80248
DIREC TIONS IN DE VELOPMENT
Poverty
Understanding Changes in Poverty
Gabriela Inchauste, João Pedro Azevedo, B. Essama-Nssah,
Sergio Olivieri, Trang Van Nguyen,
Jaime Saavedra-Chanduvi, and Hernan Winkler
Understanding Changes in Poverty
Direc tions in De velopment
Poverty
Understanding Changes in Poverty
Gabriela Inchauste, João Pedro Azevedo, B. Essama-Nssah, Sergio Olivieri,
Trang Van Nguyen, Jaime Saavedra-Chanduvi, and Hernan Winkler
© 2014 International Bank for Reconstruction and Development / The World Bank
1818 H Street NW, Washington DC 20433
Telephone: 202-473-1000; Internet: www.worldbank.org
Some rights reserved
1 2 3 4 17 16 15 14
This work is a product of the staff of The World Bank with external contributions. The findings, interpreta-
tions, and conclusions expressed in this work do not necessarily reflect the views of The World Bank, its
Board of Executive Directors, or the governments they represent. The World Bank does not guarantee the
accuracy of the data included in this work. The boundaries, colors, denominations, and other information
shown on any map in this work do not imply any judgment on the part of The World Bank concerning the
legal status of any territory or the endorsement or acceptance of such boundaries.
Nothing herein shall constitute or be considered to be a limitation upon or waiver of the privileges and
immunities of The World Bank, all of which are specifically reserved.
Rights and Permissions
This work is available under the Creative Commons Attribution 3.0 IGO license (CC BY 3.0 IGO) http://
creativecommons.org/licenses/by/3.0/igo. Under the Creative Commons Attribution license, you are free
to copy, distribute, transmit, and adapt this work, including for commercial purposes, under the following
conditions:
Attribution—Please cite the work as follows: Inchauste, Gabriela, João Pedro Azevedo, B. Essama-Nssah,
Sergio Olivieri, Trang Van Nguyen, Jaime Saavedra-Chanduvi, and Hernan Winkler. 2014.
Understanding Changes in Poverty. Directions in Development. Washington, DC: World Bank.
doi:10.1596/978-1-4648-0299-7. License: Creative Commons Attribution CC BY 3.0 IGO
Translations—If you create a translation of this work, please add the following disclaimer along with the
attribution: This translation was not created by The World Bank and should not be considered an official
World Bank translation. The World Bank shall not be liable for any content or error in this translation.
Adaptations—If you create an adaptation of this work, please add the following disclaimer along with the
attribution: This is an adaptation of an original work by The World Bank. Responsibility for the views and
opinions expressed in the adaptation rests solely with the author or authors of the adaptation and are not
endorsed by The World Bank.
Third-party content—The World Bank does not necessarily own each component of the content contained
within the work. The World Bank therefore does not warrant that the use of any third-party-owned
individual component or part contained in the work will not infringe on the rights of those third
parties. The risk of claims resulting from such infringement rests solely with you. If you wish to
re-use a component of the work, it is your responsibility to determine whether permission is needed
for that re-use and to obtain permission from the copyright owner. Examples of components can
include, but are not limited to, tables, figures, or images.
All queries on rights and licenses should be addressed to the Publishing and Knowledge Division, The
World Bank, 1818 H Street NW, Washington, DC 20433, USA; fax: 202-522-2625; e-mail: pubrights@
worldbank.org.
ISBN (paper): 978-1-4648-0299-7
ISBN (electronic): 978-1-4648-0300-0
DOI: 10.1596/978-1-4648-0299-7
Cover art: © Thomas Sennett / World Bank. Used with permission. Further permission required for reuse.
Cover design: Debra Naylor, Naylor Design, Inc.
Library of Congress Cataloging-in-Publication Data has been requested.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Contents
Acknowledgments ix
About the Authors xi
Abbreviations xv
Chapter 1 Opportunity Knocks: Deepening Our Understanding
of Poverty Reduction 1
Introduction 1
Decomposing Poverty Reduction 2
Contributions of This Volume 5
Decompositions Can Inform Policy Priorities 11
Notes 11
Bibliography 12
Chapter 2 A Simple Approach to Understanding Changes in
Poverty and Inequality 15
Introduction 15
The Size and Redistribution Effects 16
Accounting for the Contribution of Demographics and
Income Components 25
Summary and Conclusions 32
Notes 33
Bibliography 35
Chapter 3 What Accounts for Changes in Poverty over the
Past Decade? 39
Introduction 39
Growth and Poverty Reduction 40
Forces behind Poverty Reduction 44
Results 53
Summary and Conclusions 57
Annex 3A: Data Sources 58
Annex 3B: Complementary Tables 60
Notes 66
Bibliography 66
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7 v
vi Contents
Chapter 4 Counterfactual Decomposition of Changes in Poverty
Outcomes 69
Introduction 69
The Composition and Structural Effects 71
Accounting for Behavior 90
Concluding Summary and Remarks 97
Notes 99
Bibliography 104
Chapter 5 Why Has Labor Income Increased? An In-Depth
Approach to Understanding Poverty Reduction 109
Introduction 109
Modeling Strategy 111
Decomposition Approach 116
Final Remarks 123
Annex 5A: Estimating the Residual Term in
Multinomial Logit 124
Annex 5B: The Cumulative Decomposition Technique 124
Notes 128
Bibliography 128
Chapter 6 Understanding Poverty Reduction in Bangladesh,
Peru, and Thailand 131
Introduction 131
Country Context 132
The Decomposition Approach 141
Decomposition Results 143
Final Remarks 150
Annex 6A: Regression and Simulation Results 152
Notes 170
Bibliography 170
Figures
1.1 Decomposition of Changes in Moderate Poverty, by Income
Level, in Selected Developing Countries, 2000s 3
1.2 Cumulative Contributions to Moderate Poverty Reduction in
Bangladesh, Peru, and Thailand, 2000s 10
2.1 Determinants of Consumption per Capita 27
3.1 Average Real GDP Growth in Selected Developing Countries,
2000s 44
3.2 Contribution of Growth and Redistribution to Poverty Reduction
in Selected Developing Countries, by Poverty Line, 2000s 45
3.3 Change in Age-Dependency Ratio of Selected Developing
Countries, 2000s 46
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Contents vii
3.4 Change in Subsidies and Other Social Transfers in Selected
Developing Countries, 2000s 49
3.5 Change in International Remittances to Selected Developing
Countries, 2000s 50
3.6 Decomposition of Changes in Moderate Poverty, by Level, in
Selected Developing Countries, 2000s 54
5.1 Model of Contributors to Poverty Reduction,
by Stage Sequence 112
6.1 Change in Moderate Poverty Rates in Bangladesh, Peru,
and Thailand, 2000s 133
6.2 GDP in Bangladesh, Peru, and Thailand, 2000–10 134
6.3 Population Growth in Bangladesh, Peru, and Thailand,
2000–10 136
6.4 Growth Incidence Curves of Labor Income in Bangladesh,
Peru, and Thailand, 2000s 138
6.5 Nonlabor Income Growth, by Source, in Bangladesh, Peru, and
Thailand, 2000s 140
6.6 Change in Household Consumption-to-Income Ratio in
Bangladesh, Peru, and Thailand, 2000s 140
Tables
2.1 Shapley Allocations for a Three-Player Game 22
2.2 Application of Barros Methodology to Measure Contributions
of Variables to Change in Poverty 30
2.3 Proposed Methodology to Decompose Change in Poverty
along One Possible Path 31
3.1 Poverty Headcount Rates, by Benchmark, in Selected
Developing Countries, 2000s 42
3.2 Share of Adults per Household, by Poverty Level, Selected
Developing Countries, 2000s 47
3.3 Share of Working Adults per Household, by Poverty Level, in
Selected Developing Countries, 2000s 48
3.4 Share of Transfers in Total Household Income, by Poverty
Level, in Selected Developing Countries, 2000s 51
3.5 Change in Household Consumption-to-Income Ratio in
Selected Developing Countries, 2000s 52
3.6 Contributions to Declines in Moderate Poverty, by Level, in
Selected Developing Countries, 2000s 55
3A.1 Survey Sources of Data for Poverty Reduction Analysis in
Selected Developing Countries, 2000 59
3B.1 Contributions to the Decline in the $1.25-a-Day (PPP) Poverty
Headcount in Selected Developing Countries, 2000s 60
3B.2 Contributions to the Decline in the $2.50-a-Day (PPP) Poverty
Headcount in Selected Developing Countries, 2000s 62
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
viii Contents
3B.3 Contributions to the Decline in the $4.00–$5.00-a-Day (PPP)
Poverty Headcount in Selected Developing Countries, 2000s 64
6.1 Change in Poverty Rates, by Level, in Bangladesh, Peru, and
Thailand, 2000s 133
6.2 Growth and Redistribution Decomposition of Moderate
Poverty Rate Changes in Bangladesh, Peru,
and Thailand, 2000s 135
6.3 Population and Labor Force Characteristics in Bangladesh,
Peru, and Thailand, 2000s 137
6.4 Household Consumption-to-Income Ratio, by Income Decile, in
Bangladesh, Peru, and Thailand, 2000s 141
6.5 Marginal Contributions to Poverty Reduction in Bangladesh,
Peru, and Thailand, 2000s 144
6.6 Contributions to Poverty Reduction by Returns to
Endowments in Bangladesh, Peru, and Thailand, 2000s 145
6.7 Cumulative Contributions to Poverty Reduction in Bangladesh,
Peru, and Thailand, 2000s 146
6A.1 Simulating the Changing Characteristics of Households in
Bangladesh, 2000–10 152
6A.2 Simulating the Changing Characteristics of Households in Peru,
2005–09 153
6A.3 Simulating the Changing Characteristics of Households in
Thailand, 2000–09 154
6A.4 Multinomial Logit on Occupational Choice of Working-Age
Population, by Household Status, in Bangladesh,
2000 and 2010 155
6A.5 Multinomial Logit on Occupational Choice of Working-Age
Population, by Household Status, in Peru, 2004 and 2010 157
6A.6 Multinomial Logit on Occupational Choice of Working-Age
Population, by Household Status, in Thailand,
2000 and 2009 159
6A.7 Earnings Regressions for Nonfarm Working-Age Population in
Bangladesh, 2000 and 2010 161
6A.8 Earnings Regressions for Nonfarm Working-Age Population in
Peru, 2004 and 2010 163
6A.9 Earnings Regressions for Nonfarm Working-Age Population in
Thailand, 2000 and 2009 165
6A.10 Net Revenue Regressions for Farm Households in Bangladesh,
2000 and 2010 167
6A.11 Net Revenue Regressions for Farm Households in Peru,
2004 and 2010 168
6A.12 Net Revenue Regressions for Farm Households in Thailand,
2000 and 2009 169
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Acknowledgments
This book was cofunded by the Poverty Reduction and Economic Management
Network and the Multi-Donor Trust Fund for the Diagnostic Facility on Shared
Growth.
The team was led by Gabriela Inchauste (task team leader and senior econo-
mist, Poverty Reduction and Equity Department [PRMPR]) under the guidance
of Jaime Saavedra-Chanduvi (Director, PRMPR) and Christina Malmberg Calvo
(Acting Director, PRMPR). Members of the team and tasks were as follows:
Chapter 1 Gabriela Inchauste
Jaime Saavedra-Chanduvi
Chapter 2 João Pedro Azevedo
B. Essama-Nssah
Gabriela Inchauste
Sergio Olivieri
Chapter 3 João Pedro Azevedo
Gabriela Inchauste
Trang Van Nguyen
Sergio Olivieri
Hernan Winkler
Chapter 4 B. Essama-Nssah
Chapter 5 Gabriela Inchauste
Sergio Olivieri
Hernan Winkler
Chapter 6 Gabriela Inchauste
Sergio Olivieri
Hernan Winkler
The team is grateful to Mary A. Anderson (consultant, PRMPR) for provid-
ing feedback on organization and content as well as for formatting and editing
the manuscript. The authors acknowledge analytical inputs provided by Reena
Badiani-Magnusson (economist, EASHS); Reno Dewina (consultant, EASPW);
Jenny Lighthart (consultant, EAP); Minh Cong Nguyen (economist, ECSP3);
and Viviane Sanfelice (consultant, LCSPP).
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7 ix
x Acknowledgments
The authors thank colleagues for their insightful suggestions, in particular,
Kathleen Beegle (lead economist, AFRCE); Francisco Ferreira (chief economist,
AFRCE); Samuel Freije (lead economist and sector leader, LCSPR); and Emmanuel
Skoufias (lead economist, LCSPP). The authors are grateful to colleagues from the
World Bank for their comments at various stages in developing the concept note
and the drafts and presentations of the book. The authors thank all of them, in
particular Dean Joliffe (senior economist, DECPI); Ambar Narayan (lead econo-
mist, PRMPR); Luis Servén (senior advisor, DECMG); Erwing Tiongson (senior
economist, LCSPP); and Renos Vakis (lead economist, DECPI).
The research also benefited from presentations of draft papers at the Latin
American and Caribbean Economic Association meetings in Lima, 2012; at the
Network of Inequality and Poverty meetings at Columbia University, 2012; at
Tulane University, 2013; and at World Bank workshops between 2011 and 2013.
The authors gratefully acknowledge the insightful comments and suggestions of
participants at these workshops.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
About the Authors
João Pedro Azevedo is a senior economist at the World Bank, where he currently
works for the Europe and Central Asia region, focusing on Turkey and Central
Asia and leading the region’s Statistical Team. Azevedo has focused much of his
work on helping developing countries to improve their systems for evidence-
based decision making. Before joining the World Bank, he served as the superin-
tendent of monitoring and evaluation at the Secretary of Finance for the State of
Rio de Janeiro, Brazil. He was also a research fellow at the Brazilian Institute of
Applied Economic Research. He is a former chairman of the Latin American
Network on Inequality and Poverty. He holds a PhD in economics from the
University of Newcastle upon Tyne.
B. Essama-Nssah worked for 17 years for the World Bank in Washington, DC,
before he retired as a senior economist in 2011. During his tenure at the World
Bank, he performed economic analyses, prepared policy research and technical
papers, and conducted an annual training course on impact evaluation method-
ologies for staff from the World Bank and client countries. Before joining the
World Bank, he worked for 2 years as a senior research associate on the Food and
Nutrition Program at Cornell University, and for 6 years as head of the Economics
Department and vice dean of the Faculty of Law and Economics of the University
of Yaoundé, Cameroon. He currently works as a consultant focusing on poverty
and growth incidence analysis, program evaluation, and analysis of the social
impact of public policy. He holds a PhD in economics from the University of
Michigan in Ann Arbor.
Gabriela Inchauste is a senior economist in the Poverty Reduction and Equity
Department at the World Bank. Her research interests include poverty, macro-
micro linkages, and the distributional impact of fiscal policy. She has published
in academic volumes and peer-reviewed journals in the areas of fiscal policy, the
informal sector, the impact of crises on the poor, and the role of remittances in
developing countries. Before joining the World Bank, she was a senior economist
at the Inter-American Development Bank and the International Monetary Fund,
where she contributed to operational and analytical activities in numerous
countries on topics including macroeconomic forecasting, public expenditure
policy, poverty and social impact analysis, fiscal and debt sustainability analysis,
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7 xi
xii About the Authors
postdisaster needs assessments, and subsidy reform. Inchauste holds a PhD in
economics from the University of Texas at Austin.
Trang Van Nguyen is a senior economist in the Poverty Reduction and Equity
Department at the World Bank. Previously, she worked in the East Asia and
Pacific, Africa, and Europe and Central Asia Regions of the World Bank, in the
areas of labor, poverty and inequality, education, health, social protection, and
gender and development. Nguyen is one of the lead authors of the World Bank’s
regional flagship report, East Asia Pacific at Work: Employment, Enterprise, and
Well-Being. She is also coauthor of two other World Bank flagship publications:
Toward Gender Equality in East Asia and the Pacific: A Companion to the World
Development Report (2012) on gender equality and development and International
Migration and Development in the East Asia and Pacific Region. She holds a PhD
in economics from the Massachusetts Institute of Technology.
Sergio Olivieri is an economist in the Poverty Reduction and Equity Department
at the World Bank. His main research areas are ex ante analysis of the distributional
impact of macroeconomic shocks, understanding how economic growth affects
poverty, poverty measurement, and multidimensional poverty. He has published
articles in peer-reviewed journals on labor informality, polarization, and multidi-
mensional poverty and has contributed to edited volumes on inequality, poverty,
and social cohesion. Previously, he was an assistant professor of labor economics
in the Department of Economics of University of La Plata, Argentina, and as a
researcher in the university’s Center of Distributional, Labor and Social Studies.
Jaime Saavedra-Chanduvi is Minister of Education in Peru. He was previously
the director of the Poverty Reduction and Equity Department at the World
Bank. His major areas of interest include education policy, poverty reduction,
inequality, labor markets, and social policies. He was executive director and prin-
cipal researcher at Grupo de Análisis para el Desarollo, a nonpartisan think tank
based in Lima, and a principal advisor to the Ministry of Labor and Social
Promotion in Peru. He has been a consultant and researcher for the World Bank,
the Economic Commission for Latin America and the Caribbean, the Inter-
American Development Bank, and the International Labour Organization. He
has been president of the executive committee of the Network on Inequity and
Poverty of the Inter-American Development Bank–World Bank–Latin American
and Caribbean Economic Association as well as a board member of Latin
American and Caribbean Economic Association, the Nutrition Research Institute,
and the National Council of Labor in Peru. He has held teaching positions at
Pontificia Universidad Católica del Peru and Universidad del Pacífico in Peru, and
has been a visiting researcher at the University of Toronto. Saavedra-Chanduvi
holds a PhD in economics from Columbia University.
Hernan Winkler is an economist in the Office of the Chief Economist for Europe
and Central Asia at the World Bank. He specializes in applied microeconomics,
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
About the Authors xiii
with a particular focus on labor market issues and the sources and consequences
of inequality and poverty. He was part of the core team of the World Bank’s
regional flagship report, Diversified Development: Making the Most of Natural
Resources in Eurasia (2014), and also on the core team of the next regional flag-
ship report on the aging populations of Europe and Central Asia. Before joining
the World Bank, Winkler was a researcher at the Center of Distributional, Labor
and Social Studies at the University of La Plata, Argentina. He holds a PhD in
economics from the University of California at Los Angeles.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Abbreviations
ATET average treatment effect on the treated
ENAHO Encuesta Nacional de Hogares (Peru National Institute of Statistics
and Information)
FAO Food and Agriculture Organization of the United Nations
GDP gross domestic product
GIC growth incidence curve
HIES Household Income and Expenditure Survey (Bangladesh Bureau
of Statistics)
PPP purchasing power parity
RIGA Rural Income-Generating Activities (datasets compiled jointly by
the Food and Agriculture Organization of the United Nations and
the World Bank)
SEDLAC Socio-Economic Database for Latin America and the Caribbean
SES Household Socio-Economic Survey (Thailand National Statistical
Office)
Currency is in U.S. dollars unless specified otherwise.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7 xv
Chapter 1
Opportunity Knocks: Deepening
Our Understanding of Poverty
Reduction
Introduction
A 2013 cover of The Economist virtually declared victory: “Towards the End of
Poverty” (Economist 2013). Given international benchmarks, a battle has indeed
been won—to cut in half the share of the world’s population living in extreme
poverty. In fact, the achievement of this 2015 Millennium Development Goal
was met with time to spare.1 By 2013, the percentage of developing-country
populations living under the direst conditions—measured by the extreme pov-
erty line, typically set at $1.25 a day2—decreased from 43 percent in 1990
(1.9 billion people) to 21 percent by 2010 (1.2 billion people). Clearly there is
still a long way to go, with 1.2 billion people still struggling to get enough to eat.
But what can we learn from the recent success?
For example, how much do we truly understand about the specific drivers of
such momentous changes in poverty—be they reductions or increases, in low-
income or middle-income countries, at extreme or moderate poverty levels? To
be sure, 21st-century poverty reduction is largely a growth story, and one that is
indisputable by either national or international poverty lines.3 Even knowing that
growth correlates strongly with poverty reduction (Ravallion and Chen 2007),
economists and policy makers want to “drill down” into the data to answer many
more questions:
• What role have demographic changes and lower dependency ratios had on the
reduction in poverty?
• What was the role of higher employment and higher real earnings? Did higher
labor productivity lead to higher real earnings? Or did earnings growth result
either from improved human capital (in education, training, or experience) or
from changes in the returns to those characteristics?
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7 1
2 Opportunity Knocks: Deepening Our Understanding of Poverty Reduction
• To what extent did changes in the sectoral composition of employment lead to
higher productivity and higher earnings? For instance, how did movements
from agriculture to the services or industry sectors contribute to poverty
reduction?
• What was the role of transfers from the government, including the new gen-
eration of social safety net programs that have begun to proliferate?
• What was the role of private transfers in the form of international
remittances?
International bodies and financial institutions are setting ambitious new pov-
erty reduction targets for 2030. Similarly, many developing countries have clear
development goals and national plans to reduce poverty and promote inclusive
growth. The methods and results presented in this book are envisioned to con-
tribute to the evidence base that governments can use in setting the policy
agenda.
Decomposing Poverty Reduction
The links between economic growth, income redistribution, and poverty reduc-
tion have long interested economists. Growth, at the end of the day, is a means
to an end. Bolstered by decades of previous work, recent methods can be used to
decompose the contributions to poverty reduction.
Decomposition methods originated in the seminal papers of Oaxaca (1973)
and Blinder (1973), which aimed to decompose changes in wages over time.
Since then, the increase in wage inequality observed in the United States and
several other countries since the late 1970s has led to the development of new
methods, including those introduced by Juhn, Murphy, and Pierce (1993) and
Bourguignon, Ferreira, and Lustig (2005). Although these methods were focused
on better understanding distributional changes, they can also be used to better
understand changes in poverty.
In particular, we focus on two questions: What was the main contributor to
poverty reduction? What was behind the poverty reduction due to labor
income growth in particular? The first question can be addressed with a
simple accounting approach that quantifies the contributions to poverty
reduction on account of changes in demographics, changes in employment and
labor income, and changes in nonlabor income (including remittances, public
transfers, and other private transfers). The second question requires a more
complex approach to further distinguish between distributional changes
because of changes in endowments or returns to those endowments; changes
in occupational choice; and changes in the geographical, age, and gender struc-
ture of the population, along with the nonlabor dimensions mentioned above.
Such an approach can, for instance, help distinguish whether improvements in
human capital (via a more educated work force) have been accompanied
by increases in the returns to education (and potential improvements in
productivity).
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Opportunity Knocks: Deepening Our Understanding of Poverty Reduction 3
What Was the Main Contributor to Poverty Reduction?
By covering and expanding upon existing decomposition methods, chapters
2 and 3 describe and apply the first of these methods to identify the forces that
account for substantial declines4 in moderate poverty over the past decade5 in a
selected set of 21 developing nations. The main results are as follows:
• Labor income growth (comprising growth of employment and earnings) clearly
contributed the most to poverty reduction in the studied countries (figure 1.1).
Being the main asset of the poor, increased labor income accounts for
more than half of the poverty reduction in 12 of the 21 countries with
substantial poverty reduction. In another 6 countries, it accounted for more
than 40 percent of the reduction in poverty. Moreover, in most cases, it was
the growth in income per worker that contributed the most to poverty reduc-
tion, rather than an increase in employment (measured by the share of work-
ing adults).
• Demographic changes (primarily increases in the share of adults per household)
led to declining age-dependency rates, which in turn can lead to increases in
per capita income and consumption. This effect occurred in most of the
Figure 1.1 Decomposition of Changes in Moderate Poverty, by Income Level, in Selected Developing
Countries, 2000s
140
Contributions to poverty reduction (%)
120
100
80
60
40
20
0
–20
–40
009
0
–10
8
–10
0
11
05
09
–10
9
–09
0
003
10
9
09
09
0
9
0
201
0–0
2–1
1–0
4–1
1–0
4–1
6–0
7–1
07–
–20
02–
03–
01–
00–
9–2
000
000
001
000
6–2
99–
200
200
, 200
, 200
200
, 200
200
, 200
a, 20
998
a, 20
r, 20
il, 20
e, 20
199
sh, 2
na, 2
va, 2
d, 2
9
y, 19
l, 19
,
bia,
,
nes,
Rica
ania
nam
ama
na, 1
Peru
ia
ado
goli
ank
ilan
Braz
Chil
,
lade
bod
uras
do
enti
m
a
gua
ippi
Rom
Nep
Pan
ta
Viet
Ecu
Sri L
Mon
Gha
Tha
Mol
Colo
Cam
Arg
g
d
Cos
Para
Phil
Ban
Hon
≤ $1.25 a day $4.00 a day $5.00 a day
Nonlabor income Employment + earnings Share of working-age family members Consumption-to-income ratio
Sources: Data from SEDLAC, various years; FAO n.d.; and national household surveys.
Note: “Nonlabor income” refers to public and private transfers (including remittances), pensions, capital, and other nonlabor income.
“Employment + earnings” refers to the combined change in employment and earnings per working-age adult (aged 15–64 years). The
“consumption-to-income ratio” represents the ratio of measured consumption to measured income. Changes in this ratio capture changes in
savings patterns of households as well as measurement errors in household consumption and income. Consumption-based measures of poverty
are used in Bangladesh, Ghana, Moldova, Nepal, Peru, Romania, and Thailand. The remaining countries use income-based measures of poverty.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
4 Opportunity Knocks: Deepening Our Understanding of Poverty Reduction
countries considered here over the past decade and was especially important
in Costa Rica, Paraguay, and the Philippines. However, in most of the countries
studied, the magnitude of this effect was small relative to the effect of labor
income growth.
• Nonlabor income (such as government spending for subsidies or public
transfers)6 grew, as a share of gross domestic product (GDP), more than sixfold
in Ghana and by more than 60 percent in Bangladesh, Moldova, Mongolia, and
Romania. Remittances also grew strongly, especially in Honduras (to an aver-
age of 17.8 percent of GDP in 2009) and Nepal (to an average of 12 percent
of GDP in 2003). Although transfers contributed to poverty reduction, they
played a relatively smaller role in explaining declines in moderate poverty for
most countries in the sample. Three notable exceptions were Moldova,
Mongolia, and Romania. Moldova’s poor benefited mostly from an increase in
international remittances, whereas in Romania, the increase was related to
both transfers and capital income. However, public transfers clearly play an
increasingly important role in reducing extreme poverty. In other words,
increases in transfers to the extremely poor were critical in reducing the sever-
ity of poverty.
Why Did Labor Income Grow?
Although the results clearly point to growth in income per worker as the main
contributor to poverty reduction in most of the countries studied, the method is
unable to discern why labor income grew. In many developing countries, poverty
reduction has coincided with the labor force’s increasing education and health—
as well as, in some cases, more equitable distribution of land or other productive
assets. Did earnings increase because of changes in the endowments of the popu-
lation (such as higher educational levels or increases in other productive assets)?
Or did marketplace premiums (that is, the returns to endowments) rise for
workers with those endowments?
The rest of the book explores alternative methods to better understand the
hapters
roots of these changes. After a review of the literature, the book’s final two c
describe and implement a second method based on Bourguignon, Ferreira, and
Lustig (2005) and impose an underlying labor model and greater structure to
understand why earnings increased. This decomposition approach is applied to
Bangladesh, Peru, and Thailand. The results show that changes in individual char-
acteristics (such as education, work experience, and region of residence) were
influential, but that, overall, labor income grew mainly because of higher returns
to endowments—signaling an increase in the marginal value of work as a result of
increases in productivity or in the relative price of labor.
In Bangladesh and Peru, this increase in the marginal value of work was not
driven by higher returns to education, but rather by higher returns to labor with
low levels of education. In the case of Bangladesh, the increase in the marginal
value of work seems to have been associated with higher food prices in the farm
sector, while in Peru it was associated with higher premiums in the service sector,
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Opportunity Knocks: Deepening Our Understanding of Poverty Reduction 5
potentially resulting from improving productivity. Thailand, in contrast, demon-
strated that greater specialization and higher returns to education can boost the
marginal value of work, potentially through productivity increases.
The past decade affords us a fantastic opportunity to study the most signifi-
cant factors that worked in favor of the poor. The decomposition methods
reviewed in this volume provide a window to quantify the contributions to
changes in poverty at a finer level of detail than ever before. The results point to
the centrality of jobs in reducing poverty—particularly through increases in labor
productivity—that will lead to sustained income growth for the poor. It is the
authors’ hope that the decompositions developed in this volume can expand the
evidence base to establish the necessary conditions for future poverty reduction
and therefore guide policy to enable these conditions to take hold.
Contributions of This Volume
The decomposition methods developed in this volume make some distinct
contributions, further described in the chapter synopses below and summarized
as follows:
• Focus on consumption as a measure of welfare. Up until now, decomposition
methods often focused on income-based measures of welfare. These methods
are adapted in this volume to focus on consumption, given that most develop-
ing countries use a consumption aggregate to measure poverty.
• Address path dependence by calculating the Shapley-Shorrocks estimate of each
component. Like most micro-decomposition approaches in the literature, the
methods proposed in this volume suffer from path dependence. In other
words, the order in which the cumulative effects are calculated matters. The
best-known remedy for path dependence is to calculate the decomposition
across all possible paths and then take the average between them. Following
the algorithm proposed by Azevedo, Nguyen, and Sanfelice (2012a, 2012b),
we calculate the Shapley-Shorrocks (Shapley 1953; Shorrocks 1999) estimate
of each component.
• Apply to a wide set of countries experiencing poverty reduction around the world.
This volume presents the most comprehensive application of these techniques
across countries that have experienced substantial declines in poverty over the
past decade. Previous efforts typically included only a limited number of
countries. The ease of use of this technique now allows for replication in a
variety of contexts and can serve to decompose changes in poverty, inequality,
or any other distributional change.
• Propose a structural model that allows for household and individual decision
making. Previous approaches typically modeled only individual decision
making. The proposed approach models farm income at the household level,
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
6 Opportunity Knocks: Deepening Our Understanding of Poverty Reduction
including earnings of farm household members who have secondary occupa-
tions, thus recognizing not only that farm households typically make labor
decisions as a unit but also that these households can be highly diversified.
The chapter summaries below offer only snapshots of the issues and tech-
niques involved in understanding changes in poverty, allowing the full chapters
to speak for themselves.
Chapter 2: A Simple Approach to Understanding Changes in Poverty
and Inequality
Chapter 2 provides a unifying framework and a theoretical foundation for the
decomposition methods commonly used in the literature. The chapter begins
with the simplest decompositions—in which poverty changes can be decom-
posed into the growth effect and the redistribution effect (also known as the size
and redistribution effects)—often referred to as the Datt and Ravallion (1992)
decomposition. The chapter discusses the underpinnings of this decomposition
and notes that in its original form, the Datt-Ravallion method includes a residual
term. Moreover, the value of any effect (size or redistribution) depends on the
chosen period of reference. To solve this problem, the standard practice is to
compute the decomposition in both ways and then take the average.
As it turns out, this average Datt-Ravallion decomposition is precisely the
procedure proposed by Shorrocks (1999) based on the Shapley (1953) value.
Indeed, the Shapely allocation rule, based in game theory, provides a framework
that can be extended to more complex decompositions, including those that
decompose changes in poverty on account of changes within and between groups.
Although these decompositions are interesting, analysts often want to go
beyond these summary statistics. We describe an approach that quantifies the
contributions to poverty reduction on account of changes in demographics,
employment, earnings, and public and private transfers based on a simple
accounting identity. The method is adapted from Barros et al. (2006) and has the
distinct advantage of being easy to apply and easy to understand. In contrast with
the methods that use aggregate measures, such as “growth” and “redistribution”
effects, this approach generates entire counterfactual distributions, allowing us to
quantify the contributions to poverty reduction stemming from changes in
demographics, employment, earnings, and public and private transfers by chang-
ing each one of these elements at a time.7 To avoid path dependence, the pro-
posed method calculates the Shapley-Shorrocks estimates of each component to
find the contribution of each component.
Chapter 3: What Accounts for Changes in Poverty over the Past Decade?
Chapter 3 implements both the Datt-Ravallion decomposition and the method
proposed in chapter 2 for a sample of 21 countries where moderate poverty had
declined substantially between 2000 and 2010. The chapter quantifies, by
generating entire counterfactual distributions, contributions to poverty reduction
from changes in labor income, nonlabor income, and demographic characteristics.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Opportunity Knocks: Deepening Our Understanding of Poverty Reduction 7
Because most countries measure welfare through household expenditures or
consumption (as opposed to income), this chapter modifies the Barros
et al. (2006) methodology in three important respects: It decomposes consump-
tion-based measures of poverty, computes a cumulative counterfactual
distribution by adding one variable at a time, and it calculates the Shapley-
Shorrocks estimates of each component.
The decompositions across the 21 countries resulted in the following general
findings:
• Labor income growth, by far, was the main driver of poverty reduction.
Moreover, an increase in workers’ earnings was relatively more important in
reducing poverty than the increases in the share of employed adults per
household.
• Demographic changes—increased percentages of working-age adults per house-
hold (and thus declining dependency ratios)—also mattered in poverty reduc-
tion, particularly in seven of the countries. In general, however, the magnitude
of this effect is small relative to labor income growth.
• Nonlabor income increases, such as public and private transfers, were important,
but, for most countries in the sample, played the smallest role in explaining
declines in moderate poverty. However, they were far more important in
reducing extreme poverty. From decomposition of the extreme-poverty head-
count, poverty gap, and poverty severity, transfers and pensions contributed a
relatively higher share to poverty reduction than labor income growth.8 These
results point to the crucial role that social protection systems play in reaching
the ultrapoor.
Although these methods are useful to distinguish the main contributors to
poverty reduction, they cannot explain why workers’ earnings increased. To
resolve that issue, the subsequent chapters consider alternative decomposition
techniques that impose an underlying labor model and greater structure.
Chapter 4: Counterfactual Decomposition of Changes in Poverty Outcomes
Chapter 4 presents a literature review of commonly used micro-decomposition
methods to identify key drivers of changes in poverty. The chapter begins with
the standard Oaxaca-Blinder decomposition, which aimed to understand changes
in between-group differences in average wages, and reviews how this technique
has since been extended to the analysis of variation in other distributional statis-
poverty, and inequality measures.
tics, such as quantiles,
The chapter first reviews methods to identify and estimate the composition
(or endowment) effect and structural (or price) effect associated with variation
in a general distributional statistic. The methods underlying this decomposition
are considered statistical to the extent that they do not involve causal models of
the observed outcomes. In this context, nonparametric estimation has the advan-
tage of not imposing a functional form on the relationship between the outcome
and its determinants. Because these statistical methods cannot shed light on the
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
8 Opportunity Knocks: Deepening Our Understanding of Poverty Reduction
causal mechanism driving the outcome, their ability to inform policymaking is
somewhat limited.
Therefore, the review also considers methods that account for behavioral
responses to changes in the socioeconomic environment. These methods rely on
the specification and estimation of a microeconometric model based on some
theory of individual (or household) behavior and social interaction. These
methods go a step further toward identifying factors associated with structural
elements underpinning the observed changes in poverty outcomes. These struc-
tural methods—combining economics and statistics—can be predictive. In this
context, chapter 5 proposes one such structural approach.
Finally, the chapter discusses the analogy between these decomposition meth-
ods and treatment effect analysis. The chapter also notes that these decomposi-
tion methods can be used in the study of the distributional and poverty impacts
of an assigned intervention.
Chapter 5: Why Has Labor Income Increased? An In-Depth Approach to
Understanding Poverty Reduction
The literature review presented in chapter 4 concludes that structural decompo-
sition methods enable an analyst to study the distributional and poverty impacts
of an assigned intervention by taking into account how agents respond to changes
in their socioeconomic environments. This chapter presents a structure for mod-
eling distributional changes over time that allows us to account for the contribu-
tions to poverty reduction stemming from both changes in endowments and
changes in the returns to those endowments.
The proposed approach9 estimates an educational choice model, a sectoral
choice model, an activity choice model, and individual and household earnings
equations. Once all of these models are estimated for individuals and households
in each time period, the estimated coefficients from one year can be replaced
with the estimates from another year to simulate the impact of changes in each
element at a time. A series of counterfactual income distributions can then be
constructed from which a counterfactual poverty measure can be estimated. This
new poverty measure is then compared with the observed outcome while hold-
ing everything else constant. By changing one e lement at a time, these decompo-
sitions allow us to account for the observed changes in poverty.
In addition, we present a method to estimate these counterfactuals cumula-
tively, thereby accounting for the impact of concurrent changes. Although the
method presented here draws heavily from Bourguignon and Ferreira (2005) and
Bourguignon, Ferreira, and Leite (2008), it also offers some innovations:
• It assumes that welfare is measured using a consumption aggregate and
accounts for the contribution of changes in the consumption-to-income ratio.
• It models farm household income at the household level and models the earn-
ings of individuals in those farm households who have a secondary occupation,
thus recognizing not only that farm households typically make labor decisions
as a unit but also that these households can be highly diversified.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Opportunity Knocks: Deepening Our Understanding of Poverty Reduction 9
• It ensures that changes in the composition of activities, sectors, and education
of the work force are consistent with the counterfactual choices.
Chapter 6: Understanding Poverty Reduction in Bangladesh, Peru, and
Thailand
The method proposed in chapter 5 is applied in the volume’s concluding chapter
to understand the drastic reduction in poverty in three countries between 2000
and 2010: Bangladesh, Peru, and Thailand. While diverse in their levels of income
and urbanization, these three economies each reduced moderate national
poverty headcount rates by more than 14 percentage points.
In each case, GDP growth remained high, well exceeding 4 percent between
2002 and 2008. Peru and Thailand had sharp deceleration in 2009 because of the
global financial crisis but rebounded quickly in 2010. Bangladesh got through the
crisis virtually unscathed given the country’s weaker integration into global finan-
cial markets. In addition to healthy economic growth, the countries had further
similarities in potential poverty-reduction factors:
• Increases in employment, public social transfers, and remittances
• Changes in the populations’ occupational structure, with workers moving
away from farm and daily work and toward more salaried employment and to
jobs of relatively high productivity
• Workers’ sharp shifts away from agriculture and toward the higher-productivity
manufacturing and service sectors
• Improvements in the work force’s educational composition for several dimen-
sions over the past 10 years, a result of each country’s higher investment in
education in previous decades.
Once the decompositions were performed, the main result, consistent with
the findings mentioned above, was that the largest contributions to poverty
reduction in all three countries were labor market–related factors, including
changes in the sectoral, occupational, and educational structure of the labor
force, as well as changes in the returns to individual and household characteris-
tics. These changes cumulatively reduced moderate poverty levels in Bangladesh
by 80 percent, in Peru by 69 percent, and in Thailand by 62 percent (figure 1.2).
Furthermore, within these results, we can finally explain why labor income
increased. Through the micro-decomposition methods employed here, it was
ascertained that labor income grew mainly because of higher returns to human
capital endowments. In other words, the marginal value of work increased, and
that’s what made such a difference. This could signal increases in productivity, a
higher relative price of labor, or both.
In Bangladesh and Peru, this increase in the marginal value of work was not
driven by higher returns to education, but rather by higher returns to workers
with low levels of education. To the extent that these higher returns to low-
educated workers were accompanied by changes in the sectoral and occupational
structure of the work force, these premiums may be reflecting improvements in
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
10 Opportunity Knocks: Deepening Our Understanding of Poverty Reduction
Figure 1.2 Cumulative Contributions to Moderate Poverty Reduction in Bangladesh, Peru,
and Thailand, 2000s
140
120
Contribution to moderate poverty reduction (%)
100
80
60
40
20
0
–20
–40
Bangladesh Peru Thailand
2000–10 2004–10 2000–09
Nonlabor income Sector Demographics
Returns: farm Occupation Residuals
Returns: nonfarm Education Others
Sources: Calculations derived from household survey data from Bangladesh (HIES 2000–10), Peru (ENAHO 2004–10), and
Thailand (SES 2000–09).
Note: “Moderate poverty” refers to each country’s moderate poverty line. “Sector” refers to the sectoral composition of the
workforce, including agriculture, industry, services. “Occupation” refers to the activity type of the workforce including daily workers,
wage workers, self-employed, and unemployed. “Demographics” refers to changes in the age, gender, and regional composition
of the population. “Others” includes changes in the consumption-to-income ratio and unexplained portion. HIES = Household
Income and Expenditure Survey (Bangladesh Bureau of Statistics); ENAHO = Encuesta Nacional de Hogares (Peru National
Institute of Statistics and Information); SES = Household Socio-Economic Survey (Thailand National Statistical Office).
the productivity of these unskilled workers. However, these higher returns might
also simply reflect higher food prices, which have led to relative increases in the
earnings of agricultural workers.
In contrast, in Thailand, poverty decreased because of both improvements in
the education of the work force and higher returns to education. In this case, the
marginal value of work increased through higher labor productivity increases.
Beyond increases in the returns to labor, all three countries showed reductions
in poverty as a result of
• A falling earnings penalty for living outside of the capital city
• A shift in sectoral choices away from agriculture and toward services—reducing
poverty only slightly in Bangladesh and Peru and a bit more in Thailand
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Opportunity Knocks: Deepening Our Understanding of Poverty Reduction 11
• A shift in occupational choices among nonfarm workers away from daily-wage
and self-employed work and toward salaried jobs, all of which contributed to
poverty reduction, particularly in Peru.
Beyond the effects of labor income growth, the decomposition method
showed that a greater share of working-age adults helped to reduce poverty,
particularly in Bangladesh.
Finally, although most poverty reduction was the result of labor income
growth, it is important to recognize that nonlabor income in the form of transfers
did play a role: International remittances accounted for 11 and 17 percent of
poverty reduction in Bangladesh and Thailand, respectively. Public transfers
accounted for about 9 percent of Peru’s poverty reduction, while Thailand’s
generous new pension scheme, combined with various private and other trans-
fers, accounted for more than 40 percent of its poverty reduction.
Decompositions Can Inform Policy Priorities
International bodies and financial institutions are setting ambitious new poverty
reduction targets for 2030. Similarly, many developing countries have clear
development goals and national plans to reduce poverty and promote inclusive
growth. Some are facing clear challenges to improve well-being through employ-
ment; some are facing the question of how to best design the tax-transfer system;
others are undergoing tremendous demographic changes.
A first step in putting together a policy agenda is to frame the issues with
respect to achieving these poverty reduction goals. As such, understanding
poverty trends and the factors underlying these trends is critical. In particular, a
better understanding of the roles of growth vis-à-vis distributional changes and
the role of labor versus nonlabor income in explaining poverty reduction in the
past 5–10 years can usefully shed light on what has worked to date and where
there is room for improvement. The methods presented in this book can be
applied to analyze changes in poverty, inequality, or any other distributional
change over time. Both the methods and the emerging results presented here are
envisioned to contribute to the evidence base that governments can use in setting
the policy agenda.
Notes
1. For more information about the United Nations’ Millennium Development Goals
(MDGs), see the website http://www.un.org/millenniumgoals/.
2. The lower-bound poverty line, being initially set at a 1990 baseline rate of $1 a day
(at 1993 purchasing power), was increased after economists gathered a new set of
global poverty measures that spanned 116 countries (Ravallion, Chen, and Sangraula
2008). The new international poverty line (of $1.25 a day in 2005 prices) was
calculated from “the average of the 15 poorest countries’ own poverty lines in 2005
prices, adjusted for differences in purchasing power” (Economist 2013).
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
12 Opportunity Knocks: Deepening Our Understanding of Poverty Reduction
3. For more details about global changes in poverty and how poverty is measured, see
Chen and Ravallion 2008 and World Bank 2012.
4. A “substantial” decline in poverty is defined as an average reduction of 1 percentage
point per year over the data period.
5. “Moderate” poverty is a country-specific poverty line referring to the international
poverty line that is closest to the country’s moderate poverty rate (in some cases $1.25
a day, in others $2.50, and in still others $4–$5 per day.
6. Note that this exercise does not take into account the payoff of increased access to
many public services that are not part of household income, nor does it account for
the poverty impact of improvements in the quality of public services.
7. Counterfactual distributions are obtained by changing one determining factor at a
time while holding all the other factors fixed (a straight application of ceteris paribus
variation).
8. The poverty headcount, poverty gap, and severity of poverty refer to the Foster, Greer,
and Thorbecke (1984) measures. The poverty headcount index is the proportion of
the population that is counted as poor because their consumption or income is below
a certain poverty line. The poverty gap adds up the extent to which individuals on
average fall below the poverty line and expresses it as a percentage of the poverty line.
The severity of poverty is calculated as the poverty gap index squared, which implic-
itly puts more weight on observations that fall well below the poverty line.
9. Our model takes into account the Roy (1951) model of choice and consequences
(which stems from the optimization principle and applies to discrete choice prob-
lems), using it to model individuals’ educational levels, sectors, and activity choices.
We then adapt the Bourguignon, Ferreira, and Leite (2008) methodology to distin-
guish between distributional changes on account of (a) changes in endowments and
the returns to those endowments; (b) changes in occupational and sectoral choice; and
(c) changes in geographical, age, and gender structure of the population, along with
the nonlabor dimensions.
Bibliography
Azevedo, Joao Pedro, Minh Cong Nguyen, and Viviane Sanfelice. 2012a. “Adecomp:
Stata Module to Estimate Shapley Decomposition by Components of a Welfare
Measure.” Statistical Software Components S457562, Boston College, Department
of Economics.
———. 2012b. “Shapley Decomposition by Components of a Welfare Aggregate.” Policy
Research Working Paper 6414, World Bank, Washington, DC.
Barros, Ricardo Paes de, Mirela de Carvalho, Samuel Franco, and Rosane Mendonça. 2006.
“Uma Análise das Principais Causas da Queda Recente na Desigualdade de Renda
Brasileira.” Revista Econômica 8 (1): 117–47.
Blinder, Alan S. 1973. “Wage Discrimination: Reduced Form and Structural Estimates.”
Journal of Human Resources 8 (4): 436–55.
Bourguignon, François, and Francisco H. G. Ferreira. 2005. “Decomposing Changes in
the Distribution of Household Incomes: Methodological Aspects.” In The
Microeconomics of Income Distribution Dynamics in East Asia and Latin America,
edited by François Bourguignon, Francisco Ferreira, and Nora Lustig, 17–46.
Washington, DC: World Bank.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Opportunity Knocks: Deepening Our Understanding of Poverty Reduction 13
Bourguignon, François, Francisco H. G. Ferreira, and Phillippe G. Leite. 2008. “Beyond
Oaxaca-Blinder: Accounting for Differences in Household Income Distributions.”
Journal of Economic Inequality 6 (2): 117–48.
Bourguignon, François, Francisco H. G. Ferreira, and Nora Lustig, eds. 2005. The
Microeconomics of Income Distribution Dynamics in East Asia and Latin America.
Washington, DC: World Bank.
Chen, Shaohua, and Martin Ravallion. 2008. “The Developing World Is Poorer Than We
Thought, But No Less Successful in the Fight against Poverty.” Policy Research
Working Paper 4703, World Bank, Washington, DC.
Datt, Gaurav, and Martin Ravallion. 1992. “Growth and Redistribution Components of
Changes in Poverty Measures: A Decomposition with Applications to Brazil and India
in the 1980s.” Journal of Development Economics 38 (2): 275–96.
Economist. 2013. “Not Always with Us.” June 1–7.
FAO (Food and Agriculture Organization of the United Nations). n.d. Rural Income
Generating Activities (RIGA). Database, FAO, Rome; World Bank and American
University, Washington, DC. http://www.fao.org/economic/riga/riga-database/en/.
Ferreira, Francisco. 2010. “Distributions in Motion: Economic Growth, Inequality and
Poverty Dynamics.” Policy Research Working Paper 5424, World Bank, Washington, DC.
Foster, James, Joel Greer, and Erik Thorbecke. 1984. “A Class of Decomposable Poverty
Measures.” Econometrica 52 (3): 761–65.
Juhn, Chinhui, Kevin M. Murphy, and Brooks Pierce. 1993. “Wage Inequality and the Rise
of Returns to Skill.” Journal of Political Economy 101 (3): 410–42.
Oaxaca, Ronald L. 1973. “Male-Female Wage Differentials in Urban Labor Markets.”
International Economic Review 14 (3): 693–709.
Ravallion, Martin, and Shaohua Chen. 2007. “China’s (Uneven) Progress against Poverty.”
Journal of Development Economics 82 (1): 1–42.
Ravallion, Martin, Shaohua Chen, and Prem Sangraula. 2008. “A Dollar a Day Revisited.”
Policy Research Working Paper 4620, World Bank, Washington, DC.
Roy, Arthur D. 1951. “Some Thoughts on the Distribution of Earnings.” Oxford Economic
Papers 3 (2): 135–46.
SEDLAC (Socio-Economic Database for Latin America and the Caribbean). Various
years. Statistical database produced by CEDLAS (Center for Distributional, Labor
and Social Studies) of the Universidad Nacional de la Plata, Buenos Aires; and World
Bank, Washington, DC. http://sedlac.econo.unlp.edu.ar/eng/.
Shapley, Lloyd S. 1953. “A Value for n-Person Games.” In Contributions to the Theory of
Games, Vol. 2, edited by H. W. Kuhn and A. W. Tucker, 307–17. Princeton,
NJ: Princeton University Press.
Shorrocks, Anthony F. 1999. “Decomposition Procedures for Distributional Analysis:
A Unified Framework Based on Shapley Value.” Paper for the University of Essex and
Institute for Fiscal Studies.
World Bank. 2012. “An Update to the World Bank’s Estimates of Consumption Poverty in
the Developing World.” Briefing note prepared by Shaohua Chen and Martin
Ravallion, Development Research Group, World Bank, Washington, DC.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
CHAPTER 2
A Simple Approach to Understanding
Changes in Poverty and Inequality
Introduction
The link between economic growth and poverty reduction has long been of
interest to economists. The literature has proposed counterfactual decomposi-
tion methods to identify the relative contributions of different factors to varia-
tions in overall poverty.1 These decompositions include the Datt and Ravallion
(1992) method, which decomposes observed changes in poverty into a distri-
bution-neutral growth effect and a redistribution effect. Kolenikov and
Shorrocks (2005) decompose changes in poverty into growth, distribution,
and price components, while Ravallion and Huppi (1991) offer a way of
decomposing changes in poverty over time into intrasectoral effects and
population shifts.
However, the usefulness of these decomposition methods is severely
limited in policy making because they explain changes in poverty on the basis
of changes only in summary statistics that are hard to target with policy
instruments. For instance, it is hard to see what role demographics played in
reducing poverty—or what the roles of employment and labor income
were—relative to the roles of remittances and public transfers. Unfortunately,
methods such as Datt-Ravallion (1992) are unable to make these explicit
links because growth, inequality, and poverty measures are actually just three
different aggregations of information about individual income dynamics.
Moreover, they are jointly determined, such that cross-country estimates are
unlikely to shed much light on the fundamental factors underlying distribu-
tional change (Ferreira 2012).
In addition to methods that focus on aggregate summary statistics, this chapter
reviews a simple method to generate entire counterfactual distributions, allowing
us to quantify the contributions to poverty reduction stemming from changes
in demographics, changes in the share of employed adults, changes in labor
income, and changes in nonlabor income, such as public transfers and remittances.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7 15
16 A Simple Approach to Understanding Changes in Poverty and Inequality
Although it follows the approach first discussed by Barros et al. (2006), the pro-
posed method is new in three important respects:
• It extends the approach to cases where poverty is measured by consumption,
as opposed to income.
• It calculates the cumulative effect of all of these changes and ensures that the
sum of the components is equal to the total change in poverty.
• To avoid path dependence, every possible path is considered, and the average
contribution of each factor is computed over all possible paths, leading to what
is known as Shapley-Shorrocks estimates of the contribution of each compo-
nent (Shapley 1953; Shorrocks [1999] 2013).
It is important to explain at the outset that these decompositions are essen-
tially accounting exercises and do not allow for the identification of causal effects.
For example, increases in cash transfers or noncontributory pensions may in some
circumstances deter participation in the labor market, thus affecting labor
incomes. Similarly, increases in labor income can make some households ineligible
for transfer programs. For those reasons, we caution against interpreting changes
in labor income (or, for that matter, changes in pensions or transfers) as “causing”
changes in the poverty rate. Still, the decompositions are useful in identifying
empirical regularities and, as accounting tools, can help focus attention on factors
that are quantitatively more important in describing distributional changes.
The rest of the chapter is organized as follows: The next section provides
background on decomposition methods used to account for variation in poverty
outcomes in terms of changes in the mean and in relative inequality of the under-
lying outcome distribution. In particular, we review the Datt and Ravallion
(1992) and Shapley (1953) decomposition methods. This is followed by a
decomposition methodology that aims to quantify the contributions of labor
income, transfers, and demographic effects to the observed changes in poverty,
based on the Shapley approach. The last section reviews the contributions and
limitations of the decomposition methods discussed in this chapter.
The Size and Redistribution Effects
This section first frames the decomposition problem by providing a unifying
framework and a theoretical foundation for the decomposition methods com-
monly used in the literature. In particular, we note that changes in the distribu-
tion can be characterized in terms of the level of growth and the degree of
relative inequality. The most common method stemming from this approach is
the Datt and Ravallion (1992) decomposition, to which we turn first. However,
as described in detail below, this method faces at least two important method-
ological issues: whether to use the initial or end period as the reference period,
and how to interpret the residual change in poverty that is not accounted for.
Indeed, all decomposition methods seek to identify the contributions of fac-
tors influencing the distribution, as measured by a particular statistic, such as
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
A Simple Approach to Understanding Changes in Poverty and Inequality 17
poverty or inequality. Ideally, these contributions should exhaust the observed
distributional change, and it should not matter whether one chooses the initial
or end period as a reference. However, this is not always the case, because a
residual may be left unaccounted for, and the decomposition results could suffer
from path dependence. Therefore, in the second part of this section, we present
the Shapley (1953) method, which provides a unified framework to assess the
relative importance of determining factors (Shorrocks [1999] 2013). The final
part of this section considers application of the Shapley method to the decom-
position of change in poverty over time.
Framing the Decomposition
Given that poverty measures are computed on the basis of a distribution of living
standards that is fully characterized by its location parameter m (the mean of the
distribution) and the associated Lorenz curve L, it is reasonable to express a
poverty measure P as a function of these elements along with the poverty line, z,
as follows: Pt = P(mt, Lt, Zt). The overall variation in poverty, as measured by Pt,
from a base period (t = 0) to an end period (t = 1) can be written as follows:2
1
dP
P
∆O = P ( µ1 , L1 , z1 ) − P ( µ0 , L0 , z0 ) =
∫ dt dt. (2.1)
0
Equivalently, we have
1 1 1
∂P ∂ µ ∂P ∂L ∂P ∂ z
P
∆ =
O
∫0
∂ µ ∂t
dt +
∫
0
∂L ∂t
dt +
∫ ∂z ∂t dt. (2.2)
0
Equation (2.2) clearly indicates that we can think of the variation in poverty
over time as consisting of three basic components: the size effect associated with
the change in the mean of the underlying distribution, the redistribution effect
resulting from changes in the Lorenz curve (an indicator of relative inequality),
and a third component linked to the variation in the poverty line.
When working with real income (or expenditure) as an indicator of economic
welfare, the poverty line is considered fixed so that we write the overall level of
poverty as a function only of mean real income (or expenditure) and the Lorenz
function. With this assumption, the overall change in poverty defined by equa-
tion (2.2) can be written as
1 1
∂P ∂ µ ∂P ∂L
∆ =P
O
∫
0
∂ µ ∂t
dt +
∫ ∂L ∂t dt , (2.3)
0
so that any variation in poverty consists only of the size and the redistribution
effects. Note that these effects are defined in terms of partial derivatives of the
poverty measure. For a differentiable function of several variables, it is well
known that the partial derivative of that function at a particular point is the rate
of change of the function near the point with respect to one of the variables with
the other variables held constant. This notion of ceteris paribus variation that
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
18 A Simple Approach to Understanding Changes in Poverty and Inequality
underlies the definition of the decomposition terms presented in equation (2.3)
also suggests a clear strategy for the identification (and hence the estimation) of
those effects. The size effect, for instance, is defined as a variation in the poverty
outcome with respect to a variation in the average income, holding inequality
constant. Similarly, the redistribution effect is defined under the assumption of
size neutrality. The identification strategy underlying all decomposition methods
designed to implement equation (2.3) empirically relies on these definitions.
Decomposing changes in a variable over time can be viewed as a problem of
integral approximation (Müller 2008).3 Therefore, equation (2.1) provides a
unifying framework and a clear theoretical foundation for the decomposition
methods commonly discussed in the literature. In particular, its expression by
equation (2.3) provides an analytical framework for the study of the poverty
implications of a growth pattern, given that such a pattern can be characterized
in terms of the level of growth and the degree of relative inequality.
For example, Kraay (2006) proposes a decomposition of changes in poverty
that is consistent with the fundamental equation for dynamic decomposition
that is equation (2.1). Consider the class of additive poverty measures defined by
z
∫
P = y ( y | z ) f ( y ) dy , where y is the outcome variable (for example, a living-
0
standard indicator, such as income or consumption expenditure), z is the poverty
line, and f(y) is the density function associated with the distribution of y. The
term y ( y | z) is a convex and decreasing function measuring deprivation for an
individual with a level of economic welfare equal to y. This function is equal to
zero when the welfare indicator is greater than or equal to the poverty line.
Applying Leibnitz’s rule of differentiation under the integral sign and rear-
ranging terms, one gets the following expression for the proportionate change in
poverty,
z
dP 1
P
=
P ∫
yy ′ ( y | z ) g ( y ) f ( y ) dy , (2.4)
0
where y ' is the first-order derivative of the indicator of individual deprivation
and g(y) is the growth incidence curve4 (GIC) as defined by Ravallion and Chen
(2003). Equation (2.4) reveals that for the class of additively separable poverty
measures, a change in poverty over time can be written as a weighted sum of
points along the GIC.5 This fact implies that variation in poverty outcomes, as
measured by the class of additively separable poverty measures, inherits the
decomposability of the GIC. In particular, using the neutral element for addition,
one can split the GIC into one component showing the growth rate of average
income γ = dµ and another showing the deviation of each point on the curve
µ
from the overall growth rate. We therefore have g ( y) = g + [g ( y) − g ]. In fact, the
first component is the rate of growth that would be experienced at every quan-
tile if the growth process were distribution neutral. This is essentially the size
effect. It can be shown that the second component is equal to the change in the
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
A Simple Approach to Understanding Changes in Poverty and Inequality 19
slope of the Lorenz curve between the base and end periods (Ravallion and
Chen 2003). Thus, this component measures the redistribution effect. Equivalently,
we can express equation (2.4) as follows:
z z
dP γ 1
P
=
P ∫
yy ′ ( y | z ) f ( y ) dy +
0
P ∫
yy ′ ( y | z )
0
g ( y) − γ
f ( y ) dy.
(2.5)
Equation (2.5) therefore shows that the proportionate change in poverty over
time can be split into two components representing the size and redistribution
effects.6 It is clear from this equation that when growth is distribution neutral—
that is, g ( y) = g —the redistribution effect disappears.
The size effect is a product of two terms: the growth rate of per capita income
and the responsiveness of the chosen poverty measure to variations in incomes.
Ultimately, this responsiveness depends on the value judgments underpinning
poverty measurement. Such judgments are reflected in the specification of the
individual deprivation function,7 y ( y | z). The size effect measures the extent to
which poverty would have changed had the growth process been distribution
neutral. Similarly, the redistribution effect measures the extent to which poverty
would have changed had growth left per capita income unchanged (that is, had
the growth process been size neutral). This effect, too, depends on the elasticity
of poverty with respect to income.
Economic growth is pro-poor if it leads to poverty reduction for some
choice of poverty measure. On the basis of equation (2.5), Kraay (2006) identi-
fies three potential sources of pro-poor growth: (a) growth in income per
capita (g ); (b) the responsiveness of poverty to growth in average income (the
coefficient of h P in the size effect; see endnote 4); and (c) the pattern of relative
inequality.
The Datt-Ravallion Method
When the poverty line is held constant over time, the overall change in poverty
from the base period to the end period is equal to
P
∆O = P ( µ1, L1 ) − P ( µ0 , L0 ). (2.6)
Datt and Ravallion (1992) propose a threefold decomposition procedure
that allows an analyst to express that variation in poverty in terms of three
components: the size effect, the redistribution effect, and a residual interpreted
as an interaction effect. According to the ceteris paribus strategy, the identifica-
tion of each of these effects entails a comparison of observed outcomes with
counterfactual ones. For instance, the size effect is the change in poverty result-
ing from a variation in the mean while the Lorenz curve is held at some refer-
ence level. The identification of this effect requires that we compare the
observed poverty outcome with a counterfactual based on the same level of
inequality. In other words, the only difference between the two states is the
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
20 A Simple Approach to Understanding Changes in Poverty and Inequality
mean income. Similarly, the redistribution effect is the change in poverty result-
ing from a change in the Lorenz curve while holding the mean constant at some
reference level. Both the observed and counterfactual outcomes are based on
the same mean but different Lorenz curves.
A key methodological issue here is this: Which period should be used as refer-
ence for the factor that must be common to both the counterfactual and the
observed state of affairs? In principle, one could choose either the base period or
the end period as reference. However, Datt and Ravallion (1992) argue that the
base period is a natural choice for the decomposition and conduct their analysis
on that basis. Within that framework, the size effect is equal to the following
expression:
µ = P ( µ1 , L0 ) − P ( µ0 , L0 ) . (2.7)
∆P
Similarly, the redistribution effect is
L = P ( µ0 , L1 ) − P ( µ0 , L0 ) . (2.8)
∆P
Note that these two expressions describe counterfactual outcomes. The size
effect entails distribution neutral growth (the Lorenz curve does not change). The
redistribution effect implies that growth is size neutral (the mean does not
change).
To obtain the Datt-Ravallion decomposition, we can add and subtract these
counterfactual outcomes to and from the right-hand side of equation (2.6).
Upon rearranging terms, we get the following:
P
∆O P ( µ1, L1 ) − P ( µ1, L0 ) − P ( µ0 , L1 ) − P ( µ0 , L0 ) . (2.9)
µ + ∆L +
= ∆P P
The third term in brackets on the right-hand side of equation (2.9) is the
residual interpreted as the interaction effect. It is the difference between two
ways of computing the redistributive effect depending on whether one fixes the
end-period mean or the base-period mean. In other words, the residual is the
difference between the redistribution effect computed on the basis of the end-
period mean and the same effect evaluated at the initial mean (Datt and
Ravallion 1992; Ravallion 2000). This is an indicator of the extent to which
changes in per capita income matter to the responsiveness of poverty to
inequality.
Interestingly, we can rearrange terms within the residual and get the following
equivalent expression:
∆P P ( µ1, L1 ) − P ( µ0 , L1 )
R = P ( µ1, L0 ) − P ( µ0 , L0 )
− . (2.10)
This expression reveals that the residual is also equal to the difference
between the size effect computed on the basis of the end-period Lorenz
curve and the same effect evaluated at the initial-period Lorenz curve. Thus
the residual also shows the importance of inequality in determining the effect
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
A Simple Approach to Understanding Changes in Poverty and Inequality 21
of changes in per capita income on poverty. The structure of the residual
revealed by equations (2.9) and (2.10) led Datt and Ravallion (1992) to
interpret this residual as the interaction effect between the size and redistribu-
tion effects.
Indeed, if the size effect depends on the reference Lorenz curve, or the redis-
tribution effect on the reference mean, the residual would not equal zero. Thus,
the interaction term would vanish if the poverty measure is additively separable
between m and L.8 The residual would also vanish if one were to use the initial
year as the reference point in the computation of the size effect and switch to
the final year for the redistribution effect.
Finally, no residual would be involved if one performed the Datt-Ravallion
decomposition twice and took the average result. The first decomposition uses
the initial year as reference, while the second uses the end year (Ravallion 2000).
As it turns out, this average Datt-Ravallion decomposition is precisely the pro-
cedure proposed by Shorrocks ([1999] 2013) based on the Shapley (1953)
value. Kakwani (2000) proposes the same procedure but does not refer to the
Shapley value. Instead, he invokes a series of axioms that the decomposition
must satisfy.9
The Shapley Decomposition
In general, all decomposition methods used in distributional analysis seek to
identify and assign contributions to factors influencing some distributional statis-
tic. Ideally, the assignment is done in a manner that exhausts the quantity under
consideration. The Shapley method provides a unified framework to deal not
only with the identification and estimation of the size and redistribution effects
but also with many other situations in applied economics where there is a need
to assess the relative importance of determining factors (Shorrocks [1999]
2013). In this section, we first review the definition of the Shapley value that
motivates the method. Then we consider its application to the decomposition of
change in poverty over time.
The Shapley Value
The Shapley value provides a formula for dividing a joint cost or a jointly pro-
duced output on the basis of a fair assessment of individual contributions to
the formation of total cost or the production of a surplus. Thus, it can be
viewed as an interpretation of the reward principle of distributive justice
(Moulin 2003).10
Formally, the Shapley value is a solution to a cooperative game with transfer-
able utility (Shapley 1953). The problem of the commons is used often to explain
the nature of such games. A commons is a technology that is jointly owned and
operated by a group of agents. Consider a fixed set of agents engaged in such
a venture. Their problem is to share fairly the fruit of this collaboration. Starting
from scratch, let the agents join the venture one at a time in some predetermined
order. As each agent joins, the value (to be shared) increases. Thus the contribu-
tion of a given agent is the value added when he or she joins the venture.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
22 A Simple Approach to Understanding Changes in Poverty and Inequality
The Shapley value of a partner is her or his average contribution to value over all
possible orderings of the partners.
The Shapley allocation rule respects the following restrictions: (a) there
must be symmetry or anonymity; (b) the result must be an exact and additive
decomposition; and (c) the contribution of each factor is taken to be equal to its
(first-round) marginal impact. Symmetry means that the value assigned to play-
ers does not depend on their traits or identities. This property stems from the
fact that the value of a coalitional game to a player depends only on the char-
acteristic function that maps the set of all possible coalitions of players to a set
of payoffs. This function determines the payoff a group of players can expect
by forming a coalition.
To illustrate the basic idea underlying the Shapley value, consider a coalition
game with a set A of three players, A = {a, b, c}, and a characteristic (or value)
function u. The value of the grand coalition to be distributed among the three
players is equal to u (A). The number of all possible coalitions is equal to 23 = 8,
the number of all possible subsets of A (including the empty set). Thus there are
seven possible nonempty coalitions. By definition, the payoff to an empty coalition
is zero (u (Ø) = 0). Table 2.1 shows how to derive the Shapley value of each player.
Given an ordering, the computation of the marginal contribution depends
on the rank of the player in that ordering. Take, for instance, the first ordering
in table 2.1: (a, b, c). Since a is first, the marginal contribution of a is
ma = u ({a}) − u (Ø) = u ({a}). This is what he or she can achieve alone. We
interpret this situation as joining an empty coalition. Since b is in second
position, the corresponding marginal contribution is what he or she can
achieve in a coalition with a, minus what a can achieve alone. The same rea-
soning applies to the case of c. The Shapley payoff to each player equals the
arithmetic mean of her or his marginal contributions over the six possible
orderings. Thus, to obtain the Shapley value for each player, we add up the
marginal contributions in that player’s column and divide the result by six. In
the case of player a, the Shapley value is equal to the following expression:
1 1
Sa =
3
( )
v { a , b, c } +
6
( ) ( )
v {a , b} + v {a , c } − 2v ( b, c )
(2.11)
1
+
6
( ) ( ) ( )
2v {a } − v {b} − v {c }
.
Table 2.1 Shapley Allocations for a Three-Player Game
Ordering a b c
(a, b, c) u ({a}) u ({a, b}) − u ({a}) u ({a, b, c}) − u ({a, b})
(a, c, b) u ({a}) u ({a, b, c}) − u ({a, c}) u ({a, c}) − u ({a})
(c, a, b) u ({a, c}) − u ({c}) u ({a, b, c}) − u ({a, c}) u ({c})
(c, b, a) u ({a, b, c}) − u ({b, c}) u ({b, c}) − u ({c}) u ({c})
(b, c, a) u ({a, b, c}) − u ({b, c}) u ({b}) u ({b, c}) − u ({b})
(b, a, c) u ({a, b}) − u ({b}) u ({b}) u ({a, b, c}) − u ({a, b})
Shapley value Sa Sb Sc
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
A Simple Approach to Understanding Changes in Poverty and Inequality 23
We have written equation (2.11) in a way that makes transparent the fact
that it involves all nonempty coalitions that can be formed out of three players.
It is consistent with the general Shapley formula involving combinatorial analy-
sis. Let A be a fixed set of n players. Moulin (2003) explains that the Shapley
value can be viewed as a translation of the reward principle of fairness into an
explicit allocation of u (A) on the basis of (2n − 1) values u (C), for all non-
empty coalitions of the n players involved. Let Bj be the set of coalitions that
exclude player j, and Bj(k) the subset of Bj containing all coalitions of size k =
0, 1, 2, …, n−1. The Shapley value of player j is given by the following
formula:11
k !( n − k − 1) !
n −1
Sj = ∑ ∑( ) n!
(( ) )
v C ∪ { j} − v (C ) . (2.12)
k = 0 C ∈Bj k
The term mj = u (C ∪ { j}) − u (C) is the marginal contribution of player j if she
joins coalition C. The coefficient of the marginal value in equation (2.12) repre-
sents the probability that the coalition C of size k contains exactly all the players
preceding j in a random ordering of the grand coalition A. All marginal contribu-
tions involved in that expression are counterfactual outcomes. We now apply this
logic to the decomposition of change in poverty over time.
Application to Change in Poverty Over Time
To see how the above principle translates into a decomposition procedure, con-
sider a distributional statistic, such as the overall level of poverty or inequality.
For instance, let the poverty be a function of m contributory factors, which
together account for the value of the indicator. The m factors are thus analogous
to the players in a cooperative game. The decomposition approach proposed by
Shorrocks ([1999] 2013) is based on the marginal effect on the value of the
indicator resulting from eliminating sequentially each of the contributory factors
and computing the corresponding marginal change in the statistic. The method
then assigns to each factor the (arithmetic) average of its marginal contributions
in all possible elimination sequences.
Consider again the problem of decomposing a change in poverty over time
into a size and a redistribution effect using the Shapley method to implement
equation (2.3). We rewrite the overall variation in poverty as a function, H, of
two factors as follows:
P
∆O = H ( ∆µ , ∆L ). (2.13)
In other words, the overall change in poverty is fully determined by two
ontributory factors12—namely, the change in the mean of the distribution
c
Δm = m1 − m0 and the change in the Lorenz curve ΔL = L1 − L0. As seen in the
Datt-Ravallion method, the value of any effect (size or redistribution) depends
on the chosen period of reference. This path dependence violates the anonymity
constraint that the Shapley method must respect. We therefore need to consider
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
24 A Simple Approach to Understanding Changes in Poverty and Inequality
all possible sequences of elimination and the associated marginal contributions
that must be averaged in the end. In this simple case, we have only two possible
sequences: either we eliminate the size factor first by setting Δm = 0 and then the
redistribution factor by setting ΔL = 0, or we start with the redistribution factor
to end with the size factor.
The Shapley contribution of the size factor to change in poverty is equal to
the average (over the two possible elimination sequences) of the relevant mar-
ginal contributions. That is,
1
Sµ = P ( µ1, L1 ) − P ( µ0 , L1 ) P ( µ1, L0 ) − P ( µ0 , L0 )
+
. (2.14)
2
Similarly, the Shapley contribution of the redistribution factor to change in
poverty is equal to
1
SL = P ( µ1, L1 ) − P ( µ1, L0 ) P ( µ0 , L1 ) − P ( µ0 , L0 )
+
. (2.15)
2
P
The overall change in poverty can therefore be expressed as ∆O = Sµ + SL ,
which depends on the contribution of growth and redistribution with no remain-
ing residual.
Beyond the standard Datt-Ravallion partitioning, the decomposition of a
change in poverty over time in terms of the size and redistribution effects
can be embedded in a broader context to account for the effect of popula-
tion shifts or for the relative importance of components of an aggregate
living- standard indicator. The flexibility of the Shapley method makes it a
strong candidate for handling these integrated decompositions. Let the total
population of a given country be partitioned exhaustively into m socioeco-
nomic groups. For instance, these groups could represent different groups
according to geographic location, ethnicity, or other socioeconomic dimen-
sions. Let wkt be the share of population in group k at time t for t = {0, 1}
and Pkt the level of poverty in that group at the same time. For additively
decomposable poverty measures, overall poverty at time t can be written as
follows:
m
Pt = ∑w
k =1
P . (2.16)
kt kt
The change in aggregate poverty over time can now be written as follows:
m
P
∆O = ∑[w
k =1
P − w k 0 Pk 0 ]. (2.17)
k1 k1
Suppose we are interested in accounting for the overall change in poverty
P
∆O in terms of changes in within-group poverty, ΔPk = Pk1 − Pk 0, and the
population shifts between groups, Δwk = wk1 − wk 0 . We note that the
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
A Simple Approach to Understanding Changes in Poverty and Inequality 25
contribution of group k in the change of aggregate poverty is equal to the fol-
lowing expression:
Ck = wk1Pk1 − wk 0 Pk 0. (2.18)
If the population share of this group were fixed at the baseline level, the con-
tribution of this group to overall poverty change would be wk 0 ΔPk . We can add
and subtract this counterfactual to equation (2.18), rearrange terms, and sum
over k. We get the following decomposition presented in Bourguignon and
Ferreira (2005):13
m m
P
∆O = ∑
k =1
w k 0 ∆Pk + ∑P ∆w
k =1
k1 k. (2.19)
According to equation (2.19), the overall change in poverty can be split into
two components: one representing the contribution of changes in within-group
poverty and the other the contribution of population shifts.
The decomposition presented in equation (2.19) is path dependent since
changes in poverty are weighted by the base-year population shares while
changes in population shares are weighted by the end-year poverty level. The
Shapley principle leads to the following twofold decomposition (Shorrocks
[1999] 2013):14
m m
w k 0 + w k1 Pk 0 + Pk1
P
∆ =
O ∑k =1
2
∆Pk + ∑
k =1
2
∆w k. (2.20)
Equation (2.20) therefore allows for a decomposition of changes in poverty
on account of changes in within-group poverty and changes in population
between groups.
Accounting for the Contribution of Demographics and Income
Components
So far we have seen decompositions of changes in poverty on account of
changes in growth and redistribution and decompositions of within-group
changes in poverty as opposed to changes in population between groups.
However, there is also interest in decomposing changes in poverty into the
contributions that changes in demographics, employment, public transfers, and
remittances could have made. In this section, we describe an approach to
undertake these sorts of decompositions involving the Shapley method
described above.
As proposed by Azevedo et al. (2013), we begin with a household consump-
tion identity whereby household consumption, Ch, per capita is defined by
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
26 A Simple Approach to Understanding Changes in Poverty and Inequality
Y
Ch = q h nh , (2.21)
where Yh is total household income, n is the number of household members, and
q h is the consumption-to-income ratio. Because poverty depends on the distribu-
tion of consumption, changes in any of the factors on the right-hand side of
equation (2.21) will lead to changes in poverty: demographic changes (n),
growth in labor and nonlabor income (which make up Yh), and changes in con-
sumption patterns (q h).
Following Barros et al. (2006), household per capita income can be modeled as
n
∑y . (2.22)
Y 1
Y pc = nh = n i
i =1
Income per capita is based on the sum of each individual’s income as well as
the number of household members, n. If we recognize that only individuals
15 years and older contribute to family income, income per capita depends on
the number of adults in the family, nA, so income per capita can be written as
1 n
∑
nA
Y pc = yi . (2.23)
n
nA i =1
Income per adult includes labor income, yiL, and nonlabor income, yiNL, where
nonlabor income includes public social transfers, pensions, remittances, and other
private transfers. Specifying each type of income, we have
1 n n
∑ ∑y
nA 1
yiL + . (2.24)
NL
Y pc = i
n
nA nA
i ∈A i ∈A
Finally, not all adults in the household are occupied (working), and household
labor income per capita depends on the income of employed adults. Therefore,
we can decompose the labor income per occupied adult as
nA 1
n n
Y pc = no 1
n nA n ∑ yiL +
nA
∑ yiNL , (2.25)
o i ∈A i ∈A
where no is the number of occupied adults.
Note that official poverty rates in some countries are calculated on the basis
of household income. In these cases, equation (2.25) is sufficient to decompose
the contribution of demographic factors, labor income, and nonlabor income to
observed poverty reduction. However, most countries measure the distribution
of welfare, and poverty in particular, using household consumption. Therefore,
we modify the Barros et al. (2006) approach by mapping consumption to
income. In particular, we refer to the household consumption identity in (2.21).
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
A Simple Approach to Understanding Changes in Poverty and Inequality 27
Combining (2.21) and (2.25) above, we can express household consumption per
capita, Cpc, as
n 1 n
1 n
∑ ∑
n
C pc = q h A o yiL + yiNL . (2.26)
n nA n
nA
o i ∈A i ∈A
With this framework, whether countries measure welfare by per capita house-
hold income or consumption, we can separate the demographic, labor, and non-
labor components discussed earlier. In addition, we can separate the contribution
of changes in consumption patterns over time in poverty reduction. The deter-
minants of per capita consumption are summarized in figure 2.1.
Contributions of Determinants of Consumption to Poverty Reduction
Let F(.) be the cumulative distribution function (CDF) of the distribution
of welfare. Since the factors in equation (2.6) determine the changes in the
CDF of consumption underlying changes in poverty outcomes, we can trace the
effect of those factors on poverty. As a result, any poverty measure can be writ-
ten as a function of each of these components. Therefore, the contribution of
Figure 2.1 Determinants of Consumption per Capita
Propensity to
consume and
measurement
error = θh
Number of Share of
Consumption
household occupied
per capita
members = n Share of adults = n0/nA
adults =
nA/n Labor
income
per adult
Income Labor income
per capita per occupied
adult
=1/n0∑yiL
Income
per adult
Nonlabor
income
per adult
=1/nA∑yiNL
Source: Adaptation of Barros et al. 2006.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
28 A Simple Approach to Understanding Changes in Poverty and Inequality
each component toward changes in poverty or distribution can be expressed as
afunction of these indicators in the initial and end periods.
Following Barros et al. (2006), we can then simulate the distribution of
welfare by changing each of these components, one at a time, to calculate its
contribution to the observed changes in poverty. In particular, let v be a measure
of poverty, inequality, or any other distributional statistic. Then, this measure will
be a function of the cumulative density function, F(.), which in turn depends on
each of the factors above, as follows:
n n NL
ϑ = Φ F C pc q h , A , o , yPO
L
, yPA , (2.27)
n nA
where
n
∑y
L 1 L
yPO = i
no
i ∈A
and
n
∑y
1
NL
yPA = NL
i .
nA
i ∈A
Given that the distribution of per capita consumption for period 0 and
period 1 are known, we can construct counterfactual distributions for period 1 by
substituting the observed level of the indicators in period 0, one at a time. For
each counterfactual distribution, we can compute the poverty measure and inter-
pret those counterfactuals as the poverty that would have prevailed in the
absence of a change in that indicator. For example, to see the impact of the change
in the share of occupied adults, we compute ϑ ˆ , where we substitute the value of
no
n observed in period 0 to the observed distribution in period 1 as follows,
A
ϑˆ = Φ F C pc q h , n A , n
ˆ0 L NL
, yPO , yPA
,
n nA (2.28)
such that the contribution of the share of occupied adults is the difference
between the observed ϑ in period 1 and the estimated counterfactual, ϑ ˆ.
Similarly, each of the other components in the consumption per capita distribu-
tion in period 1 can be substituted by its value in period 0 so that its contribution
to changes in poverty can be computed.
Since we don’t have panel data, we do not observe period 1 households
in period 0. Therefore, we use a rank-preserving transformation to assign first-
period characteristics to the second-period observations. This method uses an
idea first proposed by Juhn, Murphy, and Pierce (1993), who decomposed
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
A Simple Approach to Understanding Changes in Poverty and Inequality 29
changes in wages by running Mincer-type ordinary least squares (OLS) regres-
sions that make it possible to decompose labor income inequality, using any
measure of inequality, into three parts:
• Quantity effects, referring to the distribution of observable workers’ character-
istics, such as education and labor market experience, and are included as
regressors in the equation;
• Price effects, which capture changes in returns to observed characteristics
through the regression’s coefficients;
• The regression residual (unobservables), which reflects changes in inequality
within education and experience groups.
Counterfactuals for the quantity effects can be created by assigning the mean
observable characteristic from one period to the other, and the counterfactual for
the price effects can be created by substituting regression coefficients from one
period to another. However, to complete that analysis, the authors needed to
assign a value to the residuals in each period. So they created a counterfactual by
ordering households by their earnings in each period and then taking the average
residual value in each quantile from the first period and assigning it to all house-
holds in the same quantile in the second period.
In this case, instead of running a Mincer regression, we create counterfactuals
by ordering households by their level of welfare (measured by either household
per capita consumption or income) and then taking the average value of each
characteristic in equation (2.25) for each quantile in period 0 and assigning it to
each household in that same quantile in period 1. For example, if we are decom-
posing the effect of labor income, we order households into quantiles by their
observed total household income in periods 0 and 1. Then, for every quantile in
period 1, we replace the period 1 labor income with the average labor income in
period 0 from households that were in the same quantile.
Barros et al. (2006) compute each counterfactual simulation in a nested fash-
ion (table 2.2). They identify the contribution that interactions between variables
have in poverty reduction by first computing the joint impact of a subset of
variables, and then subtracting the marginal impact of each variable, one at a
time. For instance, in step 2 in table 2.2, they first compute the joint impact of
inserting both the share of adults and the income per adult from the first period
into the distribution of the second period. They then compute the impact of
changing only the share of adults and take the difference of these two simulations
to approximate the marginal impact that changing the share of adults had on the
distribution. However, in step 4, instead of computing the impact of income per
adult on its own, they compute the impact of changing both the labor and non-
labor income per adult. This is done because the sum of labor and nonlabor
income should be equivalent to changing total income per adult. The results of
these two simulations are different, however, and the simulation of labor income
is not done explicitly, but rather ends up being a “residual” in step 8 to ensure
that the cumulative effect adds up to the total distributional change.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
30 A Simple Approach to Understanding Changes in Poverty and Inequality
Table 2.2 Application of Barros Methodology to Measure Contributions of Variables to Change in Poverty
Step Estimate Variable or interaction being measured
1. Initial poverty rate is J0.
n n NL
ϑ 0 = Φ F Y pc A , o , yPO
L
, yPA
n n A
2. Contribution of the interaction between share
n
ϑ a1 = Φ F Y pc A , yPA of adults and income per adult is ϑ a1 − ϑ 0 .
n
3. Contribution of share of household adults is
n
ϑ nA = Φ F Y pc A , yPA ϑ nA − ϑ a1.
n
4. Contribution of the interaction between labor
n n NL
ϑ a 2 = Φ F Y pc A , o , yPO
L
, yPA and nonlabor income is ϑ a 2 − ϑ nA.
n nA
5. n n NL
Contribution of nonlabor income is ϑ NL − ϑ a1.
ϑ NL = Φ F Y pc A , o , yPO
L
, yPA
n nA
6. Contribution of the interaction between labor
n n income and the share of occupied adults is
ϑ a 3 = Φ F Y pc A , o , yPO
L NL
, yPA
n nA
ϑ a 3 − ϑ NL.
7. Contribution of the share of occupied adults is
n n
ϑ no = Φ F Y pc A , o , yPO
L NL
, yPA ϑ no − ϑ a 3.
n nA
8. Final poverty rate is JF. The contribution of
n n NL
L
labor income, y PO, is calculated as a
ϑ F = Φ F Y pc A , o , yPO
L
, yPA
n nA
residual: ϑ F − ϑ a 3.
Azevedo et al. (2013) modify this procedure in three important ways:
• They focus on consumption as a measure of welfare.
• They compute a cumulative counterfactual distribution by adding one vari-
able at a time.
• They compute Shapley-Shorrocks estimates of each component.
The first change—the focus on consumption—is made because most develop-
ing countries use a consumption aggregate to measure poverty. Second, in contrast
to the Barros et al. (2006) approach, the proposed method does not separately
identify the contribution of the interaction between variables in the observed
distributional changes; doing so is partial at best, given that changing any variable
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
A Simple Approach to Understanding Changes in Poverty and Inequality 31
can potentially affect all other variables. Instead, the impact of changes in each
variable and its interactions with all other variables is calculated as the difference
between the cumulative counterfactuals. Table 2.3 shows an example for one pos-
sible path, taking into account the fact that nonlabor income is made up of pen-
sions, transfers, capital income, and other income.
The third methodological change—computing Shapley-Shorrocks estimates
of each component—addresses the fact that this methodology suffers from path
dependence, described in the previous section. In other words, the order in which
the cumulative effects are calculated matters.15 The best-known remedy for this
is to calculate the cumulative decomposition in every possible order and then
Table 2.3 Proposed Methodology to Decompose Change in Poverty along One Possible Path
Step Estimate Variable measured
1. Initial poverty rate is ϑ 0 .
n n NL
ϑ 0 = Φ F C pc q h , A , o , yPO
L
, yPA
n nA
2. Contribution of share of
n n
ϑ1 = Φ F C pc q h , A , o , yPO
L NL
, yPA household adults is ϑ1 − ϑ 0 .
n nA
3. Contribution of the share of
n n
ϑ 2 = Φ F C pc q h , A , o , yPO
L NL occupied adults is ϑ 2 − ϑ1.
, yPA
n nA
4. Contribution of pensions is
n n
ϑ 3 = Φ F C pc q h , A , o , yPO
L Pens
, yPA Trans
, yPA , yCap Oth NL
PA , y PA ϑ 3 − ϑ 2.
n nA
5. Contribution of transfers is
n n
ϑ 4 = Φ F C pc q h , A , o , yPO
L Pens
, yPA Trans
, yPA , yCap Oth NL
PA , y PA ϑ 4 − ϑ 3.
n nA
6. Contribution of capital income is
n n
ϑ 5 = Φ F C pc q h , A , o , yPO
L Pens
, yPA Trans
, yPA , yCap
PA , y Oth NL
PA ϑ 5 − ϑ 4.
n nA
7. Contribution of other nonlabor
n n
ϑ 6 = Φ F C pc q h , A , o , yPO
L Pens Trans
, yCap Oth NL income is ϑ 6 − ϑ5 .
, yPA , yPA PA , y PA
n nA
8. Final poverty rate is ϑ F .
n n NL
ϑ F = Φ F C pc q h , A , o , yPO
L
, yPA Contribution of labor income is
n nA
ϑ F − ϑ 6.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
32 A Simple Approach to Understanding Changes in Poverty and Inequality
average the results for each component to get the Shapley-Shorrocks estimate of
the contribution of each component, as described earlier (Shapley 1953;
Shorrocks [1999] 2013).
Summary and Conclusions
This chapter provides a unifying framework and a theoretical foundation for the
decomposition methods commonly used in the literature. We began with the
simplest decompositions, noting that changes in poverty can be characterized in
terms of the level of growth and the degree of relative inequality.
The most common method stemming from this view is the Datt and
Ravallion (1992) decomposition. Ideally, the contribution of growth and
inequality should exhaust the observed change in poverty, and the results should
be the same regardless of the reference period. However, this is not the case,
because a residual is typically left unaccounted for in the standard Datt-
Ravallion method, and the choice of reference period can affect the results. As
described in this chapter, the Shapley method provides a unified framework to
assess the relative importance of determining factors (Shorrocks [1999] 2013)
and addresses both of these concerns. Use of the Shapley method can be
extended to more complex decompositions, such as one that accounts for
changes to poverty resulting from changes within socioeconomic groups as
opposed to population shifts between groups.
Although these decompositions are interesting, analysts often want to go
beyond these summary statistics to decompose the contributions that changes in
demographics, employment, public transfers, and remittances could have made
toward poverty reduction. We described an approach to undertake these sorts of
decompositions, as proposed by Azevedo et al. (2013), that involves the Shapley
method described above. In contrast to the simpler methods, this approach rec-
ognizes that poverty is a function of total household per capita consumption,
which can be written as an accounting identity that depends on the number of
household members, the consumption-to-income ratio, and household income.
Based on these identities, we can simulate changes in the distribution of welfare
by changing each component, one at a time, to calculate its contribution to the
observed changes in poverty.
Assuming that only cross-sectional data are available, counterfactuals can be
created nonparametrically, by ordering households according to their total
income, creating quantiles for the initial and end periods, and then taking the
average value of each component for each quantile in one period and assigning it
to each household in that same quantile in the other period. When the cumula-
tive impact of each of the components is analyzed, the methodology suffers
from path dependence. Therefore, the proposed method calculates the Shapley-
Shorrocks estimates of each component to find the contribution of each
component.
With chapter 2 having described a set of relatively simple decompositions,
chapter 3 will apply the proposed methods to a large set of countries that have
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
A Simple Approach to Understanding Changes in Poverty and Inequality 33
seen significant declines in poverty over the past decade. However, before pro-
ceeding, it is important to point out two remaining caveats in this approach, the
first of which is addressed later in the book.
First, although the decomposition method proposed here is useful to distin-
guish the main contributors to poverty reduction, its main limitation is the fact
that it cannot shed light on whether the decline in poverty was the result of
changes in the endowments of the population (such as higher educational levels
or increases in other productive assets) or the result of changes in returns to those
endowments. For this, one must turn to alternative decomposition techniques
that impose an underlying labor model and greater structure (such as the meth-
ods proposed in chapters 4 and 6), compared with the nonparametric approach
adopted here.
The second caveat to this approach is that the counterfactual income distri-
butions on which these decompositions rely suffer from equilibrium inconsis-
tency. Because the decomposition modifies only one element at a time, the
counterfactuals do not result from an economic equilibrium, but rather from a
ceteris paribus exercise in which we assume that we can, in fact, modify one fac-
tor at a time and keep everything else constant. To address this concern, one
would need to employ general equilibrium modeling, which is outside the scope
of this book.
Notes
1. For a recent review of micro-decomposition methods, see Fortin, Lemieux, and Firpo
2011; Essama-Nssah 2012; and chapter 6 of this volume.
2. By the fundamental theorem of calculus, we know that if a primitive function F of
b b
another function f is known for a ≤ x ≤ b , then
(Kaplan 1991).
∫
a
∫
f ( x ) dx = F ′ ( x ) dx = F ( b ) − F ( a )
a
3. Let Q(t) = Q(x1(t), x2(t), …, xm(t)) be a quantity of interest that is a function of
m time-dependent variables xk(t), k = 1, 2, …, m. The fundamental equation
for decomposition analogous to equation (2.1) or (2.2) can be written as
m 1
∂Q ∂ x k
∆Q = ∑ ∫ ∂x
k =1 0
k ∂t
dt. Any component of this expression that contains the deriva-
tive with respect to xk is taken to represent the contribution of changes in this factor
to the overall change in Q. A successful implementation of this decomposition
requires the ability to solve the integrals involved or to approximate them. Müller
(2008) explains that terms in this expression can be approximated by their values
at the upper boundary, computing each derivative as the slope of the straight line
joining the endpoints. He further points out that even though static decomposition
of cross-sectional variations is formally equivalent to dynamic decomposition, for
static decomposition to be meaningful there would need to be a continuous range
of the variables involved between spatial or socioeconomic groups. Thus, the
framework of integral approximation may not be appropriate for static decomposi-
tion. The notion of a path connecting socioeconomic groups does not necessarily
make sense.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
34 A Simple Approach to Understanding Changes in Poverty and Inequality
4. Suppose that y is continuously distributed over the population of interest. Denote by
Ft(y) the cumulative distribution function (CDF) of income showing the proportion,
y
∫
τ = f ( y ) dy , of the population with income less than y at time t. The income level at
0
the t th quantile is given by the inverse of the CDF: yt (τ ) = Ft (τ ). The growth rate
−1
y1 (τ )
of income at the t th quantile between t = 0 and t = 1 is equal to g (τ ) = − 1 .
y0 (τ )
The growth incidence curve is obtained by letting t vary from zero to one and plot-
ting against the corresponding values of g(t ). The quantiles involved are based on the
ranking of individuals in an increasing order of their baseline income.
5. See Essama-Nssah and Lambert (2009) for an alternative derivation of this result.
6. Kakwani (1993) proposes a similar decomposition of a proportionate change in pov-
erty based on the responsiveness of the chosen poverty measure to changes in mean
income and to variation in the Gini coefficient.
7. Members of the additively separable poverty measures include, among others, the
Watts (1968) index (W) and the Foster, Greer, and Thorbecke (1984) family (FGT).
The associated deprivation functions are y W ( y | z ) = ln ( z y ) for the Watts index, and
α
y
y FGT ( y | z ) = 1 − , α ≥ 0, for the FGT family. When α = 0, the deprivation leads
z
to the headcount index: the proportion of individuals with living standards below the
poverty line. When α = 1, we get the normalized poverty deficit. Finally, when the
same parameter is equal to 2, we get the “squared poverty gap.”
8. Ravallion (2000) clarifies this point by noting that, in general, if a variable v is a func-
tion of two variables x and y and if this function is additively separable in x and y, then
we can write: v = g(x) + h(y). In these circumstances, the change in v when x changes
holding y constant depends only on the initial and final values of x. Without this addi-
tive separability, we should expect the variation in v to depend on the particular value
of y chosen.
9. In particular, Kakwani (2000) stipulates the following three axioms: First, the size and
redistribution effects must exhaust the total change in poverty. Second, if both the
size and the redistribution effects have the same sign (either [a] less than or equal to
zero, or [b] greater than or equal to zero), then the overall change in poverty must
have the same sign; otherwise, the total change in poverty must depend on the mag-
nitude of the size and the redistribution effects. Third, both the size and the redistri-
bution effects must be symmetric with respect to the base and end years. The latter
axiom means that if changes in per capita income, for instance, induce poverty reduc-
tion as we go from the base to the end year, we must accept that moving from the end
state to the initial state would increase poverty. A similar logic applies to the redistri-
bution effect.
10. Moulin (2003) argues that the concept of fairness can be interpreted in terms of four
basic principles: exogenous rights, compensation, reward, and fitness. An exogenous right
is a normative postulate that dictates how a resource must be distributed among
claimants. Equal treatment of equals is an example of such a postulate. In general,
exogenous rights set claims to resources independently of the use of such resources
and of the contribution to their production, while compensation and reward relate
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
A Simple Approach to Understanding Changes in Poverty and Inequality 35
fairness to individual characteristics relevant to the use or production of the resources
under consideration. The compensation principle advocates giving extra resources to
people who find themselves in unfortunate circumstances for which they cannot be
held (morally) responsible. The reward principle bases allocation on individual
behavior to the extent that it affects the overall burden or advantage under distribu-
tion. Finally, according to the fitness principle, resources must go to the person who can
make the best use of them.
11.
The Shapley value of player j can be equivalently expressed as
k !( n − k − 1) !
Sj = ∑{ } n!
(( ) )
v C ∪ { j} − v (C ) , where the sum is taken over all possible
C⊆ A \ j
coalitions (including the empty one) that can be built on the basis of the members of
A.
12. Thus H(0,0) = 0.
13. It is possible to further transform this twofold decomposition as follows. Consider a
counterfactual situation in which within-group poverty does not change. On the basis
of equation (2.19), the contribution of group k to change in poverty can be written as
Ck = w k 0 ( Pk1 − Pk 0 ) + Pk1 ( w k1 − w k 0 ) . Fixing group-level poverty at the base level
reduces this contribution to Pk 0 ( w k1 − w k 0 ). We can add and subtract this cou
nterfactual outcome to and from Ck, rearrange terms, and sum up over k to get
the threefold decomposition proposed by Ravallion and Huppi (1991):
m m m
P
∆O = ∑
k =1
w k 0 ∆Pk + ∑
k =1
Pk 0 ∆w k + ∑∆w ∆P . This
k =1
k k expression shows that change in
aggregate poverty over time can be decomposed into three components representing,
respectively, within-group effects, the effect associated with population shifts, and
interaction effects. This is analogous to the Datt-Ravallion decomposition for the size
and redistribution effects.
14.
Son (2003) proposes a similar expression in percentage change:
P m m
∆ Pk 0 w k 0 + w k1 ∆Pk w k 0 Pk 0 + Pk1 ∆w k
O
P0
= ∑P
k =1
0
2
Pk 0
+ ∑P
k =1
0
2
wk0
.
15.
Path dependence is common in the micro-decomposition literature. For recent
reviews of the literature, see Chapter 6 of this volume; Essama-Nssah 2012; Fortin,
Lemieux, and Firpo 2011; and Ferreira 2012.
Bibliography
Azevedo, João Pedro, Gabriela Inchauste, Sergio Olivieri, Jaime Saavedra, and Hernan
Winkler. 2013. “Is Labor Income Responsible for Poverty Reduction? A Decomposition
Approach.” Policy Research Working Paper 6414, World Bank, Washington, DC.
Barros, Ricardo Paes, Mirela de Carvalho, Samuel Franco, and Rosane Mendonça. 2006.
“Uma Análise das Principais Causas da Queda Recente na Desigualdade de Renda
Brasileira.” Revista Econômica 8 (1): 117–47.
Bourguignon François, and Francisco H. G. Ferreira. 2005. “Decomposing Changes in the
Distribution of Household Incomes: Methodological Aspects.” In The Microeconomics
of Income Distribution Dynamics in East Asia and Latin America, edited by François
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
36 A Simple Approach to Understanding Changes in Poverty and Inequality
Bourguignon, Francisco H. G. Ferreira, and Nora Lustig, 17–46. Washington, DC:
World Bank.
Datt, Gaurav, and Martin Ravallion. 1992. “Growth and Redistribution Components of
Changes in Poverty Measures: A Decomposition with Applications to Brazil and India
in the 1980s.” Journal of Development Economics 38 (2): 275–96.
Essama-Nssah, B. 2012. “Identification of Sources of Variation in Poverty Outcomes.”
Policy Research Working Paper 5954, World Bank, Washington, DC.
Essama-Nssah, B., and Peter J. Lambert. 2009. “Measuring Pro-Poorness: A Unifying
Approach with New Results.” Review of Income and Wealth 55 (3): 752–78.
Ferreira, Francisco H. G. 2012. “Distributions in Motion: Economic Growth, Inequality,
and Poverty Dynamics.” In The Oxford Handbook of the Economics of Poverty, edited by
Philip N. Jefferson, 427–62. New York: Oxford University Press.
Fortin, Nicole, Thomas Lemieux, and Sergio Firpo. 2011. “Decomposition Methods in
Economics.” In Handbook of Labor Economics, Vol. 4A, edited by Orley Ashenfelter
and David Card, 1–102. Amsterdam, the Netherlands: North-Holland.
Foster, James, Joel Greer, and Erik Thorbecke. 1984. “A Class of Decomposable Poverty
Measures.” Econometrica 52 (3): 761–65.
Juhn, Chinhui, Kevin M. Murphy, and Brooks Pierce. 1993. “Wage Inequality and the Rise
of Returns to Skill.” Journal of Political Economy 101 (3): 410–42.
Kakwani, Nanak. 1993. “Poverty and Economic Growth with Application to Côte
d’Ivoire.” Review of Income and Wealth 39 (2): 121–39.
———. 2000. “On Measuring Growth and Inequality Components of Poverty with
Application to Thailand.” Journal of Quantitative Economics 16 (1): 67–79.
Kaplan, Wilfred. 1991. Advanced Calculus. Redwood City, CA: Addison-Wesley.
Kolenikov, Stanislav, and Anthony Shorrocks. 2005. “A Decomposition Analysis of
Regional Poverty in Russia.” Review of Development Economics 9 (1): 25–46.
Kraay, Aart. 2006. “When Is Growth Pro-Poor? Evidence from a Panel of Countries.”
Journal of Development Economics 80 (1): 198–227.
Moulin, Hervé. 2003. Fair Division and Collective Welfare. Cambridge, MA: MIT Press.
Müller, Adrian. 2008. “Clarifying Poverty Decomposition.” Conference Paper 30,
Proceedings of the German Development Economics Conference, Zurich.
Ravallion, Martin. 2000. “On Decomposing Changes in Poverty into ‘Growth’ and
‘Redistribution’ Components.” Journal of Quantitative Economics 16 (1): 105–18.
Ravallion, Martin, and Monika Huppi. 1991. “Measuring Changes in Poverty: A
Methodological Case Study of Indonesia during an Adjustment Period.” World Bank
Economic Review 5 (5): 57–82.
Ravallion, Martin, and Shaohua Chen. 2003. “Measuring Pro Poor Growth.” Economic
Letters 78 (1): 93–99.
Shapley, Lloyd. S. 1953. “A Value for n-Person Games.” In Contributions to the Theory of
Games, Vol. 2, edited by H. W. Kuhn and A. W. Tucker, 307–17. Princeton, NJ:
Princeton University Press.
Shorrocks, Anthony. F. (1999) 2013. “Decomposition Procedures for Distributional
Analysis: A Unified Framework Based on Shapley Value.” Journal of Economic
Inequality 11 (1): 1–28. doi: 10.1007/s10888-011-9214-z.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
A Simple Approach to Understanding Changes in Poverty and Inequality 37
Son, Hyun Hwa. 2003. “A New Poverty Decomposition.” Journal of Economic Inequality
1 (2): 181–87.
Watts, Harold. 1968. “An Economic Definition of Poverty.” In On Understanding Poverty:
Perspectives from the Social Sciences, edited by Daniel P. Moynihan, 316–29. New York:
Basic Books.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Chapter 3
What Accounts for Changes in
Poverty over the Past Decade?
Introduction
The reduction in poverty observed during the 2000s throughout the developing
world provides an opportunity to study the most significant factors at work in
favor of the poor. Did poverty decrease because of
• Demographic changes that led to lower dependency ratios;
• Labor market improvements that boosted employment or labor income;
• Improved and more effective social policies; or
• Increased remittances to poorer countries?
To answer these questions, and to contribute to the evidence base for future
policy, this chapter focuses on a subsample of 21 countries where poverty
declined substantially in the decade from 2000 to 2010. Particularly in some
Latin American countries, debate focuses on whether better job opportunities or
improved transfer policies can best explain the observed reductions in poverty
and inequality. In some South Asian and Eastern European countries, observers
question whether it was better job opportunities or higher remittances that
reduced poverty. As for several East Asian countries—where poverty reduction
has coincided with strong growth and job creation—questions arise about
whether social policy should focus more on redistribution.
This chapter quantifies, based on a series of counterfactual simulations, the
contribution of labor income to changes in poverty across countries. Based on the
“simple” methodology described in chapter 2—which improved on methods that
focus on aggregate summary statistics—this chapter generates entire counterfac-
tual distributions, allowing us to quantify the relative contributions to poverty
reduction as a result of changes in labor income, nonlabor income, and demo-
graphic characteristics. Because most countries measure welfare through house-
hold expenditures or consumption (as opposed to income), this chapter modifies
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7 39
40 What Accounts for Changes in Poverty over the Past Decade?
existing methods and proposes a decomposition methodology for consumption-
based measures of poverty.
The next section, “Growth and Poverty Reduction,” describes the evolution of
poverty across our 21-country sample, highlighting the links between poverty,
growth and redistribution outcomes. “The Forces behind Poverty Reduction”
discusses four potentially influential factors: demographic change, increased labor
income, increased nonlabor income, and changed consumption patterns. We then
present the data and the results for each country, highlighting similarities and
differences. The concluding section summarizes the findings and assesses the
decomposition methodology and its limitations. Annex 3A lists the specific data
sources for each country in the sample.
Growth and Poverty Reduction
We focus here on 21 countries that exhibited substantial declines in moderate
poverty (defined below) using comparable consumption or income data in the
decade from 2000 to 2010. The countries included in this analysis are
Argentina, Bangladesh, Brazil, Cambodia, Chile, Colombia, Costa Rica, Ecuador,
Ghana, Honduras, Moldova, Mongolia, Nepal, Panama, Paraguay, Peru, the
Philippines, Romania, Sri Lanka, Thailand, and Vietnam.
Data Sources and Distinctions
The analysis focuses on reductions in poverty during the 2000s. Most Latin
American countries in the sample use income-based measures of poverty. For
these countries, data come from national household surveys that have been har-
monized and compiled in the Socio-Economic Database for Latin America and
the Caribbean (SEDLAC).1
Other countries in the sample use consumption-based measures of poverty,
with data provided by national household surveys. For Bangladesh, Moldova,
Peru, the Philippines, Romania, Sri Lanka, and Thailand, household surveys
were standardized by the World Bank. For Ghana’s and Nepal’s consumption-
based measures, we use the Rural Income Generating Activities (RIGA)
datasets, a harmonized database of household surveys compiled jointly by the
Food and Agriculture Organization of the United Nations (FAO) and the World
Bank (FAO n.d.).2 Table 3A.1 (in annex 3A) provides more detail about the
countries, exact years, and surveys included in this study. Regarding Cambodia,
Mongolia, Vietnam, and the Philippines, household surveys were harmonized by
the World Bank to generate comparable income data and income-based poverty
measures. The official national poverty measurements in Cambodia, Mongolia,
and Vietnam are based on the consumption welfare aggregate.
Poverty Reduction in Sample Countries
All countries in our sample had substantial poverty-reduction episodes, defined
as an average decline in moderate poverty of 1 percentage point or more per year
over a period of four years or more. Because the national moderate poverty line
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
What Accounts for Changes in Poverty over the Past Decade? 41
varies from country to country, we base the analysis on the international poverty
line that is closest in magnitude to the national moderate poverty line in that
country. For example, the moderate poverty line in Bangladesh is closer to the
$1.25-a-day poverty line, while the moderate poverty line in Peru is closer to the
$4-a-day poverty line, as shown in table 3.1.3
GDP Growth and Poverty Reduction
Poverty reduction in each of the sample countries was accompanied by strong
economic growth (figure 3.1), albeit at different rates, ranging from an average of
3 percent in real gross domestic product (GDP) in Paraguay to an average of
8.4 percent in Mongolia. The 2008–09 global financial crisis increased volatility
and vulnerability in middle-income countries, such as Romania, Thailand, and
several countries in Latin America, while other countries, such as Bangladesh,
enjoyed continued, almost uninterrupted real GDP growth of about 6 percent a
year throughout the decade.
For the countries whose moderate poverty lines came closest to the
$1.25-a-day level, the poverty reduction rate varied from an average of
percentage points a year in the Philippines to 2.9 percentage points a year
1.3
in Nepal (figure 3.2). Among countries whose national moderate poverty
lines approach the $4-a-day level, the decline varied from an average of
1 percentage point per year in Paraguay to an average of 2.8 percentage
points per year in Colombia.
The link between economic growth and poverty reduction has long inter-
ested economists. As detailed in Ferreira (2012), the cross-country literature
has found considerable evidence that economic growth is strongly and nega-
tively correlated with changes in poverty (Ravallion and Chen 2007). In addi-
tion, the higher a country’s initial level of inequality, the higher the growth rate
needed to reduce poverty by a given amount (World Bank 2005; Ravallion and
Chen 2007).
One common way to assess these relations is by using the Datt and
Ravallion (1992) decomposition, which splits the change in poverty into
distribution-neutral growth and redistribution effects. Using this method, we
found that growth explains most of the observed reduction in moderate
poverty for 17 of the 21 countries in this study (as shown in figure 3.2).
Redistribution, which can be thought of as a reduction in inequality, was
found to be more important in the cases of Argentina, Mongolia, Paraguay, and
the Philippines.
Poverty Effects beyond Aggregate Economic Growth
In the places where most of the poverty reduction was the result of growth, some
obvious questions arise: How did growth lead to poverty reduction? Were redis-
tributional changes associated with the introduction of public transfers, or were
they a result of market forces? Unfortunately, the Datt-Ravallion (1992) method
cannot make these explicit links because growth, inequality, and poverty mea-
sures are just three different aggregations of information about individual income
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
42
Table 3.1 Poverty Headcount Rates, by Benchmark, in Selected Developing Countries, 2000s
a. Income-based poverty headcount ratea
$4.00-a-day PPP c $2.50-a-day PPP $1.25-a-day PPP
Initial Final Total Annual Initial Final Total Annual Initial Final Total Annual
period period reduction change period period reduction change period period reduction change
(%) (%) (ppts) (%) (%) (%) (ppts) (%) (%) (%) (ppts) (%)
Argentina, 2000–10 27.5 14.6 −13.0 −6.2 14.2 6.4 −7.8 −7.7 5.1 1.8 −3.3 −9.8
Brazil, 2001–09 43.1 27.6 −15.5 −5.4 27.4 15.1 −12.3 −7.2 11.8 6.1 −5.7 −7.9
Cambodia, 2007–10b 75.1 75.7 0.5 0.2 57.3 60.1 2.8 1.6 34.0 29.6 −4.4 −4.5
Chile, 2000–09 23.2 11.8 −11.4 −7.2 9.0 4.3 −4.7 −7.9 2.3 1.3 −0.9 −5.6
Colombia, 2002–10 61.6 39.5 −22.1 −5.4 42.3 22.0 −20.3 −7.8 20.7 8.2 −12.6 −11.0
Costa Rica, 2000–08 29.2 18.9 −10.2 −5.3 14.7 7.6 −7.1 −7.9 5.5 2.4 −3.1 −9.8
Ecuador, 2003–10 51.5 33.4 −18.1 −6.0 31.5 15.9 −15.6 −9.3 12.2 4.6 −7.6 −13.0
Honduras, 1999–2009 66.1 52.1 −14.1 −2.4 47.9 36.2 −11.6 −2.7 24.9 17.8 −7.1 −3.3
Mongolia, 2007–11b 56.6 44.3 −12.3 −6.0 35.7 22.5 −13.2 −10.9 13.1 6.0 −7.1 −17.8
Panama, 2001–09 43.4 29.9 −13.5 −4.6 28.7 16.1 −12.6 −6.9 15.4 4.6 −10.8 −14.0
Paraguay, 1999–2010 43.3 32.8 −10.6 −2.5 26.7 18.4 −8.3 −3.3 14.0 7.2 −6.8 −5.9
Philippines, 2006–09b 78.3 77.9 −0.4 −0.2 62.2 60.3 −1.9 −1.0 31.7 27.6 −4.1 −4.5
Vietnam, 2004–10b 79.9 58.9 −21.0 −5.0 56.4 33.2 −23.2 −8.5 18.1 8.7 −9.4 −11.5
table continues next page
Table 3.1 Poverty Headcount Rates, by Benchmark, in Selected Developing Countries, 2000s (continued)
b. Consumption-based poverty headcount ratec
$4.00 or $5.00-a-day PPP d $2.50-a-day PPP $1.25-a-day PPP
Initial Final Total Annual Initial Final Total Annual Initial Final Total Annual
period period reduction change period period reduction change period period reduction change
(%) (%) (ppts) (%) (%) (%) (ppts) (%) (%) (%) (ppts) (%)
Bangladesh, 2000–10 — — — — 89.2 84.0 −5.2 −0.6 57.7 40.3 −17.4 −3.5
Ghana, 1998–2005 — — — — 71.8 58.3 −13.5 −2.9 38.2 23.5 −14.7 −6.7
Moldova, 2001–10 93.8 58.7 −35.1 −5.1 71.4 12.9 −58.5 −17.3 27.5 0.5 −27.0 −35.4
Nepal, 1996–2003 — — — — 94.3 84.9 −9.5 −1.5 54.0 25.9 −28.2 −10.0
Peru, 2004–10 45.8 30.0 −15.8 −6.8 22.9 11.7 −11.2 −10.6 3.5 0.8 −2.6 −21.1
Romania, 2001–09 75.3 33.2 −42.1 −9.7 23.7 4.2 −19.5 −19.6 2.6 0.0 −2.6 −100.0
Sri Lanka, 2002–09 84.7 78.09 −6.6 −1.2 61.9 45.1 −16.8 −4.4 14.8 4.7 −10.1 −15.1
Thailand, 2000–09 31.3 16.6 −14.7 −6.8 7.9 2.5 −5.3 −11.8 3.7 1.4 −2.3 −10.3
Sources: Data for Ghana and Nepal from FAO n.d. Data for Bangladesh, Moldova, Peru, Romania, and Thailand from national household surveys. Data for Cambodia, Mongolia, the Philippines, and Vietnam from the
World Bank’s East Asia and Pacific region harmonized household surveys.
Note: PPP = purchasing power parity; — = not available. Latin American countries typically measure poverty using a household income aggregate, while most other countries around the world use a consumption
aggregate. Because these measures are not comparable, we present them separately.
a. The decomposition analysis for countries listed in panel a use income-based poverty estimates.
b. The income-based figures for Cambodia, Mongolia, the Philippines, and Vietnam are not the official poverty figures because these countries (except the Philippines) use consumption-based poverty estimates.
However, the analysis was undertaken using income measures of poverty because of the large discrepancy between consumption and income in these countries.
c. The decomposition analysis for countries listed in panel b use consumption-based poverty estimates.
d. Moldova and Romania measure moderate poverty at rates close to $5-a-day, while Peru, Sri Lanka, and Thailand measure moderate poverty at rates close to $4-a day.
43
44 What Accounts for Changes in Poverty over the Past Decade?
Figure 3.1 Average Real GDP Growth in Selected Developing Countries, 2000s
9
8
Average annual growth (%)
7
6
5
4
3
2
1
0
0 0
11
Ho Ch 200 0
ur , 2 09
99 9
nt 200 9
Ph pal, 00 9
p 96 10
lo , 2 03
ua 20 9
st r, 2 –10
an 20 0
8
i L , 2 09
la , 2 09
na 20 0
Pe 00 0
na 200 9
on m, 2 –10
, 2 10
d 200 5
m a, 2 –1
01
Th , 19 0–0
0
Ne ina, 0–0
Ec ia, 6–0
Gh ica, 3–1
–0
Pa sh, 7–1
2 1
–0
M ia, 200
7–
nd ile 1–
Ar land –20
ilip 19 0–
Co ines –20
Sr ova 1–
ng dia 02–
–
lia 4–
Ca nk 001
Co do 02
00
m 00
1
4
az –2
00
as 00
00
0
de 00
go 00
an 8–
Br 99
aR 0
2
m 99
9
il,
,
a,
Vi ru,
,1
Ro a, 1
b
ay
Ba bo
m
a
ai
et
gu
ol
ge
ra
M
Pa
Source: World Bank 2013.
Note: GDP = gross domestic product.
dynamics. Moreover, they are jointly determined, such that cross-country esti-
mates are unlikely to shed much light on the fundamental factors underlying
distributional change (Ferreira 2012).
Therefore, instead of relying on summary measures of poverty, one could
better understand distributional changes by using full distributions of income or
consumption expenditures from representative household surveys. Instead of
focusing on economic growth—which can also be thought of as the propor-
tional change in the mean of the income distribution—it is best to analyze how
the entire distribution changes over time. Moreover, given the richness of data
available in household income and expenditure surveys, one can further disag-
gregate the observed distributional changes by decomposing the factors that
underlie these distributions. The rest of the chapter focuses on applying the
method discussed in detail in chapter 2 to further disaggregate these distribu-
tional changes.
Forces behind Poverty Reduction
We begin by noting that at least four factors could have influenced poverty
reduction:
• Demographic change, particularly the share of adults per household
• Growth in labor income, either because more people are employed or because
their earnings have increased
• Growth in nonlabor income, in the form of public or private transfers
• Changes in consumption or savings patterns
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
What Accounts for Changes in Poverty over the Past Decade? 45
Figure 3.2 Contribution of Growth and Redistribution to Poverty Reduction in Selected Developing
Countries, by Poverty Line, 2000s
Romania, 2001–09 (poverty−5.3 ppts/yr)
Moldova, 2001–10 (poverty−3.9 ppts/yr)
Costa Rica, 2000–08 (poverty−1.3 ppts/yr)
Countries, by poverty line, and average poverty reductions per year, ppts
Honduras, 1999–2009 (poverty−1.4 ppts/yr)
Thailand, 2000–09 (poverty−1.6 ppts/yr)
Peru, 2004–10 (poverty−2.6 ppts/yr)
Colombia, 2002–10 (poverty−2.8 ppts/yr)
Ecuador, 2003–10 (poverty−2.6 ppts/yr)
Panama, 2001–09 (poverty−1.7 ppts/yr)
Chile, 2000–09 (poverty−1.3 ppts/yr)
Brazil, 2001–09 (poverty−1.9 ppts/yr)
Paraguay, 1999–2010 (poverty−1.0 ppts/yr)
Argentina, 2000–10 (poverty−1.3 ppts/yr)
Vietnam, 2004–10 (poverty−1.6 ppts/yr)
Nepal, 1996–2003 (poverty−2.9 ppts/yr)
Ghana, 1998–2005 (poverty−2.1 ppts/yr)
Bangladesh, 2000–10 (poverty−1.7 ppts/yr)
Sri Lanka, 2002–09 (poverty−1.4 ppts/yr)
Philippines, 2006–09 (poverty−1.3 ppts/yr)
Mongolia, 2007–11 (poverty−1.8 ppts/yr)
–40 0 40 80 120 160
Contributions to total poverty reduction (%)
Growth Redistribution
Sources: Data from SEDLAC, various years; FAO n.d.; and national household surveys.
Note: ppts = percentage points; PPP = purchasing power parity. Countries are listed in order of growth’s contribution to the decline of poverty.
“Growth” refers to the distribution neutral mean growth of household consumption (also known as the size effect, in which the Lorenz curve does
not change), and the “redistribution” effect measures the change in poverty resulting from a change in the Lorenz curve while holding the mean
constant. Shapely-Shorrocks estimates are used, see chapter 2 for a full explanation.
Demographic Change
Demographic change could play a role by affecting the dependency ratio: the
number of earners relative to the number of consumers in a household. Among
the countries considered here, the population of Bangladesh grew by 25 percent
between 2000 and 2010, adding 19 million people to its total, while Brazil
has added 18 million (a 16 percent increase) during the same time period.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
46 What Accounts for Changes in Poverty over the Past Decade?
Figure 3.3 Change in Age-Dependency Ratio of Selected Developing Countries, 2000s
100
90
Dependents per 100 working-age
80
70
60
people (%)
50
40
30
20
10
0
n 200 0
20 09
on , 2 10
Ch 20 9
a R , 20 11
an 00 9
B , 20 08
lo l, 2 09
, 2 09
Ba nam 00 0
de 200 0
Pe 200 9
20 10
Pa bod 200 0
gu ia, 2 –10
Ho ppi 99– 0
ur , 2 10
an 99 9
pa 98 9
03
6– 5
d, 1–1
0
0
Pa na, 2–1
a, 0–1
0
–1
19 –1
Gh , 19 6–0
Ne , 19 200
99 00
Ro am, 0–
M nia 04–
lia 1–
ile 07–
Sr ica, 0–
ka 0–
Co razi 02–
Ar bia 01–
sh 1–
–
nd es 20
20
0
Ca ado 04
3
Ph ay, 007
l, 1 –2
00
go 00
0
00
as 00
–
0
,2
2
2
,
,
Ec u,
m r,
va
r
n
n
do
i
a
la
a
nt
m
m
ai
et
u
iL
la
ol
ge
st
Th
Vi
ng
ili
M
ra
Co
Initial Final
Source: World Bank 2012.
Note: The age-dependency ratio is the ratio of dependents (people younger than 15 or older than 64) to the working-age population (those aged
15–64 years). Data are shown as the proportion of dependents per 100 working-age population.
Despite these increases in population, the rate of population growth has deceler-
ated enough to begin shrinking the proportion of dependents relative to the
working-age population across almost all countries in our sample, as shown in
figure 3.3. The exception is Sri Lanka, which has seen higher births after the end
of many years of conflict in 2009.
Similarly, taking into account the fact that the elderly often continue to work
beyond age 64, the share of adults (ages 15 and above) per household has
increased in all countries in the sample (table 3.2). However, this overall similar-
ity still masks the heterogeneity across households within each country. Most
important, the share of adults per poor household decreased in Mongolia, Peru,
Romania, Colombia, Panama, and Moldova—potentially pointing to an unequal-
izing force in these countries.
Growth in Labor Income
Growth in labor income could be the main driver in the observed poverty
reductions—either because of higher employment rates or because of increased
earnings. The share of occupied (working) adults per household increased in
12 of the 21 countries in our sample, as table 3.3 illustrates. In some countries,
increased female employment was an important factor. For example, in Costa Rica,
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
What Accounts for Changes in Poverty over the Past Decade? 47
Table 3.2 Share of Adults per Household, by Poverty Level, Selected Developing Countries, 2000s
Poor households Poor households
All households (≤ $1.25-a-day PPP) (≤ $2.50-a-day PPP)
Annual Annual Annual
Initial Final average Initial Final average Initial Final average
year year change year year change year year change
Country and years of data (%) (%) (%) (%) (%) (%) (%) (%) (%)
Countries measuring poverty based on income
Argentina, 2000–10 72.3 76.2 0.5 51.5 62.5 2.0 53.5 57.7 0.8
Brazil, 2001–09 71.3 75.7 0.8 50.9 56.8 1.4 55.0 55.7 0.2
Chile, 2000–09 72.6 78.0 0.8 57.5 68.4 1.9 58.8 65.6 1.2
Colombia, 2002–10 68.0 71.1 0.6 62.1 59.6 −0.5 61.4 59.1 −0.5
Costa Rica, 2000–08 67.1 73.3 1.1 58.1 64.3 1.3 56.3 60.0 0.8
Ecuador, 2003–10 66.2 72.1 1.2 56.4 61.5 1.2 56.8 60.4 0.9
Honduras, 1999–2009 57.2 63.5 1.1 48.7 53.5 0.9 50.3 55.3 1.0
Panama, 2001–09 68.1 70.0 0.3 53.7 52.2 −0.3 56.0 54.0 −0.5
Paraguay, 1999–2010 59.8 68.2 1.2 46.3 57.2 1.9 48.6 56.6 1.4
Countries measuring poverty based on consumption
Bangladesh, 2000–10 60.4 65.3 0.8 55.4 58.4 0.5 58.9 63.6 0.8
Cambodia, 2007–10 66.3 68.6 1.1 60.8 61.0 0.1 62.2 64.3 1.1
Ghana, 1998–2005 56.1 60.1 1.0 49.5 51.3 0.5 52.3 54.2 0.5
Moldova, 2001–10 78.8 81.8 0.4 73.1 68.1 −0.8 76.4 73.5 −0.4
Mongolia, 2007–11 72.5 72.9 0.1 64.9 60.4 −1.8 66.0 62.0 −1.5
Nepal, 1996–2003 57.9 61.1 0.8 55.1 55.0 0.0 57.0 58.6 0.4
Peru, 2004–10 68.3 69.2 0.2 51.5 51.2 −0.1 55.0 50.8 −1.3
Philippines, 2006–09 65.4 67.5 1.1 55.1 57.2 1.3 59.7 61.8 1.2
Romania, 2001–09 82.2 84.8 0.4 60.9 — — 71.9 67.1 −0.9
Sri Lanka, 2002–09 74.2 76.3 0.4 65.3 68.8 0.7 71.3 73.1 0.4
Thailand, 2000–09 74.2 77.8 0.5 56.9 53.6 −0.7 61.4 65.6 0.7
Vietnam, 2004–10 73.4 76.3 0.7 64.6 67.6 0.8 69.6 70.9 0.3
Sources: Data from SEDLAC, various years; FAO n.d.; World Bank East Asia and Pacific regional unit household surveys and national household
surveys.
Note: — = not available; PPP = purchasing power parity.
both labor force participation and employment of women increased by about
percent between 2000 and 2008 (World Bank 2012).
23
However, this trend is not homogeneous across countries or within them.
Indeed, the share of working adults (15 years of age or older) per household
declined (in order of most to least) in Mongolia, Sri Lanka, Ghana, and Honduras
across the distribution. Such a change could be positive if the main breadwinners
were earning higher incomes, thus enabling the youths to continue in school and
the older adults to retire. However, even within countries, the data show some
important differences across the distribution in how the share of working adults
has changed over time. For example, the share of working adults declined signifi-
cantly among the poor in Brazil, Chile, and Costa Rica, even though these
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
48 What Accounts for Changes in Poverty over the Past Decade?
Table 3.3 Share of Working Adults per Household, by Poverty Level, in Selected Developing
Countries, 2000s
Average of all Poor households Poor households
households (≤ $1.25-a-day PPP) (≤ $2.50-a-day PPP)
Annual Annual Annual
Initial Final average Initial Final average Initial Final average
Year year change Year year change Year year change
Country and years of data (%) (%) (%) (%) (%) (%) (%) (%) (%)
Countries measuring poverty based on income
Argentina, 2000–10 48.8 55.7 1.3 26.3 27.0 0.3 32.8 34.7 0.6
Brazil, 2001–09 60.4 62.7 0.5 42.8 34.8 −2.6 50.9 46.7 −1.1
Chile, 2000–09 49.4 49.6 0.1 14.9 6.9 −8.3 27.4 15.7 −6.0
Colombia, 2002–10 55.1 60.5 1.2 40.2 39.1 −0.3 46.0 46.2 0.1
Costa Rica, 2000–08 53.7 56.9 0.7 22.5 16.5 −3.8 31.4 26.3 −2.2
Ecuador, 2003–10 61.5 59.4 −0.5 49.4 50.7 0.4 53.7 51.7 −0.5
Honduras, 1999–2009 63.2 59.6 −0.6 55.1 41.3 −2.8 57.7 50.3 −1.4
Panama, 2001–09 51.7 59.7 1.8 45.7 50.3 1.2 44.6 51.8 1.9
Paraguay, 1999–2010 61.7 64.1 0.3 51.2 49.1 −0.4 53.6 53.8 0.0
Countries measuring poverty based on consumption
Bangladesh, 2000–10 46.9 48.2 0.3 49.7 50.5 0.2 47.8 49.1 0.3
Cambodia, 2007–10 81.3 85.1 1.5 81.9 84.6 1.1 81.8 85.0 1.3
Ghana, 1998–2005 41.4 39.2 −0.8 38.1 32.0 −2.5 39.3 35.6 −1.4
Moldova, 2001–10 65.3 66.1 0.1 64.7 76.0 1.8 64.1 71.0 1.1
Mongolia, 2007–11 62.8 54.9 −3.3 55.7 45.9 −4.7 61.4 52.7 −3.8
Nepal, 1996–2003 32.7 33.8 0.5 32.7 34.0 0.6 32.7 34.2 0.6
Peru, 2004–10 69.4 72.5 0.7 82.5 78.3 −0.9 79.1 80.1 0.2
Philippines, 2006–09 59.5 58.7 −0.4 64.1 63.6 −0.3 62.1 61.4 −0.4
Romania, 2001–09 83.2 82.1 −0.2 79.5 n.a. n.a. 83.6 74.1 −1.5
Sri Lanka, 2002–09 53.7 49.6 −1.1 53.1 45.0 −2.3 53.8 49.4 −1.2
Thailand, 2000–09 74.7 74.4 −0.1 60.6 72.0 1.9 78.2 71.3 −1.0
Vietnam, 2004–10 79.1 78.5 −0.1 50.8 52.2 0.5 54.7 54.9 0.1
Sources: Data from SEDLAC, various years; FAO n.d.: World Bank East Asia Pacific regional unit harmonized household surveys and national
household surveys.
Note: “Working adults” are defined as household members (aged 15–64 years) who are occupied in work for pay. PPP = purchasing power parity.
ountries had a higher share of working adults on average—indicating that the
c
increasing share of working adults was occurring mostly at the top end of the
distribution.
In most of the countries, however, the rise of working-adult populations
might not only indicate an increase in workers per household but also reflect
evidence that labor incomes per adult increased at the bottom of the distri-
bution in many of these countries. Unfortunately, we cannot determine
whether the poor are earning more because of higher hourly wages or
because of more hours worked. In any case, labor incomes have usually
increased among the poor.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
What Accounts for Changes in Poverty over the Past Decade? 49
Figure 3.4 Change in Subsidies and Other Social Transfers in Selected Developing Countries, 2000s
25
Subsidies and transfers, % of GDP
20
15
10
5
0
09 a
09 a
09 a
9a
08 a
09 a
9
09
11
0
09
9
5
–0
–1
–0
00
2–
7–
–0
0–
06
0–
20
4–
2–
07
01
1–
–2
00
00
02
00
00
20
00
00
9–
20
20
00
98
,2
,2
20
,2
,2
99
,2
,2
,2
19
s,
a,
il,
ka
lia
ile
a,
ne
az
sh
di
ru
ia
va
,1
a,
an
go
bi
Ch
an
bo
Pe
Br
pi
de
do
ay
an
m
iL
on
ilip
m
m
gu
la
ol
lo
Gh
Sr
Ro
M
ng
Ca
M
Co
Ph
ra
Ba
Pa
Initial Final
Source: World Bank 2012.
Note: GDP = gross domestic product.
a. The reported subsidy and transfer data are for a year that is one year earlier than the household survey.
Growth in Nonlabor Income
A third major factor in poverty reduction could have been strongly related to
growth in nonlabor incomes. For example, for countries where data is readily
available, figure 3.4 shows that public subsidies and other social transfers have
increased in several countries over the past decade. Government spending for
subsidies and transfers increased more than sixfold in Ghana and by more than
60 percent as a share of GDP in Bangladesh, Moldova, Mongolia, and Romania.
In addition to public sources of transfers, private transfers, such as remittances,
have grown strongly in many of the countries (figure 3.5), as these cases
exemplify:
• In Nepal, remittances grew from 1 percent of GDP in 1996 to 12 percent
in 2003.
• In Moldova, they increased from 14 percent of GDP in 2001 to 22 percent
in 2010.
• In Honduras, they nearly tripled, going from 6 percent of GDP in 1999 to
17.8 percent in 2009.
How important were these increased transfers—whether public or private—
to poverty reduction? For instance, if remittances mostly reached higher-income
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
50 What Accounts for Changes in Poverty over the Past Decade?
Figure 3.5 Change in International Remittances to Selected Developing Countries, 2000s
25
20
Remittances, % of GDP
15
10
5
0
10
0 0
na 0 0
Ba Lan , 20 –09
Ph lade , 20 10
pi , 20 –09
nd al, 006 0
9
do 9 3
00 9
Br , 20 –09
la 0 0
na , 20 –09
Ca na, 200 9
Co od 98– 9
Co e 200 10
M Ric 000 0
Pa g 20 –10
a 2 08
Ro dor 9–2 1
m 20 05
ua 99 –1
m , 2 01
et , 2 –1
Ho p , 2 –1
ur 199 –0
ol 99 00
, 2 00
ai l, 2 –1
–0
0
2 1
1–
–
m 19 1–
a, 7–
st ru, 2–
–
lo ia, 20
in 00
Th azi 00
Pa nd 01
Gh ma 00
ra oli 00
Ec y, 1 007
Vi nia 03
Sr m 01
ng ka 04
ilip sh 02
Ne nes 00
M s, 1 6–2
va –2
bi 0
nt 20
ge le,
,
on a,
gu a,
a
Ar Chi
a
P
a
b
a
a
i
Initial Final
Source: World Bank 2012.
Note: GDP = gross domestic product.
households, one would not expect poverty to decline. As table 3.4 shows, most
countries in our sample did see an increase in transfers as a share of total
household income, but the impact of these among the poor varied. Transfers
were especially important for poor households in Argentina, Brazil, Colombia,
Moldova, and Peru. However, the share of transfers in total household income
actually declined among poor households in some countries (Cambodia, Ghana,
Honduras, and Sri Lanka), partly reflecting relatively weaker public social protec-
tion programs, but also highlighting the fact that remittances do not always help
the poor (table 3.4).
Changes in Consumption or Savings Patterns
Finally, in the absence of measurement error, changes in consumption-based
poverty could also be related to changes in consumption and savings patterns. In
the context of growing incomes, households could either increase consumption
proportionately or they could increase their savings. However, given measure-
ment errors in income and expenditure aggregates in household surveys, it is
difficult to differentiate between household consumption changes resulting from
real behavioral shifts and those resulting from measurement errors.
Whatever the cause—or causes—of apparent household consumption changes,
table 3.5 shows a full range of patterns.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
What Accounts for Changes in Poverty over the Past Decade? 51
Table 3.4 Share of Transfers in Total Household Income, by Poverty Level, in Selected Developing
Countries, 2000s
Poor households Poor households
All households (≤ $1.25-a-day PPP) (≤ $2.50-a-day PPP)
Annual Annual Annual
Initial Final average Initial Final average Initial Final average
Year year change Year year change Year year change
Country and years of data (%) (%) (%) (%) (%) (%) (%) (%) (%)
Countries measuring poverty based on income
Argentina, 2000–10 4.7 6.5 3.3 12.7 48.9 14.4 8.7 32.9 14.3
Brazil, 2001–09 1.2 5.8 21.3 4.9 47.0 32.6 2.8 24.8 31.1
Chile, 2000–09 — 6.4 — — 42.3 — — 27.7 —
Colombia, 2002–10 4.7 9.2 8.6 9.2 30.8 16.4 6.8 20.7 14.9
Costa Rica, 2000–08 — 4.5 — — 28.6 — — 19.8 —
Ecuador, 2003–10 9.0 10.6 2.3 24.8 35.3 5.1 15.1 24.4 7.1
Honduras, 1999–2009 9.1 7.1 –2.5 11.9 3.7 –11.0 10.1 2.8 –12.0
Panama, 2001–09 10.1 12.3 2.4 26.1 41.2 5.9 21.8 32.3 5.0
Paraguay, 1999–2010 8.4 9.0 0.7 8.3 16.9 6.7 10.4 16.3 4.2
Countries measuring poverty based on consumption
Bangladesh, 2000–10 2.7 3.5 2.5 1.8 2.0 0.8 2.5 3.1 2.1
Cambodia, 2007–10 2.6 1.7 –12.6 3.2 2.6 –7.4 3.0 2.1 –11.7
Ghana, 1998–2005 5.1 5.1 0.1 3.6 2.8 –3.2 3.8 3.8 –0.2
Moldova, 2001–10 4.5 24.0 20.5 7.2 22.5 13.5 6.2 19.0 13.2
Mongolia, 2007–11 25.7 29.3 3.4 49.4 69.9 9.1 35.4 48.1 8.0
Nepal, 1996–2003 3.0 4.9 7.5 2.7 3.7 4.5 2.9 4.7 7.0
Peru, 2004–10 7.6 5.4 –5.4 4.2 10.4 16.3 3.9 8.5 14.0
Philippines, 2006–09 21.1 22.5 2.2 23.0 25.1 2.9 19.9 21.3 2.3
Romania, 2001–09 8.1 8.7 0.8 29.7 … — 19.4 33.8 7.2
Sri Lanka, 2002–09 2.2 2.1 –0.2 3.2 1.8 –8.0 2.3 1.8 –3.6
Thailand, 2000–09 9.9 10.7 0.9 18.6 19.1 0.3 11.7 14.3 2.2
Vietnam, 2004–10 11.1 9.6 –2.5 9.8 10.8 1.6 9.7 10.6 1.6
Sources: Data from SEDLAC, various years; FAO n.d.: World Bank’s East Asia Pacific regional unit harmonized household surveys and national
household surveys.
Note: — = not available; … = negligible; PPP = purchasing power parity.
• In Bangladesh, Ghana, and Peru, the consumption-to-income ratio increased
among households at the bottom of the income distribution, while it fell
among those at the top.
• In Nepal, Sri Lanka, and Thailand, this ratio remained more or less flat across
the distribution.
• In Moldova and Romania, the consumption-to-income ratio fell more among
households at the bottom of the income distribution than it did among those
at the top.
Although we cannot differentiate between measurement error and changing
behavior as drivers of change in the consumption-to-income ratios, we can show
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
52
Table 3.5 Change in Household Consumption-to-Income Ratio in Selected Developing Countries, 2000s
Bangladesh Ghana Moldova Nepal Peru Romania Sri Lanka Thailand
2000–10 1998–2005 2001–10 1996–2003 2005–09 2001–09 2002–09 2000–09
Annual Annual Annual Annual Annual Annual Annual Annual
Initial Final average Initial Final average Initial Final average Initial Final average Initia Final average Initial Final average Initial Final average Initial Final average
Income year year change year year change year year change year year change year year change year year change year year change year year change
decile (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%)
1 1.8 2.3 3 22.6 34.8 6 15.3 18.4 2 3.3 2.9 –2 1.6 1.9 5 4.51 2.37 –8 3.1 3.0 0 1.5 1.2 –3
2 1.2 1.5 2 5.9 8.3 5 4.0 1.9 –8 1.9 1.6 –2 1.3 1.4 2 1.89 1.16 –6 1.5 1.5 0 1.1 1.0 –2
3 1.0 1.3 2 3.3 4.6 5 2.9 1.5 –7 1.5 1.4 –1 1.2 1.2 1 1.44 0.99 –5 1.3 1.3 0 1.0 0.9 –1
4 1.0 1.1 1 2.5 3.4 5 2.2 1.3 –6 1.2 1.2 0 1.2 1.1 –1 1.26 0.88 –4 1.2 1.2 1 0.9 0.9 –1
5 1.0 1.0 0 2.2 2.5 2 1.8 1.2 –5 1.1 1.1 0 1.1 1.0 –1 1.10 0.82 –4 1.1 1.1 –1 0.9 0.8 –1
6 1.0 0.9 –1 1.7 2.2 3 1.6 1.1 –4 1.0 0.9 –1 1.1 1.0 –2 1.01 0.74 –4 1.0 1.0 0 0.8 0.8 0
7 1.0 0.8 –2 1.5 1.6 0 1.3 1.0 –3 1.0 1.0 0 1.0 0.9 –1 0.92 0.71 –3 1.0 0.9 –1 0.8 0.8 0
8 0.9 0.7 –2 1.5 1.2 –2 1.1 0.9 –2 0.9 0.9 0 0.9 0.9 –2 0.85 0.65 –3 0.9 0.9 –1 0.7 0.7 0
9 0.9 0.6 –4 1.3 1.0 –3 0.9 0.9 –1 0.9 0.8 –1 0.8 0.8 –1 0.77 0.60 –3 0.8 0.8 –1 0.6 0.6 0
10 0.8 0.5 –5 0.8 0.5 –6 0.7 0.7 0 0.8 0.8 –1 0.7 0.7 0 0.64 0.49 –3 0.7 0.5 –3 0.5 0.5 0
Sources: FAO n.d.; national household surveys.
Note: The eight countries measured represent those out of the 21-country sample set that measure poverty by consumption rather than by income.
What Accounts for Changes in Poverty over the Past Decade? 53
how changes in this ratio affect overall changes in poverty. In short, each source
of change described above could have contributed to the observed reductions in
poverty over the past decade. How large of a contribution each of these forces
made is the next question.
Results
Main Driver of Poverty Reduction: Labor Income Growth
Which factor—demographics, labor income, public transfers, or remittances—
contributed the most to observed reductions in poverty? One key result stands
out: the most important contributor to poverty reduction has been the growth
in earnings, as a result of higher employment and higher labor incomes.
In 12 of the 21 countries with substantial declines in poverty, the combined
effect of higher labor income and higher employment rates explain more than
half of the poverty reduction; in another 6 countries, those two factors account
for more than 40 percent of the reduction (as shown in figure 3.6 and table 3.6).
This result holds true regardless of the decomposition path taken, as these results
are an average of every possible decomposition path, as described in chapter 2,
following Shapley (1953) and Shorrocks (1999).
Interestingly, in most cases, the greatest poverty reducer throughout the 2000s
was labor income growth, rather than an increase in the share of working adults.
Indeed, such growth alone cut poverty by more than half in 10 out of the
21 countries. In contrast, growth in the share of employed adults cut poverty by
more than 20 percent in only 4 countries.
Importance of Demographic Changes
Although increases in labor income are the main contributors to reductions in
poverty in most countries, demographics also mattered. In particular, a higher
share of working-age adults in the household made the largest contribution to
poverty reduction in Costa Rica, Paraguay, and the Philippines, but it was also
important in Bangladesh, Cambodia, Chile, and Honduras.
Changes in the share of adults per household were also relatively important
in explaining declines in moderate poverty in Bangladesh, Cambodia, Chile,
Costa Rica, Ecuador, Honduras, the Philippines, and Paraguay (as shown in and
figure 3.6 and table 3.6). In contrast, in Mongolia, the share of adults per house-
hold among the poor decreased, therefore increasing poverty and pointing to an
unequalizing force there (table 3.6).
In general, increased percentages of working-age adults contributed positively
to poverty reduction. However, the magnitude of this effect is small relative to
the effect of labor income growth.
Roles of Employment Growth and Labor Income
Earnings growth, resulting from higher employment and higher labor incomes,
was the most important contributor to poverty reduction. However, of these two
elements, it was the growth in labor incomes that made most of the difference.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
54 What Accounts for Changes in Poverty over the Past Decade?
Figure 3.6 Decomposition of Changes in Moderate Poverty, by Level, in Selected Developing
Countries, 2000s
140
120
Contribution to change in poverty (%)
100
80
60
40
20
0
–20
–40
1
0
03
9
09
0
09
10
8
10
10
9
9
0
–09
0
9
05
–10
09
0
1
0–1
4–1
0–0
200
1–0
201
1–0
7–1
6–0
4–1
07–
–20
02–
00–
02–
01–
03–
–20
01–
000
000
200
, 200
, 200
200
, 200
0
200
99–
99–
, 200
a, 20
996
a, 20
e, 20
a, 20
a, 20
r, 20
0
998
il, 20
d, 2
ia, 2
na, 2
s, 19
y, 19
esh,
ica,
nes,
ania
nam
ama
Peru
al, 1
na, 1
ado
goli
ank
mbi
dov
Chil
Braz
ilan
bod
enti
ta R
dura
glad
ippi
gua
Nep
Rom
Pan
Viet
Ecu
Sri L
Mon
Tha
Mol
Gha
Colo
Cam
Arg
Cos
Phil
Para
Ban
Hon
≤$1.25-a-day PPP $4.00-a-day PPP $5.00-a
-day PPP
Countries, by poverty line
Nonlabor income Share of working-age family members
Employment + earnings Consumption-to-income ratio
Sources: Data from SEDLAC, various years; FAO n.d.; and national household surveys.
Note: PPP = purchasing power parity. “Nonlabor income” refers to public and private transfers (including remittances), pensions, capital, and other
nonlabor income. “Employment + earnings” refers to the combination of increased employment among working-age adults (aged 15–64 years)
and increases in their earnings. Consumption-based measures of poverty are used in Bangladesh, Ghana, Nepal, Peru, Thailand, Moldova, and
Romania. The remaining countries use income-based measures of poverty.
Although higher employment typically contributed to poverty reduction, lower
employment among the poor partially offset poverty reduction in some countries
(table 3.6). For instance, as shown in table 3.3, employment declined among the
poor in more than one-third of the countries in our sample: Ecuador, Ghana,
Honduras, Mongolia, the Philippines, Sri Lanka, Thailand, and Vietnam. It is no
wonder that reductions in employment actually offset poverty reduction in these
countries (table 3.6).
Overall, the increase in workers’ earnings was relatively more important in
reducing poverty than the increase in the number of workers or in the number
of jobs. Although we cannot disentangle whether earnings increased because
of improvements in the quality of jobs, changes in productivity, or simply
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Table 3.6 Contributions to Declines in Moderate Poverty, by Level, in Selected Developing Countries, 2000s
≤ $1.25-a-day PPP ≤$4.00-a-day PPP ≤$5.00-a-day PPP
Honduras, 1999–2009
Paraguay, 1999–2010
Bangladesh, 2000–10
Philippines, 2006–09
Cambodia, 2007–10
Costa Rica, 2000–08
Argentina, 2000–10
Colombia, 2002–10
Mongolia, 2007–11
Ghana, 1998–2005
Romania, 2001–09
Sri Lanka, 2002–09
Thailand, 2000–09
Moldova, 2001–10
Panama, 2001–09
Vietnam, 2004–10
Nepal, 1996–2003
Ecuador, 2003–10
Brazil, 2001–09
Chile, 2000–09
Peru, 2004–10
Poverty rate
Initial period (%) 57.7 34.0 38.2 13.1 73.7 31.7 14.8 18.1 27.5 43.1 23.2 61.6 29.2 51.5 66.1 43.4 43.3 45.8 31.4 93.8 75.3
Final period (%) 40.3 29.6 23.5 6.0 53.1 27.6 4.7 8.7 14.6 27.6 11.8 39.5 18.9 33.4 52.1 29.9 32.8 30.0 16.6 58.7 33.2
Total change (ppts) –17.4 –4.4 –14.7 –7.1 –20.6 –4.1 –10.1 –9.4 –13.0 –15.5 –11.4 –22.1 –10.2 –18.1 –14.1 –13.5 –17.3 –15.8 –14.8 –35.1 –42.2
Contributions to poverty reduction
Consumption-to-
income ratio (%) –25.6 n.a. 17.7 n.a. –6.9 n.a. –0.19 n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. –0.4 –26.4 –20.4 16.0
Share of adults in the
household (%) 37.2 34.9 21.4 –1.0 14.2 47.6 19.5 15.9 22.0 16.4 31.0 12.1 34.4 27.1 32.0 13.5 59.5 1.8 22.8 6.9 25.1
Share of occupied
(working) adults (%) 20.5 33.3 –3.6 –18.2 13.4 –0.9 –21.8 –4.3 16.7 10.9 –0.1 15.8 14.2 –3.3 –4.1 29.1 11.6 17.0 –9.7 –0.4 21.7
Labor income
(earnings) per
adult (%) 60.2 46.4 49.6 62.6 50.4 23.2 62.9 75.4 35.2 41.6 48.2 38.3 17.7 53.3 55.8 30.2 33.2 67.5 80.9 44.2 –10.9
Capital (%) 7.8 — 5.0 — 5.9 — 0.0 — –5.4 –0.7 –4.2 4.3 3.9 –0.7 3.8 –0.2 0.0 3.4 1.5 1.0 21.6
Pension (%) 0.0 –5.2 0.0 3.2 0.0 2.2 1.5 2.9 14.9 17.9 15.8 4.6 23.6 6.2 3.2 10.4 –3.0 16.0 36.8 23.0 6.3
Transfers (%) 15.7 –8.1 10.0 52.2 23.0 27.6 –1.6 8.0 7.3 9.1 41.8 15.6 22.4 13.6 3.4 16.4 –1.0 –0.9 2.3 35.0 9.7
Other nonlabor
income (%) –15.9 –1.5 0.0 1.2 0.0 0.2 39.8 2.1 9.3 4.8 –32.5 9.4 –16.1 3.8 6.0 0.5 –0.3 –4.3 –8.2 10.7 10.6
Total contributions (%) 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0
Sources: SEDLAC, various years; FAO n.d.; World Bank harmonized household surveys for East Asia and Pacific countries and national household surveys.
Note: PPP = purchasing power parity; ppts = percentage points; — = not available; n.a. = not applicable. “Occupied (working) adults” is defined as those 15–64 years of age. Consumption-based measures of
poverty are used in Bangladesh, Ghana, Moldova, Nepal, Peru, Romania, and Thailand. Income-based measures of poverty are used in Argentina, Brazil, Cambodia, Chile, Colombia, Costa Rica, Ecuador, Honduras,
Mongolia, Panama, Paraguay, the Philippines, and Vietnam.
55
56 What Accounts for Changes in Poverty over the Past Decade?
longer hours, we can conclude that it was workers’ labor income growth that
made the difference rather than higher employment rates.
Roles of Nonlabor Income and Consumption
Although nonlabor income, such as public and private transfers, was also impor-
tant, its contribution toward declines in moderate poverty in most of the sample
countries was relatively small. The exceptions were Moldova, Mongolia, and
Romania, where transfers contributed relatively more to reductions in poverty
than they did elsewhere. In Romania, this was related to changes in transfers and
capital income; in Mongolia, to a significant increase in social transfers with the
introduction of the Human Development Fund; and in Moldova, mostly to
increased international remittances.
Finally, for countries where poverty is measured by consumption, these
decompositions suggest that reductions in the consumption-to-income ratio gen-
erally helped to reduce poverty in Ghana and Romania, where the consumption-
income ratio increased at the bottom of the distribution. However, in all other
instances, decreases in the consumption-to-income ratio during the past decade
did not reduce poverty as much as it would have if consumption had remained
a constant share of income.4
Further Decompositions by Poverty Line
Annex 3B presents more-detailed decompositions of poverty reductions in the
sample countries according to different international poverty lines ($1.25 a day,
in table 3B.1; $2.50 a day, in table 3B.2; and $4.00–$5.00 a day, in table 3B.3).
Moreover, each of the annex 3B tables includes decompositions along three
dimensions defined by Foster, Greer, and Thorbecke (1984):
• The poverty headcount rate (the proportion of the population that is poor),
represented as FGT0
• The poverty gap (the distance between the incomes of the poor and the pov-
erty line), represented as FGT15
• The poverty severity (a measure that gives greater weight to those furthest
away from the poverty line), represented as FGT26
For middle-income countries, the $2.50-a-day poverty line is close to national
extreme poverty lines. In such cases, nonlabor incomes are relatively more impor-
tant (table 3B.1) because transfers play a larger role in poverty reduction.7 For
instance, transfers account for about 87 percent in Chile; 58 percent of poverty
reduction in Thailand; 40 percent in the Philippines; and 20–30 percent in
Colombia, Costa Rica, Ecuador, Moldova, and Panama.
Regardless of which poverty line is used, transfers play a more important role
for those who live the furthest below the poverty line, the extreme poor. Transfers
account for a greater share of the declines in both the poverty gap (FGT1)
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
What Accounts for Changes in Poverty over the Past Decade? 57
and poverty severity (FGT2), as the data show in tables 3B.1, 3B.2, and 3B.3.
Particularly in Argentina, Brazil, Chile, Costa Rica, Moldova, the Philippines,
Thailand, and Vietnam, increases in cash transfers and pensions jointly account
for a larger share of the decline in poverty severity than do changes in labor
income (table 3B.2). This finding is consistent with the improvements those
countries have made in their social protection systems, which are typically tar-
geted at the bottom of the distribution and have increased in performance over
the past decade (World Bank 2013).8
Summary and Conclusions
This chapter has implemented a distinct approach to account for the relative
contributions of demographics, labor income, and nonlabor income to the signifi-
cant poverty reductions observed in 21 countries during the past decade. In
contrast to methods that focus on aggregate summary statistics as detailed in
chapter 2, the method adopted here generates entire counterfactual distributions,
allowing us to identify the contributions of these factors to observed distribu-
tional changes. Another contribution is the chapter’s application of the
best-known remedy for path dependence, which is to calculate the decomposi-
verage among them.
tion across all possible paths and then take the a
From Useful Findings, More Questions
For most of the 21 countries in our sample, the most important contributor to
reductions in moderate poverty has been the growth in labor income. In particu-
lar, among 12 of them with substantial declines in poverty, changes in labor
income and employment explain more than half of the change in poverty; in
another 6 countries, the same changes account for more than 40 percent of
reduced poverty.
Demographic changes had their role: the swelling numbers of working-age
adults translated into larger numbers of occupied adults per household, and the
increased employment reduced poverty. As the data confirmed, however, it was
not so much increased employment but increased earnings per occupied adult
that made the largest contribution to poverty reduction. Although we cannot
distinguish whether the earnings increased because of higher hourly wages,
better-quality jobs, higher productivity, or greater number of hours worked, the
point is that higher labor incomes appear to be the key factor behind reductions
in poverty observed in the past decade.
Declining dependency ratios also pointed to the importance of demographic
changes in alleviating poverty, especially in Bangladesh, Cambodia, Chile, Costa
Rica, Honduras, Paraguay, and the Philippines. In most cases, however, these
effects were smaller than the effect of labor income growth.
As for employment rates, we find that overall employment gains typically
contributed to poverty reduction, but in places where employment decreased
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
58 What Accounts for Changes in Poverty over the Past Decade?
specifically among the poor, poverty naturally increased. Even in cases where
employment growth helped to reduce poverty, the increase in workers’ earnings
had relatively more impact on reducing poverty than the increase in the share of
employed adults.
Finally, both public and private transfers contributed significantly to reduction
of moderate poverty, albeit less so than labor income growth. They became far
more important when accounting for changes in extreme poverty. From decom-
position of the extreme-poverty headcount, poverty gap, and poverty severity
(table 3B.1), we find that transfers and pensions contributed a relatively higher
share than labor income growth. These results point to the crucial role that social
protection systems play in reaching the extreme poor.
Limitations and Next Steps
The decomposition method applied here is quite useful for distinguishing the
main contributors to poverty reduction. Its main limitation is that it cannot
shed light on why workers’ earnings increased. Any number of alternatives arise:
Did earnings increase because of changes in the endowments of the population,
such as higher educational levels or increases in other productive assets? Did
marketplace premiums rise for workers with those endowments? In many
developing countries, poverty reduction has coincided with the labor force’s
increasing education and health—as well as, in some cases, more equitable dis-
tribution of land or other productive assets. Therefore, such distinctions as the
effects of changing endowments—and the returns from those endowments—
are important if we are to fully understand changes in poverty and how to
further reduce it.
To resolve this issue, we must consider alternative decomposition techniques
that impose an underlying labor model and greater structure than the nonpara-
metric approach adopted here (Juhn, Murphy, and Pierce 1993; Bourguignon,
Ferreira, and Lustig 2005; Bourguignon, Ferreira, and Leite 2008). It is to this task
that the remainder of the book is dedicated.
Annex 3A: Data Sources
For the data on Latin American countries, we use the SEDLAC dataset,
which covers all countries in mainland Latin America and four of the largest
countries in the Caribbean. Most household surveys included in the sample
are nationally representative. For comparability purposes, this dataset com-
putes income using a common method across countries and years. In particu-
lar, it constructs a common household income variable that includes all the
ordinary sources of income and estimates of the imputed rent from home
ownership. (For further methodological details, see the “Methodology” link on
the SEDLAC website: http://sedlac.econo.unlp.edu.ar/eng/.) Only the Latin
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
What Accounts for Changes in Poverty over the Past Decade? 59
American countries have available data on hours of work, allowing for further
decomposition in the report.
For the data on Bangladesh, Moldova, Peru, Romania, Sri Lanka, and Thailand,
we use the national Household Income and Expenditure Survey (HIES) for each
year. We make temporal and spatial adjustments for comparability reasons. For
Ghana and Nepal, we use the RIGA dataset (FAO n.d.).
Table 3A.1 Survey Sources of Data for Poverty Reduction Analysis in Selected
Developing Countries, 2000
Sampled countries Initial period and survey Final period and survey
Countries that use income-based poverty measures
Argentina 2000 EPH-C 2010 EPH-C
Brazil 2001 PNAD 2009 PNAD
Chile 2000 CASEN 2009 CASEN
Colombia 2002 GEIH 2010 GEIH
Costa Rica 2000 EHPM 2008 EHPM
Ecuador 2003 ENEMDU 2010 ENEMDU
Honduras 1999 EPHPM 2009 EPHPM
Panama 2001 EH 2009 EH
Paraguay 1999 EPH 2010 EPH
Countries that use consumption-based poverty measures
Bangladesh 2000 HIES 2010 HIES
Cambodia 2007 CSES 2010 CSES
Ghana 1998 RIGA 2005 RIGA
Moldova 2001 HBS 2010 HBS
Mongolia 2007 HSES 2011 HSES
Nepal 1996 RIGA 2003 RIGA
Peru 2004 ENAHO 2010 ENAHO
Romania 2001 HBS 2009 HBS
Sri Lanka 2002 HIES 2009 HIES
Thailand 2000 SES 2009 SES
Philippines 2006 FIES 2009 FIES
Vietnam 2004 VHLSS 2010 VHLSS
Note: EPH-C = Encuesta Permanente de Hogares-Continua (Argentina); PNAD = Pesquisa Nacional por
Amostra de Domicilios (Brazil); CASEN = Encuesta de Caracterización Socioeconómica Nacional (Chile);
GEIH = Gran Encuesta Integrada de Hogares (Colombia); EHPM = Encuesta de Hogares de Propósitos
Múltiples (Costa Rica); ENEMDU = Encuesta de Empleo, Desempleo y Subempleo (Ecuador);
EPHPM = Encuesta Permanente de Hogares de Propósitos Múltiples (Honduras); EH = Encuesta de Hogares
(Panama); EPH = Encuesta Permanente de Hogares (Paraguay); HIES = Household Income and Expenditure
Survey (Bangladesh); CSES = Cambodia Socio-Economic Survey; RIGA = Rural Income Generating Activities
(Nepal); HBS = Household Budget Survey (Romania); HSES = Household Socio-Economic Survey (Mongolia);
ENAHO = Encuesta Nacional de Hogares (Peru); SES = Household Socio-Economic Survey (Thailand);
FIES = Family Income and Expenditure Survey (Philippines); VHLSS = Vietnam Household Living
Standards Survey.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Annex 3B: Complementary Tables
Table 3B.1 Contributions to the Decline in the $1.25-a-Day (PPP) Poverty Headcount in Selected
Developing Countries, 2000s
Consumption-based measure of welfare
Bangladesh,
1998–2005
1996–2003
Sri Lanka,
Moldova,
2000–10
2002–09
2004–10
2001–10
Ghana,
Nepal,
Peru,
Poverty headcount rate (FGT0)a
Initial period (%) 57.7 14.8 38.2 73.7 3.5 27.5
Final period (%) 40.3 4.7 23.5 53.1 0.8 0.5
Total change (ppts) −17.4 −10.1 −14.7 −20.6 −2.6 −27.0
Decomposition of FGT0 a
Consumption-to-income ratio (%) −25.6 −0.2 17.7 −6.9 −9.3 −20.2
Adult population (%) 37.2 19.5 21.4 14.2 −14.9 5.7
Occupation share (%) 20.5 −21.8 −3.6 13.4 −3.1 −1.9
Labor income (%) 60.2 62.9 49.6 50.4 115.7 27.7
Capital (%) 7.8 0.0 5.0 5.9 1.4 0.0
Pension (%) — 1.5 — — −4.7 35.8
Transfers (%) 15.7 −1.6 10.0 23.0 10.8 46.1
Other nonlabor income (%) −15.9 39.8 — — 4.1 6.7
Total change (%) 100.0 100.0 100.0 100.0 100.0 100.0
Decomposition in FGT1 b
Consumption-to-income ratio (%) −60.6 −37.4 0.8 −10.7 −44.0 —
Adult population (%) 47.2 24.7 28.0 12.4 −40.5 —
Occupation share (%) 25.3 −76.3 −20.9 17.0 −12.5 —
Labor income (%) 99.1 93.2 70.9 54.0 176.4 —
Capital (%) 4.9 0.0 6.5 4.8 2.4 —
Pension (%) — 2.7 — — −9.6 —
Transfers (%) 16.7 −3.7 14.6 22.5 23.8 —
Other nonlabor income (%) −32.6 96.8 — — 4.1 —
Total change (%) 100.0 100.0 100.0 100.0 100.0 —
Decomposition in FGT2 c
Consumption-to-income ratio (%) −107.9 −112.6 −20.7 −17.8 −86.2 —
Adult population (%) 59.8 36.9 36.6 10.3 −64.7 —
Occupation share (%) 30.4 −181.4 −40.2 17.9 −27.7 —
Labor income (%) 151.8 128.1 98.4 60.2 234.1 —
Capital (%) 1.5 0.0 6.0 4.7 4.4 —
Pension (%) — 6.6 — — −14.7 —
Transfers (%) 20.0 −3.2 19.9 24.7 53.5 —
Other nonlabor income (%) −55.7 225.7 — — 1.2 —
Total change (%) 100.0 100.0 100.0 100.0 100.0 —
Sources: Data on Ghana and Nepal from FAO n.d. Data on Bangladesh, Moldova, Peru, Romania, Sri Lanka, and Thailand from household surveys.
Latin American data from SEDLAC, various years. Data on Cambodia, Mongolia, the Philippines, and Vietnam from World Bank harmonized surveys
for the East Asia and Pacific region.
Note: ppts = percentage points; PPP = purchasing power parity; — = not available; n.a. = not applicable.
a. FGT0 refers to the Foster, Greer, and Thorbecke (1984) measure of the poverty headcount index, which measures the proportion of the
population that is counted as poor using the $1.25-a-day standard.
b. FGT1 refers to the measure of the poverty gap index, which adds up the extent to which individuals on average fall below the poverty line,
expressed as a percentage of the poverty line (Foster, Greer, and Thorbecke 1984).
c. FGT2 refers to the measure of poverty severity, calculated as the poverty gap index squared, which implicitly puts more weight on observations
that fall well below the poverty line (Foster, Greer, and Thorbecke 1984).
60
Income-based measure of welfare
Philippines,
1999–2009
1999–2010
Cambodia,
Costa Rica,
Argentina,
Honduras,
Colombia,
Mongolia,
Paraguay,
Panama,
Vietnam,
Ecuador,
2000–10
2001–09
2002–10
2000–08
2003–10
2001–09
2007–10
2007–11
2006–09
2004–10
Brazil,
5.1 11.8 20.7 5.5 12.2 24.9 15.4 14.0 34.0 13.1 31.7 18.1
1.8 6.1 8.2 2.4 4.6 17.8 4.6 7.2 29.6 6.0 27.6 8.7
−3.3 −5.7 −12.6 −3.1 −7.6 −7.1 −10.8 −9.3 −4.4 −7.1 −4.1 −9.4
n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a.
20.0 11.2 5.2 7.1 19.0 41.4 4.0 34.7 34.9 −1.0 47.6 15.6
−5.7 −3.7 8.7 −2.8 −15.3 3.5 14.1 6.4 33.3 −18.2 −0.9 −4.3
14.3 44.7 48.7 26.7 50.3 51.5 39.9 47.8 46.4 62.6 23.2 75.4
−19.7 −4.4 0.4 10.7 −1.7 7.7 0.5 −1.3 — — — —
29.2 6.2 −2.2 61.3 5.2 6.7 10.4 −1.4 −5.2 3.2 2.2 2.9
37.7 37.3 31.0 55.2 36.0 −20.8 31.8 9.8 −8.1 52.2 27.6 8.0
24.2 8.8 8.2 −58.2 6.5 10.0 −0.7 3.9 −1.5 1.2 0.2 2.1
100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0
n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a.
16.5 7.9 −1.5 −2.0 13.3 52.6 2.3 30.8 16.0 −7.7 44.2 8.5
−7.9 −11.4 5.2 −23.2 −20.2 1.0 8.0 3.3 23.4 −14.0 −5.3 1.9
8.4 28.1 51.0 6.6 41.6 41.9 42.4 43.9 74.9 58.6 23.0 39.0
−26.9 −9.0 −3.3 24.0 −4.4 15.9 −0.8 −0.8 — — — —
24.5 0.6 −6.7 130.3 5.0 16.6 10.9 −0.4 −3.4 2.9 2.8 34.2
54.5 68.1 45.6 114.4 55.0 −42.7 42.0 17.6 −8.5 59.6 33.5 12.0
31.0 15.6 9.7 −150.1 9.7 14.7 −4.8 5.7 −2.4 0.6 1.8 4.4
100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0
n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a.
11.8 6.0 −4.4 −7.4 9.3 78.6 0.7 30.7 10.7 −12.1 39.7 −112.6
−7.2 −14.4 3.0 −39.8 −23.1 −9.6 3.8 −0.5 19.1 −10.3 −7.1 278.4
6.5 14.2 49.5 −12.8 35.0 21.6 42.4 39.1 82.3 52.6 20.6 −696.3
−30.4 −11.2 −5.1 35.5 −7.0 32.2 −1.4 0.3 — — — —
20.0 −2.3 −8.9 188.3 5.0 33.7 11.7 −0.3 −3.1 3.0 3.2 543.6
66.0 88.2 54.4 166.9 68.3 −78.7 50.4 23.7 −7.4 66.7 39.3 60.1
33.2 19.5 11.4 −230.8 12.5 22.2 −7.7 7.0 −1.7 0.1 4.3 26.9
100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0
61
Table 3B.2 Contributions to the Decline in the $2.50-a-Day (PPP) Poverty Headcount in Selected
Developing Countries, 2000s
Consumption-based measure of welfare
Bangladesh,
1998–2005
1996–2003
Romania,
Sri Lanka,
Thailand,
Moldova,
2000–10
2002–09
2004–10
2000–09
2001–10
2001–09
Ghana,
Nepal,
Peru,
Poverty headcount rate (FGT0)a
Initial period (%) 89.2 61.9 71.8 94.3 22.9 7.9 71.4 23.7
Final period (%) 84.0 45.1 58.3 84.9 11.7 2.5 12.9 4.2
Total change (ppts) −5.2 −16.8 −13.5 −9.5 −11.2 −5.3 −58.5 −19.5
Decomposition of FGT0 a
Consumption-to-income ratio (%) −56.1 7.4 18.5 −30.9 4.8 −47.0 −2.9 15.3
Adult population (%) 46.4 14.5 21.1 18.7 7.4 35.8 7.2 48.4
Occupation share (%) 23.3 −7.7 −0.9 6.2 5.6 −10.1 6.6 24.8
Labor income (%) 85.6 51.1 52.2 62.3 70.1 61.0 26.2 −33.0
Capital (%) 5.3 0.0 1.6 9.7 1.5 5.9 0.4 38.3
Pension (%) — 1.9 — — −0.4 — 25.5 −20.2
Transfers (%) 24.8 2.0 7.3 34.0 10.0 57.8 30.1 16.2
Other nonlabor income (%) −29.3 30.9 — — 1.0 −3.4 7.0 10.2
Total change (%) 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0
Decomposition in FGT1b
Consumption-to-income ratio (%) −46.3 3.4 14.0 −14.0 0.3 −131.5 −15.9 51.3
Adult population (%) 43.9 17.6 22.7 14.9 −0.1 58.2 4.9 138.0
Occupation share (%) 22.4 −14.6 −6.1 13.5 1.8 −39.8 6.0 70.9
Labor income (%) 81.4 54.3 55.1 53.8 85.8 67.8 22.9 −154.8
Capital (%) 6.2 0.0 4.3 6.5 1.4 13.6 0.2 117.1
Pension (%) — 1.7 — — −2.0 — 29.5 −120.9
Transfers (%) 17.9 0.5 10.0 25.3 10.9 126.7 46.8 14.8
Other nonlabor income (%) −25.6 37.0 — — 1.8 5.0 5.6 −16.4
Total change (%) 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0
Decomposition in FGT2c
Consumption-to-income ratio (%) −54.5 −4.2 8.2 −12.7 −7.4 −264.8 −31.2 148.6
Adult population (%) 45.8 19.3 25.0 13.4 −8.2 95.2 2.8 320.2
Occupation share (%) 23.9 −26.2 −12.1 15.3 −1.5 −95.6 5.3 175.7
Labor income (%) 91.6 61.0 61.8 54.3 103.2 49.5 16.2 −387.5
Capital (%) 5.5 0.0 5.1 5.6 1.6 28.1 −0.2 280.8
Pension (%) — 1.9 — — −3.4 — 32.7 −336.5
Transfers (%) 17.4 −0.3 12.0 24.0 13.3 253.4 70.1 −14.5
Other nonlabor income (%) −29.6 48.5 — — 2.4 34.3 4.3 −86.7
Total change (%) 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0
Sources: Data on Ghana and Nepal from FAO n.d. Data on Bangladesh, Moldova, Peru, Romania, Sri Lanka, and Thailand from household surveys.
Latin American data from SEDLAC, various years. Data on Cambodia, Mongolia, the Philippines, and Vietnam from World Bank harmonized
surveys for the East Asia and Pacific region.
Note: ppts = percentage points; PPP = purchasing power parity; — = not available; n.a. = not applicable.
a. FGT0 refers to the Foster, Greer, and Thorbecke (1984) measure of the headcount index, which measures the proportion of the population that
is counted as poor using the $2.50-a-day standard.
b. FGT1 refers to the measure of the poverty gap index, which adds up the extent to which individuals on average fall below the poverty line,
and expresses it as a percentage of the poverty line (Foster, Greer, and Thorbecke 1984).
c. FGT2 refers to the measure of poverty severity, calculated as the poverty gap index squared, which implicitly puts more weight on
observations that fall well below the poverty line (Foster, Greer, and Thorbecke 1984).
62
Income-based measure of welfare
Philippines,
1999–2009
1999–2010
Cambodia,
Costa Rica,
Argentina,
Honduras,
Colombia,
Mongolia,
Paraguay,
Panama,
Vietnam,
Ecuador,
2000–10
2001–09
2000–09
2002–10
2000–08
2003–10
2001–09
2007–10
2007–11
2006–09
2004–10
Brazil,
Chile,
14.2 27.4 9.0 42.3 14.7 31.5 47.9 28.7 26.7 57.3 35.7 62.2 56.4
6.4 15.1 4.3 22.0 7.6 15.9 36.2 16.1 18.4 60.1 22.5 60.3 33.2
−7.8 −12.3 −4.7 −20.3 −7.1 −15.6 −11.6 −12.6 −8.3 2.8 −13.2 −1.9 −23.2
n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a.
20.6 16.5 31.9 10.5 25.3 24.5 37.4 9.2 56.8 53.0 5.4 71.2 16.0
10.4 6.9 −18.4 14.0 9.3 −6.0 −3.1 22.5 6.5 29.0 −22.0 −18.8 3.0
29.5 44.8 46.9 40.7 24.5 53.2 57.8 33.5 43.8 −162.6 75.0 8.1 68.6
−9.1 −1.1 −11.3 3.0 4.7 −1.1 3.8 0.0 −2.4 — — — —
19.9 10.7 31.9 2.7 31.5 4.7 3.7 9.8 −3.8 1.8 3.3 2.7 4.3
16.1 16.7 87.4 20.1 28.3 20.2 −6.9 24.5 −0.5 −18.6 36.4 40.7 5.2
12.7 5.4 −68.4 9.0 −23.6 4.5 7.2 0.5 −0.4 −2.6 1.9 −4.0 2.9
100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 −100.0 100.0 100.0 100.0
n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a.
18.7 12.7 42.6 5.4 15.1 21.5 43.8 5.3 41.0 54.1 0.9 49.1 14.8
1.1 0.4 −50.2 9.5 −0.8 −12.2 −1.5 14.5 6.7 47.6 −18.3 −5.3 0.2
20.9 41.7 17.2 45.9 20.2 50.3 54.3 38.4 44.3 24.9 67.4 21.4 66.7
−15.6 −3.4 −32.2 0.4 10.1 −2.2 7.4 −0.2 −1.5 — — — —
23.6 7.3 72.1 −1.4 61.5 4.7 7.4 10.3 −1.5 −4.5 3.5 2.9 8.0
31.4 32.8 191.6 31.2 53.0 32.0 −20.8 33.3 8.1 −16.9 45.0 32.1 7.4
19.8 8.5 −141.1 9.2 −59.1 6.0 9.3 −1.6 2.9 −5.2 1.5 −0.1 2.8
100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0
n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a.
17.2 10.4 64.2 1.8 7.8 18.1 48.9 3.5 35.7 27.1 −2.6 45.6 8.0
−3.3 −4.5 −95.7 7.0 −10.5 −15.9 −0.9 10.7 5.0 27.8 −16.8 −4.7 12.7
15.0 35.9 −26.9 48.3 14.3 46.8 48.5 40.4 43.8 62.1 63.8 22.1 27.7
−20.9 −5.7 −65.9 −1.5 16.0 −3.2 11.7 −0.5 −1.1 — — — —
24.1 4.4 129.3 −4.2 91.3 4.8 11.8 10.7 −1.0 −3.7 3.2 2.9 36.7
43.0 48.0 348.1 39.1 79.6 41.9 −31.7 38.7 13.1 −10.1 51.2 32.8 10.8
24.9 11.4 −253.2 9.6 −98.5 7.5 11.7 −3.5 4.4 −3.1 1.2 1.2 4.1
100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0
63
64
Table 3B.3 Contributions to the Decline in the $4.00–$5.00-a-Day (PPP) Poverty Headcount in Selected Developing Countries, 2000s
Consumption-based measure of welfare Income-based measure of welfare
$4.00-a-day poverty line $5.00-a-day poverty line $4.00-a-day poverty line
Costa Paraguay,
Peru, Sri Lanka, Thailand, Moldova, Romania, Argentina, Brazil, Chile, Colombia, Rica, Ecuador, Honduras, Panama, 1999–
2004–10 2002–09 2000–09 2001–10 2001–09 2000–10 2001–09 2000–09 2002–10 2000–08 2003–10 1999–2009 2001–09 2010
Poverty headcount rate (FGT0)
Initial period (%) 45.8 84.7 31.4 93.8 75.3 27.5 43.1 23.2 61.6 29.2 51.5 66.1 43.4 50.3
Final period (%) 30.0 78.1 16.6 58.7 33.2 14.6 27.6 11.8 39.5 18.9 33.4 52.1 29.9 33.0
Total change (ppts) −15.8 −6.6 −14.8 −35.1 −42.2 −13.0 −15.5 −11.4 −22.1 −10.2 −18.1 −14.1 −13.5 −17.3
Decomposition of FGT0 a
Consumption-to-
income ratio (%) 5.1 −9.0 −11.8 −20.9 15.8 n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a.
Adult population (%) 9.0 9.7 24.2 3.5 24.1 22.0 16.4 31.0 12.1 34.4 27.1 32.0 13.5 59.5
Occupation share (%) 10.0 −24.3 4.0 6.5 20.4 16.7 10.9 −0.1 15.8 14.2 −3.3 −4.1 29.1 11.6
Labor income (%) 62.5 75.1 47.5 37.3 1.9 35.2 41.6 48.2 38.3 17.7 53.3 55.8 30.2 33.2
Capital (%) 2.1 0.0 2.8 0.6 21.0 −5.4 −0.7 −4.2 4.3 3.9 −0.7 3.8 −0.2 0.0
Pension (%) 0.6 4.3 — 25.1 −2.8 14.9 17.9 15.8 4.6 23.6 6.2 3.2 10.4 −3.0
Transfers (%) 10.0 6.5 33.1 37.2 8.7 7.3 9.1 41.8 15.6 22.4 13.6 3.4 16.4 −1.0
Other nonlabor
income (%) 0.7 37.7 0.3 10.6 11.0 9.3 4.8 −32.5 9.4 −16.1 3.8 6.0 0.5 −0.3
Total change (%) 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0
table continues next page
Table 3B.3 Contributions to the Decline in the $4.00–$5.00-a-Day (PPP) Poverty Headcount in Selected Developing Countries, 2000s (continued)
Consumption-based measure of welfare Income-based measure of welfare
$4.00-a-day poverty line $5.00-a-day poverty line $4.00-a-day poverty line
Costa
Peru, Sri Lanka, Thailand, Moldova, Romania, Argentina, Brazil, Chile, Colombia, Rica, Ecuador, Honduras, Panama, Paraguay,
2004–10 2002–09 2000–09 2001–10 2001–09 2000–10 2001–09 2000–09 2002–10 2000–08 2003–10 1999–2009 2001–09 1999–2010
Decomposition in FGT1 b
Consumption-to-
ncome ratio (%)
i 3.4 2.8 −35.9 −10.8 20.0 n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a.
Adult population (%) 5.9 15.0 31.8 5.6 44.0 20.5 14.8 33.9 8.7 25.8 24.2 38.0 8.0 49.8
Occupation share (%) 6.0 −14.3 −5.8 6.6 27.0 9.3 5.5 −15.8 12.3 7.9 −7.7 −2.1 19.9 8.1
Labor income (%) 72.3 56.9 53.7 27.5 −25.6 28.3 42.4 41.7 42.2 19.2 52.1 55.6 35.2 41.2
Capital (%) 1.7 0.0 4.9 0.3 36.5 −9.9 −2.0 −11.9 2.1 6.5 −1.5 5.2 −0.1 −1.3
Pension (%) −0.5 2.3 — 26.5 −20.7 19.2 11.7 31.6 1.4 40.1 5.1 5.1 10.3 −2.6
Transfers (%) 10.1 2.1 51.5 36.9 11.7 18.4 20.9 85.7 24.0 35.6 22.9 −9.5 27.3 3.8
Other nonlabor
income (%) 1.2 35.1 −0.2 7.5 7.1 14.3 6.8 −65.1 9.3 −35.0 5.0 7.7 −0.7 1.1
Total change (%) 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0
Decomposition in FGT2 c
Consumption-to-
ncome ratio (%)
i 0.9 1.9 −68.4 −13.0 30.5 n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a. n.a.
Adult population (%) 2.1 16.8 41.1 5.4 72.8 19.1 13.2 39.5 5.9 18.6 22.0 42.3 5.8 43.3
Occupation share (%) 3.6 −16.5 −18.2 6.3 40.6 4.0 1.6 −34.3 10.0 1.5 −10.9 −1.6 15.4 6.6
Labor income (%) 80.8 56.4 57.1 24.6 −64.3 23.1 40.9 28.2 44.9 18.6 50.4 53.7 37.8 43.1
Capital (%) 1.6 0.0 8.1 0.2 61.1 −14.1 −3.3 −23.3 0.6 9.5 −2.1 7.3 −0.2 −1.3
Pension (%) −1.4 2.0 — 27.8 −51.5 21.7 8.5 53.2 −1.0 57.1 4.8 7.3 10.4 −1.9
Transfers (%) 10.8 1.1 78.0 42.3 11.6 27.8 30.7 143.6 30.2 49.8 29.9 −18.1 32.7 7.6
Other nonlabor
income (%) 1.6 38.3 2.4 6.5 −0.8 18.3 8.3 −106.9 9.4 −55.2 5.9 9.1 −1.8 2.5
Total change (%) 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0
Sources: Data on Ghana and Nepal from FAO n.d. Data on Bangladesh, Moldova, Peru, Romania, Sri Lanka, and Thailand from household surveys. Latin American data from SEDLAC, various years. Data on Cambodia, Mongolia,
the Philippines, and Vietnam from World Bank harmonized surveys for the East Asia and Pacific region.
Note: ppts = percentage points; PPP = purchasing power parity; — = not available; n.a. = not applicable.
a. FGT0 refers to the Foster, Greer, and Thorbecke (1984) measure of the headcount index, which measures the proportion of the population that is counted as poor.
b. FGT1 refers to the measure of the poverty gap index, which adds up the extent to which individuals on average fall below the poverty line and expresses it as a percentage of the poverty line (Foster, Greer, and Thorbecke 1984).
65
c. FGT2 refers to the measure of poverty severity, calculated as the poverty gap index squared, which implicitly puts more weight on observations that fall well below the poverty line (Foster, Greer, and Thorbecke 1984).
66 What Accounts for Changes in Poverty over the Past Decade?
Notes
1. SEDLAC is a joint effort of the Centro de Estudios Distributivos Laborales y Sociales
(CEDLAS) of the Universidad Nacional de La Plata in Buenos Aires and the World
Bank’s Latin American Poverty and Gender Group. For more information about
SEDLAC, see its website: http://sedlac.econo.unlp.edu.ar/eng/.
2. For more information about RIGA, see its page on the FAO website: http://www.fao
.org/economic/riga/riga-database/en/.
3. For a more comprehensive treatment of how global poverty measures have developed
historically, see Chen and Ravallion (2010). The World Bank’s interactive website,
PovcalNet (developed by the Bank’s Development Research Group), also allows users
to replicate calculations made by World Bank researchers to estimate the extent of
absolute poverty in the world as well as “to calculate poverty measures under different
assumptions” (PovcalNet online poverty analysis tool, http://iresearch.worldbank .org
/PovcalNet/index.htm).
4. Note that the consumption-to-income ratio is the ratio of measured consumption to
measured income. To the extent that there is measurement error in both of these,
interpretations about changes in this ratio must be treated with caution.
5. FGT1 refers to the Foster, Greer, and Thorbecke (1984) measure of the poverty gap
index, which adds up the extent to which individuals on average fall below the
poverty line, and expresses it as a percentage of the poverty line.
6. FGT2 refers to the Foster, Greer, and Thorbecke (1984) measure of poverty severity,
calculated as the poverty gap index squared, which implicitly puts more weight on
observations that fall well below the poverty line.
7. We do not report decompositions for Chile and Thailand using the $1.25-a-day
poverty line because only 1 percent or less of the population in those countries
experiences poverty at that level.
8. Also, see Fiszbein et al. (2009) for a review of conditional cash transfer programs.
Bibliography
Bourguignon, François, Francisco H. G. Ferreira, and Nora Lustig, eds. 2005. The
Microeconomics of Income Distribution Dynamics in East Asia and Latin America.
Washington, DC: World Bank.
Bourguignon François, Francisco H. G. Ferreira., and Phillippe G. Leite 2008. “Beyond
Oaxaca-Blinder: Accounting for Differences in Household Income Distributions.”
Journal of Economic Inequality 6 (2): 117–48.
Chen, Shaohua and Martin Ravallion. 2010. “The Developing World Is Poorer Than We
Thought, but No Less Successful in the Fight Against Poverty.” Quarterly Journal of
Economics 125 (4): 1577–625.
Datt, Gaurav, and Martin Ravallion. 1992. “Growth and Redistribution Components of
Changes in Poverty Measures: A Decomposition with Applications to Brazil and India
in the 1980s.” Journal of Development Economics 38: 275–96.
FAO (Food and Agriculture Organization of the United Nations). n.d. Rural Income
Generating Activities (RIGA). Database, FAO, Rome; World Bank and American
University, Washington, DC, http://www.fao.org/economic/riga/riga-database/en/.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
What Accounts for Changes in Poverty over the Past Decade? 67
Ferreira, Francisco H. G. 2012. “Distributions in Motion: Economic Growth, Inequality,
and Poverty Dynamics.” In The Oxford Handbook of the Economics of Poverty, edited by
Philip N. Jefferson, 427–62. New York: Oxford University Press.
Fiszbein, Ariel, Norbert Schady, Francisco H. G. Ferreira, Margaret Grosh, Pedro Olinto,
and E. Skoufias. 2009. Conditional Cash Transfers: Reducing Present and Future Poverty.
Washington, DC: World Bank.
Foster, James, Joel Greer, and Erik Thorbecke. 1984. “A Class of Decomposable Poverty
Measures.” Econometrica 52 (3): 761–65.
Juhn, Chinhui, Kevin M. Murphy, and Brooks Pierce. 1993. “Wage Inequality and the Rise
of Returns to Skill.” Journal of Political Economy 101 (3): 410–42.
Ravallion, Martin, and Shaohua Chen. 2007. “China’s (Uneven) Progress against Poverty.”
Journal of Development Economics 82 (1): 1–42.
SEDLAC (Socio-Economic Database for Latin America and the Caribbean). Various
years. Statistical database produced by CEDLAS (Center for Distributional, Labor
and Social Studies) of the Universidad Nacional de la Plata, Buenos Aires; and World
Bank, Washington, DC. http://sedlac.econo.unlp.edu.ar/eng/.
Shapley, Lloyd. 1953. “A Value for n-Person Games.” In Contributions to the Theory of
Games, Vol. 2, edited by H. W. Kuhn and A. W. Tucker, 307–17. Princeton, NJ:
Princeton University Press.
Shorrocks, Anthony F. 1999. “Decomposition Procedures for Distributional Analysis:
A Unified Framework Based on Shapley Value.” Unpublished manuscript, University
of Essex and Institute for Fiscal Studies.
World Bank. 2005. World Development Report 2006: Equity and Development. Washington,
DC: World Bank and Oxford University Press.
———. 2012. World Development Indicators 2012. Washington, DC: World Bank. http://
data.worldbank.org/data-catalog/world-development-indicators. Online database.
———. 2013. World Development Indicators 2013. Washington, DC: World Bank. http://
data.worldbank.org/data-catalog/world-development-indicators. Online database.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Chapter 4
Counterfactual Decomposition of
Changes in Poverty Outcomes
Introduction
This chapter reviews decomposition methods commonly used to identify key
factors that drive changes in the observed poverty outcomes. The review focuses
on a variety of microeconometric methods and delves into the identification
strategy underlying each of these methods. Macro-decomposition methods, such
as those discussed in chapter 2, explain changes in poverty in terms of changes
in aggregate statistics (such as the mean), relative inequality, subgroup popula-
tion shares, and within-group poverty.
The usefulness of such macro methods in policy making is severely limited
because these methods identify determining factors that are hard to target with
policy instruments. Ravallion (2001) explains that a better understanding of the
heterogeneity of policy impact requires a deeper empirical analysis of distribu-
tional change at the micro level. Micro-decomposition methods thus go beyond
the summary statistics that are the focus of macro methods and attempt to link
distributional changes to fundamental elements that drive these changes, such as
individual or household characteristics that facilitate policy targeting.
The logic underpinning all these methods (including the macro ones discussed
in chapter 2) can be organized around the following terms:
• Domain: the distributional change a method seeks to decompose on the basis
of a model that links the outcome of interest to its determining factors.
• Outcome model: the assumed relationship between the outcome of interest and
its determining factors, the specification of which determines the potential
scope of the decomposition.
• Scope: the set of explanatory factors that decomposition methods seek to
identify.
This chapter adapts a previous work by the same author (Essama-Nssah 2012).
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7 69
70 Counterfactual Decomposition of Changes in Poverty Outcomes
• Identification: the assumptions needed to recover, in a meaningful way, the
terms of the decomposition that affect the outcome.
• Estimation: the computation of the relevant parameters on the basis of
sample data.
The point of departure for micro-decomposition methods is the fact that a
poverty measure is a functional1 of a distribution of living standards. An
individual’s living standard is an outcome of interaction between opportunities
offered by society and the individual’s ability to identify and exploit such
opportunities. In other words, an individual’s living standard is a payoff from
participation in the life of society.
One can thus think of life in society as a game defined by a set of rules govern-
ing various interactions of the parties involved (players). These rules spell out
what the concerned parties are allowed to do and how these allowable actions
determine outcomes. An environment within which a game is played consists of
three basic elements: a set of potential participants, a set of possible outcomes, and
a set of possible participant types (players). Player types are characterized by their
preferences, capabilities, information, and beliefs (Milgrom 2004). The operation
of a game can be represented by a function that maps environments to potential
outcomes. Thus an individual payoff is a function of participation and type.
This paradigm leads us to consider an individual’s living standard to be a func-
tion of endowments, behavior, and the circumstances that determine the returns to
these endowments from any social transaction. These elements span the scope of
most micro-decomposition methods and represent some of the deep structural
drivers of the distributional changes underlying the observed variations in
poverty outcomes.
Although macro- and micro-decomposition methods differ in their scope,
they share the same fundamental identification strategy based on the notion of
ceteris paribus variation. Attribution of outcomes to policy is the hallmark of
policy impact evaluation. Variations in individual outcomes associated with a
policy’s implementation are not necessarily the result of the policy in question.
These variations could be driven by changes in confounding factors in the socio-
economic environment. At the most fundamental level, all identification strate-
gies seek to isolate an independent source of variation in policy and link it to the
outcome of interest to ascertain impact.
Similarly, macro- and micro-decomposition methods also identify the deter-
minants of differences across living-standard distributions by comparing counter-
factual distributions with the observed ones. Counterfactual distributions are
obtained by changing one determining factor at a time while holding all the other
factors fixed (a straight application of ceteris paribus identification strategy).
The linchpin of the whole process is the estimation of credible counterfactu-
als. In the context of micro-decomposition methods, for instance, one must
carefully estimate a key counterfactual: the distribution of outcomes in the base
state (t = 0) assuming the distribution of individual characteristics prevailing in
the end state (t = 1). Put another way, such a counterfactual represents the
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Counterfactual Decomposition of Changes in Poverty Outcomes 71
distribution of outcomes that would prevail in the end state if the characteristics
in that state had been treated according to the outcome structure prevailing in
initial state. Depending on the chosen functional form for the outcome equation,
there are both parametric and nonparametric ways of estimating this
counterfactual.
The rest of the chapter is divided into three sections. The next section reviews
methods to identify and estimate the composition (or endowment) effect and
structural (or price) effect associated with variation in a general distributional
statistic. Both aggregate and detailed decompositions are considered. This is fol-
lowed by a section that focuses on proposed methods to account for behavioral
responses to changes in the socioeconomic environment. These methods rely on
the specification and estimation of a microeconometric model based on some
theory of individual (or household) behavior and social interaction. The final
section summarizes the chapter’s findings.
As the chapter proceeds, the methods first reviewed are considered statistical,
while those covered in the later section accounting for behavioral responses are
structural in nature. One particular advantage of the statistical approach is that
it gives the analyst semi- and nonparametric ways to identify the aggregate
effects without having to impose a functional form on the relationship between
the outcome and its determinants. However, the statistical framework cannot
shed light on the mechanism underlying that relationship. The decomposition
results therefore lack any causal interpretation. Because these methods are based
on statistical models of conditional distributions, it is conceivable that the behav-
ioral effect is mixed up with the price effect identified by these methods.
Structural methods go a step further toward identifying factors associated with
structural elements underpinning the observed changes in poverty outcomes.
Both the statistical and structural approaches seek to model conditional outcome
distributions. A key distinction between the two approaches is that the former
relies entirely on statistics while the latter combines economics and statistics.
The Composition and Structural Effects
As noted earlier, an individual’s living standard is an outcome of participation in
the life of society. This outcome is a function of the individual’s endowments,
behavior, and the circumstances that determine the returns to these endowments
from any socioeconomic transaction. In this section, we focus on the role of
endowments and the returns to those endowments in driving the distributional
changes2 that underlie changes in poverty outcomes. We consider changes in
both distributional statistics and entire distributions. We also discuss the contri-
bution of unobservable characteristics.
All methods reviewed in this section are consistent with the logic of the basic
Oaxaca-Blinder decomposition (Blinder 1973; Oaxaca 1973). The methods
differ mainly in the ability to account for (a) heterogeneity in the composition
and structural effects along the entire outcome distribution and (b) the contribu-
tion of observable or unobservable characteristics to both types of effects.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
72 Counterfactual Decomposition of Changes in Poverty Outcomes
In particular, using a linear regression model to link the outcome variable to
individual or household characteristics allows one to perform both aggregate
and detailed decompositions of changes in distributional statistics. Decomposing
differences in density functions or across quantiles reveals how the composition
and structural effects vary along the outcome distribution. Under some speci-
fications of the outcome model (Juhn, Murphy, and Pierce 1993), it is possible
to explicitly account for the contribution of unobservables to the structural
effect. Finally, it is important to keep in mind the analogy between the Oaxaca-
Blinder decomposition and treatment effect analysis. This analogy underlies the
development of flexible methods for estimating the endowment and structural
effects.
Changes in Distributional Statistics
In the context of assessing variations in individual and social outcomes, the com-
position or endowment effect indicates the change in outcome that is due only
to changes in the distribution of observable characteristics of agents. The struc-
tural or price effect is the result of changes in the returns to those characteristics.
Although we focus on changes in poverty outcomes, it is instructive to consider
the more general approach that applies to all distributional statistics, including
poverty and inequality measures.
We frame the analysis of the underlying distributional changes within the
logic of the standard Oaxaca-Blinder decomposition based on an outcome model
that considers both individual and social outcomes (Blinder 1973; Oaxaca 1973).
We discuss, in turn, the outcome model, the identification, and the estimation of
the relevant effects. Our presentation relies heavily on Fortin, Lemieux, and
Firpo (2011).
Outcome Model
Let q stand for the social outcome of interest. This is a distributional statistic that
we view as a functional of the distribution of individual outcomes, Fy. As noted
above, an individual outcome y is a function of endowments, behavior, and returns
to endowments. We are interested in decomposing a change in the social out-
come from the base period t = 0 to the end period t = 1. Let Fy0 |t = 0 stand for the
outcome distribution observed in the initial period and Fy1|t = 1 for that observed
in the final period. We can express distributional change between states 0 and 1
by the following variation in q (Fy):
O = q ( Fy1|t = 1 ) − q ( Fy0 |t = 0 ) . (4.1)
∆q
Equation (4.1) characterizes the domain of the decomposition methods con-
sidered in this chapter. As far as the scope is concerned, most micro-decomposition
methods seek to decompose this overall difference on the basis of the relation-
ship between the outcome variable and individual or household characteristics.
The following equation represents a general expression of that relationship:
yt = jt (xt, et), t = 0,1. (4.2)
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Counterfactual Decomposition of Changes in Poverty Outcomes 73
Equation (4.2) suggests that conditional on the observable characteristics, x,
the outcome distribution depends only on the function jt(∙) and the distribution
of the unobservable characteristics e. Thus there are four potential terms in the
scope of micro-decomposition methods based on this model. Differences in out-
come distributions between the two periods may be the result of (a) differences
in the returns to observable characteristics given the functions defining the out-
come structure; (b) differences in the returns to unobservable characteristics also
defined by the structural functions; (c) differences in the distribution of observ-
able characteristics; and (d) differences in the distribution of unobservable
characteristics.
The classic Oaxaca-Blinder method seeks to decompose the overall difference
in unconditional mean outcome between two groups or time periods. In this case,
q (Fy) = m(Fy) = E(y). The overall difference in the mean outcome between the
µ
two periods can therefore be written as ∆O = [E( y1 ) − E(y 0 )]. This framework
assumes an additive linear relationship between the outcome variable y and its
determinants. The equivalent of equation (4.2) can therefore be written as
yt = xt bt + ¨t; t = 0, 1.
Identification
Given the potential scope implied by the outcome model (equation [4.2]), the
next step is to impose enough restrictions to identify the factors of interest. In
general, these restrictions are imposed on the form of the outcome functions,
jt(∙), and on the joint distribution of the observable and unobservable character-
istics, x and e. On the basis of the general outcome model represented by the
definition of the distributional statistic q, along with equation (4.2), it is impos-
sible to distinguish the contribution of the returns to observables from that of
unobservables. These two terms can therefore be lumped into a single term: the
structural effect, also called the price effect, noted ∆q q
S. Let ∆ X stand for the com-
position effect and ∆q ε for the effect associated with differences in the distribution
of unobservables. The issue now is to identify these three effects so that they
account for the overall difference described by equation (4.1).
Just as in the case of the size and redistribution effects discussed in chapter 2,
we rely on counterfactual decomposition to identify the composition and struc-
tural effects. Let y0|t = 1 be the outcomes that would have prevailed in period 1 if
individual characteristics in that period had been rewarded according to the
reward regime prevailing in period 0, that is, according to ϕ0(∙). Let Fy0 |t = 1 stand
for the corresponding distribution and q ( Fy0 |t = 1 ) for the corresponding value of
the statistic of interest. By definition, the composition effect is given by the
q ( Fy0 |t = 1 ) − q ( Fy0 |t = 0 ) .
X =
ollowing expression: ∆q
f
This term is meaningful and identifiable only if it emerges from a ceteris
aribus variation of the distribution of observable characteristics. Given the scope
p
of the decomposition implied by the underlying outcome model and the fact
that we are lumping together the returns to both observables and unobservables,
we must restrict movements in the conditional distribution of unobservable
characteristics (given observables) and the structure of individual outcomes.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
74 Counterfactual Decomposition of Changes in Poverty Outcomes
We therefore assume that the conditional distribution of unobservables is the
same in both states of the world. This is the so-called ignorability assumption
usually made in the context of observational studies of treatment effect.
We further assume that the outcome structure, ϕ( ), remains stable as we
adjust the distribution of observables to obtain the relevant counterfactual out-
come. This assumption is sometimes referred to as simple treatment assumption
or the assumption of no general equilibrium effects. Ignorability implies that ∆q ε = 0.
Combining the assumption of ignorability with that of no general equilibrium
effects secures the identification of both the composition and structural effects.
S = q ( Fy1|t = 1 ) − q ( Fy0 |t = 1 ) . The overall
The structural effect is identified by ∆q
ifference in equation (4.1) can therefore be expressed as
d
q ( Fy0 |t = 1 ) − q ( Fy0 |t = 0 ) + q ( Fy1|t = 1 ) − q ( Fy0 |t = 1 ) ,(4.3)
O =
∆q
where the first term on the right-hand side is the endowment effect and the
second is the structural effect. In the context of poverty analysis, if P stands for
the poverty measure of interest, then equation (4.3) implies that observed
changes in poverty can be decomposed as follows:
P ( Fy0 |t = 1 ) − P ( Fy0 |t = 0 ) + P ( Fy1|t = 1 ) − P ( Fy0 |t = 1 ) .(4.4)
=
P
∆O
In addition to assuming a linear model for the individual outcome, the classic
Oaxaca-Blinder decomposition assumes that the conditional mean of ε given the
observables is equal to zero. This assumption, which is stronger than ignorability,
implies that the conditional mean outcome can be written as follows: E(y|x) = xb.
Therefore b is a measure of the effect of x on the conditional mean outcome. By
the law of iterated expectations, we know that the unconditional mean outcome
is E(y) = Ex[E(y|x)] = E(x)b. This result means that b also measures the effect of
changing the mean value of x on the unconditional mean value of y. This is the
interpretation underlying the original Oaxaca-Blinder decomposition.
The domain of the standard Oaxaca-Blinder decomposition can therefore be
E ( x1 ) β1 − E ( x 0 ) β0
µ
written as ∆O = . This overall difference in means is subject
to counterfactual decomposition as follows: The average outcome for period
1 valued on the basis of the reward parameters for period 0 is equal to E(x1)b0.
This is a counterfactual outcome for period 1. We can subtract it from, and add
it back to, the above overall mean difference to get the following expression:
E ( x1 ) − E ( x 0 ) E ( x1 )( β1 − β0 )
µ
∆O = β0 +
3
. Looking at the regression coefficients
b as characterizing the returns to (or reward for) observable characteristics, this
aggregate decomposition reveals that under the identifying assumptions, the overall
µ µ
mean difference can be expressed as ∆O = ∆µ µ
X + ∆ S , where ∆ X is the composition
µ
(endowment) effect, and ∆ S is the structural (price) effect.
Estimation
Parametric and Nonparametric Approaches. There are both parametric and non-
parametric approaches to estimating the counterfactuals involved in the
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Counterfactual Decomposition of Changes in Poverty Outcomes 75
identification and hence estimation of the composition and structural effects.
DiNardo, Fortin, and Lemieux (1996) show that the counterfactual distribution,
Fy0 |t = 1, can be estimated by properly reweighing the distribution of covariates in
period 0. One can express the resulting counterfactual distribution as follows:4
Fy0 |t = 1 ( y ) =
∫F
y0 | x 0 ( y|x ) w ( x ) dFx ( x ) , (4.5)
0
dFx1 ( x ) P(t = 1| x ) 1 − π
where the reweighing factor is equal to w ( x ) = = ⋅ .
dFx0 ( x ) 1 − P(t = 1| x ) π
These weights are proportional to the conditional odds of being observed in
state 1. The proportionality factor depends on π, which is the proportion of
cases observed in state 1. One can easily compute the reweighing factor on the
basis of a probability model, such as logit or probit. Furthermore, if one is inter-
ested only in the aggregate decomposition of the variation in a distributional
statistic, then the only needs are an estimate of the relevant counterfactual dis-
tribution and the corresponding value of the statistic in question.
The decomposition presented in equation (4.3) suggests a nonparametric
identification strategy and can be estimated by the inverse probability weighting
(IPW) method implied by equation (4.5). Nonparametric methods allow ana-
lysts to decompose changes in distributional statistics into endowment and struc-
tural effects without having to assume a functional form for the outcome model.
The downside is that one cannot separate the respective contributions of the
observable and unobservable factors into the structural effect, nor can one
account for changes in agents’ behavior. Later on, we consider a way of separating
the contribution of unobservables from that of observables, and in the concluding
section we review proposed methods to account for behavioral responses.
A parametric approach to the estimation of the composition and structural
effect requires a more explicit link between the distributional statistic of interest
and individual or household characteristics (depending on the unit of analysis).
This approach entails imposing more structure on the individual outcome equa-
tion (4.2) and finding a way of expressing the distributional statistic, q, as an
explicit function of the covariates of interest. Given the assumption that the
function ϕ(∙) in equation (4.2) is linear and separable in observable covariates x
and unobservable factors ε, the classic Oaxaca-Blinder decomposition relies on
the law of iterated expectations to establish a direct link between the uncondi-
tional mean of y (the distributional statistic of interest) and individual character-
istics. The assumption that the conditional mean of the error term given the
observables is zero ensures that the ordinary least squares (OLS) method of
estimation will produce unbiased and consistent estimates of key parameters.
Use of Influence Functions. Things are not so simple for general distributional
statistics such as quantiles, poverty, and inequality measures. For instance, if one
is interested in understanding changes in welfare distribution from one period to
another, it is necessary to deal with the entire distribution. However, one can use
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
76 Counterfactual Decomposition of Changes in Poverty Outcomes
influence functions to establish a direct link between the distributional statistic
of interest and individual (or household) characteristics and make suitable
assumptions that allow counterfactual decomposition à la Oaxaca-Blinder.
The influence function of a functional q (F ) is basically its first-order direc-
tional derivative (Hampel 1974). Let G(b) be a mixture of two distributions F
and H such that G(b) = bH + (1 − b)F. This expression says that an observation
under G(∙) is randomly sampled from H with probability b or from F with prob-
ability (1 − b). The directional derivative of q at F in the direction of H tells us
how the functional q changes as G gets closer and closer to F. Formally, we write
∇qG→F = lim
( )
q G (b ) − q ( F )
=
∂
( )
q bH + (1 − b ) F |b = 0.(4.6)
b→ 0 b ∂b
Let H = ∆y, the distribution function for a probability measure that assigns
mass 1 to y in the domain of F. In other words, ∆y(u) is equal to one if y ≤ u ;
otherwise it is equal to zero.5 In this case, G(b) = b∆y + (1 − b)F, and the influence
function of the functional q (F) can be written as6
IF ( y ; q , F ) = ∇q F →∆ y .(4.7)
This is an indicator of the relative effect of a small perturbation in F on q (F).
In that sense, it is a measure of robustness.7 The influence function defined in
equation (4.7) measures the effect that a single observation has on a functional.
It is useful at this stage to note that the directional derivative of the mean of
F in the direction of H is equal to
∂
∂b
∫( )
v bh ( v ) + (1 − b ) f ( v ) dv
b = 0
=
∫ vh ( v ) dv − ∫ vf ( v ) dv. (4.8)
In other words, the directional derivative of the mean of F at F in the direction
of H is equal to ∇µF→H = µH − µF. Hence, the influence function of the mean of
F is equal to
IF ( y ; µ , F ) = ∇µF →∆ y = y − µF . (4.9)
The expected value of the influence function of a distributional statistic is
equal to zero in all cases in which the frequencies and the range of the y-values
are bounded. Firpo, Fortin, and Lemieux (2009) define the recentered or rescaled
influence function (RIF) as the leading two terms of a von Mises (1947) linear
approximation of the associated functional.8 It is equal to the functional plus the
corresponding influence function. Let IF(y; q ) stand for the influence function
of q (Fy), then RIF(y; q ) = q (Fy) + IF(y; q ). The fact that the expected value
of the influence function is equal to zero implies that the expected value of
the RIF is equal to the corresponding distributional statistic. In other words,
q (Fy) = E[RIF(y; q )]. By the law of iterated expectations, the distributional sta-
tistic of interest can be written as the conditional expectation of the recentered
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Counterfactual Decomposition of Changes in Poverty Outcomes 77
influence function (given the observable covariates, x). This is the RIF regression
that for q (Fy) can be expressed as E[RIF( y; q )|x]. The distributional statistic
q (Fy) can therefore be expressed in terms of this conditional expectation as
follows:
q ( Fy ) =
∫ E RIF ( y; q )|x dF ( x ). (4.10)
This expression suggests that to assess the impact of covariates on q (Fy), one
needs to integrate over the conditional expectation E[RIF(y; q )|x]. This can be
easily done using regression methods. At this point, one has a choice between
linear and nonlinear models.
Linear and Nonlinear Specifications. If one assumes linearity, the conditional
expectation of the RIF can be written as a linear function of observable covariates
as follows: E[RIF(y;q )|x] = xb. One can then apply OLS to the following
equation:
RIF( y;q ) = xb + e.(4.11)
Fortin, Lemieux, and Firpo (2011) explain that the expected value of the
linear approximation of the RIF regression is equal to the expected value of the
true conditional expectation because the expected value of the approximation
error is zero. This fact makes the extension of the standard Oaxaca-Blinder
decomposition to RIF regressions both simple and meaningful. Interestingly, the
influence function for the mean presented in equation (4.9) implies that the cor-
responding recentered influence function is RIF(y; m, F ) = y. Hence, the ordinary
linear regression of y on a set of covariates x is indeed an RIF regression. So the
standard Oaxaca-Blinder decomposition is in fact based on RIF regression.
Based on equation (4.11), we find that the endowment effect identified on
the basis of an RIF regression can be written as follows:
∆q E ( x|t = 1) − E ( x|t = 0 )
X = ⋅ β0 . (4.12)
The corresponding structural effect is
S = E ( x|t = 1) ⋅ ( β1 − β 0 ) . (4.13)
∆q
This decomposition may involve a bias, because the linear specification is only
a local approximation that may not hold in the case of large changes in covari-
ates.9 The solution to this problem is to combine the reweighing method
described earlier with RIF regression.10
Given the analogy between the mean and the class of additively separable
poverty measures, equation (4.9) implies that the influence function of any
member of that class is
IF( y; P, F ) = I( y ≤ z)y (y|z) − P(F; z),(4.14)
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
78 Counterfactual Decomposition of Changes in Poverty Outcomes
where I( y ≤ z) is an indicator function and P(F; z) is the poverty measure written
in a way emphasizing that this distributional statistic is a functional of F (Cowell
and Victoria-Feser 1996; Essama-Nssah and Lambert 2012). The corresponding
recentered influence function is
RIF( y; P, F ) = I( y ≤ z)y (y|z).(4.15)
For these poverty measures, one can in fact use nonlinear specifications of the
RIF regression (for example, logit or probit for the headcount index and tobit for
the other poverty measures).11 Fairlie (2005) proposes an extension of the
Oaxaca-Blinder decomposition to logit and probit models, while Sinning, Hahn,
and Bauer (2008) explain how to extend this method to nonlinear models in
general.12
There is also an indirect way of identifying (and hence estimating) the com-
position and structural effects that account for the variation in poverty measures
that are members of the additively separable class. As noted in chapter 2, varia-
tion in poverty over time can be expressed as a weighted sum of points on the
growth incidence curve (GIC). Since the GIC describes outcome variation across
quantiles, one can use RIF regression to perform decompositions of differences
across quantiles and thus decompose growth incidence into the composition and
structural effects. One would need RIFs for the quantiles under consideration.
Firpo, Fortin, and Lemieux (2009) note that the RIF of the τth quantile of the
distribution of y is the following:13
τ − I ( y ≤ qτ )
RIF ( y ; qτ ) = qτ + IF ( y ; qτ ) = qτ + ,
(4.16)
f y ( qτ )
where I(∙) is an indicator function for whether the outcome variable y is less than
or equal to the τth quantile and fy(qτ ) is the density function of y evaluated at
the τth quantile.
One can use equation (4.16) repeatedly to decompose, for instance, the first
99 quantiles (percentiles) of the outcome distribution of interest. This means
that we can decompose the GIC into a component associated with the composi-
tion (or endowment) effect and a second one related to the structural (or the
price) effect. Formally, we express this decomposition as follows:
g( y) = gx( y) + gs( y).(4.17)
Equation (4.17) implies that the elasticity of poverty with respect to the mean
z
1
outcome (ζ P ( g ) =
γP ∫
yy ′ ( y|z ) g ( y ) f ( y ) dy ) can also be decomposed into a
0
composition and a structural effect as follows:
z z
1 1
ζP (g ) =
γP ∫
yy ′ ( y|z ) g X ( y ) f ( y ) dy +
0
γP∫
0
yy ′ ( y|z ) g S ( y ) f ( y ) dy. (4.18)
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Counterfactual Decomposition of Changes in Poverty Outcomes 79
The first term on the right-hand side of equation (4.18) represents the endow-
ment effect based on the corresponding effect for the GIC. Similarly, the second
term is the structural effect based on the corresponding effect for the GIC.
It is important to note that this indirect decomposition applies not only to this
class of poverty measures but also to all additively separable social evaluation
functions such as the Atkinson (1970) and the S-Gini welfare functions.14 This
fact implies that the rate of growth of per capita income, γ, can also be decom-
posed into two components reflecting the composition and structural effects. It
can be shown that when there is no aversion to inequality, the Atkinson welfare
function ranks social states on the basis of the average outcome (such as average
income or expenditure).
Thus, the rate of change in social welfare induced by a distributional change
is captured by the per capita rate of growth, which can be expressed as a
my
∆µ y
weighted sum of points along the GIC. Indeed, γ =
µ = ∫ µ g ( y ) dF ( y ), where
0
my stands for the maximum income or expenditure, and each point on the GIC
is weighted by the slope of the Lorenz curve at that point. Clearly, this rate of
growth can be decomposed on the basis of equation (4.17) into an endowment
effect and a structural effect. Given that the level and pattern of growth depend
on factor accumulation and productivity, we can interpret the endowment effect
as an indicator of changes in factor accumulation and the structural effect as an
indicator of changes in productivity.
Some Advantages of Linearity. As noted earlier, one has the choice between non-
linear and linear specifications of the RIF regression model. Linearity has the
added advantage of making it possible to perform a detailed decomposition that
can further decompose each of these effects in terms of the contributions of the
relevant covariates. Fortin, Lemieux, and Firpo (2011) explain that a decomposi-
tion approach provides a detailed decomposition when it allows one to apportion
the composition effect or the structural effect into components attributable to
each explanatory variable. The contribution of each explanatory variable to the
composition effect is analogous to what Rothe (2010) calls a “partial composition
effect.”15
Assuming a linear RIF regression model, let xk and bk stand, respectively, for the
kth element of x and b. Then the endowment and price effects can be written in
terms of sums over the explanatory variables. For the endowment effect, we have
m
∆q
X = ∑ E ( x |t = 1) − E( x |t = 0) β
k =1
k k 0k . (4.19)
Similarly, for the structural effect, we have the following expression:
m
∆q
S = ∑E ( x |t = 1)( β
k =1
k 1k − β0 k ).(4.20)
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
80 Counterfactual Decomposition of Changes in Poverty Outcomes
Expressions (4.19) and (4.20) provide a simple way of dividing the endow-
ment and price effects into the contribution of a single covariate or a group of
covariates as needed.16 Such detailed decompositions are not easy to obtain for
nonlinear models.
Changes along the Entire Outcome Distribution
The decomposition of changes in distributional statistics generally produces
information about the aggregate composition and structural effects. Although the
linearity assumption can help identify the contribution of various covariates to
these aggregate effects, it does not tell us how these effects vary along the entire
outcome distribution.
To obtain this information, we now consider the decomposition of differ-
ences in density functions and across quantiles. The decomposition of changes
in density functions relies on nonparametric methods. In the case of quantiles,
we focus on the use of quantile regression (a parametric method). All the
methods reviewed in this section are purely statistical in the sense that they
all rely on models of the conditional distribution of outcomes given the
covariates.
Differences in Density Functions
For decomposition purposes, one needs a model that links the outcome of inter-
est to household characteristics. To focus on differences in density functions, we
maintain that the outcome variable y has a joint distribution with characteris-
tics, x. This distribution is characterized by the following joint density function:
Jt( y, x), t = 0, 1. For instance, poverty analysis relies on household consumption
expenditure (y) as a measure of welfare. Thus data from a household income and
expenditure survey characterize the joint distribution of expenditure and house-
hold characteristics (x).
Just as the standard Oaxaca-Blinder decomposition is based on the
unconditional mean, the generalization of that method considered here
requires the marginal distribution of y noted as ft( y). This marginal density
function can be obtained by integrating the covariates x out of the joint
density. Furthermore, the factorization principle allows one to write the
joint density as a product of the distribution of y conditional on x, gt( y|x),
and the joint distribution of characteristics, ht(x). These are the two factors
underpinning the decomposition. Any change in the marginal outcome
distribution induced by a variation in the distribution of observed character-
istics (ceteris paribus) represents the endowment effect, while any change in
the distribution associated with a (ceteris paribus) variation in the condi-
tional distribution is interpreted as the price-behavioral effect (Bourguignon
and Ferreira 2005).
To see clearly what is involved,17 we express the joint density function as a
product of the two underlying functions: Jt( y, x) = gt( y|x)ht(x), t = 0, 1. On the
basis of this factorization, we can write the marginal density of y in a way that
facilitates the expression and interpretation of the decomposition results, that is,
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Counterfactual Decomposition of Changes in Poverty Outcomes 81
ft ( y ) ≡ f gt
ht
( y ). Thus the observed change in the outcome distribution between
the two periods can be stated as follows:
∆f = f1 ( y ) − f0 ( y ) ≡ f gh11 ( y ) − f gh00 ( y ). (4.21)
We can add to and subtract from the difference defined in (4.21) the follow-
1
ing counterfactual:18 f gh0 ( y ). This is the marginal density function that would
obtain if the conditional distribution were that of period 0 and if the joint distri-
bution of characteristics were that prevailing in period 1. This transformation
leads us to the following generalized decomposition of changes in the marginal
density of y:
f g 0 ( y ) − f g 0 ( y )
∆f = f g1 ( y ) − f g 0 ( y )
+
h1 h0 h1 h1
. (4.22)
The configuration of the indices (subscripts and superscripts) for the marginal
distributions involved in (4.22) suggests an interpretation of the various compo-
nents of the decomposition. The first component on the right-hand side is the
endowment effect (based on changes in the joint distribution of observed charac-
teristics). The second component measures the price-behavioral effect (linked to
the change in the conditional distribution of y, which, in fact, also includes the
effect of unobservables).
Kernel Density Approaches. In their study of the role of institutional factors in
accounting for changes in the distribution of wages in the United States, DiNardo,
Fortin, and Lemieux (1996) demonstrate how to implement empirically the
above decomposition using kernel density methods to estimate the relevant den-
sity functions. The histogram is the oldest and most common density estimator
(Silverman 1986), and kernel methods may be viewed as ways of smoothing a
histogram.19
The basic idea is to estimate the density f(y) by the proportion of the sample
that is near y. One way of proceeding is to choose some interval or “band” and to
count the points in the band around each y and normalize the count by the
sample size multiplied by the bandwidth. The whole procedure can be viewed
as sliding the band (or window) along the range of y, calculating the fraction of
the sample per unit within the band, and plotting the result as an estimate of the
density at the midpoint of the band (Deaton 1997).20
The kernel estimate of the density function ft( y) can be written as follows:
nt
yit − y
∑K
1
ft ( y ) = , t = 0, 1,(4.23)
hnt h
i =1
where h is the bandwidth representing the smoothing parameter, nt is the
sample size for period t, y is the focal point where the density is estimated,
and K(∙) is the kernel function. A kernel function is essentially a weighting
function chosen to give more weight to points near y and less weight to those
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
82 Counterfactual Decomposition of Changes in Poverty Outcomes
far away. In particular, it will assign a weight of zero to points just outside and
just inside the band.
As a weighting function, the kernel function should satisfy four basic proper-
ties: (a) positive, (b) integrate to unity over the band, (c) symmetric around
zero so that points below y get the same weight as those an equal distance
above, and (d) decreasing in the absolute value of its argument. The most com-
mon kernel functions used in empirical work are the Gaussian and Epanechnikov
kernels.21
The counterfactual density function that is the linchpin of the decomposi-
tion presented in equation (4.22) can be written in a manner analogous to the
distribution functions underlying the decomposition presented in equation
(4.3).22 In other words, f gh0 1
( y ) = f y0|t = 1. This density can be estimated by
reweighing the kernel estimate for period 0 using the same function as the one
underlying the counterfactual distribution defined in equation (4.5). The
resulting expression is
n0
yi 0 − y
∑w ( x ) K
1
fˆy0 |t =1 ( y ) = . (4.24)
hn0 h
i =1
A Semiparametric Approach. Machado and Mata (2005) propose a semiparamet-
ric approach to estimating the density functions needed in the above decomposi-
tion. Their approach is based on a two-step procedure to derive marginal density
functions from the conditional quantile process that fully characterizes the con-
ditional distribution of y given the covariates x. Specifically, these authors model
the conditional distribution of y given x by a linear conditional quantile function
as follows:
qt ( yt|xt) = xt bt(t ), t ∈(0, 1), t = 0, 1. (4.25)
The second step entails estimating the marginal density function of y that is
consistent with the conditional quantile process defined by equation (4.25). This
is achieved by running the following algorithm:
1. Draw a random sample of size m from a uniform distribution on [0, 1] to
get tj for j = 1, 2, …, m.
2. For each tj, use available data to estimate the quantile regression model and
get m estimates of coefficients β t (τ j ) , j = 1, 2, …, m.
3. Given that xt is a (nt x k) matrix of data on covariates, draw a random sample
s
of size m from the rows of xt and denote each such sample by x jt .
4. The corresponding values of the outcome variable are given by
s
y jt s
≡ x jt β t (τ j ) , j = 1, 2, …, m .
The validity of this procedure stems from the probability integral transforma-
tion theorem, which states that if u is a random variable uniformly distributed
over [0, 1], then y = F −1(u) is distributed like F. Here tj is assumed to be a
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Counterfactual Decomposition of Changes in Poverty Outcomes 83
realization of Fyt |xt . Given model (4.25), the corresponding conditional quantile
regression model can be written as
| xt (τ j , x t ) = x t β t (τ j ) , t = 0, 1.
qτ j ( yt |x t ) = Fy−t1 (4.26)
A modified version of the above algorithm leads to the critical counterfactual
upon which the decomposition is based. Recall that the counterfactual of interest
is the density function of the outcome in period 1, assuming that the character-
istics of that period had been rewarded according to the system prevailing in
period 0. This counterfactual can be estimated by applying the above algorithm
to the data for period 0, except that at stage 3, covariates must be drawn from
data for period 1. On the basis of equation (4.26), the conditional regression
model associated with this counterfactual is the following:
qτ j ( y0 |x1 ) = Fy−01|x0 (τ j , x1 ) = x1β0 (τ j ). (4.27)
As noted by Fortin, Lemieux, and Firpo (2011), this approach is computa-
tionally demanding. They suggest a simplification based on the estimation of a
large number of quantile regressions (say, 99) instead of using the random pro-
cess. The conditional quantile function can then be inverted to obtain the con-
ditional cumulative distribution that must be averaged over the empirical
distribution of the covariates to yield the unconditional distribution function. In
fact, Machado and Mata (2005) acknowledge that this is a viable alternative to
their method.
Differences across Quantiles
As discussed earlier in the context of RIF regression analysis, one can also work
with quantiles, which are easier for detailed decompositions, instead of density
functions to decompose changes along the entire outcome distribution.
Because the decomposition must be based on marginal distributions, one
needs to work with marginal quantiles, not conditional ones. There are
multiple ways to go about it. An alternative to using RIF regression is to derive
marginal quantiles from equations (4.25) and (4.26)—based on the Machado
and Mata (2005) procedure—or by numerical integration as proposed by
Melly (2005).
To link conditional quantiles to marginal quantiles, Angrist and Pischke (2009)
start from the observation that the proportion of the population below qt condi-
tional on x is equal to the proportion of conditional quantiles that are below qt.
Let I(∙) be the indicator function that takes a value of one if its argument is true
) stand for the conditional cumulative distri-
and zero otherwise. Again, let Fy|x(
bution function (CDF) of y given x. Thus the proportion of the population for
1
whom the outcome y is less than qt is equal to Fy|x ( qτ |x ) = I
∫
Fy|x (τ | x ) < qτ
−1
0
dτ ,
where the term on the right-hand side is equal to the proportion of conditional
quantiles that are below qt .
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
84 Counterfactual Decomposition of Changes in Poverty Outcomes
On the basis of equation (4.25), we can rewrite this proportion as
1
∫ xβ (τ ) < qτ
Fy|x ( qτ |x ) = I
0
dτ
), from which
. The marginal distribution of y, Fy(
one derives marginal quantiles, is obtained by integrating the conditional distri-
bution over the whole range of the distribution of the covariates (Melly 2005).
1
The resulting expression is Fy ( qτ ) = I
0
x β (τ ) <
∫ ∫
q
τ dτ dFx . The sample ana-
log of this expression—based on an estimation of quantile regressions at every
percentile for a sample of size n—is given by the following expression (Angrist
and Pischke 2009):
n 1 1
∑ ∑
1
F y ( qτ ) = I ( x i βτ < qτ ) .(4.28)
n 100 τ = 0
i =1
The marginal quantile corresponding to the above estimator of the marginal
distribution of the response variable is obtained by inverting equation (4.28). We
( ) {
note these marginal quantiles as qτ x t , β t (τ ) = inf q : F y ( qτ ) . }
The generalized Oaxaca-Blinder decomposition described by equation (4.3)
can equivalently be stated in terms of these marginal quantiles. The observed
change in the marginal distribution of the response variable is now written as
( ) ( )
∆qτ = qτ x1 , β 1 (τ ) − qτ x 0 , β 0 (τ ) . To distinguish the endowment effect from
the price effect, we subtract from and add to this expression the following coun-
( )
terfactual outcome: qτ x1 , β 0 (τ ) . This counterfactual involves the characteristics
of period 1 evaluated with the prices (coefficients) of period 0. The correspond-
ing decomposition analogous to expression (4.3) is the following:
∆qτ =
( ) (
qτ x1 , β 0 (τ ) − qτ x 0 , β 0 (τ ) ) ( ) ( )
+ qτ x1 , β (τ ) − qτ x1 , β (τ ) .(4.29)
1 0
Consistent with equation (4.3), the first term on the right-hand side of equa-
tion (4.29) is the endowment effect at the τth quantile, while the second term
measures the price effect at the same location. Again, changes along the entire
outcome distribution are obtained by performing this decomposition for the first
99 percentiles.
Accounting for the Contribution of Unobservables
Recall that based on equation (4.2), the contribution of unobservable
characteristics into changes in the outcome distribution has at least two potential
components: the first relates to changes in the returns to unobservables and the
second to the distribution of these characteristics. All the decomposition meth-
ods discussed so far lump the first component together with the returns to
observables in the structural effect. Furthermore, the contribution of changes in
the distribution of unobservable characteristics is ruled out by either the ignor-
ability assumption or the zero-conditional-mean assumption. The issue now is
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Counterfactual Decomposition of Changes in Poverty Outcomes 85
this: Under what conditions can we identify these effects that up to now have
been swept under the rug, so to speak?
Further Decomposition of the Structural Effect
Juhn, Murphy, and Pierce (1993) assume additive linearity for the outcome
model and conditional rank preservation to decompose differences in outcome
distributions in a way that accounts for the contribution of unobservables.23
Under additive linearity, the function defining the outcome variable is separable
in x and e. We can therefore write the outcome model as follows:
yt = xtbt + ut, t = 0, 1, (4.30)
where ut = ζt(e), some function of unobservable characteristics.
The assumption of conditional rank preservation means that a given individual
has the same rank in the distribution of u0 as in the distribution of u1, conditional
on observable characteristics. To see this formally, let Fu |x(ut|xt) stand for the dis-
tribution of ut conditional on xt. Also, let τi0(xi) = Fu |x(ui0|xi) be the rank of indi-
vidual i with observed characteristics xi in the conditional distribution of u0 given
x, and let τi1(xi) = Fu |x(ui1|xi) be individual i’s rank in the conditional distribution
of u1 given x. Conditional rank preservation says that τi0(xi) = τi1(xi). Fortin,
Lemieux, and Firpo (2011) explain that one can secure conditional rank invari-
ance by assuming ignorability and that the functions ut are strictly increasing in e.
In other words, these functions are monotonic.24
As expected, separability allows the analyst to construct counterfactuals sepa-
rately for observables and unobservables. To see what is involved, consider the
case of a particular individual, i, with outcome yi1 = xi1b1 + ui1 in period 1. Let
v ic0 represent what the residual part of the outcome would have been had the
unobservable characteristics of this individual been treated as in the initial
period, ceteris paribus. The corresponding counterfactual for the full outcome is
yic0 = x i1β1 + v ic0. Comparing this counterfactual with the observed outcome
reveals the contribution of changes in the returns to unobservable characteristics
of individual i to the overall change in the individual’s outcome. We denote this
as ∆ S
c
y
( ) (c
) c
,σ = y i1 − y i 0 = v i1 − v i 0 . Next, we replace b1 with b0 in the expression for
yi 0. This operation yields the following counterfactual: yib0 = x i1β0 + v ic0. Let
∆S y
( )
,β = y i 0 − y i 0 . This term is equivalent to ∆ S ,β = x i1 ( β1 − β 0 ) and clearly shows
c b y
the contribution of changes in the returns to observable characteristics.
Thus, separability along with ignorability and monotonicity make it possible to
split the structural effect into (a) one component resulting from changes in the
returns to observables and (b) another linked solely to changes in returns to unob-
y y y
servables. In other words, the total structural effect25 is equal to ∆ S = ∆S , β + ∆S, σ
.
y
The composition effect ∆ X can be identified residually from the following
y
expression: ∆O y
− ∆Sy
= ∆Xy
+ ∆ε y
, where ∆O = ( yi1 − yi 0 ). The assumption of
y
conditional independence implies, however, that ∆ ε
= 0. Recall that this
assumption implies that the conditional distribution of unobservables does
not vary across groups (periods). Therefore, under the prevailing identifying
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
86 Counterfactual Decomposition of Changes in Poverty Outcomes
assumptions, the difference between the overall outcome difference and the
identifies the observable composition effect.
structural effect
The question now is, how does one identify v ic0? This is where rank preserva-
tion comes in. This assumption leads to the following imputation rule:
v ic0 = Fv−1
(
0 |x )
τ i1 ( x i ) . (4.31)
This imputation rule says that for individual i in the end period, the counter-
factual for the residual outcome is equal to the residual outcome associated with
the individual located at the same rank in the conditional distribution of residual
outcomes in the base period. In practice, one would estimate b0 and b1 using
OLS. Bourguignon and Ferreira (2005) explain that empirical implementation of
a rank-preserving transformation is complicated by the fact that both samples do
not necessarily have the same number of observations. However, if one is willing
to assume that both distributions are the same up to some proportional transfor-
mation, then the rank-preserving transformation can be approximated by multi-
plying residuals in the base period by the ratio of the standard deviation in the
end period to the one in the initial period.
Fortin, Lemieux, and Firpo (2011) point out that assuming constant returns
to unobservables and homoskedasticity allows one to write the unobserved com-
ponent of the outcome as ut = ste. Homoskedasticity implies that the conditional
variance of e is constant (and can be normalized to one). Equation (4.30) can
therefore be written as follows:
yt = xtbt + ste, t = 0, 1. (4.32)
Applications and Limitations of Juhn-Murphy-Pierce Model
As it turns out, this is the version of the model used by Juhn, Murphy, and Pierce
(1991) in their study of the evolution of the wage differential between blacks
and whites in the United States. In that context, the standard deviation of the
residuals in the wage equation stands both for within-group inequality in the
wage distribution and for the price of unobserved skills (Yun 2009).
The outcome model specified in equation (4.32) has also been used to study
the gender pay gap. In that context, t = 1 is taken to represent males while t = 0
stands for females, and the wage regime for males is considered the nondiscrimi-
natory one. The counterfactual used in the decomposition is the outcome that
female workers would have experienced if they had been paid like their male
counterparts. Care must be taken when applying this version of the model to
decompose differences in mean outcome using OLS, because the OLS residuals
sum up to zero. To see this, consider the following expression of standard
Oaxaca-Blinder decomposition that explicitly shows the residuals:
β1 + E ( x 0 )( β1 − β0 )
E ( x1 ) − E ( x 0 )
µ
∆O =
(4.33)
E (∈
+ σ 1 + E ( ε 0 )(σ 1 − σ 0 ).
1 ) − E (ε 0 )
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Counterfactual Decomposition of Changes in Poverty Outcomes 87
The terms associated with the unobservables in the right-hand side of equa-
tion (4.33) will disappear if the decomposition is based on OLS applied to each
equation separately.
To get around this issue, Juhn, Murphy, and Pierce (1991) assume that the
returns to observable characteristics are the same for both groups and apply OLS
to only one group, constructing an auxiliary equation for the other group. In the
context of gender wage gap studies, OLS is applied to the equation for males
only. The equation for female workers is constructed as follows: y0 = x0 b1 + υ0.
The implied decomposition is
β1 − E ( v 0 ) =
E ( x1 ) − E ( x 0 ) β1 − σ 1E (η0 ) , (4.34)
E ( x1 ) − E ( x 0 )
µ
∆O =
v0
where η0 = . The above expression is computed on the basis of the sample
σ1
analogs of the parameters of interest. The first term in the twofold decomposi-
tion presented in equation (4.34) represents the predicted gap, while the second
stands for the residual gap.
As Yun (2009) points out, the residual gap is equal to the structural effect in
the standard Oaxaca-Blinder decomposition. Yet, this structural effect represents
returns to observable characteristics. It is therefore hard to see how the Juhn,
Murphy, and Pierce (1991) procedure helps to identify the contribution of unob-
servables. Yun (2009) proposes instead the decomposition defined by equation
(4.33), under the assumption that the expected value of unobservable terms is
not equal to zero. However, that author does not provide an implementation
procedure corresponding to this situation.
Relationship to Treatment Effect Analysis
As Fortin, Lemieux, and Firpo (2011) point out, there is a powerful analogy
between the classic Oaxaca-Blinder decomposition method and treatment effect
analysis.26 Treatment impact analysis seeks to identify and estimate the average
effect of treatment (intervention) on the treated (those exposed to an interven-
tion) on the basis of the difference in average outcomes between the treated and
a comparison group.
In that context, t indicates treatment status. It is equal to one for the treated
and zero for the untreated (the comparison or control group). The expression,
E ( y1 ) − E ( y0 )
µ
∆O = , can therefore be interpreted as the difference in average
outcomes between the treated and untreated. Under the assumptions underlying
the basic Oaxaca-Blinder method,27 this difference is clearly the result of differ-
ences in observable characteristics (the composition effect) and in treatment
status. The part resulting from the difference in treatment status is known as the
average treatment effect on the treated (ATET) and is in fact equal to the struc-
tural or price effect.
Countering Selection Bias
Note that the conventional approach to impact evaluation also relies on ceteris
paribus variation of treatment to identify its average effect on the treated. Within
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
88 Counterfactual Decomposition of Changes in Poverty Outcomes
that logic, the composition effect is equivalent to selection bias that must be
driven to zero by the use of randomization, propensity score matching, or similar
methods.
Randomization. To ensure that the distribution of observed and unobserved char-
acteristics is the same for both the treated and the control groups, randomization
is employed. By balancing observed and unobserved characteristics between the
groups before the administration of treatment, randomization guarantees that
the average difference in outcome between the two groups is the result of treat-
ment alone, hence the causal interpretation given to this parameter under those
circumstances. In other words, the first term on the right-hand side of equation
(4.3) (that is, the endowment effect or selection bias) is equal to zero under
random assignment to treatment and full compliance.28 Clearly, randomization is
designed to implement a ceteris paribus variation in treatment.
Conditioning by Stratification. In the context of observational studies where the
investigator does not have control over the assignment of subjects to treatment,
the determination of the causal effect of treatment hinges critically on the under-
standing of the underlying treatment assignment or selection mechanism, which
must explain how people end up in alternative treatment states. The assumption
of selection on observables (also known as ignorability) is often invoked to imple-
ment ceteris paribus identification of ATET through conditioning by
stratification.
Propensity Score Matching. Basically, conditioning by stratification entails compar-
ing only those subjects with the same value of covariates x across the two groups
(treated and untreated). This type of selection of individuals from the two groups
is known as matching. A potential dimension problem is associated with match-
ing when there are many observable characteristics taking many values. Insisting
on conditioning based on exact values can lead to too few observations in each
subgroup characterized by these observables. This dimensionality problem can
be resolved by matching on the propensity score, that is, the conditional probabil-
ity of receiving treatment given observable characteristics (Rosenbaum and
Rubin 1983).
Usefulness of Treatment Effect Analysis
The analogy between treatment effect analysis and the Oaxaca-Blinder
decomposition method has been extremely useful for the development of flexi-
ble estimation methods for endowment and structural effects. As noted earlier,
selection on observables implies that the conditional distribution of unobservable
factors is the same in both groups (treated and comparison). Although this
assumption is weaker than the zero-conditional-mean assumption29 used in the
standard Oaxaca-Blinder decomposition, it is enough to secure identification and
µ
consistent estimation of the ATET and hence the structural effect, ∆ S (in the
Oaxaca-Blinder framework).
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Counterfactual Decomposition of Changes in Poverty Outcomes 89
Fortin, Lemieux, and Firpo (2011) give the example of education and
nobservable ability. They explain that if education and ability are correlated,
u
this creates an endogeneity problem that prevents a linear regression of earnings
on education to produce consistent estimates of the structural parameters mea-
suring the return to education. Yet the aggregate decomposition remains valid as
long as the correlation between ability and education is the same in both groups.
A major implication of the difference in identification assumptions between
the traditional Oaxaca-Blinder approach and treatment effect analysis is that
consistent estimators of the ATET such as IPW and matching can be used to
µ
estimate the structural effect (∆ S ) even if the underlying relationship between
the outcome and covariates is not linear. Given such an estimate, the composition
effect can be calculated as a residual from the overall mean difference as follows:
∆µ µ µ
X = ∆ O − ∆ S . In particular, decomposition methods based on this weighting
procedure are known to be efficient. It is also worth noting that the generaliza-
tion of the Oaxaca-Blinder decomposition to variation in other distributional
statistics such as quantiles, poverty, and inequality measures enables an analyst to
study the distributional and poverty impacts of an assigned intervention.
Limitations of the Analogy with Treatment Effect Analysis
Although treatment effect analysis can help with the identification and estima-
tion of the structural effect, it is notable that this effect does not necessarily
inherit the causal interpretation generally enjoyed by the ATET for two basic
reasons: (a) In many cases, group membership is not the result of a choice or an
exogenous assignment but is rather a consequence of an intrinsic characteristic,
such as gender or race, and (b) many of the observable covariates are not equiva-
lent to the pretreatment variables that are not supposed to be affected by the
treatment.
The standard Oaxaca-Blinder decomposition method has two other impor-
tant limitations: First, each covariate’s contribution to the structural effect is
highly sensitive to the choice of the omitted group when the explanatory vari-
ables include a categorical variable. (Jann [2008] discusses possible solutions to
this problem.)
Second, the decomposition provides consistent estimates only under the
assumption that the conditional expectation is linear. Under the linearity
assumption, the counterfactual average when t = 1 is simply equal to E(x1|t = 1)·b0.
This is estimated by the cross product of sample means of characteristics for t = 1
with the relevant OLS coefficients from t = 0. The corresponding estimate is x1 β 0.
The counterfactual mean outcome will not be equal to this term when linearity
does not hold.
One possible solution is to reweigh the sample for t = 0 using the inverse
probability method discussed earlier and to compute the counterfactual mean
outcome on the basis of statistics from the reweighed counterfactual sample. Let
c
c
x0 be the vector of the means of adjusted covariates in t = 0 and β 0 the corre-
sponding least squares coefficients. Then the correct counterfactual
c
mean out-
come when the linearity assumption does not hold is x 0 c
β 0. This is the term to
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
90 Counterfactual Decomposition of Changes in Poverty Outcomes
add to and subtract from the empirical version of the overall difference in mean
outcome to get the appropriate estimates of the endowment and structural
effects when the linearity assumption fails.
Accounting for Behavior
A key limitation of the decomposition methods discussed so far is that they do
not account for changes in agents’ behavior in response to changes in their socio-
economic environments—whether those changes are the result of economic
shocks, policy reform, or other causes. Given the maintained hypothesis—that
the living standard of individuals in a given society depends crucially on what
they do with their assets (innate and external) subject to the opportunities
offered by that society—this section focuses on modeling agents’ behavior to
account for their reactions to changes in their socioeconomic environment.
Standard economic theory explains behavior in terms of the principles of
optimization and market interaction. Modeling behavior entails the specification of
the following elements (Varian 1984):
• The actions a socioeconomic agent can take.
• The constraints the agent faces.
• The objective function used to evaluate feasible actions.
The assumption that the agent seeks to maximize the objective function
subject to constraints implies that the outcome variables used to represent the
consequences of behavior can be expressed as functions of parameters of the
socioeconomic environment, embedded in the constraints facing the agent. We
consider the Roy (1951) model of choice and consequences along with its inter-
pretation in the context of modeling the determinants of the living standard. This
model stems from the optimization principle and applies to discrete choice
problems. We also discuss key considerations in simulating counterfactual distri-
butions underlying any decomposition exercise.
The Roy Model of Choice and Consequences
Heckman and Honoré (1990) explain that the original Roy (1951) model was
designed for the study of occupational choice and its implications for the distri-
bution of earnings in an economy where agents are endowed with different sets
of occupation-specific skills. In the two-skill Roy model, each income-maximizing
agent has one skill usable only in one sector and a second skill usable in only
another sector (Heckman and Honoré 1990). Therefore, the two-skilled agent
can freely choose to work in only one of two activities (in this case, either fishing
or hunting) on the basis of their productivity in each.
Under this scenario, an agent with a given skill endowment will choose to
work in the sector where the potential income is higher. There are no investment
opportunities to augment sector-specific skills, nor are there costs associated with
changing sectors. These authors also show that self-selection implies a lower level
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Counterfactual Decomposition of Changes in Poverty Outcomes 91
of inequality in earnings compared with a benchmark case where workers are
randomly assigned to jobs. The fact that occupational choice has significant
implications for earnings distribution makes the Roy model a relevant framework
for analyzing agents’ behavioral responses to changes in their socioeconomic
environment.
Heckman and Sedlacek (1985) discuss an extension of the basic framework
by including nonmarket activity as an option in the choice set facing socioeco-
nomic agents who are now assumed to maximize utility instead of income. The
utility of participating in each of the sectors depends on both sector-specific
attributes (such as the wage rate, employment risk, or job status) and individual
characteristics. That we observe only sectoral choices and not the underlying util-
ity function means it is possible to identify only parameters associated with
differences in utility across sectors. These authors also consider the contribution
of self-selection to income inequality and find that in this general model, self-
selection can increase both between- and within-sector inequality compared
with a random allocation of workers to sectors.
At the most fundamental level, the Roy model is characterized by two
components: a selection mechanism and the associated potential outcomes.
These outcomes are possible consequences of the choice made through the selec-
tion mechanism. The extended version of the Roy model is consistent with
discrete choice models to the extent that utility-maximizing agents face a dis-
crete choice set.
Discrete Choice Modeling
Train (2009) characterizes a discrete choice model in terms of two fundamental
elements: the choice set and the decision process (or the decision rule). The choice
set is the collection of alternatives from which the decision maker chooses one.
This set must be exhaustive in the sense that it must include all possible alterna-
tives, the latter being mutually exclusive from the perspective of the decision
maker. Finally, the number of alternatives must be finite. In the case of discrete
models of labor supply, for instance, the choice set can be represented by a few
options, such as not working, working part time, and working full time.
Just as in the case of the consumption-leisure paradigm, the decision process
assumes utility-maximizing behavior (Essama-Nssah 2012). It is therefore
assumed that the decision maker chooses the alternative that provides the great-
est net benefit or utility. Let uhj, j = 1, 2, ... m be the utility that agent h gets from
alternative j. The decision rule implies that the agent chooses alternative k if and
only if uhk > uhj ∀j ≠ k.
This decision-making process is usually framed within the logic of the random
utility model, in which utility has two parts: The first, known as the representative
utility, is a function of some observable characteristics of the decision maker and
of the alternatives (Train 2009). The second component is a set of unobservable
random factors. Formally, the utility function is written as uhj = uhj + ehj, where
u is the representative utility and ε represents the unobserved portion of utility
that is treated as random.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
92 Counterfactual Decomposition of Changes in Poverty Outcomes
Now, the statement that alternative k is chosen if, and only if uhk > uhj ∀j ≠ k
can be equivalently expressed as follows: k is chosen if and only if (ehj − ehk) <
(uhk − uhj) ∀j ≠ k. Because of the uncertainty implied by the random part of the
utility function, one can make only probabilistic statements about the decision
maker’s choice. The probability that agent h chooses option k is defined by the
following expression:30
Phk = Pr[(ehj − ehk) < (uhk − uhj) ∀j ≠ k].(4.35)
The type of discrete choice model derived from the above probability
tatement is determined by the assumptions made about the distribution of the
s
unobserved portion of the utility function. For instance, the common logit model
assumes that the random factors are independently and identically distributed
(iid) extreme value variables for all options. In other words, each choice is inde-
pendent from the others.31
An Interpretation of the Roy Model
Coulombe and McKay (1996) provide an interesting interpretation of the Roy
model that is consistent with our maintained hypothesis that an individual’s liv-
ing standard is a payoff from participation in the life of society. Using the house-
hold as the unit of analysis, these authors argue that a household’s living standard
depends fundamentally on the socioeconomic group to which it belongs (or its
economic activity status). To frame this view within the logic of the Roy model,
the authors further argue that one needs to explain the selection mechanism
leading to the observed socioeconomic group and, conditional on that choice, the
determinants of the living standard in that group. This logic leads to a two-
equation model—one representing the selection mechanism and the second
modeling the living standard conditional on the choice of a particular socioeco-
nomic group.
Modeling the Selection Mechanism. Modeling the selection mechanism boils down
to modeling the probability defined in equation (4.35). Consistent with the ran-
dom utility framework underlying this expression, and assuming that the random
elements are generated independently by an extreme value distribution, the
multinomial logit model can be used to explain the probability of choosing an
option. Formally, we express that probability as
exp ( zhkγ k )
Phk = ,(4.36)
∑
m
1 + exp ( zhjγ j )
j=2
where zhj is the set of relevant explanatory variables, and m is the total number
of socioeconomic groups. The probability defined in (4.36) is essentially the
propensity score.
The specification of the explanatory variables requires a good understanding
of the determinants of the choice of a socioeconomic group. Three
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Counterfactual Decomposition of Changes in Poverty Outcomes 93
technological factors affect this choice in the context of the general Roy model
(Autor 2009):
• The distribution of skills and abilities;
• The correlations among these skills in the population; and
• The technologies for applying these skills.
Coulombe and McKay (1996) make a similar point in a case study of
Mauritania. They define socioeconomic groups in terms of the income- generating
opportunities available to households and their members. In particular, they con-
sider four mutually exclusive and exhaustive groups of households: (a) house-
holds working predominantly as employees (whether in the public or private
sector); (b) those engaged mostly in agricultural self-employment; (c) those
engaged mainly in nonfarm self-employment; and (d) those not in the labor
force. In essence, socioeconomic groups are determined on the basis of the house-
hold’s main economic activity or source of income.
As to the determinants of the choice of socioeconomic group, these authors
argue that the choice depends on variables (such as education, wages, or profit
rates) that affect relative returns from economic activities and consumption pref-
erences. In particular, they make the point that the extent to which household
members choose either self-employment over wage employment or to stay out
of the labor market depends on the interaction between (a) total household labor
supply within and outside the household (a consumption decision) and (b) its
total labor demand (a production decision) for both household members and
hired labor. In other words, the household’s socioeconomic classification reflects
both consumption parameters (such as its demographic composition and the
characteristics of the head of household) and production parameters relevant to
self-employment (such as fixed inputs and variable costs).
Modeling the Living Standards. Equation (4.36) models the selection mechanism.
We need an outcome equation to complete the model—the living standard
conditional on the choice of socioeconomic group—within the logic of the Roy
framework. Following Coulombe and McKay (1996), we let yhk stand for the log
of per capita expenditure for household h in socioeconomic group k, and hhk for
a random disturbance. The outcome equation associated with equation (4.36)
can be written by analogy to the standard Mincer equation (in labor economics)
as follows:
yhk = xhbk + hhk.(4.37)
Equations (4.36) and (4.37) constitute a system designed to explain house-
hold-level living standards. In their case study, Coulombe and McKay (1996)
distinguish two categories of determinants: household-level demographic factors
and group-level demographic factors. These demographic variables include
household size, household composition, and characteristics of the economic head
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
94 Counterfactual Decomposition of Changes in Poverty Outcomes
of the household (such as educational level, marital status, gender, and ethnicity).
Group-specific factors include those affecting total household income. For those
engaged in wage employment, such factors would include level of education,
sector of employment, and numbers of hours worked in a year to account for
seasonal work.
Given that such variables are difficult to measure at the household level (the
unit of analysis), one could either (a) define and measure these variables only for
the economic head of household or (b) adopt some form of aggregation over
household members. Naturally, this would entail some loss of the heterogeneity
found at the individual level. In the case of agricultural self-employment, specific
factors include land size and quality, tenure status, use of fertilizer, insecticides,
hired labor, access to extension services, and commercialization. Similar consid-
erations apply to nonagricultural self-employment. For households outside the
labor market, possible sources of livelihood include asset holdings, borrowing,
and public and private transfers.
Another important consideration here is the classification of variables as exog-
enous or endogenous. This classification hinges on the time horizon chosen. For
instance, in the long run, the living standard can affect demographic variables,
such as household size and composition (Coulombe and McKay 1996). But in
the short run, it is reasonable to think of the direction of influence as running
from demographic variables to the living standard. The Coulombe-McKay study
adopts a short-to-medium time frame so that most of the variables listed above
are considered exogenous with respect to the model described by equations
(4.36) and (4.37).
Further Extensions of the Roy Model
The model presented in Inchauste et al. (2012) and later in chapter 6 of this
volume expands this framework by modeling educational and sectoral choice as
endogenous. For nonfarm workers, employment type is modeled endogenously,
as individuals could be salaried, daily workers, nonfarm self-employed, farm
workers, or not working at all. Similarly, because farm households are likely to be
diversified, a separate model for employment type is specified for farm workers
who engage in a secondary occupation.
Bourguignon, Ferreira, and Leite (2008) go one step further and model
changes in household demographics as endogenous. In that framework,
socioeconomic group, per capita consumption, education, and household
composition are endogenous. Variations in education and in household com-
position are modeled within the discrete choice framework portrayed by
equation (4.36). In that particular application, the demand for education is
modeled on the basis of six alternatives: 0 years of schooling, 1–4 years, 5–6
years, 7–8 years, 9–12 years, and 13 or more. The highest level of education
is the excluded category. The authors considered the following variables to
be purely exogenous: number of adults in the household, region of residence,
age, race, and gender. For household demographics, the options are 0, 1, 2, 3,
4, and 5 or more children. The last category is omitted in the estimation.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Counterfactual Decomposition of Changes in Poverty Outcomes 95
Note that education is an explanatory variable in the demographic multilogit
model.
Leite, Sanchez, and Laderchi (2009) apply this extended framework to
analyze the evolution of urban inequality in Ethiopia. They, too, focus on the
household as the unit of analysis and use per capita household expenditure as the
outcome variable.
Cogneau and Robilliard (2008) use the extended Roy model to study the
implications of three targeted poverty reduction policies in Madagascar: (a) a
direct subsidy on agricultural production prices, (b) a workfare program, and
(c) a uniform untargeted per capita transfer program. While using the household
as the unit of analysis and considering consumption as the ultimate welfare indi-
cator, these authors model first the income-generating process at the level of
individual household members and then link consumption to household income.
Working-age individuals (aged 15–64 years) faced a choice set with three alterna-
tives: family work, self-employment, and wage work. Household composition
and location are exogenously given. For self-employment and wage work, an
individual’s potential earnings are equal to a task price multiplied by a given
idiosyncratic amount of efficient labor. “Efficient labor” is assumed to be a func-
tion of some observable characteristics (such as age, experience, and location)
and unobservable skills. Family work is rewarded by a reservation wage that is a
function of individual and household characteristics.32
In the absence of labor market segmentation, the simple selection rule of the
basic Roy (1951) model would base sector choice on a comparison of the reser-
vation wage and potential wages in the other two sectors. To account for labor
market segmentation, Cogneau and Robilliard (2008) define a segmentation
variable in terms of the relative cost of entry between self-employment and wage
work, and adjust the selection rule accordingly.
For policy evaluation, these authors embed the occupational choice model
into a broader microeconomic module that includes the demand system for con-
sumption goods. To keep things simple, they assume that consumption or saving
decisions are separable from labor supply decisions. They also assume a fixed
common savings rate of 5.2 percent so that aggregate consumption is equal to
the implied consumption propensity multiplied by disposable income. The latter
is the sum of farm profits, labor income, earnings from self-employment, and
nonlabor income, such as capital income and transfers. Total consumption is
allocated to three composite goods (agricultural, informal, and formal) according
to budget shares derived from available data.
The three policies considered—subsidized agricultural production prices, the
workfare program, and the uniform untargeted per capita transfer program—
have the potential of inducing large macroeconomic effects because their
collective cost represents about 5 percent of Madagascar’s gross domestic prod-
uct. To account for this, the authors link the micro module to a small three-
sector (agriculture, informal, and formal) computable general equilibrium model.
The integrated framework makes it possible to consider the macroeconomic
impact of the policy options along with their impact on inequality and poverty.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
96 Counterfactual Decomposition of Changes in Poverty Outcomes
Adding a general equilibrium model removes a key limitation of the decomposi-
tion methods discussed up to this point. These methods rely on either a purely
statistical or a microeconomic model of behavior that cannot account for general
equilibrium effects.
Simulating Counterfactual Distributions
The simulation of counterfactual distributions needed for the decomposition of
distributional changes requires estimation of some version of the Roy system
(composed of a selection equation and an outcome equation) for the initial and
end periods. Counterfactual distributions can then be simulated by switching
parameters and variables between these two estimated models one element at a
time, holding all the other factors constant.
In general, parameters of sample selection models can be estimated with two-
stage methods or the maximum-likelihood approach. We summarize here the
basic ideas underlying two-step procedures that are also known as control
function methods or generalized residual methods (Todd 2008). These methods
are commonly used in the context of an explicit model of the outcome process
involving a selection mechanism as well as an outcome equation. The selection
mechanism is usually modeled within a random utility framework, and identify-
ing assumptions are based on functional form restrictions or exclusion restric-
tions (analogous to the instrumental variable [IV] approach).
In particular, the control function approach seeks to model conditional expec-
tations of potential outcomes (given observable characteristics and occupational
or socioeconomic status) in a way that relates unobservable determinants of
outcomes to the observables, including the choice of a socioeconomic group. This
is consistent with the view that the underlying endogeneity problem is the result
of omitted variables. The control functions represent the omitted variables. The
key assumption in this method is that the observable determinants of both selec-
tion and outcomes are independent of the unobservable determinants (Heckman
and Navarro-Lozano 2004).
Combining Parametric and Nonparametric Techniques
Once the model has been estimated, counterfactual decompositions are per-
formed. Building on the statistical approaches previously discussed in “The
Composition and Structural Effects” section, Bourguignon, Ferreira, and Leite
(2008) propose the combination of parametric and nonparametric techniques in
constructing the desired counterfactuals.
To see clearly what is involved, recall that the density function characterizing
the joint distribution of the outcomes and covariates can be written as a product
of the two underlying density functions—one characterizing the conditional dis-
tribution of outcomes given the covariates, and the other the joint distribution of
covariates. Earlier, we expressed this relation as Jt ( y, x) = gt( y|x)ht(x), t = 0, 1. As
noted earlier, this factorization suggests that counterfactual distributions can be
obtained by combining the conditional outcome distribution from one period
(such as the initial period) with the joint distribution of covariates from the other
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Counterfactual Decomposition of Changes in Poverty Outcomes 97
period (such as the end period). An example of this type of combination would
be the following:
h1
Jg 0
( y, x ) = g 0 ( y|x ) h1 ( x ). (4.38)
A key distinction between the methods discussed in the section on composi-
tion and structural effects and those reviewed in this section is that the methods
in section 3 are based on statistical models of the conditional outcome distribu-
tion, while the methods discussed here rely on economic modeling of this con-
ditional distribution. Thus, equations (4.36) and (4.37) characterizing the basic
Roy model must be seen as modeling the conditional outcome distribution,
gt( y|x).
The method of Bourguignon, Ferreira, and Leite (2008) as well as the model
adopted in chapter 6 use the parametric approach to generate counterfactuals for
the conditional outcome distribution and nonparametric sample-reweighting tech-
niques to construct counterfactuals for the joint distribution of exogenous covari-
ates. These authors argue that the parametric approach for the conditional
distribution has the advantage of providing a clear economic interpretation of the
parameter estimates along with great flexibility in exchanging parameters from
one period to another (that is, from one state of the world to another).
Application in Context of the Roy Framework
To see how this works in the context of the Roy framework, use the estimated
model to write the approximation to the conditional outcome distribution as
follows:
( )
yt = s x t , zt ; γ t , βt , ε t , ηt . (4.39)
Thus a change in the conditional outcome distribution as a result of a ceteris
paribus change in the parameters of the multinomial logit model of selection can
be computed easily as follows:
( ) ( )
∆y = s x 0 , z0 ; γ 1 , β0 , ε 0 ,η0 − s x 0 , z0 ; γ 0 , β0 , ε 0 , η0 . (4.40)
When a counterfactual requires a normalization of exogenous covariates for
both periods, we can simply apply the DiNardo, Fortin, and Lemieux (1996)
approach described earlier. The handling of the residuals in this process requires
some care. In the case of the residuals associated with the outcome equations, for
instance, one can resort to the rank-preserving transformation described earlier.
Concluding Summary and Remarks
The design and implementation of effective strategies for poverty reduction
require relevant and reliable analytical input. The bedrock of this input is cer-
tainly a rich and reliable dataset (both qualitative and quantitative) to be used in
poverty measurement and analysis. In this context, there is a need for a sound
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
98 Counterfactual Decomposition of Changes in Poverty Outcomes
understanding of the fundamental factors that account for observed variations in
poverty either across space or over time.
This chapter has reviewed some of the basic decomposition methods that are
commonly used to identify sources of variation in poverty outcomes at both the
macro and micro levels. It has focused on micro-decomposition approaches
because aggregate methods fail to account for the heterogeneity of the factors
that drive the observed changes in aggregate poverty.
Decomposition as Social Impact Evaluation
The decomposition of changes in poverty is an exercise in social impact evalua-
tion that assesses changes in individual and social outcomes attributable to socio-
economic shocks or policy implementation. Outcome models that underlie
micro-decomposition methods are consistent with the view that an individual’s
living standard is a payoff from one’s participation in the life of society.
Accordingly, that payoff is a function of endowments, behavior, and the circum-
stances that determine the returns to these endowments from any social transac-
tion. These elements drive the distributional changes that define the potential
scope of micro-decomposition methods. In general, the scope of a decomposition
method is the set of explanatory factors the method tries to uncover by decom-
position. The specification of an outcome model thus determines the potential
scope of the corresponding decomposition method.
Statistical and Structural Methods
Micro-decomposition methods fall into two basic categories: statistical and struc-
tural. All seek to model the joint distribution of the outcome variable and its
determining factors. This joint distribution can be factorized into a product of
(a) the conditional outcome distribution and (b) the marginal distribution of
exogenous (independent) variables.
Statistical methods rely uniquely on statistical principles to model the condi-
tional outcome distribution, while structural methods rely on both economics and
statistics to model this object. In particular, the structural methods considered
here use utility maximization in a partial equilibrium setting to characterize
individual behavior and social interaction. Statistical methods therefore are
purely descriptive, while structural ones are considered predictive.
Identification through Counterfactual Comparison
Identification concerns the assumptions needed to recover, in a meaningful
way, the factors of interest at the population level. These assumptions involve
both the functional form of the outcome model and the joint distribution of fac-
tors that determine the outcome. Although macro- and micro-decomposition
methods differ in their scope, they share the same fundamental identification
strategy based on the notion of ceteris paribus variation. The implementation of
this idea entails the comparison of an observed outcome distribution with a
counterfactual obtained by changing one factor at a time while holding all the
other factors constant.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Counterfactual Decomposition of Changes in Poverty Outcomes 99
A key counterfactual used in the identification of endowment and structural
effects is the outcome distribution that would have prevailed in one state of the
world had individual characteristics been rewarded according to the system
applicable in the alternative state. The construction of this counterfactual relies
critically on ignorability and the absence of general equilibrium effects. When the
outcome model is separable in observables and unobservables, one can assume
rank preservation to further split the structural effect into a component due to
observables and another due to unobservables.
An Analogy: Decomposition, Estimation, and Treatment Effect Analysis
Estimation involves the computation of the relevant parameters on the basis of
sample data. There is a powerful analogy between the decomposition methods
reviewed here and treatment effect analysis. Both fields of inquiry rely on the
same fundamental identification strategy, and the structural effect is known to be
equivalent to the treatment effect on the treated.
This analogy has led to the development of flexible estimation methods for
endowment and structural effects. Nonparametric estimation methods, such as
IPW, allow an analyst to decompose distributional changes without having to
assume a functional form for the outcome model. The downside, however, is the
inability to further decompose the structural effect and to account for behavior.
Parametric methods are more suitable for these two tasks.
Toward Fuller Causal Interpretations of Decomposition Results
Although the analogy between decomposition methods and treatment effect
analysis has helped with the development of estimation methods, it does not
necessarily confer a causal interpretation to decomposition results. As noted by
counterfactual
Ferreira (2010), such an interpretation requires the construction of
outcome distributions that are fully consistent with a general equilibrium of the
economy. One way of achieving this consistency is to base decomposition on a
full structural model of behavior and social interaction. Such a model can be built
by embedding a behavioral model, such as the Roy (1951) model of choice and
consequences, in a general equilibrium framework.
Finally, we note that the generalization of the Oaxaca-Blinder decomposition
to the analysis of variation in other distributional statistics such as quantiles,
poverty, and inequality measures enables an analyst to study the distributional
and poverty impacts of an assigned intervention.
Notes
1. Roughly speaking, a functional is a function of a function. In this particular context,
it is a rule that maps every outcome distribution in its domain into a real number
(Wilcox 2005).
2. In particular, Bourguignon and Ferreira (2005) argue that the configuration of income
distribution at one point in time is determined by the following factors: (a) the distri-
bution of factor endowments and sociodemographic characteristics among the
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
100 Counterfactual Decomposition of Changes in Poverty Outcomes
population; (b) the returns to these assets and characteristics; and (c) the behavior of
socioeconomic agents with respect to resource allocation subject to prevailing institu-
tional arrangements.
3. An alternative expression is based on this counterfactual: E(x0)b1. The corresponding
decomposition is ∆O µ
β1 + E ( x 0 )
E ( x1 ) − E ( x 0 )
= β1 − β 0
.
4. To further appreciate the importance of the identifying assumptions, note that the
process of reweighing adjusts the distribution of the covariates x in period t = 0 so that
it becomes similar to that in period t = 1. For this adjustment to help us identify the
terms of the decomposition, it must be a ceteris paribus adjustment. Because y0 = j0(x, e),
the ceteris paribus condition would be violated if changing the distribution of x also
changed either the function ϕ0(∙) or the conditional distribution of ε given x. This
would confound the impact of the adjustment, and the decomposition would be
meaningless. Changes in the structural function are ruled out by the simple treatment
assumption (no general equilibrium effects), while those in the conditional distribu-
tion of ε are ruled out by the ignorability assumption. Under these circumstances,
we expect the conditional distribution of y0 given x to be invariant with respect to
adjustments in the distribution of the observable factors x. See Fortin, Lemieux, and
Firpo (2011) for a more formal presentation of this argument.
5. This can be expressed with an indicator function as follows: Dy(u) = I( y < u). Recall
that an indicator function is equal to one when its argument is true and equal to zero
∞ ∞ v
otherwise. In particular,
∫∆
0
y ( v ) f ( y ) dy = ∫I ( y ≤ v ) f ( y ) dy = ∫ f ( y ) dy = F ( v ) .
0 0
6. Essama-Nssah and Lambert (2012) show how to derive the influence function of a
functional from the associated directional derivative. They present a collection of
influence functions for social evaluation functions commonly used in assessing the
social impact of public policy.
7. Wilcox (2005) explains that continuity alone confers only qualitative robustness to
the statistic under consideration. A continuous function is relatively unaffected by
small shifts in its argument. Similarly, differentiability is related to infinitesimal
robustness in the sense that if a function is differentiable and its derivative is bounded,
then small variations in the argument will not result in large changes in the function.
Thus a search for robust statistics can focus on functionals with bounded derivatives.
8. This is analogous to the approximation of a differentiable function at a point by a
Taylor’s polynomial.
9. In particular, b1 and b0 may differ just because their estimation is based on different
distributions of the covariates x, even if the outcome structure remains unchanged
(Firpo, Fortin, and Lemieux 2009).
10. For details, see Fortin, Lemieux, and Firpo (2011).
11. Essama-Nssah, Saumik, and Bassolé (2013) use both linear and nonlinear RIF regres-
sion models in their study of growth incidence in Cameroon. They find that linear
models lead to results that are qualitatively similar to nonlinear specifications. This is
a significant methodological finding that should comfort analysts who might worry
about the quality of the linear approximation underlying the simple RIF regression
approach.
12. To see what is involved, write the conditional mean outcome as E( yt|xt;bt), t = 0,1. The
counterfactual mean outcome when endowments in period 1 are valued under the
(reward) regime of period 0 is equal to the following: E y1 c
(
|x1; β 0 . The observed )
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Counterfactual Decomposition of Changes in Poverty Outcomes 101
difference in mean outcomes can therefore be decomposed as follows:
∆Oµ
= E y1
c
( )
|x1; β 0 − E ( y0 |x 0 ; β 0 ) + E ( y1|x1; β1 ) − E y1
c
( )
|x1; β 0 . This expression is
analogous to equation (4.3). The first term on the right-hand side is the composition
effect, and the second term is the structural effect. This is an aggregate
decomposition.
13. To see where this expression comes from, let qt be the τth quantile of F. Also, let qt (b)
stand for the τth quantile of the mixed distribution G so that G = bH(qt (b)) + (1 − b)
F (qt (b)) = t and qt (0) = qt . The first-order derivative of G with respect to b, evalu-
′ ( 0 ) = 0. Hence
ated at b = 0, yields the following expression: H ( qτ ) − F ( qτ ) + f ( qτ ) qτ
the directional derivative of this quantile in the direction of H is equal to
F ( qτ ) − H ( qτ ) τ − H ( qτ )
qτ′ (0) = = . Setting H(qt ) = Dy(qt ) = I( y ≤ qt ) implies that
f ( qτ ) f ( qτ )
τ − I ( y ≤ qτ )
the influence function of the τth quantile of F is equal to IF ( y; qτ , F ) = .
f ( qτ )
For more details, see Essama-Nssah and Lambert (2012).
14. See Essama-Nssah, Saumik, and Bassolé (2013) for an application.
15. This is the effect of a counterfactual change in the marginal distribution of a single
covariate on the unconditional distribution of an outcome variable, ceteris paribus.
Rothe (2010) interprets the ceteris paribus condition in terms of rank invariance. In
other words, the counterfactual change in the marginal distribution of the relevant
covariate is constructed in such a way that the joint distribution of ranks is
unaffected.
16. The components of a detailed decomposition are easily computed by replacing (a) the
expected values with the corresponding sample means, and (b) the coefficients associ-
ated with the covariates with their OLS estimates. An estimate of the endowment
m
∑(x
q
effect is ∆ X = ( x1 − x 0 ) β 0 = 1k − x 0 k ) β 0 k. Similarly, for the structural effect, we
k =1
m
q
(
have the following expression: ∆ S = x1 β 1 − β 0 = β 11 − β 01 + ) ( ) ∑x (β 1k 1k )
− β 0k .
k=2
17. This account draws on Essama-Nssah and Bassolé (2010).
18. To clarify our notation, we consider the simplest case where x represents a single
characteristic. No loss of generality is involved. The marginal distribution of y is equal
mx
to ft ( y ) = ∫ J ( y, x ) dx, where mx stands for the maximum value of x. Equivalently,
0
t
mx
ft ( y ) = f gt
ht
( y) = ∫ g ( y|x ) h ( x ) dx. The counterfactual used in equation (4.22) is
0
t t
mx
1
therefore defined as follows: f gh0 =
∫g 0
0 ( y|x ) h1 ( x ) dx. This expression can be derived
mx
from the marginal outcome distribution in the initial period, f0 ( y ) = ∫g
0
0 ( y|x ) h0 ( x ) dx ,
by replacing h0(x) with h1(x). For this operation to lead to a meaningful counterfac-
tual, two invariance conditions must be met. The conditional distributions gt( y|x)
must be invariant with respect to changes in the marginal distribution of observables,
ht(x). This would be the case if there are no general equilibrium effects. The distribu-
tion of unobservables must be at least conditionally independent of that of observ-
ables. Ignorability guarantees this.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
102 Counterfactual Decomposition of Changes in Poverty Outcomes
19. A histogram is “a representation of a frequency distribution by means of rectangles
whose widths represent class intervals and whose areas are proportional to the cor-
responding frequencies” (online version of the Merriam-Webster Unabridged, http://
unabridged.merriam-webster.com/).
20. Deaton (1997) further explains that the size of the bandwidth is inversely related to
the sample size. The larger the sample size, the smaller the bandwidth. To obtain a
consistent estimate of the density at each point, the bandwidth must become smaller
at a slower rate than the rate at which the sample size is increasing. However, with
only a few points, we need large bands to be able to get any points in each. By widen-
ing the bands, we run the risk of biasing the estimate by bringing into the count data
that belong to other parts of the distribution. Hence, the increase in the sample size
does two things: It allows the analyst to reduce the bandwidth and hence the bias in
estimation (due to increased mass at the point of interest). It also ensures that the
variance will shrink as the number of points within each band increases.
21. Deaton (1997) argues that the choice of the bandwidth or the smoothing parameter
is more important than that of the kernel function. Essentially, estimating densities by
kernel methods is an exercise in smoothing the sample observations into an estimated
density. The bandwidth controls the amount of smoothing achieved. Oversmoothed
estimates are biased, while undersmoothed ones are too variable.
22.
f y0 |t =1 − f y0 |t = 0
The equivalent expression for the decomposition is ∆f = f y1 |t =1 − f y0 |t =1
+
f y0 |t =1 − f y0 |t = 0
∆f = f y1 |t =1 − f y0 |t =1
+ .
23. In the context of treatment effect analysis, the assumption of rank preservation, also
known as rank invariance, is used to identify quantile treatment effects (QTE). The
assumption implies that given two mutually exclusive states of the world, the out-
come at the τth quantile of the outcome distribution in one state has its counterpart
at the same quantile of the outcome distribution in the alternative state. Bitler,
Gelbach, and Hoynes (2006) explain that when this assumption fails, the QTE
approach identifies and estimates the difference between the quantiles and not the
quantiles of the difference in outcome distributions. Rank preservation is akin to
anonymity or symmetry used to base growth incidence analysis on cross-section data
instead of on panel data. Anonymity implies that when comparing two outcome dis-
tributions, the identity of the individual experiencing a particular outcome is irrele-
vant (Carneiro, Hansen, and Heckman 2002). Thus, a permutation of outcomes
between any two individuals in any of the two distributions being compared has no
effect on the comparison. One might as well then compare such distributions across
quantiles.
24. Recall that ignorability means that the conditional distribution of e is the same across
groups (or periods). Thus individuals with the same set of observable characteristics
find themselves at the same rank in both (conditional) distributions. It is well known
that a monotonic transformation preserves order. In fact, Rapoport (1999) defines a
monotone transformation as “a formula that changes the numbers of one set to the
numbers of another set while preserving their relative positions on the axis of real
numbers.” Since ut is obtained from e through a monotonic transformation zt(·), rank
preservation must therefore follow.
y b
( )
25. Note that the structural effect can also be expressed as ∆ S = yi1 − yi 0 . In the notation
associated with equation (4.2), linearity and rank preservation imply that yib0 corre-
sponds to the counterfactual outcome obtained by replacing the outcome structure
ϕ1(∙) with ϕ0(∙). In other words, yib0 is the same as y0|t = 1. This suggests that the
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Counterfactual Decomposition of Changes in Poverty Outcomes 103
Juhn-Murphy-Pierce (1993) decomposition can be performed in two steps as follows:
y
Start with the overall difference ∆O = ( yi1 − yi 0 ) and then add to and subtract from
this difference the counterfactual outcome yib0. This yields a twofold decomposition of
the overall difference into the composition and structural effects. Finally add to and
c
subtract from the structural effect the counterfactual outcome yi 0. This step leads to
the final threefold decomposition. Ignorability guarantees that the composition effect
is due solely to changes in the distribution of observables.
26. Indeed, Fortin, Lemieux, and Firpo (2011) provide a systematic interpretation of
decomposition methods within the logic of program impact evaluation.
27.
According to Fortin, Lemieux, and Firpo (2011), these assumptions include the
following: (a) The groups are mutually exclusive. (b) The outcome structure is an
additively separable function of characteristics. (c) The conditional mean for unob-
servables given observed characteristics is equal to zero. (d) There is common support
for the distributions of characteristics across groups (to rule out cases where argu-
ments of the outcome function may differ across groups. (e) There is simple counter-
factual treatment, meaning that the outcome structure of one group is assumed to be
equilibrium
a counterfactual for the other group. This last assumption rules out general
effects so that observed outcomes for one group or time period can be reasonably used
to construct counterfactuals for the other group or time period. The Oaxaca-Blinder
method therefore follows a partial equilibrium approach.
28. Heckman and Smith (1995) explain that the mean outcome of the control group
provides an acceptable estimate of the counterfactual mean if (a) randomization does
not alter the pool of participants or their behavior and (b) no close substitutes for the
experimental program are readily available. These authors further note that random-
ization does not eliminate selection bias, but rather balances it between the two
samples (participants and nonparticipants) so that it cancels out when computing
mean impact. There would be randomization bias if those who participate in an
experiment differ from those who would have participated in the absence of random-
ization. Furthermore, substitution bias would occur if members of the control group
can easily obtain elsewhere close substitutes for the treatment.
29. Recall that the identification of the two components of the aggregate Oaxaca-Blinder
decomposition relies on the zero-conditional-mean assumption for the unobservable
factors stated as E(¨|x) = 0. This condition is what allows the analyst to claim that on
average, variation in x is unrelated to variation in the unobservables, a manifestation
of ceteris paribus variation.
30. The expression of this probability can be made more precise by considering an indica-
tor function for the decision rule. The indicator is equal to one when option k is
chosen and zero otherwise. The probability that the agent chooses option k is then
equal to the expected value of this indicator function over all possible values of the
unobserved factors. In other words, Phk = I
∫ ( ) ( )
ε hj − ε hk < v hk − v hj ∀j ≠ k f ( ε h ) dε h .
This is in fact a multidimensional integral over the joint density of the random vector,
the elements of which represent unobserved factors associated with each alternative.
This probability can be interpreted as the proportion of people within the population
who face the same observable utility as h for each alternative and choose k (Train
2009).
31. The generalized extreme value model (GEV) allows correlation among unobserved
factors. The standard multinomial logit assumes that the random factors are iid with
a double exponential distribution. The probit model assumes that the random factors
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
104 Counterfactual Decomposition of Changes in Poverty Outcomes
are jointly distributed normal variables. Train (2009) points out that the identification
of discrete choice models relies heavily on the fact that only differences in utility
matter and the scale of utility is irrelevant. Hence, only parameters that capture dif-
ferences across alternatives are identifiable and therefore estimable. This also implies
that characteristics of the decision maker that do not vary across alternatives will have
no effect unless they are specified in a way that induces differences in utility over
alternatives. Glick and Sahn (2006) handle this problem by indexing the coefficients
of sociodemographic variables in the representative utility function.
32. For agricultural households, earnings are computed on the basis of a reduced farm
profit function (based on the Cobb-Douglas production function) that includes self-
consumption and accounts for hired labor. For family members participating in farm
work, the reservation wage (a measure of the value of family work) is assumed to
depend on their contribution to farm profits.
Bibliography
Angrist, Joshua D., and Jörn-Steffen Pischke. 2009. Mostly Harmless Econometrics: An
Empiricist’s Companion. Princeton, NJ: Princeton University Press.
Atkinson, Anthony B. 1970. “On the Measurement of Inequality.” Journal of Economic
Theory 2 (3): 244–63.
Autor, David H. 2009. “Lecture Note: Self-Selection — The Roy Model.” Department of
Economics, Massachusetts Institute of Technology, Cambridge, MA. http:// econ
-www.mit.edu/files/4092.
Bitler, Marianne P., Jonah B. Gelbach, and Hilary W. Hoynes. 2006. “What Mean Impacts
Miss: Distributional Effects of Welfare Reform Experiments.” American Economic
Review 96 (4): 988–1012.
Blinder, Alan S. 1973. “Wage Discrimination: Reduced Form and Structural Estimates.”
Journal of Human Resources 8 (4): 436–55.
Bourguignon, François, and Francisco H. G. Ferreira. 2005. “Decomposing Changes in the
Distribution of Household Incomes: Methodological Aspects.” In The Microeconomics
of Income Distribution Dynamics in East Asia and Latin America, edited by François
Bourguignon, Francisco H. G. Ferreira, and Nora Lustig, 17–46. Washington, DC:
World Bank.
Bourguignon, François, Francisco H. G. Ferreira, and Phillippe G. Leite. 2008. “Beyond
Oaxaca-Blinder: Accounting for Differences in Household Income Distributions.”
Journal of Economic Inequality 6 (2): 117–48.
Carneiro, Pedro, Karsten T. Hansen, and James J. Heckman. 2002. “Removing the Veil of
Ignorance in Assessing the Distributional Impacts of Social Policies.” Discussion Paper
453, Institute for the Study of Labor (IZA), Bonn, Germany.
Cogneau, Denis, and Anne-Sophie Robilliard. 2008. “Simulating Targeted Policies with
Macro Impacts: Poverty Alleviation Policies in Madagascar.” In The Impact of
Macroeconomic Policies on Poverty and Income Distribution: Macro-Micro Evaluation
Techniques and Tools, edited by François Bourguignon, Maurizio Bussolo, and Luiz A.
Pereira da Silva, 213–45. Washington, DC: World Bank.
Coulombe, Harold, and Andrew McKay. 1996. “Modeling Determinants of Poverty in
Mauritania.” World Development 24 (6): 1015–31.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Counterfactual Decomposition of Changes in Poverty Outcomes 105
Cowell, Frank A., and Maria-Pia Victoria-Feser. 1996. “Poverty Measurement with
Contaminated Data: A Robust Approach.” European Economic Review 40 (9):
1761–71.
Deaton, Angus. 1997. The Analysis of Household Surveys: A Microeconometric Approach to
Development Policy. Baltimore, MD: Johns Hopkins University Press.
DiNardo, John, Nicole Fortin, and Thomas Lemieux. 1996. “Labor Market Institutions and
the Distribution of Wages, 1973–1992: A Semiparametric Approach.” Econometrica 64
(5): 1001–44.
Essama-Nssah, B. 2012. “Identification of Sources of Variation in Poverty Outcomes.”
Policy Research Working Paper 5954, World Bank, Washington, DC.
Essama-Nssah, B., and Léandre Bassolé. 2010. “A Counterfactual Analysis of the Poverty
Impact of Economic Growth in Cameroon.” Policy Research Working Paper 5249,
World Bank, Washington, DC.
Essama-Nssah, B., and Peter J. Lambert. 2012. “Influence Functions for Policy Impact
Analysis.” In Inequality, Mobility and Segregation: Essays in Honor of Jacques Silber,
Vol. 20, edited by John A. Bishop and Rafael Salas, 135–59. Research on Economic
Inequality Series. Bradford, UK: Emerald Group Publishing.
Essama-Nssah, B., Paul Saumik, and Léandre Bassolé. 2013. “Accounting for Heterogeneity
in Growth Incidence in Cameroon Using Recentered Influence Function Regression.”
Journal of African Economies 22 (5): 757–95. doi: 10.1093/jae/ejt009.
Fairlie, Robert W. 2005. “An Extension of the Blinder-Oaxaca Decomposition Technique
to Logit and Probit Models.” Journal of Economic and Social Measurement 30 (4):
305–16.
Ferreira, Francisco H. G. 2010. “Distributions in Motion: Economic Growth, Inequality
and Poverty Dynamics.” Policy Research Paper 5424, World Bank, Washington, DC.
Firpo, Sergio, Nicole Fortin, and Thomas Lemieux. 2009. “Unconditional Quantile
Regressions.” Econometrica 77 (3): 953–73.
Fortin, Nicole, Thomas Lemieux, and Sergio Firpo. 2011. “Decomposition Methods in
Economics.” In Handbook of Labor Economics, Vol. 4A, edited by Orley Ashenfelter
and Card David, 1–102. Amsterdam, the Netherlands: North-Holland.
Foster, James, Joel Greer, and Erik Thorbecke. 1984. “A Class of Decomposable Poverty
Measures.” Econometrica 52 (3): 761–66.
Glick, Peter, and David Sahn. 2006. “The Demand for Primary Schooling in Madagascar:
Price, Quality and the Choice between Public and Private Providers.” Journal of
Development Economics 79 (1): 118–45.
Hampel, Frank R. 1974. “The Influence Curve and Its Role in Robust Estimation.” Journal
of the American Statistical Association 69 (346): 383–93.
Heckman, James J., and Bo E. Honoré. 1990. “The Empirical Content of the Roy Model.”
Econometrica 58 (5): 1121–49.
Heckman, James J., and Guilherme Sedlacek. 1985. “Heterogeneity, Aggregation, and
Market Wage Functions: An Empirical Model of Self-Selection in the Labor Market.”
Journal of Political Economy 93 (6): 1077–125.
Heckman, James J., and Jeffrey A. Smith. 1995. “Assessing the Case of Social Experiments.”
Journal of Economic Perspectives 9 (2): 85–110.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
106 Counterfactual Decomposition of Changes in Poverty Outcomes
Heckman, James J., and Salvador Navarro-Lozano. 2004. “Using Matching, Instrumental
Variables, and Control Functions to Estimate Economic Choice Models.” Review of
Economics and Statistics 86 (1): 30–57.
Inchauste, Gabriela, Sergio Olivieri, Jaime Saavedra, and Hernan Winkler. 2012. “What Is
Behind the Decline in Poverty Since 2000? Evidence from Bangladesh, Peru, and
Thailand.” Policy Research Working Paper 6199, World Bank, Washington, DC.
Jann, Ben. 2008. “The Blinder-Oaxaca Decomposition for Linear Models.” Stata Journal 8
(4): 453–79.
Juhn, Chinhui, Kevin M. Murphy, and Brooks Pierce. 1991. “Accounting for the Slowdown
in Black-White Wage Convergence.” In Workers and Their Wages: Changing Patterns in
the United States, edited by Marvin H. Kosters. Washington, DC: American Enterprise
Institute.
———. 1993. “Wage Inequality and the Rise of Returns to Skill.” Journal of Political
Economy 101 (3): 410–42.
Leite, Phillippe G., Alan Sanchez, and Caterina R. Laderchi. 2009. “The Evolution in
Urban Inequality in Ethiopia.” Paper presented at the Centre for the Study of African
Economies 2009 Annual Conference, “Economic Development in Africa,”
St. Catherine’s College, Oxford, March 22–24.
Machado, Jose A. F., and Jose Mata. 2005. “Counterfactual Decomposition of Changes in
Wage Distributions Using Quantile Regression.” Journal of Applied Econometrics 20
(4): 445–65.
Melly, Blaise. 2005. “Decomposition of Differences in Distribution Using Quantile
Regression.” Labour Economics 12 (4): 577–90.
Milgrom, Paul. 2004. Putting Auction Theory to Work. Cambridge: Cambridge University
Press.
Oaxaca, Ronald L. 1973. “Male-Female Wage Differentials in Urban Labor Markets.”
International Economic Review 14 (3): 693–709.
Rapoport, Anatol. 1999. Two-Person Game Theory. Mineola, NY: Dover Publications.
Ravallion, Martin. 2001. “Growth, Inequality and Poverty: Looking Beyond Averages.”
World Development 29 (11): 1803–15.
Rosenbaum, Paul, and Donald Rubin. 1983. “The Central Role of the Propensity Score in
Observational Studies for Causal Effects.” Biometrika 70 (1): 41–55.
Rothe, Christoph. 2010. “Decomposing Counterfactual Distributions.” Unpublished
manuscript, Toulouse School of Economics, University of Toulouse, France.
Roy, Arthur D. 1951. “Some Thoughts on the Distribution of Earnings.” Oxford Economic
Papers 3 (2): 135–46.
Silverman, Bernard W. 1986. Density Estimation for Statistics and Data Analysis. London:
Chapman and Hall.
Sinning, Mathias, Markus Hahn, and Thomas K. Bauer. 2008. “The Blinder-Oaxaca
Decomposition for Nonlinear Models.” Stata Journal 8 (4): 480–92.
Todd, Petra E. 2008. “Evaluating Social Programs with Endogenous Program Placement
and Selection of the Treated.” In Handbook of Development Economics, Vol. 4, edited by
T. Paul Schultz and John Strauss, 3847–94. Amsterdam, the Netherlands:
North-Holland.
Train, Kenneth E. 2009. Discrete Choice Methods with Simulation. 2nd ed. Cambridge:
Cambridge University Press.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Counterfactual Decomposition of Changes in Poverty Outcomes 107
Varian, Hal R. 1984. Microeconomic Analysis. 2nd ed. New York: Norton.
von Mises, Richard. 1947. “On the Asymptotic Distribution of Differentiable Statistical
Functions.” Annals of Mathematical Statistics 18 (3): 309–48.
Wilcox, Rand R. 2005. Introduction to Robust Estimation and Hypothesis Testing. 2nd ed.
Amsterdam, the Netherlands: Elsevier.
Yun, Myeong-Su. 2009. “Wage Differentials, Discrimination and Inequality: A Cautionary
Note on the Juhn, Murphy and Pierce Decomposition Method.” Scottish Journal of
Political Economy 56 (1): 114–22.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Chapter 5
Why Has Labor Income Increased?
An In-Depth Approach to
Understanding Poverty Reduction
Introduction
As chapter 3 found, labor market outcomes (in particular, labor income increases
per working adult) were the main contributors to poverty reduction in a set of
countries where poverty declined substantially during the past decade. A simple
micro-decomposition methodology showed this to be true for moderate poverty
lines and even extreme poverty lines in some cases. The next logical question is
this: Why did total labor income increase—from improved human capital char-
acteristics or from higher returns to those characteristics? To answer this ques-
tion, an underlying model with additional structure is needed.
A broad literature on micro-decomposition methodologies (which chapter 4
described in detail) aims to identify and estimate two effects associated with dis-
tributional changes: the composition (or endowment) effect and structural (or
price) effect. Either statistical or structural approaches can account for the contri-
butions of these effects. Statistical approaches use semiparametric and nonpara-
metric methods to identify the contributors to distributional changes without
having to impose a functional form on the relationship between these changes
and their determinants. However, the statistical framework cannot shed light on
the mechanism underlying that relationship. In contrast, structural methods fur-
ther identify the factors underpinning observed changes in poverty outcomes by
specifying an economic model and using statistical analysis. In particular, these
models account for behavior: how agents respond to changes in their socioeco-
nomic environment, whether due to shocks or policy reform.
The structural approach is followed below and in the next chapter (which will
apply it in depth to three countries). In particular, this chapter presents a structure
for modeling distributional changes over time. The approach rests on the typical
economic assumption that agents seek to maximize their utility subject to con-
straints. That assumption implies that the outcome variables of interest chosen to
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7 109
110 Why Has Labor Income Increased? An In-Depth Approach to Understanding Poverty Reduction
represent the consequences of behavior (such as educational or sectoral choice)
can be expressed as functions of parameters of the socioeconomic environment,
embedded in the constraints facing the agent. In particular, throughout this
chapter, we consider the Roy (1951) model of choice and consequences (which
stems from the optimization principle and applies to discrete choice problems),
using it to model individuals’ educational levels, sectors, and activity choices.1
We then adapt the Bourguignon, Ferreira, and Leite (2008) methodology to
distinguish between distributional changes on account of (a) changes in endow-
ments and the returns to those endowments; (b) changes in occupational and
sectoral choice; (c) changes in geographical, age, and gender structure of the
population; and (d) the nonlabor dimensions such as public transfers (the latter
previously explored further in chapter 3).
Therefore, the proposed model formulates an educational choice model,
a sectoral choice model, an activity choice model, and individual and household
earnings equations. Once all of these models are estimated for individuals and
households in each time period, the estimated coefficients from one year can be
replaced with the estimates from another year to simulate the impact of changes
in each element at a time. For each change, we can construct a counterfactual
income distribution and estimate a counterfactual poverty measure for compari-
son with the observed outcome while holding everything else constant. By
changing one element at a time, these decompositions allow us to account for the
observed changes in poverty.
Innovations in Cumulative Counterfactual Estimates
In addition, we present an enhanced method to estimate these counterfactuals
cumulatively, thereby accounting for the impact of concurrent changes. Although
the method presented here draws heavily from Bourguignon and Ferreira (2005)
and Bourguignon, Ferreira, and Leite (2008), it also offers some innovations:
• It models farm household income at the household level and models the earn-
ings of individuals in those farm households who have a secondary occupation,
thus recognizing not only that farm households typically make labor decisions
as a unit but also that these households can be highly diversified.
• It assumes that welfare is measured using a consumption aggregate and
accounts for the contribution of changes in the consumption-to-income ratio.
• It ensures that changes in the composition of activities, sectors, and education
of the work force are consistent with the counterfactual choices.
As covered in chapters 2 and 3, such decompositions do not identify causal
effects, but they are useful to focus attention on the elements that are quantita-
tively more important in describing changes in poverty. In particular, they can
capture the heterogeneity of impacts throughout the distribution and account
for the contributions of demographic, sectoral, occupational, and other labor and
nonlabor dimensions toward poverty reduction.
After the next section presents the underlying model, the chapter describes
the decomposition approach adopted here and concludes with a summary
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Why Has Labor Income Increased? An In-Depth Approach to Understanding Poverty Reduction 111
of the proposed model and its innovations. We implement this methodol-
ogy in three countries around the world, the results for which we report in
chapter 6.
Modeling Strategy
Following Bourguignon, Ferreira, and Leite (2008), our approach postulates a
model in which characteristics such as age, gender, and geographic location are
exogenously determined (and are taken as given), while education, employment
activity, sector, and earnings are endogenous (determined within the model), as
illustrated in figure 5.1. Therefore, the modeling strategy consists of six stages,
each of which influences the next stage in the sequence:
1. Educational choice models
2. Sectoral and activity choice models
3. Farm and nonfarm earnings
4. Total household income, based largely on earnings
5. Household consumption, determined by total household income
6. A consumption-based calculation of the poverty rate
Educational Choice
First, individuals are distributed across educational levels and, further, as a func-
tion of age, gender, area (urban versus rural), and region (defined as districts or
provinces). This is done for all working-age individuals, following the Roy (1951)
model, whereby individuals choose their educational level to maximize their
utility. The allocation of individuals across levels of education is estimated with
a multinomial logit model (McFadden 1974a, 1974b), specified as follows:
k
I hi ( )
= 1 if Zhi y k + v ik > Max 0, Zhi y m + v im , j = 1,…,K, ∀m ≠ k (5.1)
k
I hi = 0 for all k = 1,…,K if Zhi y k + v ik ≤ 0 for all k = 1,…, K,
where Zhi is a vector of characteristics specific to individual i and household h;
ψk is a matrix with vectors of coefficients for each educational level, m; and v ik
is a vector of random variables identically and independently distributed across
individuals and activities according to the law of extreme values. Within a dis-
crete utility-maximizing framework, Zhi y k + v ik is interpreted as the utility
associated with educational level k, with v ik being the unobserved utility deter-
minants of educational level k and the utility of no education being arbitrarily
set to zero.
We estimate the conditional distributions of educational levels for each survey
year based on age group, gender, region, and area. They are estimated separately
for household heads, spouses, and other working-age members. The result of this
exercise is a model that estimates the probability of individuals obtaining a
certain level of education.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
112 Why Has Labor Income Increased? An In-Depth Approach to Understanding Poverty Reduction
Figure 5.1 Model of Contributors to Poverty Reduction, by Stage Sequence
Exogenous
characteristics Endogenous characteristics
– Age 1. Educational choice
– Gender
– Area
– Region
2. Activity choice 3. Nonfarm
Nonfarm – Paid employee earnings
household – Employer/own acct. (modeled for
– Unpaid family worker each individual)
4. Total
household
income
2. Sectoral choice
Household – Agriculture
type – Industry
– Services 5. Household
consumption
3. Farm
2. Sectoral choice earnings
Farm 6. Poverty
household (of secondary (modeled as measure
occupation) net household
revenue)
Exogenous variables used to model education, Exogenous factors
activity, sector, and earnings
Type of household determined Household type
outside the model
Endogenous variables used to model earnings, Endogenous factors
income, consumption, and poverty
Note: The green boxes are numbered according to the typically sequential stages by which contributors to poverty are determined: 1. Educational
choice; 2. Sectoral and activity choice; 3. Farm and nonfarm earnings; 4. Total household income; 5. Household consumption; and
6. Consumption-based poverty rate. Farm households are those whose heads are self-employed in agriculture. “Area” refers to whether the
household is urban or rural. “Region” refers to districts or provinces. “Activity choice” refers to an individual’s occupational status, which may include
paid employment (by hourly wages or salaried); self-employment; unpaid labor (typically for family needs); or unemployment.
Sectoral and Activity Choices
The second stage of the model structure begins with the separation of house-
holds into farm and nonfarm categories, as shown in figure 5.1. Farm households
are defined as those whose household heads are self-employed in agriculture. All
other households are considered to be nonfarm households. This distinction
allows us to model the nonfarm sector at the individual level, where workers
choose the sector of work. Meanwhile, the farm sector is modeled at the house-
hold level because farm households are likely to operate as a single unit of pro-
duction, thus making labor decisions at the household level.
In developing countries, people commonly engage in nonfarm microenter-
prises and informal small-scale activities. Although these nonfarm businesses are
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Why Has Labor Income Increased? An In-Depth Approach to Understanding Poverty Reduction 113
often modeled at the household level (Haggblade, Hazell, and Reardon 2007,
2010), this modeling implies a trade-off: In particular, if earnings are modeled at
the household level, accounting for the contribution to poverty reduction of
changes in the population’s sectoral or occupational structure is impossible
because such changes can be observed only at the individual level.
Once households are divided between farm and nonfarm households, the
second stage in the model is the allocation of individuals across work-activity
types. In particular, following Roy (1951), individuals choose the sector and type
of activity they are engaged in to maximize their utility. As above, this is esti-
mated with a multinomial logit model, where j = {salaried, nonfarm self-
employed, unpaid family worker, not employed}. Sectoral choice is also estimated
this way, where j = {agriculture, industry, services}—modeled as a function of age,
gender, area, region, and educational attainment.
Recognizing that farm households often include members who work on non-
farm activities (Davis et al. 2010), we model those individuals’ decisions to
undertake secondary activities in order to capture this diversification into non-
farm activities. We assume that the residuals are independently and identically
distributed according to a logistic function (a logit model being the estimator of
the diversification choice to have a secondary occupation or not) for all house-
hold heads who are self-employed in agriculture. The probability of undertaking
a secondary activity is modeled as a function of a vector of characteristics that
includes individual and household variables such as age, gender, educational level,
region, and areas, among others. Random terms are drawn conditional on the
initial choice.
Because both the sectoral and activity choices depend on individuals’ educa-
tional levels, any simulated change in education will imply a change in the activ-
ity and sectoral composition of the work force. Although the choice of sector and
activity are likely simultaneous decisions, here they are modeled independently
for tractability of the model.
Earnings Models
The third portion of the model structure is a set of earnings equations that serve
to construct a counterfactual income distribution. For nonfarm households, we
model the heterogeneity in individual earnings in each activity j by a log-linear
Mincer model:
j j
log( yhi ) = QXhi Ω j + ε hi , (5.2)
for i = 1, …, nh, and j = {salaried, nonfarm self-employed, not employed}. Qhi is
a vector of individual characteristics, including (a) those determined outside the
model (such as gender, area, and region), which we call Zhi, and (b) those deter-
mined within the model (including education and sector), which we call Xhi.
Ω j is a matrix of coefficients, and ε hi
j
are random variables assumed to be distrib-
uted identically and independently across individuals according to the standard
ormal law.2
n
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
114 Why Has Labor Income Increased? An In-Depth Approach to Understanding Poverty Reduction
Individual earnings equations for nonfarm workers are estimated separately
for household heads, spouses, and other members who are self-employed and
salaried. The set of characteristics considered in the specification include age,
gender, and education (among others) as well as characteristics of other house-
hold members. For instance, in the case of spouses and other members, the speci-
fications include characteristics of the household head (educational level,
whether employed, and so on). Similarly, for members of farm households at the
individual level, earnings from secondary occupations are e stimated as a function
of the members’ individual characteristics (such as age, gender, educational level,
and economic sector where they perform their secondary job).
j
In both cases, changes in income ( yhi ) could be due to changes in observable
endowments (Xhi) or changes in the returns to those endowments (Ω j ). However,
they could also occur because of changes in unobservables that are captured in
the residual term. To capture these changes, we rely on the assumption that the
residual terms are drawn from a standard normal distribution.
Earnings of farm households are modeled as the net revenue at the household
level:
F
log π h = Wh ΩF + ε h
F
, (5.3)
where Wh = (Kh, Xh) includes endowments and household characteristics, Ω F is
a vector of coefficients, and ε h F
are random variables distributed as a standard
normal.
Farm households’ net revenues are estimated using ordinary least squares. The
vector of characteristics includes endowments such as land and irrigation as well as
the household head’s individual and household characteristics, including educa-
tional level, gender, civil status, and number of members involved in the farm activ-
ity, among other characteristics. To capture changes in unobservables, we rely on the
assumption that the residual terms are drawn from a standard normal distribution.
Total Income
Finally (and as in chapter 2), the conditional distribution of nonlabor income is
estimated nonparametrically—both as a total and by its different components
such as remittances, public transfers, and other private transfers. For this purpose,
we create cells of household heads of the same educational level, gender, and area
(urban or rural). Inside of each cell, we create quantiles of nonlabor income, to
which we then ascribe the mean value of each nonlabor income component in
each quantile-cell in period s to its counterpart in period t.
Given the labor and nonlabor incomes described above, total household
income can be written as
yh + y h + π h + yh
Yh = w se F NL
, (5.4)
w
where Y is household income per capita; yh and yh SE
are total incomes from
F
salaried labor and self-employed nonfarm labor, respectively; π h
is the farm
NL
y
household net revenue function; and h is household nonlabor income.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Why Has Labor Income Increased? An In-Depth Approach to Understanding Poverty Reduction 115
Given the choice models described above, this becomes
n n
yh =
∑
i =1
w w
I hi ( ) ∑I
yhi Qhi , Ωw +
i =1
se
hi
se
yhi( ) (
Qhi , Ω se + π h
F
)
Wh , ΩF + yh
NL
, (5.5)
w se
where I hi and I hi are indicator variables that are equal to one if individual i in
w
household h is a salaried or self-employed worker. yhi and yhi se
are the correspond-
ing earnings of individual i in household h that depend on individual and house-
hold endowments (Qhi) and the returns to those endowments (Ω), which vary
w SE
across types of activity. Individuals with earnings yh and yh are in the “nonfarm”
sector, although this category comprises all salaried workers—including those in
F
agriculture—plus the nonfarm self-employed. For those in the “farm” sector, π h
is household net revenue in farm activities, which depends on household endow-
ments (Whi) and the returns to those endowments (Ω F ).
Household Consumption
Since household welfare is typically measured by consumption expenditures, we
can write
ϑh w
Ch = yh + yh
se F
+ πh NL
+ yh
, (5.6)
n
where Ch is household consumption per capita, n is the number of household
members, and ϑh is the consumption-to-income ratio. Note that once we have
defined a way to construct household consumption per capita at the household
level, we can construct a distribution of consumption across households and then
measure the poverty headcount rate or any other distributional measure.
Poverty Headcount Rate
The poverty headcount index measures the proportion of the population whose
consumption falls below the poverty line. Formally,
N
P0 =
1
N ∑ I(C
h =1
h < z ), (5.7)
where N is the total population and I(·) is an indicator function that takes on a
value of 1 if the bracketed expression is true and 0 otherwise. If consumption
expenditure (Ch) is less than the poverty line (z), then I(·) equals 1 and the
household would be counted as poor. Whenever any of the elements in the
models described above changes, a new consumption distribution can be gener-
ated, and therefore a new poverty rate can be calculated. Similarly, any other
distributional statistic can be calculated, including measures of inequality, such as
the Gini or the Theil index, as well as other poverty measures, such as the pov-
erty gap and the severity of poverty.
Therefore, equations (5.1)–(5.6) fully characterize the underlying reduced-
form models that will allow for the micro-decompositions of poverty. Next, there
are two important steps. The first is the estimation strategy, and the second is the
construction of counterfactual distributions. We turn to each of these steps in turn.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
116 Why Has Labor Income Increased? An In-Depth Approach to Understanding Poverty Reduction
Decomposition Approach
Estimation Strategy
After each of these reduced-form models has been estimated for years s and t, we
decompose the distributional changes by substituting each of the parameters
estimated for one year with the parameters of the other year and then by formu-
lating the appropriate counterfactual distribution of income and consumption.
Specifically, from equation (5.5) above, we estimate the components of house-
hold income for times s and t:
nh nh
∑ I Qhi Ω + ε w
∑I Q Ω se + ε se
w
log( yh ) = t w ˆ hi + se ˆ hi
hi
hi
hi
i =1 i =1
(5.8)
F NL
ˆ ) + log( y h ),
+ (Wh Ω + ε F
h
which, for simplicity, we express as follows:
log( yh )t = f NF Ω NF , Zhi , ε hi ,O Ψ , Zhi
t ˆt t
t t
t
, Hhi t
ˆ hi
, X 2thi , v ,
(5.9)
F ΩF ,Wht , H h
t ˆt
ˆ NL |t ,
t
,ε h , y
h
t
where NF(·) = nonfarm earning equations, and Ω NF refers to the set of esti-
mated parameters;
t
Zhi =exogenous variables such as age, gender, region, and area that
are used for the earnings and choice models estimated at the
individual level;
t
O(·) = activity choice equations, and Ψ refers to the set of estimated
( )
t t
parameters;
H Xhi t
,q hi ,φ hi = the underlying models for the sectoral and educational struc-
t
ture, where Xhi is a vector of endogenous variables including
t
sector ( X1hi ) and education ( X 2thi ), which are estimated at the
individual level and then used in the activity choice model,
t t
with q1hi and q 2hi being the respective set of estimated
parameters;
t
F(·) = net farm revenue equations, and ΩF are the set of estimated
parameters;
Wht = exogenous variables such as age, gender, region, and area for the
farm net revenue model estimated at the household level;
t t t t
error terms for earning equations for nonfarm and farm sectors
ε hi , ε h , v hi ,φ hi =
and endogenous variables such as educational structure, activity
choice, and economic sector; and
NL t
yh | = nonlabor income.
From here, we can perform marginal decompositions that consist of changing
one component at a time, keeping everything else constant. After describing
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Why Has Labor Income Increased? An In-Depth Approach to Understanding Poverty Reduction 117
how this is done for each element, we briefly discuss the cumulative approach,
which changes each additional component and adds to the total effect until all
components are accounted for. It is important to note that all decompositions
are performed considering s as the initial year and then considering t as the
initial year. The average of these decompositions is the final result reported in
the analysis.3
Construction of Counterfactual Distributions
Changes in Poverty as a Result of Changes in Demographics
The first decomposition consists of altering the joint distribution of exogenous
household characteristics such as age, gender, region, and area of each individual
in the household. Because these variables do not depend on any other variables
in the model, one can think of this simulation as the one assuming the greatest
degree of exogeneity.
The simulation is performed simply by recalibrating the population of one
year by the weights corresponding to the joint distribution of these attributes in
the target year. In other words, the demographic characteristics of year t are
weighted such that their structure replicates the demographic characteristics of
year s. For example, if the share of women in year s is higher than in year t, then
the weights in year t are modified so that in the simulation they replicate the
structure observed in year s.
Because demographic variables are determinants of the activity, sector, and
educational choice models, the reweighed structure of these variables have direct
and indirect effects on household income. The indirect effects are calculated by
substituting the reweighed demographic variables into the estimated multino-
mial logit equations to forecast a counterfactual activity, sector, and educational
composition of the work force. These new sectors, educational levels, and activi-
ties are then fed into the estimated earnings equations, along with the direct
effect of changes in the demographic variables. Formally, the vector of variables
Zhi in equation (5.9) for year t is substituted for that of year s as follows:4
log ( yh )X . Z ,W
s →t
(5.10)
( (
= f NF Ω
t
NF
s
, Zhi t ˆt
, Hhi ) ( t
, ε hi ,O y , Zhi
s
) (
ˆ hi
, X 2thi , v t
t
, F Ω ,Whs , Hh
F
t ˆt
)
, ε h , yh
NL t
| )
The resulting income distribution is then transformed to a new distribution of
consumption using equation (5.6) above, where the consumption-to-income
ratio is kept constant. Given this new counterfactual distribution of consump-
tion, any new distributional statistic can be computed.
In this case, because we are interested in the contributions to poverty reduc-
tion, we apply the poverty line for period t and calculate the counterfactual
headcount poverty rate, P ( yh )s →t
{
X . Z ,W
h
}
. The contribution of demographic
changes to the observed change in poverty will be the difference between the
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
118 Why Has Labor Income Increased? An In-Depth Approach to Understanding Poverty Reduction
poverty rate for year t and the counterfactual generated by equation (5.10),
which can be expressed as ∆PXs →t t
, Z ,W = P yh − P ( yh )s { }
→t
X . Z ,W
h
. { }
h
Changes in Poverty as a Result of Changes in Structure of Activity
Next, to account for changes in the structure of activity, the coefficients of the
activity multinomial logit models for year t are replaced with those of year s. As
a result, individuals are reallocated into different activities to conform to the
structure observed in year s. To allow for individuals to change activities in the
simulation, we must estimate the residual terms of the multinomial logit model
(v is ) in equation (5.1), which are unobserved. As annex 5A describes in detail,
these residuals must be drawn from an extreme value distribution in a way that
is consistent with observed choices. In contrast to previous papers (Bourguignon,
Ferreira, and Lustig 2005), we use the analytical solution to this problem derived
by Train and Wilson (2008).
With the new simulated structure of activity in year t, labor income is
projected using the estimated earnings equations for year t and the residuals
drawn from a standard normal distribution. Specifically, the set of parameters
t s
estimated at time t, y , are substituted by those estimated at time s, y , maintain-
ing everything else constant. This new structure of activities is then used to
obtain a counterfactual distribution of income as follows:
log ( yh )Ψ
s →t
( ( t
t
= f NF Ω NF , Zhi t
, Hhi ˆ hi
,ε t
) (
t
,O Ψ , Zhi
S
ˆ hi
, X 2thi , v t
),
(5.11)
( t
F Ω ,W , H , ε
F ˆ h
t t
h
NL
h ) , y | ).
NL t
h
The resulting counterfactual total household income distribution can then be
transformed to a new distribution of consumption using equation (5.6) above,
where the consumption-to-income ratio is kept constant. This counterfactual
distribution of consumption can be compared with the actual distribution in
(5.9). We calculate the counterfactual poverty rate and take the difference from
the poverty rate found in period t to obtain the contribution to poverty reduc-
tion: ∆Pys →t
{ }t
= P yh
h
{
s →t
− P ( yh )y ) . }h
Changes in Poverty as a Result of Changes in Education and Sectoral
Composition
Similarly, to account for changes in the educational structure or in the sectoral
composition of employment, we use the coefficients from the estimated multi-
nomial logit models. For instance, to account for changes in poverty due to
t
changes in the sectoral composition of the work force, we substitute the q1hi
s
parameters in the sectoral equation for time t, with q1hi . To allow for individuals
to change sectors in the simulation, the residual terms in the sectoral
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Why Has Labor Income Increased? An In-Depth Approach to Understanding Poverty Reduction 119
multinomial logit model are taken from an extreme value distribution, as
described in annex 5A. This will produce a simulated sectoral structure, which
will in turn affect nonfarm incomes as well as farm incomes given that we allow
for individual choice in secondary occupation. We obtain the counterfactual
income distribution as follows:
s →t
( (
log ( yh )Θ1 = f NF Ω NF , Zhi
t S
, Hhi
t
(
Θ1hi , ε hi , O Ψ , Zhi
t S
) ) ( t
, X 2thi , v hi ,
S t
)
(5.12)
( t
F Ω , Wht , Hh
F
S
Θ1h
s
,ε ( ) t
h ), y | ),
NL t
h
which can be transformed to a new distribution of consumption using equation
(5.6) above, as before, thus generating a new poverty rate that can be compared
with the actual.
To account for changes in poverty due to changes in education, the process
is slightly more complicated because education affects both occupational struc-
ture and earnings. As a result, we substitute the parameters for the educational
s t
choice equation estimated in time s, q 2hi , with those estimated for time t, q 2hi ,
in the H function. As before, to allow for individuals to change educational
levels in the simulation, the residual terms in the educational multinomial logit
model are taken from an extreme value distribution, as described in annex 5A.
Because education affects the choice of activity, the composition of activity
across the distribution needs to be simulated. The resulting new distribution
of activities is then introduced into the earnings functions, along with the
new educational structure, and we obtain the following income counterfactual
distribution:
s →t
( (
log ( yh )Θ2 = f NF Ω NF , Zhi , Hhi Θ 2hi , ε
t S
t
ˆ hi
t t(
,O Ψ , Zhi hi , Θ 2hi v
, X 2S
S
) ) ( ˆ hi
t
, (5.13)
S
( S
) )
( t
F ΩF ,Wht , H h
t ˆh
Θ 2h , ε t ( s
NL t
, yh | . ) ) )
This counterfactual distribution can then be transformed to a new distribution
of consumption using equation (5.6), as before, and compared with the actual
distribution in equation (5.9). The contribution of the change in educational
structure to the change in poverty between s and t can be estimated by the dif-
ference between poverty indices of the actual (equation [5.9]) and counterfac-
tual (equation [5.13]) distribution: ∆PΘ s →t
2 = P yh
t
− P ( yh )Θ2 .
s →t
{ }h { }
h
The difference between the distributions generated by the series of changes
in demographics, educational structure, sector choices, and activity choices is
comparable to the endowment effect in the standard Oaxaca (1973) and
Blinder (1973) decomposition. This difference is that, in each case, a new
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
120 Why Has Labor Income Increased? An In-Depth Approach to Understanding Poverty Reduction
counterfactual distribution is generated, and as a result we can look at the
contributions to any distributional summary statistic, including changes in
poverty.
Changes in Poverty due to Changes in Returns to Endowments
Poverty rates may have changed not only because of changes in endowments
but also because changing market conditions may have changed the returns to
existing endowments. For instance, higher educational attainment would likely
increase incomes and therefore reduce poverty. However, if the supply of edu-
cated workers outpaces the demand for such workers, the premium for having
higher educational levels would fall, and thus the returns to higher education
likely would also fall. This idea is often associated with Tinbergen’s (1975)
“race” between technological progress (which he saw as raising the demand for
skills) and the expansion of formal education, which raises the supply of skills
(Ferreira 2012).
To ascertain the extent to which changes in the returns to endowments con-
tributed to poverty reduction, we simulate the counterfactual household income
distribution by substituting the estimated returns to individual and household
characteristics (Ω) computed for period s into the earnings of every household at
time t, holding everything else constant:
( (
log ( yh )Ω = f NF Ω NF , Zhi
s →t S
t t
, Hhi
t
) (
, ε hi ,O Ψ , Zhi
t
t
, X 2thi , v hi ,
t
)
(5.14)
( S
F Ω ,W , H , ε , y
F h
t t
h
t
h ) NL t
h | . )
This simulation yields the earnings of each household in the sample if the
returns to each observed characteristic had been those observed at time s rather
than the actual returns observed at time t, keeping everything else constant.5 The
contribution to the overall change in the distribution assigned to a change in
returns (Ω s →t ) between periods s and t can be obtained by comparing equation
(5.9) with equation (5.14). Therefore, the effect of a change in endowment
returns on a change in poverty is ∆PΩ s →t t
= P yh { }
h
s →t
− P ( yh )Ω
h
{
. }
The difference between this simulated distribution of household incomes,
{ i }Ω
y
s →t
, and the actual distribution is equivalent to the price effect in the standard
Oaxaca (1973) and Blinder (1973) decomposition.
Changes in Poverty due to Unobservable Factors
Note that, up to this point, we have accounted for changes in the distribution
due to observable factors. However, it is likely that factors that are unobserv-
able at the household and individual levels nevertheless affect the distribu-
tion of consumption and therefore affect changes in poverty. Although we
cannot completely capture these effects, we can simulate the effect of
changes in the residuals in the earnings equations.6 To do so, we rescale the
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Why Has Labor Income Increased? An In-Depth Approach to Understanding Poverty Reduction 121
estimated residuals of the earning and net revenue equations for nonfarm and
farm workers of time t by the ratio of their standard deviations. This counter-
factual is defined as follows:
log ( yh )ε
s →t t t
= f NF Ω NF , Zhi
t ˆt σ ε
, H hi
ˆS
, ε hi t ,O Ψ , Z hi
σˆ
t
t
( ˆ hi
, X 2thi , v t
, )
ε
(5.15)
t ˆ S NL
t ˆt σ ε
F ΩF ,Wht , H h , ε h t , yh |t .
σˆε
The contribution to a change in poverty from a change in unobservable factors
(ε s →t ) can be obtained by calculating the poverty rate for the counterfactual
distribution generated by equation (5.15) with the original poverty rate at time
t as follows: ∆Pεs →t = P yh
t
{ } s →t
− P ( yh )ε
h
. { } h
Changes in Poverty as a Result of Changes in Nonlabor Income
To account for changes in nonlabor income, the nonparametric technique first
described in chapter 2 can be used. To do so, cells of household heads with the
same educational level, gender, and area (urban or rural) must be created. Then,
quantiles of nonlabor income must be created for each cell.
The counterfactual distribution of nonlabor income in year t is estimated by
assigning the mean value of nonlabor income of quantile q in cell c in year s, to
the same quantile and cell in year t. In other words, we rank the two distributions
by per capita household nonlabor income, and if q is the rank of a household with
NL
income yh at time t, we replace it with the nonlabor income of the household
with the same rank at time s. This counterfactual distribution can be expressed
formally as follows:
s →t
( (
log ( yh ) y NL = f NF Ω NF , Zhi
S
t
t
, Hhi
t
, ε hi O Ψ , Z hi
S
) (
, X 2S
t
ˆ hi
hi , v
t
, )
h
(5.16)
( t
F Ω ,WhS , Hh
t
F ,ε
t
h ), y NL S
h q )
| .
As before, we can compare with the actual distribution described in equation
(5.9), calculate poverty indices, and obtain the contribution of nonlabor income
→t
to poverty change between years s and t as follows: ∆PysNL t
= P yh
h
{ } h {
− P ( yh )sy→
NL
h
t
}.
h
Changes in Poverty as a Result of Changes in Consumption-Income Ratio
Finally, it is important to note that each of the counterfactual distributions simu-
lated so far assumes that the consumption-to-income ratio in period t remains
constant. However, in practice, this ratio could change either because of changes
in households’ savings rates or changes in measurement error.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
122 Why Has Labor Income Increased? An In-Depth Approach to Understanding Poverty Reduction
To account for changes in the consumption-income ratio, the nonparametric
technique described above can be used. In particular, cells of household
heads with the same level of education, gender, and area (urban or rural) must
be created. Then, quantiles of consumption must be created for each cell. The
counterfactual distribution of the consumption-to-income ratio in year t is
estimated by assigning the mean value of this ratio for quantile q in cell c in
year s to the same quantile and cell in year t. In other words, we rank the two
distributions by consumption, and if q is the rank of a household with con-
sumption-to-income ratio ϑ ht
at time t, we replace it with the consumption-to-
income ratio of the household with the same rank at time s. Equation (5.6)
then becomes
ϑh
s
s →t
Ch = yh
t
, (5.17)
n
which creates a counterfactual distribution from which a new poverty rate can
be calculated and compared with the poverty rate obtained in period t.
The Cumulative Decomposition Technique
All of the decompositions described above can be done on their own, holding
everything else constant. We refer to the results from that analysis as the marginal
effects. However, as mentioned before, all of these changes are likely to occur
over the course of a decade. Moreover, the effects of interaction between these
elements could be important in accounting for the changes in poverty. For
instance, changes in the educational composition could reinforce changes in the
sectoral composition of employment. Therefore, it is important to take these
potential interactions into account.
The cumulative decomposition technique allows us to account for these inter-
actions by calculating each effect successively and cumulating them into a set of
counterfactuals that contain the cumulative effects of multiple changes. We
attribute all of the additional contribution to poverty change to each specific
factor successively being added.
As described in chapters 2 and 3, it is important to note that the
magnitude of the contribution will depend on the path chosen for the
decomposition. However, the large number of factors involved in calculating
Shapley values (from s to t and vice versa) is beyond the scope of this chap-
ter. Instead, we use theory to better inform the path to be adopted.
In particular, we follow Bourguignon, Ferreira, and Leite (2008) by first
calculating the effects of changes in the characteristics of the population—
beginning with the most exogenous variables (such as age, gender, region,
and area) and following with changes in the population’s sectoral, educa-
tional, and activity structure. With these results, we then calculate the
changes in farm and nonfarm earnings resulting from changes in the returns
to these characteristics, the changes in nonlabor incomes, and the changes in
the consumption-to- income ratio.7
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Why Has Labor Income Increased? An In-Depth Approach to Understanding Poverty Reduction 123
Final Remarks
This chapter has presented a methodology to account for the finding in c hapter 3
that total labor income increased across countries where poverty declined sub-
stantially over the past decade. As chapter 4 then described in detail, there are
both statistical and structural approaches to account for the contribution of
endowment and price effects in explaining distributional changes. In contrast to
statistical methods, structural methods aim to account for the factors underpin-
ning observed changes in poverty outcomes by specifying an economic model
while also incorporating statistical analysis. In particular, these structural models
take into account the behavior of agents in response to changes in their socioeco-
nomic environments.
The approach proposed in this chapter rests on the typical economic assump-
tion that agents seek to maximize their utility subject to constraints. In particular,
throughout this chapter, we consider the Roy (1951) model of choice and con-
sequences, which stems from the optimization principle and applies to discrete
choice problems. We use it to model individuals’ educational levels, sectoral
(industry) choices, and activity choices or occupational statuses.
These models are set up sequentially, so that changes in education affect sectoral
and activity choice. All of these, in turn, affect earnings, which are modeled separately
for nonfarm individuals (in the form of standard Mincer equations) and as net revenue
functions for farm households. This setup allows us to distinguish between distri-
butional changes resulting from changes in endowments and those resulting from
returns to those endowments (following Bourguignon, Ferreira, and Leite [2008]).
Once all of these models are estimated for individuals and households in each
time period, the estimated coefficients from one year can be replaced with the
estimates from another year to simulate the impact of changes in each element
at a time. The resulting series of counterfactual income distributions can then be
used to estimate a counterfactual poverty measure for comparison with the
observed outcome while holding everything else constant. By changing one ele-
ment at a time, these decompositions allow us to account for the observed
changes in poverty. In addition, we present a method to estimate these counter-
factuals cumulatively, thereby accounting for the impact of concurrent changes.
The method presented here makes some important innovations relative to
previous work. First, it models farm income at the household level and models
the earnings of individuals in those farm households who have secondary occu-
pations, thus recognizing both that farm households typically make labor deci-
sions as a unit and that these households can be highly diversified. Second, it
measures welfare using a consumption aggregate, thus accounting for the contri-
bution of changes in the consumption-to-income ratio. Finally, it ensures that
changes in the composition of the work force’s activity, sectoral, and educational
choices are consistent with the counterfactual choices.
The next chapter concludes the volume by applying the methodology pre-
sented in this chapter to three countries that reduced poverty drastically between
2000 and 2010: Bangladesh, Peru, and Thailand.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
124 Why Has Labor Income Increased? An In-Depth Approach to Understanding Poverty Reduction
Annex 5A: Estimating the Residual Term in Multinomial Logit
The allocation of individuals across activities is represented through a multino-
mial logit model. To calculate the utility of activity s and therefore allow for
people to change activities in the simulation exercise when either Zhi or y s
changes, we must estimate the residual terms of the choice model (v is ), which
are unobserved.
The residual terms must be drawn from extreme value distributions in a way
that is consistent with the observed choices. Train and Wilson (2008) define the
distribution functions of the extreme value errors conditional on the chosen
alternative. In particular, assume that the alternative zero is chosen (j = 0), and
J
∑ exp −V ,
j 0j 0j
j 0
denote Zhi Ψ = V for j = 0,…, J. Define V = V − V and D =
0 j
hi
hi hi hi hi hi
j=0
0 0
where Phi = 1 Dhi is the logit choice probability. Then the cumulative distribu-
0
tion function (cdf ) for the alternative chosen v hi is
( 0
F v hi ) 0
|alternative 0 is chosen = exp( − Dhi 0
exp Vhi . (5A.1)( )
Calculating the inverse of this distribution, we have
( ) ( )
− ln − ln ( µ ) , (5A.2)
0
0
v hi = ln Dhi
where μ is drawn from a uniform distribution between zero and one. Error terms
j
for other alternatives v hi with j ≠ 0 must be calculated conditioned on the error
terms of the alternative chosen (v 0 hi
j
). The distribution for these errors is
( j
F v | alternative 0 is chosen, j ≥ 1 = )
( j
exp −exp v hi ( ))
hi
0
(5A.3)
exp −exp − V hi + v hi
0j
j 0j 0
for v hi < V hi + v hi . The inverse of this distribution is
j
v hi = − ln − ln m v hi µ , (5A.4)
0
( )
( )
0
where m v hi = exp − exp − V hi + v hi , and m is drawn from a uniform dis-
0j 0
tribution between zero and one. We repeat this same method when an alterna-
tive other than zero is chosen. A similar approach is taken for the sectoral and
educational choice models.
Annex 5B: The Cumulative Decomposition Technique
The cumulative decomposition technique allows us to account for these interac-
tions by calculating each effect successively and cumulating them into a set of
counterfactuals that contain multiple changes. We attribute all of the additional
contribution to poverty change to each specific factor being added. However, the
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Why Has Labor Income Increased? An In-Depth Approach to Understanding Poverty Reduction 125
magnitude of that contribution will depend on the path chosen for the decom-
position. We follow the path suggested by Bourguignon, Ferreira, and Leite
(2008)—creating a cumulative counterfactual distribution from which poverty
headcount rates can be calculated at every point in the adopted path and com-
pared with the actual poverty rates to estimate the contribution of each addi-
tional change on the observed change in poverty.
We begin by calculating the effects of changes in the characteristics of the
population, beginning with what we consider to be the most exogenous variables,
s
including age, gender, region, and area ( Zhi , Whs ). Specifically, first we reweigh
the demographic characteristics of year t in such a way that their structure rep-
licates the demographic characteristics of year s as follows:
{( (
∆P s →t = P f NF Ω NF , Z hi
t t ˆt
, Hhi , ε hi ,O Ψ , Z hi
t
t
t
, H hi ˆ hi
,v t
, ) ( t
)
( t ˆt
F ΩF ,Wht , Hh
t
, ε h , yh
NL t
|q ) )} h
(5B.1)
{( ( s
− P f NF Ω , Z hi t ˆt
, Hhi , ε hi ,O Ψ , Z hi
s
t
, H hi
NF
t
ˆ hi
,v t
, ) ( t
)
( t ˆt
F ΩF ,Whs , Hh
t
, ε h , yh
NL t
|q h . ) }
Second, keeping the demographic effects, we add the education structure
s
change (Θ2hi ):
{( ( s
= P f NF Ω NF , Zhi , Hhi
t
t ˆt
, ε hi ,O Ψ , Zhi
s t
, Hhi ˆ hi
,v t
, ) ( t
)
( t
t ˆt
F ΩF ,Whs , Hh , ε h , yh
NL t
|q ) )} h
(5B.2)
{( (
− P f NF Ω
t
NF ,Z ,Hs
hi
s
hi (Θ1 ,Θ2 ) ,εˆ ) ,O ( Ψ , Z
t
hi
s
hi
t
hi
t
s
hi
s
, H , Θ2 v
hi ˆ
s
t
hi hi ),
( t
F ΩF ,Whs , Hh
s ˆ hi
Θ1hi , Θ2hi , ε t
( NL t
, yh |q
t s
) ) )} . h
Third, preserving the previous changes, we include the change in occupation
( )
structure Ψ s :
{( ( s
= P f NF Ω NF , Zhi s
, Hhi
t
ˆ hi
Θ1hi , Θ2hi , ε t
(
s
,O Ψ , Zhi s
, Hhi
t
ˆ hi
Θ2hi , v t
,
s
) ) ( t
( s
) )
( t
F ΩF ,Whs , Hh
s ˆh
Θ1hi , Θ2hi , ε (
t NL t
, yh |q
t s
) ) )} h
{( (
(5B.3)
− P f NF Ω
t
NF
s
,Z ,H
hi
s
hi (Θ1 ,Θ2 ) ,ε ) ,O ( Ψ , Z
t
hi
s
hi
t
hi
s
s
hi
s
, H , Θ2
hi
( s
hi ) , vˆ ) ,
t
hi
( t
F ΩF ,Whs , Hh
s ˆh
Θ1hi , Θ2hi , ε (
t NL t
, yh |q
t s
) ) )} . h
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
126 Why Has Labor Income Increased? An In-Depth Approach to Understanding Poverty Reduction
s
Fourth, we add the change in the structure of economic sectors (Θ1hi ):
{( ( s
= P f NF Ω NF , Zhi s
t
, Hhi ˆ hi
Θ1hi , Θ2hi , ε t s
,O Ψ , Zhi s
, Hhi ( ˆ hi
Θ2hi , v t
t
,
s
) ) ( s
( s
) )
( s
F ΩF ,Whs , Hh
t
ˆh
Θ1hi , Θ2hi ε t NL t
, yh |q ( t s
) ) )} h
(5B.4)
{( (
− P f NF Ω
t
NF ,Z ,H s
hi
s
hi (Θ1 ,Θ2 ) ,εˆ ) ,O ( Ψ , Z
s
hi
s
hi
t
hi
s
s
hi ,H s
hi
(Θ2 ) , vˆ )) ,
s
hi
t
hi
( t
F ΩF ,Whs , Hh
s ˆh
Θ1hi , Θ2hi ε t NL t
, yh |q ( s s
) ) )} . h
s
Fifth, we include the returns to the nonfarm sector (Ω NF ):
{( ( s
= P f NF Ω NF , Zhi s
, Hhi ˆ hi
Θhi , ε t
t
s
,O Ψ , Zhi s
, Hhi ˆ hi
Θ2hi , v t ( ) ) ( s s
( s
) )
( t
s
F ΩF ,Whs , Hh ˆh
Θhi , ε t NL t
, yh |q ( ) ) s
)} h
(5B.5)
{( (
− P f NF Ω
s
NF ,Z ,Hs
hi
s
hi
(Θ ) ,εˆ ) ,O ( Ψ , Z
s
hi
t
hi
s
s
hi ,H s
hi
(Θ2 ) , vˆ ) ,s
hi
t
hi
( s
F ΩF ,Whs , Hh
t
ˆh
Θhi , ε t NL t
, yh |q ( ) ) s
)} . h
s
Sixth, we change the returns to the farm sector (ΩF ):
{( ( s
= P f NF Ω NF , Zhi s
, Hhi ˆ hi
Θhi , ε t
s
s
,O Ψ , Zhi s
, Hhi ˆ hi
Θ2hi , v t
, ( ) ) ( s s
( s
) )
( s
F ΩF ,Whs , Hh
t
ˆh
Θhi , ε t NL t
, yh |q ( ) ) s
)} h
(5B.6)
{( ( s
− P f NF Ω NF , Zhi s
, Hhi ˆ hi
Θhi , ε t
s
s
,O Ψ , Zhi s
, Hhi ˆ hi
Θ2hi , v t
, ( ) ) ( s s
( s
) )
( s
F ΩF ,Whs , Hh
s
ˆh
Θhi , ε t NL t
, yh |q ( ) ) s
)} . h
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Why Has Labor Income Increased? An In-Depth Approach to Understanding Poverty Reduction 127
Seventh, we change residuals of both earnings and net revenues equations:
{( ( s
s
= P f NF Ω NF , Zhi s
, Hhi ( ) ) (
ˆ hi
Θhi , ε t
s
s
,O Ψ , Zhi s
s
, Hhi ˆ hi
Θ2hi , v t
, ( s
) )
( s
s
F ΩF ,Whs , Hh ( ) )
ˆh
Θhi , ε
s
t NL t
, yh |q )}h
s
s
− P f NF Ω NF , Zhi s
, Hhi ( )
s ˆs
t σε
ˆ hi
Θhi , ε
σ t
ˆε
s
(
s
,O Ψ , Zhi s
, Hhi
s
(ˆ hi
Θ2hi , v t ) )
,
(5B.7)
s s
F ΩF ,Whs , Hh
s
( ) ˆh
Θhi , ε t σε
ˆ s NL
ˆεt
, yh
|tq .
σ h
Eighth, we add the change in nonlabor income components and the consump-
tion ratio. The latter is not formally displayed in this example:
s ˆs s s
s t σε t
Θhi , ε s
Θ2hi ,ν
s s
s
= P f NF Ω NF , Zhi , Hhi ˆ hi t
,O y , Zhi , Hhi ˆ hi ,
ˆ
σε
s s
ˆ s NL t
t σε
Θhi , ε
s
F ΩF ,Whs , Hhi ˆh t
, yh |q
σˆ
ε
h
(5B.8)
s ˆs s s
t σε s t
s
Θhi , ε Θ2hi ,ν
s s
s
− P f NF Ω NF , Zhi , Hhi ˆ hi t
,O y , Zhi , Hhi ˆ hi ,
ˆ
σε
s s
ˆ s NL s
t σε
Θhi , ε
s
F ΩF ,Whs , Hhi ˆh t
, yh |q .
ˆ
σε h
Finally, we take into account changes in the consumption-to-income ratio as
described in the main text. The cumulative counterfactual change in poverty and
the observed change in poverty are compared. Any change that is not accounted
for is described as a residual change.
As before, the cumulative decomposition technique is performed both consid-
ering s as the initial year and then considering t as the initial year. The average of
these decomposition effects is the final result reported in the analysis.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
128 Why Has Labor Income Increased? An In-Depth Approach to Understanding Poverty Reduction
Notes
1. By “activity choice,” we refer to the individual’s occupational status, which may
include paid employment, self-employment, unpaid labor, or unemployment.
2. Note that earnings may be underestimated to the extent that individuals opt out of
the labor force because their reservation wage is lower than their market wage
(Heckman 1979). Although this is a well-known bias, we do not attempt to correct
for it given the complexity of the decompositions that follow.
3. Bear in mind that this approach does not solve all path-dependence problems.
Shapley values must be estimated to tackle this difficulty.
4. The notation s → t refers to using the parameters estimated in period s in the equation
for time t.
ˆ t are assumed
5. The returns to the unobserved characteristics behind the residual term ε
to be unchanged.
6. The estimated error terms for the multinomial logit models are not rescaled.
7. For a full description of the cumulative approach, please refer to annex 5B.
Bibliography
Blinder, Alan S. 1973. “Wage Discrimination: Reduced Form and Structural Estimates.”
Journal of Human Resources 8 (4): 436–55.
Bourguignon, François, and Francisco H. G. Ferreira. 2005. “Decomposing Changes in the
Distribution of Household Incomes: Methodological Aspects.” In The Microeconomics
of Income Distribution Dynamics in East Asia and Latin America, edited by François
Bourguignon, Francisco H. G. Ferreira, and Nora Lustig, 17–46. Washington, DC:
World Bank.
Bourguignon, François, Francisco H. G. Ferreira, and Nora Lustig, eds. 2005. The
Microeconomics of Income Distribution Dynamics in East Asia and Latin America.
Washington, DC: World Bank.
Bourguignon, François, Francisco H. G. Ferreira, and Phillippe G. Leite. 2008. “Beyond
Oaxaca-Blinder: Accounting for Differences in Household Income Distributions.”
Journal of Economic Inequality 6 (2): 117–48.
Davis, Benjamin, Paul. Winters, Calogero Carletto, Katia Covarrubias, Esteban J. Quiñones,
Alberto Zezza, Carlo Azzarri, Kostas Stamoulis, and Stefania DiGiuseppe. 2010.
“A Cross-Country Comparison of Rural Income Generating Activities.” World
Development 38 (1): 48–63.
Ferreira, Francisco H. G. 2012. “Distributions in Motion: Economic Growth, Inequality,
and Poverty Dynamics.” In The Oxford Handbook of the Economics of Poverty, edited by
Philip N. Jefferson, 427–62. New York: Oxford University Press.
Haggblade, Steven, Peter B. R. Hazell, and Thomas Reardon. 2007. Transforming the Rural
Non-Farm Economy: Opportunities and Threats in the Developing World. Baltimore, MD:
Johns Hopkins University Press.
———. 2010. “The Rural Non-Farm Economy: Prospects for Growth and Poverty
Reduction.” World Development 38 (10): 1429–41.
Heckman, James J. 1979. “Sample Selection Bias as a Specification Error.” Econometrica
47 (1): 153–61.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Why Has Labor Income Increased? An In-Depth Approach to Understanding Poverty Reduction 129
McFadden, Daniel. 1974a. “Conditional Logit Analysis of Qualitative Choice Behavior.” In
Frontiers in Econometrics, edited by Paul Zarembka, 105–42. New York: Academic
Press.
———. 1974b. “The Measurement of Urban Travel Demand.” Journal of Public Economics
3 (4): 63–82.
Oaxaca, Ronald L. 1973. “Male-Female Wage Differentials in Urban Labor Markets.”
International Economic Review 14 (3): 693–709.
Roy, Arthur D. 1951. “Some Thoughts on the Distribution of Earnings.” Oxford Economic
Papers 3 (2): 135–46.
Tinbergen, Jan. 1975. Income Differences: Recent Research. Amsterdam, the Netherlands:
North-Holland.
Train, Kenneth E., and Wesley W. Wilson. 2008. “Estimation on Stated-Preference
Experiments Constructed from Revealed-Preference Choices.” Transportation Research
Part B: Methodological 42 (3): 191–203.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Chapter 6
Understanding Poverty Reduction
in Bangladesh, Peru, and Thailand
Introduction
The decomposition method proposed in chapter 2 and implemented in c hapter 3
distinguished labor income growth as the main contributor to poverty reduction.
However, the method’s main limitation is that it cannot explain why labor
incomes increased: Did they grow because the populations increased their own
human capital endowments (such as higher educational levels or other produc-
tive assets) or because of changes in returns to those endowments? For this
answer, we have turned to an alternative decomposition technique that imposes
an underlying labor model and greater structure.
This chapter implements chapter 5’s proposed approach: to distinguish
between distributional changes on account of (a) changes in endowments and the
returns to those endowments; (b) changes in occupational and sectoral choice;
(c) changes in the population’s geographical, age, and gender structure; and (d) the
contribution of the nonlabor income dimensions (such as public transfers, remit-
tances, and other private transfers) previously explored in chapter 3. This micro-
decomposition approach is adapted from Bourguignon, Ferreira, and Leite (2008).
In focusing on Bangladesh, Peru, and Thailand, the chapter highlights their
impressive reductions in poverty while taking advantage of the availability, com-
parability, and transparency of their microdata. The moderate national poverty
headcount rate in each of these countries fell by more than 12 percentage points
during the past decade, partly because of high gross domestic product (GDP)
growth, which was well above 4 percent per year during the 2002–08 period.1
In the cases of Peru and Thailand, GDP sharply decelerated in the wake of the
financial crisis in 2009, only to rebound quickly the following year. In contrast,
GDP growth in Bangladesh got through the crisis unscathed.
In all three countries, employment and public social transfers increased, as did
remittances. However, the changing patterns of income distribution varied across
countries, as did the roles of different factors in reducing poverty. Moreover, the
countries progressed from very different starting points: Bangladesh, despite
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7 131
132 Understanding Poverty Reduction in Bangladesh, Peru, and Thailand
strong growth, remains a low-income country with a GDP per capita of $1,710.
Peru and Thailand are firmly in the middle-income ranks, with GDP per capita
of $10,439 and $9,630, respectively (all figures in purchasing power parity
terms). In addition, Peru is highly urbanized, as opposed to Bangladesh and
Thailand, whose shares of urban population remain below 30 percent.
To understand how these diverse economies each reduced poverty so substan-
tially over a similar period of time, we applied the structural decomposition
approach progressively developed in the previous chapters of this volume.
Chapter 5 had proposed a method to estimate an educational choice model,
a sectoral choice model, a work-activity choice model, and individual and house-
hold earnings equations. Once all of these models are estimated for individuals
and households in each time period, the estimated coefficients from one year can
be replaced with the estimates from another year to simulate the impact of
changes in each element at a time.
This process enables us to build a series of counterfactual income distributions
to estimate a counterfactual poverty measure for comparison with the observed
outcome while holding everything else constant. By changing one element at
a time, these decompositions allow us to account for the observed changes in
poverty. In addition, we can estimate these counterfactuals cumulatively, thereby
accounting for the impact of concurrent changes.
The chapter next describes the evolution of poverty and economic growth in
Bangladesh, Peru, and Thailand—highlighting the countries’ similarities and differ-
ences in their initial- and end-period outcomes. “The Decomposition Approach”
then presents the results of the decomposition exercise for each country. The
chapter ends the volume by summarizing our findings and laying out the possi-
bilities for further application.
Country Context
For all their differences, Bangladesh, Peru, and Thailand have at least one
thing in common: they drastically reduced poverty rates over the past decade
(figure 6.1). The share of households living at or below the international
poverty line corresponding to $2.50 per day (as shown in table 6.1) fell by an
average of 1.8 percentage points per year in Bangladesh (2000–10), 2.7 per-
centage points per year in Peru (2004–10), and 1.6 percentage points per year
in Thailand (2000–09).
To calculate the share of population living below the moderate poverty line, the
three countries use the “cost of basic needs” approach, although the “basket” for
basic needs and specific lines are different.2 Moreover, this reduction in poverty
coincided with strong GDP growth over the decade, which averaged 5.8 percent
in Bangladesh, 6 percent in Peru, and 4.4 percent in Thailand ( figure 6.2).
There is considerable evidence that economic growth is strongly and nega-
tively correlated with changes in poverty (Ravallion and Chen 2007). Indeed,
in these three countries, using the standard Datt and Ravallion (1992) decom
position, growth does indeed explain most of the observed poverty reduction
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Understanding Poverty Reduction in Bangladesh, Peru, and Thailand 133
Figure 6.1 Change in Moderate Poverty Rates in Bangladesh, Peru, and
Thailand, 2000s
60
50
40
Population (%)
30
20
10
0
Bangladesh Peru Thailand
2000–10 2004–10 2000–09
Initial Final
Sources: Household survey results from Bangladesh (HIES 2000–10), Peru (ENAHO 2004–10), and Thailand
(SES 2000–09).
Note: “Moderate poverty” headcounts are based on each country’s national moderate poverty line. HIES =
Household Income and Expenditure Survey (Bangladesh Bureau of Statistics); ENAHO = Encuesta Nacional
de Hogares (Peru National Institute of Statistics and Information); SES = Household Socio-Economic Survey
(Thailand National Statistical Office).
Table 6.1 Change in Poverty Rates, by Level, in Bangladesh, Peru, and Thailand, 2000s
Poverty lines
National National
$1.25 $2.50 $4.00 extreme moderate
Poverty rate in Bangladesh, 2000–10
Initial period (%) 57.7 89.2 — 34.5 49.1
Final period (%) 40.3 84.0 — 17.6 31.5
Change (ppts) −17.4 −5.2 — −16.9 −17.6
Average annual change (ppts) −1.7 −0.5 — −1.7 −1.8
Annualized percentage change (%) −3.5 −0.6 — −6.5 −4.3
Poverty rate in Peru, 2004–10
Initial period (%) 3.5 22.9 45.8 17.4 49.1
Final period (%) 0.8 11.7 30.0 10.5 33.0
Change (ppts) −2.6 −11.2 −15.8 −6.9 −16.2
Average annual change (ppts) −0.4 −1.9 −2.6 −1.2 −2.7
Annualized percentage change (%) −21.1 −10.6 −6.8 −8.1 −6.4
table continues next page
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
134 Understanding Poverty Reduction in Bangladesh, Peru, and Thailand
Table 6.1 Change in Poverty Rates, by Level, in Bangladesh, Peru, and Thailand, 2000s (continued)
Poverty lines
National National
$1.25 $2.50 $4.00 extreme moderate
Poverty rate in Thailand, 2000–09
Initial period (%) 0.2 7.9 31.4 — 23.9
Final period (%) 0.0 2.5 16.6 — 9.8
Change (ppts) −0.2 −5.3 −14.8 — −14.1
Average annual change (ppts) … −0.6 −1.6 — −1.6
Annualized percentage change (%) −16.8 −11.8 −6.8 — −9.4
Sources: Household survey results from Bangladesh (HIES 2000–10), Peru (ENAHO 2004–10), and Thailand (SES 2000–09).
Note: … = negligible; — = not available; ppts = percentage points; HIES = Household Income and Expenditure Survey
(Bangladesh Bureau of Statistics); ENAHO = Encuesta Nacional de Hogares (Peru National Institute of Statistics and
Information). SES = Household Socio-Economic Survey (Thailand National Statistical Office). The $1.25, $2.50, and $4.00
poverty lines are daily household income amounts that correspond to international measurements of varying poverty levels
in low- and moderate-income countries. The “National extreme” and “National moderate” poverty levels refer to country-
based poverty lines calculated based on the “cost of basic needs,” such as specified calorie levels per person per day, and
other essentials, such as clothing and shelter.
Figure 6.2 GDP in Bangladesh, Peru, and Thailand, 2000–10
12
10
8
Real GDP growth (%)
6
4
2
0
–2
–4
2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010
Bangladesh Thailand Peru
Source: World Bank 2011.
Note: GDP = gross domestic product.
(table 6.2). In each case, more than 80 percent of the reduction in poverty is
explained by growth in mean income, whereas better income distribution
explains less than 20 percent of the reduction.
Although these estimates of the reduced-form relationships between eco-
nomic growth, inequality, and poverty have been useful in identifying empirical
regularities, they cannot explicitly link how growth and poverty reduction are
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Understanding Poverty Reduction in Bangladesh, Peru, and Thailand 135
Table 6.2 Growth and Redistribution Decomposition of Moderate Poverty Rate Changes in
Bangladesh, Peru, and Thailand, 2000s
Percent
Bangladesh Peru Thailand
2000 vs. 2010 2004 vs. 2010 2000 vs. 2009
Poverty headcount rate: FGT(0)a
t0 48.9 (1.2) 49.1 (0.2) 23.9 (0.2)
t1 31.5 (0.9) 33.0 (0.2) 9.8 (0.1)
Actual percentage change −17.3 (1.5) −16.1 (0.3) −14.1 (0.2)
Change resulting from growth −15.8 (1.6) −13.8 (0.4) −14.2 (0.3)
Change resulting from redistribution −1.6 (1.5) −2.4 (0.3) 0.1 (0.3)
Contributions to the decline in FGT(0)
Growth 91 85 101
Redistribution 9 15 −1
Poverty gap FGT(1)b
t0 12.8 (0.5) 16.4 (0.1) 5.5 (0.1)
t1 6.5 (0.2) 9.3 (0.1) 1.8 (0.0)
Actual percentage change −6.2 (0.5) −7.1 (0.1) −3.6 (0.1)
Change resulting from growth −5.6 (0.6) −5.9 (0.2) −3.7 (0.1)
Change resulting from redistribution −0.6 (0.6) −1.1 (0.2) 0.0 (0.1)
Contributions to the decline in FGT(1)
Growth 90 84 103
Redistribution 10 16 −3
Sources: Household survey results from Bangladesh (HIES 2000–10), Peru (ENAHO 2004–10), and Thailand (SES 2000–09).
Note: Poverty rates are based on per capita consumption. The Datt and Ravallion (1992) decomposition reported here uses
the Shapley (1953) approach described in chapter 2. Standard errors shown in parentheses. t0 = the initial year and t1 = the
last year. HIES = Household Income and Expenditure Survey (Bangladesh Bureau of Statistics); ENAHO = Encuesta Nacional
de Hogares (Peru National Institute of Statistics and Information); SES = Household Socio-Economic Survey (Thailand National
Statistical Office).
a. FGT(0) = the poverty headcount rate. FGT refers to the Foster, Greer, and Thorbecke (1984) metric, a generalized measure of
poverty in an economy.
b. FGT(1) = the poverty gap, which estimates the depth of poverty, defined as the average percentage distance between the
incomes of the incomes of the poor and the poverty line (Foster, Greer, and Thorbecke 1984).
related (Ferreira 2010). In particular, we would like to capture the heterogeneity
of impacts throughout the distribution and to be able to account for the contri-
butions of demographics, sectoral, occupational, and other labor and nonlabor
dimensions toward reducing poverty.
Elements That Could Affect Poverty Reduction
Changes in poverty can be decomposed into those resulting from changes in
demographics, changes in labor incomes, changes in nonlabor incomes, and
changes in consumption-to-income ratios. Changes in the population’s demo-
graphics include changes in age, gender, and regional structure. Changes in labor
incomes could result from improvements in human capital (such as education)
and other endowments (such as greater access to productive assets), all of which
would have increased labor productivity. Alternatively, labor incomes could have
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
136 Understanding Poverty Reduction in Bangladesh, Peru, and Thailand
risen simply because demand for certain characteristics outstripped supply—
representing increased returns to a population’s endowments. In addition,
increases in nonlabor incomes could have reduced poverty through higher remit-
tances or public transfers. Finally, households are likely to change the share of
income they dedicate to consumption, which affects the very measurements of
consumption from which poverty rates are calculated. We look at each of these
types of changes in turn.
Demographic Changes
Population growth slowed considerably in each of the countries considered over
the past decade (figure 6.3). The change has been significant enough that in
Bangladesh and Peru, the youth bulge observed in earlier periods has now
reached working age (15–64 years old).3 As a result, the share of adults per
household increased in each case (table 6.3). In other demographic changes, the
share of employed women in all three countries increased slightly. Finally, there
have been population shifts across regions, with a clear tendency toward greater
urbanization in Bangladesh and Thailand.
Changes in Labor Income
Labor income growth is a primary factor to examine regarding the observed
poverty reduction. The growth incidence curves in figure 6.4 show that labor
income in the 2000s grew across the income distribution in each country.
Figure 6.3 Population Growth in Bangladesh, Peru, and Thailand, 2000–10
a. Total population growth b. Population growth
of working-age adults
2.0 75
1.8
Population in 15–64 age group (%)
1.6 70
Population growth (%)
1.4
1.2 65
1.0
0.8 60
0.6
0.4 55
0.2
0 50
01
01
03
00
02
03
09
10
0
02
08
09
10
04
08
04
05
06
05
06
07
07
0
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
Bangladesh Peru Thailand
Source: World Bank 2011.
Note: “Working-age adults” are defined as those aged 15–64 years.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Understanding Poverty Reduction in Bangladesh, Peru, and Thailand 137
Table 6.3 Population and Labor Force Characteristics in Bangladesh, Peru, and Thailand, 2000s
Bangladesh Peru Thailand
2000 2010 2004 2010 2000 2009
Population and demographics
Total (millions) 71.0 89.8 27.6 31.4 59.8 64.7
Men (%) 50.4 48.4 49.7 49.2 47.8 47.9
Women (%) 49.6 51.6 50.3 50.8 52.2 52.1
Urban (%) 21.8 28.0 64.5 63.5 30.9 31.0
Rural (%) 78.2 72.0 35.5 36.5 69.1 69.0
Average household size (no. of members) 5.2 4.5 4.4 4.2 3.6 3.3
Average adults per household (no.) 3.1 2.9 2.7 2.6 2.5 2.4
Average adults per household (%) 63.4 68.4 68.0 70.7 75.2 78.9
Average occupied adults per household 48.4 50.1 73.2 75.7 77.7 77.4
(% of no. of adults)
Labor force participation (% of working-age population)
All 49.4 49.2 73.5 76.2 80.4 80.5
Men 83.3 84.9 83.1 84.4 87.0 86.7
Women 14.9 15.6 64.1 68.4 74.5 74.9
Employment rate (% of working-age population)
All 46.6 47.8 71.2 74.9 78.3 79.8
Men 81.6 82.9 80.9 83.4 84.3 85.8
Women 11.1 14.8 61.6 66.8 73.0 74.4
Unemployment rate (% of labor force)
All 5.6 1.4 5.1 3.8 2.5 0.9
Men 2.0 2.0 4.5 3.3 3.1 1.0
Women 25.9 0.8 5.8 4.4 2.0 0.7
Educational level (% of working-age population)
Illiterate and incomplete primary 57.2 47.1 19.5 16.4 44.3 34.4
Primary and lower secondary 33.8 43.4 31.6 29.4 32.2 39.0
Higher secondary and tertiary 8.9 9.4 47.4 52.2 23.6 26.7
Labor relationship (% of employed population)
Daily workersa 33.5 32.4 — — — —
Self-employed 46.3 42.3 61.3 57.0 58.8 57.4
Wage workers 20.2 25.4 38.7 43.0 41.2 42.6
Economic sector (% of employed population)
Agriculture 49.2 41.8 35.9 30.0 49.3 40.6
Manufacturinga 18.5 19 — — — —
Industry 4.4 5.5 9.8 10.4 17.3 20.1
Services 27.9 33.7 46.0 50.8 24.6 30.2
Area (% of employed population)
Rural 78.6 71.6 36.5 34.6 68.9 68.7
Urban 21.4 28.4 63.5 65.4 31.1 31.3
Sources: Household survey results from Bangladesh (HIES 2000–10), Peru (ENAHO 2004–10), and Thailand (SES 2000–09).
Note: “Working-age population” is defined as adults aged 15–64 years. — = not available; HIES = Household Income and Expenditure Survey
(Bangladesh Bureau of Statistics); ENAHO = Encuesta Nacional de Hogares (Peru National Institute of Statistics and Information); SES = Household
Socio-Economic Survey (Thailand National Statistical Office).
a. The statistics in Peru and Thailand make no distinctions between “daily” and “wage” workers or between the “manufacturing” and “industrial” sectors.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
138 Understanding Poverty Reduction in Bangladesh, Peru, and Thailand
Figure 6.4 Growth Incidence Curves of Labor Income in Bangladesh, Peru, and Thailand, 2000s
a. Bangladesh, 2000–10 b. Peru, 2004–10
6 0.14
5 0.12
Change in income (%)
Change in income (%)
4 0.10
3 0.08
2 0.06
1 0.04
0 0.02
–1 0
1 11 21 31 41 51 61 71 81 91 100 1 11 21 31 41 51 61 71 81 91 100
Percentile Percentile
c. Thailand, 2000–09
0.8
0.7
Change in income (%)
0.6
0.5
0.4
0.3
0.2
0.1
0
1 11 21 31 41 51 61 71 81 91 100
Percentile
Labor income per capita Total income per capita
Sources: Household survey results from Bangladesh (HIES 2000–10), Peru (ENAHO 2004–10), and Thailand (SES 2000–09).
Note: The smoothing algorithm is a polynomial regression. The degrees of the polynomial vary by country and were chosen to maximize the fit of
the data. “Total income” includes both labor and nonlabor sources of income. HIES = Household Income and Expenditure Survey (Bangladesh
Bureau of Statistics); ENAHO = Encuesta Nacional de Hogares (Peru National Institute of Statistics and Information); SES = Household Socio-
Economic Survey (Thailand National Statistical Office).
In Bangladesh, labor incomes grew faster for those at the top of the distribu-
tion than for those at the bottom, while in Peru the opposite was true. In
Thailand, incomes of the poor grew faster than those of the higher deciles
(except for those in the poorest decile).
Given the magnitude of the changes, labor income increases likely had an
influence in moving people out of poverty. What accounts for these increases?
What accounts for the different rates of growth across the distribution?
Labor incomes could have grown for several reasons pertaining to changing
labor force characteristics in all three countries:
• Increased, and better-paying, employment. In the economically active popula-
tion, increases in employment, particularly in better-paying jobs, would lead to
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Understanding Poverty Reduction in Bangladesh, Peru, and Thailand 139
declines in poverty. Although labor force participation remained relatively
stable in Bangladesh and Thailand over the 2000s, employment increased
slightly. In Peru, both labor force participation and employment increased,
particularly among women.
• Increases in wage employment. In addition to higher employment, all three
countries saw a clear movement away from self-employed work and toward
wage employment. In Bangladesh, there was also a movement away from
daily-wage work toward wage work. All of these movements are likely to have
led to better-quality employment.
• Sectoral shifts. Similarly, employment shifted away from agriculture and toward
the manufacturing and services sectors, which tend to be better paying. In
Bangladesh, there was also a slight movement toward industry, which is also
better paying than agriculture.
• Educational advancement. The improved educational composition of the work
force over the past 10 years also could have increased labor incomes. In all
three countries, a smaller share of the population was illiterate by the end of
the decade. In Bangladesh and Thailand, higher shares of the work force had
completed primary and lower secondary school; and in Peru and Thailand,
higher shares of the population had completed secondary and tertiary school
(table 6.3).
Changes in Nonlabor Income
Although labor income has clearly increased, growth in nonlabor income could
also be driving the observed reductions in poverty. Indeed, figure 6.5a shows
that both public and private transfers steadily increased in Bangladesh, Peru,
and Thailand during the 2000s. Although the countries differed widely in
terms of the magnitude of public social spending relative to the size of their
economies, public spending in each case increased by at least 25 percent during
the 2000s.
Whether public transfers are an important means of reducing poverty—and,
if so, how important—depends on how effective this spending has been, particu-
larly in terms of targeting the poor. As for private transfers, international remit-
tances have tripled in Bangladesh over the past decade, have grown modestly in
Peru, but have declined in Thailand (figure 6.5b). The question is how important
have these changes been to poverty reduction?
Changes in Consumption-to-Income Ratio
Finally, in the context of growing overall incomes, households are likely to change
the share of income they dedicate to consumption, given different marginal pro-
pensities to consume. Figure 6.6 shows that in Bangladesh and Thailand, the
consumption-to-income ratio fell over the course of the decade—a large part
of which occurred in households below the poverty line, as table 6.4 shows.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
140 Understanding Poverty Reduction in Bangladesh, Peru, and Thailand
Figure 6.5 Nonlabor Income Growth, by Source, in Bangladesh, Peru, and Thailand, 2000s
a. Subsidies and other social transfers b. International remittances
9 14
8 12
7
10
6
8
GDP (%)
GDP (%)
5
4 6
3
4
2
2
1
0 0
00
01
02
03
04
05
06
07
08
09
10
00
01
02
03
04
05
06
07
08
09
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
Peru Thailand Bangladesh
Source: World Bank 2011.
Note: GDP = gross domestic product. Data on subsidies and other social transfers were unavailable for Bangladesh preceding 2001 and for
Thailand preceding 2003.
Figure 6.6 Change in Household Consumption-to-Income Ratio in Bangladesh,
Peru, and Thailand, 2000s
10
Change in household consumption-to-
8
6
4
income ratio (%)
2
0
–2
–4
–6
–8
–10
–12
Bangladesh Peru Thailand
2000–10 2004–10 2000–09
Sources: Household survey results from Bangladesh (HIES 2000–10), Peru (ENAHO 2004–10), and Thailand
(SES 2000–09).
Note: HIES = Household Income and Expenditure Survey (Bangladesh Bureau of Statistics); ENAHO =
Encuesta Nacional de Hogares (Peru National Institute of Statistics and Information); SES = Household
Socio-Economic Survey (Thailand National Statistical Office).
In contrast, the consumption-to-income ratio increased over the course of the
decade in Peru, particularly at the bottom of the distribution.
As a result, the observed changes in consumption may be less dramatic than
what we would have otherwise expected in Bangladesh and Thailand, but more
dramatic in Peru, had the consumption-to-income ratio remained constant.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Understanding Poverty Reduction in Bangladesh, Peru, and Thailand 141
Table 6.4 Household Consumption-to-Income Ratio, by Income Decile, in Bangladesh, Peru,
and Thailand, 2000s
Units of consumption per units of income
Bangladesh Peru Thailand
2000 2010 2004 2010 2000 2009
Average 1.26 1.22 1.13 1.22 0.98 0.89
Income deciles (per capita)
1 3.69 3.46 1.96 2.52 2.74 3.01
2 1.39 1.53 1.33 1.50 2.24 1.64
3 1.14 1.29 1.22 1.31 1.77 1.5
4 1.07 1.13 1.18 1.18 1.76 1.46
5 1.05 1.03 1.12 1.13 1.74 1.29
6 1.01 0.92 1.05 1.05 1.56 1.31
7 0.97 0.83 1.00 1.01 1.49 1.21
8 0.95 0.76 0.93 0.92 1.49 1.22
9 0.88 0.67 0.84 0.85 1.59 1.15
10 0.76 0.5 0.71 0.71 1.5 1.1
Spearman ranka 0.79 0.56 0.86 0.83 0.81 0.83
Correlation 0.62 0.18 0.76 0.73 0.62 0.51
Coefficient of variation
Consumption 0.87 0.78 0.97 0.85 0.96 0.99
Income 1.06 2.99 1.35 1.26 1.33 1.49
Sources: Household survey results from Bangladesh (HIES 2000–10), Peru (ENAHO 2004–10), and Thailand (SES 2000–09).
Note: HIES = Household Income and Expenditure Survey (Bangladesh Bureau of Statistics); ENAHO = Encuesta Nacional de
Hogares (Peru National Institute of Statistics and Information); SES = Household Socio-Economic Survey.
a. “Spearman rank” refers to Spearman’s rank correlation coefficient (named after Charles Spearman), a nonparametric measure
of statistical dependence between two variables. The closer the correlation is to one, the more perfect the correlation is.
Since poverty is measured by consumption, actual poverty rates are higher in the
final period than they would have been had the consumption-to-income ratio
remained constant in Bangladesh and Thailand. In contrast, actual poverty will be
lower in the final period in Peru simply because poor households are consuming
a higher share of their income.
The Decomposition Approach
As described in chapter 4, we implement the decomposition in four stages:
1. We estimate the determinants of occupational choice, sectoral choice, and
educational level for two periods during the past decade.
2. We estimate the earnings regressions for each period for household heads
and other household members (distinguishing between wage workers, self-
employed, and daily workers) and for net farm revenue for farm households.
3. We use the coefficients from these regressions to simulate counterfactual
distributions by replacing one element at a time.
4. We compare these counterfactuals to the observed changes in distribution.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
142 Understanding Poverty Reduction in Bangladesh, Peru, and Thailand
Estimate Educational, Occupational, and Sectoral Choice Structures
As summarized in the first step above, we first estimate the educational, occupa-
tional, and sectoral choice models in each period for which data are available. In
annex 6A, tables 6A.1, 6A.2, and 6A.3 present simulations for the educational
structure, occupation, and economic sectors using these regressions compared
with the actual structures during the early and late part of the decade.
Also in the annex, tables 6A.4, 6A.5, and 6A.6 present the multinomial logit
regression results for occupational choice in Bangladesh, Peru, and Thailand,
respectively. Given the considerable diversification of income sources that is
common in rural households (Davis et al. 2010), we estimate the sectoral choice
model for the secondary occupation of individuals in farm households, as well as
for all nonfarm work.
The simulated structures are close to the true structures, with the errors being
relatively small in all three countries. In the case of Bangladesh, the simulated
2010 nonfarm occupational structure has slightly fewer wage workers and more
daily workers than the true values, but the regression results show a relatively
high coefficient of determination (R2). This gives us confidence that we can use
the coefficients from these regressions to simulate shifts in the labor force struc-
ture one at a time.
Estimate Earnings Equations
Next, we separate labor income into farm and nonfarm income to estimate the
earnings equations.
Nonfarm Household Income
Tables 6A.7, 6A.8, and 6A.9 present the results for individuals engaged in non-
farm activities in Bangladesh, Peru, and Thailand, respectively. The results show
that the models fit the data relatively well, with coefficients being statistically
significant and of the right sign. In all cases, higher individual earnings are associ-
ated with being male, having higher education and experience, living in urban
areas, and belonging to the manufacturing sector.
Farm Household Income
Tables 6A.10, 6A.11, and A6A.12 present results of net revenue for farm house-
holds in Bangladesh, Peru, and Thailand, respectively.4 Net revenue for farmers
increases with experience, land holdings, access to irrigation, and the number of
household members participating in farm work.5
Simulate Counterfactual Distributions
In the next step, we use the estimated coefficients from the educational, sectoral,
labor, and earnings models to simulate counterfactual distributions. For instance,
because we estimated the returns to education in two periods, we can take the
estimated parameters in the first period (say, 2000) and evaluate the earnings
equations with the 2010 levels of education. This generates counterfactual
earnings at the individual level, which we aggregate to get the corresponding
household income, then the corresponding level of consumption, and finally a
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Understanding Poverty Reduction in Bangladesh, Peru, and Thailand 143
counterfactual poverty rate. In this way—changing one parameter at a time or
one characteristic at a time—we can generate multiple counterfactual distribu-
tions and recalculate a new poverty rate.
Compare Counterfactuals with Distributional Changes
Finally, we compare these counterfactual poverty rates with the observed pov-
erty rates to quantify the impact of each element being considered. Because
replacing the first-period parameters into last-period data will yield results that
are different from doing it the other way around, we calculate the counterfactual
by doing it both ways and then take the average (in line with Bourguignon,
Ferreira, and Lustig 2005; Bourguignon, Ferreira, and Leite 2008).
We first calculate the effects on poverty by changing one element at a time
and leaving everything else constant. However, given that changes in multiple
factors could have interaction effects, we also calculate the cumulative effect of
these decompositions. For this purpose, we adhere to the literature (Bourguignon,
Ferreira, and Leite 2008) as follows:
• Begin by calculating the effects on poverty as a result of changes in the popula-
tion’s characteristics, beginning with the most exogenous: age, gender, and
region.
• Then calculate the effects on poverty as a result of changes in the population’s
educational, occupational, and sectoral structure, all of which depend on age,
gender, and region.
• Use the preceding results to calculate the effects on poverty as a result of
–– Changes in farm and nonfarm earnings due to changes in the returns to
these characteristics;
–– Changes in nonlabor incomes (first private transfers, then public ones); and
–– Changes in the consumption-to-income ratio.
Decomposition Results
Chapter 3 found that, in a set of 21 countries with large declines in poverty, most
of these declines could be attributed to increases in labor income. The results of
this chapter are in line with that result, while adding a modeling structure that
enables us to further conclude that the main contributor to increased labor
incomes—and thus to poverty reduction—was the improvement in the returns
to individual and household characteristics and endowments.
Endowments and the Returns on Endowments
That these increased returns contributed the most to poverty reduction, rather
than changes in the populations’ endowments or characteristics, points to an
increase in the relative price of labor and higher productivity as the main
contributors to poverty reduction in each case. In particular, returns to farm
and nonfarm endowments amount to 44 percent of poverty reduction in Peru,
more than 48 percent in Thailand, and 60 percent in Bangladesh (table 6.5). In
both Thailand and Bangladesh, this effect is concentrated in the farm sector,
whereas Peru’s increased returns were slightly larger in the nonfarm sector.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
144 Understanding Poverty Reduction in Bangladesh, Peru, and Thailand
The overall conclusion—that increased returns to endowments was the largest
factor in poverty reduction—holds true whether we (a) calculate one counter-
factual at a time, holding all else constant (as shown by the marginal contribu-
tions in tables 6.5 and 6.6) or (b) adopt the cumulative decomposition path
described (as shown in table 6.7). Although the cumulative results are slightly
different in magnitude, the overall messages are the same.
Table 6.5 shows the net effect of changes in returns. However, these net
effects reflect the sum of price changes that the labor market assigns to each
Table 6.5 Marginal Contributions to Poverty Reduction in Bangladesh, Peru, and Thailand, 2000s
Bangladesh, Peru, Thailand,
2000–10 2004–10 2000–09
Change Change Change Change Change Change
(ppts) (% of total) (ppts) (% of total) (ppts) (% of total)
Nonfarm labor income −4.56 26 −9.35 58 −3.46 27
Returns to characteristics −3.52 20 −4.93 31 −1.25 10
Occupational choice −1.61 9 −3.44 21 0.08 −1
Economic sector −0.48 3 −0.08 1 −1.01 8
Education −0.55 3 −0.25 2 −1.34 10
Unobservable factors 1.59 −9 −0.65 4 0.06 0
Farm income −6.02 35 −2.74 17 −4.91 38
Returns to characteristics −6.98 40 −2.04 13 −4.83 38
Occupational choice 0.56 −3 −0.25 2 1.31 −10
Economic sector — — −0.14 1 −1.11 9
Education 0.13 −1 −0.08 1 −0.56 4
Unobservable factors 0.26 −2 −0.23 1 0.28 −2
Nonlabor income 1.05 −6 −2.28 14 −5.80 45
Total private transfers 1.05 −6 −0.85 5 −3.30 26
International private transfers −1.94 11 0.19 −1 −2.19 17
Domestic private transfers 1.10 −6 0.24 −1 — —
Private donations — — −0.70 4 — —
Other private transfers 0.58 −3 0.01 0 −1.12 9
Capital 1.31 −8 −0.58 4 0.02 0
Total public transfers 0.00 0 −1.38 9 −2.51 20
Public donations — — −0.45 3 — —
Public transfers — — −0.90 6 — —
Pensions — — 0.01 0 −2.51 20
Other nonlabor income — — −0.04 0 — —
Other −7.50 43 −0.78 5 2.50 −20
Age-gender-regional structure −3.48 20 −1.17 7 −1.18 9
Consumption-to-income ratio 0.93 −5 −1.73 11 3.43 −27
Unexplained −4.95 29 2.12 −13 0.26 −2
Total −17.34 100 −16.13 100 −12.84 100
Sources: Calculations derived from household survey data from Bangladesh (HIES 2000–10), Peru (ENAHO 2004–10), and Thailand (SES 2000–09).
Note: ppts = percentage points; — = not available; HIES = Household Income and Expenditure Survey (Bangladesh Bureau of Statistics);
ENAHO = Encuesta Nacional de Hogares (Peru National Institute of Statistics and Information); SES = Household Socio-Economic Survey (Thailand
National Statistical Office).
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Table 6.6 Contributions to Poverty Reduction by Returns to Endowments in Bangladesh, Peru, and Thailand, 2000s
Bangladesh Peru Thailand
2000–10 2004–10 2000–09
Nonfarm households Farm households Nonfarm households Farm households Nonfarm households Farm households
Change Change Change Change Change Change Change Change Change Change Change Change
(ppts) (%) (ppts) (%) (ppts) (%) (ppts) (%) (ppts) (%) (ppts) (%)
Total parameter effects −3.52 20 −6.98 40 −4.93 31 −2.04 13 −1.25 10 −4.83 38
Education 1.77 −10 0.78 −4 0.54 −3 … … −0.28 2 −1.11 9
Age 3.9 −23 4.19 −24 1.0 −6 0.96 −6 −0.90 7 −0.54 4
Female −0.44 3 −0.20 1 0.77 −5 −0.08 1 −0.01 … −0.33 3
Urban 0.11 −1 0.45 −3 0.20 −1 0.12 −1 −0.40 3 −0.09 1
Region −0.87 5 −1.82 10 −4.79 30 −0.18 1 −0.27 2 −1.62 13
Sector 5.85 −34 0.21 −1 −1.05 7 −0.19 1 3.50 −27 0.32 −3
Land n.a. n.a. −7.25 42 n.a. n.a. −3.18 20 n.a. n.a. — —
Irrigation n.a. n.a. 0.30 −2 n.a. n.a. −0.11 1 n.a. n.a. — —
Other members n.a. n.a. 1.48 −9 n.a. n.a. 1.07 −7 n.a. n.a. — —
Constant −14.87 86 −7.74 45 −3.38 21 … … −2.71 21 −1.11 9
Sources: Calculations derived from household survey data from Bangladesh (HIES 2000–10), Peru (ENAHO 2004–10), and Thailand (SES 2000–09).
Note: ppts = percentage points; … = negligible; — = not available; n.a. = not applicable. “Region” refers to residence in provinces, states, or other regional subgroups. “Sector” refers to changes in the sectoral choice
of workers, including agriculture, manufacturing, industry, and services. “Other members” refers to the number of household members engaged in farm activities. HIES = Household Income and Expenditure Survey
(Bangladesh Bureau of Statistics). ENAHO = Encuesta Nacional de Hogares (Peru National Institute of Statistics and Information). SES = Household Socio-Economic Survey (Thailand National Statistical Office).
145
146 Understanding Poverty Reduction in Bangladesh, Peru, and Thailand
Table 6.7 Cumulative Contributions to Poverty Reduction in Bangladesh, Peru,
and Thailand, 2000s
Bangladesh, 2000–10 Peru, 2004–10 Thailand, 2000–09
Change Change Change Change Change Change
(ppts) (%) (ppts) (%) (ppts) (%)
Demographics −4.41 25 −0.88 5 −0.41 3
Education −0.88 5 −0.31 2 −1.74 14
Occupation −1.38 8 −3.90 24 0.52 −4
Sector −0.51 3 0.02 0 −1.75 14
Returns nonfarm −2.93 17 −5.06 31 −1.40 11
Returns farm −8.18 47 −1.95 12 −3.59 28
Residuals 1.33 −8 0.05 0 0.06 0
Nonlabor income −1.93 11 −0.13 1 −3.59 28
Others 1.55 −9 −3.96 25 −0.93 7
Total −17.34 100 −16.13 100 −12.84 100
Sources: Calculations derived from household survey data from Bangladesh (HIES 2000–10), Peru (ENAHO
2004–10), and Thailand (SES 2000–09).
Note: ppts = percentage points. “Demographics” refers to the exogenous structure of the population’s age,
gender, and regional makeup. For a detailed description of the methodology, refer to chapter 5.
HIES = Household Income and Expenditure Survey (Bangladesh Bureau of Statistics); ENAHO = Encuesta
Nacional de Hogares (Peru National Institute of Statistics and Information); SES = Household Socio-Economic
Survey (Thailand National Statistical Office).
characteristic or endowment. For example, the price the labor market assigns
for each additional level of education or for living in a particular region in the
country may be components of this net effect.
Table 6.6 breaks down these changes in returns. The results show that for self-
employed farmers, most of the increase in returns has to do with a large increase
in the constant, which measures average real farm wages holding everything else
constant. This increase could be the result of either higher productivity or higher
relative prices in agriculture—potentially resulting from food price increases.
As we have seen, poverty rates changed because of both changes in endow-
ments and changes in the returns to those endowments. However, these two
forces could act in opposite directions. For instance, if higher education raises
incomes, it also would tend to lower poverty. However, if the supply of educated
workers outpaces the demand for such workers, then the returns to, or premium
for, higher education will likely fall. This idea is often associated with Tinbergen’s
(1975) “race” between technological progress—which he saw as raising the
demand for skills—and the expansion of formal education—which raises
the supply of skills (Ferreira 2012). Below, we summarize the results regarding
the net impacts of changes in these opposing forces.
Impact of Demographic Changes
Changes in the structure of the populations’ age, gender, area (urban or rural),
and regional (district) accounted for about 25 percent of the observed poverty
reduction in Bangladesh, 5 percent of the reduction in Peru, and 3 percent of the
reduction in Thailand when holding all else constant (table 6.7).
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Understanding Poverty Reduction in Bangladesh, Peru, and Thailand 147
These are important contributions that are consistent with the results described
in chapter 3, which follow an alternative methodology. Given the increase in
the economically active population, the size of the effect is not surprising—
particularly in Bangladesh, where fertility rates have declined while a large genera-
tion of young people has joined the work force. However, the entrance of young
people has coincided with a decline in the premium for experience (as proxied by
age), particularly for daily and self-employed workers in Bangladesh. The decline
in the returns to experience (shown as a –23 percent decline in the age parameter
for Bangladesh in table 6.6) has counteracted the progress in poverty reduction
(shown as a combined positive effect of 25 percent for Bangladesh in table 6.7).
In contrast, in Thailand, both the increase in the working-age population
and the greater returns to age (interpreted as experience) contributed to pov-
erty reduction. In particular, demographics contributed 3 percent to poverty
reduction (table 6.7), while greater returns to experience in the farm and non-
farm households contributed 4 percent and 7 percent, respectively (table 6.6).
A consistent result in all three countries is that the earnings penalty for living
outside of the capital city fell over the past decade. This finding points to either
an increase in the relative price of labor, higher productivity outside the capitals
that helped to reduce poverty, or both. This result is most evident in Peru,
where the reduced penalty for living outside of Lima (which table 6.6 reflects
through the impact of residency in the “regions”) accounted for 30 percent of
the reduction in poverty for nonfarm households.6 In Bangladesh, the penalty
for living outside of the capital also declined, accounting for 10 percent and
5 percent of poverty reduction for farm and nonfarm households, respectively.
Thailand saw a 13 percent and 2 percent in poverty reduction among farm and
nonfarm households, respectively, even though the share of people living out-
side the capital remained more or less constant.
Impact of Changes in Labor Income
Education
As expected, a more-educated population helped to reduce poverty in all three
countries, particularly in the nonfarm sector (table 6.5). However, this effect was
modest—being more than offset by a decline in the returns to education in
Bangladesh and Peru (table 6.6). In Thailand, in contrast, wage premiums for
education increased in farm and nonfarm households, further reducing poverty.
The results of greater education and of returns to education were as follows
in the three countries:
• In Bangladesh, the large increase in the share of the population who had com-
pleted primary school (from 33.8 percent in 2000 to 43.4 percent in 2010, as
previously shown in table 6.3) led to only a slight (3 percent) reduction in
poverty in the nonfarm sector (table 6.5) because it drove down the premium
for completing primary school for all but wage workers.
• In Peru, similarly, a more-educated population accounted for 2 percent of the
decline in poverty in the nonfarm sector (table 6.5). However, this effect was
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
148 Understanding Poverty Reduction in Bangladesh, Peru, and Thailand
more than offset by a decline in the returns to education (table 6.6), because
the greater supply of more-educated household heads was not met with
greater demand for these workers; therefore, the wage premium for education
fell. As a result, the increase in education reduced poverty less than expected.
• In Thailand, a more-educated population accounted for 10 percent of the
reduction in poverty in nonfarm households, and a 4 percent reduction in farm
households (table 6.5). In contrast to Peru and Bangladesh, increased returns
to education reinforced these gains in poverty reduction, as wage premiums
for education and experience increased in both farm and nonfarm households
(table 6.6). These results point to greater demand for specialized workers in
Thailand than in Bangladesh and Peru.
Collectively, these results point to a relative increase in the price of labor for
uneducated workers, an increase that strongly drove poverty reduction in
Bangladesh and Peru. Capturing this effect is the large contribution to poverty
reduction from the left-out category—the constant—which includes the returns
to labor for individuals with no schooling (table 6.6). In Bangladesh, this was true
in both the farm and nonfarm sectors for daily and self-employed workers, whose
additional years of education did not help to reduce poverty (see the table 6A.7).
However, the opposite is true for wage workers, whose returns to education
increased. This seeming contrast implies that the population’s educational levels
grew faster than the rate at which job creation could absorb them. The different
job-related returns to education also reflect a sizable increase in the relative price
of unskilled labor, which could be driven, at least partly, by higher productivity.
Occupation
Changes in occupational choice were critical to poverty reduction in all three
countries among nonfarm workers, who aimed to benefit from better work
opportunities. This effect was most important in Peru, where occupational
shifts—from being unpaid family workers to being wage workers—accounts for
21 percent of poverty reduction (table 6.5). In the nonfarm sector in Bangladesh,
the shift from daily and self-employed work toward wage employment accounted
for almost 10 percent of the observed poverty reduction.7
In contrast, for workers who remained in agriculture, we find greater special-
ization on the part of farmers. By the end of the decade, they were less likely to
diversify into a secondary occupation, either because they saw increased returns
from farm activities or because they lacked the skills to take up a second occu-
pation. For example, the share of Thai farmers with a secondary occupation
declined from 32 percent in 2000 to 23 percent in 2009. Meanwhile, in
Bangladesh, the drop was even more dramatic: the share of farmers with a
secondary occupation fell from 30 percent in 2000 to 10 percent in 2010.
This lower diversification increased poverty by 10 percent and 3 percent in
Thailand and Bangladesh, respectively (see the “Occupational choice” row
under “Farm income” in table 6.5).8
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Understanding Poverty Reduction in Bangladesh, Peru, and Thailand 149
Sector of Work
Changes in work sector also affected poverty reduction, particularly shifts away
from agriculture and into services. However, in both Bangladesh and Thailand,
any related poverty reduction was more than offset by reduced returns to work-
ing in the service sector, whereas Peru showed the opposite dynamic:
• In Bangladesh, the nonfarm sector’s shift into the services sector accounted for
3 percent of the observed poverty reduction (table 6.5). However, this was
more than offset by a reduction in the service sector wage premium in non-
farm households, leading to a 34 percent higher poverty rate than would have
otherwise been expected (table 6.6).
• In Thailand, similarly, many workers’ movement into manufacturing and ser-
vices accounted for a 9 percent poverty decrease among farm households and
an 8 percent decrease among nonfarm households (table 6.5). However, these
effects were countered by a decline in the returns to working in those sectors
in farm and nonfarm households (table 6A.9), which led to a 3 percent and
27 percent higher poverty rate, respectively, than would have otherwise been
expected (table 6.6).
• In Peru, in contrast, despite the increasing share of workers in the service
sector, there were increases in the returns to working in the service sector
(see table 6A.8) that accounted for 7 percent of the reduction in poverty
(table 6.6). This is astonishing, given that Peru has a much larger share of
service sector workers than either of the other countries (see table 6.3).
Rural Assets
Among farm households in Bangladesh and Peru, we find some evidence of
increased returns to agriculture. In particular, for farm households in Bangladesh,
the most important change was the increase in the returns to land, accounting
for 42 percent of the reduction in national poverty when holding all other factors
constant (table 6.6). These returns increased because average land per capita
declined from 0.8 acres to 0.6 acres between 2000 and 2010 given the popula-
tion increase.
In contrast, in the case of Peru, the average land size for farm households
increased while the returns to land increased as well (table 6.6), a boon that
accounted for 20 percent of the reduction in poverty. This was complemented
by better access to irrigation in Peru, accounting for another 1 percent of the
poverty reduction. However, the returns to additional agricultural workers
declined. In Bangladesh, although both access to irrigation and the number of
agricultural workers increased over the course of the decade, the returns to
irrigation and having additional household members employed in farming fell, so
neither effect helped to reduce poverty.9
In considering the increase to rural assets, it is important to note that we
cannot disentangle the effects as a result of increased real productivity (real out-
put per worker) from the increase in the real value of output per worker. Given
that this period was characterized by an increase in the prices of commodities,
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
150 Understanding Poverty Reduction in Bangladesh, Peru, and Thailand
this factor might have been an important driver of agricultural returns, through
its effect on the real value of agricultural production.
Impact of Changes in Nonlabor Income
Although poverty was reduced primarily because of labor income growth,
increases in nonlabor income also played a role that varied by country:
• In Bangladesh, the increase in international remittances contributed 11 p ercent
to the decline in poverty. Offsetting this effect, however, was a decline in
domestic transfers that led to a slightly higher poverty rate than if both remit-
tance sources had remained constant (table 6.5).
• In Peru, public transfers and donations accounted for 9 percent of the reduc-
tion in poverty (table 6.5), while capital and private donations accounted for
nearly 5 percent of the reduction in poverty (table 6.5).
• In Thailand, nonlabor income was important, particularly through private
transfers in the form of international remittances and other private transfers
(17 percent and 9 percent, respectively) and pensions (20 percent), poten-
tially reflecting the introduction of a new pension scheme in the late 1990s
(table 6.5). These results are consistent with those using the simple approach,
but they have been further disaggregated to show the relative importance of
the different types of sources.
Final Remarks
The past decade affords us a fantastic opportunity to study the most significant
factors that worked in favor of the poor. This chapter accounts for the contribu-
tion of changes in demographics, labor income, and nonlabor income in the
significant poverty reductions observed in Bangladesh, Peru, and Thailand dur-
ing the 2000s. In contrast to methods that focus on aggregate summary statis-
tics, the methods adopted here generate entire counterfactual distributions,
allowing us to identify more precisely the contributions to the observed distri-
butional changes.
Link of Labor Income to Marginal Value of Work
The results show that labor income growth has been the most important con-
tributor to poverty reduction over the past decade. Further—through the micro-
decomposition methods employed here—we ascertained that the growth in
labor income was mainly the result of higher returns to endowments, signaling
an increase in the marginal value of work, resulting from increases in either pro-
ductivity or relative prices of labor.
In Bangladesh and Peru, this increase in the marginal value of work was not
driven by higher returns to education, but rather by higher returns to unskilled
labor. Thailand, in contrast, demonstrated that greater specialization and higher
returns to human capital can boost the marginal value of work, potentially through
productivity increases. All three countries showed these results consistently:
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Understanding Poverty Reduction in Bangladesh, Peru, and Thailand 151
• A falling earnings penalty for living outside of the capital city served to reduce
poverty over the past decade—accounting for as much of 31 percent of the
reduction in Peru, 15 percent in Bangladesh, and 15 percent in Thailand.
• A shift in the sectoral composition of the work force—away from agriculture and
toward services—reduced poverty only slightly in Bangladesh and Peru and
a bit more in Thailand. However, this increase in service sector workers led
to a declining premium for working in the service sector in Bangladesh and
Thailand. In Peru, however, the higher share of service workers was accom-
panied by increased premiums even though Peru’s service sector makes up a
far larger share of total employment than it does in the other countries.
• Occupational choices among nonfarm workers shifted away from daily-wage
and self-employed work and toward wage jobs, all of which contributed to
poverty reduction, particularly in Peru.
Roles of Demographic Change and Nonlabor Income Growth
Beyond the effects of labor income growth, the decomposition method adopted
in this chapter (as well as the method adopted in chapter 3) showed that a greater
share of working-age adults helped to reduce poverty, particularly in Bangladesh.
Finally, although most of the reduction in poverty was the result of labor
income growth, it is important to recognize that nonlabor income in the form of
transfers did play a role: International remittances accounted for 11 percent and
17 percent of poverty reduction in Bangladesh and Thailand, respectively. Public
transfers accounted for about 9 percent of Peru’s poverty reduction, particularly
for those at the very bottom of the income distribution. And Thailand’s generous
new pension scheme, combined with various private and other transfers, accounts
for more than one-third of its poverty reduction.
An Agenda for Further Research
This volume has proposed two distinct micro-decomposition approaches to
understanding changes in poverty over the last decade. The proposed approaches
are complementary—the first being quite simple and easy to apply across coun-
tries with relatively limited data requirements, and the second providing a more
in-depth analysis based on a structural approach.
The approach implemented in this chapter allows for movement across
sectors and occupations for nonfarm households. In addition, it allows individuals
in farm households to choose a secondary occupation and sector. However, the
proposed structure does not model the choice of a household to be either a farm
or a nonfarm household. Further research could explore the extent to which this
could be incorporated into the decomposition analysis.
In addition, to the extent that demographic changes are important, further
work could include a fertility choice model, particularly in settings where there
are substantial shifts over time. Finally, as described in chapter 4, other decom-
position techniques could be explored and perhaps combined with the two
approaches highlighted in this volume.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
152
Annex 6A: Regression and Simulation Results
Table 6A.1 Simulating the Changing Characteristics of Households in Bangladesh, 2000–10
Household head Other members
2000 2010 2000 2010
Actual Simulated Actual Simulated Actual Simulated Actual Simulated
(1) (2) (2) – (1) (1) (2) (2) – (1) (1) (2) (2) – (1) (1) (2) (2) – (1)
(%) (%) (ppts) (%) (%) (ppts) (%) (%) (ppts) (%) (%) (ppts)
Education structure
Illiterate and incomplete primary 63.2 61.9 −1.3 57.0 57.8 0.8 54.4 54.5 0.1 42.1 42.3 0.2
Primary and low secondary 29.0 29.6 0.6 33.1 33.2 0.1 36.2 36.0 −0.3 48.6 48.2 −0.4
Complete tertiary 7.7 8.5 0.8 9.9 9.0 −0.9 9.5 9.6 0.1 9.3 9.4 0.2
Occupation
Nonemployed 9.6 10.1 0.5 8.0 7.7 −0.4 79.8 83.0 3.2 78.0 80.1 2.2
Daily workers 45.5 40.7 −4.9 43.4 49.9 6.5 8.6 4.7 −3.9 8.2 5.6 −2.7
Self-employed—nonagriculture 24.7 25.1 0.3 24.1 23.9 −0.2 4.4 4.8 0.3 4.1 4.4 0.3
Wage workers 20.2 24.2 4.0 24.5 18.6 −5.9 7.2 7.5 0.3 9.7 9.9 0.2
Economic sectors
Daily workers
Agriculture 58.8 56.4 −2.4 51.8 51.8 0.0 54.8 53.4 −1.4 42.7 44.1 1.4
Manufacturing 12.3 12.8 0.5 12.6 13.1 0.6 19.0 18.3 −0.8 21.5 20.4 −1.1
Industry 7.9 8.5 0.6 11.9 12.3 0.4 8.4 8.9 0.6 15.8 16.4 0.5
Services 21.0 22.4 1.4 23.7 22.8 −0.9 17.8 19.4 1.6 20.0 19.2 −0.8
Self-employed
Manufacturing 34.9 35.3 0.4 17.6 17.0 −0.6 41.1 45.7 4.7 26.8 22.1 −4.8
Industry 6.2 7.8 1.6 1.7 1.9 0.1 7.1 6.9 −0.2 3.1 2.6 −0.5
Services 58.9 57.0 −2.0 80.7 81.1 0.4 51.8 47.4 −4.5 70.1 75.4 5.3
Wage workers
Agriculture 8.6 7.3 −1.3 5.4 6.9 1.4 4.4 3.4 −1.0 3.0 3.7 0.8
Manufacturing 27.7 26.8 −0.9 33.4 34.3 0.9 42.6 42.4 −0.2 47.3 47.5 0.3
Industry 3.7 3.8 0.1 3.9 4.0 0.1 2.5 2.0 −0.5 2.6 2.7 0.1
Services 60.0 62.2 2.1 57.3 54.9 −2.4 50.5 52.2 1.7 47.2 46.0 −1.2
Source: Household survey results from Bangladesh (HIES 2000–10).
Note: ppts = percentage points. HIES = Household Income and Expenditure Survey (Bangladesh Bureau of Statistics).
Table 6A.2 Simulating the Changing Characteristics of Households in Peru, 2005–09
Household head Other members
2004 2010 2004 2010
Actual Simulated Actual Simulated Actual Simulated Actual Simulated
(1) (2) (2) – (1) (1) (2) (2) – (1) (1) (2) (2) – (1) (1) (2) (2) – (1)
(%) (%) (ppts) (%) (%) (ppts) (%) (%) (ppts) (%) (%) (ppts)
Education type
Less than primary 55.3 55.4 0.1 49.7 49.9 0.3 50.3 50.8 0.6 45.4 44.6 −0.8
Primary 28.2 28.1 −0.1 32.4 32.3 −0.1 36.5 35.6 −0.8 39.7 40.7 1.0
Secondary and above 16.5 16.5 0.0 18.0 17.8 −0.2 13.3 13.6 0.3 15.0 14.7 −0.2
Occupation
Nonemployed 16.8 17.3 0.5 13.1 12.8 −0.3 57.9 58.7 0.9 49.5 50.4 0.9
Wage workers 48.2 48.1 0.0 50.2 50.0 −0.2 26.2 25.7 −0.5 32.2 31.6 −0.6
Self-employed (nonagriculture) 35.0 34.5 −0.5 36.7 37.2 0.5 15.9 15.6 −0.4 18.3 18.0 −0.3
Economic sectors
Wage workers
Agriculture 31.8 29.9 0.0 25.9 27.6 0.1 17.0 17.7 −0.5 14.8 13.9 0.7
Industry 12.1 12.1 1.6 12.0 12.1 −1.0 13.2 12.8 −1.4 13.2 13.9 1.6
Services 35.2 36.8 0.2 40.7 39.7 −1.0 52.8 51.4 1.2 56.7 58.3 −1.4
Public sector 20.9 21.2 0.0 21.5 20.5 0.0 16.9 18.1 0.0 15.3 14.0 0.0
Source: Household survey results from Peru (ENAHO 2004–10).
Note: ppts = percentage points. ENAHO = Encuesta Nacional de Hogares (Peru National Institute of Statistics and Information).
153
154
Table 6A.3 Simulating the Changing Characteristics of Households in Thailand, 2000–09
Household head Other members
2000 2009 2000 2009
Actual Simulated Actual Simulated Actual Simulated Actual Simulated
(1) (2) (2) – (1) (1) (2) (2) – (1) (1) (2) (2) – (1) (1) (2) (2) – (1)
(%) (%) (ppts) (%) (%) (ppts) (%) (%) (ppts) (%) (%) (ppts)
Education (years)
<6 59.7 64.8 5.1 46.9 40.8 −6.0 36.6 42.3 5.7 27.3 22.1 −5.2
6–11 23.2 19.3 −3.8 29.3 34.2 4.9 37.1 32.9 −4.2 44.4 49.0 4.6
>11 17.1 15.9 −1.3 23.8 25.0 1.2 26.3 24.8 −1.5 28.3 28.9 0.7
Occupation
Nonemployed 20.4 21.5 1.1 20.7 19.4 −1.2 49.4 47.0 −2.5 47.3 50.6 3.3
Salaried 57.3 56.4 −0.9 53.9 55.7 1.8 41.0 42.8 1.8 41.9 39.5 −2.4
Self-employed (nonagriculture) 22.3 22.1 −0.2 25.4 24.8 −0.6 9.5 10.3 0.7 10.8 9.9 −0.9
Economic sectors
Self-employed
Industry 15.6 14.4 −1.2 17.1 16.7 −0.4 20.0 18.4 −1.6 18.7 16.1 −2.6
Services 84.4 85.6 1.2 82.9 83.3 0.4 80.0 81.7 1.6 81.3 83.9 2.6
Wage workers
Agriculture 20.6 16.5 −4.0 11.6 15.1 3.5 20.7 18.9 −1.8 9.1 11.0 1.9
Industry 30.7 27.1 −3.6 35.3 37.3 1.9 36.0 32.2 −3.8 38.8 41.2 2.4
Services 23.6 22.8 −0.7 29.3 30.5 1.2 27.2 25.6 −1.5 34.7 35.9 1.2
Public sector 25.1 33.5 8.4 23.8 17.2 −6.6 16.1 23.3 7.2 17.4 11.9 −5.4
Source: Household survey results from Thailand (SES 2000–09).
Note: ppts = percentage points. SES = Household Socio-Economic Survey (Thailand National Statistical Office).
Understanding Poverty Reduction in Bangladesh, Peru, and Thailand 155
Table 6A.4 Multinomial Logit on Occupational Choice of Working-Age Population, by Household Status,
in Bangladesh, 2000 and 2010
2000 2010
Daily workers Self-employeda Wage worker Daily workers Self-employeda Wage worker
a. Household heads
Primary & lower
secondary −1.204*** −0.0182 0.614*** −1.082*** 0.122 0.682***
education (0.160) (0.161) (0.166) (0.145) (0.147) (0.147)
Higher secondary −2.408*** 0.301 2.002*** −3.461*** −0.599*** 1.227***
& tertiary education (0.368) (0.294) (0.287) (0.267) (0.203) (0.194)
Age 0.0408 0.126** 0.114** 0.109** 0.281*** 0.238***
(0.0510) (0.0544) (0.0556) (0.0423) (0.0449) (0.0440)
Age squared −0.00141** −0.00228*** −0.00208*** −0.00211*** −0.00382*** −0.00360***
(0.000572) (0.000611) (0.000627) (0.000473) (0.000504) (0.000496)
Urban −0.354** 0.421*** 0.795*** −1.150*** −0.411*** −0.00785
(0.160) (0.161) (0.162) (0.134) (0.135) (0.133)
Barisal 0.0892 −0.0402 −0.313 0.101 −0.330 −0.310
(0.263) (0.275) (0.288) (0.255) (0.262) (0.258)
Chittagong −0.490*** −0.310* −0.216 0.228 −0.525*** −0.264
(0.179) (0.184) (0.186) (0.161) (0.168) (0.161)
Khulna 0.432* 0.273 0.0118 1.127*** 0.352 0.128
(0.231) (0.240) (0.248) (0.217) (0.224) (0.223)
Rajsahi 1.735*** 1.400*** 0.834*** 1.641*** 1.141*** 0.390**
(0.227) (0.234) (0.242) (0.190) (0.194) (0.197)
Sylhet 0.541* −0.0682 0.126 0.564** 0.467* −0.0835
(0.281) (0.319) (0.323) (0.262) (0.270) (0.276)
Attends school −3.549*** −3.393*** −2.742*** −0.263 −14.93 −1.508
(0.753) (0.866) (0.724) (1.439) (565.3) (1.281)
Remittances −0.646*** −1.014*** −0.715*** −0.129 −0.00592 −0.178
(0.172) (0.178) (0.181) (0.214) (0.215) (0.218)
Female −2.213*** −2.993*** −1.178*** −1.903*** −2.802*** −1.019***
(0.409) (0.489) (0.444) (0.376) (0.444) (0.379)
Remittances × Female −0.328 −0.985* −1.450*** −1.534*** −1.583*** −2.358***
(0.321) (0.595) (0.431) (0.338) (0.506) (0.380)
Married 0.558 0.683* 0.593 1.295*** 1.471*** 1.235***
(0.367) (0.384) (0.389) (0.355) (0.375) (0.355)
Married × Female −2.269*** −1.432** −1.876*** −2.703*** −1.877*** −1.897***
(0.496) (0.651) (0.569) (0.467) (0.590) (0.464)
Other member −0.521*** −0.385*** −0.513*** −0.395*** −0.492*** −0.363***
employed (0.0700) (0.0717) (0.0762) (0.0730) (0.0760) (0.0736)
Number of children 0.246*** 0.267*** 0.213*** 0.114 0.0960 −0.115
(0.0484) (0.0503) (0.0521) (0.0697) (0.0713) (0.0722)
Constant 3.198*** 0.0640 −0.443 1.710* −3.336*** −2.206**
(1.030) (1.094) (1.113) (0.915) (0.976) (0.943)
Observations 4,974 4,974 4,974 7,862 7,862 7,862
Pseudo R2 0.233 0.233 0.233 0.259 0.259 0.259
table continues next page
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
156 Understanding Poverty Reduction in Bangladesh, Peru, and Thailand
Table 6A.4 Multinomial Logit on Occupational Choice of Working-Age Population, by Household Status,
in Bangladesh, 2000 and 2010 (continued)
2000 2010
Daily workers Self-employeda Wage worker Daily workers Self-employeda Wage worker
b. Other household members
Primary & lower
secondary −0.964*** −0.0334 0.259*** −0.802*** 0.321*** 0.240***
education (0.0977) (0.117) (0.0965) (0.0739) (0.0986) (0.0704)
Higher secondary −2.894*** 0.0292 1.227*** −2.253*** 0.114 1.460***
& tertiary education (0.358) (0.197) (0.155) (0.264) (0.171) (0.111)
Age 0.180*** 0.231*** 0.190*** 0.154*** 0.336*** 0.183***
(0.0239) (0.0311) (0.0255) (0.0207) (0.0281) (0.0194)
Age squared −0.00286*** −0.00345*** −0.00308*** −0.00243*** −0.00459*** −0.00303***
(0.000345) (0.000454) (0.000381) (0.000296) (0.000407) (0.000285)
Urban −0.172 0.623*** 0.781*** −0.134 0.358*** 0.966***
(0.108) (0.112) (0.0884) (0.0826) (0.0907) (0.0617)
Barisal 0.219 −0.117 −0.284 0.0287 −0.406** −0.677***
(0.179) (0.206) (0.174) (0.157) (0.183) (0.136)
Chittagong 0.107 −0.0791 0.0725 0.189* −0.704*** −0.429***
(0.119) (0.136) (0.103) (0.100) (0.124) (0.0766)
Khulna 0.504*** 0.155 −0.540*** 0.902*** 0.219* −0.506***
(0.141) (0.166) (0.155) (0.109) (0.128) (0.0996)
Rajsahi 1.508*** 0.433*** −0.141 0.976*** 0.0781 −0.748***
(0.109) (0.142) (0.121) (0.0901) (0.108) (0.0847)
Sylhet 0.414** 0.0886 0.193 0.791*** 0.0883 −0.629***
(0.171) (0.222) (0.168) (0.130) (0.163) (0.132)
Attends school −3.253*** −3.346*** −3.866*** −3.964*** −3.359*** −3.996***
(0.281) (0.287) (0.228) (0.251) (0.239) (0.151)
Remittances −0.319** −0.440*** −0.367*** −0.516*** −0.295** −0.394***
(0.127) (0.139) (0.125) (0.110) (0.123) (0.106)
Female −3.038*** −3.288*** −1.753*** −3.197*** −3.580*** −1.563***
(0.161) (0.269) (0.136) (0.146) (0.256) (0.0973)
Remittances × Female 0.143 −0.0907 0.0784 0.0908 0.140 −0.352**
(0.201) (0.329) (0.189) (0.202) (0.251) (0.166)
Married 0.822*** 1.240*** 0.932*** 0.817*** 0.794*** 0.639***
(0.148) (0.159) (0.149) (0.116) (0.128) (0.114)
Married × Female −2.151*** −2.283*** −2.376*** −1.856*** −1.477*** −2.115***
(0.202) (0.309) (0.189) (0.176) (0.276) (0.138)
Employed Hd head −0.413*** −0.412*** −0.331*** 0.170 −0.0691 0.0817
(0.121) (0.135) (0.112) (0.123) (0.135) (0.0982)
Household head
with primary &
lower secondary −0.952*** −0.244** −0.379*** 1.414*** 0.171 0.213**
education (0.104) (0.115) (0.0948) (0.240) (0.160) (0.106)
Household head with
higher secondary & −1.844*** −0.492** −0.465*** 0.630*** 0.0284 −0.0465
tertiary education (0.314) (0.204) (0.156) (0.244) (0.156) (0.100)
Constant −1.512*** −3.638*** −2.838*** −3.080*** −5.940*** −2.808***
(0.364) (0.475) (0.387) (0.414) (0.482) (0.329)
Observations 14,058 14,058 14,058 21,715 21,715 21,715
Pseudo R2 0.362 0.362 0.362 0.359 0.359 0.359
Source: Household survey results from Bangladesh (HIES 2000–10).
Note: “Working-age” = 15–64 years; Hd = household. Dhaka is the base region. “Nonemployed” is the base category.
Remittances × Female = interaction effect between remittances and female. Married × Female = interaction effect between married and female.
Standard errors in parentheses. HIES = Household Income and Expenditure Survey (Bangladesh Bureau of Statistics).
a. Nonagricultural self-employment.
*p < 0.1, **p < 0.05, ***p < 0.01
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Table 6A.5 Multinomial Logit on Occupational Choice of Working-Age Population, by Household Status, in Peru, 2004 and 2010
2004 2010
Head Spouse Other Head Spouse Other
Salaried Self- Salaried Self- Salaried Self- Salaried Self- Salaried Self- Salaried Self-
job employed job employed job employed job employed job employed job employed
At least high school 0.609*** 0.416*** 0.809*** 0.527***
education (0.0476) (0.0690) (0.0510) (0.0748)
Head or spouse with 0.395*** −0.0337 0.547*** 0.0410 −0.323*** −0.572*** 0.185* −0.128 0.675*** 0.119 −0.324*** −0.384***
at least high school (0.0882) (0.0881) (0.0940) (0.0685) (0.0517) (0.0730) (0.102) (0.103) (0.0917) (0.0763) (0.0562) (0.0787)
education
Age (years)
26–35 0.737*** 1.052*** 0.709*** 0.693*** 0.777*** 1.305*** 0.401* 0.844*** 0.439*** 0.856*** 0.863*** 1.181***
(0.187) (0.205) (0.165) (0.125) (0.0567) (0.0774) (0.233) (0.257) (0.161) (0.154) (0.0648) (0.0866)
36–45 0.673*** 1.208*** 0.859*** 0.746*** 0.564*** 1.312*** 0.286 0.894*** 0.570*** 0.946*** 0.796*** 1.425***
(0.176) (0.194) (0.172) (0.128) (0.0941) (0.113) (0.224) (0.246) (0.165) (0.154) (0.0997) (0.122)
46–55 0.127 0.778*** 0.192 0.179 0.372*** 1.211*** −0.189 0.603** 0.296* 0.757*** 0.316** 1.458***
(0.179) (0.196) (0.190) (0.139) (0.144) (0.167) (0.229) (0.250) (0.177) (0.162) (0.154) (0.163)
56–65 −1.287*** −0.128 −0.479* −0.200 −0.790*** 0.455** −1.307*** −0.204 −0.727*** 0.260 −1.248*** 0.911***
(0.182) (0.198) (0.272) (0.153) (0.215) (0.227) (0.232) (0.250) (0.210) (0.173) (0.274) (0.222)
At least one family 0.0356 1.314*** −1.118*** 0.689*** −0.261*** 0.218** −0.0482 1.003*** −0.844*** 0.811*** −0.202*** 0.221**
worker (0.239) (0.208) (0.234) (0.0909) (0.0633) (0.0938) (0.330) (0.314) (0.188) (0.0940) (0.0668) (0.0984)
Farm household −0.315** −0.900*** −0.829*** −0.480*** −0.317*** −0.568*** −0.162 −0.573*** −0.927*** −0.591*** −0.244*** −0.402***
(0.150) (0.170) (0.134) (0.103) (0.0711) (0.104) (0.171) (0.185) (0.143) (0.107) (0.0771) (0.116)
Female −1.640*** −0.949*** −2.191*** −0.974*** −0.712*** −0.643*** −1.412*** −1.193*** −1.975*** −0.736*** −0.685*** −0.623***
(0.129) (0.127) (0.231) (0.245) (0.0448) (0.0630) (0.152) (0.155) (0.224) (0.234) (0.0477) (0.0661)
Attends school −0.773*** −1.019*** 0.995*** −0.00837 −1.110*** −0.803*** 0.0414 −0.673** 0.714** −0.461 −1.163*** −0.917***
(0.206) (0.252) (0.298) (0.340) (0.0600) (0.0840) (0.275) (0.301) (0.279) (0.350) (0.0578) (0.0842)
Married −0.141 −0.0760 0.00307 0.267*** −0.185 −0.236 −0.109 0.138
(0.128) (0.128) (0.0608) (0.0833) (0.155) (0.158) (0.0689) (0.0866)
table continues next page
157
158
Table 6A.5 Multinomial Logit on Occupational Choice of Working-Age Population, by Household Status, in Peru, 2004 and 2010 (continued)
2004 2010
Head Spouse Other Head Spouse Other
Salaried Self- Salaried Self- Salaried Self- Salaried Self- Salaried Self- Salaried Self-
job employed job employed job employed job employed job employed job employed
Head employed −0.00833 −0.431*** −0.0719 0.0156 0.203 −0.212* −0.0477 0.117
(0.155) (0.106) (0.0586) (0.0775) (0.163) (0.125) (0.0706) (0.0889)
Spouse employed 0.0886* 0.0435 0.0948* 0.0138
(0.0496) (0.0688) (0.0513) (0.0733)
Costa −0.00486 0.0614 −0.153 0.307*** −0.0886 0.201** −0.0765 0.121 0.0767 0.367*** −0.366*** −0.0823
(0.103) (0.103) (0.107) (0.0891) (0.0586) (0.0808) (0.118) (0.119) (0.108) (0.0990) (0.0671) (0.0904)
Sierra −0.0484 0.0569 −0.150 0.340*** −0.418*** −0.0714 −0.0881 0.256** −0.0680 0.289*** −0.477*** −0.275***
(0.109) (0.110) (0.113) (0.0921) (0.0625) (0.0861) (0.125) (0.127) (0.112) (0.100) (0.0715) (0.0944)
Selva −0.190 0.105 −0.0935 0.200** −0.388*** −0.0687 −0.0160 0.304** 0.0924 0.383*** −0.457*** −0.0230
(0.128) (0.125) (0.120) (0.101) (0.0715) (0.0987) (0.141) (0.143) (0.117) (0.105) (0.0763) (0.0987)
Rural 0.314*** −0.414*** −0.457*** −0.674*** −0.0716 −0.570*** −0.00734 −0.609*** −0.308** −0.737*** −0.0120 −0.543***
(0.113) (0.121) (0.129) (0.0973) (0.0716) (0.103) (0.148) (0.152) (0.139) (0.106) (0.0774) (0.118)
Child ≤5 0.0589 0.176* −0.300*** −0.349*** −0.0138 −0.0140 −0.313*** −0.179***
(0.0971) (0.102) (0.0670) (0.0496) (0.0861) (0.0898) (0.0621) (0.0418)
Constant 1.349*** 0.477** 0.594* 0.289 −0.0442 −1.454*** 2.166*** 1.253*** 0.687** −0.165 0.315*** −1.089***
(0.204) (0.220) (0.329) (0.291) (0.0809) (0.114) (0.273) (0.288) (0.306) (0.299) (0.100) (0.131)
Observations 10,087 10,087 12,004 12,004 21,720 21,720 8,312 8,312 9,327 9,327 16,824 16,824
Pseudo R2 0.0879 0.0879 0.0914 0.0914 0.118 0.118 0.0599 0.0599 0.0950 0.0950 0.136 0.136
Source: Household survey results from Peru (ENAHO 2004–10).
Note: “Working age” = 15–64 years. Standard errors in parentheses. “Other” = household members who are neither household heads nor spouses. ENAHO = Encuesta Nacional de Hogares (Peru National Institute of
Statistics and Information).
*p < 0.1, **p < 0.05, ***p < 0.01
Table 6A.6 Multinomial Logit on Occupational Choice of Working-Age Population, by Household Status, in Thailand, 2000 and 2009
2000 2009
Head Spouse Other Head Spouse Other
Wage Self- Wage Self- Wage Self- Wage Self- Wage Self- Wage Self-
worker employed worker employed worker employed worker employed worker employed worker employed
6–11 yrs educ. 0.452*** 0.541*** 0.655*** 0.581***
(0.0846) (0.133) (0.0775) (0.107)
>11 yrs educ. 0.610*** 0.443*** 1.336*** 0.887***
(0.0881) (0.142) (0.0796) (0.114)
Head of household education (years)
6–11 0.128 0.0573 −0.157* 0.321*** −0.121 −0.141 0.194** 0.0547 0.0444 0.313*** −0.0690 −0.145
(0.123) (0.133) (0.0898) (0.123) (0.0876) (0.152) (0.0819) (0.0855) (0.0716) (0.0867) (0.0624) (0.120)
>11 0.797*** −0.256* −0.422*** 0.309** 0.173* −0.224 0.616*** −0.221** 0.0599 0.434*** −0.0296 −0.190
(0.121) (0.140) (0.116) (0.153) (0.103) (0.210) (0.0842) (0.0886) (0.0839) (0.103) (0.0681) (0.125)
Spouse education
<6 0.316** 0.581*** −0.252*** −0.285*** 0.540*** 0.581*** −0.258*** −0.169*
(0.157) (0.172) (0.0659) (0.109) (0.120) (0.123) (0.0563) (0.0914)
6–11 0.829*** 1.204*** 0.240** −0.124 −0.363*** −0.636*** 0.478*** 0.807*** 0.237*** 0.0360 −0.390*** −0.307**
(0.188) (0.206) (0.0977) (0.126) (0.119) (0.234) (0.126) (0.132) (0.0757) (0.0867) (0.0787) (0.138)
>11 0.660*** 1.021*** 1.164*** −0.552*** −0.188 −0.274 0.598*** 0.881*** 0.842*** −0.200* −0.422*** −0.358**
(0.200) (0.217) (0.127) (0.175) (0.163) (0.278) (0.125) (0.134) (0.0879) (0.113) (0.0905) (0.165)
Age (years)
26–35 0.936*** 1.936*** 0.00906 1.252*** 0.597*** 1.524*** 1.159*** 2.171*** 0.157 0.743*** 0.758*** 1.352***
(0.178) (0.244) (0.146) (0.256) (0.0681) (0.131) (0.156) (0.227) (0.114) (0.245) (0.0556) (0.117)
36–45 0.591*** 1.977*** −0.328** 1.421*** 0.482*** 1.888*** 0.619*** 2.028*** −0.0610 1.234*** 0.549*** 1.760***
(0.171) (0.238) (0.149) (0.258) (0.0988) (0.166) (0.150) (0.219) (0.111) (0.237) (0.0663) (0.123)
46–55 −0.270 1.263*** −0.881*** 1.221*** −0.0748 1.667*** −0.222 1.427*** −0.611*** 1.070*** 0.225** 1.778***
(0.182) (0.243) (0.160) (0.269) (0.140) (0.206) (0.140) (0.212) (0.118) (0.240) (0.0972) (0.147)
56–65 −1.997*** −0.121 −2.002*** 0.282 −2.409*** 0.237 −1.934*** 0.102 −1.705*** 0.240 −1.330*** 0.740***
(0.178) (0.240) (0.188) (0.285) (0.218) (0.300) (0.147) (0.214) (0.138) (0.249) (0.143) (0.184)
Female −1.169*** −0.649*** −0.972*** −0.108 −0.459*** −0.199** −1.020*** −0.563*** −1.214*** −0.394*** −0.531*** −0.326***
(0.111) (0.120) (0.141) (0.177) (0.0532) (0.0923) (0.0719) (0.0734) (0.0762) (0.0953) (0.0426) (0.0709)
table continues next page
159
160
Table 6A.6 Multinomial Logit on Occupational Choice of Working-Age Population, by Household Status, in Thailand, 2000 and 2009 (continued)
2000 2009
Head Spouse Other Head Spouse Other
Wage Self- Wage Self- Wage Self- Wage Self- Wage Self- Wage Self-
worker employed worker employed worker employed worker employed worker employed worker employed
Central −0.0226 0.0373 0.0240 0.0409 −0.404*** −0.153 0.278*** 0.302*** 0.109 0.248** 0.0418 0.0641
(0.125) (0.133) (0.107) (0.141) (0.0940) (0.167) (0.0865) (0.0923) (0.0788) (0.107) (0.0639) (0.112)
North 0.0286 0.0767 0.382*** 0.579*** −0.343*** 0.269 −0.188** 0.295*** −0.403*** 0.417*** −0.407*** 0.110
(0.129) (0.139) (0.114) (0.146) (0.102) (0.170) (0.0944) (0.0983) (0.0886) (0.109) (0.0735) (0.121)
Northeast −0.502*** −0.502*** −0.0292 0.249* −0.755*** −0.347** −0.690*** −0.0869 −0.469*** 0.255** −0.649*** −0.0379
(0.125) (0.133) (0.110) (0.144) (0.0956) (0.173) (0.0908) (0.0954) (0.0864) (0.109) (0.0693) (0.119)
South 0.0191 0.255* −0.358*** 0.372** −0.874*** −0.161 0.172 0.420*** −0.230** 0.518*** −0.105 0.339**
(0.140) (0.149) (0.116) (0.145) (0.102) (0.178) (0.108) (0.114) (0.0931) (0.116) (0.0801) (0.136)
Urban 0.107 0.842*** −0.345*** −0.0744 −0.162*** 0.109 0.0459 0.728*** −0.277*** 0.0603 −0.151*** 0.144**
(0.0766) (0.0855) (0.0686) (0.0821) (0.0534) (0.0936) (0.0569) (0.0597) (0.0498) (0.0587) (0.0420) (0.0686)
Head employed 0.575*** −0.277** 0.00257 −0.130 0.599*** −0.217*** −0.220*** −0.406***
(0.111) (0.119) (0.0620) (0.100) (0.0814) (0.0837) (0.0498) (0.0784)
Spouse employed 0.185*** 0.240* 0.276*** 0.146
(0.0710) (0.130) (0.0566) (0.105)
Married −0.660*** −0.630*** 0.111* 0.449*** −0.640*** −0.569*** 0.190*** 0.711***
(0.138) (0.155) (0.0574) (0.0986) (0.112) (0.116) (0.0470) (0.0734)
Attends school −2.980*** −2.716*** −0.649 0.399 −3.489*** −3.428*** −2.325*** −2.473*** −0.127 −2.850*** −3.016*** −3.195***
(0.235) (0.399) (0.546) (0.502) (0.110) (0.353) (0.283) (0.286) (0.271) (0.656) (0.0823) (0.254)
Number of children −0.371*** −0.167** −0.525*** −0.398*** −0.350*** −0.286*** −0.575*** −0.386***
< 5 years old (0.0668) (0.0689) (0.0582) (0.0689) (0.0550) (0.0562) (0.0488) (0.0574)
Constant 2.002*** −0.884*** 0.969*** −1.936*** 0.810*** −2.567*** 1.825*** −1.034*** 1.009*** −1.653*** 0.145 −2.818***
(0.226) (0.281) (0.249) (0.357) (0.141) (0.245) (0.178) (0.238) (0.176) (0.289) (0.113) (0.189)
Observations 15,221 15,221 11,386 11,386 19,603 19,603 25,341 25,341 19,176 19,176 31,564 31,564
Pseudo R2 0.163 0.163 0.0883 0.0883 0.266 0.266 0.145 0.145 0.106 0.106 0.262 0.262
Source: Household survey results from Thailand (SES 2000–09).
Note: “Working age” = 15–64 years; “Other” = household members who are neither heads of households nor spouses. Sample includes individuals who are nonfarmers in main occupation. The omitted occupational
category includes family workers, unemployed, and inactive individuals (that is, individuals with zero earnings). Standard errors in parentheses. SES = Household Socio-Economic Survey (Thailand National Statistical
Office).
*p < 0.1, **p < 0.05, ***p < 0.01
Understanding Poverty Reduction in Bangladesh, Peru, and Thailand 161
Table 6A.7 Earnings Regressions for Nonfarm Working-Age Population in Bangladesh, 2000 and 2010
2000 2010
Self- Self-
Daily workers employeda Wage worker Daily workers employeda Wage worker
a. Household heads
Primary & lower 0.132*** 0.370*** 0.330*** 0.0530*** 0.312*** 0.357***
secondary education (0.0392) (0.0484) (0.0499) (0.0150) (0.0387) (0.0387)
Higher secondary & 0.420** 0.994*** 0.796*** 0.248*** 0.632*** 0.826***
tertiary education (0.164) (0.0810) (0.0551) (0.0654) (0.0625) (0.0420)
Age 0.0510*** 0.0591*** 0.0529*** 0.0224*** 0.0411*** 0.0916***
(0.00989) (0.0186) (0.0161) (0.00426) (0.0149) (0.0117)
Age squared −0.000666*** −0.000664*** −0.000570*** −0.000267*** −0.000541*** −0.00107***
(0.000119) (0.000219) (0.000191) (5.09e-05) (0.000172) (0.000138)
Female −0.894*** −1.477*** −0.956*** −0.688*** −1.080*** −0.709***
(0.0621) (0.166) (0.0849) (0.0266) (0.138) (0.0609)
Urban 0.0863* 0.234*** 0.108** −0.0217 0.264*** 0.117***
(0.0447) (0.0495) (0.0421) (0.0164) (0.0396) (0.0315)
Barisal 0.0601 −0.193** −0.0421 0.0120 −0.0592 −0.0989
(0.0641) (0.0966) (0.0880) (0.0293) (0.0857) (0.0671)
Chittagong −0.0891* −0.0956 −0.0181 0.0488** −0.152** −0.0456
(0.0461) (0.0622) (0.0521) (0.0194) (0.0597) (0.0405)
Khulna 0.103** −0.108 −0.0193 −0.216*** −0.212*** −0.184***
(0.0478) (0.0747) (0.0677) (0.0194) (0.0621) (0.0506)
Rajsahi −0.182*** −0.242*** −0.195*** −0.179*** −0.314*** −0.112**
(0.0364) (0.0589) (0.0610) (0.0160) (0.0466) (0.0454)
Sylhet −0.0841 −0.183 −0.370*** −0.0994*** 0.302*** −0.0124
(0.0604) (0.131) (0.111) (0.0298) (0.0846) (0.0812)
Manufacturing 0.432*** 0.446*** 0.0674*** 0.107
(0.0456) (0.0780) (0.0196) (0.0699)
Industry 0.178*** 0.00876 0.433*** 0.272*** −0.0833 0.250**
(0.0547) (0.0962) (0.121) (0.0201) (0.146) (0.0969)
Services 0.429*** −0.0220 0.403*** 0.111*** −0.0844* 0.0882
(0.0379) (0.0472) (0.0740) (0.0156) (0.0475) (0.0670)
Public job 0.178*** 0.320***
(0.0452) (0.0393)
Constant 6.395*** 6.636*** 6.256*** 7.190*** 7.164*** 5.698***
(0.198) (0.386) (0.334) (0.0872) (0.317) (0.248)
Observations 2,092 1,269 1,134 3,390 2,007 1,820
R2 0.216 0.240 0.401 0.289 0.183 0.373
Adjusted R2 0.211 0.232 0.393 0.286 0.177 0.368
b. Other members
Primary & lower 0.133** 0.449*** 0.459*** 0.0492** 0.0246 0.292***
secondary education (0.0554) (0.0831) (0.0674) (0.0211) (0.0676) (0.0372)
Higher secondary & 0.114 0.975*** 0.837*** 0.466*** 0.331*** 0.954***
tertiary education (0.270) (0.124) (0.0854) (0.0992) (0.107) (0.0478)
Age 0.0514*** 0.0693*** 0.0478*** 0.0384*** 0.124*** 0.0693***
(0.0131) (0.0222) (0.0178) (0.00593) (0.0167) (0.00975)
Age squared −0.000633*** −0.000774** −0.000444 −0.000470*** −0.00170*** −0.000846***
(0.000203) (0.000350) (0.000282) (9.17e-05) (0.000246) (0.000153)
table continues next page
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
162 Understanding Poverty Reduction in Bangladesh, Peru, and Thailand
Table 6A.7 Earnings Regressions for Nonfarm Working-Age Population in Bangladesh, 2000 and 2010 (continued)
2000 2010
Self- Self-
Daily workers employeda Wage worker Daily workers employeda Wage worker
Female −1.078*** −1.244*** −0.682*** −0.691*** −1.132*** −0.462***
(0.0698) (0.121) (0.0656) (0.0301) (0.0841) (0.0331)
Urban −0.0135 0.0209 −0.147** −0.104*** 0.287*** −0.0407
(0.0737) (0.0803) (0.0619) (0.0276) (0.0615) (0.0316)
Barisal 0.216* −0.0186 −0.228* 0.0885* 0.156 −0.0396
(0.111) (0.151) (0.124) (0.0502) (0.127) (0.0752)
Chittagong −0.200*** −0.123 −0.178** 0.101*** −0.149* −0.0281
(0.0756) (0.101) (0.0704) (0.0321) (0.0883) (0.0388)
Khulna −0.107 −0.0433 −0.215* −0.204*** −0.297*** −0.237***
(0.0855) (0.121) (0.117) (0.0338) (0.0907) (0.0562)
Rajsahi −0.227*** −0.396*** −0.295*** −0.103*** −0.00253 −0.0741
(0.0665) (0.103) (0.0910) (0.0281) (0.0756) (0.0481)
Sylhet −0.231** −0.481*** −0.636*** 0.0448 0.125 −0.0857
(0.103) (0.166) (0.125) (0.0389) (0.112) (0.0741)
Manufacturing 0.300*** 0.463*** −0.169*** 0.149*
(0.0674) (0.145) (0.0274) (0.0904)
Industry 0.448*** −0.0366 0.440** 0.174*** 0.700*** 0.0853
(0.0897) (0.148) (0.221) (0.0296) (0.171) (0.126)
Services 0.469*** 0.0881 0.297** 0.0930*** 0.195*** −0.0847
(0.0692) (0.0777) (0.144) (0.0276) (0.0707) (0.0905)
Public job 0.555*** 0.437***
(0.0921) (0.0522)
Constant 6.270*** 6.379*** 6.198*** 6.974*** 5.456*** 6.227***
(0.192) (0.334) (0.285) (0.0892) (0.273) (0.163)
Observations 1,153 673 1,104 1,857 973 1,984
R2 0.307 0.317 0.285 0.383 0.343 0.348
Adjusted R2 0.299 0.304 0.275 0.379 0.334 0.343
Source: Household survey results from Bangladesh (HIES 2000–10).
Note: “Working-age” = 15–64 years. Dhaka is the base region. “Illiterate and incomplete primary” education is the base for educational levels.
“Agriculture” is the base sector for “Daily” and “wage” workers, while “Manufacturing” is the base sector for “self-employed.” Standard errors in
parentheses. HIES = Household Income and Expenditure Survey (Bangladesh Bureau of Statistics).
a. Nonagricultural self-employment.
*p < 0.1, **p < 0.05, ***p < 0.01
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Table 6A.8 Earnings Regressions for Nonfarm Working-Age Population in Peru, 2004 and 2010
2004 2010
Head Spouse Other Head Spouse Other
Salaried Self- Salaried Self- Salaried Self- Salaried Self- Salaried Self- Salaried Self-
job employed job employed job employed job employed job employed job employed
Education type
Primary 0.132*** 0.296*** 0.0984** 0.148*** 0.220*** 0.142***
(0.0383) (0.0637) (0.0417) (0.0367) (0.0623) (0.0457)
Secondary 0.350*** 0.558*** 0.238*** 0.271*** 0.519*** 0.278***
(0.0382) (0.0663) (0.0415) (0.0359) (0.0633) (0.0451)
College 0.837*** 0.937*** 0.570*** 0.627*** 0.825*** 0.578***
(0.0404) (0.0727) (0.0452) (0.0381) (0.0662) (0.0477)
Age (years)
26–35 0.243*** 0.314*** 0.280*** 0.291*** 0.321*** 0.389*** 0.169*** 0.305*** 0.249*** 0.457*** 0.283*** 0.556***
(0.0441) (0.0842) (0.0875) (0.0955) (0.0206) (0.0486) (0.0422) (0.0946) (0.0754) (0.110) (0.0198) (0.0539)
36–45 0.358*** 0.428*** 0.318*** 0.372*** 0.313*** 0.482*** 0.192*** 0.390*** 0.254*** 0.591*** 0.285*** 0.705***
(0.0436) (0.0818) (0.0876) (0.0946) (0.0307) (0.0634) (0.0413) (0.0922) (0.0731) (0.108) (0.0289) (0.0691)
46–55 0.405*** 0.277*** 0.357*** 0.329*** 0.396*** 0.339*** 0.167*** 0.253*** 0.241*** 0.554*** 0.251*** 0.545***
(0.0446) (0.0824) (0.0922) (0.0983) (0.0484) (0.0935) (0.0419) (0.0919) (0.0759) (0.110) (0.0462) (0.0899)
56–65 0.305*** 0.0801 0.541*** 0.128 0.548*** 0.220 0.0896** −0.00376 0.405*** 0.418*** 0.223** 0.711***
(0.0504) (0.0845) (0.107) (0.106) (0.0968) (0.156) (0.0456) (0.0940) (0.0907) (0.116) (0.0994) (0.133)
Female −0.427*** −0.582*** −0.166*** −0.601*** −0.435*** −0.781*** −0.236*** −0.859***
(0.0287) (0.0338) (0.0182) (0.0416) (0.0235) (0.0337) (0.0172) (0.0461)
Informal −0.373*** −0.554*** −0.539*** −0.568*** −0.336*** −0.538*** −0.383*** −0.596*** −0.731*** −0.829*** −0.373*** −0.580***
(0.0248) (0.0323) (0.0504) (0.0592) (0.0194) (0.0505) (0.0229) (0.0317) (0.0425) (0.0554) (0.0187) (0.0545)
Manufacturing 0.0471 0.327*** 0.299*** −0.00185 0.497*** 0.153***
(0.0383) (0.0924) (0.0347) (0.0371) (0.0856) (0.0339)
Services −0.0832*** −0.00266 0.402*** 0.855*** 0.266*** 0.478*** 0.0911*** 0.113*** 0.599*** 0.723*** 0.209*** 0.534***
(0.0322) (0.0396) (0.0770) (0.0569) (0.0297) (0.0648) (0.0312) (0.0414) (0.0708) (0.0575) (0.0287) (0.0713)
Public sector −0.188*** 0.352*** 0.337*** −0.0215 0.460*** 0.287***
(0.0357) (0.0804) (0.0387) (0.0351) (0.0736) (0.0368)
table continues next page
163
164
Table 6A.8 Earnings Regressions for Nonfarm Working-Age Population in Peru, 2004 and 2010 (continued)
2004 2010
Head Spouse Other Head Spouse Other
Salaried Self- Salaried Self- Salaried Self- Salaried Self- Salaried Self- Salaried Self-
job employed job employed job employed job employed job employed job employed
Costa −0.360*** −0.491*** −0.392*** −0.575*** −0.459*** −0.682*** −0.233*** −0.234*** −0.347*** −0.176*** −0.255*** −0.298***
(0.0257) (0.0345) (0.0502) (0.0526) (0.0222) (0.0501) (0.0231) (0.0353) (0.0457) (0.0543) (0.0219) (0.0563)
Sierra −0.342*** −0.632*** −0.570*** −0.563*** −0.599*** −0.730*** −0.233*** −0.212*** −0.486*** −0.305*** −0.370*** −0.386***
(0.0283) (0.0393) (0.0537) (0.0553) (0.0257) (0.0612) (0.0254) (0.0391) (0.0494) (0.0581) (0.0241) (0.0676)
Selva −0.452*** −0.521*** −0.409*** −0.490*** −0.442*** −0.645*** −0.209*** −0.201*** −0.444*** −0.0545 −0.230*** −0.232***
(0.0381) (0.0501) (0.0667) (0.0686) (0.0339) (0.0784) (0.0347) (0.0512) (0.0638) (0.0703) (0.0315) (0.0787)
Rural −0.186*** −0.200*** −0.104 −0.496*** −0.0519* −0.524*** −0.208*** −0.196*** −0.196*** −0.513*** −0.112*** −0.348***
(0.0339) (0.0595) (0.0675) (0.0540) (0.0289) (0.0722) (0.0318) (0.0575) (0.0629) (0.0573) (0.0259) (0.0739)
Has two jobs 0.243*** 0.229*** 0.324*** 0.539*** 0.316*** 0.448*** 0.0652*** 0.190*** 0.163*** 0.403*** 0.111*** 0.361***
(0.0284) (0.0435) (0.0575) (0.0601) (0.0317) (0.0700) (0.0228) (0.0387) (0.0454) (0.0543) (0.0262) (0.0637)
Head/spouse with at 0.201*** 0.156*** −0.241*** 0.0844*** 0.103** −0.266***
least secondary (0.0312) (0.0432) (0.0457) (0.0327) (0.0466) (0.0477)
At least one family 0.307*** 0.160*** −0.00777 0.246*** 0.283*** −0.00123
worker (0.0525) (0.0535) (0.0601) (0.0554) (0.0552) (0.0652)
At least secondary 0.178*** 0.410***
(0.0482) (0.0540)
Constant 6.248*** 6.709*** 5.268*** 5.133*** 5.813*** 5.949*** 6.495*** 6.764*** 5.462*** 5.114*** 5.992*** 5.530***
(0.0605) (0.0954) (0.115) (0.127) (0.0484) (0.0978) (0.0592) (0.107) (0.105) (0.137) (0.0506) (0.109)
Observations 4,758 3,651 1,665 3,022 6,374 2,375 5,483 4,250 2,326 3,586 7,963 2,817
R2 0.370 0.300 0.444 0.264 0.414 0.380 0.271 0.265 0.437 0.198 0.300 0.327
Sigma 0.650 0.846 0.761 1.049 0.722 1.025 0.649 0.921 0.801 1.206 0.764 1.162
Source: Household survey results from Peru (ENAHO 2005–09).
Note: “Working age” = 15–64 years. Standard errors in parentheses. “Other” = those who are neither household heads nor spouses. ENAHO = Encuesta Nacional de Hogares (Peru National Institute of Statistics and
Information).
*p < 0.1, **p < 0.05, ***p < 0.01.
Table 6A.9 Earnings Regressions for Nonfarm Working-Age Population in Thailand, 2000 and 2009
2000 2009
Head Spouse Others Head Spouse Others
Wage Self- Wage Self- Wage Self- Wage Self- Wage Self- Wage Self-
worker employed worker employed worker employed worker employed worker employed worker employed
Education (years)
6–11 yrs 0.375*** 0.289*** 0.330*** 0.240*** 0.428*** 0.378*** 0.369*** 0.129*
(0.0240) (0.0365) (0.0321) (0.0788) (0.0176) (0.0243) (0.0264) (0.0716)
>11 yrs 0.944*** 1.104*** 0.783*** 0.454*** 1.089*** 1.058*** 0.888*** 0.376***
(0.0256) (0.0413) (0.0348) (0.0863) (0.0180) (0.0261) (0.0269) (0.0733)
Head of household education (years)
6–11 0.185*** 0.171*** 0.0273 0.211*** 0.194*** 0.0753
(0.0387) (0.0558) (0.0969) (0.0306) (0.0459) (0.0654)
>11 0.534*** 0.316*** 0.291** 0.508*** 0.321*** 0.0711
(0.0493) (0.0733) (0.117) (0.0355) (0.0530) (0.0767)
Spouse education (years)
<6 0.108** 0.105* 0.177*** −0.00364
(0.0480) (0.0592) (0.0357) (0.0459)
6–11 0.285*** 0.0293 0.287* 0.270*** 0.154*** −0.114
(0.0523) (0.0637) (0.150) (0.0351) (0.0488) (0.0899)
>11 0.460*** 0.434*** 0.0742 0.344*** 0.267*** −0.0660
(0.0631) (0.0942) (0.168) (0.0384) (0.0612) (0.107)
Age (years)
26–35 0.304*** 0.140 0.211*** −0.0875 0.341*** 0.485*** 0.259*** 0.205** 0.209*** 1.186*** 0.391*** 0.367***
(0.0326) (0.108) (0.0451) (0.168) (0.0210) (0.0771) (0.0258) (0.102) (0.0361) (0.146) (0.0144) (0.0671)
36–45 0.534*** 0.277*** 0.341*** 0.203 0.497*** 0.745*** 0.457*** 0.309*** 0.361*** 1.331*** 0.573*** 0.592***
(0.0337) (0.107) (0.0500) (0.169) (0.0322) (0.0901) (0.0252) (0.0991) (0.0358) (0.140) (0.0181) (0.0691)
46–55 0.597*** 0.414*** 0.440*** 0.199 0.612*** 0.696*** 0.656*** 0.244** 0.577*** 1.364*** 0.699*** 0.508***
(0.0367) (0.109) (0.0576) (0.174) (0.0578) (0.126) (0.0266) (0.0999) (0.0391) (0.144) (0.0292) (0.0885)
56–65 0.395*** 0.0838 0.141* 0.0322 0.348*** −0.0913 0.541*** −0.00722 0.487*** 1.169*** 0.643*** 0.185
(0.0436) (0.113) (0.0746) (0.182) (0.130) (0.183) (0.0311) (0.102) (0.0489) (0.150) (0.0597) (0.126)
table continues next page
165
166
Table 6A.9 Earnings Regressions for Nonfarm Working-Age Population in Thailand, 2000 and 2009 (continued)
2000 2009
Head Spouse Others Head Spouse Others
Wage Self- Wage Self- Wage Self- Wage Self- Wage Self- Wage Self-
worker employed worker employed worker employed worker employed worker employed worker employed
Female −0.283*** −0.271*** −0.246*** −0.291*** −0.0594*** −0.0789 −0.242*** −0.254*** −0.243*** −0.278*** −0.0650*** −0.166***
(0.0193) (0.0448) (0.0443) (0.109) (0.0191) (0.0568) (0.0130) (0.0286) (0.0207) (0.0541) (0.0124) (0.0417)
Services 1.159*** 0.0232 1.136*** 0.665*** 1.026*** 0.357*** 0.635*** −0.0124 0.649*** 0.613*** 0.518*** 0.314***
(0.0284) (0.0407) (0.0402) (0.0573) (0.0315) (0.0703) (0.0223) (0.0299) (0.0307) (0.0456) (0.0257) (0.0533)
Central −0.0730*** −0.102** −0.0775* −0.254*** −0.115*** −0.268*** −0.196*** −0.206*** −0.149*** −0.131** −0.278*** −0.393***
(0.0253) (0.0438) (0.0412) (0.0842) (0.0313) (0.0978) (0.0183) (0.0353) (0.0280) (0.0660) (0.0210) (0.0765)
North −0.578*** −0.558*** −0.646*** −0.674*** −0.696*** −0.420*** −0.499*** −0.403*** −0.598*** −0.507*** −0.629*** −0.750***
(0.0306) (0.0525) (0.0481) (0.0845) (0.0360) (0.105) (0.0232) (0.0405) (0.0338) (0.0692) (0.0247) (0.0814)
Northeast −0.694*** −0.630*** −0.692*** −0.690*** −0.786*** −0.726*** −0.611*** −0.411*** −0.627*** −0.436*** −0.694*** −0.662***
(0.0292) (0.0490) (0.0468) (0.0838) (0.0346) (0.102) (0.0217) (0.0373) (0.0319) (0.0673) (0.0227) (0.0773)
South −0.247*** −0.248*** −0.305*** −0.326*** −0.379*** −0.172 −0.279*** −0.237*** −0.237*** −0.236*** −0.500*** −0.342***
(0.0322) (0.0536) (0.0515) (0.0893) (0.0401) (0.116) (0.0228) (0.0414) (0.0340) (0.0710) (0.0251) (0.0853)
Urban 0.148*** 0.346*** 0.141*** 0.117** 0.0608** 0.157** 0.268*** 0.410*** 0.247*** 0.360*** 0.254*** 0.325***
(0.0202) (0.0337) (0.0311) (0.0516) (0.0237) (0.0635) (0.0140) (0.0254) (0.0205) (0.0399) (0.0151) (0.0488)
Manufacturing 1.067*** 1.041*** 1.071*** 0.636*** 0.612*** 0.482***
(0.0261) (0.0365) (0.0289) (0.0214) (0.0295) (0.0249)
Public sector 1.489*** 1.458*** 1.269*** 0.976*** 1.116*** 0.775***
(0.0302) (0.0478) (0.0380) (0.0242) (0.0356) (0.0292)
Constant 7.168*** 8.545*** 7.183*** 8.228*** 7.107*** 7.745*** 7.486*** 8.633*** 7.498*** 6.876*** 7.480*** 8.219***
(0.0454) (0.127) (0.0810) (0.218) (0.0491) (0.156) (0.0368) (0.111) (0.0561) (0.169) (0.0393) (0.129)
Observations 8,251 4,216 4,138 1,929 7,755 1,350 13,016 7,564 7,216 3,421 13,185 2,587
R2 0.624 0.236 0.627 0.203 0.473 0.176 0.537 0.213 0.522 0.210 0.438 0.175
Sigma 0.725 0.959 0.762 0.975 0.816 1.005 0.667 0.933 0.700 0.978 0.687 1.001
Source: Household survey results from Thailand (SES 2000–09).
Note: “Working age” = 15–64 years; “Others” = those who are neither heads of households nor spouses. Standard errors in parentheses. SES = Household Socio-Economic Survey (Thailand National Statistical Office).
*p < 0.1, **p < 0.05, ***p < 0.01.
Understanding Poverty Reduction in Bangladesh, Peru, and Thailand 167
Table 6A.10 Net Revenue Regressions for Farm Households in Bangladesh, 2000 and 2010
2000 2010
Primary & lower secondary education 0.132** (0.0633) 0.0331 (0.0489)
Higher secondary & tertiary education 0.295** (0.147) −0.00290 (0.119)
Age 0.142*** (0.0214) 0.0795*** (0.0172)
Age squared −0.00175*** (0.000245) −0.000843*** (0.000190)
Female −2.243*** (0.123) −0.926*** (0.0605)
Urban 0.197 (0.123) −0.161** (0.0816)
Barisal −0.285** (0.123) 0.270*** (0.0920)
Chittagong −0.240** (0.0977) 0.164** (0.0680)
Khulna 0.405*** (0.0955) 0.204*** (0.0751)
Rajshahi 0.0125 (0.0763) 0.303*** (0.0629)
Sylhet 0.0104 (0.126) 0.253** (0.108)
Land (low) −0.107 (0.0986) 0.743*** (0.0666)
Land (high) 0.457*** (0.111) 1.333*** (0.0786)
Irrigation 0.458*** (0.0941) 0.408*** (0.0707)
Household members (no.) 0.525*** (0.0507) 0.429*** (0.0525)
Constant 4.007*** (0.462) 4.780*** (0.381)
Observations 1,678 2,956
R2 0.340 0.323
Adjusted R2 0.334 0.320
Source: Household survey results from Bangladesh (HIES 2000–10).
Note: Results represent net revenue for farm households in Bangladesh. Net revenue was calculated using the available
information on total revenue stemming from agricultural production and the cost of inputs. The sample comprises those
farm household members of working age (15–64 years) who are self-employed in agriculture. Dhaka is the base region.
“Illiterate and incomplete primary” is the base for education. “No-land” is the base category for land tenure. Standard errors
appear in parentheses. HIES = Household Income and Expenditure Survey (Bangladesh Bureau of Statistics).
*p < 0.1, **p < 0.05, ***p < 0.01.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
168 Understanding Poverty Reduction in Bangladesh, Peru, and Thailand
Table 6A.11 Net Revenue Regressions for Farm Households in Peru, 2004 and 2010
2004 2010
Head/spouse with at least secondary 0.133*** (0.0296) 0.128*** (0.0273)
Age (years)
26–35 0.267*** (0.0610) 0.231*** (0.0736)
36–45 0.412*** (0.0598) 0.333*** (0.0722)
46–55 0.473*** (0.0600) 0.272*** (0.0724)
56–65 0.373*** (0.0607) 0.214*** (0.0736)
Female −0.420*** (0.0349) −0.385*** (0.0318)
Head farmer 0.421*** (0.0409) 0.304*** (0.0351)
More than 1 farmer in household 0.688*** (0.0571) 0.513*** (0.0453)
Land size (ha)
0.5–2 −0.0127 (0.0377) 0.437*** (0.0281)
2–5 0.288*** (0.0443) 0.799*** (0.0348)
>5 0.604*** (0.0468) 1.055*** (0.0371)
Share of adults >50 percent 0.0476* (0.0254) 0.0206 (0.0251)
Improved irrigation system 0.130*** (0.0340) 0.177*** (0.0249)
Owns land −0.218*** (0.0336) 0.0647** (0.0252)
Sierra −0.151*** (0.0336) 0.00957 (0.0326)
Selva 0.0414 (0.0404) −0.0303 (0.0422)
Rural 0.0757** (0.0302) 0.0935*** (0.0303)
Constant 4.249*** (0.0784) 4.182*** (0.0845)
Observations 6,870 7,117
R2 0.123 0.238
Sigma 0.922 0.894
Source: Household survey results from Peru (ENAHO 2005–09).
Note: ha = hectares (1 hectare = about 100 acres or 10,000 square meters). Results represent net revenue for farm households
in Bangladesh. Net revenue was calculated using the available information on total revenue stemming from agricultural
production and the cost of inputs. Standard errors in parentheses. “Other” = those who are neither household heads nor
spouses. ENAHO = Encuesta Nacional de Hogares (Peru National Institute of Statistics and Information).
*p < 0.1, **p < 0.05, ***p < 0.01.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Understanding Poverty Reduction in Bangladesh, Peru, and Thailand 169
Table 6A.12 Net Revenue Regressions for Farm Households in Thailand, 2000 and 2009
2000 2009
Head of household education (years)
6–11 0.0575 (0.0556) 0.0474 (0.0387)
>11 0.273*** (0.105) 0.0814 (0.0565)
Spouse education (years)
<6 0.189** (0.0742) 0.343*** (0.0429)
6–11 0.271*** (0.0890) 0.421*** (0.0497)
>11 0.743*** (0.154) 0.629*** (0.0724)
Age (years)
26–35 0.177 (0.167) 0.369** (0.170)
36–45 0.426** (0.166) 0.573*** (0.165)
46–55 0.455*** (0.169) 0.612*** (0.167)
56–65 0.545*** (0.170) 0.512*** (0.168)
Female −0.334*** (0.0761) −0.158*** (0.0388)
North −0.816*** (0.0607) −0.512*** (0.0443)
Northeast −1.078*** (0.0556) −1.104*** (0.0413)
South −0.0996 (0.0675) 0.220*** (0.0508)
Urban −0.171*** (0.0651) −0.0914* (0.0485)
Household owns land 0.132 (0.101) 0.167*** (0.0585)
Household members older than 18 years of age (%) −0.161* (0.0947) −0.0815 (0.0686)
Constant 7.392*** (0.205) 7.394*** (0.177)
Observations 5,801 10,118
R2 0.119 0.151
Sigma 1.393 1.348
Source: Household survey results from Thailand (SES 2000–09).
Note: HH = household. Standard errors in parentheses. Figure results represent net revenue for farm households in Thailand.
Net revenue was calculated using the available information on total revenue stemming from agricultural production and the
cost of inputs. Unfortunately, unlike the surveys for Bangladesh and Peru, the household survey for Thailand does not contain
information on landholding and access to irrigation. SES = Household Socio-Economic Survey (Thailand National Statistical
Office).
*p < 0.1, **p < 0.05, ***p < 0.01.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
170 Understanding Poverty Reduction in Bangladesh, Peru, and Thailand
Notes
1. As first noted in chapter 1, “moderate poverty” is a country-specific poverty line refer-
ring to the international poverty line that is closest to the country’s moderate poverty
rate (in some cases $1.25 a day, in others $2.50; and in still others $4–$5 per day).
2. The widely used “cost of basic needs” approach to drawing consumption-based pov-
erty lines “first estimates the cost of acquiring enough food for adequate nutrition—
usually 2,100 calories per person per day—and then adds the cost of other essentials
such as clothing and shelter” (Haughton and Khandker, 2009, 39).
3. Despite this deceleration of population growth, Bangladesh has added 19 billion
people to its total, a 15 percent increase between 2000 and 2010. Over the same
period, Peru has added 3.2 million (a 12 percent increase) and Thailand 6 million
(a 9 percent increase).
4. Net revenue was calculated using the available information on total revenue stem-
ming from agricultural production and the cost of inputs.
5. Unfortunately, the household survey for Thailand does not contain information on
landholding and access to irrigation.
6. As mentioned earlier, this decomposition method cannot identify the reason behind
this change. However, one possible theory is that improved road networks reduced
transportation costs and thus allowed for greater returns in investing outside the
capital. An alternative explanation is that internal migration toward the capital led to
a greater scarcity of workers in other regions, and therefore relatively better earnings.
Given the size of this effect, further research on this would be useful to identify the
source of change.
7. The cumulative effects are larger because they include all occupational changes,
including the choice between daily, wage, and self-employed work for nonfarm
workers, and the choice to have a secondary occupation for farm workers.
8. Note that this is consistent with the findings using the simple approach (see table 6.5).
9. Similar analysis about the returns to agriculture in Thailand was not possible because
of lack of similar variables in the database used.
Bibliography
Bourguignon, François, Francisco H. G. Ferreira, and Nora Lustig, eds. 2005. The
Microeconomics of Income Distribution Dynamics in East Asia and Latin America.
Washington, DC: World Bank.
Bourguignon, François, Francisco H. G. Ferreira, and Phillippe G. Leite. 2008. “Beyond
Oaxaca-Blinder: Accounting for Differences in Household Income Distributions.”
Journal of Economic Inequality 6 (2): 117–48.
Datt, Gaurav, and Martin Ravallion. 1992. “Growth and Redistribution Components of
Changes in Poverty Measures: A Decomposition with Applications to Brazil and India
in the 1980s.” Journal of Development Economics 38 (2): 275–96.
Davis, Benjamin, Paul Winters, Calogero Carletto, Katia Covarrubias, Esteban J. Quiñones,
Alberto Zezza, Carlo Azzarri, Kostas Stamoulis, and Stefania DiGiuseppe. 2010.
“A Cross-Country Comparison of Rural Income Generating Activities.” World
Development 38 (1): 48–63.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Understanding Poverty Reduction in Bangladesh, Peru, and Thailand 171
Ferreira, Francisco H. G. 2010. “Distributions in Motion: Economic Growth, Inequality
and Poverty Dynamics.” Policy Research Working Paper 5424, World Bank.
———. 2012. “Distributions in Motion: Economic Growth, Inequality, and Poverty
Dynamics.” In The Oxford Handbook of the Economics of Poverty, edited by Philip N.
Jefferson, 427–62. New York: Oxford University Press.
Foster, James, Joel Greer, and Erik Thorbecke. 1984. “A Class of Decomposable Poverty
Measures.” Econometrica 52 (3): 761–66.
Government of Thailand. 2003. “Thailand’s Official Poverty Lines.” National Economic
and Social Development Board, Government of Thailand, Bangkok.
Haughton, Jonathan, and Shahidur R. Khandker. 2009. Handbook on Poverty + Inequality.
World Bank Training Series. Washington, DC: World Bank.
Kakwani, Nanak, and Medhi Krongkaew. 1998. “Poverty in Thailand: Defining, Measuring
and Analyzing.” Working Paper N4, Office of the National Economic and Social
Development Board, Development Evaluation Division, Bangkok.
Ravallion, Martin, and Shaohua Chen. 2007. “China’s (Uneven) Progress against Poverty.”
Journal of Development Economics 82 (1): 1–42.
Shapley, Lloyd. S. 1953. “A Value for n-Person Games.” In Contributions to the Theory
of Games, Vol. 2, edited by H. W. Kuhn and A. W. Tucker, 307–17. Princeton,
NJ: Princeton University Press.
Shorrocks, Anthony F. 1999. “Decomposition Procedures for Distributional Analysis:
A Unified Framework Based on Shapley Value.” University of Essex and Institute for
Fiscal Studies, London.
Tinbergen, Jan. 1975. Income Differences: Recent Research. Amsterdam, the Netherlands:
North-Holland.
World Bank. 2011. World Development Indicators 2011. Washington, DC: World Bank.
Online database.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
Environmental Benefits Statement
The World Bank Group is committed to reducing its environmental footprint.
In support of this commitment, the Publishing and Knowledge Division lever-
ages electronic publishing options and print-on-demand technology, which is
located in regional hubs worldwide. Together, these initiatives enable print runs
to be lowered and shipping distances decreased, resulting in reduced paper
consumption, chemical use, greenhouse gas emissions, and waste.
The Publishing and Knowledge Division follows the recommended standards
for paper use set by the Green Press Initiative. Whenever possible, books are
printed on 50 percent to 100 percent postconsumer recycled paper, and at least
50 percent of the fiber in our book paper is either unbleached or bleached using
Totally Chlorine Free (TCF), Processed Chlorine Free (PCF), or Enhanced
Elemental Chlorine Free (EECF) processes.
More information about the Bank’s environmental philosophy can be found
at http://crinfo.worldbank.org/wbcrinfo/node/4.
Understanding Changes in Poverty • http://dx.doi.org/10.1596/978-1-4648-0299-7
The 2015 Millennium Development Goal to reduce by 50 percent the share of the world’s population living
in extreme poverty was met early. The number of individuals in developing countries who live in extreme
poverty had decreased from 43 percent in 1990 to 21 percent by 2010. Yet, with 1.2 billion people still
struggling today, we have a long way to go. What can we learn from the recent success of reducing extreme
poverty?
Understanding Changes in Poverty brings together different methods to decompose the contributions to
poverty reduction. A simple approach quantifies the contribution of changes in demographics, employment,
earnings, public transfers, and remittances to poverty reduction. A more complex approach quantifies the
contributions to poverty reduction from changes in individual and household characteristics, including
changes in the sectoral, occupational, and educational structure of the workforce, as well as changes in
the returns to individual and household characteristics.
Understanding Changes in Poverty implements these approaches and finds that labor income growth—that
is, growth in income per worker rather than an increase in the number of employed workers—was the
largest contributor to moderate poverty reduction in 21 countries experiencing substantial reductions in
poverty over the past decade. Changes in demographics, public transfers, and remittances helped, but made
relatively smaller contributions to poverty reduction. Further decompositions in three countries find that
labor income grew mainly because of higher returns to human capital endowments, signaling increases in
productivity, higher relative price of labor, or both.
Understanding Changes in Poverty will be of particular relevance to development practitioners interested in
better understanding distributional changes over time. The methods and tools presented in this book can
also be applied to better understand changes in inequality or any other distributional change.
ISBN 978-1-4648-0299-7
SKU 210299