnitudes suggest the prominence of mining

contributes to large international differences in health and education outcomes

across countries.

In Panel A of Table 2.2, Columns 1 and 2 look at infant mortality. The estimated

coefficients in Column 2 suggest doubling the share of mining in the economy

related to greater initial fossil fuel endowments corresponds to, on average, an 18

percent increase in the infant mortality rate in the long run. Doubling the size of

the mining sector in a country with the mean infant mortality rate, i.e., 34 per 1000,

is expected to lead that country to settle at an infant mortality rate of 40 per 1000,

holding all else constant. The OLS estimate in Column 1 is slightly biased towards

zero, as expected.19 The rest of Panel A in Table 2.2 reports the effect of mining

on national educational outcomes for youth. Column 3 presents the OLS estimate

on average years of educational attainment: a small and statistically insignificant

negative effect. The IV estimate in Column 4 is less ambiguous. Doubling the

share of mining due to greater initial fossil fuel endowments leads to around

21 percent fewer years of schooling, or a reduction in educational attainment

from 7.7 to 6.2 years in the mean-educated country. Columns 5 and 6 look at

the percentage of the young people with no education, picking up completely

19Coefficients on mining share obtained from my IV estimates in Table 2.2 are best consideredlower bounds for four reasons. Firstly, I control for (hold constant) key potential channels.Secondly, IV estimates in finite samples tend to be biased the same direction as OLS (downwards,in this case). Third, between estimation detects only contemporaneous and long-run effects,making it difficult for treatment to affect the outcome variable in countries with recent miningexpansions. Finally, estimates on cohort-specific indicators tend to underestimate long-termpopulation effects, although I have used extensive-margin indicators (e.g., infant mortality, youtheducational attainment) to minimise this problem.

§2.4 Results 37

Tabl

e2.

2:H

ealt

han

dEd

ucat

ion

Effe

ctso

fMin

ing

Dep

ende

nce

Pane

lA:E

ffect

son

child

ren

and

yout

hLo

gde

pend

entv

aria

ble

Infa

ntm

ort.

(dea

thsp

er’0

00)

Avg

.yrs

educ

.,yo

uth

Perc

enta

geof

pop.

w/

noed

uc.,

yout

hEs

timat

orO

LSIV

OLS

IVO

LSIV

Col

umn

12

34

56

Log

min

ing

shar

e0.

108*

**0.

181*

**-0

.018

-0.2

08**

*0.

260*

**1.

095*

**(0

.018

)(0

.04)

(0.0

12)

(0.0

47)

(0.0

77)

(0.2

46)

Log

real

GD

Ppe

rcap

ita-0

.553

***

-0.5

94**

*0.

216*

**0.

338*

**-0

.800

***

-1.3

20**

*(0

.044

)(0

.064

)(0

.042

)(0

.072

)(0

.159

)(0

.28)

Latit

ude

-0.7

27**

*-0

.435

0.24

1*-0

.239

-1.8

98**

1.41

5(0

.199

)(0

.28)

(0.1

27)

(0.3

71)

(0.8

32)

(1.6

95)

Gov

ernm

ente

ffect

iven

ess

-0.2

13**

*-0

.171

**-0

.045

-0.2

05**

0.33

90.

838*

*(0

.059

)(0

.085

)(0

.034

)(0

.087

)(0

.228

)(0

.405

)

Wild

cats

(*10

00)

0.04

0***

0.03

0***

0.00

50.

010*

-0.1

40**

*-0

.190

***

(0.0

05)

(0.0

06)

(0.0

03)

(0.0

09)

(0.0

3)(0

.05)

Firs

t-sta

ge

Perc

apita

foss

ilfu

els1

971

0.00

3***

0.00

3***

0.00

3***

(0.0

01)

(0.0

01)

(0.0

01)

Excl

uded

-F16

.43

16.7

216

.72

Cou

ntrie

s15

113

312

711

212

611

2Pa

nelB

:Effe

ctso

nth

etot

alpo

pula

tion

Log

depe

nden

tvar

iabl

eLi

feex

p.(y

rs)

Avg

.yrs

educ

.Pe

rcen

tage

ofpo

p.w

/no

educ

.Es

timat

orO

LSIV

OLS

IVO

LSIV

Col

umn

78

910

1112

Log

min

ing

shar

e-0

.012

**-0

.042

***

-0.0

21*

-0.1

99**

*0.

120*

*0.

671*

**(0

.005

)(0

0.01

)(0

.013

)(0

.05)

(0.0

57)

(0.1

89)

Log

real

GD

Ppe

rcap

ita0.

010*

**0.

123*

**0.

242*

**0.

352*

**-0

.464

***

-0.7

89**

*(0

.012

)(0

.018

)(0

.045

)(0

.076

)(0

.12)

(0.2

09)

Latit

ude

0.07

20.

050.

351*

*-0

.192

-3.9

94**

*-1

.526

(0.0

46)

(0.0

8)(0

.139

)(0

.379

)(0

.619

)(1

.269

)

Gov

ernm

ente

ffect

iven

ess

-0.0

07-0

.039

-0.0

58-0

.189

**0.

128

0.43

5(0

.015

)(0

.025

)(0

.039

)(0

.089

)(0

.181

)(0

.302

)

Wild

cats

(*10

00)

-0.0

03**

-0.0

010.

010*

**0.

020*

**-0

.110

***

-0.1

50**

*(0

.001

)(0

.002

)(0

.003

)(0

.01)

(0.0

2)(0

.04)

Firs

t-sta

gePe

rcap

itafo

ssil

fuel

s197

10.

003*

**0.

003*

**0.

003*

**(0

.001

)(0

.001

)(0

.001

)Ex

clud

ed-F

16.5

216

.55

16.5

5C

ount

ries

150

132

122

107

122

107

Star

sde

note

stat

istic

alsi

gnifi

canc

eat

the

10,5

,and

1pe

rce

ntle

vels

.Sa

mpl

eis

the

larg

estp

ossi

ble

2005

cros

s-se

ctio

nfo

rth

ese

varia

bles

.H

eter

oske

dast

icity

-rob

usts

tand

ard

erro

rsin

pare

nthe

ses,

and

coeffi

cien

tson

cons

tant

san

dfir

st-s

tage

cont

rols

are

not

repo

rted

.IV

estim

ates

inst

rum

entl

ogm

inin

gsh

are

with

perc

apita

foss

ilfu

elre

serv

esin

1971

.The

rele

vant

criti

calv

alue

fort

heK

leib

erge

n-Pa

apW

alk

rkF

stat

istic

srep

orte

d(i.

e.ex

clud

ed-F

)ist

heSt

ock-

Yogo

criti

calv

alue

of16

.38c

alcu

late

dfo

rone

endo

geno

usre

gres

sor,

one

inst

rum

ent,

10%

max

imum

IVre

lativ

ebi

as,a

ndi.i

.der

rors

.

§2.4 Results 38

discouraged, disengaged, or excluded young people. The OLS coefficient in

Column 5 is positive and statistically significant at the one percent level, rising to

1.095 in Column 6. Single equation estimates not accounting for the endogenous

response of mining to human capital appear to tend towards zero, consistent with

earlier conflicting findings (Gylfason 2001; Stijns, 2006).

In Panel B, I turn to life expectancy and educational attainment indicators for

the whole population. Columns 7 and 8 show countries with more mining in their

economies tend to have lower life expectancies than other countries at a similar

income level. Results for years of education in Columns 9 and 10 of Panel B

are similar to youth, as expected exploiting between country variation. Results

for the percentage of the population (Columns 11 and 12) with no education

are of a lesser magnitude than for youth, although economically quite large.

Applied to the sample mean of 17 percent, doubling the mining share results

in the uneducated share of the population rising to almost 30 percent in the

long run. The Kleibergen-Paap rk Wald F statistic of 16.43 exceeds the highest

Stock-Yogo critical value (16.38) across all estimates in Table 2.2, providing no

evidence of a weak IV. Per capita income is the most important control, statistically

significant at the one percent level and of a similar magnitude in all estimates.

Coefficients on constants, controls, and the excluded instrument are not reported

for the remainder of this chapter.

§2.4 Results 39

Tabl

e2.

3:Se

nsit

ivit

yA

naly

sis

1970

cont

rol

Regi

onFE

sLo

gG

DP

p.c.

Log

IMR

Log

year

sedu

c.D

emoc

racy

Exec

.con

st.

Pres

.C

olon

yFr

.leg

alLo

gse

t.mor

t.A

llN

orm

anIV

ASP

OIV

Col

umn

12

34

56

78

910

1112

Pane

lA:L

ogin

fant

mor

talit

y(d

eath

sper

’000

)

Log

min

ing

shar

e0.

15**

*0.

17**

*0.

14**

*0.

17**

*0.

17**

*0.

18**

*0.

21**

*0.

17**

*0.

19**

*0.

14*

0.16

***

0.08

**(0

.05)

(0.0

5)(0

.05)

(0.0

5)(0

.04)

(0.0

4)(0

.05)

(0.0

3)(0

.03)

(0.0

8)(0

.04)

-0.0

4Ex

clud

ed-F

2.32

58.

5110

.45

10.3

54.

665.

079.

785.

8416

.15

23.7

414

.23

12.1

3C

ount

ries

133

133

121

107

109

109

130

114

136

7513

312

5Pa

nelB

:Log

aver

agey

ears

ofed

ucat

iona

latta

inm

ent

Log

min

ing

shar

e-0

.16*

*-0

.13*

*-0

.15*

*-0

.03

-0.1

0*-0

.10*

-0.1

3***

-0.1

3***

-0.1

3***

-0.0

1-0

.13*

**-0

.06*

(0.0

8)(0

.07)

(0.0

6)(0

.05)

(0.0

6)(0

.05)

(0.0

4)(0

.05)

(0.0

3)(0

.03)

(0.0

5)(0

.03)

Excl

uded

-F3.

1510

.11

10.9

510

.35

5.94

6.47

13.1

65.

6116

.46

17.8

814

.17

11.9

4C

ount

ries

107

107

102

107

9494

108

9710

960

107

100

Pane

lC:L

ogpe

rcen

tage

ofpo

pula

tion

with

noed

ucat

ion

Log

min

ing

shar

e0.

51*

0.35

*0.

42*

0.23

0.21

***

0.21

***

0.61

***

0.52

***

0.59

***

0.18

0.46

**0.

32**

(0.2

8)(0

.19)

(0.2

3)(0

.20)

(0.0

8)(0

.08)

(0.1

9)(0

.19)

(0.1

5)(0

.17)

(0.1

9)(0

.13)

Excl

uded

-F3.

1510

.11

10.9

510

.35

5.94

6.47

13.1

65.

6116

.46

17.8

814

.17

11.9

4C

ount

ries

107

107

102

107

9494

108

9710

960

107

100

Star

sden

otes

tatis

tical

sign

ifica

ncea

tthe

10,5

,and

1pe

rcen

tlev

els.

Sam

plei

sthe

larg

estp

ossi

ble2

005

cros

s-se

ctio

nfo

rthe

seva

riabl

es.H

eter

oske

dast

icity

-rob

usts

tand

ard

erro

rsar

ein

pare

nthe

ses,

and

coeffi

cien

tson

cons

tant

s,co

ntro

ls,a

ndfir

st-s

tage

regr

essi

ons

are

notr

epor

ted.

All

estim

ates

incl

ude

log

real

GD

Ppe

rca

pita

,abs

olut

edi

stan

cefr

omth

eeq

uato

r,an

inde

xof

gove

rnm

ent

effec

tiven

ess,

and

the

tota

lnum

ber

ofw

ildca

tsdr

illed

inth

e20

thce

ntur

yas

cont

rols

.Ex

tra

cont

rols

are

allv

alue

sin

1970

,exc

eptf

orse

ttler

mor

talit

y.A

lles

timat

esus

eth

elim

ited

info

rmat

ion

max

imum

likel

ihoo

d(L

IML)

estim

ator

with

aFu

llerp

aram

eter

ofon

e.C

olum

ns1–

10in

stru

men

tlog

min

ing

shar

ew

ithpe

rcap

itafo

ssil

fuel

rese

rves

in19

71,C

olum

n11

uses

the

sum

ofpe

rcap

itaoi

l,ga

s,co

al,a

ndm

iner

alre

serv

esfr

omN

orm

an(2

009)

,i.e

.,in

clud

ing

min

eral

s,an

dC

olum

n12

uses

actu

ales

timat

edoi

l“en

dow

men

t”fr

omC

otet

and

Tsui

(201

3).T

here

leva

ntcr

itica

lval

uefo

rth

eK

leib

erge

n-Pa

apW

alk

rkF

stat

istic

srep

orte

dis

the

Stoc

k-Yo

gocr

itica

lval

ueof

16.3

8ca

lcul

ated

foro

neen

doge

nous

regr

esso

r,on

ein

stru

men

t,10

%m

axim

umIV

rela

tive

bias

,and

i.i.d

erro

rs.

§2.4 Results 40

2.4.2 Robustness

If the exclusion restrictions for my IV estimates are valid, potential omitted

variable problems are overcome. Satisfying the exclusion restriction relies on

the assumption that the covariates control for all relevant omitted variables

and systematic error in measurement of observed reserves. Table 2.3 presents

IV estimates with additional controls arguably related to potential error in

observed fossil fuels. I use the median-unbiased limited information maximum

likelihood (LIML) estimator with a Fuller parameter of one for improved inference

with weaker instruments (Fuller, 1977). Column 1 includes regional fixed

effects, Columns 2–4 deeply lagged values of income, health, and education,

and Columns 5–10 institutional factors potentially affecting the detection and

measurement of reserves. Columns 1–10 of Panel A show how further restricting

the variation in reserves exploited has little effect on the size or direction of the

infant mortality estimates. Column 11 instruments mining share with total per

capita reserves from Norman (2009), i.e., including minerals. Results are similar,

but with weaker identification. Column 12 shows that a similar result is obtained

instrumenting mining share with an alternative measure of oil endowment from

the Association for the Study of Peak Oil and Gas, estimated by geologists and

factoring in cumulative discovery and local geological conditions (Tsui, 2011;

Cotet and Tsui, 2013).

Panels B and C of Table 2.3 present sensitivity analysis for the main education

estimates. Including a deep lag for years of education (Column 4) renders

the main result insignificant for both indicators; the global mining sector was

already developed in 1971 and the distribution of years of schooling across

countries is persistent and has not changed significantly since. In Panel B, other

initial controls and institutional indicators do not significantly affect the result

for years of schooling, except for settler mortality, which halves the sample.

Instrumenting mining share with total per capita mineral, oil, coal, and natural

§2.4 Results 41

gas reserves (Column 11) gives a similar result, and the coefficient obtained

from instrumenting with the alternative oil endowment measure (Column 12) is

less significant but of the same sign. In Panel C of Table 2.3, estimates for the

percentage of the population with no schooling are similar to those for years of

schooling. Second-stage coefficients are relatively stable to the inclusion of a wide

range of additional covariates, suggesting further omitted variables are unlikely

to have any major effects on my results. Further robustness checks are provided

in the Chapter Appendix, including estimates using alternative periods, country

averages, alternative measures of the size of the mining sector, and dropping

resource-rich mini-states and other outliers.

2.4.3 Health and education elasticities of income, by sector

Until now I have focused on the share of mining in the economy, holding

per capita incomes constant. The net effects of mining are ambiguous if we

move from partial to general equilibrium, as crowding-out may be more than

one-for-one (e.g., as in the Dutch disease case), or learning-by-doing and upstream

spillovers may occur (as was the case for Norway (Torvik; 2001; Mideksa, 2013)).

I now compare the health and education elasticities of mining income with

income from other sectors by extending my main approach to level terms. I

multiply mining share by total per capita income in PPP terms to obtain per capita

mining income deflated by the economy-wide GDP deflator (c.f., sector-specific

deflators). I replace the per capita GDP control with per capita non-mining

income to hold income from other sectors constant. Per capita agricultural income,

non-agricultural income, manufacturing income, and non-manufacturing income

are constructed with the same UN (2014) national accounts to compare across

sectors. I instrument mining income with per capita fossil fuel reserves in

1971 and agricultural income with per capita arable land (World Bank, 2014).

Manufacturing is not instrumented.

§2.4 Results 42

The long-term health and education elasticities of income from mining,

agriculture, and manufacturing are presented in Figure 2.6.20 The bars represent

point estimates and the whiskers 95 percent robust confidence intervals. With

narrow confidence intervals, mining income, on average, appears to be of

significant net harm to long-term health and education outcomes when income

from other sectors is held constant. Coefficients on agricultural income show

imprecise negative effects, and non-mining and manufacturing income are

associated with better health and education outcomes. The large gap between

non-mining income and manufacturing income suggests services likely play a

major role (services are not a clear category in UN national accounts data, so

omitted from the analysis). When it comes to mining income, wealthier is

not necessarily healthier (c.f., Pritchett and Summers, 1996). Income from the

non-mining sectors tends to be better for health and educational development.

20The formal regression results behind each of the bars in Figure 2.6 are presented in theChapter Appendix. Bars represent the coefficients on sectoral income with robust 95 percentconfidence intervals and the standard covariates are included in all estimates. Mining andagricultural income coefficients are obtained from IV estimation and manufacturing from OLS.Mining is instrumented with per capita fossil fuel reserves in 1971 and agriculture with log percapita arable land, both strong instruments exceeding the highest Stock-Yogo critical value. Sampleperiod is 2005.

§2.4 Results 43Fi

gure

2.6:

Hea

lth

and

Educ

atio

nEl

asti

citi

esof

Inco

me,

BySe

ctor

§2.4 Results 44

2.4.4 Heterogeneity

Torvik (2009) argues “the most interesting aspect of the paradox of plenty is

not the average effect of natural resources, but its variation. For every Nigeria

or Venezuela there is a Norway or Botswana.” Two main types of heterogeneity

could be masked by my main results: different effects among those countries

with fossil fuel reserves, and different effects for different types of primary

commodities. I estimate my main results by sub-samples to explore variation by

region and institutional characteristics used to explain winners and losers in the

natural resource lottery (Torvik, 2009). Health results are consistent across regions

and institutions, but education results are more varied (see Chapter Appendix).

Estimated coefficients turn positive or insignificant in parliamentary democracies,

in countries with less than 80 percent of the population of one faith, and excluding

the Middle East and Africa, suggesting this region drives the main education

results.

My main IV estimates identify the local average partial effects of mining related

to fossil fuel abundance. While including minerals gives similar results (Table 2.3),

oil and natural gas are special point resources known to exacerbate the resource

curse. To explore potential heterogeneity in mining dependence related to each

type of reserves, I instrument mining share with separate per capita oil, gas,

mineral, and coal reserves. Oil and gas extraction have stronger negative health

and education effects than the main estimates and coal mining appears harmful

for health (see Chapter Appendix). Conclusions cannot be drawn in relation to

mining dependence arising from mineral abundance due to weak identification.

§2.4 Results 45

2.4.5 Potential channels

I now empirically test for direct links between mining dependence and

non-mining income, human capital investment, and institutions (dotted lines

and clear boxes in Figure 3). I use the same IV approach, with proxies for

these potential channels as dependent variables. Table 2.4 presents the results.

Column 1 of Panel A shows that doubling the mining share corresponds to a 10

percent reduction in the level of non-mining income. The non-mining economy

is responsible for, on average, 11 times more value-added than mining, making

magnitudes economically significant, particularly given the opportunity cost of

other welfare-improving modern sectors (UN, 2013). Columns 2 and 3 find more

mining corresponds to countries investing less in health per capita and as a share

of total government spending. Columns 4 and 5 provide no evidence a larger

mining share leads to less investment in education per capita or as a share of total

government spending, perhaps reflecting the recent policy shift by resource-rich

countries towards education investment and contradicting prior studies showing

that natural resources divert financial resources away from education (Gylfason,

2001).

In Panel B of Table 2.4 I examine governance-related institutional indicators:

government effectiveness, corruption, and gender equality. Columns 6 and

7 show that countries with a larger mining share score lower on indices of

government effectiveness and control of corruption (Kaufman et al, 2013).

Column 8 shows Transparency International’s Corruption Perceptions Index is

also considerably lower in mining-dependent countries. Gender equality has

been a focal area of modern governance reforms and anti-corruption policies, and

resource wealth could allow the persistence of cultural norms altering women’s

roles in society. In Columns 9 and 10 (Panel B) I look at gender equality, estimating

the effect of a larger mining share on the proportion of seats held by women

in national parliaments in Column 9 and on the World Bank’s CPIA gender

§2.4 Results 46

Tabl

e2.

4:N

atio

nal-

leve

lM

echa

nism

s

Pane

lA:I

ncom

ein

othe

rsec

tors

and

publ

ichu

man

capi

tali

nves

tmen

tLo

gde

pend

entv

aria

ble

Perc

apita

non-

min

ing

inco

me

Perc

apita

heal

thex

pend

iture

Hea

lthex

pend

iture

shar

ePe

rcap

itaed

ucat

ion

expe

nditu

reEd

ucat

ion

expe

nditu

resh

are

Col

umn

12

34

5

Log

min

ing

shar

e-0

.09*

**-0

.22*

*-0

.16*

-0.0

80.

03(0

.04)

(0.0

9)(0

.09)

(0.0

8)(0

.11)

Excl

uded

-F16

.37

16.4

316

.43

17.5

18.6

4C

ount

ries

133

133

133

108

122

Pane

lB:Q

ualit

yof

gove

rnm

enta

ndge

nder

equa

lity

Dep

ende

ntva

riabl

eG

over

nmen

teffe

ctiv

enes

sC

ontr

olof

corr

uptio

nC

orru

ptio

npe

rcep

tions

Fem

ale

parli

amen

taria

nsC

PIA

gend

erC

olum

n6

78

910

Log

min

ing

shar

e-0

.17*

**-0

.14*

**-0

.31*

**-4

.90*

**-0

.17*

**(0

.04)

(0.0

4)(0

.10)

(1.1

0)(0

.03)

Excl

uded

-F16

.05

16.7

318

.09

6.07

71.6

1C

ount

ries

134

137

103

132

55Pa

nelC

:Con

flict

and

stab

ility

Dep

ende

ntva

riabl

eYe

arso

fcon

flict

Stab

ility

inde

xPo

lity

IVsc

ore

Num

bero

fcou

psLo

gm

ilita

rysh

are

Col

umn

1112

1314

15

Log

min

ing

shar

e0.

29-0

.08*

-5.5

2***

0.74

***

0.44

***

(0.2

4)(0

.04)

(1.2

0)(0

.26)

(0.1

0)Ex

clud

ed-F

16.7

316

.73

13.2

316

.73

5.4

Cou

ntrie

s13

713

711

913

711

6

Star

sden

otes

tatis

tical

sign

ifica

ncea

tthe

10,5

,and

1pe

rcen

tlev

els.

Sam

plei

sthe

larg

estp

ossi

ble2

005

cros

s-se

ctio

nfo

rthe

seva

riabl

es.H

eter

oske

dast

icity

-rob

usts

tand

ard

erro

rsin

pare

nthe

ses,

and

coeffi

cien

tsco

nsta

nts,

cont

rols

,and

first

-sta

gere

gres

sion

sare

notr

epor

ted.

Pane

lAco

ntro

lsin

clud

elo

gre

alG

DP

perc

apita

,lat

itude

,gov

ernm

ente

ffect

iven

ess,

and

wild

cats

drill

edin

the

20th

cent

ury.

Pane

lsB

and

Cco

ntro

lfor

incl

ude

log

real

GD

Ppe

rca

pita

,lat

itude

,and

wild

cats

,i.e

.,go

vern

men

teffe

ctiv

enes

sis

omitt

edbe

caus

eit

ishi

ghly

corr

elat

edw

ithth

ein

stitu

tiona

lde

pend

entv

aria

bles

ofin

tere

st.A

lles

timat

esus

eth

elim

ited

info

rmat

ion

max

imum

likel

ihoo

d(L

IML)

estim

ator

with

aFu

llerp

aram

eter

ofon

e.A

lles

timat

esin

stru

men

tlog

min

ing

shar

ew

ithpe

rcap

itafo

ssil

fuel

rese

rves

in19

71an

dth

ere

leva

ntcr

itica

lval

uefo

rthe

Kle

iber

gen-

Paap

Wal

krk

Fst

atis

tics

repo

rted

isth

eSt

ock-

Yogo

criti

calv

alue

of16

.38

calc

ulat

edfo

rone

endo

geno

usre

gres

sor,

one

inst

rum

ent,

10%

max

imum

IVre

lativ

ebi

as,a

ndi.i

.der

rors

.

§2.4 Results 47

equality index in Column 10. I find that at a given level of income, doubling

the mining share corresponds to almost 50 percent less women in parliament and

a much lower score on the gender equality index. While more mining-oriented

economies appear to have higher levels of institutionalised gender inequality,

estimating my main results (Table 2.2) separately for males and females reveals

no statistically significant gender differences (see Chapter Appendix). Other

measures of institutional quality yield similar results and the direct institutional

“curse” of natural resources appears directly relevant for economic dependence

on mining. Partialling-out institutional effects in my main results (i.e., holding

government effectiveness constant) implies larger general equilibrium effects.

In Panel C of Table 2.4 I turn to conflict and stability. In Column 11, I

estimate my main equation using the number of years since 1970 (i.e., when the

instrument is measured) that a country has been classified as in a state of civil

war in the Uppsala Conflict Data Program / International Peace Research Institute

(UCDP/PRIO) Armed Conflict Dataset as a dependence variable (Centre for the

Study of Civil War, 2009). I find no evidence of any relationship between mining

dependence and conflict, consistent with related studies by Brunnschweiler and

Bulte (2009), Arezki and Gylfason (2013), and Blattman and Bazzi (2014). In

Column 12 I find a weakly significant relationship between mining dependence

and government stability (Kaufman et al, 2013). Columns 13 and 14 find

mining-dependent countries, after controlling for income level and distance from

the equator, tend to be less democratic and more likely to experience a coup d’etat.

This result is reinforced by the final estimate in Column 15, showing that doubling

the mining share of a country corresponds to around a 50 percent increase in the

military expenditure share of the national budget. My results for conflict, coups,

and military expenditure line up firmly behind the idea that resource rents can

be used to “buy” peace and stability (Arezki and Gylfason, 2013; Cotet and Tsui,

2013).

§2.5 Within-country evidence from Indonesia 48

2.5 Within-country evidence from Indonesia

Looking for similar patterns in different contexts is a useful way to gauge

generality. In this section I present new evidence from a large cross-section of

Indonesian districts in 2009, holding country- and province-specific observable

and unobservable factors constant, and using an estimation strategy analogous

to my international estimates.21 Indonesia’s large mining boom of the 2000s is

coming to an end and its long-term development implications are still not well

understood (Garnaut, 2015; Hill et al., 2008).

Consider the equation:

ln(yd) = αln(Md) + γp + βX ′d + ud (2.2)

where yd is a health or education outcome in district d in 2009. While my

international estimates look at national educational attainment, sub-national

data allow me to also explore participation and quality using enrolment rates

and test scores. I use net enrolment ratios for each district from the high

quality, district representative socioeconomic survey (SUSENAS) carried out

by Indonesia’s central statistics agency, Badan Pusat Statistik (BPS). I also use

average examination test scores (out of 100) for each district from the Ministry of

Education, reported by each school to the district education office and on-reported

up to the ministry. To my knowledge, there are no reliable district-level child

mortality or life expectancy data.22 Instead I look at (a) the percentage of the

births attended by a skilled health worker and (b) average household health and

education expenditures, both derived from SUSENAS.

21All data used in this section are taken from the Indonesia Database for Policy and EconomicResearch (World Bank, 2015), freely available for download and easily replicable.

22Child death is a rare event for the average household. With roughly 40 deaths out of every1000 births, a very large sample is needed to capture sufficient incidence and variation at thedistrict level. Existing household surveys are simply not large enough; census data must be used.


Md is the mining and quarrying share of regional gross domestic product

(RGDP) in district d in 2009 from BPS.23 Consistent with the UN variable used

in my cross-country estimates, mining covers oil, natural gas, coal, and minerals;

quarrying refers to the quarrying of surface rocks, sand, and soil. District-level

mining dependence across Indonesia is presented in Figure 7.

γp is a fixed effect for each of Indonesia’s 33 provinces, capturing

province-specific factors jointly affecting mining dependence and social outcomes

and restricting my comparison to districts within the same province. X ′ includes

log per capita RGDP (combining BPS district RGDP and unpublished population

data) and a categorical variable (i.e., fixed effect) for the assessment given to each

district by the Indonesia Audit Board for their sub-national budgets, to proxy the

quality of local institutions. ud is a heteroskedasticity-consistent error term.

α has a causal interpretation if mining shares are exogenous to the outcomes of

interest conditional on province fixed effects, per capita incomes, and institutions,

i.e., there are no problematic omitted variables correlated with both mining share

and the outcomes of interest within provinces. Such factors cannot be ruled out,

so the following estimates are best interpreted as robust correlations.

Sub-national estimates are presented in Table 2.5. Panel A considers education,

health, and poverty, and Panel B household human capital expenditures. Column

1 of Panel A asks whether, at a given level of income and institutional quality,

children are less likely to participate in school in mining-dependent districts

than in neighbouring districts in the same province. I focus on senior secondary

enrolments because mandatory enrolment policies up to junior secondary remove

variation at the lower levels. Columns 1 and 2 find that a doubling in district

mining share corresponds to a three percent decrease in the net enrolment ratio,

and slightly lower test scores. Both are precisely estimated, statistically significant

23Sub-national accounts are imperfect and likely have some measurement and imputationerror, but Indonesian statistics are comparatively better than many other development countries’national accounts. See McCulloch and Sjahrir (2008) and McCulloch and Malesky (2011) for moredetailed discussions on Indonesian sub-national accounts.


Figu

re2.

7:M

inin

gSh

are

ofD

istri

ctRe

gion

alG

ross

Dom

esti

cPr

oduc

tin

Indo

nesia

,200

9


Tabl

e2.

5:Su

b-na

tion

alEv

iden

cefr

omIn

done

sian

Dist

rict

s

Pane

lA:H

ealth

,edu

catio

n,an

dpo

vert

y

Log

depe

nden

tvar

iabl

eSn

r.En

rolm

ent(

%)

Snr.

test

scor

es(/

100)

Skill

edbi

rth

atte

ndan

t(%

)Po

vert

yra

te(%

)

Col

umn

12

34

Log

min

ing

shar

e-0

.032

***

-0.0

04**

*-0

.024

***

0.05

2***

(0.0

07)

(0.0

01)

(0.0

08)

(0.0

12)

Dis

tric

ts44

745

544

744

5

Pane

lB:H

ouse

hold

(HH

)hum

anca

pita

linv

estm

ent

Log

depe

nden

tvar

iabl

eH

Hed

ucat

ion

expe

nditu

reH

H.h

ealth

expe

nditu

re

Leve

lSh

are

ofto

tal

Leve

lSh

are

ofto

tal

Col

umn

56

78

Log

min

ing

shar

e-0

.070

***

-0.0

37**

*-0

.049

***

-0.0

16*

(0.0

11)

(0.0

08)

(0.0

10)

(0.0

09)

Dis

tric

ts44

743

344

743

3

Star

sde

note

stat

istic

alsi

gnifi

canc

eat

the

10,5

,and

1pe

rcen

tlev

els.

Sam

ple

isth

ela

rges

tpos

sibl

e20

09cr

oss-

sect

ion

ofIn

done

sian

dist

ricts

avai

labl

efo

rthe

seva

riabl

es,u

sing

2009

dist

rictd

efini

tions

and

boun

darie

s.A

lles

timat

esco

ntro

lfor

dist

rictl

ogpe

rcap

itaG

DP,

sub-

natio

nalb

udge

taud

itas

sess

men

tsco

res(

asac

ateg

oric

alva

riabl

e),a

nd33

prov

ince

-spe

cific

fixed

effec

ts.L

east

squa

resi

suse

dth

roug

hout

and

hete

rosk

edas

ticity

-rob

usts

tand

ard

erro

rsar

epr

ovid

edin

pare

nthe

ses.

Coe

ffici

ents

onco

nsta

nts,

cont

rols

,and

fixed

effec

tsar

eno

trep

orte

d.Su

b-na

tiona

ldat

aus

edto

prod

uce

this

tabl

ear

efr

eely

avai

labl

efr

omth

eW

orld

Bank

(201

5)In

done

sia

Dat

abas

efo

rPol

icy

and

Econ

omic

Rese

arch

(DA

POER

).

§2.6 Concluding remarks 52

at the one percent level.24 Column 3 shows that people in more mining-dependent

districts are much less likely to have births attended by skilled health workers.

A district with twice as much of its output from mining, on average, also has

a poverty rate over five percent higher (Column 4), consistent with Bhattacharya

and Resosudarmo’s (2015) finding that non-mining economic growth significantly

reduces poverty in Indonesia but mining growth does not.

Panel B of Table 2.5 extends the national-level mechanism analysis to look

at average household human capital investment in mining-dependent districts.

Columns 5–8 show that households in mining-dependent districts tend to spend

less on education and health-related expenditures both in level terms and

as a share of total household expenditure, although the estimate for health

expenditure share is only statistically significant at the 10 percent level. Doubling

the mining share is associated with seven percent less household spending on

education and five percent less spending on health. The patterns observed across

countries thus appear more widely applicable and a promising area for more

detailed empirical study at the micro level.

2.6 Concluding remarks

This chapter documents how countries with larger mining sectors tend

to diverge below the Preston curve, with lower levels of general health and

educational attainment than expected for their income level. By instrumenting

the relative size of the mining sector with the natural geological variation in

countries’ historical fossil fuel endowments, I provide evidence suggestive of a

causal relationship. Similar patterns between mining, health, and education are

observed across Indonesian districts. My results provide support for a growing

body of evidence linking mining to poorer average living standards, particularly

24A greater mining share corresponds to poorer test scores across all levels of schooling:primary, junior secondary, and senior secondary (estimates available on author request).

§2.6 Concluding remarks 53

vis-a-vis other types of income (Bulte et al., 2005; Daniele, 2011; Gamu et al., 2015).

But a more systematic investigation of causal mechanisms and the conditions

under which mining can be less harmful for health and education is also needed.

Chapter 4 begins down this track, examining how district poverty, average

household expenditure, and different sectors of the local economy respond to coal

mining and natural gas booms in Indonesia. I hope that by highlighting the links

between mining intensity and long-term health and education outcomes at the

international and district levels, I encourage others to focus on these important

aspects of well-being when considering the impacts of mining, particularly in

comparison to other economic development strategies.

§2.7 Chapter 2 Appendix 54

2.7 Chapter 2 Appendix


Figu

re2.

8:G

loba

lIn

fant

Mor

tali

ty,2

005


Figu

re2.

9:G

loba

lYe

arso

fEdu

cati

onal

Att

ainm

ent,

2005


Tabl

e2.

6:M

ain

Resu

ltsI

nstr

umen

ting

wit

hM

iner

al,O

il,G

as,a

ndCo

alRe

serv

es

Log

dep.

varia

ble

Inf.

mor

t.A

vg.y

rsed

uc.,

yout

hPe

rcen

tpop

.w/

noed

uc.,

yout

hLi

feex

p.A

vg.y

rsed

uc.

Perc

entp

op.w

/no

educ

.

Col

umn

12

34

56

Log

min

ing

shar

e0.

17**

*-0

.16*

**0.

87**

*-0

.03*

**-0

.14*

**0.

51**

(0.0

4)(0

.05)

(0.2

7)(0

.01)

(0.0

5)(0

.20)

Excl

uded

-F14

.23

14.3

614

.36

14.1

714

.17

14.1

7

Cou

ntrie

s13

311

211

213

210

710

7

This

tabl

esh

owst

hatt

hem

ain

resu

ltsar

esi

mila

rift

heag

greg

ated

Nor

man

(200

9)pe

rcap

itare

serv

es,i

nclu

ding

min

eral

sast

heIV

.The

smal

lerc

oeffi

cien

tsca

nbe

inte

rpre

ted

asa

com

bina

tion

ofw

eake

ride

ntifi

catio

nst

reng

than

dpo

tent

ially

mor

ehar

mfu

leffe

ctsf

rom

oil,

gas,

and

coal

than

min

eral

s.St

arsd

enot

est

atis

tical

sign

ifica

nce

atth

e10

,5,a

nd1

per

cent

leve

ls.

Sam

ple

isth

ela

rges

tpos

sibl

e20

05cr

oss-

sect

ion

for

thes

eva

riabl

es.

Het

eros

keda

stic

ity-r

obus

tst

anda

rder

rors

inpa

rent

hese

s,an

dco

effici

ents

onco

ntro

ls,c

onst

ants

,and

first

-sta

geco

effici

ents

are

notr

epor

ted.

IVes

timat

esin

stru

men

tlog

min

ing

shar

ew

ithto

talp

erca

pita

oil,

gas,

min

eral

,and

coal

rese

rves

in19

71fr

omN

orm

an(2

009)

.Th

ere

leva

ntcr

itica

lval

uefo

rth

eK

leib

erge

n-Pa

apW

alk

rkF

stat

istic

sre

port

ed(i.

e.ex

clud

ed-F

)is

the

Stoc

k-Yo

gocr

itica

lval

ueof

16.3

8ca

lcul

ated

foro

neen

doge

nous

regr

esso

r,on

ein

stru

men

t,10

%m

axim

umIV

rela

tive

bias

,and

i.i.d

erro

rs.


Tabl

e2.

7:M

ain

Resu

ltsI

nstr

umen

ting

wit

hon

lyO

ilan

dG

asRe

serv

es

Log

dep.

var.

Inf.

mor

t.A

vg.y

rsed

uc.,

yout

hPe

rcen

tpop

.w/

noed

uc.,

yout

hLi

feex

p.A

vg.y

rsed

uc.

Perc

entp

op.w

/no

educ

.

Col

umn

12

34

56

Log

min

ing

shar

e0.

18**

*-0

.22*

**1.

15**

*-0

.04*

**-0

.20*

**0.

69**

*

(0.0

4)(0

.05)

(0.2

5)(0

.01)

(0.0

5)(0

.19)

Excl

uded

-F17

.05

17.2

17.2

17.1

416

.99

16.9

9

Cou

ntrie

s13

311

211

213

210

710

7

This

tabl

esh

ows

the

mai

nre

sults

are

sim

ilar

ifon

lype

rca

pita

oila

ndga

sre

serv

esar

eus

edas

the

IV,i

.e.,

coal

isdr

oppe

d.Th

ela

rger

coeffi

cien

tsca

nbe

inte

rpre

ted

asa

com

bina

tion

ofim

prov

edin

stru

men

tstr

engt

han

dpo

tent

ially

mor

eha

rmfu

leffe

cts

from

oila

ndga

sth

anco

al.S

tars

deno

test

atis

tical

sign

ifica

ncea

tthe

10,5

,and

1pe

rcen

tlev

els.

Sam

plei

sthe

larg

estp

ossi

ble2

005

cros

s-se

ctio

nfo

rthe

seva

riabl

es.H

eter

oske

dast

icity

-rob

usts

tand

ard

erro

rsin

pare

nthe

ses,

and

coeffi

cien

tson

cons

tant

s,co

ntro

ls,a

ndfir

st-s

tage

coeffi

cien

tare

notr

epor

ted.

IVes

timat

esin

stru

men

tlog

min

ing

shar

ew

ithto

talp

erca

pita

oila

ndga

sres

erve

sin

1971

from

Nor

man

(200

9).T

here

leva

ntcr

itica

lval

uefo

rthe

Kle

iber

gen-

Paap

Wal

krk

Fst

atis

ticsr

epor

ted

(i.e.

excl

uded

-F)i

sth

eSt

ock-

Yogo

criti

calv

alue

of16

.38

calc

ulat

edfo

rone

endo

geno

usre

gres

sor,

one

inst

rum

ent,

10%

max

imum

IVre

lativ

ebi

as,a

ndi.i

.der

rors

.


Tabl

e2.

8:M

ain

Resu

ltsu

sing

Alt

erna

tive

Tim

ePe

riod

s

Sam

ple

1995

2000

Betw

een

estim

ator

Ave

rage

min

ing

Estim

ator

OLS

IVO

LSIV

OLS

IVO

LSIV

Col

umn

12

34

56

78

Pane

lA:D

epen

dent

varia

ble,

log

infa

ntm

orta

lity

(dea

thsp

er’0

00)

Log

min

ing

shar

e0.

09**

*0.

22**

*0.

10**

*0.

22**

*0.

11**

*0.

22**

0.12

***

0.19

***

(0.0

2)(0

.05)

(0.0

2)(0

.05)

(0.0

2)(0

.09)

(0.0

2)(0

.04)

Excl

uded

-F17

.73

15.9

416

.65

Cou

ntrie

s14

913

315

113

321

0718

6115

213

4

Pane

lB:D

epen

dent

varia

ble,

log

aver

agey

ears

ofed

ucat

iona

latta

inm

ent

Log

min

ing

shar

e-0

.01

-0.2

4***

-0.0

2-0

.22*

**-0

.02

-0.2

2**

-0.0

2-0

.20*

**

(0.0

2)(0

.06)

(0.0

2)(0

.06)

(0.0

2)(0

.11)

(0.0

1)(0

.05)

Excl

uded

-F16

.53

15.2

716

.79

Cou

ntrie

s12

210

812

310

836

732

312

310

8

This

tabl

esh

ows

that

the

mai

nre

sults

are

unch

ange

dw

hen

diffe

rent

time

perio

dsan

dco

untr

yav

erag

esar

eus

edin

stea

dof

the

2005

cros

s-se

ctio

n.St

arsd

enot

est

atis

tical

sign

ifica

nce

atth

e10

,5,a

nd1

perc

entl

evel

s.Sa

mpl

eis

the

larg

estp

ossi

ble

2005

cros

s-se

ctio

nfo

rthe

seva

riabl

es.H

eter

oske

dast

icity

-rob

usts

tand

ard

erro

rsin

pare

nthe

ses,

and

coeffi

cien

tsco

nsta

nts,

cont

rols

,an

dfir

st-s

tage

regr

essi

onsa

reno

trep

orte

d.A

lles

timat

esin

clud

elo

gre

alG

DP

perc

apita

,abs

olut

edi

stan

cefr

omth

eeq

uato

r,an

inde

xof

gove

rnm

ente

ffect

iven

ess,

and

the

tota

lnum

ber

ofw

ildca

tsdr

illed

inth

e20

thce

ntur

yas

cont

rols

.IV

estim

ates

inst

rum

entl

ogm

inin

gsh

are

with

per

capi

tafo

ssil

fuel

rese

rves

in19

71.

The

rele

vant

criti

calv

alue

for

the

Kle

iber

gen-

Paap

Wal

krk

Fst

atis

ticsr

epor

ted

(i.e.

excl

uded

-F)i

sthe

Stoc

k-Yo

gocr

itica

lval

ueof

16.3

8ca

lcul

ated

foro

neen

doge

nous

regr

esso

r,on

ein

stru

men

t,10

%m

axim

umIV

rela

tive

bias

,and

i.i.d

erro

rs.


Tabl

e2.

9:M

ain

Resu

ltsu

sing

Alt

erna

tive

Mea

sure

sofM

inin

g

Log

depe

nden

tvar

iabl

eIn

f.m

orta

lity

Avg

.yrs

educ

.,yo

uth

Perc

entp

op.w

/no

educ

.,yo

uth

(dea

thsp

er’0

00)

Estim

ator

OLS

IVO

LSIV

OLS

IV

Col

umn

12

34

56

Pane

lA:M

inin

gin

com

eexp

lana

tory

varia

ble(

level)

Log

min

ing

inco

me

perc

apita

0.11

***

0.18

***

-0.0

2*-0

.20*

**0.

27**

*1.

05**

*

(0.0

2)(0

.04)

(0.0

1)(0

.05)

(0.0

7)(0

.24)

Excl

uded

-F15

.69

16.2

416

.24

Cou

ntrie

s15

013

212

611

112

511

1

Pane

lB:M

inin

gex

port

ssha

reex

plan

ator

yva

riabl

e

Log

min

ing

shar

eof

expo

rts

0.04

***

0.23

***

-0.0

2-0

.40*

*0.

28**

*2.

39**

*

(0.0

2)(0

.06)

(0.0

2)(0

.16)

(0.1

0)(0

.84)

Excl

uded

-F13

.09

6.44

6.44

Cou

ntrie

s13

311

511

710

211

610

2

This

tabl

esh

ows

that

usin

gm

inin

gin

com

ean

dm

inin

gex

port

ssh

are

ofto

tale

xpor

tsas

expl

anat

ory

varia

bles

yiel

dses

timat

esof

sim

ilar

sign

san

dco

mpa

rativ

em

agni

tude

s.St

ars

deno

test

atis

tical

sign

ifica

nce

atth

e10

,5,a

nd1

perc

entl

evel

s.Sa

mpl

eis

the

larg

estp

ossi

ble

2005

cros

s-se

ctio

nfo

rthe

seva

riabl

es.H

eter

oske

dast

icity

-rob

usts

tand

ard

erro

rsin

pare

nthe

ses,

and

coeffi

cien

tsco

nsta

nts,

cont

rols

,and

first

-sta

gere

gres

sion

sar

eno

trep

orte

d.A

lles

timat

esin

clud

elo

gre

alG

DP

per

capi

ta,a

bsol

ute

dist

ance

from

the

equa

tor,

anin

dex

ofgo

vern

men

teff

ectiv

enes

s,an

dth

eto

taln

umbe

rof

wild

cats

drill

edin

the

20th

cent

ury

asco

ntro

ls.

IVes

timat

esin

stru

men

tlo

gm

inin

gin

com

epe

rca

pita

(UN

)and

log

min

ing

shar

eof

expo

rts

(Wor

ldBa

nk)w

ithpe

rca

pita

foss

ilfu

elre

serv

esin

1971

.Th

ere

leva

ntcr

itica

lval

uefo

rth

eK

leib

erge

n-Pa

apW

alk

rkF

stat

istic

srep

orte

d(i.

e.ex

clud

ed-F

)ist

heSt

ock-

Yogo

criti

calv

alue

of16

.38

calc

ulat

edfo

rone

endo

geno

usre

gres

sor,

one

inst

rum

ent,

10%

max

imum

IVre

lativ

ebi

as,a

ndi.i

.der

rors

.


Tabl

e2.

10:M

ain

Resu

ltsw

itho

utRe

sour

ce-r

ich

Min

i-sta

tesa

ndO

ther

Out

lier

s

Log

depe

nden

tvar

iabl

eIn

f.m

orta

lity

(dea

ths/

’000

Avg

.yrs

educ

.,yo

uth

Perc

entp

op.w

/no

educ

.,yo

uth

Estim

ator

OLS

IVO

LSIV

OLS

IV

Col

umn

12

34

56

Pane

lA:D

ropp

ing

oil-r

ichm

ini-s

tate

sfro

mth

esam

ple

Log

min

ing

shar

e0.

11**

*0.

13**

*-0

.01

-0.1

5**

0.19

**0.

47**

*

(0.0

2)(0

.04)

(0.0

1)(0

.07)

(0.0

8)(0

.13)

Cou

ntrie

s14

612

812

310

812

210

8

Pane

lB:D

ropp

ing

oil-r

ichm

ini-s

tate

sand

outli

ersf

rom

thes

ampl

e

Log

min

ing

shar

e0.

11**

*0.

18**

*-0

.01

-0.1

8***

0.15

*0.

41**

(0.0

2)(0

.03)

(0.0

1)(0

.06)

(0.0

8)(0

.17)

Cou

ntrie

s14

112

311

910

411

810

4

This

tabl

esh

owst

hatd

ropp

ing

reso

urce

-ric

hm

ini-s

tate

sand

othe

rout

liers

from

the

sam

ple

does

nota

ltert

hem

ain

resu

lts.S

tars

deno

test

atis

tical

sign

ifica

nce

atth

e10

,5,a

nd1

per

cent

leve

ls.

Sam

ple

isth

ela

rges

tpos

sibl

e20

05cr

oss-

sect

ion

for

thes

eva

riabl

es.

Het

eros

keda

stic

ity-r

obus

tst

anda

rder

rors

inpa

rent

hese

s,an

dco

effici

ents

cons

tant

s,co

ntro

ls,a

ndfir

st-s

tage

regr

essi

onsa

reno

trep

orte

d.A

lles

timat

esin

clud

elog

real

GD

Ppe

rcap

ita,a

bsol

ute

dist

ance

from

the

equa

tor,

anin

dex

ofgo

vern

men

teffe

ctiv

enes

s,an

dth

eto

taln

umbe

rofw

ildca

tsdr

illed

inth

e20

thce

ntur

yas

cont

rols

.IV

estim

ates

inst

rum

entl

ogm

inin

gsh

are

with

per

capi

tafo

ssil

fuel

rese

rves

in19

71.

Pane

lAdr

ops

Qat

ar,U

AE,

Brun

ei,K

uwai

t,Ba

hrai

n,an

dO

man

.Pan

elB

drop

sQat

ar,U

AE,

Brun

ei,K

uwai

t,Ba

hrai

n,O

man

Saud

iAra

bia,

Liby

a,Eq

uato

rialG

uine

a,Ir

aq,a

ndVe

nezu

ela.


Tabl

e2.

11:H

ealt

han

dEd

ucat

ion

Elas

tici

ties

ofIn

com

efr

omD

iffer

entS

ecto

rs

Log

depe

nden

tvar

iabl

eIn

fant

mor

talit

y(d

eath

sper

’000

)A

vg.y

ears

ofed

ucat

ion

Perc

ento

fpop

.w/

noed

ucat

ion

Estim

ator

IVIV

OLS

IVIV

OLS

IVIV

OLS

Col

umn

12

34

56

78

9

Log

perc

apita

min

ing

leve

lval

ue-a

dded

0.07

***

-0.1

0***

0.44

***

(0.0

2)(0

.02)

(0.1

0)

Log

perc

apita

agric

ultu

re(A

HFF

)val

ue-a

dded

0.23

*-0

.06

-0.1

7(0

.12)

(0.0

5)(0

.18)

Log

perc

apita

man

ufac

turin

gva

lue-

adde

d-0

.23*

**0.

10**

-0.4

5***

(0.0

6)(0

.05)

(0.1

5)

Log

perc

apita

valu

e-ad

ded

inre

stof

econ

omy

-0.6

0***

-0.4

1***

-0.2

0***

0.39

***

0.03

**-0

.08

-1.0

9***

-0.0

50.

47**

*(0

.06)

(0.0

5)(0

.06)

(0.0

7)(0

.01)

(0.0

5)(0

.23)

(0.0

6)(0

.17)

Excl

uded

-F21

.16*

**38

.61*

**24

.1**

*53

.94*

**24

.1**

*53

.94*

**C

ount

ries

133

161

161

9912

712

799

127

127

This

tabl

epr

esen

tsth

ere

gres

sion

estim

ates

forF

igur

e2.

6in

the

mai

nar

ticle

.Sta

rsde

note

stat

istic

alsi

gnifi

canc

eat

the

10,5

,and

1pe

rcen

tlev

els.

Het

eros

keda

stic

ity-r

obus

tsta

ndar

der

rors

are

inpa

rent

hese

s,co

effici

ents

onco

nsta

nts

and

first

-sta

geco

effici

ents

are

notr

epor

ted,

and

gove

rnm

ente

ffect

iven

ess

and

latit

ude

cont

rols

incl

uded

inal

lest

imat

es.

“Res

tofe

cono

my”

ÂĂ

Âİr

efer

sto

tota

lval

uead

ded

min

usth

ein

com

eal

read

yin

clud

edin

the

estim

ate

(i.e.

,min

ing,

agric

ultu

re,o

rman

ufac

turin

g).M

inin

gis

inst

rum

ente

dw

ithpe

rcap

itafo

ssil

fuel

rese

rves

in19

71an

dag

ricul

ture

with

log

perc

apita

arab

lela

nd.T

here

leva

ntcr

itica

lval

uefo

rthe

Kle

iber

gen-

Paap

Wal

krk

Fst

atis

ticsr

epor

ted

(i.e.

excl

uded

-F)i

sthe

Stoc

k-Yo

gocr

itica

lval

ueof

16.3

8ca

lcul

ated

foro

neen

doge

nous

regr

esso

r,on

ein

stru

men

t,10

%m

axim

umIV

rela

tive

bias

,and

i.i.d

erro

rs.T

heex

clud

edF

stat

istic

isco

mpa

red

toth

eSt

ock-

Yogo

criti

calv

alue

of16

.38

foro

neen

doge

nous

regr

esso

r,on

ein

stru

men

t,an

d10

%m

axim

umIV

rela

tive

bias

.


Tabl

e2.

12:R

egio

nal

Sub-

sam

ples

Cou

ntry

grou

pO

ECD

No

OEC

DM

iddl

eEa

stan

dA

fric

aN

oSu

b-Sa

hara

nA

fric

aN

oM

iddl

eEa

stan

dA

fric

a

Estim

ator

OLS

IVO

LSIV

OLS

IVO

LSIV

OLS

IV

Col

umn

12

34

56

78

910

Pane

lA:D

epen

dent

varia

ble,

Log

infa

ntm

orta

lity

Log

min

ing

shar

e0.

07**

0.08

**0.

11**

*0.

17**

*0.

09**

0.02

0.10

***

0.18

***

0.09

***

0.08

***

(0.0

3)(0

.04)

(0.0

2)(0

.04)

(0.0

4)(0

.09)

(0.0

2)(0

.04)

(0.0

2)(0

.03)

Excl

uded

-F7.

7813

.37

1.4

17.4

712

.78

Cou

ntrie

s31

2612

010

761

5910

891

9074

Pane

lB:D

epen

dent

varia

ble,

Log

aver

agey

ears

ofed

ucat

iona

latta

inm

ent

Log

min

ing

shar

e0.

020.

06**

-0.0

4**

-0.2

1***

-0.0

2-0

.21*

-0.0

2-0

.13*

**0.

02-0

.03

(0.0

1)(0

.03)

(0.0

2)(0

.06)

(0.0

3)(0

.11)

(0.0

1)(0

.04)

(0.0

1)(0

.03)

Excl

uded

-F7.

7811

.17

3.33

13.1

211

.81

Cou

ntrie

s31

2691

8145

4493

7877

63

This

tabl

epr

esen

tsth

em

ain

estim

ates

from

Tabl

e2.

2in

the

artic

lees

timat

edby

regi

onal

sub-

sam

ples

.Th

eco

ntin

ento

fAfr

ica

iscl

early

driv

ing

muc

hof

the

aver

age

effec

tin

Tabl

e2.

Star

sde

note

stat

istic

alsi

gnifi

canc

eat

the

10,5

,and

1pe

rce

ntle

vels

.Sa

mpl

eis

the

larg

estp

ossi

ble

2005

cros

s-se

ctio

nfo

rth

ese

varia

bles

.H

eter

oske

dast

icity

-rob

usts

tand

ard

erro

rsin

pare

nthe

ses,

and

coeffi

cien

tsco

nsta

nts,

cont

rols

,and

first

-sta

gere

gres

sion

sar

eno

trep

orte

d.A

lles

timat

esco

ntro

lfor

incl

ude

log

real

GD

Ppe

rca

pita

,lat

itude

,gov

ernm

ente

ffect

iven

ess,

and

wild

cats

drill

edin

the

20th

cent

ury,

and

use

the

limite

din

form

atio

nm

axim

umlik

elih

ood

(LIM

L)es

timat

orw

itha

Fulle

rpa

ram

eter

ofon

e.A

lles

timat

esin

stru

men

tlog

min

ing

shar

ew

ithpe

rca

pita

foss

ilfu

elre

serv

esin

1971

and

the

rele

vant

criti

calv

alue

fort

heK

leib

erge

n-Pa

apW

alk

rkF

stat

istic

srep

orte

dis

the

Stoc

k-Yo

gocr

itica

lval

ueof

16.3

8ca

lcul

ated

foro

neen

doge

nous

regr

esso

r,on

ein

stru

men

t,10

%m

axim

umIV

rela

tive

bias

,and

i.i.d

erro

rs.


Tabl

e2.

13:S

ub-s

ampl

eEs

tim

atio

nby

Inst

itut

ion

Type

Cat

egor

ical

dum

my

Dem

ocra

cyPr

esid

entia

lsys

tem

Fren

chle

gals

yste

mC

olon

yM

ono-

faith

NY

NY

NY

NY

NY

Col

umn

12

34

56

78

910

Pane

lA:D

epen

dent

varia

ble,

log

infa

ntm

orta

lity

(dea

thsp

er’0

00

Log

min

ing

shar

e0.

19*

0.13

***

0.16

***

0.26

***

0.17

***

0.13

***

0.10

***

0.14

***

0.15

***

0.20

***

(0.1

0)(0

.05)

(0.0

5)(0

.08)

(0.0

5)(0

.04)

(0.0

4)(0

.04)

(0.0

5)(0

.04)

Excl

uded

-F1.

7936

.726

.65.

39.

2111

.78.

227.

5812

.28

14.9

2

Cou

ntrie

s52

7140

8866

6725

8774

59

Pane

lB:D

epen

dent

varia

ble,

log

aver

agey

ears

ofed

ucat

iona

latta

inm

ent

Log

min

ing

shar

e-0

.15

0.03

**0.

04**

*-0

.23*

**-0

.09*

*-0

.19*

**-0

.02

-0.1

8***

-0.0

4-0

.14*

**

(0.1

0)(0

.02)

(0.0

1)(0

.07)

(0.0

4)(0

.05)

(0.0

2)(0

.07)

(0.0

4)(0

.05)

Excl

uded

-F3.

7625

.69

25.3

28.

516.

2115

.21

12.8

26.

2612

.58

13.3

9

Cou

ntrie

s38

6634

7255

5223

7259

48

This

tabl

epr

esen

tsth

em

ain

estim

ates

from

Tabl

e2.

2in

the

artic

lees

timat

edby

sub-

sam

ples

cate

goris

edby

bina

ryin

stitu

tiona

lcha

ract

eris

tics

com

mon

lyus

edto

expl

ain

the

hete

roge

neou

sex

perie

nces

ofre

sour

ceric

hco

untr

ies.

Star

sde

note

stat

istic

alsi

gnifi

canc

eat

the

10,5

,and

1pe

rcen

tlev

els.

Sam

ple

isth

ela

rges

tpos

sibl

e20

05cr

oss-

sect

ion

fort

hese

varia

bles

.Het

eros

keda

stic

ity-r

obus

tsta

ndar

der

rors

inpa

rent

hese

s,an

dco

effici

ents

cons

tant

s,co

ntro

ls,a

ndfir

st-s

tage

regr

essi

ons

are

notr

epor

ted.

All

estim

ates

cont

rolf

orin

clud

elo

gre

alG

DP

perc

apita

,lat

itude

,gov

ernm

ente

ffect

iven

ess,

and

wild

cats

drill

edin

the

20th

cent

ury,

and

use

the

limite

din

form

atio

nm

axim

umlik

elih

ood

(LIM

L)es

timat

orw

itha

Fulle

rpa

ram

eter

ofon

e.A

lles

timat

esin

stru

men

tlog

min

ing

shar

ew

ithpe

rcap

itafo

ssil

fuel

rese

rves

in19

71an

dth

ere

leva

ntcr

itica

lval

uefo

rthe

Kle

iber

gen-

Paap

Wal

krk

Fst

atis

ticsr

epor

ted

isth

eSt

ock-

Yogo

criti

calv

alue

of16

.38

calc

ulat

edfo

rone

endo

geno

usre

gres

sor,

one

inst

rum

ent,

10%

max

imum

IVre

lativ

ebi

as,a

ndi.i

.der

rors

.


Tabl

e2.

14:L

ocal

Aver

age

Part

ial

Effe

cts,

ByCo

mm

odit

y

Inst

rum

ent

Oil

Gas

Coa

lM

iner

als

All

Estim

ator

LIM

LLI

ML

LIM

LLI

ML

CU

E-G

MM

Col

umn

12

34

5Pa

nelA

:Dep

ende

ntva

riabl

e,lo

gin

fant

mor

talit

y(d

eath

sper

’000

)Lo

gm

inin

gsh

are

0.16

***

0.18

***

0.23

***

0.19

*0.

16**

*(0

.04)

(0.0

4)(0

.08)

(0.0

6)(0

.04)

Excl

uded

-F7.

3714

.88

3.36

5.18

10.4

1C

ount

ries

133

133

133

133

133

Pane

lB:D

epen

dent

varia

ble,

log

aver

agey

ears

ofed

ucat

iona

latta

inm

ent

Log

min

ing

shar

e-0

.16*

**-0

.16*

**0.

09*

0.14

**-0

.16*

**(0

.04)

(0.0

6)(0

.05)

(0.0

7)(0

.04)

Excl

uded

-F7.

2714

.38

2.9

3.13

8.48

Cou

ntrie

s10

710

710

710

710

7

This

tabl

epr

esen

tsth

elo

cala

vera

gepa

rtia

leffe

ctof

min

ing

depe

nden

ceco

nditi

onal

onea

chty

peof

natu

ralr

esou

rce

rese

rves

inN

orm

an(2

009)

.N

ote

that

coal

and

min

eral

sre

serv

esar

epr

actic

ally

unid

entifi

edan

dth

eref

ore

unin

terp

reta

ble.

Star

sden

ote

stat

istic

alsi

gnifi

canc

eat

the

10,5

,and

1pe

rcen

tlev

elsa

ndsa

mpl

epe

riod

2005

.Het

eros

keda

stic

ity-r

obus

tsta

ndar

der

rors

inpa

rent

hese

s,an

dco

effici

ents

cons

tant

s,co

ntro

ls,a

ndfir

st-s

tage

regr

essi

onsa

reno

trep

orte

d.A

lles

timat

esin

clud

edas

cont

rols

log

real

GD

Ppe

rcap

ita,l

atitu

de,g

over

nmen

teffe

ctiv

enes

s,an

dw

ildca

tsdr

illed

inth

e20

thce

ntur

y.C

olum

ns1–

4ar

ees

timat

edw

ithth

elim

ited

info

rmat

ion

max

imum

likel

ihoo

d(L

IML)

estim

ator

with

aFu

ller

para

met

erof

one

and

Col

umn

5us

esth

eco

ntin

uous

lyup

date

dge

nera

lised

met

hod

ofm

omen

tses

timat

or(C

UE-

GM

M).

The

inst

rum

ent

inth

eto

pro

wre

fer

toth

epe

rca

pita

rese

rves

ofth

isty

pein

2001

,and

Col

umns

5in

clud

esal

lfo

urty

pesa

ssep

arat

ein

stru

men

ts.T

here

leva

ntcr

itica

lval

uefo

rthe

Kle

iber

gen-

Paap

Wal

krk

Fst

atis

ticsr

epor

ted

aret

heSt

ock-

Yogo

criti

calv

alue

sof1

6.38

forC

olum

ns1–

4an

d5.

44fo

rCol

umn

5,bo

thfo

r10%

max

imum

IVre

lativ

ebi

asan

di.i

.der

rors

.


Tabl

e2.

15:N

atio

nal-

leve

lM

echa

nism

s—O

LSan

dIV

Pane

lA:I

ncom

ein

othe

rsec

tors

and

publ

ichu

man

capi

tali

nves

tmen

t

Log

depe

nden

tvar

iabl

ePe

rcap

itano

n-m

inin

gin

com

ePe

rcap

itahe

alth

expe

nditu

reH

ealth

expe

nditu

resh

are

Perc

apita

educ

atio

nex

pend

iture

Educ

atio

nex

pend

iture

shar

e

Estim

ator

OLS

LIM

LO

LSLI

ML

OLS

LIM

LO

LSLI

ML

OLS

LIM

L

Log

min

ing

shar

e-0

.06*

**-0

.09*

**-0

.04*

*-0

.22*

*-0

.06*

*-0

.16*

-0.0

2-0

.08

0.01

0.03

(0.0

2)(0

.04)

(0.0

2)(0

.09)

(0.0

3)(0

.09)

(0.0

2)(0

.08)

(0.0

2)(0

.11)

Excl

uded

-F16

.37

16.4

316

.43

17.5

18.6

4

Cou

ntrie

s15

113

315

113

315

113

311

910

813

912

2

Pane

lB:Q

ualit

yof

gove

rnm

enta

ndge

nder

equa

lity

Dep

ende

ntva

riabl

eG

over

nmen

teffe

ctiv

enes

sind

exC

ontr

olof

corr

uptio

nin

dex

Cor

rupt

ion

perc

eptio

nsin

dex

Fem

ale

parl.

(p.c

.)C

PIA

gend

erin

dex

Estim

ator

OLS

LIM

LO

LSLI

ML

OLS

LIM

LO

LSLI

ML

OLS

LIM

L

Log

min

ing

shar

e-0

.11*

**-0

.17*

**-0

.15*

**-0

.14*

**-0

.31*

**-0

.31*

**-1

.09*

**-4

.90*

**-0

.10*

**-0

.17*

**

(0.0

3)(0

.04)

(0.0

3)(0

.04)

(0.0

7)(0

.10)

(0.4

0)(1

.1)

(0.0

3)(0

.03)

Excl

uded

-F16

.05

16.7

318

.09

6.06

671

.61

Cou

ntrie

s15

213

415

613

711

710

315

113

263

55

Pane

lC:C

onfli

ctan

dst

abili

ty

Dep

ende

ntva

riabl

eYe

arso

fcon

flict

Stab

ility

inde

xPo

lity

IVsc

ore

Num

bero

fcou

psLo

gm

ilita

ryex

pend

iture

shar

e

OLS

LIM

LO

LSLI

ML

OLS

LIM

LO

LSLI

ML

OLS

LIM

L

Log

min

ing

shar

e0.

190.

29-0

.12*

**-0

.08*

-1.3

1***

-5.5

2***

0.27

*0.

74**

*0.

12**

*0.

44**

*

(0.2

2)(0

.24)

(0.0

3)(0

.04)

(0.3

1)(1

.2)

(0.1

6)(0

.26)

(0.0

4)(0

.1)

Excl

uded

-F16

.73

16.7

313

.23

16.7

35.

4

Cou

ntrie

s15

613

715

613

713

711

915

613

713

511

6

This

tabl

epr

esen

tsth

ena

tiona

l-lev

elm

echa

nism

anal

ysis

from

the

artic

le(T

able

2.4)

expa

nded

toin

clud

eth

ele

asts

quar

eses

timat

ebe

side

IVes

timat

e,fo

rcom

paris

on.S

tars

deno

test

atis

tical

sign

ifica

nce

atth

e10

,5,a

nd1

perc

entl

evel

s.Sa

mpl

eis

the

larg

estp

ossi

ble

2005

cros

s-se

ctio

nfo

rthe

seva

riabl

es.H

eter

oske

dast

icity

-rob

usts

tand

ard

erro

rsin

pare

nthe

ses,

and

coeffi

cien

tsco

nsta

nts,

cont

rols

,and

first

-sta

gere

gres

sion

sar

eno

trep

orte

d.Pa

nelA

cont

rols

incl

ude

log

real

GD

Ppe

rcap

ita,l

atitu

de,g

over

nmen

teffe

ctiv

enes

s,an

dw

ildca

tsdr

illed

inth

e20

thce

ntur

y.Pa

nels

Ban

dC

cont

rolf

orin

clud

elo

gre

alG

DP

per

capi

ta,l

atitu

de,a

ndw

ildca

ts,i

.e.,

gove

rnm

ente

ffect

iven

ess

isom

itted

.A

lles

timat

esus

eth

elim

ited

info

rmat

ion

max

imum

likel

ihoo

d(L

IML)

estim

ator

with

aFu

ller

para

met

erof

one.

All

estim

ates

inst

rum

entl

ogm

inin

gsh

are

with

perc

apita

foss

ilfu

elre

serv

esin

1971

and

the

rele

vant

criti

calv

alue

fort

heK

leib

erge

n-Pa

apW

alk

rkF

stat

istic

srep

orte

dis

the

Stoc

k-Yo

gocr

itica

lval

ueof

16.3

8ca

lcul

ated

for

one

endo

geno

usre

gres

sor,

one

inst

rum

ent,

10%

max

imum

IVre

lativ

ebi

as,a

ndi.i

.der

rors

.


Tabl

e2.

16:M

ain

Resu

lts,

ByG

ende

r

Log

dep.

varia

ble

Inf.

mor

t.(de

aths

/’00

0Li

feex

pect

.(ye

ars)

Avg

.yrs

educ

.Pe

rcen

tpop

.w/

noed

uc.

Estim

ator

OLS

IVO

LSIV

OLS

IVO

LSIV

Pane

lA:F

emal

eC

olum

n1

23

45

67

8Lo

gm

inin

gsh

are

0.09

***

0.22

***

-0.0

1***

-0.0

6***

-0.0

1-0

.19*

**0.

14**

0.61

***

(0.0

2)(0

.04)

(0.0

1)(0

.01)

(0.0

2)(0

.05)

(0.0

6)(0

.18)

Excl

uded

-F15

.94

16.5

216

.72

16.7

2C

ount

ries

151

133

150

132

126

112

126

112

Pane

lB:M

ale

Col

umn

910

1112

1314

1516

Log

min

ing

shar

e0.

09**

*0.

21**

*-0

.01*

*-0

.03*

**-0

.01

-0.1

8***

0.17

***

0.74

***

(0.0

2)(0

.04)

(0.0

1)(0

.01)

(0.0

1)(0

.04)

(0.0

6)(0

.19)

Excl

uded

-F15

.94

16.5

216

.72

16.7

2C

ount

ries

151

133

150

132

126

112

126

112

This

tabl

esh

ows

the

mai

nre

sults

are

virt

ually

unch

ange

dif

disa

ggre

gate

dby

gend

er.

Star

sde

note

stat

istic

alsi

gnifi

canc

eat

the

10,5

,and

1pe

rcen

tlev

els.

Sam

ple

isth

ela

rges

tpos

sibl

e20

05cr

oss-

sect

ion

fort

hese

varia

bles

.H

eter

oske

dast

icity

-rob

ust

stan

dard

erro

rsin

pare

nthe

ses,

and

coeffi

cien

tsco

nsta

nts,

cont

rols

,an

dfir

st-s

tage

regr

essi

ons

are

notr

epor

ted.

All

estim

ates

incl

ude

log

real

GD

Ppe

rca

pita

,abs

olut

edi

stan

cefr

omth

eeq

uato

r,an

inde

xof

gove

rnm

ente

ffect

iven

ess,

and

the

tota

lnum

bero

fwild

cats

drill

edin

the

20th

cent

ury

asco

ntro

ls.I

Ves

timat

esin

stru

men

tlog

min

ing

shar

ew

ithpe

rcap

itafo

ssil

fuel

rese

rves

in19

71.T

here

leva

ntcr

itica

lval

uefo

rthe

Kle

iber

gen-

Paap

Wal

krk

Fst

atis

ticsr

epor

ted

(i.e.

excl

uded

-F)i

sthe

Stoc

k-Yo

gocr

itica

lval

ueof

16.3

8ca

lcul

ated

for

one

endo

geno

usre

gres

sor,

one

inst

rum

ent,

10%

max

imum

IVre

lativ

ebi

as,a

ndi.i

.der

rors

.

Chapter 3

Is plantation agriculture good for the

poor?

Abstract

I examine the poverty impacts of the largest modern plantation sector expansion,

Indonesian oil palm in the 2000s. Combining administrative data on local oil

palm acreage at the district level with survey-based estimates of poverty into

a balanced district panel, I estimate long-differences spanning the expansion.

Causal effects are identified through an instrumental variable strategy exploiting

detailed geo-spatial data on crop-specific agro-climatic suitability. The results

suggest that increasing the oil palm share of land in a district by ten percentage

points leads to around a forty per cent reduction in its poverty rate. Of the

more than 10 million Indonesians lifted from poverty over the 2000s, my most

conservative estimate suggests that at least 1.3 million people have escaped

poverty due to growth in the oil palm sector. Different panel data techniques

are used to assess short-run dynamics. I observe similar effects across different

oil palm producing regions, for industrial and smallholder plantations, and at the

province level. Oil palm expansion tends to be followed by a small but sustained

boost to the value of district-level agricultural output, manufacturing output, and

total output.

68


3.1 Introduction

Palm oil is the world’s most consumed vegetable oil. Crude palm oil is

derived from the reddish pulp of the fruit of the oil palm, a plantation-based,

labour-intensive cash crop originating from Africa (elaeis guineensis) and the

Americas (elaeis oleifera), mostly grown in developing countries today.25 Millions

of people across Asia, South America, and Africa earn income from oil palms,

yielding more oil per hectare than any other crop from relatively little inputs. Oil

palm is one of the most economically attractive uses for land in humid lowland

tropics (Butler et al, 2009) but the palm oil industry is one of the world’s most

socially contested due to deforestation, forest fires, endangered wildlife, displaced

people, and local conflicts.26

In this chapter I ask whether the world’s largest modern plantation-based

agricultural expansion has been pro-poor. I estimate the impacts of the

remarkable expansion in palm oil production in Indonesia on poverty over the

2000s using rich new longitudinal data. Blending administrative information on

local oil palm acreage at the district (kabupaten) level with survey-based estimates

of district poverty, I relate decadal changes in oil palm plantation area to changes

in district poverty over the same period, comparing the poverty elasticity of

oil palm land against alternative uses for land (e.g., rice and forestry). Causal

effects are identified through a novel instrumental variable (IV) strategy exploiting

detailed geo-spatial data on agro-climatic suitability for every field in Indonesia.

By controlling for potential yields of other crops that could share agro-climatic

suitability characteristics with oil palm, I ensure the identifying variation relates

only to oil palm and not other types of agriculture.

25A cash crop is typically an agricultural crop grown to sell rather than consume, usually toglobal export markets (c.f., locally consumed food and subsistence crops).

26See Corley and Tinker (2003) for history and physiology, and Rival and Levang (2014) andSayer et al (2012) for physiology and recent developments in Asia. Dennis et al (2005), Koh andWilcove (2007, 2008), Busch et al (2015), and Miriam et al (2015) discuss environmental impacts,and Barr and Sayer (2012), McCarthy et al (2011), Rist et al (2010), Cramb (2013), Gellert (2015),and Cramb and McCarthy (2016) discuss local social impacts.


The key finding is that districts with larger oil palm expansion have achieved

more poverty reduction than otherwise similar districts without oil palm

expansion. The magnitude of the estimated poverty reduction from increasing

the district share of oil palm land by ten percentage points from my preferred IV

estimator is around 40% of the poverty rate. A simple policy simulation based

on my most conservative estimate suggests that at least 1.3 million out of the

approximately 10 million people lifted from poverty over the 2000s have escaped

poverty due to growth in the oil palm sector. Poverty gaps significantly narrow,

suggesting not only those near the poverty line are being lifted up. I assess

short-term effects and dynamics using standard panel estimators with distributed

lags: dynamic effects reflect the oil palm life cycle. I find no evidence of any major

effect heterogeneity when I disaggregate by large plantations and smallholders,

despite the starkly different characteristics of the two sectors. Similar effects

are also observed across Indonesia’s major palm oil producing regions and at

the province level. I find some evidence of spillovers to other local economic

activities, with oil palm expansion usually followed by a small but sustained boost

to agriculture, manufacturing, and total district output.

The key contribution of this chapter is to provide new causal evidence on

impacts of growth in Indonesia’s palm oil sector on poverty. In providing these

estimates, I shed new light on the poverty elasticity of plantation-based cash

crop production. The role of the agricultural sector in economic development

and poverty reduction has been widely studied (Dercon (2009) and Dercon and

Gollin (2014) provide reviews), but little attention has been paid to plantation

agriculture or cash crops despite their ubiquity in developing countries (Barbier,

1989; Maxwell and Fernando, 1989; Pryor, 1982; Tiffen and Mortimore, 1990).

Agricultural growth tends to be pro-poor (Ravallion and Chen, 2003; Kraay,

2006; Ravallion and Chen, 2007; Christiaensen et al, 2012), but large-scale

agricultural development remains contested and plantation-based cash crops


have starkly different characteristics to other forms of agriculture (Quizon and

Binswanger, 1986; Anriquez and Lopez, 2007; Maertens and Swinnen, 2009;

Hayami, 2010). Unlike food crops and subsistence agriculture, plantation-based

cash crops seldom feed those employed in modern sectors (c.f., Lewis, 1954;

Schultz, 1964) and the potential for agricultural demand-led industrialisation

is ambiguous (c.f., Johnston and Mellor, 1961; Ranis and Fei, 1961; Adelman,

1984). On one hand, consumption linkages may be greater than other agriculture,

due to higher yields and profits. On the other hand, low technology, skill,

and processing requirements for most cash crops imply limited production

linkages. Plantation-based cash crops, however, are unique in this regard. The

plantation system arises due to the need for closer coordination between farm

production and large-scale processing due to the need to process some crops

shortly after harvest. Examples include black tea, sisal, and palm oil, which

must be milled within 24 hours of harvest (c.f., green tea, cocoa, coconuts, and

copra do not require much further processing or marketing, so are suitable

to independent smallholders and family farms). So although several key

mechanisms responsible for past agriculture-led poverty reduction and the “green

revolution”—agricultural technology growth, initial agricultural infrastructure,

and human capital conditions (Ravallion and Datt, 2002; Gollin et al, 2002)—are

generally less applicable for cash crops, this may not be the case for the plantation

system due to its infrastructure requirements (Gollin (2010) provides a useful

review of the theory and evidence linking agricultural production to poverty

reduction). To my knowledge, this is the first nationwide study of the link between

plantation-based cash crops and poverty.

§3.2 Indonesia’s oil palm expansion 72

I provide new evidence against a particularly salient policy debate on palm

oil across the developing world. While the environmental costs of oil palm

have been widely documented, whether Indonesia’s dramatic shift in land use

towards oil palm has brought benefits to the poorest is the subject of much

speculation but yet no systematic quantitative inquiry (McCarthy et al, 2011; Rist

et al, 2010; Cramb and Sujang, 2013; Cahyadi and Waibel, 2013; Budidarsono et

al, 2012). Existing qualitative narratives and geographically-narrow case studies

provide a rich source of descriptive evidence but little basis for causal inference,

as they tend to have weak and narrow internal validity. Although I focus

on Indonesia, findings are informative for other developing countries looking

towards plantation agriculture for poverty reduction. An additional contribution

is my use of a novel IV approach to study causal effects of agricultural sector

growth.

The chapter proceeds as follows. The next section provides a brief background

on Indonesia’s recent oil palm expansion and possible links to poverty. Section

3.3 explains the data and Section 3.4 my empirical approach. Section 3.5 presents

the main results. Section 3.6 estimates short-run effects using annual panel data.

Section 3.7 considers alternative explanations for the main results. Section 3.8

explores effect heterogeneity and spillovers to other sectors. Section 3.9 concludes.

3.2 Indonesia’s oil palm expansion

The world’s largest plantation-based agricultural expansion is taking place

in Indonesia. The third most populous developing country after China and

India, Indonesia supplied more than 40 per cent of the 60.54 million metric

tons of palm oil produced in 2014–15. Global palm oil production has doubled

every decade since the 1960s, surpassing soy bean oil in 2007 to become the

dominant vegetable oil (US Department of Agriculture, 2015). With a comparative

advantage in unskilled labour-intensive goods and proximity to India and China


(the largest purchasers), Indonesia was well-placed to capitalise on the growing

demand. Palm oil has been Indonesia’s largest agricultural export for the last two

decades, with its rapid increase in production coming almost exclusively through

land area expansion (92 per cent) rather than intensification and higher yields

(Gaskell, 2015). From 2001–2009 (my study period) oil palm land increased from

around five million hectares to just under 17 million hectares, or 8.7 per cent of

Indonesia.27 Oil palm expansion has had large opportunity costs, particularly in

terms of the environment (Busch et al, 2015; Carlson et al, 2013; Fargione et al,

2008; Gibbs et al, 2010; Hunt, 2010; Koh et al, 2011; Rival and Levang, 2014; and

Wheeler et al, 2013). Land use is the central policy issue but there is a paucity of

credible evidence on the welfare impacts of changing land use patterns.

Three decades of economic growth and structural change since the 1970s saw

broad-based benefits and poverty reduction across Indonesia (Hill, 1996). Rural

poverty reduction appears to have been mostly driven by agricultural growth,

including through the Green Revolution (Suryahadi et al, 2009; de Silva and

Sumarto, 2014; Rada et al, 2011).28 Since the Asian Financial Crisis and the fall

of Suharto in 1997, economic growth and poverty reduction have both slowed

(Pepinski and Wihardja, 2011). Rapid structural change and a steadily rising

manufacturing share of gross domestic product (GDP) slowed to a halt with

the contemporaneous mining and palm oil booms of the 2000s. The poverty

headcount has continued to fall, but it is unclear how much progress can be

attributed to oil palm. Almost 100 million Indonesians still lived below or near

the poverty line in 2014, of which 28 million, or 11.4% of the population, lived

below the poverty line.29

27Gatto et al (2015) discuss the links between the oil palm boom and land-use dynamics.28See Booth (1988), Fuglie (2010), and Rada et al (2011) for overviews of Indonesian agricultural

development in this period.29Hill, 2014; Manning (2010), Miranti (2010), Suryahadi et al (2003), Manning and Sumarto

(2011), and Wetterberg et al (1999) provide comprehensive accounts of the evolution anddeterminants of poverty in Indonesia.


3.2.1 Linking oil palms to poverty

The poverty elasticity of economic growth in different sectors depends on

sectors’ relative importance to the economy and poor people (Loayza and

Raddatz, 2010). Economic expansion in the oil palm sector is thus likely to be

pro-poor if poor people (a) are employed, (b) have access to land to become

smallholders, or (c) benefit from related economic development.

Labour intensity shapes the poverty elasticity of sectoral growth in most

countries and any poverty benefits from oil palm expansion could be a purely

labour income story for new smallholders or people working on plantations

(Thorbecke and Jung, 1996). Oil palm is a labour-intensive cash crop requiring

little skill or capital to grow and harvest. Farmers and plantation workers typically

earn more than other low skilled workers, with returns to labour estimated to

be 2–7 times the average agricultural wage (Budidarsono et al, 2012). Large

plantations employ roughly two people for every five hectares. In 2010, 1.7

million Indonesians worked on oil palm plantations (Burke and Resosudarmo,

2012). However almost half of Indonesia’s reported oil palm plantation area is

managed by smallholders, usually with 1–2 hectares each, generating significantly

more jobs per hectare. Smallholder plantation area has grown much faster than

company and state-owned plantation area since 2000 (McCarthy et al, 2011; Gatto

et al, 2015), accounting for a larger share of the oil palm-related labour market.

Smallholders tend to report improved yields, profits, nutrition, and incomes after

entering the sector (Budidarsono et al, 2012). While those living below the poverty

line are more likely to be landless and unable to legally become smallholders, they

often work on large industrial plantations. Existing studies typically argue oil

palm expansions bring little benefits to local communities (Obidzinski et al, 2014),

but palm oil is unique compared many other cash crops, combining high returns

to labour with the need for initial infrastructural outlays and processing facilities

(capital needed to actually grow oil palm and sell fresh fruit bunches is minimal

§3.3 Data 75

though). New roads and electrification needed to run palm oil plantations and

mills could alleviate some constraints to rural development. The question of

whether Indonesia’s large oil palm expansion has been good for the poor is

ultimately an empirical one.

3.3 Data

3.3.1 Oil palm

My main explanatory variable is official district oil palm acreage, measured

in hectares and taken from the Tree Crop Statistics of Indonesia for Oil Palm.

Produced by the Department of Agriculture annually since 1996, data cover land

of varying condition (damaged, immature, and mature plantation) and ownership

(private, state, and smallholder).30 While data on official oil palm land are likely

imperfect, focusing on plantation land declared by the Indonesian Government

has greatest tractability.31 I convert oil palm land area to a share of total district

area to focus on changing compositions: comparing oil palm land to other land

uses. As oil palm expansion has been predominantly in rural districts, the

comparison tends to be against other types of agriculture and rural livelihoods

(e.g., rice, rubber, coffee, and forestry). Oil palm land as a share of total district in

2009 is shown in Figure 3.1. Districts across Sumatra and Kalimantan use a greater

share of land for oil palm than those in Sulawesi, Java, and eastern Indonesia.

30Districts with no oil palm land are missing values in the original data, so I recode them aszeros to retain the baseline and control districts. Before recoding as zeros, I cross-checked dataagainst other sources for official plantation figures and gained strong anecdotal evidence frompublic officials that data are more or less nationally exhaustive. There are no large jumps from theimputed zero values. All increase gradually. Similar results are obtained if I drop all districts withno oil palm, focusing only on changes in districts with oil palm land. Unless otherwise stated,subnational data are taken from the World Bank (2015).

31Alternative satellite data are ill-suited for this study, as they cannot distinguish betweenmature oil palm plantations and natural or other forests. For the parts of Indonesia where satellitedata on plantation areas have been field-verified, strong anecdotal evidence from NGOs currentlyassessing these issues suggests small unofficial, informal, and illegal oil palm developments tendto locate alongside and proportional to officially declared plantations, as the same supply chaininfrastructure is needed.

§3.3 Data 76

Figu

re3.

1:O

ilPa

lmLa

ndas

aSh

are

ofD

istri

ctA

rea,

2009

§3.3 Data 77

3.3.2 Poverty

My primary outcome variable is the district poverty rate from 2002 to 2010,

taken from Indonesia’s central statistics agency, Badan Pusat Statistik (BPS). The

poverty rate is a key social policy target for Indonesian governments, defined

as the share of total district population living below an expenditure-based

poverty line that varies by district and period, linked by a universal

consumption requirement (mostly caloric). Poverty figures are derived from

the consumption module of BPS’s district-representative national socio-economic

survey (SUSENAS), implemented at least annually and covering almost two

million people across all provinces in 2010. SUSENAS is agnostic to whether

consumption goods are purchased in formal or informal markets and a consistent

method has been used to calculate poverty rates for the period under study (i.e.,

the method changed in 1998 and 2011). The distribution of household expenditure

can be steep around the poverty line, so I also estimate impacts on the depth

of poverty measured by the poverty gap index: the average gap between the

expenditure of poor people and the poverty line. This allows me to assess whether

only people near the poverty line are affected or those further below, although the

depth of poverty is interesting in its own right. District poverty rates in 2010 are

presented in Figure 3.2. Most of the poor live in Java and poverty rates are highest

in the eastern periphery away from the north-western islands producing most of

Indonesia’s palm oil.

§3.3 Data 78

Figu

re3.

2:D

istri

ctPo

vert

yRa

tes,

2010

§3.3 Data 79

3.3.3 Pemekaran

Indonesian districts (kabupaten) are clearly defined legal and geographical

units with district-level administrations reflecting local economies. A district

panel provides temporal and spatial variation suitable to identifying aggregate

district-level impacts (Sparrow et al, 2015). Indonesia underwent one of the

world’s largest reconfigurations of a modern state with the fall of President

Suharto in 1997, democratising and decentralising power to around 300 district

governments. New political and fiscal powers drove the number of districts

to proliferate from 292 in 1998 to 514 in 2014, a process known as pemekaran.

Pierskalla (forthcoming) and Fitriani, Hofman, and Kaiser (2005) provide detailed

accounts of pemekaran, highlighting how district splits followed sub-district

(kecamatan) boundaries and did not affect neighbouring districts’ borders. I

combine the SUSENAS-derived estimates of district poverty with the official

district oil palm statistics and apply year-2001 district boundaries to obtain a

nationally-exhaustive balanced panel of 341 constant geographic units.32

32In most Indonesian data, districts retain the original names and codes after splitting andreducing in size. Care is needed to avoid applying district fixed effects to such units. Ininternational data, this equates to letting the USSR series continue without its former membersinstead of creating a new series for Russia. Alternative district definitions yield similar results,but constant land area units allow an uninterrupted panel dataset better suited to my researchquestion. Panel summary statistics are provided in the Chapter Appendix.


3.4 Empirical approach

I relate changes in shares of district area used for oil palms to poverty with the

long-difference equation:

ln(yd,2010) − ln(yd,2002) = β(Pd,2009 − Pd,2001) + δi + γXd,2000 + εd (3.1)

where ln(yd,2010) − ln(yd,2002) denotes the change in the log poverty rate from

2002–2010 in district d. Differenced log dependent variables are used to better

compare relative changes in poverty prevalence between districts with high and

low poverty rates.33 Pd,2009 − Pd,2001 is the 2001–2009 change in the share of

district area used for oil palm plantations, lagged by one year because poverty is

measured in the middle of the year and oil palm at the end. Palm oil land shares

are not logged to retain zero values. Differencing removes any time-invariant level

sources of bias jointly affecting land use and poverty (e.g., geography, climate,

history, institutions, and culture). β is the effect of an additional percentage

point of oil palm land as a share of total district land on the district poverty

rate (i.e., a semi-elasticity). δi are island fixed effects, capturing region-specific

factors and allowing different regional trends (e.g., due to different patterns

of economic development or large regional infrastructure investments). Island

groups are defined as Java, Sumatra, Kalimantan, and Sulawesi, with remaining

eastern islands grouped together. γXd,2000 includes initial log poverty and per

capita output, capturing convergence across regions with higher poverty rates and

allowing variable trends by initial conditions.34 Standard errors are adjusted for

33The logged dependent variable ensures districts with relatively low levels of poverty makingsimilar proportional gains to districts with higher relative levels of poverty are accountedfor similarly. Results are similar using linear-linear functional form (see Chapter Appendix),suggesting districts with relatively low levels of poverty are not driving my results.

34The presence of a lagged dependent variable in a short panel could bias coefficients onthe convergence term upwards and oil palm land towards zero (Barro, 2015). Estimating across-section of long differences appears to minimise Hurwicz-Nickell bias, as results are similar(a) without the lagged dependent variable, and (b) with the dependent variable and without islandfixed effects (available on request).


heteroskedasticity.

I opt for long differences because any poverty impacts arising from additional

oil palm land are not likely to be fully realised immediately. Plantation

companies must establish the necessary infrastructure, hire workers, prepare

land, plant oil palms, then harvest the first fruit at least two years later. Likewise

smallholder farmers need time to switch livelihood, prepare land, and plant oil

palms, although smallholders commonly intercrop and adopt mixed livelihood

strategies. It takes five to seven years for oil palm to reach a productive state and

the price paid for a fresh fruit bunch increases with tree maturity.35

A causal interpretation of β̂ relies on an assumption of parallel trends, common

to all difference-in-difference-type approaches. Consistent estimates are obtained

if no time-varying omitted variables systematically shift poverty trends within

island groups after allowing for differential trends by initial income and poverty

levels. While oil palm expansion is governed by administrative processes subject

to a high degree of randomness (discussed further in Section 4.6), unobservable

heterogeneity is impossible to rule out in non-randomised observational studies,

so I turn to IV estimation to identify causal effects.

3.4.1 Instrumental variable strategy

My main source of identifying variation is a rich geo-spatial dataset on

agricultural productivity: the Food and Agriculture Organisation’s (FAO)

Global Agro-Ecological Zones (GAEZ) data. I instrument the change in the

share of district area used for oil palm from 2001–2009 with average district

agro-climatically attainable oil palm yield. Exploiting the variation in oil palm

expansion arising from crop-specific agro-climatic suitability isolates the effect of

developing oil palm on areas where it is makes the most sense to develop it.

35Prices are set weekly and published in local newspapers, such that per hectare income likelyincreases with time. Differencing and allowing differential regional trends is likely sufficient tocapture any systematic differences across local markets.


Figure 3.3: Attainable Palm Oil Yield Across Indonesia


The GAEZ dataset uses state-of-the-art agronomic models and high-resolution

data on geographic characteristics and climatic conditions to predict attainable

yields for 1.7 million grid cells covering the Earth’s surface. Estimates are available

for different crops on every piece of land regardless of whether the land is

cultivated or growing the crop, informing farmers and policy makers on how

productive they would be at crops they are not currently growing. Key inputs

to the model are exogenous variables known for every grid cell: soil types and

conditions, elevation, land gradient, rainfall, temperature, humidity, wind speed,

and sun exposure. Inputs feed into agronomic models predicting how inputs

affect the micro-foundations of each crop’s growth processes, explaining how a

given set of growing conditions map to potential yields at each grid cell.36 GAEZ

provide different sets of productivity predictions for different input scenarios.

I opt for the median options: medium man-made inputs and rain-fed water

supply.37

Pixel-level data for attainable palm oil yield across every field in Indonesia is

presented in Figure 3.3. Each major region has some districts suitable for palm oil

production with only rainfall irrigation and a medium level of inputs, particularly

in the low-lying tropical parts of Sumatra, Kalimantan, and eastern Indonesia best

suited to tropical oil seed crops (i.e., around the equator). I mapped the gridded

data on attainable yield of each of Indonesia’s main agricultural commodities

to Indonesia’s district boundaries using geographical information systems (GIS)

then calculated district’s mean. The granularity of the data and the continuous

nature of this variable gives a rich source of variation: a different value for every

potential palm oil producing district.

36Time-varying variables (i.e., humidity, temperature, rainfall, windspeed) are measured at ahigh frequency and their levels and variation over time are used in the models. Predictions foryields at the end of the 20th century and beyond are based on a large number of past realisationsof these variables over the 20th century.

37Rain-fed irrigation minimises measurement error from historical changes in irrigationintensity and technologies (Nunn and Qian (2011). Alternative input assumptions give a similarspatial distribution, so do not affect my results. See Fischer et al (2002), Nunn and Qian (2011),and Costinot et al (2016) for further details on FAO GAEZ data.


Instrument relevance and strength

Oil palms only grow under certain agro-climatic conditions—humid low-land

tropics—and potential yields and profits in each district affect the likelihood

that district will have oil palms planted. The IV is thus theoretically relevant.

First-stage coefficients on potential palm oil yield are positive and statistically

significant at the 0.1 per cent level (presented with main results). As a weak

IV problem can be present even with highly significant first-stage coefficients

(Bound et al, 1995), I report the Kleibergen and Paap (2006) rk Wald F statistic

(the heteroskedasticity-robust analogue of the Cragg and Donald (1993) test

statistic) against the relevant Stock and Yogo (2005) critical values and use the

Fuller (1977) median-unbiased limited information maximum likelihood (LIML)

estimator for all IV estimates.38 Additional confidence intervals and hypothesis

tests are provided using Moreira’s (2003) conditional likelihood ratio (CLR)

procedures, which significantly outperform the traditional Anderson and Rubin

(1949) weak-instrument-robust-inference tests (Andrews, Moreira, and Stock,

2004, 2006).

Exogeneity and exclusion

A causal interpretation is only obtained if average attainable district palm

oil yields do not affect changes in poverty through any channel other than oil

palm expansion. GAEZ potential yield predictions do not involve estimating any

sort of statistical relationship between observed inputs, outputs and agro-climatic

conditions, so are exogenously determined with respect to district economic

and poverty conditions. The two theoretically endogenous factors shaping

GAEZ data—irrigation and man-made inputs—are set equal for all districts, so

uncorrelated with poverty trends across districts.

38I prefer the Fuller estimator over the standard two-stage least squares (2SLS) IV estimatorbecause (a) a few IV estimates have scope for a weak IV problem and LIML point estimates aremore reliable for inference under weak IV (Murray, 2006), and (b) I prefer to use the same estimatorthroughout. 2SLS gives similar results (available on request).

§3.5 Main results 85

The main concern is that a key input to the oil palm GAEZ productivity model

could affect productivity of similar tropical crops and therefore welfare through

agricultural productivity in other sectors. This is a common challenge with

external instruments, particularly those relating to weather and climate (Bazzi

and Clemens, 2013; Sarsons, 2015). Using a crop-specific instrument reduces this

threat, but I can go a step further. GAEZ attainable yield data are available for

most of Indonesia’s major agricultural crops. By controlling for other key crops’

potential yields, I further restrict the identifying variation to that relating only to

oil palms and not shared suitability characteristics with other crops (i.e., tropical,

humid, non-mountainous, lowlands with sufficient rainfall that are less suitable

for other tropical cash crops that could likely be grown in similar areas to oil palm).

Second-stage coefficient stability to the inclusion of attainable yields for all rice

types, tea, coffee, cocoa, and cassava, suggest the exclusion restriction is likely

satisfied.

3.5 Main results

My main result is presented in Figure 3.4 and Table 3.1. Districts that converted

more of their land to oil palm plantations in the 2000s have achieved more rapid

poverty reduction than districts of similar initial poverty levels and per capita

incomes in the same region. Oil palm is on average a better rural land use than

alternatives for poverty alleviation.39

39More precisely, my specification using the oil palm share of district area compares the effectof using additional oil palm land relative to the average of all other possible uses for land.


Tabl

e3.

1:Po

vert

yIm

pact

soft

he20

01–2

010

Oil

Palm

Expa

nsio

n

Dep

ende

ntva

riabl

e∆

log

dist

rictp

over

tyra

teEs

timat

orO

LSIV

IVIV

IVIV

Col

umn

12

34

56

∆oi

lpal

mla

nd/

dist

ricta

rea

-0.0

12**

*-0

.034

***

-0.0

40**

*-0

.046

***

-0.0

78**

*-0

.033

***

(0.0

03)

(0.0

13)

(0.0

12)

(0.0

12)

(0.0

23)

(0.0

13)

CLR

95%

confi

denc

ein

terv

alN

/A[-0

.19,

-0.0

03]

[-0.1

33,-

0.00

8][-0

.12,

-0.0

2][-0

.20,

0.00

9][-0

.17,

0.00

6]C

LRp-

valu

e(H

0:B

=0)

N/A

0.03

0.01

00.

080.

08Is

land

fixed

effec

tY

NN

NN

NPo

tent

ialr

ice

and

casa

vayi

elds

NN

YN

YN

Pote

ntia

lcoff

ee,c

ocoa

,tea

yiel

dsN

NN

YY

NFi

rst-s

tage

coeffi

cient

sand

diag

nost

icsLo

gpo

tent

ialp

alm

oily

ield

N/A

0.69

0***

0.69

0***

0.89

3***

0.58

4***

Log

palm

oils

uita

bilit

yin

dex

N/A

0.50

6***

Kle

iber

gen-

Paap

Wal

drk

Fst

atN

/A13

.41

13.1

220

.25

9.08

10.8

330

perc

entm

axFu

llerb

iasc

ritic

alva

lue

N/A

12.7

112

.71

12.7

112

.71

12.7

110

perc

entm

axFu

llerb

iasc

ritic

alva

lue

N/A

19.3

619

.36

19.3

619

.36

19.3

6O

bser

vatio

ns33

530

830

330

330

330

9

Star

sde

note

stat

istic

alsi

gnifi

canc

eat

the

10,5

,and

1pe

rce

ntle

vels

.A

llIV

estim

ates

use

Fulle

r’slim

ited

info

rmat

ion

max

imum

likel

ihoo

des

timat

orw

ithaF

ulle

rpar

amet

erof

one.

Sam

plei

sthe

long

-diff

eren

cecr

oss-

sect

ion

ofal

lava

ilabl

edis

tric

tsfr

om20

02–2

010.

2001

dist

rictb

ound

arie

sar

eus

ed,w

ithne

wdi

stric

tsco

llaps

edin

toye

ar-2

001

pare

ntdi

stric

ts.

Cha

nges

insa

mpl

essi

zear

edu

eto

data

avai

labi

lity.

Oil

palm

land

isla

gged

one

perio

d(2

001–

2009

).H

eter

oske

dast

icity

-rob

usts

tand

ard

erro

rsar

ein

pare

nthe

ses

and

cond

ition

allik

elih

ood

ratio

-bas

edco

nfide

ncei

nter

vals

insq

uare

brac

kets

.Isl

and

grou

ping

sare

defin

edas

Java

,Sum

atra

,Kal

iman

tan,

and

Sula

wes

i,w

ithre

mai

ning

dist

ricts

grou

ped

toge

ther

.Eac

hre

gres

sion

incl

udes

log

pove

rty

and

log

perc

apita

outp

utin

the

initi

alpe

riod

asco

ntro

lvar

iabl

es.


Figure 3.4: Poverty Impacts of the 2001–2010 Oil Palm Expansion

-.15

-.1-.0

50

Least squares Oil palm yield IVOil palm yield IV + rice Oil palm yield IV + cash cropsOil palm yield IV + all crops Oil palm suitability index IV

Column 1 of Table 3.1 presents Equation 1 estimated with least squares. A

district that experienced a ten percentage point increase in the share of land

used for oil palm over the 2000s, at the mean (e.g., from 10 to 20 per cent of

district area), has a poverty rate 12 per cent lower than otherwise similar districts

in 2010. Columns 2–5 of Table 3.1 present the IV estimates, dropping island

dummies for stronger identification. Positive first-stage coefficients confirm oil

palm expansion has been most pronounced where most productive. Column 2

shows that a ten percentage point increase in district oil palm land share over the

2000s corresponds to over a thirty per cent greater reduction in the poverty rate.

That the estimate in Column 2 is almost three times the magnitude of least squares

is not surprising, as oil palm is likely to have more pronounced effects where more

productive.40 The CLR confidence interval reported under the main coefficient

40While oil palm in less suitable areas—notably in a few poorer mountainous areas—suchgrowers (i.e., “non-compliers”) account for a minuscule component of total oil palm area andproduction.


does not overlap with zero, rejecting the null hypothesis that the coefficient equals

zero with 97% confidence.

In Column 3 of Table 3.1 I control for the attainable yield of two of Indonesia’s

most important non-cash crop agricultural commodities: rice (wet- and dry-land)

and cassava. The first-stage coefficient in Column 3 is virtually the same and

the second-stage coefficient slightly larger; this is expected, as rice is typically

produced in slightly different regions with different agro-climatic conditions

and so should not share much of the identifying variation with oil palms (e.g.,

compare Java to Kalimantan, or the rice-growing deltas of south-east Asia to

neighbouring tropical islands growing cash crops. See Hayami (2010) for further

discussion). In Column 4 I separately include average district-specific attainable

yields for three of Indonesia’s key tropical cash crops: cocoa, coffee, and tea.41

The first-stage coefficient increases to 0.89 and the excluded-F statistic to 20.25,

exceeding the Stock-Yogo critical value of 19.36 for ten per cent maximum Fuller

bias. The second-stage coefficient restricting the identifying variation to suitability

to oil palms but none of Indonesia’s other major cash crops (i.e., controlling

for agro-climatic suitability for cocoa, coffee, and tea) is 0.046, suggesting an

additional ten percentage point increase in oil palm land share where it is most

suitable can almost halve the poverty rate. The CLR test rejects the null that

the coefficient equals zero with almost 100% confidence. In Column 5, I include

potential yield for all six additional crops as controls. Identification is significantly

weaker, with an excluded-F statistic of 9.08, and the estimated coefficient on oil

palm expansion much larger at 0.078. The CLR test however still rejects the

null hypothesis that the coefficient is equal to zero at the ten per cent level.

Instrumenting oil palm expansion with GAEZ’s oil palm suitability index instead

of potential yield gives a similar result in Column 6. Figure 3.4 illustrates how

similar confidence intervals (at the 95% level) emerge from these alternative IV

specifications, most overlapping those from least squares.41To my knowledge, agro-climatic suitability data for rubber is unavailable.


Tabl

e3.

2:Im

pact

soft

he20

01–2

010

Oil

Palm

Expa

nsio

non

the

Pove

rty

Gap

Dep

ende

ntva

riabl

e∆

log

dist

rictp

over

tyga

pin

dex

Estim

ator

OLS

IVIV

IVIV

IVC

olum

n1

23

45

6

∆oi

lpal

mla

nd/

dist

ricta

rea

-0.0

15**

*-0

.038

**-0

.038

**-0

.054

***

-0.1

32**

*-0

.046

**(0

.005

)(0

.019

)(0

.018

)(0

.018

)(0

.049

)(0

.020

)C

LR95

%co

nfide

nce

inte

rval

N/A

[-0.1

30,0

.013

][-0

.110

,0.0

06]

[-0.1

46,-

0.01

1][-0

.282

,0.0

17]

[-0.2

44,0

.017

]C

LRp-

valu

e(H

0:B=

0)N

/A0.

122

0.08

50.

016

0.08

40.

12Is

land

fixed

effec

tY

NN

NN

NPo

tent

ialr

ice

and

casa

vayi

elds

NN

YN

YN

Pote

ntia

lcoff

ee,c

ocoa

,tea

yiel

dsN

NN

YY

NFi

rst-s

tage

coeffi

cient

sand

diag

nost

icsLo

gpo

tent

ialp

alm

oily

ield

N/A

0.69

5***

0.82

***

0.90

2***

0.58

4***

Log

palm

oils

uita

bilit

yin

dex

N/A

0.50

8***

Kle

iber

gen-

Paap

Wal

drk

Fst

atN

/A13

.54

13.7

720

.53

9.08

10.8

330

perc

entm

axFu

llerb

iasc

ritic

alva

lue

N/A

12.7

112

.71

12.7

112

.71

12.7

110

perc

entm

axFu

llerb

iasc

ritic

alva

lue

N/A

19.3

619

.36

19.3

619

.36

19.3

6O

bser

vatio

ns33

530

830

330

330

330

9

Star

sden

otes

tatis

tical

sign

ifica

ncea

tthe

10,5

,and

1pe

rcen

tlev

els.

All

IVes

timat

esus

eFul

ler’s

limite

din

form

atio

nm

axim

umlik

elih

ood

estim

ator

with

aFu

llerp

aram

eter

ofon

e.Sa

mpl

eis

the

long

-diff

eren

cecr

oss-

sect

ion

ofal

lava

ilabl

edi

stric

tsfr

om20

02–2

010.

2001

dist

rict

boun

darie

sare

used

,with

new

dist

ricts

colla

psed

into

year

-200

1pa

rent

dist

ricts

.Cha

nges

insa

mpl

essi

zear

edu

eto

data

avai

labi

lity.

Oil

palm

land

isla

gged

one

perio

d(2

001–

2009

).H

eter

oske

dast

icity

-rob

usts

tand

ard

erro

rsar

ein

pare

nthe

ses

and

cond

ition

allik

elih

ood

ratio

-bas

edco

nfide

nce

inte

rval

sin

squa

rebr

acke

ts.

Isla

ndgr

oupi

ngs

are

defin

edas

Java

,Sum

atra

,Kal

iman

tan,

and

Sula

wes

i,w

ithre

mai

ning

dist

ricts

grou

ped

toge

ther

.Ea

chre

gres

sion

incl

udes

log

pove

rty

and

log

per

capi

taou

tput

inth

ein

itial

perio

das

cont

rol

varia

bles

.


The reduction in the poverty rate observed in Table 3.1 could be due to

people near the poverty line being lifted just above, with little effect on those

further down the income distribution. In this case, the poverty rate would fall

but the gap between the average poor person and the poverty line (i.e., poverty

depth) increase. I estimate impacts on poverty depth in Table 3.2. Results are

similar to Table 3.1. Oil palm expansion corresponds to reductions in poverty

depth. Decreases in the rate and depth of poverty confirm benefits from oil palm

expansion tend to reach the average person living below the poverty line.

Table 3.3 explores the robustness of the main long-difference estimate to

controlling for economic growth and changes in the natural environment,

partialling out effects through these two channels. In Column 1 of Table 3.3 I

control for the decadal change in log per capita output. The coefficient on RGDP

growth is statistically insignificant, and the coefficient on oil palm land share is

unchanged, implying palm oil production has been a particularly pro-poor (i.e.,

redistributive) economic activity.

Like many equatorial developing countries, Indonesia was mostly tropical

forest half a century ago and oil palms are planted on areas once primary

forest. The forestry landscape can often change alongside oil palm expansion.

Such changes could bias estimates through potential omitted variables.42 In

Column 2 I control for the initial level and 2000–2010 change in tree cover using

pixel-level Moderate Resolution Imaging Spectroradiometer (MODIS) satellite

imagery data.43 Results are similar to Column 1 of Table 3.1 and the coefficients

on tree cover variables statistically insignificant, suggesting conversion of primary

forest into more “economically productive” use is not driving my result (i.e.,

42For example, income from forestry and logging taking place in the same districts as oilpalm expansion could bias my estimates downwards, and social harms like conflict and malariaassociated with deforestation could bias estimates upwards. Note that if such factors arise due tooil palm expansion, this is included in the net effect in my main estimates.

43Data are taken from Wheeler et al (2013). While MODIS data cannot disentangle primaryforest from plantations (i.e., it is distinctly not a measure of deforestation in the Indonesiancontext), it is still a useful proxy for observed changes in forest and the natural environment.Burgess et al (2012) and Wheeler et al (2013) discuss the MODIS data in detail.


Tabl

e3.

3:A

ddit

iona

lCo

vari

ates

and

Robu

stne

ss

Dep

ende

ntva

riabl

e:∆

log

dist

rictp

over

tyra

te

Col

umn

12

34

∆oi

lpal

mla

nd/

dist

ricta

rea

-0.0

12**

*-0

.011

***

-0.0

13**

*-0

.014

***

(0.0

03)

(0.0

04)

(0.0

04)

(0.0

04)

∆lo

gdi

stric

tGD

Ppe

rcap

ita(ID

R)-0

.048

(0.0

74)

∆tr

eeco

ver(

pixe

ls)

-0.0

001

(0.0

001)

Late

first

year

NN

YN

Early

final

year

NN

NY

Obs

erva

tions

335

335

336

334

Star

sden

ote

stat

istic

alsi

gnifi

canc

eat

the

10,5

,and

1pe

rcen

tlev

els.

Sam

ple

isth

elo

ng-d

iffer

ence

cros

s-se

ctio

nof

all

avai

labl

edi

stric

tsfr

om20

02–2

010,

with

oil

palm

land

lagg

edon

epe

riod.

2001

dist

rict

boun

darie

sar

eus

ed,w

ithne

wdi

stric

tsco

llaps

edin

toye

ar-2

001

pare

ntdi

stric

ts.

Het

eros

keda

stic

ity-r

obus

tst

anda

rder

rors

are

inpa

rent

hese

s.A

lles

timat

esus

eor

dina

ryle

ast

squa

res

and

cont

rolf

oris

land

fixed

effec

ts,i

nitia

lpov

erty

,and

initi

alpe

rca

pita

inco

me.

Isla

ndgr

oupi

ngsa

rede

fined

asdi

stric

tsfr

omJa

va,S

umat

ra,K

alim

anta

n,an

dSu

law

esi,

with

rem

aini

ngdi

stric

tsgr

oupe

dto

geth

er.C

hang

ein

tree

cove

rref

erst

oth

ech

ange

pixe

lsof

tree

cove

rmea

sure

dby

MO

DIS

sate

llite

data

;ini

tialt

ree

cove

risa

lso

incl

uded

asa

cont

rolv

aria

ble

inC

olum

n2.

§3.6 Short-run and dynamic impacts 92

similar poverty impacts from oil palm land expansion are observed holding tree

cover constant). Finally, as long-differences can sometimes be sensitive to start

and finish year, I use alternative start and finish years in Columns 3 and 4 of Table

3.3. Results are similar.44

3.6 Short-run and dynamic impacts

I have focused on the total changes in oil palm plantation land and poverty

over the 2000s. But the relationship between growth in the palm oil sector and

poverty could vary over the crop’s life cycle. In this section I use alternative panel

estimators to examine short-run impacts.

My preferred panel estimator takes the form:

ln(yd,t) = βPd,t−1 + δd + τi,t + γXd,t−1 + εd,t (3.2)

yd,t denotes poverty in district d at time t. Pd,t−1 is the oil palm land percentage

of total district area, with additional lags in some estimates.45 β is the effect of

an additional percentage point of oil palm land on poverty. δd are district fixed

effects, removing time-invariant district-specific sources of confoundedness. τi,t

are island–year fixed effects capturing time-varying shocks common to each island

group (e.g., economic growth and business cycles, international commodity

prices for an island’s commodities, political shocks, regional infrastructure

investments, and other major policy changes).46 Island-year fixed effects focus my44Following the common heuristic that coefficient stability to additional controls can be

informative about omitted variable bias (Oster, 2015; Bellows and Miguel, 2009; Altonji et al, 2005),the estimated parameter of interest is also similar if I include electricity-related variables proxyingeconomic capacity (Sparrow et al, 2015), fiscal and political variables, and the battery of othercorrelates of poverty available in the Indonesia Database for Policy and Economic Research (WorldBank, 2015). Such additional estimates are available on author request. Further robustness checksare provided in the Chapter Appendix, including using district per capita palm oil production intons instead of the district share of oil palm land, splitting the sample period, and omitting theisland of Java.

45Estimates including lead values of palm oil land, as in-time placebo tests, are included in theChapter Appendix.

46Social policy in Indonesia is strongly targeted towards the poor, but its spatial is relatively


comparison to districts within the same island group, relaxing the parallel trends

assumption to more flexible regional trends. γXd,t−1 is a vector of potential time-

and district-varying controls. Standard errors are adjusted for heteroskedasticity

and clustered by district to allow arbitrary correlation within districts over time.47

β̂ in Equation Two has a causal interpretation if there are no time and district

varying omitted variables correlated with yd,t and Pd,t−1 influential enough to

systematically shift poverty trends within island groups. Assuming changes

in official oil palm land are exogenous to changes in district poverty outcomes

conditional on district and island-by-year fixed effects is reasonable for two

reasons. Firstly, Equation Two focuses on the poverty response from the timing

of district oil palm expansions, so the critical issue is what determines the timing.

Oil palm land declared by the Department of Agriculture reflects plantation sector

land use decisions made through the large, decentralised bureaucracy: each step

in this process is influenced by idiosyncratic factors resulting highly unpredictable

delays.48 The second reason is that island-year fixed effects appear to eliminate

selection bias into oil palm in the short run: a wide range of time-varying

correlates of poverty do not explain changes in oil palm land when included in

the same panel regression model as island-year fixed effects, consistent with the

short-term changes in oil palm land—the timing—being subject to some degree

of randomness.49

unchanged from 2001–2010 and mostly captured by district fixed effects. New social programswere mostly implemented nationally (e.g., the Raskin rice subsidy, PNPM, unconditional cashtransfers, and scholarships) so captured by island–year fixed effects, or piloted in a few villagesbefore national roll-out.

47Bertrand et al (2004) discuss problems arising in panel estimates when serial correlationis unaddressed. I consider larger cluster robust errors a more conservative basis for inferenceand hypothesis testing, with weaker assumptions and better finite sample properties than moreefficient counterparts.

48Indonesian land use regulations are complicated. The Regional Autonomy Laws 1999 sawdistrict forest departments become answerable to bupatis (district heads) instead of the centralgovernment. Bupatis apply to the central government for approval to convert land into oil palmplantations, a process involving identifying areas for plantations, attracting investors, gainingdistrict parliament approval, making a formal request to the central government, central agenciesworking through the request, the district receiving approval, and land being converted. Burgesset al (2012) similarly highlight how administrative lags from central to district governments renderdistrict splits exogenous to province and district outcomes.

49Estimates provided in the Chapter Appendix show how many poverty correlates are


Tabl

e3.

4:Sh

ort-

run

and

Dyn

amic

Pove

rty

Impa

cts

Dep

ende

ntva

riabl

e:lo

gdi

stric

tpov

erty

rate

Estim

ator

Firs

t-diff

eren

ceW

ithin

FEFE

IV

Col

umns

12

34

56

78

Oil

palm

land

/di

stric

tare

a(%

)-0

.003

**-0

.004

***

-0.0

04**

-0.0

07**

*-0

.005

**-0

.007

***

-0.0

06**

*-0

.037

***

(0.0

01)

(0.0

02)

(0.0

02)

(0.0

02)

(0.0

02)

(0.0

02)

(0.0

02)

(0.0

14)

Firs

tlag

-0.0

03-0

.002

-0.0

03

(0.0

02)

(0.0

03)

(0.0

03)

Seco

ndla

g-0

.009

***

-0.0

08**

-0.0

07**

(0.0

03)

(0.0

04)

(0.0

03)

Third

lag

-0.0

06**

*-0

.005

*-0

.005

(0.0

02)

(0.0

03)

(0.0

03)

Σco

effici

ents

-0.0

22**

*-0

.019

***

-0.0

20**

*

Firs

tsta

geco

effici

ent

0.20

7***

Kle

iber

gen-

Paap

Wal

drk

Fst

at8.

79

Dis

tric

tfixe

deff

ects

NN

YY

YY

YY

Isla

nd-s

peci

fictim

etr

ends

NN

NN

NY

NN

Prov

ince

-spe

cific

time

tren

dsN

NN

NN

NY

N

Obs

erva

tions

3040

2371

2371

3386

2717

3386

3386

3386

Star

sde

note

stat

istic

alsi

gnifi

canc

eat

the

10,

5,an

d1

per

cent

leve

ls.

Het

eros

keda

stic

ity-r

obus

tst

anda

rder

rors

are

inpa

rent

hese

s,cl

uste

red

atth

edi

stric

tlev

el.

Sam

ple

isan

annu

al34

1-di

stric

tpan

elfr

om20

02–2

010.

Oil

palm

land

isla

gged

one

perio

d.20

01di

stric

tbo

unda

ries

are

used

,with

new

dist

ricts

colla

psed

into

year

-200

1pa

rent

dist

ricts

.C

hang

esin

sam

ple

size

are

due

toda

taav

aila

bilit

y.Is

land

-yea

rfix

edeff

ects

are

incl

uded

thro

ugho

ut,w

ithis

land

grou

ping

sde

fined

Java

,Sum

atra

,Kal

iman

tan,

Sula

wes

i,an

dth

ere

st.T

hew

ithin

estim

ator

refe

rsto

the

mea

n-di

ffere

nced

(with

in-d

istr

ict)

fixed

effec

tses

timat

or.

The

FEIV

estim

ator

isa

mea

n-di

ffere

nced

Fulle

rlim

ited

info

rmat

ion

max

imum

likel

ihoo

dIV

estim

ator

,w

here

the

inst

rum

enti

sdi

stric

t-spe

cific

linea

rtre

nds

base

don

the

leve

lofi

nitia

loil

palm

land

inea

chdi

stric

t.Si

gnifi

canc

ere

port

edfo

rsum

ofth

eco

effici

ents

rela

test

oth

ete

stth

atth

esu

mof

the

coeffi

cien

tson

oilp

alm

iseq

ualt

oze

ro.


I estimate Equation Two with first-differences, mean-deviations (i.e., within

estimation), distributed lags, and IV. Column 1 of Table 3.4 presents the annual

first-difference. Assuming an effect within the same year, a ten percentage point

increase in the district share of land used for oil palm in one year corresponds to

a three per cent reduction in the poverty rate the next year, statistically significant

at the five per cent level. Assuming the land data is accurate and timely in its

reporting, such immediate effects must come through channels other than the

production and sale of the crop (e.g., payments to communities or waged labour

to establish plantations). In Column 2 I include the first three lags of the annual

first-difference. The second and third lags have much larger coefficients, reflecting

the oil palm life cycle. The sum of the coefficients on oil palm land is 0.022,

between long-difference estimates obtained from least squares and IV. As the

evolution of oil palm land has been gradual, in Column 3 I take first differences

and include district-fixed effects to extract the “shock” component of the changes

in oil palm land (Ciccone, 2011). Coefficients are similar.

In Columns 4–7 of Table 3.4 I adopt the mean-differenced “within” estimator.

Unlike the first-differences in Columns 1–3, coefficients reflect the effect of

variation over time within each district (c.f., at a particular point in time across

districts). As within estimation also picks up level effects, this is a more

appropriate flexible estimator than first-differences (i.e., due to the lags in the

palm oil production process). Column 4 of Table 3.4 presents my preferred within

panel estimate. A ten percentage point increase in the share of land used for oil

palm at the mean corresponds to a seven per cent reduction in the poverty rate

in the short-run. Column 5 includes lags, summing exactly to the least squares

long-difference in Column 1 of Table 3.1. Columns 6 and 7 include island- and

province-specific time trends. Results are almost identical with these rich control

vectors.

statistically significant determinants of changes oil palm land in pooled least squares and withinestimators with district and year fixed effects (Columns 1 and 2), but including island-year fixedeffects in Columns 3 renders them all statistically insignificant.

§3.7 A migration story? 96

The final column of Table 3.4 presents a panel fixed effects IV estimate,

exploiting demonstration effects arising from early adoption. I instrument the

level shares of oil palm land with a district-specific linear trend increasing in

districts’ initial oil palm land, as initial conditions matter for future growth

trajectories. Oil palms historically expanded more in areas where plantations

were already established due to better access to pre-existing knowledge networks,

materials, processing facilities, and other necessary infrastructure. Column 8 of

Table 3.4 shows that a ten percentage point increase in the share of district land

used for oil palm at the mean—exclusively due to that district’s steeper oil palm

land expansion trajectory because of existing oil palm activity—corresponds to

an almost 40 per cent reduction in the poverty rate. The larger IV estimates

suggest early adopters on average achieved more rapid poverty reduction, with

a magnitude similar to my long-difference IV estimates.50

Table 3.5 presents the same set of estimates for poverty depth. The main

coefficient is negative, of similar magnitude, and statistically significant across

all estimates. Panel estimates in Section 4.6 withstand in-time placebo tests and

the inclusion of a wide range of time-varying covariates (see Chapter Appendix).

3.7 A migration story?

District poverty rates can fall either due to real consumption growth for

the poor, or through changes in population. Population changes that would

contaminate my interpretation include inward migration of non-poor people

and outward migration of poor people. Both would also alter poverty rates

in my comparison pool if migration is to and from districts without oil palm

expansion. Critics of the palm oil sector highlight a story of displacement, where

“land-grabbing” drives forest-dwellers, indigenous people, and poor farmers off

50This local average partial effect is quite widely applicable, with 71 of the 341 districts acrossfour of the five island groups using some share of their land for oil palm plantations in 2000.


Tabl

e3.

5:Sh

ort-

run

and

Dyn

amic

Impa

ctso

nPo

vert

yD

epth

Dep

ende

ntva

riabl

e:lo

gpo

vert

yga

pin

dex

Estim

ator

Firs

t-diff

eren

ceW

ithin

FEFE

IV

Col

umns

12

34

56

78

Oil

palm

land

/di

stric

tare

a(%

)-0

.010

***

-0.0

09**

*-0

.012

**-0

.014

***

-0.0

09**

-0.0

14**

*-0

.006

***

-0.0

21**

*

(0.0

03)

(0.0

03)

(0.0

05)

(0.0

04)

(0.0

04)

(0.0

04)

(0.0

02)

(0.0

07)

Firs

tlag

0.00

0-0

.005

0.00

1

(0.0

04)

(0.0

07)

(0.0

04)

Seco

ndla

g-0

.006

-0.0

10-0

.003

(0.0

04)

(0.0

07)

(0.0

05)

Third

lag

-0.0

11**

-0.0

14**

-0.0

10**

(0.0

04)

(0.0

06)

(0.0

05)

Σco

effici

ents

-0.0

26**

*-0

.041

***

-0.0

23**

*

Firs

tsta

geco

effici

ent

0.47

8***

Kle

iber

gen-

Paap

Wal

drk

Fst

at12

.17

Dis

tric

tfixe

deff

ects

NN

YY

YY

YY

Isla

nd-s

peci

fictim

etr

ends

NN

NN

NY

NN

Prov

ince

-spe

cific

time

tren

dsN

NN

NN

NY

N

Obs

erva

tions

2705

2036

2036

3051

2382

3051

3386

3051

Star

sde

note

stat

istic

alsi

gnifi

canc

eat

the

10,

5,an

d1

per

cent

leve

ls.

Het

eros

keda

stic

ity-r

obus

tst

anda

rder

rors

are

inpa

rent

hese

s,cl

uste

red

atth

edi

stric

tlev

el.

Sam

ple

isan

annu

al34

1-di

stric

tpan

elfr

om20

02–2

010.

Oil

palm

land

isla

gged

one

perio

d.20

01di

stric

tbou

ndar

iesa

reus

ed,w

ithne

wdi

stric

tsco

llaps

edin

toye

ar-2

001

pare

ntdi

stric

ts.C

hang

esin

sam

ple

size

ared

ueto

data

avai

labi

lity.

Isla

nd-y

earfi

xed

effec

tsar

einc

lude

dth

roug

hout

,with

isla

ndgr

oupi

ngsd

efine

dJa

va,S

umat

ra,

Kal

iman

tan,

Sula

wes

i,an

dth

eres

t.Th

ewith

ines

timat

orre

fers

toth

emea

n-di

ffere

nced

(with

in-d

istr

ict)

fixed

effec

tses

timat

or.

The

FEIV

estim

ator

isa

mea

n-di

ffere

nced

Fulle

rlim

ited

info

rmat

ion

max

imum

likel

ihoo

dIV

estim

ator

,whe

reth

ein

stru

men

tis

dist

rict-s

peci

ficlin

eart

rend

sbas

edon

the

leve

lofi

nitia

loil

palm

land

inea

chdi

stric

t.Si

gnifi

canc

ere

port

edfo

rsum

ofth

eco

effici

ents

rela

test

oth

ete

stth

atth

esu

mof

the

coeffi

cien

tson

oilp

alm

iseq

ualt

oze

ro.


their previously occupied land (Gellert, 2015; Cramb and McCarthy, 2016). I do

not dispute the existence of such cases and have heard them first hand. But could

population movements–by choice or by force–explain the reductions in district

poverty rates documented in this Chapter?

Understanding the scale and scope of local migration in Indonesia is difficult,

as reliable internal migration data are collected only in the decadal population

census with questions only on inter-province movements and no information

on income level.51 I investigate the plausibility of a migration-based alternative

explanation in three main steps. I first identify the quantum of migration needed

to explain my main estimates in the context of official internal migration statistics

and relevant contextual information gathered from two field visits. I then pursue

province-level estimates, less exposed to the issue of migration given the larger

size of provinces vis-a-vis districts. I finally estimate district-level impacts of oil

palm expansion on population change and on the number of poor people in a

district (c.f., the poverty rate).

The 2010 Indonesian Population Census reported 2.5 per cent of the population

living in a different province to where they lived at the time of the previous census.

In resource-rich provinces, the rate can be higher (around 6 per cent in Riau and

East Kalimantan) or closer to the national average, even below (4 and 1.8 per

cent in Jambi and South Sumatra). The highest rate is in West Papua, the least

densely populated province, but still under 8 per cent. It is important to note

that magnitude of migration flows in the average oil palm district would have

to be around four times the national average rate of recent migrants to explain

my long-difference estimate (i.e., 10 per cent), a further four times that to explain

my preferred IV estimate, and predominately involve poor people leaving or

non-poor people coming. Contrast this to the tendencies of lower-income people

to move to booming regions seeking economic opportunities, and of wealthy51SUSENAS and the labour market surveys SAKERNAS do not include information on

migration. The Indonesian Family Life Surveys are not district-representative. Meng et al (2010)study recent internal migration patterns in Indonesia.


beneficiaries of natural resource sectors to be based in capital cities.

Two other contextual issues bear a mention. First, the popular displacement

narrative relates to agro-industrial frontier expansion, but smallholders manage

around half of Indonesia’s planted oil palm area. The increase in planted oil palm

area over the period of this study was mostly from smallholders, thus accounting

for most of the identifying variation. Independent smallholders tend to be local

people without much government or company support, to be less affluent, but

also to be far more hesitant to move. Plasma scheme smallholders mostly moved

in during the transmigration program, which ceased in 2000.52

Second, a district is a large geographic unit, on average comprising over 200

villages. When villages are forcefully moved or formal relocation agreements

reached, communities tend to be relocated nearby or incorporated into plantation

activities within the same sub-district (kecamaten) or the existing village area if

large–often on unfavourable terms. Relocation to other districts is rare, and a

displaced poor individual is unlikely to move farther than the district or provincial

capital, in no small part due to financial constraints.

Estimating analogous models at a greater level of spatial aggregation

is a useful way to remove the influence of any within-province migration.

Province-level estimates are presented in Table 3.6. Columns 1 and 2 present

short-run effects, focusing on changes within each province over time. Column

1 includes island-specific poverty trends and Column 2 island-year fixed effects.

The magnitude of the estimate in Column 2 is similar to that from the analogous

district-level within estimator (Column 4 of Table 3.4). A long-difference estimate

with island fixed effects is presented in Column 3. Provinces with a ten percentage

point increase in their share of oil palm land have experienced, on average, a 13

per cent greater reduction in the poverty rate from 2002–2010. Province-level

estimates are similar to district-level estimates, suggesting that intra-province

migration is not substantially affecting my findings.52Bazzi et al (2016) detail Indonesia’s transmigration program.


Table 3.6: Province level Results

Estimator FE FE LD

Column 1 2 3

Oil palm land / district area (%)-0.014* -0.007** -0.013**

(0.006) (0.003) (0.004)

Linear island trends Y N N

Island–year fixed effects N Y N

Island fixed effects N Y Y

Observations 319 319 30

Stars denote statistical significance at the 10, 5, and 1 per centlevels. Sample is an annual balanced panel of Indonesianprovinces from 2002–2010, with oil palm land lagged one period.Estimates are the within estimator with province fixed effects (FE)and the long difference estimator (LD). Heteroskedasticitiy-robuststandard errors are in parentheses, clustered at the province levelfor FE estimates and the island level for LD. Data taken from theWorld Bank (2015).

In Table 3.7 I present results from least squares fixed effects and long-difference

estimators (Equations 1 and 2) using logged population (Columns 1 and 2) and

logged number of poor people (Columns 3 and 4) as dependent variables. Column

1 provides no evidence of any short-term change in population size arising from

oil palm land expansion. Column 2 shows that over the nine years, districts with

greater oil palm expansion tend to now have slightly larger populations, although

this effect is statistically significant only at the ten per cent level. Columns 3 and

4 show more oil palm land corresponds to a large reduction in the total number

of poor people in each district.53 I cannot rule out poor people systematically

leaving oil palm districts and being replaced by non-poor inward migrants, but

the evidence presented above suggests that this is highly unlikely to fully explain

the falling poverty rates identified in this Chapter.

53Note that estimates in Columns 1–4 are simple decompositions of estimates in Tables 3.1 and3.4 using the log poverty rate as the dependent variable.


Tabl

e3.

7:Po

pula

tion

,Poo

rPe

ople

,and

Prod

ucti

on

Dep

ende

ntva

riabl

eLo

gdi

stric

tpop

ulat

ion

Log

num

bero

fpoo

rLo

gdi

stric

tpov

erty

rate

Estim

ator

FELD

FELD

FELD

Col

umn

12

34

56

Oil

palm

land

/di

stric

tare

a(%

)-0

.000

70.

003*

-0.0

11**

*-0

.009

***

-0.0

04*

-0.0

07

(0.0

007)

(0.0

02)

(0.0

03)

(0.0

03)

(0.0

02)

(0.0

04)

Perc

apita

palm

oilp

rodu

ctio

n(to

ns)

-0.1

90**

*-0

.125

**

(0.0

54)

(0.0

54)

Dis

tric

tfixe

deff

ects

YN

YN

YN

Isla

nd-y

earfi

xed

effec

tsY

NY

NY

N

Initi

alco

nditi

onco

ntro

lsN

YN

YN

Y

Dis

tric

ts34

133

534

133

534

133

5

Obs

erva

tions

3689

335

3045

335

3386

335

Star

sde

note

stat

istic

alsi

gnifi

canc

eat

the

10,5

,and

1pe

rce

ntle

vels

.Sa

mpl

eis

anan

nual

341

dist

rictp

anel

from

2002

–201

0.O

ilpa

lmva

riabl

esar

ela

gged

one

perio

d.20

01di

stric

tbou

ndar

iesa

reus

ed,w

ithne

wdi

stric

tsco

llaps

edin

toye

ar-2

001p

aren

tdis

tric

ts.V

aria

tions

inth

esam

ples

izea

redu

eto

data

avai

labi

lity.

Estim

ator

sare

with

infix

edeff

ects

(FE)

and

long

diffe

renc

es(L

D).

Het

eros

keda

stic

ity-r

obus

tsta

ndar

der

rors

arei

npa

rent

hese

s,cl

uste

red

atth

edis

tric

tlev

elfo

rpan

eles

timat

ors.

Initi

alco

nditi

onco

ntro

lsre

fert

olo

gdi

stric

tper

capi

tain

com

e,lo

gpo

vert

yra

te,a

ndan

isla

ndgr

oup

fixed

effec

t.

§3.8 Heterogeneity and wider impacts 102

3.8 Heterogeneity and wider impacts

3.8.1 Heterogeneity by region

Existing qualitative studies examining the poverty implications of oil palm

expansion in Indonesia emphasise context-specific heterogeneity (McCarthy,

2010). I now explore potential heterogeneity by region and sector. Figure 3.5

presents the long-difference point estimates for each of Indonesia’s five main

regions. Tabulated results are presented in Table 3.8, for the full-sample within

and long-difference estimates, interacting the island dummies with my main oil

palm land share variable to provide marginal effects by region.54

Figure 3.5: Regional Heterogeneity

-.1-.0

50

.05

.1.1

5Se

mi-e

last

icity

of d

istri

ct p

over

ty ra

te to

oil

palm

land

sha

re

Java Sumatra Kalimantan Sulawesi Other

54I drop the main (not interacted) effects to allow a more straightforward interpretation ofthe estimated coefficients. Main results in Tables 3.1 and 3.2 are weighted averages of these.Sub-sample estimates are provided in the Chapter Appendix.


Table 3.8: Regional Heterogeneity

Dependent variable Log poverty rate Log poverty gap

Estimator FE LD FE LD

Column 1 2 3 4

Java*oil palm land share0.013 -0.040 -0.063 -0.122*

(0.015) (0.041) (0.040) (0.065)

Sumatra*oil palm land share-0.006*** -0.010*** -0.012*** -0.011**

(0.002) (0.003) (0.003) (0.004)

Kalimantan*oil palm land share-0.014*** -0.029*** -0.027*** -0.039***

(0.004) (0.008) (0.007) (0.012)

Sulawesi*oil palm land share-0.045*** -0.052*** -0.044*** -0.031***

(0.007) (0.017) (0.008) (0.008)

Other*oil palm land share-0.036 0.060 -0.162 0.098

(0.128) (0.062) (0.174) (0.118)

District and year fixed effects Y N Y N

Initial conditions controls N Y N Y

Observations 3386 335 3051 335

Stars denote statistical significance at the 10, 5, and 1 per cent levels. Samplesample is an annual district panel from 2002–2010. Oil palm land is laggedone period. 2001 district boundaries are used, with new districts collapsed intoyear-2001 parent districts. Island groupings are defined as districts from Java,Sumatra, Kalimantan, Sulawesi, and with remaining islands grouped together.Estimators are within fixed effects estimator (FE) with district and year fixedeffects, and the long difference estimator (LD) with initial log poverty andlog per capita income controls. Heteroskedasticity-robust standard errors arein parentheses, clustered at the district level for FE estimates. Island*palminteraction terms interact the island dummy for each island with the main oilpalm land share variable. Main effects (not interacted) are dropped for a morestraightforward interpretation.


Column 1 of Table 3.8 presents the within estimate including district and year

fixed effects. The coefficients on the interaction terms for regions with little oil

palm (i.e., Java and eastern Indonesia) are statistically insignificant. Across the

main oil palm growing regions of Sumatra, Kalimantan, and Sulawesi, districts

have experienced short-run poverty reductions as a result of oil palm expansion.

Districts in Sulawesi experienced the largest reductions in district poverty rates,

highlighting how it is not just regions with relatively low poverty driving

my results, but some districts with high poverty rates making commensurate

proportional poverty reductions. The long-difference estimate presented in

Column 2 paints a similar picture to the panel estimates. The only difference is

magnitudes for Kalimantan and Sulawesi are similar in the short and long runs,

but for Sumatra the long-run effects are twice the magnitude of the short-run

effects. A plausible explanation for this difference is that in Kalimantan and

Sulawesi most recent oil palm expansion has come through large industrial-scale

plantations, whereas Sumatran oil palm expansion has been predominantly

smallholders (discussed further below). Columns 3 and 4 of Table 3.8 present

impacts on the poverty gap by region. Results are similar in the short and longer

term across all palm oil producing regions.

3.8.2 Heterogeneity by sector

Indonesian smallholders are reported to have per hectare yields up to 40

per cent lower than industrial estates, struggle to exploit economies of scale,

and use inefficient practices restricting yields and incomes (Hasnah et al, 2004;

Burke and Resosudarmo, 2012; Lee et al, 2013; Alwarritzi et al, 2015). Industrial

plantations are usually between 5,000–20,000 hectares and intensively managed

to maximise efficiency (Corley and Tinker, 2003). Naturally the sectors could have

heterogeneous effects.


Table 3.9: Heterogeneity by Plantation Type

Sector State Private Smallholder

Panel width Annual 4-year Annual 4-year Annual 4-year

Column 1 2 3 4 5 6

Panel A: log poverty rate

Oil palm land/district area (%)-0.011** -0.011** -0.012*** -0.011*** -0.004 -0.011**

(0.004) (0.004) (0.003) (0.004) (0.002) (0.004)

Panel B: log poverty gap index

Oil palm land/district area (%)-0.012** -0.015** -0.014*** -0.011** -0.004 -0.014**

(0.006) (0.006) (0.004) (0.005) (0.003) (0.006)

Observations 3009 1004 3009 1004 3009 1004

Stars denote statistical significance at the 10, 5, and 1 per cent levels. Sample is an annual 341 districtpanel (2002–2010) Oil palm land is lagged one period. 2001 district boundaries are used, with newdistricts collapsed into year-2001 parent districts. Heteroskedasticity-robust standard errors are inparentheses, clustered at the district level. A within estimator with district and island–year fixedeffects is used throughout.

Table 3.9 compares the poverty impacts of additional state, industry, and

smallholder-managed oil palm land. Data are taken from the Tree Crop Statistics

for Indonesia from the Department of Agriculture. Sub-sectoral oil palm land

data are strongly unbalanced, so I use the within estimator (i.e., Equation

Two) and shift between an annual and four-yearly panel to assess dynamics.

Columns 1–4 of Table 3.9 show similar coefficient magnitudes and dynamics

in large state-owned and private plantations, consistent in magnitude with the

long-difference presented in the Column 1 of Table 3.1. Effects for smallholders

are more variable. In Column 5, there is no detectable short-run relationship

between more smallholder land and district poverty rates. Large state and

company plantations, on the other hand, immediately hire labour to establish

and work on the plantations, often building local infrastructure and community

facilities for their workers. Independent smallholders bear these costs and usually

see little profit for two years. Column 6 extends the time-to-effect to four years:

the coefficient is similar to other sectors.


Across all four-yearly panel estimates I find no evidence of any differential

impacts on the rate or depth of poverty, despite the different nature of the sectors

and their varying direct engagement with the poor. This is shown clearly for

the poverty rate in Figure 3.6, with similar point estimates and overlapping

confidence intervals for the three palm oil sub-sectors.55

Figure 3.6: Sector Heterogeneity

-.05

-.04

-.03

-.02

-.01

0Se

mi-e

last

icity

of d

istri

ct p

over

ty ra

te to

oil

palm

land

sha

re

State-owned Private Industry Smallholder

55Estimates exploring further heterogeneity by sector and by land quality and provided in theChapter Appendix.


3.8.3 Wider impacts

In this section I use similar long-difference and fixed effects estimators to

assess whether oil palm expansion has generated district-level aggregate demand

spillovers. Table 3.10 estimates the local economic impacts of oil palm expansion

on the value of district agricultural, manufacturing, and aggregate output (all

official BPS data). First, note that while the estimated coefficients are mostly

positive and statistically significant, magnitudes are not large. Oil palm expansion

does not systematically correspond to local economic booms. For the two

sectors most directly involved—agriculture and manufacturing, which accounts

for milling—Columns 1–4 show small, persistent, and statistically significant

increases in the value of output. A ten percentage point increase in the share

of land used for oil palm plantations corresponds to a seven per cent increase in

the value of agricultural output and a four per cent increase for manufacturing.

Considering aggregate output in Column 5, an annual panel fixed effects estimate

finds no statistically significant immediate effect, implying short-run crowding

out and reallocation of factors of production. The long-difference estimate in

Column 6 shows increasing the share of district land used for oil palm by 10

percentage points corresponds to an average increase in non-oil and gas real

output of 2.4 per cent relative to districts without oil palm expansion. Any

crowding-out of other local economic activity appear at least fully offset in the

medium term, with net economic effects positive but small.


Tabl

e3.

10:E

ffec

tsof

Oil

Palm

Expa

nsio

non

Sect

oral

and

Tota

lD

istri

ctG

DP

RGD

Pde

pend

entv

aria

ble(

IDR)

Log

agric

ultu

ralo

utpu

tLo

gm

anuf

actu

ring

outp

utLo

gG

DP

(exc

l.oi

l,ga

s)

Estim

ator

FELD

FELD

FELD

Oil

palm

land

/di

stric

tare

a(%

)0.

004*

*0.

007*

**0.

005*

**0.

004*

0.00

10.

002*

*

(0.0

02)

(0.0

02)

(0.0

02)

(0.0

02)

(0.0

01)

(0.0

01)

Dis

tric

tfixe

deff

ects

YN

YN

YN

Isla

nd–y

earfi

xed

effec

tsY

NY

NY

N

Initi

alco

nditi

onsc

ontr

ols

NY

NY

NY

Obs

erva

tions

3410

342

3410

342

3410

342

Star

sden

ote

stat

istic

alsi

gnifi

canc

eat

the

10,5

,and

1pe

rcen

tlev

els.

Sam

ple

isan

annu

al34

1di

stric

tpan

elfr

om20

02–2

010.

Oil

palm

land

isla

gged

one

perio

d.20

01di

stric

tbou

ndar

ies

are

used

,with

new

dist

ricts

colla

psed

into

year

-200

1pa

rent

dist

ricts

.Es

timat

ors

are

with

infix

edeff

ects

(FE)

and

long

diffe

renc

e(L

D)

estim

ator

s.H

eter

oske

dast

icity

-rob

usts

tand

ard

erro

rsar

ein

pare

nthe

ses,

clus

tere

dat

the

dist

rictl

evel

forp

anel

estim

ator

s.In

itial

cond

ition

sco

ntro

lsre

fer

tolo

gdi

stric

tper

capi

tain

com

ean

dlo

gpo

vert

yra

tein

2000

,and

isla

ndfix

edeff

ects

.

§3.9 Conclusion 109

3.9 Conclusion

This chapter’s objective was to quantify the contribution of oil palm

expansion to local poverty reduction in Indonesia. While there have

been clear environmental consequences associated with Indonesia’s rapid

increase in palm oil production, Indonesian districts using more land for oil

palm tend to experience more rapid poverty reduction. Indonesia’s recent

smallholder-led oil palm expansion provides an important case study of how

geographically-dispersed pro-poor growth can reach remote rural regions. But

how significant is this contribution for national poverty reduction?

Table 3.11: Estimated Contribution to Poverty Reduction

District ∆ oil palm/areaPoverty rate, 2010

∆ poorActual Counterfactual

Column 1 2 3 4

Rokan Hulu 36 13 20 -28,526

Asahan 34 12 13 -47,012

Labuhan Batu 34 13 12 -43,457

Tanah Laut 24 5 7 -4,509

Deli Serdang 21 7 9 -50,281

Simalungun 20 11 15 -26,463

Kampar 19 10 13 -16,303

Kuantan Singingi 19 13 19 -11,003

Pasaman 17 10 12 -13,077

Langkat 17 11 16 -29,751

Σ estimated poverty reduction for all districts (no. poor people) -1,319,369

Districts are ten largest oil palm expansions, as measured by the 2001–2009change in district area allocated to oil palm and defined by 2001 districtboundaries. Counterfactual poverty rates are estimated by predicting eachdistrict poverty rate with oil palm expansion set to zero using the mostconservative least squares estimator (Column 1, Table 3.1). The estimatedpoverty reduction is calculated from the difference between the estimatedpoverty rate and its counterfactual. The sum in the final row is for all districtsfor which data are available.


Table 3.11 presents the ten districts with the largest proportional oil palm

land expansions. Columns 3 and 4 compare the actual poverty rate to a

simulated counterfactual poverty rate without oil palm expansion based on my

first least squares estimate (i.e., setting oil palm expansion to zero and using a

semi-elasticity of 0.012). All but one of these districts reduced poverty below

its estimated counterfactual poverty rate in the absence of oil palm expansion.

Of the more than 10 million Indonesians lifted from poverty over the 2000s, my

most conservative estimate suggests that at least 1.3 million people have escaped

poverty exclusively due to growth in the oil palm sector.

In this chapter I focused on the macro-level, reduced-form impacts of oil palm

expansion on local poverty. My focus on effects within the same district tends to

miss spillovers across regions or nationally, positive and negative. My findings do

not imply that an oil palm boom is the best way to reduce national poverty. Detailed

mechanism analysis using individual-level data is now needed to understand

whether the observed poverty reduction is purely a labour income story for

those employed in the agricultural sector, or whether there are wider economic

spillovers. Recent work on local multipliers (e.g., Moretti, 2010; Hornbeck

and Keskin, 2015) provides a useful framework for such analysis and, to my

knowledge, has not yet been extended to a developing country context (i.e.

with imperfect substitutability between imports and local consumables, immobile

factors of production, and abundant unskilled labour). Moreover, Indonesia

has continued to rapidly urbanise since its 1998 decentralisation without much

further industrialisation—a phenomenon common to many resource-dependent

countries (Vollrath, Gollin, and Jedwab, 2015). Most palm oil companies are

based in capital cities and general equilibrium effects are not well understood,

particularly consumption linkages to cities’ non-tradable sectors where profits are

mostly spent. The longer-term economic and social consequences of pro-poor

primary sector growth also warrant further study.


My main finding that oil palm expansion has tended to reduce local poverty

in Indonesia should be considered with four more widely-known facts. First,

given the immense environmental costs associated with converting tropical forests

to oil palm plantations, emissions and biodiversity loss in particular, a strong

environmental case can be made for any future expansion to focus on existing

agricultural or degraded land already identified as suitable for oil palm (Gingold

et al, 2012; Austin et al, 2015). Second, oil palms are one of the most productive

uses for land in humid low-lying tropics. There are large gains to be made

from farmers continuing to switch to more productive crops. Market failures

inhibiting crop switching, for example incomplete credit markets, insufficient

public infrastructure, or restrictive land use practices (e.g., relating to food

self-sufficiency policies) could be promising areas for further research or policy

development. Third, large differences in productivity remain (between Indonesia

and Malaysia, and within Indonesia) and improving smallholder productivity

is often as simple as adopting improved agricultural practices.56 Thus there

is scope for further oil palm-related poverty reduction through these three

avenues (extensification, intensification, and crop switching) without the large

environmental costs that have characterised the sector to date. Finally, Indonesia’s

uniquely large share of smallholders engaged in plantation-based agriculture are

central to this story. Generalising my findings to other countries with different

levels of smallholder engagement would be injudicious.

56Knowledge of good agricultural practices for growing oil palms are not widespread forsmallholders and the transfer of knowledge between nucleus and plasma schemes has beenproblematic.



Table 3.12: Panel Summary Statistics

Variable 2002 2010 All years Mean difference

Palm oil land / district area (%)

Mean 0.58 2.65 1.3 2.07

SD 1.61 6.35 4.01 4.74

N 341 341 3386

District poverty rate (%)Mean 19.94 13.82 16.74 -6.12

SD 11.57 7.3 9.5 -4.27

N 335 341 3386

Summary statistics are for the balanced panel of constant geographic units, where districtboundaries are reset to those at the start of the panel period for consistency. Palm oil landas a share of district area is lagged by one year, as it is in my estimates. Data are officialIndonesian Government data, obtained through the World Bank’s Indonesian Databasefor Policy and Economic Research online public portal.


Tabl

e3.

13:M

ain

Resu

lts—

Line

ar-l

inea

rFu

ncti

onal

Form

Dep

ende

ntva

riabl

ePo

vert

yra

te

Estim

ator

OLS

IVIV

IVIV

Col

umn

12

34

5

Palm

oill

and

/di

stric

tare

a(%

)-0

.108

***

-0.6

43**

-0.5

51**

*-0

.535

***

-0.6

08**

(0.0

37)

(0.2

61)

(0.2

12)

(0.1

74)

(0.2

58)

Isla

ndfix

edeff

ect

YN

NN

N

Initi

alco

nditi

onsc

ontr

ols

YY

Yy

Y

Potn

tialr

ice

and

casa

vayi

elds

NN

YN

N

Pote

ntia

lcoff

ee,c

ocoa

,tea

yiel

dsN

NN

YN

Suita

bilit

yin

dex

IVN

NN

NY

Kle

iber

gen-

Paap

Wal

drk

Fst

atN

/A12

.70

12.7

620

.97

10.7

6

Obs

erva

tions

335

308

303

303

309

This

tabl

esh

ows

the

that

the

mai

nre

sults

are

sim

ilar

ifa

linea

r-lin

ear

func

tiona

lfor

mis

used

(i.e.

,no

tlo

gpo

vert

y).

Star

sde

note

stat

istic

alsi

gnifi

canc

eat

the

10,

5,an

d1

per

cent

leve

ls.

All

IVes

timat

esus

eth

etw

o-st

age

leas

tsqu

ares

inst

rum

enta

lvar

iabl

ees

timat

or.

Sam

ple

isth

elo

ng-d

iffer

ence

cros

s-se

ctio

nof

alla

vaila

ble

dist

ricts

from

2002

–201

0.20

01di

stric

tbou

ndar

ies

are

used

,with

new

dist

ricts

colla

psed

into

year

-200

1pa

rent

dist

ricts

.Cha

nges

insa

mpl

essi

zear

edu

eto

data

avai

labi

lity.

Oil

palm

land

isla

gged

onep

erio

d(2

001–

2009

).H

eter

oske

dast

icity

-rob

ust

stan

dard

erro

rsar

ein

pare

nthe

ses.

Isla

ndgr

oupi

ngs

are

defin

edas

Java

,Sum

atra

,Kal

iman

tan,

and

Sula

wes

i,w

ithre

mai

ning

dist

ricts

grou

ped

toge

ther

.Ea

chre

gres

sion

incl

udes

log

pove

rty

and

log

perc

apita

outp

utin

the

initi

alpe

riod

asco

ntro

lvar

iabl

es.


Tabl

e3.

14:M

ain

Resu

lts—

Prod

ucti

onIn

stea

dof

Land

Dep

ende

ntva

riabl

eLo

gdi

stric

tpov

erty

rate

(%)

Log

dist

rictp

over

tyga

pin

dex

(IDR)

Estim

ator

FEFE

LIM

LLD

LDLI

ML

FEFE

LIM

LLD

LDLI

ML

Col

umn

12

34

56

78

Perc

apita

palm

oilp

rodu

ctio

n(to

ns)

-0.2

12**

*-0

.594

***

-0.1

95**

*-0

.427

***

-0.2

41**

*-0

.813

***

-0.2

62**

*-0

.565

***

(0.0

54)

(0.1

59)

(0.0

43)

(0.1

18)

(0.0

54)

(0.2

29)

(0.0

69)

(0.2

06)

Dis

tric

tand

year

FEs

YY

NN

YY

NN

Isla

nd–y

earF

EsY

NN

NY

NN

N

Initi

alco

nditi

onsc

ontr

ols

NN

YY

NN

YY

Excl

uded

-Fst

atis

tic8.

0312

.01

8.3

12.0

1

Dis

tric

ts34

134

133

527

434

134

133

527

4

Obs

erva

tions

3386

3386

335

274

3051

3051

335

274

This

tabl

esh

ows

the

effec

tsob

serv

edfo

rad

ditio

nalp

alm

oill

and

asa

shar

eof

tota

ldis

tric

tare

aca

rry

over

toac

tual

palm

oilp

rodu

ctio

nin

tons

,and

conv

erte

dto

perc

apita

term

sto

scal

e.St

arsd

enot

est

atis

tical

sign

ifica

nce

atth

e10

,5,a

nd1

perc

entl

evel

s.Sa

mpl

eis

anan

nual

341

dist

rictp

anel

,200

2-20

10.

Palm

oill

and

isla

gged

one

perio

d(i.

e.,2

001-

2009

).20

01di

stric

tbou

ndar

ies

are

used

,with

new

dist

ricts

colla

psed

into

year

2001

pare

ntdi

stric

ts.

LIM

Lre

fers

toth

elim

ited

info

rmat

ion

max

imum

likel

ihoo

din

stru

men

talv

aria

ble

estim

ator

,FE

fixed

effec

ts,

and

LDlo

ng-d

iffer

ence

s.Het

eros

keda

stic

ity-r

obus

tsta

ndar

der

rors

are

inpa

rent

hese

s,cl

uste

red

atth

edi

stric

tlev

elin

forF

Ees

timat

ean

dat

the

prov

ince

leve

lfor

LDs.

FELI

ML

inst

rum

ents

oilp

alm

prod

uctio

nw

ithth

ein

itial

dist

ricts

hare

ofpa

lmoi

llan

din

tera

cted

with

atim

etr

end,

and

LDLI

ML

dist

ricta

gro-

clim

atic

suita

bilit

yfo

roi

lpal

m.

Initi

alco

nditi

onco

ntro

lsar

ein

itial

pove

rty

rate

s,in

itial

per

capi

tain

com

es,a

ndis

land

dum

mie

s.Ex

clud

ed-F

refe

rsto

the

Kle

iber

gen-

Paap

Wal

drk

Fst

atis

ticob

tain

edfr

omfir

st-s

tage

regr

essi

ons.


Table 3.15: Determinants of Changing Oil Palm Land Shares

Dependent variable Palm oil land / district area (%)

Estimator Pooled OLS Within FE Within FE

Column 1 2 3

Lag electricity capacity-0.0004** -0.0002 -0.0002

(0.0002) (0.0002) (0.0002)

Lag electricity capacity nearby-0.0003* -0.0003* 0001

(0.0001) (0.0002) (0.0001)

Lag access to electricity-0.010 0.046*** 0.016

(0.010) (0.015) (0.019)

Lag human development index-0.180*** 0.243*** 0.220

(0.058) (0.086) (0.243)

Lag child immunisation rate-0.168*** 0.003 -0.0001

(0.036) (0.010) (0.010)

Lag adult literacy rate0.245*** -0.088** -0.036

(0.026) (0.036) (0.036)

Lag skilled birth0.046*** 0.001 0.005

(0.015) (0.013) (0.013)

District FEs N Y Y

Island-year FEs N N Y

Observations 1019 1019 1019

R-squared 0.14 0.10 0.20

Stars denote statistical significance at the 10, 5, and 1 per centlevels. Sample is an annual 341 district panel, 2002–2010. Palm oilland is lagged one period (i.e., 2001–2009). 2001 district boundariesare used, with new districts collapsed into year-2001 parent districts.Heteroskedasticity-robust standard errors are in parentheses, clustered atthe district level. Covariates are all taken from the World Bank (2015)Indonesia Database for Economic and Policy Research.


Tabl

e3.

16:R

obus

tnes

s—Pa

nel

Fixe

dEf

fect

s

Dep

ende

ntva

riabl

ePa

lmla

ndLo

gdi

stric

tpov

erty

rate

Col

umn

12

34

56

78

9

Palm

oill

and

/di

stric

tare

a0.

001

-0.0

06**

*-0

.008

***

-0.0

10**

*-0

.009

***

-0.0

07**

*-0

.011

***

-0.0

04*

(0.0

01)

(0.0

02)

(0.0

02)

(0.0

03)

(0.0

03)

(0.0

02)

(0.0

03)

(0.0

02)

Lag

log

pove

rty

rate

0.23

50.

412*

**

(0.3

62)

(0.0

57)

2nd

lag

log

pove

rty

rate

0.45

1

(0.6

17)

In-ti

me

plac

ebo

NY

NN

NN

NN

N

Lag

pove

rty

cont

rols

NN

YN

NN

NN

N

Elec

tric

ityco

ntro

lsN

NN

YN

NN

NN

Polit

ical

,fisc

al,o

il,ga

scon

trol

sN

NN

NY

NN

NN

Fore

st&

polit

ical

cont

rols

NN

NN

NY

NN

N

Inco

me

and

reve

nue

cont

rols

NN

NN

NN

YN

N

Cor

rela

teso

fpov

erty

cont

rols

NN

NN

NN

NY

N

Dis

tric

t-by-

dist

rictt

ime

tren

dsN

NN

NN

NN

NY

Dis

tric

tand

isla

nd-y

earF

EsY

YY

YY

YY

YY

Obs

erva

tions

2359

2704

2699

2321

1445

1445

3386

1333

3386

Star

sden

ote

stat

istic

alsi

gnifi

canc

eat

the

10,5

,and

1pe

rcen

tlev

els.

Sam

ple

isan

annu

al34

1di

stric

tpan

el,2

002–

2010

.Pal

moi

llan

dis

lagg

edon

epe

riod

(i.e.

,200

1–20

09).

2001

dist

rictb

ound

arie

sar

eus

ed,w

ithne

wdi

stric

tsco

llaps

edin

toye

ar-2

001

pare

ntdi

stric

ts.

With

ines

timat

orw

ithdi

stric

tand

isla

nd–y

earF

Esus

edfo

rall

estim

ates

.Het

eros

keda

stic

ity-r

obus

tsta

ndar

der

rors

arei

npa

rent

hese

s,cl

uste

red

atth

edi

stric

tlev

el.C

olum

n1

regr

esse

spov

erty

lags

onpa

lmla

ndan

dth

ein

-tim

epla

cebo

test

inC

olum

n2

uses

futu

repa

lmoi

llan

dva

lues

.Ele

ctric

ityco

ntro

lsre

fert

odi

stric

tand

neig

hbou

ring

dist

rictp

ower

capa

city

,tak

enfr

omSp

arro

wet

al(2

015)

.Pol

itica

l,fis

cal,

oil,

and

gas

cont

rols

are

take

nfr

omBu

rges

set

al(2

012)

and

fore

stco

ntro

lsfr

omW

heel

eret

al(2

013)

.All

othe

rcov

aria

tes

are

take

nfr

omth

eW

orld

Bank

(201

5)In

done

sia

Dat

abas

efo

rEco

nom

ican

dPo

licy

Rese

arch

.


Tabl

e3.

17:R

obus

tnes

s–A

lter

nati

veSa

mpl

es

Dep

var:

log

dist

rictp

over

tyra

te

Sam

ple

No

Java

2006

–201

020

01–2

005

Estim

ator

FEFE

LIM

LLD

LDLI

ML

FEFE

LIM

LFE

FELI

ML

Col

umn

12

34

56

78

Palm

oill

and

/di

stric

tare

a(%

)-0

.007

***

-0.0

39**

-0.0

12**

*-0

.012

**-0

.006

**-0

.056

*-0

.002

-0.1

16**

*

(0.0

02)

(0.0

16)

(0.0

03)

(0.0

06)

(0.0

03)

(0.0

34)

(0.0

02)

(0.0

34)

Dis

tric

tand

year

fixed

effec

tsY

YN

NY

YY

Y

Isla

nd–y

earfi

xed

effec

tsY

NN

NY

NY

N

Add

ition

alco

ntro

lsN

NY

YN

NN

N

Excl

uded

-Fst

atis

tic7.

770

14.9

84.

185

17.0

6

Obs

erva

tions

2295

2295

230

175

1683

1683

2044

2044

Col

umns

1–4

show

that

my

mai

nes

timat

esin

Tabl

e3.

2an

dth

ead

ditio

nali

nstr

umen

talv

aria

ble

estim

ates

are

robu

stto

incl

udin

gIn

done

sia’

sm

ostp

opul

ous

isla

ndw

ithlit

tlepa

lmoi

llan

d,Ja

va.

Col

umns

5–8

use

the

first

and

seco

ndha

lfof

the

sam

ple

perio

d,as

ther

ew

asan

incr

ease

inov

eral

lpal

moi

lpro

duct

ivity

betw

een

the

perio

ds.

Star

sde

note

stat

istic

alsi

gnifi

canc

eat

the

10,5

,and

1pe

rce

ntle

vels

.Sa

mpl

eis

anan

nual

341

dist

rictp

anel

,200

2-20

10.

Palm

oill

and

isla

gged

one

perio

d(i.

e.,2

001-

2009

).20

01di

stric

tbo

unda

ries

are

used

,with

new

dist

ricts

colla

psed

into

year

2001

pare

ntdi

stric

ts.

LIM

Lre

fers

toth

elim

ited

info

rmat

ion

max

imum

likel

ihoo

din

stru

men

talv

aria

ble

estim

ator

,FE

fixed

effec

ts,a

ndLD

long

-diff

eren

ces.H

eter

oske

dast

icity

-rob

usts

tand

ard

erro

rsar

ein

pare

nthe

ses,

clus

tere

dat

the

dist

rictl

evel

info

rFE

estim

ate

and

atth

epr

ovin

cele

velf

orLD

s.FE

LIM

Lin

stru

men

tsoi

lpal

mpr

oduc

tion

with

the

initi

aldi

stric

tsha

reof

palm

oill

and

inte

ract

edw

itha

time

tren

d,an

dLD

LIM

Ldi

stric

tagr

o-cl

imat

icsu

itabi

lity

foro

ilpa

lm.I

nitia

lcon

ditio

nco

ntro

lsar

ein

itial

pove

rty

rate

s,in

itial

perc

apita

inco

mes

,and

isla

nddu

mm

ies.

Excl

uded

-Fre

fers

toth

eK

leib

erge

n-Pa

apW

ald

rkF

stat

istic

obta

ined

from

first

-sta

gere

gres

sion

s.


Table 3.18: Heterogeneity–By Region

Dependent variable: log district poverty rateSample Island Island AllEstimator FE LD FEColumn 1 2 3Panel A: Java

Palm oil land / district area 0.015 -0.035 -0.007***(0.016) (0.047) (0.002)

Island–palm interaction 0.021(0.015)

N observations 1091 105 3386Panel B: Sumatra

Palm oil land / district area -0.007*** -0.011*** -0.016***(0.002) (0.004) (0.005)

Island–palm interaction 0.010**(0.005)

N observations 960 96 3386Panel C: Kalimantan

Palm oil land / district area -0.007* -0.009** -0.006***(0.003) (0.004) (0.001)

Island–palm interaction -0.008**(0.004)

N observations 379 37 3386Panel D: Sulawesi

Palm oil land / district area -0.05*** -0.05*** -0.006***(0.008) (0.016) (0.002)

Island–palm interaction -0.039***(0.007)

N observations 450 45 3386Panel E: Other islands

Palm oil land / district area -0.039 0.345** -0.007***(0.130) (0.171) (0.002)

Island–palm interaction -0.026(0.128)

N observations 506 51 3386

This table provides consonant sub-samples estimates by region tosupplement to the full-sample estimates with interaction termspresented in the paper. Stars denote statistical significance atthe 10, 5, and 1 per cent levels. Full sample (Column 3) is anannual 341 district panel, 2002–2010. Palm oil land is lagged oneperiod (i.e., 2001–2009). 2001 district boundaries are used, withnew districts collapsed into year 2001 parent districts. Estimatorsare the within estimator (FE) with district and year fixed effects,and the long-difference (LD) estimator with initial log povertyand log per capita income controls. Heteroskedasticity-robuststandard errors are in parentheses, clustered at the district level.


Tabl

e3.

19:H

eter

ogen

eity

–By

Sect

or

Dep

var:

log

pove

rty

rate

Pane

lwid

thA

nnua

l4-

year

lyA

nnua

l4-

year

lyA

nnua

l4-

year

lyA

nnua

l4-

year

ly

Col

umn

12

34

56

78

Stat

e-ow

ned

palm

oill

and

-0.0

11**

-0.0

11**

-0.0

03-0

.002

(0.0

04)

(0.0

04)

(0.0

05)

(0.0

06)

Priv

ate

palm

oill

and

-0.0

12**

*-0

.011

***

-0.0

12**

*-0

.009

**

(0.0

03)

(0.0

03)

(0.0

03)

(0.0

05)

Smal

lhol

derp

alm

oill

and

-0.0

04-0

.011

**0.

002

-0.0

04

(0.0

02)

(0.0

05)

(0.0

02)

(0.0

06)

Dis

tric

tand

isla

nd-y

earF

EsY

YY

YY

YY

Y

Obs

erva

tions

3009

1004

3009

1004

3009

1004

3009

1004

This

tabl

epr

esen

tsth

esa

me

sect

oral

estim

ates

asth

ose

inth

epa

per,

buta

dditi

onal

Col

umns

7an

d8,

whi

chsh

owth

atth

epo

vert

yim

pact

sofp

rivat

epa

lmoi

llan

ddo

min

ate

thos

efr

omot

hers

ecto

rsw

hen

alls

ecto

rsar

ein

clud

edin

the

one

regr

essi

on.

Palm

oill

and

varia

bles

are

allp

lant

atio

nar

eash

ares

ofto

tald

istr

icta

reas

.Sta

rsde

note

stat

istic

alsi

gnifi

canc

eat

the

10,5

,and

1pe

rcen

tlev

els.

Sam

ple

isan

annu

al34

1di

stric

tpan

el,2

002-

2010

.Pa

lmoi

llan

dis

lagg

edon

epe

riod

(i.e.

,200

1-20

09).

2001

dist

rictb

ound

arie

sar

eus

ed,w

ithne

wdi

stric

tsco

llaps

edin

toye

ar20

01pa

rent

dist

ricts

.Het

eros

keda

stic

itiy-

robu

stst

anda

rder

rors

are

inpa

rent

hese

s,cl

uste

red

atth

edi

stric

tlev

el.P

anel

fixed

effec

tses

timat

orus

edth

roug

hout

,with

dist

rict,

year

,and

isla

nd-y

earfi

xed

effec

ts.


Tabl

e3.

20:H

eter

ogen

eity

–By

Land

Qua

lity

Dep

ende

ntva

riabl

e:lo

gdi

stric

tpov

erty

rate

Palm

oill

and

qual

ityD

amag

edIm

mat

ure

Mat

ure

Pane

lwid

thA

nnua

l2-

year

ly4-

year

lyA

nnua

l2-

year

ly4-

year

lyA

nnua

l2-

year

ly4-

year

ly

Col

umn

12

34

56

78

9

Palm

oill

and/

dist

ricta

rea

(%)

-0.1

42**

-0.1

41**

-0.0

74-0

.025

***

-0.0

25**

*-0

.035

***

-0.0

14**

*-0

.014

***

-0.0

11**

(0.0

66)

(0.0

69)

(0.0

83)

(0.0

09)

(0.0

09)

(0.0

13)

(0.0

04)

(0.0

05)

(0.0

04)

Obs

erva

tions

3009

1674

1004

3009

1674

1004

3009

1674

1004

This

tabl

esh

owsh

owin

crea

sing

the

shar

eof

dist

rictl

and

used

ford

amag

edpl

anta

tions

hast

hela

rges

tsho

rt-te

rmim

pact

son

dist

rictp

over

ty,

tape

ring

offov

ertim

e.Im

mat

ure

and

mat

ure

plan

tatio

nsha

vebr

oadl

yco

nsis

tent

impa

cts

over

4ye

ars:

the

form

ersl

ight

lyin

crea

ses

(per

haps

from

“lea

rnin

gby

doin

gan

dth

ein

crea

sed

prod

uctiv

ityfr

omtr

eesc

omin

gof

age)

and

the

latte

rslig

htly

decr

ease

s(as

pove

rty

gain

swou

ldha

veal

read

ybe

enm

ade)

.St

ars

deno

test

atis

tical

sign

ifica

nce

atth

e10

,5,a

nd1

per

cent

leve

ls.

Sam

ple

isan

annu

al34

1di

stric

tpan

el,2

002-

2010

.Pa

lmoi

llan

dis

lagg

edon

epe

riod

(i.e.

,200

1-20

09).

2001

dist

rict

boun

darie

sar

eus

ed,w

ithne

wdi

stric

tsco

llaps

edin

toye

ar20

01pa

rent

dist

ricts

.Het

eros

keda

stic

itiy-

robu

stst

anda

rder

rors

are

inpa

rent

hese

s,cl

uste

red

atth

edi

stric

tlev

el.T

hew

ithin

pane

lest

imat

orw

ithdi

stric

tan

dis

land

-by-

year

fixed

effec

tsis

used

inal

lest

imat

es.

Chapter 4

Local impacts of resource booms

Abstract

I study the local economic and welfare impacts of three of Indonesia’s largest

natural resource sector booms. Applying the synthetic control method to

district-level data over the 2000s, I construct non-parametric counterfactual

estimates of the local impacts of resource booms in three districts experiencing

some of the largest and steepest increases in resource sector output. I examine

palm oil in Sumatra, coal mining in Kalimantan, and natural gas extraction in

West Papua. All three resource booms boosted total economic output. Oil

palm expansion in Riau raised agricultural, industrial, and services output. Coal

mining in South Kalimantan reduced agriculture and services output. Oil palm

and coal mining booms appear to have delivered strong local poverty reduction.

The Tangguh natural gas project in West Papua delivered a massive increase in

industry and aggregate output, but appears to have had indiscernible impacts on

household welfare or poverty.

121


4.1 Introduction

Indonesia is the world’s largest exporter of coal and palm oil, but the long-term

development implications of Indonesia’s sustained resource-driven growth are

still not well understood (Burke and Resosudarmo, 2012; Garnaut, 2015; Hill

et al., 2008; van der Eng, 2014). With some of the world’s most decentralised

governance arrangements, understanding region-specific impacts of booming

resource sectors is critically important for economic, social, and environmental

policy and longer-term development strategies.

This chapter asks how rapid natural resource sector expansion affects a local

district economy and its residents’ welfare. I present a quantitative case study

on each of Indonesia’s three largest natural resource exports–coal ($22.9B; 11% of

merchandise exports in 2014), natural gas ($17.4B; 8.5%), and palm oil ($16.5B;

8.1%)–and exploit some of the largest and most sudden increases in district-level

production for each commodity: palm oil in Indragiri Hilir, Riau; coal mining in

Tapin, South Kalimantan; and a giant natural gas project in the Bintuni Bay of

West Papua. All three sectors have been argued to be economic enclaves, but have

starkly different characteristics. Oil palm is labour- and land-intensive and spread

amongst government, private, and smallholder farmers. Natural gas extraction

and coal mining are capital-intensive, generate relatively little employment, and

have highly concentrated rents. My key hypothesis is that the more diffuse oil

palm sector should generate broader-based benefits than point resource sectors,

particularly natural gas. An appropriate control group is needed to compare the

observed district outcomes with the same district’s unobserved counterfactual

outcomes (“untreated”). I use a relatively new method—synthetic control

modelling—to construct a “synthetic” comparison district for each resource boom

district, allowing me to compare the booming districts’ observed outcomes with

reasonable counterfactuals. Each synthetic control district is a weighted average

of untreated districts with similar pre-treatment observable characteristics and


outcome behaviour to the treated district. Optimal weights are determined

through a data-driven algorithm minimising the mean squared prediction error

over the pre-treatment period, resulting in a synthetic control more closely

resembling the treated district than any other single comparison district or group

of districts.

I find that all three resource booms boosted total economic output and

significantly altered the structure of the local economy. Oil palm expansion

in Indragiri Hilir raised agricultural, industrial, and services output. The

poverty rate declined substantially while average household expenditure fell

below its counterfactual—a puzzle I discuss further in the chapter. A major

coal mining expansion in Tapin reduced agriculture and services sector output

while delivering strong poverty reduction. Natural gas extraction in West

Papua delivered a giant boost to mining, industry, and total output in one of

Indonesia’s most developmentally lagging regions. Small, positive services sector

spillovers are observed, but agricultural output diverged below its estimated

counterfactual. No change in the pre-existing poverty and average household

expenditure trends implies negligible welfare impacts in light of the massive ramp

up in district per capita output. My three case studies collectively suggest that

more diffuse natural resources tend to generate broad-based, inequality-reducing

economic development. The benefits from extracting point resources appear

more concentrated, suggesting a greater role for social policy and economic

diversification in areas dominated by these sectors.

Existing studies on the local impacts of natural resources typically fall into

two categories: small-scale qualitative case studies (see, e.g., McCarthy et al

(2010; 2011), Rist et al (2010), and Budidarsono et al (2012) for palm oil), or

larger-scale statistical studies (see Cust and Poelhekke (2015), Torvik (2009),

van der Ploeg (2011), and Wick and Bulte (2009) for mining). Qualitative

narratives and geographically-narrow case studies provide a rich source of


descriptive evidence, but are generally unable to quantify impacts and provide

little basis for causal inference. Econometric studies—while sometimes providing

a causal interpretation of the data—can often lack context, assume homogeneous

treatments, and mask heterogeneity across treated units. One of the few points

of agreement in the resource curse literature is that the development impacts of

natural resources have been heterogeneous across time, commodities, country

characteristics, and local conditions.

My main contribution in this chapter is to provide new quantitative case

study evidence of the local economic and welfare impacts of rapid resource sector

expansion in Indonesia’s three largest export commodities: coal, natural gas, and

palm oil. In doing this, I provide new micro-level evidence on the poverty and

welfare elasticity of resource sector growth, and on its potential inter-sectoral

linkages (i.e., spillovers). With sub-national panel data becoming increasingly

available, I also show how the synthetic control method can be a useful tool to

analyse the impacts of major policies and economic shocks to single administrative

units, particularly in Indonesia where many decisions are made at the district and

province levels.

The chapter is structured as follows. Section 4.2 explains synthetic control

modelling and its application to sub-national data from Indonesia. Section 4.3

presents the results for my three case studies. Section 4.4 concludes.

§4.2 Synthetic control approach 125

4.2 Synthetic control approach

To make causal inferences about the local impacts of palm oil plantations,

mines, or other resource-related activities through case study-based research

we must ask what the area would look like without these activities—a difficult

question to answer. Many factors affect local economic outcomes, from

government policies, to private investment, to weather. Finding a single

comparison unit free of the treatment but still sufficiently similar to the treated

unit is often problematic. Before-and-after analyses tend to be contaminated by

other changes in the local economy, and differencing out trends cannot account

for systematic differences in growth paths.

One approach to deal with these problems is the synthetic control method

proposed by Abadie and Gardeazabal (2003). The synthetic control method is a

systematic way of choosing case study comparison units, allowing quantitative

causal inferences in small samples, usually a single “treated” unit. It uses

data from multiple comparison units to construct a single synthetic comparison

unit—a weighted average—that most closely resembles the unit of interest before

the event under study, i.e., exhibiting the same pre-treatment dynamics. If a

synthetic district can be generated based on a set of similar districts and valid

predictor variables, and a resource boom has substantial impacts on the local

economy, then the economic trajectories of the district experiencing the resource

boom and its synthetic control should be similar before the resource boom but

diverge after.

The usefulness of the synthetic control method has been demonstrated

across a variety of applications. At the national level, Pepinski and Wihardja

(2011) examine the economic impact of Indonesia’s decentralisation in 1999.

Mideksa (2013) quantifies the positive economic impact of Norway’s natural

resource endowment over a long time period. Smith (2015) applies the same

approach to many resource-rich countries, finding that countries that have


become resource-rich have subsequently attained higher per capita incomes.57 My

study relates closely to Mideksa (2013) and Smith (2015) but at a sub-national level

and focusing on indicators beyond per capita output.

Within-country applications are becoming popular as well. Synthetic control

analysis is particularly well-suited to micro-level studies, as within-country units

share some of the same external shocks (e.g., related to national-level policies

and the economy). In the seminal study introducing the method, Abadie and

Gardeazabal (2003) estimate the economic effects of terrorism in the Basque

Country of northern Spain. The closest studies to mine are those of Munasib

and Rickman (2015), who use the synthetic control method to look at the regional

economic and welfare impacts of shale gas in a few US counties, and Sills et al

(2015), who use sub-national data from Brazilian municipalities to examine the

effectiveness of a local anti-deforestation policy.58 To my knowledge, mine is the

second application of the synthetic control method to sub-national data from a

developing country after Sills et al.’s (2015) Brazilian study.

4.2.1 Estimation and inference

Consider d = 1 as the district experiencing a resource boom (the treated

district) in a sample ofD+ 1 districts. d = 2 to d = D+ 1 are potential comparison

units in a comparison pool. Yd,t is the outcome of interest for district d at time

57Additional national level studies include Abadie, Diamond and Hainmueller’s (2014) studyof the impact of German reunification in 1990 on West Germany’s economic growth, Horiuchiand Mayerson’s (2015) study of the economic costs of conflict in Israel following from the SecondIntifada, Billmeier and Nannicini’s (2013) study of the impacts of trade liberalisation on per capitaoutput, and Karlsson and Pichler’s (2015) work on the demographic impacts of HIV.

58Other sub-national applications include Montalvo’s (2011) study of the electoral impacts ofthe 2004 Madrid bombings, Abadie, Diamond and Hainmueller’s (2010) study of the impact of alarge-sale tobacco control program on cigarette sales in California, and Bohn et al’s (2014) study onthe impact of the 2007 Legal Arizona Workers Act on the state’s population composition. Chan et al(2014) construct synthetic academics and estimate the impact of the John Bates Clark Medal andthe Fellowship of the Econometric society on academic performance, Bauhoff (2014) look at thedietary impacts of nutritional standards in schools, Barone and Mocetti (2014) look at the regionaleconomic impacts of two earthquakes in Italy, Pinotti (2015) examines the impacts of mafia activityin Italy, Ando (2015) looks at the local impacts of new power plants in Japan, and Krief et al (2015)re-examine a difference-in-difference evaluation of a major UK health policy without assumingparallel trends.


t. A synthetic control district for d = 1 is constructed as a weighted average of

comparison pool districts, using a (D × 1) vector of weights:

W = (w2, ..., wd+1)′ (4.1)

with 0 ≤ wd ≤ 1 for d = 2, ..., D + 1 and w2 + ...+wd+1 = 1. W ∗ is selected through

an optimisation procedure minimising:

(X1 −X0W )′V (X1 −X0W ) (4.2)

where X1 is a vector of pre-treatment variables for my case study district, X0 is

the corresponding matrix of the same variables for D comparison pool districts,

and V a diagonal matrix reflecting the relative importance of the different Xs.

A data-driven algorithm minimises the pre-treatment differences between the

outcomes of interest and results in a single comparison district resembling

the treated district—in outcome levels, outcome behaviour, and observable

covariates—better than any single comparison district or group of districts. If a

good pre-treatment fit between the case study district and its synthetic control is

achieved, differences in post-treatment outcomes can be assumed to be a result of

the treatment. The treatment effect for district d = 1 in the post-boom period is:

Y1,t − ΣD+1d=2 w

∗dYd,t (4.3)

where w∗d are the optimal weights.

A key strength of the synthetic control method is that it is a generalised

difference-in-difference approach allowing quantitative causal inference on a

single treated unit. But unlike difference-in-difference and panel fixed effects

estimators capturing only time-invariant confounders, well-constructed synthetic

controls do not rely on a parallel trends assumption and are robust to time-varying


unobservable confounders (Abadie et al, 2010). A principal reason to use synthetic

controls is to control for the effect of unobservable factors potentially influencing

time trends in the treated and control units (Acemoglu et al, 2014; Smith,

2015). Unlike regression and matching on baseline covariates, synthetic controls

compare units similar in terms of baseline covariates plus the behaviour of the

outcome of interest in the pre-treatment period and how the covariates shape

this behaviour. The credibility of the method depends on being able to construct

a close comparison unit falling within the convex hull of the treated unit, then

arguing the divergence in outcomes between the treated unit and the synthetic

control is due to the intervention of interest (i.e., the treatment is the only major

change affecting the treated unit).59

The key limitation of the synthetic control approach is that conventional

methods of assessing statistical significance (i.e., based on large-sample inferential

techniques) are not available due to the smaller number of time periods and

control pool observations. The three cases in this Chapter are no exception.

However, non-parametric treatment effects obtained from synthetic control

analyses can be subjected to a range of falsification tests, leading to what Abadie et

al (2010) term “exact inference” from examining the full distribution of treatment

effects obtained by iteratively applying the treatment to each unit in the sample

(akin to random permutation testing and inference). Such placebo tests can be

conducted across time periods (Bertrand et al, 2004), across donor pool units, on

the synthetic control itself, and on alternative outcomes plausibly unaffected by

the treatment.

59See Abadie and Gardeazabal (2003) and Abadie, Diamond, and Hainmueller (2010; 2011;2014) for more technical treatments of the synthetic control method, including proofs andalternative approaches to constructing V .


Applying the synthetic control method to district-level resource booms in

Indonesia is arguably more challenging than previous applications for two

reasons. Sub-national data from developing countries tend to be more porous

and volatile than the national accounts and within-country data used in previous

studies. This is important because the approach has historically required clean,

slow-moving data for the optimisation procedure to function well and separate the

treatment effect from idiosyncratic variations in the outcome of interest. Abadie

et al (2010) go as far as recommending smoothing volatile outcome variables

to minimise idiosyncratic variation. Given the imperfections already laden in

the data, I avoid this practice.60 Secondly, many districts of Indonesia have

experienced resource booms of varying size and scope. Care must be taken to

identify resource expansions appropriately interpreted as discrete shocks (i.e.,

dichotomous treatments), and large restrictions must be placed on the comparison

pool to ensure sufficiently similar but untreated districts.

4.2.2 Data

A balanced panel of Indonesian districts

The constrained optimisation procedure to obtain the W ∗ weights requires

data for matrices X1 and X0. Balanced panel data are needed for the outcome

variables for the treated unit and all potential control units in the comparison

pool. Additional covariates (predictors) must also be available for all units for at

least one pre-treatment period. All data for this study are taken from the World

Bank’s (2015) Indonesia Database for Policy and Economic Research (DAPOER).

DAPOER is a public sub-national database covering over 200 socio-economic

variables across Indonesia’s 33 provinces and over 500 districts (kabupatens).

60McCulloch and Sjahrir (2008) McCulloch and Malesky (2011) discuss some of themeasurement issues associated with Indonesian subnational accounts data. Indonesian statisticscan however be considered relatively reliable by developing country standards.


My three case studies are conducted at the district level to identify the

aggregate impacts of resource booms on the local economy and residents’ welfare.

Indonesian districts are clearly defined legal and geographical units with district

administrations reflecting local economies, but the number of districts has not

remained constant over time. Power was decentralised to around 300 district

governments with the fall of Suharto in 1998, driving the number of districts to

proliferate from 292 in 1998 to over 500 in 2015 (Fitriani et al, 2005). To obtain

a nationally-exhaustive balanced panel of constant geographic units, I apply

year-2001 district boundaries to my dataset. The average district in my balanced

panel had a population of 1.4 million people in 2011, with the smallest district

having a population of just 21,000.

The study period is as long as the data permit for each estimate,

as longer treatment periods improve the accuracy of estimation (Abadie,

Diamond, and Hainmueller, 2010).61 Most district-level data are only available

post-decentralisation, so I focus on resource expansions from 2005 to allow a

sufficient pre-treatment window to fit the synthetic controls. Post-treatment

periods are extended as long as data permit to allow treatment effects to emerge

gradually and to examine dynamics.62

61Recall the credibility of the approach depends on how well the synthetic control tracks thetreated unit before the treatment, with this demonstrated for as many periods as possible. Aneffective synthetic control does not only match the level and the trend of the treated unit in thepre-treatment period, but its behaviour and sensitivity to time-varying conditions (e.g., businesscycles).

62The study period includes commodity price booms and the global financial crisis. Districtswithout commodities would not benefit from their prices, so the total treatment effect naturallyincludes some price effects. Commodity price booms and financial crises only bias results ifresponsible for a structural break in how unobservables map to the outcome after the treatment.


Outcomes of interest

My economic outcomes of interest are real regional gross domestic product

(RGDP) per capita and its components. I follow standard United Nations

sectoral classifications. The primary sector comprises agriculture, forestry, fishing

(one item in Indonesian sub-national accounts), and mining and quarrying.

Manufacturing, construction, and electricity, gas, water, and sanitary services

(utilities) are the secondary sector (industry). The tertiary sector—broadly

defined as services—consists of transport, storage, and communications,

wholesale and retail trade, banking, insurance and real estate, ownership of

dwellings, public administration and defence, services, and all other activities. I

am interested in the effects of resource booms on the rest of the primary sector (i.e.,

agriculture in the case of mining), industry, and services (excluding government).

Examining impacts on different sectors allows me to gauge any local structural

change and spillovers through: (a) production linkages (producing commodity

sector inputs or processing raw materials); (b) consumption linkages (greater

demand for non-resource sector output from resource sector income); or (c)

crowding-out non-resource sectors (i.e., (a) and (b) in reverse).63 Output variables

are converted to per capita terms to take population size into account and to

better facilitate matching. RGDP in million Indonesian rupiah (IDR, constant

2000 prices) and total district population data are taken from Indonesia’s central

statistics agency, Badan Pusat Statistik (BPS). While most national accounts are

subject to some measurement and imputation error, Indonesian statistics are

generally better than many other developing countries’ national accounts data

(McCulloch and Sjahrir, 2008; McCulloch and Malesky, 2011).

63Fiscal linkages, the third of Hirschman’s (1981) linkages in addition to production andconsumption linkages, are beyond the scope of this study.


To assess broader welfare and distributional impacts I also examine average

per capita monthly household expenditure and the district poverty rate. Per

capita household expenditures (monthly IDR) are derived from the consumption

module of BPS’s high quality, district-representative national socio-economic

survey (SUSENAS). SUSENAS is implemented at least annually (usually at a

similar time each year) and covered around 1.2 million people in 2011.64 The

poverty rate is the share of the total district population whose expenditure level is

below the poverty line: the central social policy target for Indonesian governments

and development agencies. The expenditure-based poverty line varies by district

and period, linked by a universal consumption requirement.

A key concern when selecting outcome indicators is that sufficient time is

allowed for impacts to be quantitatively seen. Although the full impacts of

any resource booms will not likely be realised for years, output, poverty, and

household expenditure should be relatively responsive. The synthetic control

method allows me to evaluate their short-run dynamics over time. Long-term

effects may be different.

4.2.3 Identifying appropriate case studies

Resource booms must be of sufficient magnitude to cause noticeable impacts

on the outcomes of interest relative to any idiosyncratic shocks, particularly in

comparative case studies where small effects are often indistinguishable from

random shocks.

I separately sorted district-year observations by the two-year change in mining

and quarrying output, oil and gas output, and palm oil production to identify the

district-level resource boom for each of Indonesia’s three largest natural resource

64I leave household expenditure in nominal terms. This may be misleading if a resource boomleads to local (district-specific) inflation, for example due to local Dutch disease dynamics. As pricedata are available infrequently for only major cities, I cannot assess year-to-year inflation in mycase study districts and in comparison with other districts. However, there is no reason to expectthe law of one price not to hold, and cities (kotas) nearest to my case study districts (Pekanbaru,Banjarmasin, and Manokwari) show no abnormalities relative to other cities and national trends.


export commodities most suitable for synthetic control analysis.65

The synthetic control approach relies on a sharp discontinuity or event (i.e., a

structural break) to split the observed outcome from the simulated counterfactual.

To create a dichotomous treatment environment from my continuous treatment

variables, I identify districts with consistently low or non-existent resource

production in the early 2000s and a large and rapid production scale-up followed

by persistently high production. I visually inspect the time series for districts with

the largest short-term changes in mining, oil and gas, and palm oil production to

restrict potential case studies to those where there is just the one major production

increase during the period under study. Assuming I fit valid synthetic controls,

differences between the treated districts’ outcomes and the synthetic controls in

the post-treatment periods can be interpreted as causal effects if the resource

boom is the only major event affecting the treated unit. Thus districts with

resource booms for other commodities (e.g., with oil and palm oil, as is the case in

some districts in Riau province) and other district-specific economic shocks (e.g.,

natural disasters) are removed.

I identify (a) palm oil expansion in the Indragiri Hilir district of Riau province,

(b) coal mining expansion in Tapin district of South Kalimantan, and (c) the

commencement of natural gas extraction in Manokwari, West Papua, as the three

district-level resource booms most suitable for synthetic control study (see Figure

4.1). Table 4.1 compares case study districts’ characteristics to each other and

national averages in 2005, the last common pre-boom year.

65Sorting was done in level changes in production output and checked against growth rates(i.e., the percentage change in the second lag). Mining and quarrying, and oil and gas output,are official BPS sub-national accounts data. Mining captures economic activity extracting oil,natural gas, coal, and minerals, and preparing them for further processing. Quarrying refers tothe quarrying of chemical elements, mineral, and recess rock sediments just below the earth’ssurface, excluding metal, coal, petroleum, and natural gas. Manufacturing includes processingthese materials mechanically, chemically, or manually into finished or semi-finished products.District-level palm oil production in tons is taken from the Tree Crop Statistics of Indonesia forOil Palm produced annually by the Department of Agriculture.


Figu

re4.

1:C

ase

Stud

yD

istri

cts—

Loca

tion

s


Tabl

e4.

1:C

ase

Stud

yD

istri

cts—

Des

crip

tive

Stat

istic

s

Dis

tric

tIn

drag

iriH

ilir

Tapi

nM

anok

war

iN

at.a

vg.

Trea

tmen

tyea

r20

0820

0520

08N

/A

Com

mod

ityO

ilpa

lmC

oal

Nat

ural

gas

N/A

Pove

rty

rate

(%)

168

5118

Avg

.mon

thly

HH

expe

nditu

re(ID

R)21

2,93

641

5,68

525

8,74

326

0,58

9

Out

put(

mill

ion

IDR)

4,65

4,04

568

5,34

21,

215,

282

4,77

8,09

6

Out

putp

erca

pita

(mill

ion

IDR)

7.4

4.6

5.2

7.5

Agr

icul

ture

outp

utsh

are

(%)

5049

5232

Are

a(k

m)

12,6

152,

701

37,9

015,

636

Popu

latio

n63

0,86

315

0,00

023

5,60

863

0,12

0

Rura

lpop

ulat

ion

shar

e(%

)88

8595

67

Agr

icul

ture

empl

oym

ents

hare

(%)

3464

6545

Adu

ltlit

erac

yra

te(%

)97

9587

90

All

2005

exce

ptth

eem

ploy

men

tsha

re(2

008)

.A

vera

gein

clud

esal

l334

inth

eba

lanc

edpa

nelw

ith20

01di

stric

tbou

ndar

ies.

Ave

rage

mon

thly

hous

ehol

dex

pend

iture

isex

pres

sed

inno

min

alID

Rte

rms.

Out

puta

ndco

mpo

nent

sare

alli

nre

alte

rms.


Figure 4.2: Case Study Districts—Treatments

0.5

11.

5

Out

put p

er c

apita

(ID

R)

020

0000

4000

0060

0000

Pal

m o

il pr

oduc

tion

(ton

s)

2002 2005 2008 2011Year

Indragiri Hilir palm oil (LHS) Tapin mining (RHS)Manokwari oil & gas /10 (RHS)

Figure 4.2 presents each case study’s time series for resource sector output.

Oil and gas output per capita for Manokwari is by far the largest boom, so

I divide it by ten to present the three together. By identifying the resource

booms through a component of output prior to assessing impacts, I validate

a dose-response relationship between resource production and the economy.

This implies the resource “shock” is the key driver of any observed changes,

rather than some other factor coinciding with the boom (Mideksa, 2013). By

focusing on production, the main challenges is that the two possible states for the

treated district are pre-production and post-production. For the mining booms,

pre-production periods may be different to periods with no mining activity (e.g.,

due to construction and other activities, reflecting the mining life cycle). However

if a tight match is achieved on the pre-production period, the synthetic control

comparisons are valid: the counterfactual is just based on a synthetic control

mimicking a pre-production state rather than a purely untreated state. Relatedly


reverse causation can also threaten the validity of comparisons. Resource booms,

and in particular, private investment in the resource sector, could be triggered by

anticipation of future growth prospects. If such growth expectations are captured

by the unobservable heterogeneity included in the model—demonstrated by a

tight pre-treatment fit and the synthetic control and the actual outcome moving

together—this should not bias estimates. Using earlier treatment years—to

include pre-production activities in the estimated treatment effects and account

potential anticipation effects—gives similar results.66

66Although the synthetic control method handles unobservable heterogeneity, resource shocksarguably can be treated as exogenous in any case. The geological placement of naturalresource reserves is fundamentally random (Carmignani, 2013; Edwards, 2016; van der Ploegand Poelhekke, 2010) and the development of Indonesia’s natural resources tends to proceedregardless of local socioeconomic conditions (Resosudarmo, 2005). Likewise the process of oilpalm expansion is subject to a high degree of arbitrariness, similar to that described in Burgess etal. (2012).


4.2.4 Constructing each synthetic control

Identifying relevant predictors

Optimal weights used to construct synthetic controls (W ∗) depend critically

on V , so predictor variables in X0 and X1 must be considered carefully. Valid

synthetic controls for my three case study districts must be based on predictors

effectively (a) describing the three districts’ pre-treatment economic profiles, and

(b) predicting their post-treatment trajectories. I identify output per capita, the

adult literacy rate, population density, agricultural employment share, industry

employment share, average household expenditures, the poverty rate, the poverty

gap index, the agricultural share of output, manufacturing share of output, and

lagged outcome variables as relevant predictors.67 I ensure lagged outcome

variables are omitted for a few pre-treatment periods (to allow sufficient influence

of the other predictors) and follow Abadie and Gardeazabal (2003) to choose

V through the purely data-driven method described earlier.68 There are likely

other relevant predictors not included in my model, but tight pre-treatment fits

(i.e., the synthetic control tracking the movements of the actual outcome) and

strong predictor balance suggest my predictors capture most of the unobserved

heterogeneity in the pre-treatment periods.

Refining the pool of potential comparison districts

The comparison unit is supposed to approximate the counterfactual of the

unit of interest without any treatment, so selecting a comparison pool of districts

sufficiently similar to the districts of interest is critical to construct a valid synthetic

67Predictors are taken from the World Bank (2015), statistically relevant according to simplepooled ordinary least squares and panel fixed effects estimates, and identified from existingresearch on Indonesian sub-national development, e.g., Arndt (1984), Hill (1996), Hill et al (2012),Suryahadi et al (2003), Suryahadi et al (2009), Wetterberg et al (1999), Resosudarmo (2005), Miranti(2010), Manning (2010), Manning and Sumarto (2011), and De Silva and Sumarto (2014).

68Including all outcome lags as separate predictors is common to improve the pre-treatmentfit; see, e.g., Bohn et al (2012) and Billmeier and Nanncini (2012). I avoid this practice, as it canlead to over-fitting and poorer post-treatment performance, rendering other predictors irrelevantirrespective of how important they are in describing the pre-treatment (Abadie et al, 2010).


control. With 336 districts in 2001, my balanced panel provides rich cross-sectional

variation to draw an appropriate comparison pool. However using an excessive

number of observations tends to give unreasonably large weights to observations

similar to the treated unit in the pre-treatment period in some characteristics, but

wildly different in others. Abadie et al (2010; 2014) emphasise the importance

of limiting the comparison pool to units whose outcomes are driven by the

same structural process as the outcome of the unit of interest and not subject

to structural shocks to the outcome in the period of the study. Restricting the

comparison pool ensures appropriate comparison units are used, helps avoid

such interpolation biases, and minimises over-fitting the data to idiosyncratic

variations in a large sample of unaffected units.

Three empirical issues must be addressed in restricting the comparison

pool. Districts in the comparison pool must be: (a) unaffected by similar

treatments or any other large idiosyncratic shocks that could affect the outcome

of interest during the period of the study; (b) not affected by spill-overs from the

treatment on the treated unit (otherwise known as the stable unit treatment value

(SUTVA) assumption or no interference); and (c) sufficiently similar observable

characteristics and outcomes to the treated districts.

To ensure the districts in the comparison pool are not “treated” by similar

resource sector shocks, I restrict the comparison pool to districts with an

average total (not per capita) mining output of less than 10,000 million IDR

and palm oil production less than 1,000 tons over the whole period. I use the

full sample period for this restriction (c.f., pre-treatment period) because other

treated districts cannot be in the control group. My difference-in-difference-style

comparisons are therefore between districts with little to no resource production

expanding rapidly and staying expanded, and a synthetic district with little to no

resource production without any expansion. While this procedure raises concerns

about the comparability of resource-rich districts with non-resource-rich districts,


ensuring potential comparison units are untreated is the first-order issue.69

Oil palm can be processed in mills in neighbouring districts and companies

tend to be head-quartered in provincial capitals (Pekanbaru in this case) and

Jakarta. Indonesia’s unique resource revenue sharing arrangements often see

mining revenues distributed to other districts in the same province (see Agustina

et al (2012), Fadliya and McLeod (2010), and Lewis and Smoke (2015) for

detailed discussions of Indonesia’s fiscal arrangements). Both scenarios could bias

estimates. To minimise contamination from economic spillovers in potentially

affected neighbouring districts, I remove districts from the same province as the

treated district from each comparison pool.

Comparison districts sufficiently similar in observable characteristics and past

outcome behaviour are selected using the data-driven algorithm described earlier

(i.e., dissimilar units receive zero weights). Ideally I would also restrict the

comparison pool to districts in the same region (i.e., island group) to ensure

“common geographic support”, but this is only feasible for oil palm expansion in

Indragiri Hilir. The goal is to strike a balance between how similar the observable

characteristics of districts within the comparison pool are and the size of the

comparison pool. Tapin is a relatively small district and other non-resource-rich

districts on the island are not sufficiently comparable. Likewise Manokwari is a

national outlier in many respects and it is necessary to draw from a larger pool

to get a good match. Losing place-based comparability and running the risk

of interpolating across dissimilar regions, I gain the benefit of drawing from a

sufficiently large comparison pool with districts more comparable to Tapin and

Manokwari in observables and pre-treatment outcome levels and behaviour. My

three case study districts are rural but I retain cities (kotas) in the comparison pools

to better capture local business cycle conditions.

69The synthetic control method also deals with this concern by generating a comparison unitfor the untreated district with the natural resources but without the boom in natural resourceproduction. Natural gas and coal reserves were known for the whole study period, as was the factthat oil palm could plausibly grow in the areas where it was later planted.

§4.3 Results 141

4.3 Results

4.3.1 Oil palm expansion in Sumatra

Palm oil is the world’s most consumed vegetable oil and Indonesia its

largest producer. Over the past few decades the Indonesian landscape has been

transformed by oil palm expansion, with most of the increase in production

coming from land conversion (Gaskell, 2015).70 Indragiri Hilir is one such district

that has been recently transformed by oil palm. Indragiri Hilir’s palm oil mill

(Teluk Bakau POM, in Rotan Semelur village) was recently certified by the

Roundtable on Sustainable Palm Oil (RSPO) and the district area land used for

oil palm plantations almost tripled in from 2007–2008. Production then increased

sevenfold from 2007–2009 (see Figure 4.2).

Table 4.2: Impacts of Oil Palm Expansion in Sumatra

Treatment minus synthetic controlRMSPE N

Pre-T T T+1 T+2 T+3 T+4

Column 1 2 3 4 5 6 7 8

Total output -0.01 0.16 0.28 0.85 0.67 0.87 0.04 35

Agriculture 0.00 0.29 0.36 0.76 0.78 0.87 0.02 35

Industry 0.00 0.07 0.12 0.25 0.25 0.32 0.02 35

Services 0.00 0.06 0.10 0.31 0.25 0.35 0.02 35

Poverty (%) 0.01 1.18 -0.19 -1.55 -2.57 -2.07 0.00 100

HH exp. -21 2764 7713 -77372 6415.00 225

Poverty matched from 2002 up to the pre-treatment period 2007. All other variablesmatched from 2001–2007. Pre-T is the average difference between the observedoutcome and the synthetic control in the pre-treatment matching period, RMSPEis the root mean squared prediction error, and N is the number of districts in thecomparison pool. HH exp. refers to average monthly household expenditure percapita in nominal IDR.

70Corley and Tinker (2003) discuss the oil palm’s history and physiology, and Rival and Levang(2014) recent developments in Asia. Dennis et al (2005) and Koh and Wilcove (2007, 2008) discussenvironmental impacts. McCarthy et al (2010; 2011) and Rist et al (2010) discuss local socialimpacts.

§4.3 Results 142

Figure 4.3: Impacts of Oil Palm Expansion in Sumatra

(a) Aggregate Output

02

46

810

12

Per

cap

ita R

GD

P (

mill

ion

IDR

)

2001 2003 2005 2007 2009 2011Year

Indragiri Hilir synthetic control

(b) Agriculture

01

23

45

Per

cap

ita a

gric

ultu

re R

GD

P (

mill

ion

IDR

)2001 2003 2005 2007 2009 2011

Year


(c) Industry

0.5

11.

52

2.5

Per

cap

ita in

dust

ry R

GD

P (

mill

ion

IDR

)

2001 2003 2005 2007 2009 2011Year


(d) Services

01

23

4P

er c

apita

ser

vice

s R

GD

P (

mill

ion

IDR

)

2001 2003 2005 2007 2009 2011Year


(e) Household expenditure

010

0000

2000

0030

0000

4000

00

Avg

. mon

thly

hou

seho

ld e

xpen

ditu

res

(ID

R)

2001 2003 2005 2007 2009Year


(f) Poverty

02

46

810

1214

1618

Pov

erty

rat

e (%

)

2002 2004 2006 2008 2010 2012Year


§4.3 Results 143

Economic impacts

The local impacts of oil palm expansion in Indragiri Hilir are presented in

Table 4.2. The units for total output, its components, and average household

expenditure are per capita Indonesian rupiah. Poverty is in percentage points.

2008 is the treatment year, and the synthetic control is constructed using data from

2001–2007. Results are presented up to 2012 or as far as the data permit (2010

for household expenditure). The first column of Table 4.2 presents the average

difference between the observed Indragiri Hilir and its estimated synthetic control

over the pre-treatment period. Columns 2–6 show the difference between the

observed Indragiri Hilir and its estimated synthetic control in the treatment and

post-treatment years. The pre-treatment differences between actual Indragiri Hilir

and its synthetic control and the root mean squared prediction error (loss of fit)

are small relative to the outcome and its variation and grow in the post-treatment

period, as should be the case if valid synthetic controls have been constructed and

impacts increase over time. The final column presents the number of districts in

each restricted comparison pool.

Figure 4.3 graphically compares Indragiri Hilir’s observed outcomes for the

period 2001–2012 with the outcomes of counterfactual Indragiri Hilir. Panel A

shows output per capita increased steadily above its counterfactual, suggesting

the oil palm expansion was not offset by any local “Dutch disease” effects, at least

in aggregate terms.

Table 4.3 compares the average values of the predictor variables for Indragiri

Hilir and the estimated synthetic control for per capita output. Outcomes and

predictors for the synthetic control closely approximate Indragiri Hilir in the

pre-treatment period: coupled with the tight pre-treatment fit, this suggests the

synthetic control captures unobservable heterogeneity.

§4.3 Results 144

Table 4.3: Predictor Balance: Indragiri Hilir

Predictor Treated Synthetic

Output per capita (2008) 8.72 8.56



Average output per capita 7.15 7.15

Adult literacy rate 97.51 96.94

Agricultural employment share 4.076 3.26

Average Household expenditures 215, 516 261, 980

Poverty rate 16.77 17.31

Poverty gap index 3.27 3.05

Industry employment share 8.77 2.77

Agriculture output share 0.50 0.41

Manufacturing output share 0.16 0.15

RMSPE: 0.076

Table 4.3 compares the predictor variables for Indragiri Hilirand its synthetic control for the impact estimate on district percapita output in Panel A of Figure 4.3 and row one of Table4.2. Variables other than lagged outcomes are averaged overthe pre-intervention period 2001–2008. RMSPE refers to the rootmean squared prediction error.

§4.3 Results 145

Weights used to construct the synthetic control in Panel A of Figure 4.3 are

presented in Table 4.4. Dairi (North Sumatra) and Palembang, (South Sumatra)

account for most of synthetic Indragiri Hilir. The synthetic control comprising

just a few districts with large weights suggests the model is not overfitted (c.f.,

most of the sample getting very small weights suggests a potential overfitting

problem). None of the districts contributing to the comparison group in Table

4.4 have experienced a rapid increase in oil palm production and they span the

island of Sumatra, as intended. I do not present predictor balance and weights

for the other 17 synthetic controls in the chapter.

Table 4.4: Synthetic Indragiri Hilir Weights

Weight District name Province

0.5 Dairi North Sumatra

0.31 Palembang South Sumatra

0.06 Sabang Nanggroe Aceh Darussalam

0.06 Tanggamus Lampung

0.03 Banda Aceh Nanggroe Aceh Darussalam

0.03 Medan North Sumatra

0.01 Aceh Besar Nanggroe Aceh Darussalam

0.01 Lampung Timur Lampung

Table 4.4 presents the weights used to calculate the total per capitaoutput synthetic control for Indragiri Hilir, Riau, behind Panel Aof Figure 4.3 and the top row of Table 4.2. The comparison pool is36 districts in Sumatra but outside Riau province. 28 districts areassigned zero weights.

§4.3 Results 146

Panels B and C of Figure 4.3 show substantial impacts on per capita agriculture

and industry output, the two sectors capturing the growing and milling of palm

oil.71 Industry classifications in Indonesian sub-national accounts do not allow

me to disentangle whether any of the positive impact on industry output is due to

activities beyond processing palm oil at the mills. In Panel D I turn to services

output and find limited evidence of any major economic spillovers to formal

services, with only a small increase relative to the counterfactual. Given the size of

the informal services sector in rural Indonesia and its systematic underestimation

in official data, I cannot rule out further consumption spillovers in the informal

economy.

Welfare impacts

Panels E and F of Figure 4.3 present the broader welfare impacts of palm oil

expansion in Indragiri Hilir. Average per capita household expenditures and the

district poverty rate tell a different although not necessarily inconsistent story.

Average household expenditures track the synthetic control until the final year,

when it plateaus off below the synthetic control (Panel E). The pace of poverty

reduction picks up with around a four percentage point reduction relative to the

counterfactual four years into the oil palm expansion (Panel F).72 But why do

average nominal household expenditure and the poverty rate both decline, rather

than move in opposite directions? The poverty rate falling without an increase

in the average implies changes to the distribution of consumption (e.g., holding

the distribution constant and substantially increasing the average consumption,

the poverty rate must fall). First, household expenditure and the poverty rate

are derived from the same source so data issues are unlikely to explain the

71The synthetic control for agriculture also diverges from its pre-treatment trend wheneach district in the comparison pool is sequentially dropped and when a wider comparisonpool is used (c.f., Sumatra). This highlights the risk of assuming linear or parallel trends indifference-in-difference studies, as pre-trends can indeed change during the treatment period.

72I draw from a larger nationwide comparison pool to attain a closer pre-treatment fit for thesetwo estimates (c.f., restricting to Sumatra).

§4.3 Results 147

divergence.73 Second, growth in the palm oil sector could just be pro-poor

and redistributive, as rural poor mostly work in agriculture and agricultural

sector growth has historically been pro-poor (Suryahadi et al, 2009; de Silva

and Sumarto, 2014). Third, the cumulative distribution function of household

expenditure can also be quite steep around the poverty line, so a small change

in the distribution can correspond to a large change in the poverty rate. This is

indeed what is observed in the individual-level data. Figure 4.4 plots cumulative

distribution function of household expenditure before and during the palm oil

boom, and Figure 4.5 plots the two corresponding kernel density estimates.74

The curves both shift slightly to the right and flatten out a little.

Figure 4.4: Consumption in Indragiri Hilir

Poverty line

0

.2

.4

.6

.8

1

Cum

ulat

ive

Prob

abilit

y

0 1000000 2000000 3000000Per capita household monthly expenditure (constant 2004 IDR)

2004 2010

73Counting the wealthiest Indonesians in SUSENAS has been a persistent problem and leadsto substantial measurement error when studying consumption inequality, but this is a challengein most years and regions so unlikely to drive my results.

74Figures 4.4 and 4.5 do not present any counterfactual, only the pre-post comparison. Percapital household monthly expenditure is adjusted into constant 2004 terms using the nationalGDP deflator.

§4.3 Results 148

Figure 4.5: Kernel Density Estimate–Consumption Distribution in IndragiriHilir

0 1000000 2000000 3000000Per capita household monthy expenditure (constant 2004 IDR)

2004 2010

Results presented in Figure 4.3, Table 4.2, and results to follow are similar

across a wide variety of sensitivity analyses and falsification exercises, which

I omit for brevity due to the 6 × 3 outcomes analysed in this Chapter. These

robustness checks are summarised in the Chapter Appendix. Some smaller

treatment effects are statistically insignificant at conventional levels (i.e., one and

five per cent) when ranked against the full distribution of treatment effects. This

is as expected given that my analysis utilises such small donor groups.

§4.3 Results 149

4.3.2 Coal mining in Kalimantan

My second case study is of coal mining. Flying over the giant island of

Kalimantan (Borneo) you cannot help but notice the landscape is littered with

large open-cut mines. Many of these mines have been active for decades or are in

districts with contemporaneous oil, gas, and palm oil booms, precluding them

from this study. However in the Tapin district of South Kalimantan—home to

one of the region’s largest coal terminals—mining and quarrying output more

than tripled in 2005 due to a large expansion in coal mining operations (Figure

4.2). Like many districts in Kalimantan, there has also been an oil palm expansion

since 2009, so I focus on the impacts of the mining boom in the years leading up

to oil palm production in 2009.

Table 4.5: Impacts of Coal Mining in Kalimantan


Pre-T T T+1 T+2 T+3 T+4

Column 1 2 3 4 5 6 7 8

Total output 0.01 0.01 0.72 0.76 0.74 0.73 0.00 93

Agriculture 0.00 -0.06 -0.05 -0.07 -0.18 -0.03 0.02 93

Industry 0.00 0.00 0.03 0.04 0.09 0.11 0.00 93

Services 0.00 0.01 -0.04 -0.01 -0.05 -0.06 0.00 93

Poverty -0.09 -0.83 0.67 -0.26 -2.62 -3.10 0.00 105

HH exp. 386 289 93267 178072 -73769 28291 0 98

Poverty matched from 2002 up to the pre-treatment period 2004. All other variablesmatched from 2001–2004. Pre-T is the average difference between the observed outcomeand the synthetic control in the pre-treatment matching period, RMSPE is the root meansquared prediction error, and N is the number of districts in the comparison pool. HHexp. refers to average monthly household expenditure per capita in nominal IDR.

§4.3 Results 150

Figure 4.6: Impacts of Coal Mining in Kalimantan


02

46

Per

cap

ita R

GD

P (

mill

ion

IDR

)

2001 2003 2005 2007 2009Year

Tapin synthetic control

(b) Agriculture

0.5

11.

52

2.5

Per

cap

ita a

gric

ultu

re R

GD

P (

mill

ion

IDR

)2001 2003 2005 2007 2009

Year


(c) Industry

0.2

.4.6

Per

cap

ita in

dust

ry R

GD

P (

mill

ion

IDR

)

2001 2003 2005 2007 2009Year


(d) Services

0.5

11.

52

Per

cap

ita s

ervi

ces

RG

DP

(m

illio

n ID

R)

2001 2003 2005 2007 2009Year


(e) Household Expenditure

020

0000

4000

0060

0000

Avg

. mon

thly

hou

seho

ld e

xpen

ditu

re (

IDR

)

2001 2003 2005 2007 2009Year


(f) Poverty

02

46

810

Pov

erty

rat

e (%

)

2002 2004 2006 2008Year


§4.3 Results 151

Economic impacts

Results for Tapin are presented in Table 4.5 and Figure 4.6. A good

pre-treatment fit is achieved throughout, even though there is just four years lead

to fit the synthetic control before the treatment year 2005. Aggregate and industry

output increase with the rapid scale-up of mining production, as expected. Just

like palm oil milling in the previous case study, an increase in industry output

does not necessarily provide evidence of production linkages. Washing and

transporting the rocks at Tapin’s large coal facility is captured in industry output.

Agricultural and services output better identify spillovers. Both sectors declined

below the synthetic control, consistent with localised Dutch disease-type impacts

on non-tradable sectors (Corden, 1984; Matsuyama, 1992). Negative impacts on

local agriculture from coal mining are also consistent with Aragon and Rud’s

(2015) study of gold mines in Ghana, and Fleming and Measham’s (2014) study

finding no significant manufacturing or agricultural spillovers from Australian

mining. Agriculture recovers with Tapin’s first oil palm yield in 2009, but the

small negative impact on services appears more permanent.

Welfare impacts

Panels E and F of Figure 4.6 present effects on poverty and average household

expenditures. Panel E of Figure 4.6 shows how average household expenditure

moved above then below synthetic Tapin during the mining boom, suggesting

limited sustained impacts on the welfare of the average Tapin resident. Tapin’s

poverty rate (Panel F) diverges away from its synthetic control in the treatment

period (2005), spikes in 2006, then diverges off far below the counterfactual until

2009, halving from ten to five percent. A positive poverty impact is different to

what has been reported on average across Indonesia, as mining growth typically

does not correspond to poverty reduction (Resosudarmo and Bhattacharya, 2015)

and mining-dependent districts tend to have higher poverty rates (Chapter 2). In

§4.3 Results 152

a related study, Salami et al (2014) highlight the health impacts of mining in South

Kalimantan, finding children have decreased pulmonary function due to greater

intake of harmful particulates. Studies like this complement mine, highlighting

the health/wealth trade-off often faced by mining communities in developing

countries.

4.3.3 Natural Gas Extraction in West Papua

My final case study examines one of the world’s archetypical resource

enclaves: West Papua. I study the British Petroleum (BP) Tangguh project,

Indonesia’s first fully vertically integrated (i.e., from raw material to final product)

liquified natural gas (LNG) operation. Two unmanned offshore production

platforms tap the giant Tangguh LNG gas fields in Bintuni Bay (the “Bird’s head”

in Figure 4.1), pump the gas through subsea pipelines to the LNG processing

facility located in the village of Tanah Merah, then deliver LNG to markets across

Asia and the United States. The gas fields contain over 500 billion cubic meters

of proven natural gas reserves (with an extra 300 billion cubic meters estimated)

and are being developed by an international consortium led by BP. The facility has

operated at full capacity (seven million tons per year) since it became operational

in 2009 and is expected to expand with the addition of a third LNG train before

2020.

West Papua provides a unique case study in many respects. It is one of the least

populated and poorest parts of Indonesia. Papua and Irian Jaya were retained

by the Dutch following independence in 1945, seized by Indonesia in 1961, and

today remain one of Indonesia’s last areas of active separatism and civil unrest. In

2000, West Papua province contained three districts, but like many resource-rich

areas of Indonesia, the three districts Balkanised into twelve today. Using 2001

district boundaries (the district named Manokwari in 2001), my estimates relate

to a generous geographic area covering several modern-day districts. Despite the

§4.3 Results 153

larger area, the relative economic importance of this economic shock to the region

dwarves my other two case studies. Manokwari extracted no oil or natural gas

prior to the Tangguh project (see Figure 4.2), and the natural resource riches in the

Bintuni Bay could potentially transform one of Indonesia’s most developmentally

challenged regions.

Production at the Tangguh facility officially began in 2009. I use 2008 as the

treatment year to capture related economic activity prior to the facility becoming

operational (e.g., advance company payments to communities, building the

necessary infrastructure around the facility) and because the longer lead time

allows. The synthetic control is fitted from from 2001–2007.75

Table 4.6: Impacts of Gas Extraction in West Papua


Pre-T T T+1 T+2 T+3 T+4

Column 1 2 3 4 5 6 7 8

Total output 0.02 0.02 1.67 8.34 14.91 19.26 0.15 96

Agriculture 0.00 0.14 0.13 0.00 -0.05 -0.16 0.06 96

Industry -0.01 0.14 1.58 7.72 14.14 17.07 0.03 96

Services 0.00 -0.01 0.01 0.14 0.12 0.04 0.02 96

Poverty 0.47 -0.02 -0.42 -3.85 -0.08 -3.38 2.13 108

HH exp. 533 14276 53997 10235 10190 101

Poverty matched from 2002 up to the pre-treatment period 2007. All other variablesmatched from 2001–2007. Pre-T is the average difference between the observedoutcome and the synthetic control in the pre-treatment matching period, RMSPE is theroot mean squared prediction error, and N is the number of districts in the comparisonpool. HH exp. refers to average monthly household expenditure per capita in nominalIDR.

75The land was acquired in 1999, with many resettlement agreements. The village relocationwas completed in 2004, so unlikely to affect my synthetic control modelling.

§4.3 Results 154

Figure 4.7: Impacts of Gas Extraction in West Papua


05

1015

2025

Per

cap

ita R

GD

P (

mill

ion

IDR

)

2001 2003 2005 2007 2009 2011Year

Manokwari synthetic control

(b) Agriculture

0.5

11.

52

2.5

33.

5

Per

cap

ita a

gric

ultu

re R

GD

P (

mill

ion

IDR

)2001 2003 2005 2007 2009 2011

Year


(c) Industry

05

1015

20

Per

cap

ita in

dust

ry R

GD

P (

mill

ion

IDR

)

2001 2003 2005 2007 2009 2011Year


(d) Services

0.5

11.

52

2.5

3

Per

cap

ita s

ervi

ces

RG

DP

(m

illio

n ID

R)

2001 2003 2005 2007 2009 2011Year



010

0000

2000

0030

0000

4000

0050

0000

Avg

. mon

thly

hou

seho

ld e

xpen

ditu

res

(ID

R)

2001 2003 2005 2007 2009Year


(f) Poverty

010

2030

4050

Pov

erty

rat

e (%

)

2002 2004 2006 2008 2010 2012Year


§4.3 Results 155

Economic impacts

The local economic and welfare impacts of the Tangguh project are presented

in Table 4.6 and Figure 4.7. Panels A and C of Figure 4.7 show dramatic boosts

to total and industry output, expected given the facility’s vertically-integrated

nature. Output in the agricultural sector—the largest sector before the mining

boom and still accounting for most Papuans’ livelihoods (Resosudarmo et al,

2013)—paints a less positive picture of the Tangguh project (Panel B). Agricultural

output breaks from its long-term trend as the large gas project became operational.

Panel D shows a small increase in services sector output relative to the

counterfactual, although some services may be part of the project (e.g., hotel

accommodation for expatriates).

Welfare impacts

Manokwari’s poverty rate has steadily declined from 2002–2012, falling by an

impressive 20 percentage points (Panel F of Figure 4.7). But there is no sharp

discontinuity or change in trajectory as the LNG facility became fully operational

in 2009. Similarly Panel E of Figure 4.7 shows how the commencement of the

facility’s operations corresponded to an immediate boost in average household

expenditures, but convergence with the synthetic control shortly after. The

quadrupling of per capita output has not been met by a commensurate increase

in welfare or poverty reduction: any impacts are slight in context of the massive

scale of the output ramp up. Unlike for coal mining in Tapin, this result is

consistent with the higher poverty rates that are observed on average across

mining-dependent districts in Chapter 2.


4.4 Conclusion

The objective of this chapter was to quantify the local economic and welfare

impacts of resource booms in three Indonesian districts. I studied Indonesia’s

three largest primary exports and used the synthetic control method to construct

counterfactual paths for districts’ per capita economic output and its components,

as well as average household expenditure and poverty. Each resource boom

substantially altered the structure of the local economy. Palm oil expansion in

Indragiri Hilir delivered a small boost to all sectors of the economy and strong

poverty reduction. Coal mining in Tapin reduced agricultural and services sector

output, but also delivered strong poverty reduction relative to the estimated

counterfactual. The Tangguh natural gas project in West Papua delivered a

massive increase in local economic output, but no major impacts on average

household welfare and poverty and a contraction in the agricultural sector.

According to the Indonesian constitution, Indonesia’s natural resources are

to “be controlled and utilised by the State for the maximum prosperity of the

people” (Gandataruna and Haymon, 2011). My three case studies highlight how

natural resource sectors can make important contributions to increasing district

GDP per capita and reducing poverty in Indonesia. Relative to their size, sectors

with more concentrated rents (e.g., natural gas) appear to provide fewer benefits to

local residents than more diffuse, labour-intensive sectors. For regions depending

on resource sectors with highly concentrated rents, natural resource sector-led

economic growth alone should not be expected to improve the average residents’

welfare and reduce poverty. More active fiscal and social policies may be needed.



The 18 synthetic control estimates presented in the chapter were subject to the

robustness checks / sensitivity analyses summarised below. Results are omitted

for brevity, but similar and readily produced from the public data or available on

request.

Sensitivity analyses

• Leave-one out analysis. I estimate the synthetic control iteratively leaving

out each of the three comparison pool districts assigned the largest weights

to gauge sensitivity to a particular unit being included in the comparison.

Like permutation testing, a distribution of leave-one-out treatment effects is

evaluated.

• Population effects. I checked that changes in population are not driving the

results on per capita economic output and that similar results are obtained

using total and log levels (i.e., not in per capita terms).

• Alternative comparison pools:

– Regional (island) group vs. full sample; and

– Dropping districts where the average outcome of interest is fifty

per cent larger or smaller than that for the treatment district in the

pre-treatment period, as in Horiuchi and Mayerson (2015).

• Varying the predictor set:

– Parsimonious predictors, as used in my main estimates;

– All lagged outcomes only; and

– A more extensive set of predictors taken from DAPOER (World Bank,

2015), allowing the data-driven synthetic control procedure to instead

assess performance and relevance.


• Using alternative techniques to estimate the V matrix.

Placebo / falsification tests

• In-time placebo tests involve treating years without the oil palm expansion

as the treatment years. See Bertrand et al (2004).

• In-space placebo tests or permutation inference, where the treatment is

iteratively applied to every unit in the comparison pool to compare the

estimated treatment effect on the unit of interest against the distribution

of treatment effects obtained for untreated units in the comparison pool.

This leads to what Abadie et al (2010) call “exact inference”, akin to random

permutation testing. Although this test is becoming standard practice, its

interpretation is ambiguous when large restrictions are placed on the donor

pool (as I have done), so I have omitted it from the chapter. This exercise also

showed that prediction errors for the untreated units are roughly similar in

the pre-and post-treatment periods, providing no evidence of over-fitting.

• On estimated synthetic controls and on alternative outcomes plausibly

unaffected by the treatment, i.e., fiscal transfers unrelated to natural

resources, precipitation and temperature, and lagged outcomes.

Replications

Replications using similar cases identified through my case study selection

procedure—the runner-ups—gave broadly similar results. Replication cases

include: palm oil in Kalimantan and Sumatra; oil and gas in East Java; mining

in Sulawesi and Kalimantan; and West Papua estimated at the province level (c.f.,

district).


Main Results Graphs with Non-zero Y Axes


Figure 4.8: Appendix–Impacts of Oil Palm Expansion in Sumatra, non-zero Y axis


67

89

1011

Per

cap

ita R

GD

P (

mill

ion

IDR

)

2001 2003 2005 2007 2009 2011Year


(b) Agriculture

33.

54

4.5

5

Per

cap

ita a

gric

ultu

re R

GD

P (

mill

ion

IDR

)2001 2003 2005 2007 2009 2011

Year


(c) Industry

11.

52

2.5

Per

cap

ita in

dust

ry R

GD

P (

mill

ion

IDR

)

2001 2003 2005 2007 2009 2011Year


(d) Services

1.5

22.

53

3.5

4P

er c

apita

ser

vice

s R

GD

P (

mill

ion

IDR

)

2001 2003 2005 2007 2009 2011Year


(e) Household expenditure

1500

0020

0000

2500

0030

0000

3500

0040

0000

Avg

. mon

thly

hou

seho

ld e

xpen

ditu

res

(ID

R)

2001 2003 2005 2007 2009Year


(f) Poverty

810

1214

1618

Pov

erty

rat

e (%

)

2002 2004 2006 2008 2010 2012Year



Figure 4.9: Appendix–Impacts of Coal Mining in Kalimantan, non-zero Y axis


44.

55

5.5

6

Per

cap

ita R

GD

P (

mill

ion

IDR

)

2001 2003 2005 2007 2009Year


(b) Agriculture

2.1

2.2

2.3

2.4

2.5

2.6

Per

cap

ita a

gric

ultu

re R

GD

P (

mill

ion

IDR

)2001 2003 2005 2007 2009

year


(c) Industry

.4.4

5.5

.55

.6.6

5

Per

cap

ita in

dust

ry R

GD

P (

mill

ion

IDR

)

2001 2003 2005 2007 2009Year


(d) Services

1.4

1.5

1.6

1.7

1.8

1.9

Per

cap

ita s

ervi

ces

RG

DP

(m

illio

n ID

R)

2001 2003 2005 2007 2009Year



1000

0020

0000

3000

0040

0000

5000

0060

0000

Avg

. mon

thly

hou

seho

ld e

xpen

ditu

re (

IDR

)

2001 2003 2005 2007 2009Year


(f) Poverty

56

78

910

Pov

erty

rat

e (%

)

2002 2004 2006 2008Year



Figure 4.10: Appendix–Impacts of Gas Extraction in West Papua, non-zero Y axis


510

1520

25P

er c

apita

RG

DP

(m

illio

n ID

R)

2001 2003 2005 2007 2009 2011Year


(b) Agriculture

2.6

2.8

33.

23.

43.

6

Per

cap

ita a

gric

ultu

re R

GD

P (

mill

ion

IDR

)2001 2003 2005 2007 2009 2011

Year


(c) Industry

05

1015

20

Per

cap

ita in

dust

ry R

GD

P (

mill

ion

IDR

)

2001 2003 2005 2007 2009 2011Year


(d) Services

11.

52

2.5

3

Per

cap

ita s

ervi

ces

RG

DP

(m

illio

n ID

R)

2001 2003 2005 2007 2009 2011Year



1000

0020

0000

3000

0040

0000

5000

00

Avg

. mon

thly

hou

seho

ld e

xpen

ditu

res

(ID

R)

2001 2003 2005 2007 2009Year


(f) Poverty

3035

4045

5055

Pov

erty

rat

e (%

)

2002 2004 2006 2008 2010 2012Year


Chapter 5

Conclusion

This thesis has implications for policy and suggests several promising avenues

for future research. Each chapter shows that natural resource sectors have

important effects on other sectors of the economy. These effects differ for

different types of natural resource sectors. At the international level, more

mining income in the economy leads to lower levels of non-mining income in

the long run (Chapter 2). Within Indonesia, oil palm expansion does not appear

to crowd out other economic activity at the district level (Chapters 3). In a

case study of a booming oil palm district in Riau (Chapter 4), I find small but

positive impacts on all sectors, consistent with the pattern observed nationally

in Chapter 3. For Indonesia’s mining sectors, two case studies suggest that coal

mining and natural gas extraction have increased total economic output and

output in sectors directly involved in resource extraction and processing, but

substantially reduced agricultural output (Chapter 4). These findings highlight

how Indonesia’s resource boom has led to major local structural change across its

diverse regions. Growth in the oil palm sector exhibits minor positive effects on

other sectors and total economic output. Mining sector growth, on the other hand,

crowds-out non-tradable agriculture, suggesting a local Dutch disease. Future

research could investigate possible links between natural resource sectors and

urbanisation, employment and labour market outcomes, and other aspects of

structural change in Indonesia or elsewhere.

That natural resource sectors alter the structure of national and sub-national

economies has implications for social development trajectories. Chapter 2 showed

163

164

that non-mining income is on average better for health and education than income

from the mining sector. Mining-sector growth is likely to undermine national and

sub-national economic growth and prosperity due to the mining sector’s weaker

links to human capital development and lower social productivity than alternative

sectors. Countries and sub-national regions committed to mining sector-led

economic growth may wish to consider strategies to diversify the economy

towards more human capital-intensive modern sectors or better utilise mining

rents for broad social outcomes. Further investigation of the mechanisms and the

conditions under which mining can be more beneficial for health and education

outcomes could consider returns to education and skills under mining booms.

Examining the long-term health and education impacts of more labour-intensive

natural resource sectors like palm oil could also be worthwhile.

The first Sustainable Development Goal is to “end poverty in all its forms

everywhere”. In this thesis I find that natural resource sectors can make important

contributions to poverty reduction, but these contributions depend on the type of

natural resource activity and the concentration of resource rents.

Growth in Indonesia’s palm oil sector has, on average, made strong

contributions to poverty reduction over the period 2001–2010 (Chapter 3). This

finding may influence debates on oil palm and development. In 2011 former

World Bank President Robert Zoellick directed the World Bank Group to cease any

financing and operations related to the palm oil industry due to environmental

issues. My thesis encourages policy-makers to instead seek a middle ground

balancing environmental conservation with the pro-poor growth opportunities

presented by oil palms. As mentioned Chapter 3, there are ways to further

increase palm oil production without clearing pristine primary rainforest. Future

research could investigate how different palm oil sector business models affect the

diffusion of benefits through local communities, compare plantations and mills

that have attained sustainability certification with those that have not, and extend

165

my analysis to other countries with growing palm oil sectors (e.g., Ghana and

Cameroon).

In contrast to the poverty impacts of palm oil expansion, districts more

dependent on the mining sector tend to have significantly higher poverty rates

(Chapter 2). Looking beyond these averages, I find that a large coal mining

expansion in South Kalimantan reduced poverty while a giant natural gas project

in West Papua had no discernible effect (Chapter 4).

The local economic benefits of natural resource sectors thus appear to be

spread in accordance with the concentration of the rents. For natural resource

sectors with highly concentrated resource rents, expansions in output alone

should not be expected to reduce poverty. Policies to ensure broad-based benefits

might be needed. Further research into how such policies can be designed

and implemented in the developing country context may be beneficial. Further

research could also examine the impacts of natural resource sectors on poverty

using international data or alternative poverty measures (e.g., electrification,

nutrition and food security, or asset poverty).

There is still much to learn about how natural resource sectors affect human

development outcomes across and within countries. Theory and empirical

evidence on how natural resource sectors affect poverty and inequality, labour

markets and human capital, and industrialisation remains thin. I hope the

research presented in this thesis prompts other researchers to investigate these

issues further, develop new theories to explain the effects documented in this

thesis, and extend my analysis using rich new spatial and micro datasets.

Integrated assessments factoring in environmental and climate impacts beyond

the economic and social impacts documented in this thesis will also be important.

References

[1] Abadie, A. and Gardeazabal, J. (2003), “The economic costs of conflict: A case

study of the Basque Country.” American Economic Review, 93:1, pp. 113–132.

[2] Abadie, A., Diamond, A. and Hainmueller, J. (2010), “Synthetic control

methods for comparative case studies: Estimating the effect of California’s

Tobacco Control program.” Journal of the American Statistical Association,

105:490, pp. 493–505.

[3] Abadie, A., Diamond, A. and Hainmueller, J. (2014), “Comparative Politics and

the Synthetic Control Method.” American Journal of Political Science, 29:2, pp.

495–510.

[4] Abadie, A., Diamond, A. and Hainmueller, J. (2011), “Synth: An R Package for

Synthetic Control Methods in Comparative Case Studies.” Journal of Statistical

Software 42:13, pp. 1–17.

[5] Acemoglu, D., Johnson, S., and Robinson, J. (2001), “The Colonial Origins of

Comparative Development: An Empirical Investigation”, 91:5, pp. 1369–1401.

[6] Acemoglu, D., Finkelstein, A., and Notowidigdo, M. (2013), “Income and

health spending: evidence from oil price shocks”, Review of Economics and

Statistics, 95:4, pp. 1079–1095.

[7] Adelman, I. (1984), “Beyond export-led growth”, World Development, 12, pp.

937–949.

[8] Agustina, C.D., Fengler, W., and Schulze, G.G. (2012), “The regional effects

of Indonesia’s oil and gas policy: options for reform”, Bulletin of Indonesian

Economic Studies, 48:3, pp. 367–395.

166

167

[9] Alexeev, M. and Conrad, R. (2009,) “The elusive curse of oil”, Review of

Economics and Statistics, 91:3, pp. 586–598.

[10] Altonji, J., Conley, T., Elder, T., and Taber, C. (2005), “Selection on Observed and

Unobserved Variables: Assessing the Effectiveness of Catholic Schools,” Journal

of Political Economy, 113:1, pp. 151–184.

[11] Alwarritzi, W., Nanseki, T., and Chomei, Y. (2015), “Analysis of the factors

influencing the technical efficiency among oil palm smallholder farmers in

Indonesia”, Procedia Environmental Sciences, 28, pp. 630–638.

[12] Anderson, T.W. and Rubin, H. (1949), “Estimators of the Parameters of a Single

Equation in a Complete Set of Stochastic Equations”, Annals of Mathematical

Statistics, 21, pp. 570–582.

[13] Ando, M. (2015), “Dreams of urbanization: Quantitative case studies on the

local impacts of nuclear power facilities using the synthetic control method”,

Journal of Urban Economics, 85, pp. 68–85.

[14] Andrews, D., Moreira, M., and Stock, J. (2006), “Optional Two-Sided Invariant

Similar Tests for Instrumental Variables Regression”, Econometrica, 74:3, pp.

715–752.

[15] Andrews, D., Moreira, M., and Stock, J. (2004), “Performance of Conditional

Wald Tests in IV Regression with Weak Instruments”, Journal of Econometrics,

139, pp. 116–132.

[16] Anriquez, G., and Lopez, R. (2007), “Agricultural growth and poverty in an

archetypical middle income country: Chile”, Agricultural Economics, 36:2, pp.

1910–202

[17] Aragon, F.M. and Rud, J.P. (2013), “Natural resources and local communities:

evidence from a Peruvian gold mine”, American Economic Review: Economic

Policy 5: 2, pp. 1–25.

168

[18] Aragon, F.M. and Rud, J.P. (2015), “Polluting Industries and Agricultural

Productivity: Evidence from Mining in Ghana”, Economic Journal, forthcoming.

[19] Arezki, R., and Gylfason, T. (2013), “Resource Rents, Democracy, Corruption,

and Conflict: Evidence from Sub-Saharan Africa”, Journal of African Economies,

22: 4, pp. 552–569.

[20] Arezki, R., Gylfason, T., and Sy, A. (2011), “Beyond the Resource Curse:

Policies to Harness the Power of Natural Resources”, International Monetary

Fund, Washington, D.C.

[21] Arndt, H.W. (1984), “The Indonesian Economy: Collected Papers”, Chopman

Publishers, Singapore.

[22] Austin, K., Kasibhatla, P., Urban, D., Stolle, F., and Vincent, J. (2015),

“Reconciling Oil Palm Expansion and Climate Change Mitigation in Kalimantan,

Indonesia”, PLoS ONE 10:5, e0127963.

[23] Auty, R.M. (1997), “Natural Resource Endowment: The State and Development

Strategy”, Journal of International Development, 9:4, pp. 651–663.

[24] Auty, R.M and Mikesell, R. (1998), “Sustainable Development in Mineral

Economies”, Oxford University Press, Oxford.

[25] Auty, R.M. (2001), “Resource Abundance and Economic Development”,

Oxford University Press, Oxford.

[26] Auty, R.M. (2001), “The political economy of resource-driven growth”,

European Economic Review, 45, pp. 839–846.

[27] Ayres, R., and Warr, B. (2005), “Accounting for growth: the role of physical

work”, Structural Change and Economic Dynamics, 16, pp. 181–209.

[28] Ayres, R., and Warr, B. (2009), “The Economic Growth Engine: How Energy

and Work Drive Material Prosperity”, Edward Elgar, Cheltenham.

169

[29] Baltagi, B. H. and Griffin, J. M. (1984), “Short and Long Run Effects in Pooled

Models”, International Economic Review, 25:3, pp. 631–645.

[30] Barbier, E., (1989), “Cash Crops, Food Crops, and Sustainability: The Case of

Indonesia”, World Development, 19:6, pp. 879-895.

[31] Barone, G., and Mocetti, S. (2014), “Natural disasters, growth, and institutions:

a tale of two earthquakes”, 84, pp. 52–66.

[32] Barr, C.M., and Sayer, J.A. (2012), “The political economy of reforestation

and forest restoration in Asia-Pacific: critical issues for REDD+”, Biological

Conservation, 154, pp. 9–19.

[33] Barro, R.J. (2015), “Convergence and Modernisation”, Economic Journal, 125,

pp. 911–942

[34] Barro, R. J. and Lee, J. W. (1993), “International comparisons of educational

attainment”, Journal of Monetary Economics, 32, pp. 363–394.

[35] Barro, R. J. and Lee, J. W (2010), “A new data set of education attainment in the

world, 1950–2010”, Journal of Development Economics, 104, pp. 184–198.

[36] Bauhoff, S. (2014), “The effect of school nutrition policies on dietary intake and

overweight: a synthetic control approach”, Economics and Human Biology, 12,

pp. 44–55.

[37] Bazzi, S., Gaduh, A., Rothenberg, A., and Wong, M. (2016), “Skill

Transferability, Migration, and Development: Evidence from Population

Resettlement in Indonesia”, American Economic Review, forthcoming.

[38] Bazzi, S. and Clemens, M.A. (2013), “Blunt instruments: Avoiding common

pitfalls in identifying the cause of economic growth”, American Economic

Journal: Macroeconomics, 5:2, pp. 152–86.

170

[39] Bellows, J., and Miguel, E. (2009), “War and local collective action in Sierra

Leone,” Journal of Public Economics, 93:11-12, pp. 1144–1157.

[40] Bertrand, M., Duflo, E., Mullainathan, S. (2004) “How much should we trust

difference-in-differences estimates?” Quarterly Journal of Economics, 119:1, pp.

249-275.

[41] Besley, T., and Kudamatsu, M. (2006), “Health and Democracy”, American

Economic Review, 96:2, pp. 313–318.

[42] Bhattacharyya, S. and Resosudarmo, B.P. (2015), “Growth. Growth

Accelerations, and the Poor: Lessons from Indonesia”, World Development, 66,

154–165.

[43] Bhattacharyya, S. and Williamson, J. (2013), “Distributional Impact of

Commodity Price Shocks: Australia over a Century”, CSAE Working Paper

Series, 2013–11, Centre for the Study of African Economies, University of Oxford.

[44] Billmeier, A. and Nannicici, T. (2013), “Assessing Economic Liberalization

Episodes: A Synthetic Control Approach.” Review of Economics and Statistics

95:3, pp. 983–1001.

[45] Black, D., McKinnish, T., and Sanders, S. (2005), “The economic impact of the

coal boom and bust”, Economic Journal, 115: 503, pp. 449–476.

[46] Blanco, L. and Grier, R. “Natural resource dependence and the accumulation

of physical and human capital in Latin America”, Resources Policy, 37:3, pp.

281–295.

[47] Blattman, C., and Bazzi, S. (2014), “Economic Shocks and Conflict: Evidence

from Commodity Prices”, American Economic Journal: Macroeconomics, 6:4,

pp. 1–38.

171

[48] Booth, A. (1988), “Agricultural Development in Indonesia”, Allen and Unwin,

Sydney.

[49] Bohn, S., Lofstrom, M., Raphael, S. (2014), “Did the 2007 Legal Arizona

Workers Act reduce the state’s unauthorized immigrant population?” Review

of Economics and Statistics, 96:2, pp. 258–269.

[50] Boschini, A., Pettersson, J., and Roine, J. (2013) “The Resource Curse and its

Potential Reversal”, World Development, 43, pp. 19–41.

[51] Bound, H., Jaeger, D., and Baker, R. (1995), “Problems with Instrumental

Variables Estimation when the Correlation Between the Instruments and the

Endogenous Explanatory Variables is Weak”, Journal of the American Statistical

Association, 90:430, pp.443–450.

[52] Brunnschweiler, C. and Bulte, E. H. (2008), “The resource curse revisited

and revised: A tale of paradoxes and red herrings”, Journal of Environmental

Economics and Management, 55:3, pp. 248–264.

[53] Brunnschweiler, C. and Bulte, E. H. (2009), “Natural resources and violent

conflict: Resource abundance, dependence, and the onset of civil wars”, Oxford

Economic Papers, 61:4, pp.651–674.

[54] Budidarsono, S., Dewi, S., Sofiyuddin, M., Rahmanulloh, A. (2012),

“Socioeconomic Impact Assessment of Palm Oil Production”, Technical Brief No.

27: palm oil series. Bogor, Indonesia. World Agroforestry Center–ICRAF, SEA

Regional Office.

[55] Budidarsono, S., Rahmanulloh, A., and Sofiyuddin, M. (2012), “Economic

assessment of Palm Oil Production”, Technical Brief No. 26: palm oil series.

Bogor, Indonesia. World Agroforestry Center–ICRAF, SEA Regional Office.

172

[56] Burgess, R., Hansen, M., Olken, B. A., Potapov, P., and Sieber, S. (2012),

“The Political Economy of Deforestation in the Tropics”, Quarterly Journal of

Economics, 122:1, pp. 73–117.

[57] Bulte, E., Damania, R., and Deacon, R. (2005), “Resource Intensity, Institutions,

and Development”, World Development, 33:7, pp. 1029–1044.

[58] Burke, P. and Nishitateno, S. (2015), “Gasoline prices and road fatalities:

international evidence”, Economic Inquiry, 53:3, pp. 1437–1450.

[59] Burke, P. and Resosudarmo, B.P. (2012), “Survey of Recent Developments”,

Bulletin of Indonesian Economic Studies, 48:3, pp. 299–324.

[60] Busch, J., Ferreti-Gallon, Engelmann, J., Wright, M., Austin, K., Stolle, F.,

Turubanova, S., Potapov, P., Margono, B., Hansen, M., and Baccini, A. (2015),

“Reduction in emissions from deforestation from Indonesia’s moratorium on

new oil palm, timber, and logging concessions”, Proceedings of the National

Academy of Sciences, 112:5.

[61] Butler, R.A., Koh, L.P., and Ghazoul, J. (2009), “ REDD in the red: palm oil could

undermine carbon payment schemes”, Conservation Letters, 2:2, pp. 67–73.

[62] Cahyadi, E.R. and Waibel, H. (2013), Is contract farming in the Indonesian oil

palm industry pro-poor? Journal of Southeast Asian Economies, 30:1, pp. 62–76.

[63] Carlson, K., Curran, L., Asner, G., McDonald Pittman, A., Trigg, S., Adeney,

M. (2013), “Carbon emissions from forest conversion by Kalimantan oil palm

plantations”, Nature Climate Change, 3, pp. 283–287.

[64] Carmignani, F. (2013), “Development outcomes, resource abundance, and

the transmission through inequality”, Resource and Energy Economics, 35, pp.

412–428.

173

[65] Carmignani, F. and Avom, D. (2010), “The social development effects of

primary commodity export dependence”, Ecological Economics, 70:2, pp.

317–330.

[66] Caselli, F. (2005), “Accounting for Cross-Country Income Differences, in

Handbook of Economic Growth, Vol 1., ed. Aghion, P. and Durlauf, S., Chapter

9, pp. 679–741, Elsevier.

[67] Caselli, F. and Michaels, G. (2013), “Do Oil Windfalls Improve Living

Standards? Evidence from Brazil”, American Economic Journal: Applied

Economics, 5:1, pp. 208–238.

[68] Cassing, J. and Warr, P. G. (1985), “The distributional impacts of a resource

boom”, Journal of International Economics, 18, pp. 301–319.

[69] Centre for the Study of Civil War (2009), “Uppsala Conflict Data

Program/Peace Research Institute Oslo Armed Conflict Dataset”, Version

4-2009.

[70] Chan, H., Frey, B., Gallus, J., and Torgler, B. (2014), “Academic honors and

performance”, Labour Economics, 31, pp. 188–204.

[71] Cramb, R. and McCarthy, J. (2016), “The Oil Palm Complex: Smallholders,

Agribusiness and the State in Indonesia and Malaysia”, National University of

Singapore Press.

[72] Christiaensen, L., Demery, L., and Kuhl, J. (2011), “The (evolving) role

of agriculture in poverty reduction—an empirical perspective”, Journal of

Development Economics, 96:2, pp. 239–54.

[73] Ciccone, A. (2011), “Economic Shocks and Civil Conflict: A Comment”,

American Economic Journal: Applied Economics, 3, pp. 215–215.

174

[74] Cleveland, C. J., Costanza, R., Hall, C., and Kaufmann, R. (1984), “Energy and

the U.S. economy: a biophysical perspective”, Science, 225, pp. 890–897.

[75] Collier, P. and Goderis, B. (2012), “Commodity prices and growth: an empirical

investigation”, European Economic Review, 56, pp. 1241–1260.

[76] Corden, W.M. (1984), “Booming Sector and Dutch Disease Economics: Survey

and Consolidation”, Oxford Economic Papers, 36:3, pp. 359–380.

[77] Corden, W.M. and Neary, P. (1982), “Booming sector and de-industrialisation

in a small open economy”, Economic Journal, 92: 368, pp. 825–848.

[78] Corley, R., and Tinker, R. (2003), “The Oil Palm”, 4th Edition, Blackwell,

Oxford.

[79] Cotet, A. and Tsui, T. (2013), “Oil and Conflict: What Does the Cross Country

Evidence Really Show?” American Economic Journal: Macroeconomics, 5:1, pp.

49–80.

[80] Cotet, A. and Tsui, T. (2013), “Oil, Growth, and Health: What Does the

Cross-Country Evidence Really Show?” Scandinavian Journal of Economics,

115:4, pp. 1107–1136.

[81] Costinot, A., Donaldson, D., Smith, C. (2016), “Evolving Comparative

Advantage and the Impact of Climate Change on Agricultural Markets: Evidence

form 1.7 Million Fields around the World”, Journal of Political Economy,

forthcoming.

175

[82] Coxhead, I. and Li, M. (2008), “Prospects for skills-based exports in

resource-rich developing countries: Indonesia in comparative perspective”,


[83] Coxhead, I. (2014), “Indonesia: returns to occupations, education, and ability

during a resource export boom”, in Ahtukorala, Patron, and Resosudarmo

(ends), “Trade development and Political Economy in East Asia: Essays in

Honour of Hal Hill”, Singapore, ISEAS.

[84] Cragg, J.G. and Donald, S.G. (1994), “Testing Identifiability and Specification

in Instrumental Variable Models”, Econometric Theory, 9, pp. 222–240.

[85] Cramb, R.A. (2013), “Palmed off: incentive problems with joint-venture

schemes for oil palm development on customary land”, World Development, 43,

pp. 84–99.

[86] Cramb, R.A., and Sujang, P.S. (2013), “The mouse deer and the crocodile: oil

palm smallholders and livelihood strategies in Sarawak, Malaysia“, Journal of

Peasant Studies, 40:1, pp. 129–154.

[87] Cust, J., and Poelhekke, S. (2015), “The Local Economic Impacts of Natural

Resource Extraction”, Annual Review of Resource Economics, 7:5.1–5.18.

[88] Daniele, V. (2011), “Natural resources and the quality of development”, Journal

of Development Studies, 47:4, pp. 545–573.

[89] De Silva, I., and Sumarto, S. (2014), “Does Economic Growth Really Benefit

the Poor? Income Distribution Dynamics and Pro-poor Growth in Indonesia,”


[90] de Soysa, I., and Gizelis, T. (2012), “The natural resource curse and the spread

of HIV/AIDS, 1990–2008”, Social Science and Medicine, 77, pp. 90–96

[91] Deaton, A. (2013), “The Great Escape”, Princeton University Press.

176

[92] Dennis, R., Mayer, J., Applegate, G., Chokkalingam, U., Colfer, C., Kurniawan,

I., Lachowski, H., Maus, P., Permana, R., Ruchiat, Y. (2005), “Fire, people and

pixels: Linking social science and remote sensing to understand underlying

causes and impacts of fires in Indonesia”, Human Ecology, 33, pp. 465–504.

[93] Dercon, S. (2009), “Rural poverty: old challenges in new contexts”, World Bank

Research Observer, 24, pp. 1–28.

[94] Dercon, S., and Gollin, D. (2014), “Agriculture in African Development:

Theories and Strategies”, Annual Review of Resource Economics, 6, pp. 471–492.

[95] Dollar, D., Kleineberg, T., and Kraay, A. (2016), “Growth still is good for the

poor”, European Economic Review, 81, pp. 68–85.

[96] Domenech, J. (2008), “Mineral Resource Abundance and Regional Growth in

Spain, 1860–2000”, Journal of International Development, 20, pp. 1122-1135.

[97] Dube, O. and Vargas, J. (2013), “Commodity Price Shocks and Civil Conflict:

Evidence from Columbia”, Review of Economic Studies, 80, pp. 1384–1421.

[98] Ebeke, C., Omgba, L., and Laajaj, R. (2015), “Oil, governance and the (mis)

allocation of talent in developing countries”, Journal of Development Economics,

114, pp. 126–141.

[99] Edwards, R.B. (2016), “Mining away the Preston curve”, World Development,

78, pp. 22–36.

[100] Eichengreem B., Park, D., and Shin, K. (2013), “Growth slowdowns redux: new

evidence on the middle-income trap”, NBER Working Papers No. 18673.

[101] Ellison, G., Glaeser, E., and Kerr, W. (2010), “What cause Industry

Agglomeration? Evidence from Conglomeration Patterns”, American Economic

Review, 100: 3, pp. 1195–1213.

177

[102] Evenson, R. E. and Gollin, D. (2003), “Assessing the impact of the Green

Revolution, 1960–2000”, Science, 300: 5620, pp. 758–762.

[103] Factor-Litvak, P., Wasserman, G., Kline, J., and Graziano, J., (1999), “The

Yugoslavia Prospective Study of Environmental Lead Exposure”, Environmental

Health Perspectives, 107:1, pp. 9–15.

[104] Fadliya and McLeod, R. (2010), “Fiscal transfers to regional governments in

Indonesia”, Arndt-Corden Department of Economics Working Papers in Trade

and Development, Working Paper No. 2010/14.

[105] Fargione, J., Hill, J., and Tilman, D. (2008), “Land clearing and the biofuel

carbon debt”, Science, 319, pp. 1235–1238.

[106] Farhadi, M., Islam, M. R., and Moslehi, S. (2015), “Economic freedom and

productivity growth in resource-rich economies”, World Development, 72, pp.

109–126.

[107] Fitriani, F., Hofman, B., and Kaiser, K. (2005), “Unity in diversity? The creation

of new local governments in decentralising Indonesia”, Bulletin of Indonesian

Economic Studies, 74:1, pp. 70–93.

[108] Fischer, G., van Nelthuizen, H., Shah, M., and Nachtergaele, F. (2002), Global

Agro-ecological Assessment for Agriculture in the 21st Century: Methodology

and Results, Food and Agriculture Organisation, Rome.

[109] Fleming, D.A., and Measham, T.G. (2014), “Local job multipliers of mining”,

Resources Policy, 41, pp. 9–15.

[110] Fuglie, K.O. (2010), “Source of growth in Indonesian agriculture”, Journal of

Productivity Analysis, 33, pp. 225–240.

[111] Fuller, W. (1977), “Some Properties of a Modification of the Limited

Information Estimator”, Econometrica, 45:4, pp. 939–954.

178

[112] Gamu, J., Le Billon, P., and Spiegel, S. (2015), “Extractive industries and

poverty: a review of recent findings and linkage mechanism”, Extractive

Industries and Society, 2, pp. 162–176.

[113] Gandataruna, K. and Haymon, K. (2011), “A dream denied? Mining legislation

and the Constitution in Indonesia”, Bulletin of Indonesian Economic Studies,

47:2, pp. 221–231

[114] Garnaut, R. (2015), “Indonesia’s Resources Boom in International Perspective:

Policy Dilemmas and Options for Continued Strong Growth”, Bulletin of

Indonesian Economic Studies, 51:2, pp. 189–212.

[115] Gaskell, J. C. (2015), “The Role of Markets, Technology, and Policy in

Generating Palm-Oil Demand in Indonesia”, Bulletin of Indonesian Economic

Studies, 51:1, pp. 29–45.

[116] Gatto, M., Wollni, M., and Qaim, M. (2015), “Oil palm boom and land use

dynamics in Indonesia: the role of policies and socioeconomic factors”, Land

Use Policy, 46, pp. 292–303.

[117] Gellert, P. (2015), “Palm Oil Expansion in Indonesia: Land Grabbing

as Accumulation by Dispossession”, in Jon Shefner (ed.) States and

Citizens: Accommodation, Facilitation and Resistance to Globalization (Current

Perspectives in Social Theory, Volume 34), Emerald Group Publishing Limited,

pp.65 –99.

[118] Gingold, B., Rosenbarger, A., Muliastra, Y., Stolle, F., Sudana, I., Manessa,

M., Murdimanto, A., Tiangga, S., Madusari, C., and Douard, P. (2012), “How

to identify degraded land for sustainable palm oil in Indonesia.” Working Paper,

World Resources Institute and Sekala, Washington D.C.

179

[119] Gibbs, H. K., Ruesch, A. S., Achard, F., Clayton, M. K., Holmgren, P.,

Ramankutty, N., and Foley, J. A. (2010), “Tropical forests were the primary

sources of new agricultural land in the 1980s and 1990s”. Proceedings of the

National Academy of Sciences, 107:38, pp. 16732–16737.

[120] Goderis, B. and Malone, S.W. (2011), “Natural resource booms and inequality:

Theory and evidence”, Scandinavian Journal of Economics, 113: 2, pp. 388–417.

[121] Gollin, D. (2010), “Agricultural Productivity and Economic Growth”, in

Handbook of Agricultural Economics, pp. 3825–3866.

[122] Gollin, D., Parente, S., and Rogerson, R. (2002), “The Role of Agriculture in

Development”, American Economic Review, 92:2, pp. 160–164.

[123] Greenstone, M., Hornbeck, R., and Moretti, E. (2010), “Identifying

Agglomeration Spillovers: Evidence from Winners and Losers of Large

Plant Openings”, Journal of Political Economy, 119: 3, pp. 536–598.

[124] Gregory, R.G. (1976), “Some implications of the growth of the mineral sector”,

Australian Journal of Agricultural Economics, 20:2, pg. 71.

[125] Griliches, Z. (1969), “Capital–skill complementarity”, Review of Economics

and Statistics, 51: 4, pp. 465–468.

[126] Gylfason, T. (2001), “Natural Resources, Education, and Economic

Development”, European Economic Review, 45, pp. 847–859.

[127] Gylfason, T., Herbertsson, T., and Zoega, G. (1999), “A mixed blessing: natural

resources and economic growth”, Macroeconomic Dynamics, 3, pp. 204–225.

[128] Gylfason, T. and Zoega, G. (2003), “Education, social equality, and economic

growth: a view of the landscape”, CESifo Economic Studies, 49: 4, pp. 557–579.

[129] Gylfason, T. and Zoega, G. (2006), “Natural Resources and Economic Growth:

The Role of Investment”, World Economy, 29:8, pp. 1091–1115.

180

[130] Hall, C., Tharakan, P., Hallock, J., Cleveland, C., and Jefferson, M. (2003)

“Hydrocarbons and the evolution of human culture“, Nature, 426, pp. 318–322.

[131] Hanson, G. (2012) “The rise of middle kingdoms: emerging economies in

global trade”, Journal of Economic Perspectives, 26: 2, pp. 41–64.

[132] Hasnah, Fleming, E., and Coelli, T. (2004), “Assessing the performance of a

nucleus estate and smallholder scheme for oil palm production in West Sumatra:

a stochastic frontier analysis”, Agricultural Systems, 79, pp. 17–30.

[133] Hayami, T. (2010), “Plantations Agriculture”, in Handbook of Agricultural

Economics, pp. 3305–3322.

[134] Heston, A., Summers, R., and Aten, B. (2012), “Penn World Table Version 7.1”,

Center for International Comparisons of Production, Income and Prices at the

University of Pennsylvania.

[135] Hill, H. (1996), “The Indonesian Economy Since 1996: Southeast Asia’s

Emerging Giant”, Cambridge University Press, London.

[136] Hill, H. (2014), “Regional Dynamics in a Decentralized Indonesia”, Vol. 501,

Institution of Southeast Asian Studies.

[137] Hill, H., Khan, M.E., and Zhuang, J. (2012), “Diagnosing the Indonesian

Economy: Toward Inclusive and Green Growth”, Anthem Press and the Asian

Development Bank, Manilla.

[138] Hill, H., Resosudarmo, B.P., and Vidyattama, Y., (2008), “Indonesia’s Changing

Economic Geography”, Bulletin of Indonesian Economic Studies, 44:3, pp.

407–435.

[139] Hirschman, A.O. (1981), “Essays in Trespassing: Economics to Politics and

Beyond”, Cambridge University Press, New York.

181

[140] Horiuchi, Y. and Mayerson, A. (2015), “The Opportunity Cost of Conflict:

Statistically Comparing Israel and Synthetic Israel”, Political Science and

Research Methods, May, pp. 1–10.

[141] Hornbeck, R., and Keskin, P. (2015), “Does Agriculture Generate Local

Economic Spillovers? Short-run and Long-Run Evidence from the Ogallala

Aquifer”, American Economic Journal: Economic Policy, 7:2, pp. 192–213.

[142] Hunt, C. (2010), “The costs of reducing deforestation in Indonesia”, Bulletin of

Indonesian Economic Studies, 46:2, pp. 187–192.

[143] Iyengar, G., and Nair, P. (2000), “Global outlook on nutrition and the

environment: meeting the challenges of the next millennium”, The Science of

the Total Environment, 249:1–3, pp. 331–346.

[144] James, A. and Aadland, D. (2011), “The curse of natural resources: an empirical

investigation of US countries”, Resource and Energy Economics, 33, pp 440–463.

[145] James, A. and James, R. (2011), “Do resource dependent regions grow slower

than they should?”, Economics Letters, 111, pp. 194–196.

[146] Jedwab, R. (2013), “Urbanisation without structural transformation: evidence

from consumption cities in Africa”, mimeo, George Washington University.

[147] Johnston, B.F., and Mellor, J.W. (1961), “The Role of Agriculture in Economic

Development”, American Economic Review, 51:4, pp. 566–593

[148] Kander, A., Malanima, P., and Warde, P. (2013), “Power to the People–Energy

and Economic Transformation of Europe over Four Centuries”, Princeton

University Press, Princeton.

[149] Kander, A., and Stern, D.I. (2014), “Economic growth and the transition from

traditional to modern energy in Sweden”, Energy Economics, 36, pp. 56–65.

182

[150] Karlsson, M., and Pichler, S. (2015), “Demographic consequences of HIV”.

Journal of Population Economics, 28, pp. 1097–1135.

[151] Kaufmann, D., Kraay, A., and Mastruzzi, M. (2013), “The Worldwide

Governance Indicators Project”, World Bank, Washington, D.C.

[152] Kleibergen, F., and Paap, R. (2006), “Generalized reduced rank test using the

singular value decomposition”, Journal of Econometrics, 127:1, pp. 97–126.

[153] Koh, L. P., Miettinen, J., Liew, S. C., and Ghazoul, J. (2011), “Remotely sensed

evidence of tropical peatland conversion to oil palm”. Proceedings of the

National Academy of Sciences, 108:12, pp. 5127–5132

[154] Koh, L., and Wilcove, D. (2007), “Cashing in palm oil for conservation”, Nature,

448, pp. 993–994.

[155] Koh, L., and Wilcove, D. (2008), “Is oil palm agriculture really destroying

tropical biodiversity?”, Conservation Letters, 1, pp. 60–64.

[156] Kraay, A. (2006), “When is growth pro-poor? Evidence from a panel of

countries”, Journal of Development Economics, 80, pp. 198–227.

[157] Krief, N., Grieve, R., Gangartner, D., Turner, A., Nikolova, S., and Sutton,

M. (2015), “Examination of the synthetic control method for evaluating health

policies with multiple treated units”, Health Economics, forthcoming.

[158] Krugman, P. (1987), “The Narrow Moving Band, the Dutch Disease, and

the Competitive Consequences of Mrs. Thatcher”, Journal of Development

Economics, 27, pp. 41–55

[159] Lederman, D. and Maloney, W. (2007), “Natural resources: neither curse nor

destiny”, Stanford University Press, Stanford, CA.

183

[160] Lee, J., Ghazoul, J., Obidzinski, K., and Koh, L. (2013), “Oil palm smallholder

yields and incomes constrained by harvesting practices and type of smallholder

management in Indonesia”, Agronomy for Sustainable Development, 34:2, pp.

510–513.

[161] Lei, Y. and Michaels, G. (2014), “Do giant oilfield discoveries fuel internal

armed conflicts?”, Journal of Development Economics, 110, pp. 139–157.

[162] Lewis, A.W. (1954), “Economic Development with Unlimited Supplies of

Labor”, Manchester School, pp. 139–91.

[163] Lewis, B. and Smoke, P. (2015), “Intergovernmental Fiscal Transfers and Local

incentives and Responses: The Case of Indonesia”, Fiscal Studies, forthcoming.

[164] Loayza, N.V., and Raddatz, C. (2010), “The composition of growth matters for

poverty alleviation”, Journal of Development Economics, 93, pp. 137–151.

[165] Maertens M, and Swinnen, J. (2009), “Trade, standards, and poverty: evidence

from Senegal”, World Development, 37:1, pp. 161–178.

[166] Manning, C. (2010), “Labor market adjustment to Indonesia’s economic crisis:

context, trends, and implications”, Bulletin of Indonesian Economic Studies, 36:1,

pp. 105–136.

[167] Manning, C., and Sumarto, S. (2011), “Employment, Living Standards, and

Poverty in Contemporary Indonesia”, Indonesia Update Series, Institute of South

East Asian Studies, Singapore.

[168] Manzano, O, and Rigobon, R. (2001), “Resource curse or debt overhang?”

NBER Working Paper 8390, July.

[169] Matsuyama, K. (1992), “Agricultural Productivity, Comparative Advantage,

and Economic Growth”, Journal of Economic Theory, 58, pp. 317–334.

184

[170] Maxwell, S., and Fernando, A. (1989), “Cash Crops in Developing Countries:

The Issues, the Facts, the Policies”, World Development, 17:11, pp. 1677–1708.

[171] Mayumi, K. (1991), ‘Temporary emancipation from land: from the Industrial

Revolution to the present time”, Ecological Economics, 4, pp. 35–56.

[172] McCarthy, J. (2010), “Processes of inclusion and adverse incorporation: oil

palm and agrarian change in Sumatra, Indonesia”, Journal of Peasant Studies,

37:4, pp. 821–850.

[173] McCarthy, J., Gillespie, P., Zen, Z. (2011), “Swimming Upstream: Local

Indonesian Production Networks in “Globalised” Palm Oil Production”, World

Development, 40:3, pp. 555–569.

[174] McCulloch, N., and Malesky, E. (2011), “Does Better Local Governance

Improve District Growth Performance in Indonesia”, IDS Working Paper 369.

[175] McCulloch, N., and Sjahrir, B. (2008), “Endowments, Location, or Luck?

Evaluating the Determinants of Sub-National Growth in Decentralized

Indonesia”, World Bank Policy Research Working Paper 4769.

[176] McKinsey Global Institute (2013), “Reverse the curse: Maximizing the

potential of resource-driven economies”, McKinsey Global Institute Report.

[177] Mehlum, H., Moene, K., and Torvik, R. (2006), “Institutions and the Resource

Curse”, Economic Journal, 116, pp. 1–20.

[178] Meng, X., Manning, C, Shi, L, and Effendi, T. (2010), “The Great Migration:

Rural?Urban Migration in China and Indonesia”, Edward Elgar Publishing.

[179] Michaels, G. (2010), “The long term consequences of resource-based

specialisation”, Economic Journal, 121, pp. 31–57.

[180] Mideksa, T. (2013), “The economic impact of natural resources”, Journal of

Environmental Economics and Management, 65, pp. 277–289.

185

[181] Mikesell, R.F. (1997), “Explaining the resource curse with special reference to

mineral-exporting countries”, Resources Policy, 23:4, pp. 191–199.

[182] Miranti, R. (2010), “Poverty in Indonesia 1984-2002: the impact of growth and

changes in inequality”, Bulletin of Indonesian Economic Studies, 46:1, pp. 79–97.

[183] Miriam, E., Marlier, I., DeFries, R., Kim, P., Koplitz, S., Jacob, D., Mickley, L.,

and Myers, S. (2015), “Fire emissions and regional air quality impacts from fires in

oil palm, timber, and logging concessions in Indonesia”, Environmental Research

Letters, 10: 8.

[184] Montalvo, J.G.,(2011) “Voting after the bombings: a natural experiment on

the effect of terrorist attacks on democratic elections”, Review of Economics and

Statistics, 93:4, pp. 1146–1154.

[185] Moreira, M.J. (2003), “A Conditional Likelihood Ratio Test for Structural

Models”, Econometrica, 71(4), pp. 1027–1048

[186] Moretti, E. (2010), “Local Multipliers”, American Economic Review: Papers

and Proceedings 100, pp. 373–377.

[187] Munasib, A. and Rickman, D. S. (2015), “Regional economic impacts of the

shale gas and tight oil boom: A synthetic control analysis”, Regional Science and

Urban Economics, 50, pp. 1-17.

[188] Murray, M.P. (2006), “Avoiding Invalid Instruments and Coping with Weak

Instruments”, Journal of Economic Perspectives, 20:4, pp.111–132.

[189] Newhouse, D., and Suryadarma, D. (2011), “The value of vocational education:

high school type and labor market outcomes in Indonesia”, Bulletin of

Indonesian Economic Studies, 25: 2, pp. 296–322

[190] Norman, C.S. (2009), “Rule of Law and the Resource Curse: Abundance Versus

Intensity”, Environmental and Resource Economics, 43, pp. 183–207.

186

[191] Nunn, N., and Qian, N. (2011), “The potato’s contribution to population

and urbanization: evidence from a historical experiment”, Quarterly Journal of

Economics, 126, pp. 593–650.

[192] Obidzinski, K., Dermawan., A., and Hadianto, A. (2014), “Oil palm plantation

investments in Indonesia’s forest frontiers: limited economic multipliers and

uncertain benefits for local communities”, Environment, Development, and

Sustainability, 16, pp. 1177–1196.

[193] Oster, E. (2015), “Unobservable Selection and Coefficient Stability: Theory and

Evidence”, Mimeo.

[194] Papyrakis, E., and Gerlagh, R. (2007), “Resource abundance and economic

growth in the United States”, European Economic Review, 51, pp. 1101-1039.

[195] Pepinski, T.B., and Wihardja, M.M. (2011), “Decentralization and Economic

Performance in Indonesia”, Journal of East Asian Studies, 11:3, pp. 337–371.

[196] Pierskalla, J.H. (2016), “Splitting the Difference? The politics of district creation

in Indonesia”, Comparative Politics, 42:2, forthcoming.

[197] Pinotti, P. (2015) “The economic costs of organised crime: evidence from

southern Italy”, Economic Journal, 125:586, pp. F203–F232.

[198] Preston, S.H. (1975), “The changing relation between mortality and level of

development”, Population Studies, 29:2, pp. 231–248.

[199] Pritchett, L. and Summers, L. H. (1996), “Wealthier Is Healthier”, Journal of

Human Resources, 31, pp. 841–868.

[200] Pryor, F. (1982), “The Plantation Economy as an Economic System”, Journal of

Comparative Economics, 7, pp. 288–317.

187

[201] Quizon, J., and Binswanger, H. (1986), “Modelling the Impact of Agricultural

Growth and Government Policy on Income Distribution in India,” World Bank

Economic Review, 1:1, pp. 103–48.

[202] Rada, N., Buccola, S.T., and Fuglie, K.O. (2011), “Government policy and

agricultural productivity in Indonesia”, American Journal of Agricultural

Economics, 93:3, pp. 867–884.

[203] Ranis, G., and Fei, J. (1961), “A theory of economic development”, American

Economic Review, 51, pp. 533–565.

[204] Ravallion M, and Chen, S. (2007), “China’s (uneven) progress against poverty”,

Journal of Development Economics, 82:1, pp. 1–42

[205] Ravallion, M., and Chen, S. (2003), “Measuring pro-poor growth”, Economic

Letters, 78, pp. 93–99.

[206] Ravallion, M., and Datt, G. (2002), “Why has economic growth been more

pro-poor in some states of India than others?”, Journal of Development

Economics, 68, pp. 381–400.

[207] Resosudarmo, B. P. (2005), “The Politics and Economics of Indonesia’s Natural

Resources”, Indonesia Update Series, Institute of South East Asian Studies,

Singapore.

[208] Restuccia, D., Yang, D. T., and Zhu, X. (2008), “Agriculture and Aggregate

Productivity: a Quantitative Cross-Country Analysis”, Journal of Monetary

Economics, 55: 2, pp. 234–250.

[209] Rist, L., Feintrenie, L., and Levang, P. (2010), “The livelihood impacts of

oil palm: smallholders in Indonesia”, Biodiversity and Conservation, 19:4, pp.

1009–1024.

188

[210] Rival, A., and Levang, P. (2014), “Palms of controversies: Oil palm and

development challenges”, Center for International Forestry Research, Bogor.

[211] Rodrik, D. (2015), “Premature industrialisation”, IAS School of Social Science,

Economics Working Papers, No. 107, Princeton University.

[212] Ross, M.L. (2001), “Does OIL hinder democracy?” World Politics, 53, pp.

325–61.

[213] Ross, M.L. (2007), “How mineral-rich states can reduce inequality”, Escaping

the resource curse, pp 237–75.

[214] Sachs, J. and Warner, A. (1995), “Natural resource abundance and economic

growth”, NBER Working Paper 5398, December.

[215] Sachs, J. and Warner, A. (2001), “The curse of natural resources”, European


[216] Saha, S., Pattanayak, E., and Sills (2011), “Undermining Health: Environmental

Justice and Mining in India”, Health and Place, 17:1, pp. 140–148.

[217] Sala-i-Martin, X., Doppelhofer, G., and Miller, R.I (2004), “Determinants

of long-term growth: a Bayesian Averaging of Classical Estimates (BACE)

approach”, American Economic Review, 94, pp. 813–835.

[218] Salami, I., Zulfikar, A., Marselina, M., and Roosmini, D. (2014), “Respiratory

health risk assessment of children living close to industrial areas in Indonesia”,

Reviews on Environmental Health, 29, pp. 1–2.

[219] Sarsons, H. (2015), “Rainfall and conflict: A cautionary tale”, Journal of

Development Economics, 115, pp. 62–72.

[220] Sayer, J., Ghazoul, J., Nelson, P., Boedhihartono, A. (2012), “Oil palm expansion

transforms tropical landscapes and livelihoods”, Global Food Security, 1, pp.

114–119.

189

[221] Schultz, T.W. (1964), “Transforming Traditional Agriculture”, Yale University

Press, New Haven, Connecticut.

[222] Sills, E., Herrera, D., Kirkpatrick, J., Brandao, A., Sickson, Y., Hall, S.,

Pattanayak, S., Shoch, D., Vedoveto, M., Young, L., Pfaff, A. (2015), “Estimating

the Impacts of Local Policy Innovation: The Synthetic Control Method Applied

to Tropical Deforestation”, PLoS ONE, 10:7, e0132590.

[223] Smith, B. (2015), “The resource curse exorcised: Evidence from a panel of

countries”, Journal of Development Economics, 117, pp. 57–73.

[224] Smith, B. and Wills, S. (2015), “Left in the Dark? Oil and Rural Poverty”,

OxCarre Research Paper 164, Oxford University.

[225] Sokoloff, K.L. and Engerman, S.L. (2000), “History Lessons: Institutions, Factor

Endowments and Paths of Development in the New World”, Journal of Economic

Perspectives, 14:3, pp. 217–232.

[226] Sparrow, R., Grimm, M., and Tasciotti, L. (2015), “Does Electrification Spur

the Fertility Transition? Evidence from Indonesia”, Demography, 52:5, pp.

1773–1796.

[227] Staiger, D. and Stock, J. H. (1997), “Instrumental variables regression with weak

instruments”, Econometrica, 65:3, pp. 556–586.

[228] Stern, D.I. (2010), “Between Estimates of the Emissions-Income Elasticity,”

Ecological Economics, 69, pp. 2173–2182.

[229] Stijns, J.P. (2006), “Natural Resource Abundance and Human Capital

Accumulation”, World Development, 34:6, pp. 1060–1083.

190

[230] Stock, J. H., and Yogo, M. (2005), “Testing for weak instruments in linear IV

regression,” in Donald W. K. Andrews and James H. Stock (Eds.), Identification

and Inference for Econometric Models: Essays in Honor of Thomas Rothenberg,

Cambridge, UK: Cambridge University Press.

[231] Suryadarma, D. and Jones, G.W. (2013), “Meeting the education challenge”,

in Suryadarma and Jones (eds), “Education in Indonesia, Singapore, ISEAS, pp.

1–14.

[232] Suryahadi, A., Sumamto, S., and Pritchett, L. (2003), “Evolution of poverty

during the crisis in Indonesia”, Asian Economic Journal, 17:3, pp. 221–241.

[233] Suryahadi, A., Suryadarma, D., and Sumarto, S. (2009), “The effects of

location and sectoral components of economic growth on poverty: Evidence from

Indonesia”, Journal of Development Economics, 89:1, pp. 109–117.

[234] Taubman, P. (1989), “Role of Parental Income in Educational Attainment”,

American Economic Review, Papers and Proceedings, May 1989, pp. 57–61.

[235] Temple, J. (2003), “Growing into trouble: Indonesia after 1966”, in Rodrik,

D. (ed.), “In search of prosperity: analytic narratives on economic growth”,

Princeton University Press, Princeton.

[236] Thorbecke, E., and Jung, H. (1996), “A Multiplier Decomposition Method

to Analyze Poverty Alleviation”, Journal of Development Economics, 48:2 pp.

253–277.

[237] Tiffen, M.m and Mortimore, M. (1990), “Theory and Practice in Plantation

Agriculture: an Economic Review”, Overseas Development Institute, Russell

Press, London.

[238] Torvik, R. (2001) “Learning by doing and the Dutch disease”, European


191

[239] Torvik, R. (2002), “Natural resources, rent seeking, and welfare”, Journal of

Development Economics, 65, pp. 455–470

[240] Torvik, R. (2009) “Why do some resource-abundant countries succeed while

others do not”, Oxford Review of Economic Policy, 25:2, pp. 241–256.

[241] Tsui, K. (2011), “More oil, less democracy: evidence from worldwide crude oil

discoveries”, Economic Journal, 121, pp. 89–115.

[242] United Nations (2014),“National Accounts Estimates of Main Aggregates”,

United Nations Statistics Division.

[243] United Nations (2013), “Contribution of mining to total value added”,

Environmental Indicators: Energy and Minerals.

[244] United States Department of Agriculture (2015), “Oilseeds: world markets and

trade”, Washington, D.C.

[245] van der Eng, P. (2009), “Capital formation and capital stock in Indonesia,

1950–2008”, Bulletin of Indonesian Economic Studies, 45:3, pp. 345–371.

[246] van der Eng, P. (2014), “Mining and Indonesia’s Economy: Institutions and

Value Adding, 1870–2010”, Poverty Reduction, Institutions, Markets and Policies

in Developing Countries (PRIMCED) Discussion Paper Series, No. 57. Institute

of Economic Research, Hitotsubashi University.

[247] van der Ploeg, F. (2011), “Natural Resources: Curse or Blessing?” Journal of

Economic Literature, 49:2, pp. 366–420.

[248] van der Ploeg, F. and Poelhekke, S. (2010), “The Pungent Smell of Red Herrings:

subsoil assets, rents, volatility and the resource curse”, Journal of Environmental

Economics and Management, 60:1, pp. 44–55.

[249] van der Ploeg, F. and Poelhekke, S. (2009), “Volatility and the natural resource

curse”, Oxford Economic Papers, 61:4, pp. 727–760.

192

[250] van Winjnbergen, S. (1984), “The Dutch Disease: a disease after all?” Economic

Journal, 94: 373, pp. 41–55.

[251] Vollrath, D., Gollin, D., and Jedwab, R. (2015), “Urbanization with and without

Industrialization”, Journal of Economic Growth, forthcoming.

[252] Wetterberg, A., Sumarto, S., and Pritchett, L. (1999), “A national snapshot of

the social impact of Indonesia’s crisis”, Bulletin of Indonesian Economic Studies,

35:3, pp. 145–152.

[253] Wheeler, D., Hammer, D., Kraft, R., Dasgupta, S., and Blankespoor, B.

(2013), “Economic dynamics and forest clearing: a spatial econometric analysis”,

Ecological Economics, 85, pp. 85–96.

[254] Wick, K. and Bulte, E. (2009), “The Curse of Natural Resources”, Annual

Review of Economics, 1:1.1-1.17.

[255] Williams, A. (2011), “Shining a light on the resource curse: an empirical

analysis of the relationship between natural resources, transparency, and

economic growth”, World Development, 39:4, pp. 490–505.

[256] World Bank (2013), “Africa’s Pulse”, The Office of the Chief Economist for the

Africa Region. October 2013, Volume 8, World Bank, Washington, D.C.

[257] World Bank (2014), “World Development Indicators”. World Bank data portal,

Washington, D.C.

[258] World Bank (2014), “Indonesia: Avoiding the Trap”, 2014 Development Policy

Review, World Bank, Jakarta.

[259] World Bank (2015,) “Indonesia Database for Economic and Policy Research”,

World Bank data portal, Washington, D.C.

[260] Wrigley, E. A. (2010), “Energy and the English Industrial Revolution”.

Cambridge University Press, Cambridge.

193

[261] Yusuf, A.A. (2013), “The evolution of inequality in Indonesia, 1990–2012”,

Department of Economics, Padjadjaran University, Working Papers in Economics

and Development Studies, No. 2013114, Bandung.

Documents

Natural Resource Sectors and Human Development ... · Natural Resource Sectors and Human Development: International and Indonesian Evidence Ryan B. Edwards a thesis submitted for